1 //===-- ARM64ISelLowering.cpp - ARM64 DAG Lowering Implementation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the ARM64TargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "arm64-lower"
16 #include "ARM64ISelLowering.h"
17 #include "ARM64PerfectShuffle.h"
18 #include "ARM64Subtarget.h"
19 #include "ARM64CallingConv.h"
20 #include "ARM64MachineFunctionInfo.h"
21 #include "ARM64TargetMachine.h"
22 #include "ARM64TargetObjectFile.h"
23 #include "MCTargetDesc/ARM64AddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Type.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 STATISTIC(NumTailCalls, "Number of tail calls");
40 STATISTIC(NumShiftInserts, "Number of vector shift inserts");
42 // This option should go away when tail calls fully work.
44 EnableARM64TailCalls("arm64-tail-calls", cl::Hidden,
45 cl::desc("Generate ARM64 tail calls (TEMPORARY OPTION)."),
49 StrictAlign("arm64-strict-align", cl::Hidden,
50 cl::desc("Disallow all unaligned memory accesses"));
52 // Place holder until extr generation is tested fully.
54 EnableARM64ExtrGeneration("arm64-extr-generation", cl::Hidden,
55 cl::desc("Allow ARM64 (or (shift)(shift))->extract"),
59 EnableARM64SlrGeneration("arm64-shift-insert-generation", cl::Hidden,
60 cl::desc("Allow ARM64 SLI/SRI formation"),
63 //===----------------------------------------------------------------------===//
64 // ARM64 Lowering public interface.
65 //===----------------------------------------------------------------------===//
66 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
67 if (TM.getSubtarget<ARM64Subtarget>().isTargetDarwin())
68 return new ARM64_MachoTargetObjectFile();
70 return new ARM64_ELFTargetObjectFile();
73 ARM64TargetLowering::ARM64TargetLowering(ARM64TargetMachine &TM)
74 : TargetLowering(TM, createTLOF(TM)) {
75 Subtarget = &TM.getSubtarget<ARM64Subtarget>();
77 // ARM64 doesn't have comparisons which set GPRs or setcc instructions, so
78 // we have to make something up. Arbitrarily, choose ZeroOrOne.
79 setBooleanContents(ZeroOrOneBooleanContent);
80 // When comparing vectors the result sets the different elements in the
81 // vector to all-one or all-zero.
82 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
84 // Set up the register classes.
85 addRegisterClass(MVT::i32, &ARM64::GPR32allRegClass);
86 addRegisterClass(MVT::i64, &ARM64::GPR64allRegClass);
87 addRegisterClass(MVT::f32, &ARM64::FPR32RegClass);
88 addRegisterClass(MVT::f64, &ARM64::FPR64RegClass);
89 addRegisterClass(MVT::f128, &ARM64::FPR128RegClass);
90 addRegisterClass(MVT::v16i8, &ARM64::FPR8RegClass);
91 addRegisterClass(MVT::v8i16, &ARM64::FPR16RegClass);
93 // Someone set us up the NEON.
94 addDRTypeForNEON(MVT::v2f32);
95 addDRTypeForNEON(MVT::v8i8);
96 addDRTypeForNEON(MVT::v4i16);
97 addDRTypeForNEON(MVT::v2i32);
98 addDRTypeForNEON(MVT::v1i64);
99 addDRTypeForNEON(MVT::v1f64);
101 addQRTypeForNEON(MVT::v4f32);
102 addQRTypeForNEON(MVT::v2f64);
103 addQRTypeForNEON(MVT::v16i8);
104 addQRTypeForNEON(MVT::v8i16);
105 addQRTypeForNEON(MVT::v4i32);
106 addQRTypeForNEON(MVT::v2i64);
108 // Compute derived properties from the register classes
109 computeRegisterProperties();
111 // Provide all sorts of operation actions
112 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
113 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
114 setOperationAction(ISD::SETCC, MVT::i32, Custom);
115 setOperationAction(ISD::SETCC, MVT::i64, Custom);
116 setOperationAction(ISD::SETCC, MVT::f32, Custom);
117 setOperationAction(ISD::SETCC, MVT::f64, Custom);
118 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
119 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
120 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
121 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
122 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
123 setOperationAction(ISD::SELECT, MVT::i32, Custom);
124 setOperationAction(ISD::SELECT, MVT::i64, Custom);
125 setOperationAction(ISD::SELECT, MVT::f32, Custom);
126 setOperationAction(ISD::SELECT, MVT::f64, Custom);
127 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
128 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
129 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
130 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
131 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
132 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
134 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
135 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
136 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
138 setOperationAction(ISD::FREM, MVT::f32, Expand);
139 setOperationAction(ISD::FREM, MVT::f64, Expand);
140 setOperationAction(ISD::FREM, MVT::f80, Expand);
142 // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
143 // silliness like this:
144 setOperationAction(ISD::FABS, MVT::v1f64, Expand);
145 setOperationAction(ISD::FADD, MVT::v1f64, Expand);
146 setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
147 setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
148 setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
149 setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
150 setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
151 setOperationAction(ISD::FMA, MVT::v1f64, Expand);
152 setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
153 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
154 setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
155 setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
156 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
157 setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
158 setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
159 setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
160 setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
161 setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
162 setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
163 setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
164 setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
165 setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
166 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
167 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
168 setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);
170 setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
171 setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
172 setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
173 setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
174 setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);
176 // Custom lowering hooks are needed for XOR
177 // to fold it into CSINC/CSINV.
178 setOperationAction(ISD::XOR, MVT::i32, Custom);
179 setOperationAction(ISD::XOR, MVT::i64, Custom);
181 // Virtually no operation on f128 is legal, but LLVM can't expand them when
182 // there's a valid register class, so we need custom operations in most cases.
183 setOperationAction(ISD::FABS, MVT::f128, Expand);
184 setOperationAction(ISD::FADD, MVT::f128, Custom);
185 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
186 setOperationAction(ISD::FCOS, MVT::f128, Expand);
187 setOperationAction(ISD::FDIV, MVT::f128, Custom);
188 setOperationAction(ISD::FMA, MVT::f128, Expand);
189 setOperationAction(ISD::FMUL, MVT::f128, Custom);
190 setOperationAction(ISD::FNEG, MVT::f128, Expand);
191 setOperationAction(ISD::FPOW, MVT::f128, Expand);
192 setOperationAction(ISD::FREM, MVT::f128, Expand);
193 setOperationAction(ISD::FRINT, MVT::f128, Expand);
194 setOperationAction(ISD::FSIN, MVT::f128, Expand);
195 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
196 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
197 setOperationAction(ISD::FSUB, MVT::f128, Custom);
198 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
199 setOperationAction(ISD::SETCC, MVT::f128, Custom);
200 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
201 setOperationAction(ISD::SELECT, MVT::f128, Custom);
202 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
203 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
205 // Lowering for many of the conversions is actually specified by the non-f128
206 // type. The LowerXXX function will be trivial when f128 isn't involved.
207 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
208 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
209 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
210 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
211 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
212 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
213 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
214 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
215 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
216 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
217 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
218 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
219 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
220 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
223 setOperationAction(ISD::ATOMIC_SWAP, MVT::i128, Custom);
224 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i128, Custom);
225 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
226 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i128, Custom);
227 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i128, Custom);
228 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i128, Custom);
229 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i128, Custom);
230 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i128, Custom);
231 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i128, Custom);
232 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i128, Custom);
233 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i128, Custom);
234 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i128, Custom);
235 // These are surprisingly difficult. The only single-copy atomic 128-bit
236 // instruction on AArch64 is stxp (when it succeeds). So a store can safely
237 // become a simple swap, but a load can only be determined to have been atomic
238 // if storing the same value back succeeds.
239 setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
240 setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Expand);
242 // Variable arguments.
243 setOperationAction(ISD::VASTART, MVT::Other, Custom);
244 setOperationAction(ISD::VAARG, MVT::Other, Custom);
245 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
246 setOperationAction(ISD::VAEND, MVT::Other, Expand);
248 // Variable-sized objects.
249 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
250 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
251 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
253 // Exception handling.
254 // FIXME: These are guesses. Has this been defined yet?
255 setExceptionPointerRegister(ARM64::X0);
256 setExceptionSelectorRegister(ARM64::X1);
258 // Constant pool entries
259 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
262 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
264 // Add/Sub overflow ops with MVT::Glues are lowered to CPSR dependences.
265 setOperationAction(ISD::ADDC, MVT::i32, Custom);
266 setOperationAction(ISD::ADDE, MVT::i32, Custom);
267 setOperationAction(ISD::SUBC, MVT::i32, Custom);
268 setOperationAction(ISD::SUBE, MVT::i32, Custom);
269 setOperationAction(ISD::ADDC, MVT::i64, Custom);
270 setOperationAction(ISD::ADDE, MVT::i64, Custom);
271 setOperationAction(ISD::SUBC, MVT::i64, Custom);
272 setOperationAction(ISD::SUBE, MVT::i64, Custom);
274 // ARM64 lacks both left-rotate and popcount instructions.
275 setOperationAction(ISD::ROTL, MVT::i32, Expand);
276 setOperationAction(ISD::ROTL, MVT::i64, Expand);
278 // ARM64 doesn't have a direct vector ->f32 conversion instructions for
279 // elements smaller than i32, so promote the input to i32 first.
280 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
281 setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
282 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
283 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
284 // Similarly, there is no direct i32 -> f64 vector conversion instruction.
285 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
286 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
287 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
288 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
290 // ARM64 doesn't have {U|S}MUL_LOHI.
291 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
292 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
294 // ARM64 doesn't have MUL.2d:
295 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
297 // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
298 // counterparts, which ARM64 supports directly.
299 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
300 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
301 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
302 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
304 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
305 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
307 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
308 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
309 setOperationAction(ISD::SREM, MVT::i32, Expand);
310 setOperationAction(ISD::SREM, MVT::i64, Expand);
311 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
312 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
313 setOperationAction(ISD::UREM, MVT::i32, Expand);
314 setOperationAction(ISD::UREM, MVT::i64, Expand);
316 // Custom lower Add/Sub/Mul with overflow.
317 setOperationAction(ISD::SADDO, MVT::i32, Custom);
318 setOperationAction(ISD::SADDO, MVT::i64, Custom);
319 setOperationAction(ISD::UADDO, MVT::i32, Custom);
320 setOperationAction(ISD::UADDO, MVT::i64, Custom);
321 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
322 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
323 setOperationAction(ISD::USUBO, MVT::i32, Custom);
324 setOperationAction(ISD::USUBO, MVT::i64, Custom);
325 setOperationAction(ISD::SMULO, MVT::i32, Custom);
326 setOperationAction(ISD::SMULO, MVT::i64, Custom);
327 setOperationAction(ISD::UMULO, MVT::i32, Custom);
328 setOperationAction(ISD::UMULO, MVT::i64, Custom);
330 setOperationAction(ISD::FSIN, MVT::f32, Expand);
331 setOperationAction(ISD::FSIN, MVT::f64, Expand);
332 setOperationAction(ISD::FCOS, MVT::f32, Expand);
333 setOperationAction(ISD::FCOS, MVT::f64, Expand);
334 setOperationAction(ISD::FPOW, MVT::f32, Expand);
335 setOperationAction(ISD::FPOW, MVT::f64, Expand);
336 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
337 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
339 // ARM64 has implementations of a lot of rounding-like FP operations.
340 static MVT RoundingTypes[] = { MVT::f32, MVT::f64, MVT::v2f32,
341 MVT::v4f32, MVT::v2f64 };
342 for (unsigned I = 0; I < array_lengthof(RoundingTypes); ++I) {
343 MVT Ty = RoundingTypes[I];
344 setOperationAction(ISD::FFLOOR, Ty, Legal);
345 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
346 setOperationAction(ISD::FCEIL, Ty, Legal);
347 setOperationAction(ISD::FRINT, Ty, Legal);
348 setOperationAction(ISD::FTRUNC, Ty, Legal);
349 setOperationAction(ISD::FROUND, Ty, Legal);
352 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
354 if (Subtarget->isTargetMachO()) {
355 // For iOS, we don't want to the normal expansion of a libcall to
356 // sincos. We want to issue a libcall to __sincos_stret to avoid memory
358 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
359 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
361 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
362 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
365 // ARM64 does not have floating-point extending loads, i1 sign-extending load,
366 // floating-point truncating stores, or v2i32->v2i16 truncating store.
367 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
368 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
369 setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
370 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Expand);
371 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
372 setTruncStoreAction(MVT::f128, MVT::f80, Expand);
373 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
374 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
375 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
376 // Indexed loads and stores are supported.
377 for (unsigned im = (unsigned)ISD::PRE_INC;
378 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
379 setIndexedLoadAction(im, MVT::i8, Legal);
380 setIndexedLoadAction(im, MVT::i16, Legal);
381 setIndexedLoadAction(im, MVT::i32, Legal);
382 setIndexedLoadAction(im, MVT::i64, Legal);
383 setIndexedLoadAction(im, MVT::f64, Legal);
384 setIndexedLoadAction(im, MVT::f32, Legal);
385 setIndexedStoreAction(im, MVT::i8, Legal);
386 setIndexedStoreAction(im, MVT::i16, Legal);
387 setIndexedStoreAction(im, MVT::i32, Legal);
388 setIndexedStoreAction(im, MVT::i64, Legal);
389 setIndexedStoreAction(im, MVT::f64, Legal);
390 setIndexedStoreAction(im, MVT::f32, Legal);
393 // Likewise, narrowing and extending vector loads/stores aren't handled
395 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
396 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
398 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,
401 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
402 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
403 setTruncStoreAction((MVT::SimpleValueType)VT,
404 (MVT::SimpleValueType)InnerVT, Expand);
405 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
406 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
407 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
411 setOperationAction(ISD::TRAP, MVT::Other, Legal);
412 setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
414 // We combine OR nodes for bitfield operations.
415 setTargetDAGCombine(ISD::OR);
417 // Vector add and sub nodes may conceal a high-half opportunity.
418 // Also, try to fold ADD into CSINC/CSINV..
419 setTargetDAGCombine(ISD::ADD);
420 setTargetDAGCombine(ISD::SUB);
422 setTargetDAGCombine(ISD::XOR);
423 setTargetDAGCombine(ISD::SINT_TO_FP);
424 setTargetDAGCombine(ISD::UINT_TO_FP);
426 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
428 setTargetDAGCombine(ISD::ANY_EXTEND);
429 setTargetDAGCombine(ISD::ZERO_EXTEND);
430 setTargetDAGCombine(ISD::SIGN_EXTEND);
431 setTargetDAGCombine(ISD::BITCAST);
432 setTargetDAGCombine(ISD::CONCAT_VECTORS);
433 setTargetDAGCombine(ISD::STORE);
435 setTargetDAGCombine(ISD::MUL);
437 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
438 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
439 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
441 setStackPointerRegisterToSaveRestore(ARM64::SP);
443 setSchedulingPreference(Sched::Hybrid);
446 MaskAndBranchFoldingIsLegal = true;
448 setMinFunctionAlignment(2);
450 RequireStrictAlign = StrictAlign;
453 void ARM64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
454 if (VT == MVT::v2f32) {
455 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
456 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);
458 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
459 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
460 } else if (VT == MVT::v2f64 || VT == MVT::v4f32) {
461 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
462 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);
464 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
465 AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
468 // Mark vector float intrinsics as expand.
469 if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
470 setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
471 setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
472 setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
473 setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
474 setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
475 setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
476 setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
477 setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
478 setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);
481 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
482 setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
483 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
484 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
485 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
486 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
487 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
488 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
489 setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
490 setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
491 setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
492 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
494 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
495 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
496 setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
497 setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
499 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
500 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
501 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
502 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
503 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
505 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
506 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
509 void ARM64TargetLowering::addDRTypeForNEON(MVT VT) {
510 addRegisterClass(VT, &ARM64::FPR64RegClass);
511 addTypeForNEON(VT, MVT::v2i32);
514 void ARM64TargetLowering::addQRTypeForNEON(MVT VT) {
515 addRegisterClass(VT, &ARM64::FPR128RegClass);
516 addTypeForNEON(VT, MVT::v4i32);
519 EVT ARM64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
522 return VT.changeVectorElementTypeToInteger();
525 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in
526 /// Mask are known to be either zero or one and return them in the
527 /// KnownZero/KnownOne bitsets.
528 void ARM64TargetLowering::computeMaskedBitsForTargetNode(
529 const SDValue Op, APInt &KnownZero, APInt &KnownOne,
530 const SelectionDAG &DAG, unsigned Depth) const {
531 switch (Op.getOpcode()) {
534 case ARM64ISD::CSEL: {
535 APInt KnownZero2, KnownOne2;
536 DAG.ComputeMaskedBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
537 DAG.ComputeMaskedBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
538 KnownZero &= KnownZero2;
539 KnownOne &= KnownOne2;
542 case ISD::INTRINSIC_W_CHAIN:
544 case ISD::INTRINSIC_WO_CHAIN:
545 case ISD::INTRINSIC_VOID: {
546 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
550 case Intrinsic::arm64_neon_umaxv:
551 case Intrinsic::arm64_neon_uminv: {
552 // Figure out the datatype of the vector operand. The UMINV instruction
553 // will zero extend the result, so we can mark as known zero all the
554 // bits larger than the element datatype. 32-bit or larget doesn't need
555 // this as those are legal types and will be handled by isel directly.
556 MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
557 unsigned BitWidth = KnownZero.getBitWidth();
558 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
559 assert(BitWidth >= 8 && "Unexpected width!");
560 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
562 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
563 assert(BitWidth >= 16 && "Unexpected width!");
564 APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
574 MVT ARM64TargetLowering::getScalarShiftAmountTy(EVT LHSTy) const {
578 unsigned ARM64TargetLowering::getMaximalGlobalOffset() const {
579 // FIXME: On ARM64, this depends on the type.
580 // Basically, the addressable offsets are o to 4095 * Ty.getSizeInBytes().
581 // and the offset has to be a multiple of the related size in bytes.
586 ARM64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
587 const TargetLibraryInfo *libInfo) const {
588 return ARM64::createFastISel(funcInfo, libInfo);
591 const char *ARM64TargetLowering::getTargetNodeName(unsigned Opcode) const {
595 case ARM64ISD::CALL: return "ARM64ISD::CALL";
596 case ARM64ISD::ADRP: return "ARM64ISD::ADRP";
597 case ARM64ISD::ADDlow: return "ARM64ISD::ADDlow";
598 case ARM64ISD::LOADgot: return "ARM64ISD::LOADgot";
599 case ARM64ISD::RET_FLAG: return "ARM64ISD::RET_FLAG";
600 case ARM64ISD::BRCOND: return "ARM64ISD::BRCOND";
601 case ARM64ISD::CSEL: return "ARM64ISD::CSEL";
602 case ARM64ISD::FCSEL: return "ARM64ISD::FCSEL";
603 case ARM64ISD::CSINV: return "ARM64ISD::CSINV";
604 case ARM64ISD::CSNEG: return "ARM64ISD::CSNEG";
605 case ARM64ISD::CSINC: return "ARM64ISD::CSINC";
606 case ARM64ISD::THREAD_POINTER: return "ARM64ISD::THREAD_POINTER";
607 case ARM64ISD::TLSDESC_CALL: return "ARM64ISD::TLSDESC_CALL";
608 case ARM64ISD::ADC: return "ARM64ISD::ADC";
609 case ARM64ISD::SBC: return "ARM64ISD::SBC";
610 case ARM64ISD::ADDS: return "ARM64ISD::ADDS";
611 case ARM64ISD::SUBS: return "ARM64ISD::SUBS";
612 case ARM64ISD::ADCS: return "ARM64ISD::ADCS";
613 case ARM64ISD::SBCS: return "ARM64ISD::SBCS";
614 case ARM64ISD::ANDS: return "ARM64ISD::ANDS";
615 case ARM64ISD::FCMP: return "ARM64ISD::FCMP";
616 case ARM64ISD::FMIN: return "ARM64ISD::FMIN";
617 case ARM64ISD::FMAX: return "ARM64ISD::FMAX";
618 case ARM64ISD::DUP: return "ARM64ISD::DUP";
619 case ARM64ISD::DUPLANE8: return "ARM64ISD::DUPLANE8";
620 case ARM64ISD::DUPLANE16: return "ARM64ISD::DUPLANE16";
621 case ARM64ISD::DUPLANE32: return "ARM64ISD::DUPLANE32";
622 case ARM64ISD::DUPLANE64: return "ARM64ISD::DUPLANE64";
623 case ARM64ISD::MOVI: return "ARM64ISD::MOVI";
624 case ARM64ISD::MOVIshift: return "ARM64ISD::MOVIshift";
625 case ARM64ISD::MOVIedit: return "ARM64ISD::MOVIedit";
626 case ARM64ISD::MOVImsl: return "ARM64ISD::MOVImsl";
627 case ARM64ISD::FMOV: return "ARM64ISD::FMOV";
628 case ARM64ISD::MVNIshift: return "ARM64ISD::MVNIshift";
629 case ARM64ISD::MVNImsl: return "ARM64ISD::MVNImsl";
630 case ARM64ISD::BICi: return "ARM64ISD::BICi";
631 case ARM64ISD::ORRi: return "ARM64ISD::ORRi";
632 case ARM64ISD::NEG: return "ARM64ISD::NEG";
633 case ARM64ISD::EXTR: return "ARM64ISD::EXTR";
634 case ARM64ISD::ZIP1: return "ARM64ISD::ZIP1";
635 case ARM64ISD::ZIP2: return "ARM64ISD::ZIP2";
636 case ARM64ISD::UZP1: return "ARM64ISD::UZP1";
637 case ARM64ISD::UZP2: return "ARM64ISD::UZP2";
638 case ARM64ISD::TRN1: return "ARM64ISD::TRN1";
639 case ARM64ISD::TRN2: return "ARM64ISD::TRN2";
640 case ARM64ISD::REV16: return "ARM64ISD::REV16";
641 case ARM64ISD::REV32: return "ARM64ISD::REV32";
642 case ARM64ISD::REV64: return "ARM64ISD::REV64";
643 case ARM64ISD::EXT: return "ARM64ISD::EXT";
644 case ARM64ISD::VSHL: return "ARM64ISD::VSHL";
645 case ARM64ISD::VLSHR: return "ARM64ISD::VLSHR";
646 case ARM64ISD::VASHR: return "ARM64ISD::VASHR";
647 case ARM64ISD::CMEQ: return "ARM64ISD::CMEQ";
648 case ARM64ISD::CMGE: return "ARM64ISD::CMGE";
649 case ARM64ISD::CMGT: return "ARM64ISD::CMGT";
650 case ARM64ISD::CMHI: return "ARM64ISD::CMHI";
651 case ARM64ISD::CMHS: return "ARM64ISD::CMHS";
652 case ARM64ISD::FCMEQ: return "ARM64ISD::FCMEQ";
653 case ARM64ISD::FCMGE: return "ARM64ISD::FCMGE";
654 case ARM64ISD::FCMGT: return "ARM64ISD::FCMGT";
655 case ARM64ISD::CMEQz: return "ARM64ISD::CMEQz";
656 case ARM64ISD::CMGEz: return "ARM64ISD::CMGEz";
657 case ARM64ISD::CMGTz: return "ARM64ISD::CMGTz";
658 case ARM64ISD::CMLEz: return "ARM64ISD::CMLEz";
659 case ARM64ISD::CMLTz: return "ARM64ISD::CMLTz";
660 case ARM64ISD::FCMEQz: return "ARM64ISD::FCMEQz";
661 case ARM64ISD::FCMGEz: return "ARM64ISD::FCMGEz";
662 case ARM64ISD::FCMGTz: return "ARM64ISD::FCMGTz";
663 case ARM64ISD::FCMLEz: return "ARM64ISD::FCMLEz";
664 case ARM64ISD::FCMLTz: return "ARM64ISD::FCMLTz";
665 case ARM64ISD::NOT: return "ARM64ISD::NOT";
666 case ARM64ISD::BIT: return "ARM64ISD::BIT";
667 case ARM64ISD::CBZ: return "ARM64ISD::CBZ";
668 case ARM64ISD::CBNZ: return "ARM64ISD::CBNZ";
669 case ARM64ISD::TBZ: return "ARM64ISD::TBZ";
670 case ARM64ISD::TBNZ: return "ARM64ISD::TBNZ";
671 case ARM64ISD::TC_RETURN: return "ARM64ISD::TC_RETURN";
672 case ARM64ISD::SITOF: return "ARM64ISD::SITOF";
673 case ARM64ISD::UITOF: return "ARM64ISD::UITOF";
674 case ARM64ISD::SQSHL_I: return "ARM64ISD::SQSHL_I";
675 case ARM64ISD::UQSHL_I: return "ARM64ISD::UQSHL_I";
676 case ARM64ISD::SRSHR_I: return "ARM64ISD::SRSHR_I";
677 case ARM64ISD::URSHR_I: return "ARM64ISD::URSHR_I";
678 case ARM64ISD::SQSHLU_I: return "ARM64ISD::SQSHLU_I";
679 case ARM64ISD::WrapperLarge: return "ARM64ISD::WrapperLarge";
683 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
684 unsigned &LdrOpc, unsigned &StrOpc) {
685 static unsigned LoadBares[] = { ARM64::LDXRB, ARM64::LDXRH, ARM64::LDXRW,
686 ARM64::LDXRX, ARM64::LDXPX };
687 static unsigned LoadAcqs[] = { ARM64::LDAXRB, ARM64::LDAXRH, ARM64::LDAXRW,
688 ARM64::LDAXRX, ARM64::LDAXPX };
689 static unsigned StoreBares[] = { ARM64::STXRB, ARM64::STXRH, ARM64::STXRW,
690 ARM64::STXRX, ARM64::STXPX };
691 static unsigned StoreRels[] = { ARM64::STLXRB, ARM64::STLXRH, ARM64::STLXRW,
692 ARM64::STLXRX, ARM64::STLXPX };
694 unsigned *LoadOps, *StoreOps;
695 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
700 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
701 StoreOps = StoreRels;
703 StoreOps = StoreBares;
705 assert(isPowerOf2_32(Size) && Size <= 16 &&
706 "unsupported size for atomic binary op!");
708 LdrOpc = LoadOps[Log2_32(Size)];
709 StrOpc = StoreOps[Log2_32(Size)];
712 MachineBasicBlock *ARM64TargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
713 MachineBasicBlock *BB,
714 unsigned Size) const {
715 unsigned dest = MI->getOperand(0).getReg();
716 unsigned ptr = MI->getOperand(1).getReg();
717 unsigned oldval = MI->getOperand(2).getReg();
718 unsigned newval = MI->getOperand(3).getReg();
719 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
720 unsigned scratch = BB->getParent()->getRegInfo().createVirtualRegister(
721 &ARM64::GPR32RegClass);
722 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
723 DebugLoc dl = MI->getDebugLoc();
725 // FIXME: We currently always generate a seq_cst operation; we should
726 // be able to relax this in some cases.
727 unsigned ldrOpc, strOpc;
728 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
730 MachineFunction *MF = BB->getParent();
731 const BasicBlock *LLVM_BB = BB->getBasicBlock();
732 MachineFunction::iterator It = BB;
733 ++It; // insert the new blocks after the current block
735 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
736 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
737 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
738 MF->insert(It, loop1MBB);
739 MF->insert(It, loop2MBB);
740 MF->insert(It, exitMBB);
742 // Transfer the remainder of BB and its successor edges to exitMBB.
743 exitMBB->splice(exitMBB->begin(), BB,
744 std::next(MachineBasicBlock::iterator(MI)), BB->end());
745 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
749 // fallthrough --> loop1MBB
750 BB->addSuccessor(loop1MBB);
757 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
758 BuildMI(BB, dl, TII->get(Size == 8 ? ARM64::SUBSXrr : ARM64::SUBSWrr))
759 .addReg(Size == 8 ? ARM64::XZR : ARM64::WZR, RegState::Define)
762 BuildMI(BB, dl, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(exitMBB);
763 BB->addSuccessor(loop2MBB);
764 BB->addSuccessor(exitMBB);
767 // strex scratch, newval, [ptr]
771 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
772 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loop1MBB);
773 BB->addSuccessor(loop1MBB);
774 BB->addSuccessor(exitMBB);
780 MI->eraseFromParent(); // The instruction is gone now.
786 ARM64TargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
787 unsigned Size, unsigned BinOpcode) const {
788 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
789 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
791 const BasicBlock *LLVM_BB = BB->getBasicBlock();
792 MachineFunction *MF = BB->getParent();
793 MachineFunction::iterator It = BB;
796 unsigned dest = MI->getOperand(0).getReg();
797 unsigned ptr = MI->getOperand(1).getReg();
798 unsigned incr = MI->getOperand(2).getReg();
799 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
800 DebugLoc dl = MI->getDebugLoc();
802 unsigned ldrOpc, strOpc;
803 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
805 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
806 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
807 MF->insert(It, loopMBB);
808 MF->insert(It, exitMBB);
810 // Transfer the remainder of BB and its successor edges to exitMBB.
811 exitMBB->splice(exitMBB->begin(), BB,
812 std::next(MachineBasicBlock::iterator(MI)), BB->end());
813 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
815 MachineRegisterInfo &RegInfo = MF->getRegInfo();
816 unsigned scratch = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
820 : RegInfo.createVirtualRegister(Size == 8 ? &ARM64::GPR64RegClass
821 : &ARM64::GPR32RegClass);
825 // fallthrough --> loopMBB
826 BB->addSuccessor(loopMBB);
830 // <binop> scratch2, dest, incr
831 // stxr scratch, scratch2, ptr
832 // cbnz scratch, loopMBB
833 // fallthrough --> exitMBB
835 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
837 // operand order needs to go the other way for NAND
838 if (BinOpcode == ARM64::BICWrr || BinOpcode == ARM64::BICXrr)
839 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(incr).addReg(dest);
841 BuildMI(BB, dl, TII->get(BinOpcode), scratch2).addReg(dest).addReg(incr);
844 BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
845 BuildMI(BB, dl, TII->get(ARM64::CBNZW)).addReg(scratch).addMBB(loopMBB);
847 BB->addSuccessor(loopMBB);
848 BB->addSuccessor(exitMBB);
854 MI->eraseFromParent(); // The instruction is gone now.
859 MachineBasicBlock *ARM64TargetLowering::EmitAtomicBinary128(
860 MachineInstr *MI, MachineBasicBlock *BB, unsigned BinOpcodeLo,
861 unsigned BinOpcodeHi) const {
862 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
863 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
865 const BasicBlock *LLVM_BB = BB->getBasicBlock();
866 MachineFunction *MF = BB->getParent();
867 MachineFunction::iterator It = BB;
870 unsigned DestLo = MI->getOperand(0).getReg();
871 unsigned DestHi = MI->getOperand(1).getReg();
872 unsigned Ptr = MI->getOperand(2).getReg();
873 unsigned IncrLo = MI->getOperand(3).getReg();
874 unsigned IncrHi = MI->getOperand(4).getReg();
875 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
876 DebugLoc DL = MI->getDebugLoc();
878 unsigned LdrOpc, StrOpc;
879 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
881 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
882 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
883 MF->insert(It, LoopMBB);
884 MF->insert(It, ExitMBB);
886 // Transfer the remainder of BB and its successor edges to exitMBB.
887 ExitMBB->splice(ExitMBB->begin(), BB,
888 std::next(MachineBasicBlock::iterator(MI)), BB->end());
889 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
891 MachineRegisterInfo &RegInfo = MF->getRegInfo();
892 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
893 unsigned ScratchLo = IncrLo, ScratchHi = IncrHi;
895 assert(BinOpcodeHi && "Expect neither or both opcodes to be defined");
896 ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
897 ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
902 // fallthrough --> LoopMBB
903 BB->addSuccessor(LoopMBB);
906 // ldxp DestLo, DestHi, Ptr
907 // <binoplo> ScratchLo, DestLo, IncrLo
908 // <binophi> ScratchHi, DestHi, IncrHi
909 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
910 // cbnz ScratchRes, LoopMBB
911 // fallthrough --> ExitMBB
913 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
914 .addReg(DestHi, RegState::Define)
917 // operand order needs to go the other way for NAND
918 if (BinOpcodeLo == ARM64::BICXrr) {
919 std::swap(IncrLo, DestLo);
920 std::swap(IncrHi, DestHi);
923 BuildMI(BB, DL, TII->get(BinOpcodeLo), ScratchLo).addReg(DestLo).addReg(
925 BuildMI(BB, DL, TII->get(BinOpcodeHi), ScratchHi).addReg(DestHi).addReg(
929 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
933 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
935 BB->addSuccessor(LoopMBB);
936 BB->addSuccessor(ExitMBB);
942 MI->eraseFromParent(); // The instruction is gone now.
948 ARM64TargetLowering::EmitAtomicCmpSwap128(MachineInstr *MI,
949 MachineBasicBlock *BB) const {
950 unsigned DestLo = MI->getOperand(0).getReg();
951 unsigned DestHi = MI->getOperand(1).getReg();
952 unsigned Ptr = MI->getOperand(2).getReg();
953 unsigned OldValLo = MI->getOperand(3).getReg();
954 unsigned OldValHi = MI->getOperand(4).getReg();
955 unsigned NewValLo = MI->getOperand(5).getReg();
956 unsigned NewValHi = MI->getOperand(6).getReg();
957 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(7).getImm());
958 unsigned ScratchRes = BB->getParent()->getRegInfo().createVirtualRegister(
959 &ARM64::GPR32RegClass);
960 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
961 DebugLoc DL = MI->getDebugLoc();
963 unsigned LdrOpc, StrOpc;
964 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
966 MachineFunction *MF = BB->getParent();
967 const BasicBlock *LLVM_BB = BB->getBasicBlock();
968 MachineFunction::iterator It = BB;
969 ++It; // insert the new blocks after the current block
971 MachineBasicBlock *Loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
972 MachineBasicBlock *Loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
973 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
974 MF->insert(It, Loop1MBB);
975 MF->insert(It, Loop2MBB);
976 MF->insert(It, ExitMBB);
978 // Transfer the remainder of BB and its successor edges to exitMBB.
979 ExitMBB->splice(ExitMBB->begin(), BB,
980 std::next(MachineBasicBlock::iterator(MI)), BB->end());
981 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
985 // fallthrough --> Loop1MBB
986 BB->addSuccessor(Loop1MBB);
989 // ldxp DestLo, DestHi, [Ptr]
990 // cmp DestLo, OldValLo
991 // sbc xzr, DestHi, OldValHi
994 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
995 .addReg(DestHi, RegState::Define)
997 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
999 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1002 BuildMI(BB, DL, TII->get(ARM64::Bcc)).addImm(ARM64CC::NE).addMBB(ExitMBB);
1003 BB->addSuccessor(Loop2MBB);
1004 BB->addSuccessor(ExitMBB);
1007 // stxp ScratchRes, NewValLo, NewValHi, [Ptr]
1008 // cbnz ScratchRes, Loop1MBB
1010 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1014 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(Loop1MBB);
1015 BB->addSuccessor(Loop1MBB);
1016 BB->addSuccessor(ExitMBB);
1022 MI->eraseFromParent(); // The instruction is gone now.
1027 MachineBasicBlock *ARM64TargetLowering::EmitAtomicMinMax128(
1028 MachineInstr *MI, MachineBasicBlock *BB, unsigned CondCode) const {
1029 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
1030 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1032 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1033 MachineFunction *MF = BB->getParent();
1034 MachineFunction::iterator It = BB;
1037 unsigned DestLo = MI->getOperand(0).getReg();
1038 unsigned DestHi = MI->getOperand(1).getReg();
1039 unsigned Ptr = MI->getOperand(2).getReg();
1040 unsigned IncrLo = MI->getOperand(3).getReg();
1041 unsigned IncrHi = MI->getOperand(4).getReg();
1042 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(5).getImm());
1043 DebugLoc DL = MI->getDebugLoc();
1045 unsigned LdrOpc, StrOpc;
1046 getExclusiveOperation(16, Ord, LdrOpc, StrOpc);
1048 MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1049 MachineBasicBlock *ExitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1050 MF->insert(It, LoopMBB);
1051 MF->insert(It, ExitMBB);
1053 // Transfer the remainder of BB and its successor edges to exitMBB.
1054 ExitMBB->splice(ExitMBB->begin(), BB,
1055 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1056 ExitMBB->transferSuccessorsAndUpdatePHIs(BB);
1058 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1059 unsigned ScratchRes = RegInfo.createVirtualRegister(&ARM64::GPR32RegClass);
1060 unsigned ScratchLo = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1061 unsigned ScratchHi = RegInfo.createVirtualRegister(&ARM64::GPR64RegClass);
1065 // fallthrough --> LoopMBB
1066 BB->addSuccessor(LoopMBB);
1069 // ldxp DestLo, DestHi, Ptr
1070 // cmp ScratchLo, DestLo, IncrLo
1071 // sbc xzr, ScratchHi, DestHi, IncrHi
1072 // csel ScratchLo, DestLo, IncrLo, <cmp-op>
1073 // csel ScratchHi, DestHi, IncrHi, <cmp-op>
1074 // stxp ScratchRes, ScratchLo, ScratchHi, ptr
1075 // cbnz ScratchRes, LoopMBB
1076 // fallthrough --> ExitMBB
1078 BuildMI(BB, DL, TII->get(LdrOpc), DestLo)
1079 .addReg(DestHi, RegState::Define)
1082 BuildMI(BB, DL, TII->get(ARM64::SUBSXrr), ARM64::XZR).addReg(DestLo).addReg(
1084 BuildMI(BB, DL, TII->get(ARM64::SBCXr), ARM64::XZR).addReg(DestHi).addReg(
1087 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchLo)
1091 BuildMI(BB, DL, TII->get(ARM64::CSELXr), ScratchHi)
1096 BuildMI(BB, DL, TII->get(StrOpc), ScratchRes)
1100 BuildMI(BB, DL, TII->get(ARM64::CBNZW)).addReg(ScratchRes).addMBB(LoopMBB);
1102 BB->addSuccessor(LoopMBB);
1103 BB->addSuccessor(ExitMBB);
1109 MI->eraseFromParent(); // The instruction is gone now.
1115 ARM64TargetLowering::EmitF128CSEL(MachineInstr *MI,
1116 MachineBasicBlock *MBB) const {
1117 // We materialise the F128CSEL pseudo-instruction as some control flow and a
1121 // [... previous instrs leading to comparison ...]
1127 // Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]
1129 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1130 MachineFunction *MF = MBB->getParent();
1131 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
1132 DebugLoc DL = MI->getDebugLoc();
1133 MachineFunction::iterator It = MBB;
1136 unsigned DestReg = MI->getOperand(0).getReg();
1137 unsigned IfTrueReg = MI->getOperand(1).getReg();
1138 unsigned IfFalseReg = MI->getOperand(2).getReg();
1139 unsigned CondCode = MI->getOperand(3).getImm();
1140 bool CPSRKilled = MI->getOperand(4).isKill();
1142 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
1143 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
1144 MF->insert(It, TrueBB);
1145 MF->insert(It, EndBB);
1147 // Transfer rest of current basic-block to EndBB
1148 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
1150 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
1152 BuildMI(MBB, DL, TII->get(ARM64::Bcc)).addImm(CondCode).addMBB(TrueBB);
1153 BuildMI(MBB, DL, TII->get(ARM64::B)).addMBB(EndBB);
1154 MBB->addSuccessor(TrueBB);
1155 MBB->addSuccessor(EndBB);
1157 // TrueBB falls through to the end.
1158 TrueBB->addSuccessor(EndBB);
1161 TrueBB->addLiveIn(ARM64::CPSR);
1162 EndBB->addLiveIn(ARM64::CPSR);
1165 BuildMI(*EndBB, EndBB->begin(), DL, TII->get(ARM64::PHI), DestReg)
1171 MI->eraseFromParent();
1176 ARM64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1177 MachineBasicBlock *BB) const {
1178 switch (MI->getOpcode()) {
1183 assert(0 && "Unexpected instruction for custom inserter!");
1186 case ARM64::ATOMIC_LOAD_ADD_I8:
1187 return EmitAtomicBinary(MI, BB, 1, ARM64::ADDWrr);
1188 case ARM64::ATOMIC_LOAD_ADD_I16:
1189 return EmitAtomicBinary(MI, BB, 2, ARM64::ADDWrr);
1190 case ARM64::ATOMIC_LOAD_ADD_I32:
1191 return EmitAtomicBinary(MI, BB, 4, ARM64::ADDWrr);
1192 case ARM64::ATOMIC_LOAD_ADD_I64:
1193 return EmitAtomicBinary(MI, BB, 8, ARM64::ADDXrr);
1194 case ARM64::ATOMIC_LOAD_ADD_I128:
1195 return EmitAtomicBinary128(MI, BB, ARM64::ADDSXrr, ARM64::ADCXr);
1197 case ARM64::ATOMIC_LOAD_AND_I8:
1198 return EmitAtomicBinary(MI, BB, 1, ARM64::ANDWrr);
1199 case ARM64::ATOMIC_LOAD_AND_I16:
1200 return EmitAtomicBinary(MI, BB, 2, ARM64::ANDWrr);
1201 case ARM64::ATOMIC_LOAD_AND_I32:
1202 return EmitAtomicBinary(MI, BB, 4, ARM64::ANDWrr);
1203 case ARM64::ATOMIC_LOAD_AND_I64:
1204 return EmitAtomicBinary(MI, BB, 8, ARM64::ANDXrr);
1205 case ARM64::ATOMIC_LOAD_AND_I128:
1206 return EmitAtomicBinary128(MI, BB, ARM64::ANDXrr, ARM64::ANDXrr);
1208 case ARM64::ATOMIC_LOAD_OR_I8:
1209 return EmitAtomicBinary(MI, BB, 1, ARM64::ORRWrr);
1210 case ARM64::ATOMIC_LOAD_OR_I16:
1211 return EmitAtomicBinary(MI, BB, 2, ARM64::ORRWrr);
1212 case ARM64::ATOMIC_LOAD_OR_I32:
1213 return EmitAtomicBinary(MI, BB, 4, ARM64::ORRWrr);
1214 case ARM64::ATOMIC_LOAD_OR_I64:
1215 return EmitAtomicBinary(MI, BB, 8, ARM64::ORRXrr);
1216 case ARM64::ATOMIC_LOAD_OR_I128:
1217 return EmitAtomicBinary128(MI, BB, ARM64::ORRXrr, ARM64::ORRXrr);
1219 case ARM64::ATOMIC_LOAD_XOR_I8:
1220 return EmitAtomicBinary(MI, BB, 1, ARM64::EORWrr);
1221 case ARM64::ATOMIC_LOAD_XOR_I16:
1222 return EmitAtomicBinary(MI, BB, 2, ARM64::EORWrr);
1223 case ARM64::ATOMIC_LOAD_XOR_I32:
1224 return EmitAtomicBinary(MI, BB, 4, ARM64::EORWrr);
1225 case ARM64::ATOMIC_LOAD_XOR_I64:
1226 return EmitAtomicBinary(MI, BB, 8, ARM64::EORXrr);
1227 case ARM64::ATOMIC_LOAD_XOR_I128:
1228 return EmitAtomicBinary128(MI, BB, ARM64::EORXrr, ARM64::EORXrr);
1230 case ARM64::ATOMIC_LOAD_NAND_I8:
1231 return EmitAtomicBinary(MI, BB, 1, ARM64::BICWrr);
1232 case ARM64::ATOMIC_LOAD_NAND_I16:
1233 return EmitAtomicBinary(MI, BB, 2, ARM64::BICWrr);
1234 case ARM64::ATOMIC_LOAD_NAND_I32:
1235 return EmitAtomicBinary(MI, BB, 4, ARM64::BICWrr);
1236 case ARM64::ATOMIC_LOAD_NAND_I64:
1237 return EmitAtomicBinary(MI, BB, 8, ARM64::BICXrr);
1238 case ARM64::ATOMIC_LOAD_NAND_I128:
1239 return EmitAtomicBinary128(MI, BB, ARM64::BICXrr, ARM64::BICXrr);
1241 case ARM64::ATOMIC_LOAD_SUB_I8:
1242 return EmitAtomicBinary(MI, BB, 1, ARM64::SUBWrr);
1243 case ARM64::ATOMIC_LOAD_SUB_I16:
1244 return EmitAtomicBinary(MI, BB, 2, ARM64::SUBWrr);
1245 case ARM64::ATOMIC_LOAD_SUB_I32:
1246 return EmitAtomicBinary(MI, BB, 4, ARM64::SUBWrr);
1247 case ARM64::ATOMIC_LOAD_SUB_I64:
1248 return EmitAtomicBinary(MI, BB, 8, ARM64::SUBXrr);
1249 case ARM64::ATOMIC_LOAD_SUB_I128:
1250 return EmitAtomicBinary128(MI, BB, ARM64::SUBSXrr, ARM64::SBCXr);
1252 case ARM64::ATOMIC_LOAD_MIN_I128:
1253 return EmitAtomicMinMax128(MI, BB, ARM64CC::LT);
1255 case ARM64::ATOMIC_LOAD_MAX_I128:
1256 return EmitAtomicMinMax128(MI, BB, ARM64CC::GT);
1258 case ARM64::ATOMIC_LOAD_UMIN_I128:
1259 return EmitAtomicMinMax128(MI, BB, ARM64CC::CC);
1261 case ARM64::ATOMIC_LOAD_UMAX_I128:
1262 return EmitAtomicMinMax128(MI, BB, ARM64CC::HI);
1264 case ARM64::ATOMIC_SWAP_I8:
1265 return EmitAtomicBinary(MI, BB, 1, 0);
1266 case ARM64::ATOMIC_SWAP_I16:
1267 return EmitAtomicBinary(MI, BB, 2, 0);
1268 case ARM64::ATOMIC_SWAP_I32:
1269 return EmitAtomicBinary(MI, BB, 4, 0);
1270 case ARM64::ATOMIC_SWAP_I64:
1271 return EmitAtomicBinary(MI, BB, 8, 0);
1272 case ARM64::ATOMIC_SWAP_I128:
1273 return EmitAtomicBinary128(MI, BB, 0, 0);
1275 case ARM64::ATOMIC_CMP_SWAP_I8:
1276 return EmitAtomicCmpSwap(MI, BB, 1);
1277 case ARM64::ATOMIC_CMP_SWAP_I16:
1278 return EmitAtomicCmpSwap(MI, BB, 2);
1279 case ARM64::ATOMIC_CMP_SWAP_I32:
1280 return EmitAtomicCmpSwap(MI, BB, 4);
1281 case ARM64::ATOMIC_CMP_SWAP_I64:
1282 return EmitAtomicCmpSwap(MI, BB, 8);
1283 case ARM64::ATOMIC_CMP_SWAP_I128:
1284 return EmitAtomicCmpSwap128(MI, BB);
1286 case ARM64::F128CSEL:
1287 return EmitF128CSEL(MI, BB);
1289 case TargetOpcode::STACKMAP:
1290 case TargetOpcode::PATCHPOINT:
1291 return emitPatchPoint(MI, BB);
1293 llvm_unreachable("Unexpected instruction for custom inserter!");
1296 //===----------------------------------------------------------------------===//
1297 // ARM64 Lowering private implementation.
1298 //===----------------------------------------------------------------------===//
1300 //===----------------------------------------------------------------------===//
1302 //===----------------------------------------------------------------------===//
1304 /// changeIntCCToARM64CC - Convert a DAG integer condition code to an ARM64 CC
1305 static ARM64CC::CondCode changeIntCCToARM64CC(ISD::CondCode CC) {
1308 llvm_unreachable("Unknown condition code!");
1332 /// changeFPCCToARM64CC - Convert a DAG fp condition code to an ARM64 CC.
1333 static void changeFPCCToARM64CC(ISD::CondCode CC, ARM64CC::CondCode &CondCode,
1334 ARM64CC::CondCode &CondCode2) {
1335 CondCode2 = ARM64CC::AL;
1338 llvm_unreachable("Unknown FP condition!");
1341 CondCode = ARM64CC::EQ;
1345 CondCode = ARM64CC::GT;
1349 CondCode = ARM64CC::GE;
1352 CondCode = ARM64CC::MI;
1355 CondCode = ARM64CC::LS;
1358 CondCode = ARM64CC::MI;
1359 CondCode2 = ARM64CC::GT;
1362 CondCode = ARM64CC::VC;
1365 CondCode = ARM64CC::VS;
1368 CondCode = ARM64CC::EQ;
1369 CondCode2 = ARM64CC::VS;
1372 CondCode = ARM64CC::HI;
1375 CondCode = ARM64CC::PL;
1379 CondCode = ARM64CC::LT;
1383 CondCode = ARM64CC::LE;
1387 CondCode = ARM64CC::NE;
1392 static bool isLegalArithImmed(uint64_t C) {
1393 // Matches ARM64DAGToDAGISel::SelectArithImmed().
1394 return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
1397 static SDValue emitComparison(SDValue LHS, SDValue RHS, SDLoc dl,
1398 SelectionDAG &DAG) {
1399 EVT VT = LHS.getValueType();
1401 if (VT.isFloatingPoint())
1402 return DAG.getNode(ARM64ISD::FCMP, dl, VT, LHS, RHS);
1404 // The CMP instruction is just an alias for SUBS, and representing it as
1405 // SUBS means that it's possible to get CSE with subtract operations.
1406 // A later phase can perform the optimization of setting the destination
1407 // register to WZR/XZR if it ends up being unused.
1408 return DAG.getNode(ARM64ISD::SUBS, dl, DAG.getVTList(VT, MVT::i32), LHS, RHS)
1412 static SDValue getARM64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1413 SDValue &ARM64cc, SelectionDAG &DAG, SDLoc dl) {
1414 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1415 EVT VT = RHS.getValueType();
1416 uint64_t C = RHSC->getZExtValue();
1417 if (!isLegalArithImmed(C)) {
1418 // Constant does not fit, try adjusting it by one?
1424 if ((VT == MVT::i32 && C != 0x80000000 &&
1425 isLegalArithImmed((uint32_t)(C - 1))) ||
1426 (VT == MVT::i64 && C != 0x80000000ULL &&
1427 isLegalArithImmed(C - 1ULL))) {
1428 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1429 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1430 RHS = DAG.getConstant(C, VT);
1435 if ((VT == MVT::i32 && C != 0 &&
1436 isLegalArithImmed((uint32_t)(C - 1))) ||
1437 (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
1438 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1439 C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
1440 RHS = DAG.getConstant(C, VT);
1445 if ((VT == MVT::i32 && C != 0x7fffffff &&
1446 isLegalArithImmed((uint32_t)(C + 1))) ||
1447 (VT == MVT::i64 && C != 0x7ffffffffffffffULL &&
1448 isLegalArithImmed(C + 1ULL))) {
1449 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1450 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1451 RHS = DAG.getConstant(C, VT);
1456 if ((VT == MVT::i32 && C != 0xffffffff &&
1457 isLegalArithImmed((uint32_t)(C + 1))) ||
1458 (VT == MVT::i64 && C != 0xfffffffffffffffULL &&
1459 isLegalArithImmed(C + 1ULL))) {
1460 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1461 C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
1462 RHS = DAG.getConstant(C, VT);
1469 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
1470 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
1471 ARM64cc = DAG.getConstant(ARM64CC, MVT::i32);
1475 static std::pair<SDValue, SDValue>
1476 getARM64XALUOOp(ARM64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
1477 assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
1478 "Unsupported value type");
1479 SDValue Value, Overflow;
1481 SDValue LHS = Op.getOperand(0);
1482 SDValue RHS = Op.getOperand(1);
1484 switch (Op.getOpcode()) {
1486 llvm_unreachable("Unknown overflow instruction!");
1488 Opc = ARM64ISD::ADDS;
1492 Opc = ARM64ISD::ADDS;
1496 Opc = ARM64ISD::SUBS;
1500 Opc = ARM64ISD::SUBS;
1503 // Multiply needs a little bit extra work.
1507 bool IsSigned = (Op.getOpcode() == ISD::SMULO) ? true : false;
1508 if (Op.getValueType() == MVT::i32) {
1509 unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1510 // For a 32 bit multiply with overflow check we want the instruction
1511 // selector to generate a widening multiply (SMADDL/UMADDL). For that we
1512 // need to generate the following pattern:
1513 // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
1514 LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
1515 RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
1516 SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1517 SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
1518 DAG.getConstant(0, MVT::i64));
1519 // On ARM64 the upper 32 bits are always zero extended for a 32 bit
1520 // operation. We need to clear out the upper 32 bits, because we used a
1521 // widening multiply that wrote all 64 bits. In the end this should be a
1523 Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
1525 // The signed overflow check requires more than just a simple check for
1526 // any bit set in the upper 32 bits of the result. These bits could be
1527 // just the sign bits of a negative number. To perform the overflow
1528 // check we have to arithmetic shift right the 32nd bit of the result by
1529 // 31 bits. Then we compare the result to the upper 32 bits.
1530 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
1531 DAG.getConstant(32, MVT::i64));
1532 UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
1533 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
1534 DAG.getConstant(31, MVT::i64));
1535 // It is important that LowerBits is last, otherwise the arithmetic
1536 // shift will not be folded into the compare (SUBS).
1537 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
1538 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1541 // The overflow check for unsigned multiply is easy. We only need to
1542 // check if any of the upper 32 bits are set. This can be done with a
1543 // CMP (shifted register). For that we need to generate the following
1545 // (i64 ARM64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
1546 SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
1547 DAG.getConstant(32, MVT::i64));
1548 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1550 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1551 UpperBits).getValue(1);
1555 assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
1556 // For the 64 bit multiply
1557 Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
1559 SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
1560 SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
1561 DAG.getConstant(63, MVT::i64));
1562 // It is important that LowerBits is last, otherwise the arithmetic
1563 // shift will not be folded into the compare (SUBS).
1564 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1565 Overflow = DAG.getNode(ARM64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
1568 SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
1569 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
1571 DAG.getNode(ARM64ISD::SUBS, DL, VTs, DAG.getConstant(0, MVT::i64),
1572 UpperBits).getValue(1);
1579 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);
1581 // Emit the ARM64 operation with overflow check.
1582 Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
1583 Overflow = Value.getValue(1);
1585 return std::make_pair(Value, Overflow);
1588 SDValue ARM64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
1589 RTLIB::Libcall Call) const {
1590 SmallVector<SDValue, 2> Ops;
1591 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1592 Ops.push_back(Op.getOperand(i));
1594 return makeLibCall(DAG, Call, MVT::f128, &Ops[0], Ops.size(), false,
1598 static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
1599 SDValue Sel = Op.getOperand(0);
1600 SDValue Other = Op.getOperand(1);
1602 // If neither operand is a SELECT_CC, give up.
1603 if (Sel.getOpcode() != ISD::SELECT_CC)
1604 std::swap(Sel, Other);
1605 if (Sel.getOpcode() != ISD::SELECT_CC)
1608 // The folding we want to perform is:
1609 // (xor x, (select_cc a, b, cc, 0, -1) )
1611 // (csel x, (xor x, -1), cc ...)
1613 // The latter will get matched to a CSINV instruction.
1615 ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
1616 SDValue LHS = Sel.getOperand(0);
1617 SDValue RHS = Sel.getOperand(1);
1618 SDValue TVal = Sel.getOperand(2);
1619 SDValue FVal = Sel.getOperand(3);
1622 // FIXME: This could be generalized to non-integer comparisons.
1623 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
1626 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
1627 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
1629 // The the values aren't constants, this isn't the pattern we're looking for.
1630 if (!CFVal || !CTVal)
1633 // We can commute the SELECT_CC by inverting the condition. This
1634 // might be needed to make this fit into a CSINV pattern.
1635 if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
1636 std::swap(TVal, FVal);
1637 std::swap(CTVal, CFVal);
1638 CC = ISD::getSetCCInverse(CC, true);
1641 // If the constants line up, perform the transform!
1642 if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
1644 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
1647 TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
1648 DAG.getConstant(-1ULL, Other.getValueType()));
1650 return DAG.getNode(ARM64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
1657 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
1658 EVT VT = Op.getValueType();
1660 // Let legalize expand this if it isn't a legal type yet.
1661 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
1664 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
1667 bool ExtraOp = false;
1668 switch (Op.getOpcode()) {
1670 assert(0 && "Invalid code");
1672 Opc = ARM64ISD::ADDS;
1675 Opc = ARM64ISD::SUBS;
1678 Opc = ARM64ISD::ADCS;
1682 Opc = ARM64ISD::SBCS;
1688 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
1689 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
1693 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
1694 // Let legalize expand this if it isn't a legal type yet.
1695 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
1698 ARM64CC::CondCode CC;
1699 // The actual operation that sets the overflow or carry flag.
1700 SDValue Value, Overflow;
1701 std::tie(Value, Overflow) = getARM64XALUOOp(CC, Op, DAG);
1703 // We use 0 and 1 as false and true values.
1704 SDValue TVal = DAG.getConstant(1, MVT::i32);
1705 SDValue FVal = DAG.getConstant(0, MVT::i32);
1707 // We use an inverted condition, because the conditional select is inverted
1708 // too. This will allow it to be selected to a single instruction:
1709 // CSINC Wd, WZR, WZR, invert(cond).
1710 SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), MVT::i32);
1711 Overflow = DAG.getNode(ARM64ISD::CSEL, SDLoc(Op), MVT::i32, FVal, TVal, CCVal,
1714 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
1715 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
1718 // Prefetch operands are:
1719 // 1: Address to prefetch
1721 // 3: int locality (0 = no locality ... 3 = extreme locality)
1722 // 4: bool isDataCache
1723 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
1725 unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
1726 unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
1727 // The data thing is not used.
1728 // unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
1730 bool IsStream = !Locality;
1731 // When the locality number is set
1733 // The front-end should have filtered out the out-of-range values
1734 assert(Locality <= 3 && "Prefetch locality out-of-range");
1735 // The locality degree is the opposite of the cache speed.
1736 // Put the number the other way around.
1737 // The encoding starts at 0 for level 1
1738 Locality = 3 - Locality;
1741 // built the mask value encoding the expected behavior.
1742 unsigned PrfOp = (IsWrite << 4) | // Load/Store bit
1743 (Locality << 1) | // Cache level bits
1744 (unsigned)IsStream; // Stream bit
1745 return DAG.getNode(ARM64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
1746 DAG.getConstant(PrfOp, MVT::i32), Op.getOperand(1));
1749 SDValue ARM64TargetLowering::LowerFP_EXTEND(SDValue Op,
1750 SelectionDAG &DAG) const {
1751 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1754 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1756 return LowerF128Call(Op, DAG, LC);
1759 SDValue ARM64TargetLowering::LowerFP_ROUND(SDValue Op,
1760 SelectionDAG &DAG) const {
1761 if (Op.getOperand(0).getValueType() != MVT::f128) {
1762 // It's legal except when f128 is involved
1767 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1769 // FP_ROUND node has a second operand indicating whether it is known to be
1770 // precise. That doesn't take part in the LibCall so we can't directly use
1772 SDValue SrcVal = Op.getOperand(0);
1773 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1774 /*isSigned*/ false, SDLoc(Op)).first;
1777 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1778 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1779 // Any additional optimization in this function should be recorded
1780 // in the cost tables.
1781 EVT InVT = Op.getOperand(0).getValueType();
1782 EVT VT = Op.getValueType();
1784 // FP_TO_XINT conversion from the same type are legal.
1785 if (VT.getSizeInBits() == InVT.getSizeInBits())
1788 if (InVT == MVT::v2f64) {
1790 SDValue Cv = DAG.getNode(Op.getOpcode(), dl, MVT::v2i64, Op.getOperand(0));
1791 return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
1794 // Type changing conversions are illegal.
1798 SDValue ARM64TargetLowering::LowerFP_TO_INT(SDValue Op,
1799 SelectionDAG &DAG) const {
1800 if (Op.getOperand(0).getValueType().isVector())
1801 return LowerVectorFP_TO_INT(Op, DAG);
1803 if (Op.getOperand(0).getValueType() != MVT::f128) {
1804 // It's legal except when f128 is involved
1809 if (Op.getOpcode() == ISD::FP_TO_SINT)
1810 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1812 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1814 SmallVector<SDValue, 2> Ops;
1815 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i)
1816 Ops.push_back(Op.getOperand(i));
1818 return makeLibCall(DAG, LC, Op.getValueType(), &Ops[0], Ops.size(), false,
1822 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1823 // Warning: We maintain cost tables in ARM64TargetTransformInfo.cpp.
1824 // Any additional optimization in this function should be recorded
1825 // in the cost tables.
1826 EVT VT = Op.getValueType();
1828 SDValue In = Op.getOperand(0);
1829 EVT InVT = In.getValueType();
1831 // v2i32 to v2f32 is legal.
1832 if (VT == MVT::v2f32 && InVT == MVT::v2i32)
1835 // This function only handles v2f64 outputs.
1836 if (VT == MVT::v2f64) {
1837 // Extend the input argument to a v2i64 that we can feed into the
1838 // floating point conversion. Zero or sign extend based on whether
1839 // we're doing a signed or unsigned float conversion.
1841 Op.getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
1842 assert(Op.getNumOperands() == 1 && "FP conversions take one argument");
1843 SDValue Promoted = DAG.getNode(Opc, dl, MVT::v2i64, Op.getOperand(0));
1844 return DAG.getNode(Op.getOpcode(), dl, Op.getValueType(), Promoted);
1847 // Scalarize v2i64 to v2f32 conversions.
1848 std::vector<SDValue> BuildVectorOps;
1849 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
1850 SDValue Sclr = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, In,
1851 DAG.getConstant(i, MVT::i64));
1852 Sclr = DAG.getNode(Op->getOpcode(), dl, MVT::f32, Sclr);
1853 BuildVectorOps.push_back(Sclr);
1856 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &BuildVectorOps[0],
1857 BuildVectorOps.size());
1860 SDValue ARM64TargetLowering::LowerINT_TO_FP(SDValue Op,
1861 SelectionDAG &DAG) const {
1862 if (Op.getValueType().isVector())
1863 return LowerVectorINT_TO_FP(Op, DAG);
1865 // i128 conversions are libcalls.
1866 if (Op.getOperand(0).getValueType() == MVT::i128)
1869 // Other conversions are legal, unless it's to the completely software-based
1871 if (Op.getValueType() != MVT::f128)
1875 if (Op.getOpcode() == ISD::SINT_TO_FP)
1876 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1878 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
1880 return LowerF128Call(Op, DAG, LC);
1883 SDValue ARM64TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
1884 // For iOS, we want to call an alternative entry point: __sincos_stret,
1885 // which returns the values in two S / D registers.
1887 SDValue Arg = Op.getOperand(0);
1888 EVT ArgVT = Arg.getValueType();
1889 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1896 Entry.isSExt = false;
1897 Entry.isZExt = false;
1898 Args.push_back(Entry);
1900 const char *LibcallName =
1901 (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
1902 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
1904 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
1905 TargetLowering::CallLoweringInfo CLI(
1906 DAG.getEntryNode(), RetTy, false, false, false, false, 0,
1907 CallingConv::Fast, /*isTaillCall=*/false,
1908 /*doesNotRet=*/false, /*isReturnValueUsed*/ true, Callee, Args, DAG, dl);
1909 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1910 return CallResult.first;
1913 SDValue ARM64TargetLowering::LowerOperation(SDValue Op,
1914 SelectionDAG &DAG) const {
1915 switch (Op.getOpcode()) {
1917 llvm_unreachable("unimplemented operand");
1919 case ISD::GlobalAddress:
1920 return LowerGlobalAddress(Op, DAG);
1921 case ISD::GlobalTLSAddress:
1922 return LowerGlobalTLSAddress(Op, DAG);
1924 return LowerSETCC(Op, DAG);
1926 return LowerBR_CC(Op, DAG);
1928 return LowerSELECT(Op, DAG);
1929 case ISD::SELECT_CC:
1930 return LowerSELECT_CC(Op, DAG);
1931 case ISD::JumpTable:
1932 return LowerJumpTable(Op, DAG);
1933 case ISD::ConstantPool:
1934 return LowerConstantPool(Op, DAG);
1935 case ISD::BlockAddress:
1936 return LowerBlockAddress(Op, DAG);
1938 return LowerVASTART(Op, DAG);
1940 return LowerVACOPY(Op, DAG);
1942 return LowerVAARG(Op, DAG);
1947 return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
1954 return LowerXALUO(Op, DAG);
1956 return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
1958 return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
1960 return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
1962 return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
1964 return LowerFP_ROUND(Op, DAG);
1965 case ISD::FP_EXTEND:
1966 return LowerFP_EXTEND(Op, DAG);
1967 case ISD::FRAMEADDR:
1968 return LowerFRAMEADDR(Op, DAG);
1969 case ISD::RETURNADDR:
1970 return LowerRETURNADDR(Op, DAG);
1971 case ISD::INSERT_VECTOR_ELT:
1972 return LowerINSERT_VECTOR_ELT(Op, DAG);
1973 case ISD::EXTRACT_VECTOR_ELT:
1974 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
1975 case ISD::BUILD_VECTOR:
1976 return LowerBUILD_VECTOR(Op, DAG);
1977 case ISD::VECTOR_SHUFFLE:
1978 return LowerVECTOR_SHUFFLE(Op, DAG);
1979 case ISD::EXTRACT_SUBVECTOR:
1980 return LowerEXTRACT_SUBVECTOR(Op, DAG);
1984 return LowerVectorSRA_SRL_SHL(Op, DAG);
1985 case ISD::SHL_PARTS:
1986 return LowerShiftLeftParts(Op, DAG);
1987 case ISD::SRL_PARTS:
1988 case ISD::SRA_PARTS:
1989 return LowerShiftRightParts(Op, DAG);
1991 return LowerCTPOP(Op, DAG);
1992 case ISD::FCOPYSIGN:
1993 return LowerFCOPYSIGN(Op, DAG);
1995 return LowerVectorAND(Op, DAG);
1997 return LowerVectorOR(Op, DAG);
1999 return LowerXOR(Op, DAG);
2001 return LowerPREFETCH(Op, DAG);
2002 case ISD::SINT_TO_FP:
2003 case ISD::UINT_TO_FP:
2004 return LowerINT_TO_FP(Op, DAG);
2005 case ISD::FP_TO_SINT:
2006 case ISD::FP_TO_UINT:
2007 return LowerFP_TO_INT(Op, DAG);
2009 return LowerFSINCOS(Op, DAG);
2013 /// getFunctionAlignment - Return the Log2 alignment of this function.
2014 unsigned ARM64TargetLowering::getFunctionAlignment(const Function *F) const {
2018 //===----------------------------------------------------------------------===//
2019 // Calling Convention Implementation
2020 //===----------------------------------------------------------------------===//
2022 #include "ARM64GenCallingConv.inc"
2024 /// Selects the correct CCAssignFn for a the given CallingConvention
2026 CCAssignFn *ARM64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
2027 bool IsVarArg) const {
2030 llvm_unreachable("Unsupported calling convention.");
2031 case CallingConv::WebKit_JS:
2032 return CC_ARM64_WebKit_JS;
2033 case CallingConv::C:
2034 case CallingConv::Fast:
2035 if (!Subtarget->isTargetDarwin())
2036 return CC_ARM64_AAPCS;
2037 return IsVarArg ? CC_ARM64_DarwinPCS_VarArg : CC_ARM64_DarwinPCS;
2041 SDValue ARM64TargetLowering::LowerFormalArguments(
2042 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2043 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2044 SmallVectorImpl<SDValue> &InVals) const {
2045 MachineFunction &MF = DAG.getMachineFunction();
2046 MachineFrameInfo *MFI = MF.getFrameInfo();
2048 // Assign locations to all of the incoming arguments.
2049 SmallVector<CCValAssign, 16> ArgLocs;
2050 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2051 getTargetMachine(), ArgLocs, *DAG.getContext());
2053 // At this point, Ins[].VT may already be promoted to i32. To correctly
2054 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2055 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2056 // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
2057 // we use a special version of AnalyzeFormalArguments to pass in ValVT and
2059 unsigned NumArgs = Ins.size();
2060 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
2061 unsigned CurArgIdx = 0;
2062 for (unsigned i = 0; i != NumArgs; ++i) {
2063 MVT ValVT = Ins[i].VT;
2064 std::advance(CurOrigArg, Ins[i].OrigArgIndex - CurArgIdx);
2065 CurArgIdx = Ins[i].OrigArgIndex;
2067 // Get type of the original argument.
2068 EVT ActualVT = getValueType(CurOrigArg->getType(), /*AllowUnknown*/ true);
2069 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
2070 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2072 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2074 else if (ActualMVT == MVT::i16)
2077 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2079 AssignFn(i, ValVT, LocVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
2080 assert(!Res && "Call operand has unhandled type");
2084 SmallVector<SDValue, 16> ArgValues;
2085 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2086 CCValAssign &VA = ArgLocs[i];
2088 // Arguments stored in registers.
2089 if (VA.isRegLoc()) {
2090 EVT RegVT = VA.getLocVT();
2093 const TargetRegisterClass *RC;
2095 if (RegVT == MVT::i32)
2096 RC = &ARM64::GPR32RegClass;
2097 else if (RegVT == MVT::i64)
2098 RC = &ARM64::GPR64RegClass;
2099 else if (RegVT == MVT::f32)
2100 RC = &ARM64::FPR32RegClass;
2101 else if (RegVT == MVT::f64 || RegVT == MVT::v1i64 ||
2102 RegVT == MVT::v1f64 || RegVT == MVT::v2i32 ||
2103 RegVT == MVT::v4i16 || RegVT == MVT::v8i8)
2104 RC = &ARM64::FPR64RegClass;
2105 else if (RegVT == MVT::v2i64 || RegVT == MVT::v4i32 ||
2106 RegVT == MVT::v8i16 || RegVT == MVT::v16i8)
2107 RC = &ARM64::FPR128RegClass;
2109 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2111 // Transform the arguments in physical registers into virtual ones.
2112 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2113 ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
2115 // If this is an 8, 16 or 32-bit value, it is really passed promoted
2116 // to 64 bits. Insert an assert[sz]ext to capture this, then
2117 // truncate to the right size.
2118 switch (VA.getLocInfo()) {
2120 llvm_unreachable("Unknown loc info!");
2121 case CCValAssign::Full:
2123 case CCValAssign::BCvt:
2124 ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
2126 case CCValAssign::SExt:
2127 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
2128 DAG.getValueType(VA.getValVT()));
2129 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2131 case CCValAssign::ZExt:
2132 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
2133 DAG.getValueType(VA.getValVT()));
2134 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
2138 InVals.push_back(ArgValue);
2140 } else { // VA.isRegLoc()
2141 assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
2142 unsigned ArgOffset = VA.getLocMemOffset();
2143 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
2144 int FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
2146 // Create load nodes to retrieve arguments from the stack.
2147 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2148 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, FIN,
2149 MachinePointerInfo::getFixedStack(FI), false,
2156 if (!Subtarget->isTargetDarwin()) {
2157 // The AAPCS variadic function ABI is identical to the non-variadic
2158 // one. As a result there may be more arguments in registers and we should
2159 // save them for future reference.
2160 saveVarArgRegisters(CCInfo, DAG, DL, Chain);
2163 ARM64FunctionInfo *AFI = MF.getInfo<ARM64FunctionInfo>();
2164 // This will point to the next argument passed via stack.
2165 unsigned StackOffset = CCInfo.getNextStackOffset();
2166 // We currently pass all varargs at 8-byte alignment.
2167 StackOffset = ((StackOffset + 7) & ~7);
2168 AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
2174 void ARM64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
2175 SelectionDAG &DAG, SDLoc DL,
2176 SDValue &Chain) const {
2177 MachineFunction &MF = DAG.getMachineFunction();
2178 MachineFrameInfo *MFI = MF.getFrameInfo();
2179 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
2181 SmallVector<SDValue, 8> MemOps;
2183 static const MCPhysReg GPRArgRegs[] = { ARM64::X0, ARM64::X1, ARM64::X2,
2184 ARM64::X3, ARM64::X4, ARM64::X5,
2185 ARM64::X6, ARM64::X7 };
2186 static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
2187 unsigned FirstVariadicGPR =
2188 CCInfo.getFirstUnallocated(GPRArgRegs, NumGPRArgRegs);
2190 static const MCPhysReg FPRArgRegs[] = { ARM64::Q0, ARM64::Q1, ARM64::Q2,
2191 ARM64::Q3, ARM64::Q4, ARM64::Q5,
2192 ARM64::Q6, ARM64::Q7 };
2193 static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
2194 unsigned FirstVariadicFPR =
2195 CCInfo.getFirstUnallocated(FPRArgRegs, NumFPRArgRegs);
2197 unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
2199 if (GPRSaveSize != 0) {
2200 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
2202 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
2204 for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
2205 unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &ARM64::GPR64RegClass);
2206 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
2208 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2209 MachinePointerInfo::getStack(i * 8), false, false, 0);
2210 MemOps.push_back(Store);
2211 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2212 DAG.getConstant(8, getPointerTy()));
2216 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
2218 if (FPRSaveSize != 0) {
2219 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
2221 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
2223 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
2224 unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &ARM64::FPR128RegClass);
2225 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::v2i64);
2227 DAG.getStore(Val.getValue(1), DL, Val, FIN,
2228 MachinePointerInfo::getStack(i * 16), false, false, 0);
2229 MemOps.push_back(Store);
2230 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
2231 DAG.getConstant(16, getPointerTy()));
2235 FuncInfo->setVarArgsGPRIndex(GPRIdx);
2236 FuncInfo->setVarArgsGPRSize(GPRSaveSize);
2237 FuncInfo->setVarArgsFPRIndex(FPRIdx);
2238 FuncInfo->setVarArgsFPRSize(FPRSaveSize);
2240 if (!MemOps.empty()) {
2241 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
2246 /// LowerCallResult - Lower the result values of a call into the
2247 /// appropriate copies out of appropriate physical registers.
2248 SDValue ARM64TargetLowering::LowerCallResult(
2249 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
2250 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2251 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
2252 SDValue ThisVal) const {
2253 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2254 : RetCC_ARM64_AAPCS;
2255 // Assign locations to each value returned by this call.
2256 SmallVector<CCValAssign, 16> RVLocs;
2257 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2258 getTargetMachine(), RVLocs, *DAG.getContext());
2259 CCInfo.AnalyzeCallResult(Ins, RetCC);
2261 // Copy all of the result registers out of their specified physreg.
2262 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2263 CCValAssign VA = RVLocs[i];
2265 // Pass 'this' value directly from the argument to return value, to avoid
2266 // reg unit interference
2267 if (i == 0 && isThisReturn) {
2268 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
2269 "unexpected return calling convention register assignment");
2270 InVals.push_back(ThisVal);
2275 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InFlag);
2276 Chain = Val.getValue(1);
2277 InFlag = Val.getValue(2);
2279 switch (VA.getLocInfo()) {
2281 llvm_unreachable("Unknown loc info!");
2282 case CCValAssign::Full:
2284 case CCValAssign::BCvt:
2285 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2289 InVals.push_back(Val);
2295 bool ARM64TargetLowering::isEligibleForTailCallOptimization(
2296 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
2297 bool isCalleeStructRet, bool isCallerStructRet,
2298 const SmallVectorImpl<ISD::OutputArg> &Outs,
2299 const SmallVectorImpl<SDValue> &OutVals,
2300 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
2301 // Look for obvious safe cases to perform tail call optimization that do not
2302 // require ABI changes. This is what gcc calls sibcall.
2304 // Do not sibcall optimize vararg calls unless the call site is not passing
2306 if (isVarArg && !Outs.empty())
2309 // Also avoid sibcall optimization if either caller or callee uses struct
2310 // return semantics.
2311 if (isCalleeStructRet || isCallerStructRet)
2314 // Note that currently ARM64 "C" calling convention and "Fast" calling
2315 // convention are compatible. If/when that ever changes, we'll need to
2316 // add checks here to make sure any interactions are OK.
2318 // If the callee takes no arguments then go on to check the results of the
2320 if (!Outs.empty()) {
2321 // Check if stack adjustment is needed. For now, do not do this if any
2322 // argument is passed on the stack.
2323 SmallVector<CCValAssign, 16> ArgLocs;
2324 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2325 getTargetMachine(), ArgLocs, *DAG.getContext());
2326 CCAssignFn *AssignFn = CCAssignFnForCall(CalleeCC, /*IsVarArg=*/false);
2327 CCInfo.AnalyzeCallOperands(Outs, AssignFn);
2328 if (CCInfo.getNextStackOffset()) {
2329 // Check if the arguments are already laid out in the right way as
2330 // the caller's fixed stack objects.
2331 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2332 ++i, ++realArgIdx) {
2333 CCValAssign &VA = ArgLocs[i];
2334 if (VA.getLocInfo() == CCValAssign::Indirect)
2336 if (VA.needsCustom()) {
2337 // Just don't handle anything that needs custom adjustments for now.
2338 // If need be, we can revisit later, but we shouldn't ever end up
2341 } else if (!VA.isRegLoc()) {
2342 // Likewise, don't try to handle stack based arguments for the
2352 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
2353 /// and add input and output parameter nodes.
2354 SDValue ARM64TargetLowering::LowerCall(CallLoweringInfo &CLI,
2355 SmallVectorImpl<SDValue> &InVals) const {
2356 SelectionDAG &DAG = CLI.DAG;
2358 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2359 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2360 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2361 SDValue Chain = CLI.Chain;
2362 SDValue Callee = CLI.Callee;
2363 bool &IsTailCall = CLI.IsTailCall;
2364 CallingConv::ID CallConv = CLI.CallConv;
2365 bool IsVarArg = CLI.IsVarArg;
2367 MachineFunction &MF = DAG.getMachineFunction();
2368 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
2369 bool IsThisReturn = false;
2371 // If tail calls are explicitly disabled, make sure not to use them.
2372 if (!EnableARM64TailCalls)
2376 // Check if it's really possible to do a tail call.
2377 IsTailCall = isEligibleForTailCallOptimization(
2378 Callee, CallConv, IsVarArg, IsStructRet,
2379 MF.getFunction()->hasStructRetAttr(), Outs, OutVals, Ins, DAG);
2380 // We don't support GuaranteedTailCallOpt, only automatically
2381 // detected sibcalls.
2382 // FIXME: Re-evaluate. Is this true? Should it be true?
2387 // Analyze operands of the call, assigning locations to each operand.
2388 SmallVector<CCValAssign, 16> ArgLocs;
2389 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
2390 getTargetMachine(), ArgLocs, *DAG.getContext());
2393 // Handle fixed and variable vector arguments differently.
2394 // Variable vector arguments always go into memory.
2395 unsigned NumArgs = Outs.size();
2397 for (unsigned i = 0; i != NumArgs; ++i) {
2398 MVT ArgVT = Outs[i].VT;
2399 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2400 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
2401 /*IsVarArg=*/ !Outs[i].IsFixed);
2402 bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
2403 assert(!Res && "Call operand has unhandled type");
2407 // At this point, Outs[].VT may already be promoted to i32. To correctly
2408 // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
2409 // i8 to CC_ARM64_AAPCS with i32 being ValVT and i8 being LocVT.
2410 // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
2411 // we use a special version of AnalyzeCallOperands to pass in ValVT and
2413 unsigned NumArgs = Outs.size();
2414 for (unsigned i = 0; i != NumArgs; ++i) {
2415 MVT ValVT = Outs[i].VT;
2416 // Get type of the original argument.
2417 EVT ActualVT = getValueType(CLI.Args[Outs[i].OrigArgIndex].Ty,
2418 /*AllowUnknown*/ true);
2419 MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
2420 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2421 // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
2423 if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
2425 else if (ActualMVT == MVT::i16)
2428 CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
2429 bool Res = AssignFn(i, ValVT, LocVT, CCValAssign::Full, ArgFlags, CCInfo);
2430 assert(!Res && "Call operand has unhandled type");
2435 // Get a count of how many bytes are to be pushed on the stack.
2436 unsigned NumBytes = CCInfo.getNextStackOffset();
2438 // Adjust the stack pointer for the new arguments...
2439 // These operations are automatically eliminated by the prolog/epilog pass
2442 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true), DL);
2444 SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, ARM64::SP, getPointerTy());
2446 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2447 SmallVector<SDValue, 8> MemOpChains;
2449 // Walk the register/memloc assignments, inserting copies/loads.
2450 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); i != e;
2451 ++i, ++realArgIdx) {
2452 CCValAssign &VA = ArgLocs[i];
2453 SDValue Arg = OutVals[realArgIdx];
2454 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2456 // Promote the value if needed.
2457 switch (VA.getLocInfo()) {
2459 llvm_unreachable("Unknown loc info!");
2460 case CCValAssign::Full:
2462 case CCValAssign::SExt:
2463 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
2465 case CCValAssign::ZExt:
2466 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
2468 case CCValAssign::AExt:
2469 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
2471 case CCValAssign::BCvt:
2472 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2474 case CCValAssign::FPExt:
2475 Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
2479 if (VA.isRegLoc()) {
2480 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
2481 assert(VA.getLocVT() == MVT::i64 &&
2482 "unexpected calling convention register assignment");
2483 assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
2484 "unexpected use of 'returned'");
2485 IsThisReturn = true;
2487 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2489 assert(VA.isMemLoc());
2490 // There's no reason we can't support stack args w/ tailcall, but
2491 // we currently don't, so assert if we see one.
2492 assert(!IsTailCall && "stack argument with tail call!?");
2493 unsigned LocMemOffset = VA.getLocMemOffset();
2494 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2495 PtrOff = DAG.getNode(ISD::ADD, DL, getPointerTy(), StackPtr, PtrOff);
2497 // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
2498 // promoted to a legal register type i32, we should truncate Arg back to
2500 if (Arg.getValueType().isSimple() &&
2501 Arg.getValueType().getSimpleVT() == MVT::i32 &&
2502 (VA.getLocVT() == MVT::i1 || VA.getLocVT() == MVT::i8 ||
2503 VA.getLocVT() == MVT::i16))
2504 Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getLocVT(), Arg);
2506 SDValue Store = DAG.getStore(Chain, DL, Arg, PtrOff,
2507 MachinePointerInfo::getStack(LocMemOffset),
2509 MemOpChains.push_back(Store);
2513 if (!MemOpChains.empty())
2514 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOpChains[0],
2515 MemOpChains.size());
2517 // Build a sequence of copy-to-reg nodes chained together with token chain
2518 // and flag operands which copy the outgoing args into the appropriate regs.
2520 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2521 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[i].first,
2522 RegsToPass[i].second, InFlag);
2523 InFlag = Chain.getValue(1);
2526 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2527 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2528 // node so that legalize doesn't hack it.
2529 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
2530 Subtarget->isTargetMachO()) {
2531 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2532 const GlobalValue *GV = G->getGlobal();
2533 bool InternalLinkage = GV->hasInternalLinkage();
2534 if (InternalLinkage)
2535 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2537 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0,
2539 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2541 } else if (ExternalSymbolSDNode *S =
2542 dyn_cast<ExternalSymbolSDNode>(Callee)) {
2543 const char *Sym = S->getSymbol();
2545 DAG.getTargetExternalSymbol(Sym, getPointerTy(), ARM64II::MO_GOT);
2546 Callee = DAG.getNode(ARM64ISD::LOADgot, DL, getPointerTy(), Callee);
2548 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2549 const GlobalValue *GV = G->getGlobal();
2550 Callee = DAG.getTargetGlobalAddress(GV, DL, getPointerTy(), 0, 0);
2551 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2552 const char *Sym = S->getSymbol();
2553 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), 0);
2556 std::vector<SDValue> Ops;
2557 Ops.push_back(Chain);
2558 Ops.push_back(Callee);
2560 // Add argument registers to the end of the list so that they are known live
2562 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2563 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2564 RegsToPass[i].second.getValueType()));
2566 // Add a register mask operand representing the call-preserved registers.
2567 const uint32_t *Mask;
2568 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2569 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2571 // For 'this' returns, use the X0-preserving mask if applicable
2572 Mask = ARI->getThisReturnPreservedMask(CallConv);
2574 IsThisReturn = false;
2575 Mask = ARI->getCallPreservedMask(CallConv);
2578 Mask = ARI->getCallPreservedMask(CallConv);
2580 assert(Mask && "Missing call preserved mask for calling convention");
2581 Ops.push_back(DAG.getRegisterMask(Mask));
2583 if (InFlag.getNode())
2584 Ops.push_back(InFlag);
2586 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2588 // If we're doing a tall call, use a TC_RETURN here rather than an
2589 // actual call instruction.
2591 return DAG.getNode(ARM64ISD::TC_RETURN, DL, NodeTys, &Ops[0], Ops.size());
2593 // Returns a chain and a flag for retval copy to use.
2594 Chain = DAG.getNode(ARM64ISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
2595 InFlag = Chain.getValue(1);
2597 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2598 DAG.getIntPtrConstant(0, true), InFlag, DL);
2600 InFlag = Chain.getValue(1);
2602 // Handle result values, copying them out of physregs into vregs that we
2604 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2605 InVals, IsThisReturn,
2606 IsThisReturn ? OutVals[0] : SDValue());
2609 bool ARM64TargetLowering::CanLowerReturn(
2610 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2611 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2612 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2613 : RetCC_ARM64_AAPCS;
2614 SmallVector<CCValAssign, 16> RVLocs;
2615 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
2616 return CCInfo.CheckReturn(Outs, RetCC);
2620 ARM64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2622 const SmallVectorImpl<ISD::OutputArg> &Outs,
2623 const SmallVectorImpl<SDValue> &OutVals,
2624 SDLoc DL, SelectionDAG &DAG) const {
2625 CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS ? RetCC_ARM64_WebKit_JS
2626 : RetCC_ARM64_AAPCS;
2627 SmallVector<CCValAssign, 16> RVLocs;
2628 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2629 getTargetMachine(), RVLocs, *DAG.getContext());
2630 CCInfo.AnalyzeReturn(Outs, RetCC);
2632 // Copy the result values into the output registers.
2634 SmallVector<SDValue, 4> RetOps(1, Chain);
2635 for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
2636 ++i, ++realRVLocIdx) {
2637 CCValAssign &VA = RVLocs[i];
2638 assert(VA.isRegLoc() && "Can only return in registers!");
2639 SDValue Arg = OutVals[realRVLocIdx];
2641 switch (VA.getLocInfo()) {
2643 llvm_unreachable("Unknown loc info!");
2644 case CCValAssign::Full:
2646 case CCValAssign::BCvt:
2647 Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
2651 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
2652 Flag = Chain.getValue(1);
2653 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2656 RetOps[0] = Chain; // Update chain.
2658 // Add the flag if we have it.
2660 RetOps.push_back(Flag);
2662 return DAG.getNode(ARM64ISD::RET_FLAG, DL, MVT::Other, &RetOps[0],
2666 //===----------------------------------------------------------------------===//
2667 // Other Lowering Code
2668 //===----------------------------------------------------------------------===//
2670 SDValue ARM64TargetLowering::LowerGlobalAddress(SDValue Op,
2671 SelectionDAG &DAG) const {
2672 EVT PtrVT = getPointerTy();
2674 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2675 unsigned char OpFlags =
2676 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
2678 assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
2679 "unexpected offset in global node");
2681 // This also catched the large code model case for Darwin.
2682 if ((OpFlags & ARM64II::MO_GOT) != 0) {
2683 SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2684 // FIXME: Once remat is capable of dealing with instructions with register
2685 // operands, expand this into two nodes instead of using a wrapper node.
2686 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
2689 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2690 const unsigned char MO_NC = ARM64II::MO_NC;
2692 ARM64ISD::WrapperLarge, DL, PtrVT,
2693 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G3),
2694 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
2695 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
2696 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
2698 // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
2699 // the only correct model on Darwin.
2700 SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2701 OpFlags | ARM64II::MO_PAGE);
2702 unsigned char LoFlags = OpFlags | ARM64II::MO_PAGEOFF | ARM64II::MO_NC;
2703 SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);
2705 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
2706 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
2710 /// \brief Convert a TLS address reference into the correct sequence of loads
2711 /// and calls to compute the variable's address (for Darwin, currently) and
2712 /// return an SDValue containing the final node.
2714 /// Darwin only has one TLS scheme which must be capable of dealing with the
2715 /// fully general situation, in the worst case. This means:
2716 /// + "extern __thread" declaration.
2717 /// + Defined in a possibly unknown dynamic library.
2719 /// The general system is that each __thread variable has a [3 x i64] descriptor
2720 /// which contains information used by the runtime to calculate the address. The
2721 /// only part of this the compiler needs to know about is the first xword, which
2722 /// contains a function pointer that must be called with the address of the
2723 /// entire descriptor in "x0".
2725 /// Since this descriptor may be in a different unit, in general even the
2726 /// descriptor must be accessed via an indirect load. The "ideal" code sequence
2728 /// adrp x0, _var@TLVPPAGE
2729 /// ldr x0, [x0, _var@TLVPPAGEOFF] ; x0 now contains address of descriptor
2730 /// ldr x1, [x0] ; x1 contains 1st entry of descriptor,
2731 /// ; the function pointer
2732 /// blr x1 ; Uses descriptor address in x0
2733 /// ; Address of _var is now in x0.
2735 /// If the address of _var's descriptor *is* known to the linker, then it can
2736 /// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
2737 /// a slight efficiency gain.
2739 ARM64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
2740 SelectionDAG &DAG) const {
2741 assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");
2744 MVT PtrVT = getPointerTy();
2745 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2748 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2749 SDValue DescAddr = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TLVPAddr);
2751 // The first entry in the descriptor is a function pointer that we must call
2752 // to obtain the address of the variable.
2753 SDValue Chain = DAG.getEntryNode();
2754 SDValue FuncTLVGet =
2755 DAG.getLoad(MVT::i64, DL, Chain, DescAddr, MachinePointerInfo::getGOT(),
2756 false, true, true, 8);
2757 Chain = FuncTLVGet.getValue(1);
2759 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2760 MFI->setAdjustsStack(true);
2762 // TLS calls preserve all registers except those that absolutely must be
2763 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2765 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2766 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2767 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2769 // Finally, we can make the call. This is just a degenerate version of a
2770 // normal ARM64 call node: x0 takes the address of the descriptor, and returns
2771 // the address of the variable in this thread.
2772 Chain = DAG.getCopyToReg(Chain, DL, ARM64::X0, DescAddr, SDValue());
2773 Chain = DAG.getNode(ARM64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
2774 Chain, FuncTLVGet, DAG.getRegister(ARM64::X0, MVT::i64),
2775 DAG.getRegisterMask(Mask), Chain.getValue(1));
2776 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Chain.getValue(1));
2779 /// When accessing thread-local variables under either the general-dynamic or
2780 /// local-dynamic system, we make a "TLS-descriptor" call. The variable will
2781 /// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
2782 /// is a function pointer to carry out the resolution. This function takes the
2783 /// address of the descriptor in X0 and returns the TPIDR_EL0 offset in X0. All
2784 /// other registers (except LR, CPSR) are preserved.
2786 /// Thus, the ideal call sequence on AArch64 is:
2788 /// adrp x0, :tlsdesc:thread_var
2789 /// ldr x8, [x0, :tlsdesc_lo12:thread_var]
2790 /// add x0, x0, :tlsdesc_lo12:thread_var
2791 /// .tlsdesccall thread_var
2793 /// (TPIDR_EL0 offset now in x0).
2795 /// The ".tlsdesccall" directive instructs the assembler to insert a particular
2796 /// relocation to help the linker relax this sequence if it turns out to be too
2799 /// FIXME: we currently produce an extra, duplicated, ADRP instruction, but this
2801 SDValue ARM64TargetLowering::LowerELFTLSDescCall(SDValue SymAddr,
2802 SDValue DescAddr, SDLoc DL,
2803 SelectionDAG &DAG) const {
2804 EVT PtrVT = getPointerTy();
2806 // The function we need to call is simply the first entry in the GOT for this
2807 // descriptor, load it in preparation.
2808 SDValue Func = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, SymAddr);
2810 // TLS calls preserve all registers except those that absolutely must be
2811 // trashed: X0 (it takes an argument), LR (it's a call) and CPSR (let's not be
2813 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2814 const ARM64RegisterInfo *ARI = static_cast<const ARM64RegisterInfo *>(TRI);
2815 const uint32_t *Mask = ARI->getTLSCallPreservedMask();
2817 // The function takes only one argument: the address of the descriptor itself
2819 SDValue Glue, Chain;
2820 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, ARM64::X0, DescAddr, Glue);
2821 Glue = Chain.getValue(1);
2823 // We're now ready to populate the argument list, as with a normal call:
2824 SmallVector<SDValue, 6> Ops;
2825 Ops.push_back(Chain);
2826 Ops.push_back(Func);
2827 Ops.push_back(SymAddr);
2828 Ops.push_back(DAG.getRegister(ARM64::X0, PtrVT));
2829 Ops.push_back(DAG.getRegisterMask(Mask));
2830 Ops.push_back(Glue);
2832 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2833 Chain = DAG.getNode(ARM64ISD::TLSDESC_CALL, DL, NodeTys, &Ops[0], Ops.size());
2834 Glue = Chain.getValue(1);
2836 return DAG.getCopyFromReg(Chain, DL, ARM64::X0, PtrVT, Glue);
2839 SDValue ARM64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
2840 SelectionDAG &DAG) const {
2841 assert(Subtarget->isTargetELF() && "This function expects an ELF target");
2842 assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
2843 "ELF TLS only supported in small memory model");
2844 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2846 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2849 EVT PtrVT = getPointerTy();
2851 const GlobalValue *GV = GA->getGlobal();
2853 SDValue ThreadBase = DAG.getNode(ARM64ISD::THREAD_POINTER, DL, PtrVT);
2855 if (Model == TLSModel::LocalExec) {
2856 SDValue HiVar = DAG.getTargetGlobalAddress(
2857 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2858 SDValue LoVar = DAG.getTargetGlobalAddress(
2859 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2861 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2862 DAG.getTargetConstant(16, MVT::i32)),
2864 TPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, TPOff, LoVar,
2865 DAG.getTargetConstant(0, MVT::i32)),
2867 } else if (Model == TLSModel::InitialExec) {
2868 TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2869 TPOff = DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, TPOff);
2870 } else if (Model == TLSModel::LocalDynamic) {
2871 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2872 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2873 // the beginning of the module's TLS region, followed by a DTPREL offset
2876 // These accesses will need deduplicating if there's more than one.
2877 ARM64FunctionInfo *MFI =
2878 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
2879 MFI->incNumLocalDynamicTLSAccesses();
2881 // Accesses used in this sequence go via the TLS descriptor which lives in
2882 // the GOT. Prepare an address we can use to handle this.
2883 SDValue HiDesc = DAG.getTargetExternalSymbol(
2884 "_TLS_MODULE_BASE_", PtrVT, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2885 SDValue LoDesc = DAG.getTargetExternalSymbol(
2886 "_TLS_MODULE_BASE_", PtrVT,
2887 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2889 // First argument to the descriptor call is the address of the descriptor
2891 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2892 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2894 // The call needs a relocation too for linker relaxation. It doesn't make
2895 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2897 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2900 // Now we can calculate the offset from TPIDR_EL0 to this module's
2901 // thread-local area.
2902 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2904 // Now use :dtprel_whatever: operations to calculate this variable's offset
2905 // in its thread-storage area.
2906 SDValue HiVar = DAG.getTargetGlobalAddress(
2907 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G1);
2908 SDValue LoVar = DAG.getTargetGlobalAddress(
2909 GV, DL, MVT::i64, 0, ARM64II::MO_TLS | ARM64II::MO_G0 | ARM64II::MO_NC);
2912 SDValue(DAG.getMachineNode(ARM64::MOVZXi, DL, PtrVT, HiVar,
2913 DAG.getTargetConstant(16, MVT::i32)),
2915 DTPOff = SDValue(DAG.getMachineNode(ARM64::MOVKXi, DL, PtrVT, DTPOff, LoVar,
2916 DAG.getTargetConstant(0, MVT::i32)),
2919 TPOff = DAG.getNode(ISD::ADD, DL, PtrVT, TPOff, DTPOff);
2920 } else if (Model == TLSModel::GeneralDynamic) {
2921 // Accesses used in this sequence go via the TLS descriptor which lives in
2922 // the GOT. Prepare an address we can use to handle this.
2923 SDValue HiDesc = DAG.getTargetGlobalAddress(
2924 GV, DL, PtrVT, 0, ARM64II::MO_TLS | ARM64II::MO_PAGE);
2925 SDValue LoDesc = DAG.getTargetGlobalAddress(
2927 ARM64II::MO_TLS | ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
2929 // First argument to the descriptor call is the address of the descriptor
2931 SDValue DescAddr = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, HiDesc);
2932 DescAddr = DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, DescAddr, LoDesc);
2934 // The call needs a relocation too for linker relaxation. It doesn't make
2935 // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
2938 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, ARM64II::MO_TLS);
2940 // Finally we can make a call to calculate the offset from tpidr_el0.
2941 TPOff = LowerELFTLSDescCall(SymAddr, DescAddr, DL, DAG);
2943 llvm_unreachable("Unsupported ELF TLS access model");
2945 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2948 SDValue ARM64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2949 SelectionDAG &DAG) const {
2950 if (Subtarget->isTargetDarwin())
2951 return LowerDarwinGlobalTLSAddress(Op, DAG);
2952 else if (Subtarget->isTargetELF())
2953 return LowerELFGlobalTLSAddress(Op, DAG);
2955 llvm_unreachable("Unexpected platform trying to use TLS");
2957 SDValue ARM64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2958 SDValue Chain = Op.getOperand(0);
2959 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2960 SDValue LHS = Op.getOperand(2);
2961 SDValue RHS = Op.getOperand(3);
2962 SDValue Dest = Op.getOperand(4);
2965 // Handle f128 first, since lowering it will result in comparing the return
2966 // value of a libcall against zero, which is just what the rest of LowerBR_CC
2967 // is expecting to deal with.
2968 if (LHS.getValueType() == MVT::f128) {
2969 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2971 // If softenSetCCOperands returned a scalar, we need to compare the result
2972 // against zero to select between true and false values.
2973 if (RHS.getNode() == 0) {
2974 RHS = DAG.getConstant(0, LHS.getValueType());
2979 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
2981 unsigned Opc = LHS.getOpcode();
2982 if (LHS.getResNo() == 1 && isa<ConstantSDNode>(RHS) &&
2983 cast<ConstantSDNode>(RHS)->isOne() &&
2984 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
2985 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
2986 assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
2987 "Unexpected condition code.");
2988 // Only lower legal XALUO ops.
2989 if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
2992 // The actual operation with overflow check.
2993 ARM64CC::CondCode OFCC;
2994 SDValue Value, Overflow;
2995 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, LHS.getValue(0), DAG);
2997 if (CC == ISD::SETNE)
2998 OFCC = getInvertedCondCode(OFCC);
2999 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3001 return DAG.getNode(ARM64ISD::BRCOND, SDLoc(LHS), MVT::Other, Chain, Dest,
3005 if (LHS.getValueType().isInteger()) {
3006 assert((LHS.getValueType() == RHS.getValueType()) &&
3007 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3009 // If the RHS of the comparison is zero, we can potentially fold this
3010 // to a specialized branch.
3011 const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
3012 if (RHSC && RHSC->getZExtValue() == 0) {
3013 if (CC == ISD::SETEQ) {
3014 // See if we can use a TBZ to fold in an AND as well.
3015 // TBZ has a smaller branch displacement than CBZ. If the offset is
3016 // out of bounds, a late MI-layer pass rewrites branches.
3017 // 403.gcc is an example that hits this case.
3018 if (LHS.getOpcode() == ISD::AND &&
3019 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3020 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3021 SDValue Test = LHS.getOperand(0);
3022 uint64_t Mask = LHS.getConstantOperandVal(1);
3024 // TBZ only operates on i64's, but the ext should be free.
3025 if (Test.getValueType() == MVT::i32)
3026 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3028 return DAG.getNode(ARM64ISD::TBZ, dl, MVT::Other, Chain, Test,
3029 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3032 return DAG.getNode(ARM64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
3033 } else if (CC == ISD::SETNE) {
3034 // See if we can use a TBZ to fold in an AND as well.
3035 // TBZ has a smaller branch displacement than CBZ. If the offset is
3036 // out of bounds, a late MI-layer pass rewrites branches.
3037 // 403.gcc is an example that hits this case.
3038 if (LHS.getOpcode() == ISD::AND &&
3039 isa<ConstantSDNode>(LHS.getOperand(1)) &&
3040 isPowerOf2_64(LHS.getConstantOperandVal(1))) {
3041 SDValue Test = LHS.getOperand(0);
3042 uint64_t Mask = LHS.getConstantOperandVal(1);
3044 // TBNZ only operates on i64's, but the ext should be free.
3045 if (Test.getValueType() == MVT::i32)
3046 Test = DAG.getAnyExtOrTrunc(Test, dl, MVT::i64);
3048 return DAG.getNode(ARM64ISD::TBNZ, dl, MVT::Other, Chain, Test,
3049 DAG.getConstant(Log2_64(Mask), MVT::i64), Dest);
3052 return DAG.getNode(ARM64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
3057 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3058 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
3062 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3064 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3065 // clean. Some of them require two branches to implement.
3066 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3067 ARM64CC::CondCode CC1, CC2;
3068 changeFPCCToARM64CC(CC, CC1, CC2);
3069 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3071 DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
3072 if (CC2 != ARM64CC::AL) {
3073 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3074 return DAG.getNode(ARM64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
3081 SDValue ARM64TargetLowering::LowerFCOPYSIGN(SDValue Op,
3082 SelectionDAG &DAG) const {
3083 EVT VT = Op.getValueType();
3086 SDValue In1 = Op.getOperand(0);
3087 SDValue In2 = Op.getOperand(1);
3088 EVT SrcVT = In2.getValueType();
3090 if (SrcVT == MVT::f32 && VT == MVT::f64)
3091 In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
3092 else if (SrcVT == MVT::f64 && VT == MVT::f32)
3093 In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0));
3095 // FIXME: Src type is different, bail out for now. Can VT really be a
3102 SDValue EltMask, VecVal1, VecVal2;
3103 if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
3106 EltMask = DAG.getConstant(0x80000000ULL, EltVT);
3108 if (!VT.isVector()) {
3109 VecVal1 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3110 DAG.getUNDEF(VecVT), In1);
3111 VecVal2 = DAG.getTargetInsertSubreg(ARM64::ssub, DL, VecVT,
3112 DAG.getUNDEF(VecVT), In2);
3114 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3115 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3117 } else if (VT == MVT::f64 || VT == MVT::v2f64) {
3121 // We want to materialize a mask with the the high bit set, but the AdvSIMD
3122 // immediate moves cannot materialize that in a single instruction for
3123 // 64-bit elements. Instead, materialize zero and then negate it.
3124 EltMask = DAG.getConstant(0, EltVT);
3126 if (!VT.isVector()) {
3127 VecVal1 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3128 DAG.getUNDEF(VecVT), In1);
3129 VecVal2 = DAG.getTargetInsertSubreg(ARM64::dsub, DL, VecVT,
3130 DAG.getUNDEF(VecVT), In2);
3132 VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
3133 VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
3136 llvm_unreachable("Invalid type for copysign!");
3139 std::vector<SDValue> BuildVectorOps;
3140 for (unsigned i = 0; i < VecVT.getVectorNumElements(); ++i)
3141 BuildVectorOps.push_back(EltMask);
3143 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT,
3144 &BuildVectorOps[0], BuildVectorOps.size());
3146 // If we couldn't materialize the mask above, then the mask vector will be
3147 // the zero vector, and we need to negate it here.
3148 if (VT == MVT::f64 || VT == MVT::v2f64) {
3149 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
3150 BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
3151 BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
3155 DAG.getNode(ARM64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);
3158 return DAG.getTargetExtractSubreg(ARM64::ssub, DL, VT, Sel);
3159 else if (VT == MVT::f64)
3160 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, VT, Sel);
3162 return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
3165 SDValue ARM64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
3166 if (DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
3167 AttributeSet::FunctionIndex, Attribute::NoImplicitFloat))
3170 // While there is no integer popcount instruction, it can
3171 // be more efficiently lowered to the following sequence that uses
3172 // AdvSIMD registers/instructions as long as the copies to/from
3173 // the AdvSIMD registers are cheap.
3174 // FMOV D0, X0 // copy 64-bit int to vector, high bits zero'd
3175 // CNT V0.8B, V0.8B // 8xbyte pop-counts
3176 // ADDV B0, V0.8B // sum 8xbyte pop-counts
3177 // UMOV X0, V0.B[0] // copy byte result back to integer reg
3178 SDValue Val = Op.getOperand(0);
3180 EVT VT = Op.getValueType();
3181 SDValue ZeroVec = DAG.getUNDEF(MVT::v8i8);
3184 if (VT == MVT::i32) {
3185 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Val);
3187 DAG.getTargetInsertSubreg(ARM64::ssub, DL, MVT::v8i8, ZeroVec, VecVal);
3189 VecVal = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);
3192 SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, VecVal);
3193 SDValue UaddLV = DAG.getNode(
3194 ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
3195 DAG.getConstant(Intrinsic::arm64_neon_uaddlv, MVT::i32), CtPop);
3198 UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
3202 SDValue ARM64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3204 if (Op.getValueType().isVector())
3205 return LowerVSETCC(Op, DAG);
3207 SDValue LHS = Op.getOperand(0);
3208 SDValue RHS = Op.getOperand(1);
3209 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3212 // We chose ZeroOrOneBooleanContents, so use zero and one.
3213 EVT VT = Op.getValueType();
3214 SDValue TVal = DAG.getConstant(1, VT);
3215 SDValue FVal = DAG.getConstant(0, VT);
3217 // Handle f128 first, since one possible outcome is a normal integer
3218 // comparison which gets picked up by the next if statement.
3219 if (LHS.getValueType() == MVT::f128) {
3220 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3222 // If softenSetCCOperands returned a scalar, use it.
3223 if (RHS.getNode() == 0) {
3224 assert(LHS.getValueType() == Op.getValueType() &&
3225 "Unexpected setcc expansion!");
3230 if (LHS.getValueType().isInteger()) {
3233 getARM64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);
3235 // Note that we inverted the condition above, so we reverse the order of
3236 // the true and false operands here. This will allow the setcc to be
3237 // matched to a single CSINC instruction.
3238 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
3241 // Now we know we're dealing with FP values.
3242 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3244 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3245 // and do the comparison.
3246 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3248 ARM64CC::CondCode CC1, CC2;
3249 changeFPCCToARM64CC(CC, CC1, CC2);
3250 if (CC2 == ARM64CC::AL) {
3251 changeFPCCToARM64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
3252 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3254 // Note that we inverted the condition above, so we reverse the order of
3255 // the true and false operands here. This will allow the setcc to be
3256 // matched to a single CSINC instruction.
3257 return DAG.getNode(ARM64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
3259 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3260 // clean. Some of them require two CSELs to implement. As is in this case,
3261 // we emit the first CSEL and then emit a second using the output of the
3262 // first as the RHS. We're effectively OR'ing the two CC's together.
3264 // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
3265 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3266 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3268 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3269 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3273 /// A SELECT_CC operation is really some kind of max or min if both values being
3274 /// compared are, in some sense, equal to the results in either case. However,
3275 /// it is permissible to compare f32 values and produce directly extended f64
3278 /// Extending the comparison operands would also be allowed, but is less likely
3279 /// to happen in practice since their use is right here. Note that truncate
3280 /// operations would *not* be semantically equivalent.
3281 static bool selectCCOpsAreFMaxCompatible(SDValue Cmp, SDValue Result) {
3285 ConstantFPSDNode *CCmp = dyn_cast<ConstantFPSDNode>(Cmp);
3286 ConstantFPSDNode *CResult = dyn_cast<ConstantFPSDNode>(Result);
3287 if (CCmp && CResult && Cmp.getValueType() == MVT::f32 &&
3288 Result.getValueType() == MVT::f64) {
3290 APFloat CmpVal = CCmp->getValueAPF();
3291 CmpVal.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven, &Lossy);
3292 return CResult->getValueAPF().bitwiseIsEqual(CmpVal);
3295 return Result->getOpcode() == ISD::FP_EXTEND && Result->getOperand(0) == Cmp;
3298 SDValue ARM64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3299 SDValue CC = Op->getOperand(0);
3300 SDValue TVal = Op->getOperand(1);
3301 SDValue FVal = Op->getOperand(2);
3304 unsigned Opc = CC.getOpcode();
3305 // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
3307 if (CC.getResNo() == 1 &&
3308 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3309 Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
3310 // Only lower legal XALUO ops.
3311 if (!DAG.getTargetLoweringInfo().isTypeLegal(CC->getValueType(0)))
3314 ARM64CC::CondCode OFCC;
3315 SDValue Value, Overflow;
3316 std::tie(Value, Overflow) = getARM64XALUOOp(OFCC, CC.getValue(0), DAG);
3317 SDValue CCVal = DAG.getConstant(OFCC, MVT::i32);
3319 return DAG.getNode(ARM64ISD::CSEL, DL, Op.getValueType(), TVal, FVal, CCVal,
3323 if (CC.getOpcode() == ISD::SETCC)
3324 return DAG.getSelectCC(DL, CC.getOperand(0), CC.getOperand(1), TVal, FVal,
3325 cast<CondCodeSDNode>(CC.getOperand(2))->get());
3327 return DAG.getSelectCC(DL, CC, DAG.getConstant(0, CC.getValueType()), TVal,
3331 SDValue ARM64TargetLowering::LowerSELECT_CC(SDValue Op,
3332 SelectionDAG &DAG) const {
3333 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3334 SDValue LHS = Op.getOperand(0);
3335 SDValue RHS = Op.getOperand(1);
3336 SDValue TVal = Op.getOperand(2);
3337 SDValue FVal = Op.getOperand(3);
3340 // Handle f128 first, because it will result in a comparison of some RTLIB
3341 // call result against zero.
3342 if (LHS.getValueType() == MVT::f128) {
3343 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3345 // If softenSetCCOperands returned a scalar, we need to compare the result
3346 // against zero to select between true and false values.
3347 if (RHS.getNode() == 0) {
3348 RHS = DAG.getConstant(0, LHS.getValueType());
3353 // Handle integers first.
3354 if (LHS.getValueType().isInteger()) {
3355 assert((LHS.getValueType() == RHS.getValueType()) &&
3356 (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));
3358 unsigned Opcode = ARM64ISD::CSEL;
3360 // If both the TVal and the FVal are constants, see if we can swap them in
3361 // order to for a CSINV or CSINC out of them.
3362 ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
3363 ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
3365 if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
3366 std::swap(TVal, FVal);
3367 std::swap(CTVal, CFVal);
3368 CC = ISD::getSetCCInverse(CC, true);
3369 } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
3370 std::swap(TVal, FVal);
3371 std::swap(CTVal, CFVal);
3372 CC = ISD::getSetCCInverse(CC, true);
3373 } else if (TVal.getOpcode() == ISD::XOR) {
3374 // If TVal is a NOT we want to swap TVal and FVal so that we can match
3375 // with a CSINV rather than a CSEL.
3376 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(1));
3378 if (CVal && CVal->isAllOnesValue()) {
3379 std::swap(TVal, FVal);
3380 std::swap(CTVal, CFVal);
3381 CC = ISD::getSetCCInverse(CC, true);
3383 } else if (TVal.getOpcode() == ISD::SUB) {
3384 // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
3385 // that we can match with a CSNEG rather than a CSEL.
3386 ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(TVal.getOperand(0));
3388 if (CVal && CVal->isNullValue()) {
3389 std::swap(TVal, FVal);
3390 std::swap(CTVal, CFVal);
3391 CC = ISD::getSetCCInverse(CC, true);
3393 } else if (CTVal && CFVal) {
3394 const int64_t TrueVal = CTVal->getSExtValue();
3395 const int64_t FalseVal = CFVal->getSExtValue();
3398 // If both TVal and FVal are constants, see if FVal is the
3399 // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
3400 // instead of a CSEL in that case.
3401 if (TrueVal == ~FalseVal) {
3402 Opcode = ARM64ISD::CSINV;
3403 } else if (TrueVal == -FalseVal) {
3404 Opcode = ARM64ISD::CSNEG;
3405 } else if (TVal.getValueType() == MVT::i32) {
3406 // If our operands are only 32-bit wide, make sure we use 32-bit
3407 // arithmetic for the check whether we can use CSINC. This ensures that
3408 // the addition in the check will wrap around properly in case there is
3409 // an overflow (which would not be the case if we do the check with
3410 // 64-bit arithmetic).
3411 const uint32_t TrueVal32 = CTVal->getZExtValue();
3412 const uint32_t FalseVal32 = CFVal->getZExtValue();
3414 if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
3415 Opcode = ARM64ISD::CSINC;
3417 if (TrueVal32 > FalseVal32) {
3421 // 64-bit check whether we can use CSINC.
3422 } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
3423 Opcode = ARM64ISD::CSINC;
3425 if (TrueVal > FalseVal) {
3430 // Swap TVal and FVal if necessary.
3432 std::swap(TVal, FVal);
3433 std::swap(CTVal, CFVal);
3434 CC = ISD::getSetCCInverse(CC, true);
3437 if (Opcode != ARM64ISD::CSEL) {
3438 // Drop FVal since we can get its value by simply inverting/negating
3445 SDValue Cmp = getARM64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
3447 EVT VT = Op.getValueType();
3448 return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
3451 // Now we know we're dealing with FP values.
3452 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3453 assert(LHS.getValueType() == RHS.getValueType());
3454 EVT VT = Op.getValueType();
3456 // Try to match this select into a max/min operation, which have dedicated
3457 // opcode in the instruction set.
3458 // NOTE: This is not correct in the presence of NaNs, so we only enable this
3460 if (getTargetMachine().Options.NoNaNsFPMath) {
3461 if (selectCCOpsAreFMaxCompatible(LHS, FVal) &&
3462 selectCCOpsAreFMaxCompatible(RHS, TVal)) {
3463 CC = ISD::getSetCCSwappedOperands(CC);
3464 std::swap(TVal, FVal);
3467 if (selectCCOpsAreFMaxCompatible(LHS, TVal) &&
3468 selectCCOpsAreFMaxCompatible(RHS, FVal)) {
3478 return DAG.getNode(ARM64ISD::FMAX, dl, VT, TVal, FVal);
3486 return DAG.getNode(ARM64ISD::FMIN, dl, VT, TVal, FVal);
3492 // If that fails, we'll need to perform an FCMP + CSEL sequence. Go ahead
3493 // and do the comparison.
3494 SDValue Cmp = emitComparison(LHS, RHS, dl, DAG);
3496 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
3497 // clean. Some of them require two CSELs to implement.
3498 ARM64CC::CondCode CC1, CC2;
3499 changeFPCCToARM64CC(CC, CC1, CC2);
3500 SDValue CC1Val = DAG.getConstant(CC1, MVT::i32);
3501 SDValue CS1 = DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);
3503 // If we need a second CSEL, emit it, using the output of the first as the
3504 // RHS. We're effectively OR'ing the two CC's together.
3505 if (CC2 != ARM64CC::AL) {
3506 SDValue CC2Val = DAG.getConstant(CC2, MVT::i32);
3507 return DAG.getNode(ARM64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
3510 // Otherwise, return the output of the first CSEL.
3514 SDValue ARM64TargetLowering::LowerJumpTable(SDValue Op,
3515 SelectionDAG &DAG) const {
3516 // Jump table entries as PC relative offsets. No additional tweaking
3517 // is necessary here. Just get the address of the jump table.
3518 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
3519 EVT PtrVT = getPointerTy();
3522 SDValue Hi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, ARM64II::MO_PAGE);
3523 SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
3524 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3525 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3526 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3529 SDValue ARM64TargetLowering::LowerConstantPool(SDValue Op,
3530 SelectionDAG &DAG) const {
3531 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
3532 EVT PtrVT = getPointerTy();
3535 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
3536 // Use the GOT for the large code model on iOS.
3537 if (Subtarget->isTargetMachO()) {
3538 SDValue GotAddr = DAG.getTargetConstantPool(
3539 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3541 return DAG.getNode(ARM64ISD::LOADgot, DL, PtrVT, GotAddr);
3544 const unsigned char MO_NC = ARM64II::MO_NC;
3546 ARM64ISD::WrapperLarge, DL, PtrVT,
3547 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3548 CP->getOffset(), ARM64II::MO_G3),
3549 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3550 CP->getOffset(), ARM64II::MO_G2 | MO_NC),
3551 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3552 CP->getOffset(), ARM64II::MO_G1 | MO_NC),
3553 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3554 CP->getOffset(), ARM64II::MO_G0 | MO_NC));
3556 // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
3557 // ELF, the only valid one on Darwin.
3559 DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
3560 CP->getOffset(), ARM64II::MO_PAGE);
3561 SDValue Lo = DAG.getTargetConstantPool(
3562 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
3563 ARM64II::MO_PAGEOFF | ARM64II::MO_NC);
3565 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3566 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3570 SDValue ARM64TargetLowering::LowerBlockAddress(SDValue Op,
3571 SelectionDAG &DAG) const {
3572 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
3573 EVT PtrVT = getPointerTy();
3575 if (getTargetMachine().getCodeModel() == CodeModel::Large &&
3576 !Subtarget->isTargetMachO()) {
3577 const unsigned char MO_NC = ARM64II::MO_NC;
3579 ARM64ISD::WrapperLarge, DL, PtrVT,
3580 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G3),
3581 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G2 | MO_NC),
3582 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G1 | MO_NC),
3583 DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_G0 | MO_NC));
3585 SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGE);
3586 SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, ARM64II::MO_PAGEOFF |
3588 SDValue ADRP = DAG.getNode(ARM64ISD::ADRP, DL, PtrVT, Hi);
3589 return DAG.getNode(ARM64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
3593 SDValue ARM64TargetLowering::LowerDarwin_VASTART(SDValue Op,
3594 SelectionDAG &DAG) const {
3595 ARM64FunctionInfo *FuncInfo =
3596 DAG.getMachineFunction().getInfo<ARM64FunctionInfo>();
3600 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3601 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3602 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
3603 MachinePointerInfo(SV), false, false, 0);
3606 SDValue ARM64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
3607 SelectionDAG &DAG) const {
3608 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3609 // Standard, section B.3.
3610 MachineFunction &MF = DAG.getMachineFunction();
3611 ARM64FunctionInfo *FuncInfo = MF.getInfo<ARM64FunctionInfo>();
3614 SDValue Chain = Op.getOperand(0);
3615 SDValue VAList = Op.getOperand(1);
3616 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3617 SmallVector<SDValue, 4> MemOps;
3619 // void *__stack at offset 0
3621 DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), getPointerTy());
3622 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3623 MachinePointerInfo(SV), false, false, 8));
3625 // void *__gr_top at offset 8
3626 int GPRSize = FuncInfo->getVarArgsGPRSize();
3628 SDValue GRTop, GRTopAddr;
3630 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3631 DAG.getConstant(8, getPointerTy()));
3633 GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), getPointerTy());
3634 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3635 DAG.getConstant(GPRSize, getPointerTy()));
3637 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3638 MachinePointerInfo(SV, 8), false, false, 8));
3641 // void *__vr_top at offset 16
3642 int FPRSize = FuncInfo->getVarArgsFPRSize();
3644 SDValue VRTop, VRTopAddr;
3645 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3646 DAG.getConstant(16, getPointerTy()));
3648 VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), getPointerTy());
3649 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3650 DAG.getConstant(FPRSize, getPointerTy()));
3652 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3653 MachinePointerInfo(SV, 16), false, false, 8));
3656 // int __gr_offs at offset 24
3657 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3658 DAG.getConstant(24, getPointerTy()));
3659 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3660 GROffsAddr, MachinePointerInfo(SV, 24), false,
3663 // int __vr_offs at offset 28
3664 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3665 DAG.getConstant(28, getPointerTy()));
3666 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3667 VROffsAddr, MachinePointerInfo(SV, 28), false,
3670 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3674 SDValue ARM64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3675 return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
3676 : LowerAAPCS_VASTART(Op, DAG);
3679 SDValue ARM64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3680 // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
3682 unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
3683 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3684 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3686 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op), Op.getOperand(1),
3687 Op.getOperand(2), DAG.getConstant(VaListSize, MVT::i32),
3688 8, false, false, MachinePointerInfo(DestSV),
3689 MachinePointerInfo(SrcSV));
3692 SDValue ARM64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
3693 assert(Subtarget->isTargetDarwin() &&
3694 "automatic va_arg instruction only works on Darwin");
3696 const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3697 EVT VT = Op.getValueType();
3699 SDValue Chain = Op.getOperand(0);
3700 SDValue Addr = Op.getOperand(1);
3701 unsigned Align = Op.getConstantOperandVal(3);
3703 SDValue VAList = DAG.getLoad(getPointerTy(), DL, Chain, Addr,
3704 MachinePointerInfo(V), false, false, false, 0);
3705 Chain = VAList.getValue(1);
3708 assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
3709 VAList = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3710 DAG.getConstant(Align - 1, getPointerTy()));
3711 VAList = DAG.getNode(ISD::AND, DL, getPointerTy(), VAList,
3712 DAG.getConstant(-(int64_t)Align, getPointerTy()));
3715 Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
3716 uint64_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
3718 // Scalar integer and FP values smaller than 64 bits are implicitly extended
3719 // up to 64 bits. At the very least, we have to increase the striding of the
3720 // vaargs list to match this, and for FP values we need to introduce
3721 // FP_ROUND nodes as well.
3722 if (VT.isInteger() && !VT.isVector())
3724 bool NeedFPTrunc = false;
3725 if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
3730 // Increment the pointer, VAList, to the next vaarg
3731 SDValue VANext = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3732 DAG.getConstant(ArgSize, getPointerTy()));
3733 // Store the incremented VAList to the legalized pointer
3734 SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
3737 // Load the actual argument out of the pointer VAList
3739 // Load the value as an f64.
3740 SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
3741 MachinePointerInfo(), false, false, false, 0);
3742 // Round the value down to an f32.
3743 SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
3744 DAG.getIntPtrConstant(1));
3745 SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
3746 // Merge the rounded value with the chain output of the load.
3747 return DAG.getMergeValues(Ops, 2, DL);
3750 return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
3754 SDValue ARM64TargetLowering::LowerFRAMEADDR(SDValue Op,
3755 SelectionDAG &DAG) const {
3756 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3757 MFI->setFrameAddressIsTaken(true);
3759 EVT VT = Op.getValueType();
3761 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3762 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, ARM64::FP, VT);
3764 FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
3765 MachinePointerInfo(), false, false, false, 0);
3769 SDValue ARM64TargetLowering::LowerRETURNADDR(SDValue Op,
3770 SelectionDAG &DAG) const {
3771 MachineFunction &MF = DAG.getMachineFunction();
3772 MachineFrameInfo *MFI = MF.getFrameInfo();
3773 MFI->setReturnAddressIsTaken(true);
3775 EVT VT = Op.getValueType();
3777 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3779 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3780 SDValue Offset = DAG.getConstant(8, getPointerTy());
3781 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
3782 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
3783 MachinePointerInfo(), false, false, false, 0);
3786 // Return LR, which contains the return address. Mark it an implicit live-in.
3787 unsigned Reg = MF.addLiveIn(ARM64::LR, &ARM64::GPR64RegClass);
3788 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
3791 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3792 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3793 SDValue ARM64TargetLowering::LowerShiftRightParts(SDValue Op,
3794 SelectionDAG &DAG) const {
3795 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3796 EVT VT = Op.getValueType();
3797 unsigned VTBits = VT.getSizeInBits();
3799 SDValue ShOpLo = Op.getOperand(0);
3800 SDValue ShOpHi = Op.getOperand(1);
3801 SDValue ShAmt = Op.getOperand(2);
3803 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3805 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3807 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3808 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3809 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3810 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3811 DAG.getConstant(VTBits, MVT::i64));
3812 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3815 emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64), dl, DAG);
3816 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3818 SDValue FalseValLo = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3819 SDValue TrueValLo = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3821 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3823 // ARM64 shifts larger than the register width are wrapped rather than
3824 // clamped, so we can't just emit "hi >> x".
3825 SDValue FalseValHi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3826 SDValue TrueValHi = Opc == ISD::SRA
3827 ? DAG.getNode(Opc, dl, VT, ShOpHi,
3828 DAG.getConstant(VTBits - 1, MVT::i64))
3829 : DAG.getConstant(0, VT);
3831 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValHi, FalseValHi, CCVal, Cmp);
3833 SDValue Ops[2] = { Lo, Hi };
3834 return DAG.getMergeValues(Ops, 2, dl);
3837 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3838 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
3839 SDValue ARM64TargetLowering::LowerShiftLeftParts(SDValue Op,
3840 SelectionDAG &DAG) const {
3841 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3842 EVT VT = Op.getValueType();
3843 unsigned VTBits = VT.getSizeInBits();
3845 SDValue ShOpLo = Op.getOperand(0);
3846 SDValue ShOpHi = Op.getOperand(1);
3847 SDValue ShAmt = Op.getOperand(2);
3850 assert(Op.getOpcode() == ISD::SHL_PARTS);
3851 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
3852 DAG.getConstant(VTBits, MVT::i64), ShAmt);
3853 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3854 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
3855 DAG.getConstant(VTBits, MVT::i64));
3856 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3857 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3859 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3862 emitComparison(ExtraShAmt, DAG.getConstant(0, MVT::i64), dl, DAG);
3863 SDValue CCVal = DAG.getConstant(ARM64CC::GE, MVT::i32);
3864 SDValue Hi = DAG.getNode(ARM64ISD::CSEL, dl, VT, Tmp3, FalseVal, CCVal, Cmp);
3866 // ARM64 shifts of larger than register sizes are wrapped rather than clamped,
3867 // so we can't just emit "lo << a" if a is too big.
3868 SDValue TrueValLo = DAG.getConstant(0, VT);
3869 SDValue FalseValLo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3871 DAG.getNode(ARM64ISD::CSEL, dl, VT, TrueValLo, FalseValLo, CCVal, Cmp);
3873 SDValue Ops[2] = { Lo, Hi };
3874 return DAG.getMergeValues(Ops, 2, dl);
3878 ARM64TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3879 // The ARM64 target doesn't support folding offsets into global addresses.
3883 bool ARM64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3884 // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
3885 // FIXME: We should be able to handle f128 as well with a clever lowering.
3886 if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
3890 return ARM64_AM::getFP64Imm(Imm) != -1;
3891 else if (VT == MVT::f32)
3892 return ARM64_AM::getFP32Imm(Imm) != -1;
3896 //===----------------------------------------------------------------------===//
3897 // ARM64 Optimization Hooks
3898 //===----------------------------------------------------------------------===//
3900 //===----------------------------------------------------------------------===//
3901 // ARM64 Inline Assembly Support
3902 //===----------------------------------------------------------------------===//
3904 // Table of Constraints
3905 // TODO: This is the current set of constraints supported by ARM for the
3906 // compiler, not all of them may make sense, e.g. S may be difficult to support.
3908 // r - A general register
3909 // w - An FP/SIMD register of some size in the range v0-v31
3910 // x - An FP/SIMD register of some size in the range v0-v15
3911 // I - Constant that can be used with an ADD instruction
3912 // J - Constant that can be used with a SUB instruction
3913 // K - Constant that can be used with a 32-bit logical instruction
3914 // L - Constant that can be used with a 64-bit logical instruction
3915 // M - Constant that can be used as a 32-bit MOV immediate
3916 // N - Constant that can be used as a 64-bit MOV immediate
3917 // Q - A memory reference with base register and no offset
3918 // S - A symbolic address
3919 // Y - Floating point constant zero
3920 // Z - Integer constant zero
3922 // Note that general register operands will be output using their 64-bit x
3923 // register name, whatever the size of the variable, unless the asm operand
3924 // is prefixed by the %w modifier. Floating-point and SIMD register operands
3925 // will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
3928 /// getConstraintType - Given a constraint letter, return the type of
3929 /// constraint it is for this target.
3930 ARM64TargetLowering::ConstraintType
3931 ARM64TargetLowering::getConstraintType(const std::string &Constraint) const {
3932 if (Constraint.size() == 1) {
3933 switch (Constraint[0]) {
3940 return C_RegisterClass;
3941 // An address with a single base register. Due to the way we
3942 // currently handle addresses it is the same as 'r'.
3947 return TargetLowering::getConstraintType(Constraint);
3950 /// Examine constraint type and operand type and determine a weight value.
3951 /// This object must already have been set up with the operand type
3952 /// and the current alternative constraint selected.
3953 TargetLowering::ConstraintWeight
3954 ARM64TargetLowering::getSingleConstraintMatchWeight(
3955 AsmOperandInfo &info, const char *constraint) const {
3956 ConstraintWeight weight = CW_Invalid;
3957 Value *CallOperandVal = info.CallOperandVal;
3958 // If we don't have a value, we can't do a match,
3959 // but allow it at the lowest weight.
3960 if (CallOperandVal == NULL)
3962 Type *type = CallOperandVal->getType();
3963 // Look at the constraint type.
3964 switch (*constraint) {
3966 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3970 if (type->isFloatingPointTy() || type->isVectorTy())
3971 weight = CW_Register;
3974 weight = CW_Constant;
3980 std::pair<unsigned, const TargetRegisterClass *>
3981 ARM64TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3983 if (Constraint.size() == 1) {
3984 switch (Constraint[0]) {
3986 if (VT.getSizeInBits() == 64)
3987 return std::make_pair(0U, &ARM64::GPR64commonRegClass);
3988 return std::make_pair(0U, &ARM64::GPR32commonRegClass);
3991 return std::make_pair(0U, &ARM64::FPR32RegClass);
3992 if (VT.getSizeInBits() == 64)
3993 return std::make_pair(0U, &ARM64::FPR64RegClass);
3994 if (VT.getSizeInBits() == 128)
3995 return std::make_pair(0U, &ARM64::FPR128RegClass);
3997 // The instructions that this constraint is designed for can
3998 // only take 128-bit registers so just use that regclass.
4000 if (VT.getSizeInBits() == 128)
4001 return std::make_pair(0U, &ARM64::FPR128_loRegClass);
4005 if (StringRef("{cc}").equals_lower(Constraint))
4006 return std::make_pair(unsigned(ARM64::CPSR), &ARM64::CCRRegClass);
4008 // Use the default implementation in TargetLowering to convert the register
4009 // constraint into a member of a register class.
4010 std::pair<unsigned, const TargetRegisterClass *> Res;
4011 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4013 // Not found as a standard register?
4014 if (Res.second == 0) {
4015 unsigned Size = Constraint.size();
4016 if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
4017 tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
4018 const std::string Reg =
4019 std::string(&Constraint[2], &Constraint[Size - 1]);
4020 int RegNo = atoi(Reg.c_str());
4021 if (RegNo >= 0 && RegNo <= 31) {
4022 // v0 - v31 are aliases of q0 - q31.
4023 // By default we'll emit v0-v31 for this unless there's a modifier where
4024 // we'll emit the correct register as well.
4025 Res.first = ARM64::FPR128RegClass.getRegister(RegNo);
4026 Res.second = &ARM64::FPR128RegClass;
4034 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4035 /// vector. If it is invalid, don't add anything to Ops.
4036 void ARM64TargetLowering::LowerAsmOperandForConstraint(
4037 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
4038 SelectionDAG &DAG) const {
4039 SDValue Result(0, 0);
4041 // Currently only support length 1 constraints.
4042 if (Constraint.length() != 1)
4045 char ConstraintLetter = Constraint[0];
4046 switch (ConstraintLetter) {
4050 // This set of constraints deal with valid constants for various instructions.
4051 // Validate and return a target constant for them if we can.
4053 // 'z' maps to xzr or wzr so it needs an input of 0.
4054 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4055 if (!C || C->getZExtValue() != 0)
4058 if (Op.getValueType() == MVT::i64)
4059 Result = DAG.getRegister(ARM64::XZR, MVT::i64);
4061 Result = DAG.getRegister(ARM64::WZR, MVT::i32);
4071 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4075 // Grab the value and do some validation.
4076 uint64_t CVal = C->getZExtValue();
4077 switch (ConstraintLetter) {
4078 // The I constraint applies only to simple ADD or SUB immediate operands:
4079 // i.e. 0 to 4095 with optional shift by 12
4080 // The J constraint applies only to ADD or SUB immediates that would be
4081 // valid when negated, i.e. if [an add pattern] were to be output as a SUB
4082 // instruction [or vice versa], in other words -1 to -4095 with optional
4083 // left shift by 12.
4085 if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
4089 uint64_t NVal = -C->getSExtValue();
4090 if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal))
4094 // The K and L constraints apply *only* to logical immediates, including
4095 // what used to be the MOVI alias for ORR (though the MOVI alias has now
4096 // been removed and MOV should be used). So these constraints have to
4097 // distinguish between bit patterns that are valid 32-bit or 64-bit
4098 // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
4099 // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
4102 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4106 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4109 // The M and N constraints are a superset of K and L respectively, for use
4110 // with the MOV (immediate) alias. As well as the logical immediates they
4111 // also match 32 or 64-bit immediates that can be loaded either using a
4112 // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
4113 // (M) or 64-bit 0x1234000000000000 (N) etc.
4114 // As a note some of this code is liberally stolen from the asm parser.
4116 if (!isUInt<32>(CVal))
4118 if (ARM64_AM::isLogicalImmediate(CVal, 32))
4120 if ((CVal & 0xFFFF) == CVal)
4122 if ((CVal & 0xFFFF0000ULL) == CVal)
4124 uint64_t NCVal = ~(uint32_t)CVal;
4125 if ((NCVal & 0xFFFFULL) == NCVal)
4127 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4132 if (ARM64_AM::isLogicalImmediate(CVal, 64))
4134 if ((CVal & 0xFFFFULL) == CVal)
4136 if ((CVal & 0xFFFF0000ULL) == CVal)
4138 if ((CVal & 0xFFFF00000000ULL) == CVal)
4140 if ((CVal & 0xFFFF000000000000ULL) == CVal)
4142 uint64_t NCVal = ~CVal;
4143 if ((NCVal & 0xFFFFULL) == NCVal)
4145 if ((NCVal & 0xFFFF0000ULL) == NCVal)
4147 if ((NCVal & 0xFFFF00000000ULL) == NCVal)
4149 if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
4157 // All assembler immediates are 64-bit integers.
4158 Result = DAG.getTargetConstant(CVal, MVT::i64);
4162 if (Result.getNode()) {
4163 Ops.push_back(Result);
4167 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
4170 //===----------------------------------------------------------------------===//
4171 // ARM64 Advanced SIMD Support
4172 //===----------------------------------------------------------------------===//
4174 /// WidenVector - Given a value in the V64 register class, produce the
4175 /// equivalent value in the V128 register class.
4176 static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
4177 EVT VT = V64Reg.getValueType();
4178 unsigned NarrowSize = VT.getVectorNumElements();
4179 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4180 MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
4183 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
4184 V64Reg, DAG.getConstant(0, MVT::i32));
4187 /// getExtFactor - Determine the adjustment factor for the position when
4188 /// generating an "extract from vector registers" instruction.
4189 static unsigned getExtFactor(SDValue &V) {
4190 EVT EltType = V.getValueType().getVectorElementType();
4191 return EltType.getSizeInBits() / 8;
4194 /// NarrowVector - Given a value in the V128 register class, produce the
4195 /// equivalent value in the V64 register class.
4196 static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
4197 EVT VT = V128Reg.getValueType();
4198 unsigned WideSize = VT.getVectorNumElements();
4199 MVT EltTy = VT.getVectorElementType().getSimpleVT();
4200 MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
4203 return DAG.getTargetExtractSubreg(ARM64::dsub, DL, NarrowTy, V128Reg);
4206 // Gather data to see if the operation can be modelled as a
4207 // shuffle in combination with VEXTs.
4208 SDValue ARM64TargetLowering::ReconstructShuffle(SDValue Op,
4209 SelectionDAG &DAG) const {
4211 EVT VT = Op.getValueType();
4212 unsigned NumElts = VT.getVectorNumElements();
4214 SmallVector<SDValue, 2> SourceVecs;
4215 SmallVector<unsigned, 2> MinElts;
4216 SmallVector<unsigned, 2> MaxElts;
4218 for (unsigned i = 0; i < NumElts; ++i) {
4219 SDValue V = Op.getOperand(i);
4220 if (V.getOpcode() == ISD::UNDEF)
4222 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4223 // A shuffle can only come from building a vector from various
4224 // elements of other vectors.
4228 // Record this extraction against the appropriate vector if possible...
4229 SDValue SourceVec = V.getOperand(0);
4230 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4231 bool FoundSource = false;
4232 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4233 if (SourceVecs[j] == SourceVec) {
4234 if (MinElts[j] > EltNo)
4236 if (MaxElts[j] < EltNo)
4243 // Or record a new source if not...
4245 SourceVecs.push_back(SourceVec);
4246 MinElts.push_back(EltNo);
4247 MaxElts.push_back(EltNo);
4251 // Currently only do something sane when at most two source vectors
4253 if (SourceVecs.size() > 2)
4256 SDValue ShuffleSrcs[2] = { DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4257 int VEXTOffsets[2] = { 0, 0 };
4259 // This loop extracts the usage patterns of the source vectors
4260 // and prepares appropriate SDValues for a shuffle if possible.
4261 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4262 if (SourceVecs[i].getValueType() == VT) {
4263 // No VEXT necessary
4264 ShuffleSrcs[i] = SourceVecs[i];
4267 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4268 // It probably isn't worth padding out a smaller vector just to
4269 // break it down again in a shuffle.
4273 // Don't attempt to extract subvectors from BUILD_VECTOR sources
4274 // that expand or trunc the original value.
4275 // TODO: We can try to bitcast and ANY_EXTEND the result but
4276 // we need to consider the cost of vector ANY_EXTEND, and the
4277 // legality of all the types.
4278 if (SourceVecs[i].getValueType().getVectorElementType() !=
4279 VT.getVectorElementType())
4282 // Since only 64-bit and 128-bit vectors are legal on ARM and
4283 // we've eliminated the other cases...
4284 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2 * NumElts &&
4285 "unexpected vector sizes in ReconstructShuffle");
4287 if (MaxElts[i] - MinElts[i] >= NumElts) {
4288 // Span too large for a VEXT to cope
4292 if (MinElts[i] >= NumElts) {
4293 // The extraction can just take the second half
4294 VEXTOffsets[i] = NumElts;
4296 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4297 DAG.getIntPtrConstant(NumElts));
4298 } else if (MaxElts[i] < NumElts) {
4299 // The extraction can just take the first half
4301 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4302 SourceVecs[i], DAG.getIntPtrConstant(0));
4304 // An actual VEXT is needed
4305 VEXTOffsets[i] = MinElts[i];
4306 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4307 SourceVecs[i], DAG.getIntPtrConstant(0));
4309 DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, SourceVecs[i],
4310 DAG.getIntPtrConstant(NumElts));
4311 unsigned Imm = VEXTOffsets[i] * getExtFactor(VEXTSrc1);
4312 ShuffleSrcs[i] = DAG.getNode(ARM64ISD::EXT, dl, VT, VEXTSrc1, VEXTSrc2,
4313 DAG.getConstant(Imm, MVT::i32));
4317 SmallVector<int, 8> Mask;
4319 for (unsigned i = 0; i < NumElts; ++i) {
4320 SDValue Entry = Op.getOperand(i);
4321 if (Entry.getOpcode() == ISD::UNDEF) {
4326 SDValue ExtractVec = Entry.getOperand(0);
4328 cast<ConstantSDNode>(Op.getOperand(i).getOperand(1))->getSExtValue();
4329 if (ExtractVec == SourceVecs[0]) {
4330 Mask.push_back(ExtractElt - VEXTOffsets[0]);
4332 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4336 // Final check before we try to produce nonsense...
4337 if (isShuffleMaskLegal(Mask, VT))
4338 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4344 // check if an EXT instruction can handle the shuffle mask when the
4345 // vector sources of the shuffle are the same.
4346 static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4347 unsigned NumElts = VT.getVectorNumElements();
4349 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4355 // If this is a VEXT shuffle, the immediate value is the index of the first
4356 // element. The other shuffle indices must be the successive elements after
4358 unsigned ExpectedElt = Imm;
4359 for (unsigned i = 1; i < NumElts; ++i) {
4360 // Increment the expected index. If it wraps around, just follow it
4361 // back to index zero and keep going.
4363 if (ExpectedElt == NumElts)
4367 continue; // ignore UNDEF indices
4368 if (ExpectedElt != static_cast<unsigned>(M[i]))
4375 // check if an EXT instruction can handle the shuffle mask when the
4376 // vector sources of the shuffle are different.
4377 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
4379 unsigned NumElts = VT.getVectorNumElements();
4382 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4388 // If this is a VEXT shuffle, the immediate value is the index of the first
4389 // element. The other shuffle indices must be the successive elements after
4391 unsigned ExpectedElt = Imm;
4392 for (unsigned i = 1; i < NumElts; ++i) {
4393 // Increment the expected index. If it wraps around, it may still be
4394 // a VEXT but the source vectors must be swapped.
4396 if (ExpectedElt == NumElts * 2) {
4402 continue; // ignore UNDEF indices
4403 if (ExpectedElt != static_cast<unsigned>(M[i]))
4407 // Adjust the index value if the source operands will be swapped.
4414 /// isREVMask - Check if a vector shuffle corresponds to a REV
4415 /// instruction with the specified blocksize. (The order of the elements
4416 /// within each block of the vector is reversed.)
4417 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4418 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4419 "Only possible block sizes for REV are: 16, 32, 64");
4421 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4425 unsigned NumElts = VT.getVectorNumElements();
4426 unsigned BlockElts = M[0] + 1;
4427 // If the first shuffle index is UNDEF, be optimistic.
4429 BlockElts = BlockSize / EltSz;
4431 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4434 for (unsigned i = 0; i < NumElts; ++i) {
4436 continue; // ignore UNDEF indices
4437 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4444 static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4445 unsigned NumElts = VT.getVectorNumElements();
4446 WhichResult = (M[0] == 0 ? 0 : 1);
4447 unsigned Idx = WhichResult * NumElts / 2;
4448 for (unsigned i = 0; i != NumElts; i += 2) {
4449 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4450 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
4458 static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4459 unsigned NumElts = VT.getVectorNumElements();
4460 WhichResult = (M[0] == 0 ? 0 : 1);
4461 for (unsigned i = 0; i != NumElts; ++i) {
4463 continue; // ignore UNDEF indices
4464 if ((unsigned)M[i] != 2 * i + WhichResult)
4471 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4472 unsigned NumElts = VT.getVectorNumElements();
4473 WhichResult = (M[0] == 0 ? 0 : 1);
4474 for (unsigned i = 0; i < NumElts; i += 2) {
4475 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4476 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + NumElts + WhichResult))
4482 /// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
4483 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4484 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4485 static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4486 unsigned NumElts = VT.getVectorNumElements();
4487 WhichResult = (M[0] == 0 ? 0 : 1);
4488 unsigned Idx = WhichResult * NumElts / 2;
4489 for (unsigned i = 0; i != NumElts; i += 2) {
4490 if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
4491 (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
4499 /// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
4500 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4501 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4502 static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4503 unsigned Half = VT.getVectorNumElements() / 2;
4504 WhichResult = (M[0] == 0 ? 0 : 1);
4505 for (unsigned j = 0; j != 2; ++j) {
4506 unsigned Idx = WhichResult;
4507 for (unsigned i = 0; i != Half; ++i) {
4508 int MIdx = M[i + j * Half];
4509 if (MIdx >= 0 && (unsigned)MIdx != Idx)
4518 /// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
4519 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4520 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4521 static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4522 unsigned NumElts = VT.getVectorNumElements();
4523 WhichResult = (M[0] == 0 ? 0 : 1);
4524 for (unsigned i = 0; i < NumElts; i += 2) {
4525 if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
4526 (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
4532 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4533 /// the specified operations to build the shuffle.
4534 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4535 SDValue RHS, SelectionDAG &DAG,
4537 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4538 unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
4539 unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);
4542 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4551 OP_VUZPL, // VUZP, left result
4552 OP_VUZPR, // VUZP, right result
4553 OP_VZIPL, // VZIP, left result
4554 OP_VZIPR, // VZIP, right result
4555 OP_VTRNL, // VTRN, left result
4556 OP_VTRNR // VTRN, right result
4559 if (OpNum == OP_COPY) {
4560 if (LHSID == (1 * 9 + 2) * 9 + 3)
4562 assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
4566 SDValue OpLHS, OpRHS;
4567 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4568 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4569 EVT VT = OpLHS.getValueType();
4573 llvm_unreachable("Unknown shuffle opcode!");
4575 // VREV divides the vector in half and swaps within the half.
4576 if (VT.getVectorElementType() == MVT::i32 ||
4577 VT.getVectorElementType() == MVT::f32)
4578 return DAG.getNode(ARM64ISD::REV64, dl, VT, OpLHS);
4579 // vrev <4 x i16> -> REV32
4580 if (VT.getVectorElementType() == MVT::i16)
4581 return DAG.getNode(ARM64ISD::REV32, dl, VT, OpLHS);
4582 // vrev <4 x i8> -> REV16
4583 assert(VT.getVectorElementType() == MVT::i8);
4584 return DAG.getNode(ARM64ISD::REV16, dl, VT, OpLHS);
4589 EVT EltTy = VT.getVectorElementType();
4591 if (EltTy == MVT::i8)
4592 Opcode = ARM64ISD::DUPLANE8;
4593 else if (EltTy == MVT::i16)
4594 Opcode = ARM64ISD::DUPLANE16;
4595 else if (EltTy == MVT::i32 || EltTy == MVT::f32)
4596 Opcode = ARM64ISD::DUPLANE32;
4597 else if (EltTy == MVT::i64 || EltTy == MVT::f64)
4598 Opcode = ARM64ISD::DUPLANE64;
4600 llvm_unreachable("Invalid vector element type?");
4602 if (VT.getSizeInBits() == 64)
4603 OpLHS = WidenVector(OpLHS, DAG);
4604 SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, MVT::i64);
4605 return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
4610 unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
4611 return DAG.getNode(ARM64ISD::EXT, dl, VT, OpLHS, OpRHS,
4612 DAG.getConstant(Imm, MVT::i32));
4615 return DAG.getNode(ARM64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4617 return DAG.getNode(ARM64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4619 return DAG.getNode(ARM64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4621 return DAG.getNode(ARM64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4623 return DAG.getNode(ARM64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4625 return DAG.getNode(ARM64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS, OpRHS);
4629 static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
4630 SelectionDAG &DAG) {
4631 // Check to see if we can use the TBL instruction.
4632 SDValue V1 = Op.getOperand(0);
4633 SDValue V2 = Op.getOperand(1);
4636 EVT EltVT = Op.getValueType().getVectorElementType();
4637 unsigned BytesPerElt = EltVT.getSizeInBits() / 8;
4639 SmallVector<SDValue, 8> TBLMask;
4640 for (int Val : ShuffleMask) {
4641 for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
4642 unsigned Offset = Byte + Val * BytesPerElt;
4643 TBLMask.push_back(DAG.getConstant(Offset, MVT::i32));
4647 MVT IndexVT = MVT::v8i8;
4648 unsigned IndexLen = 8;
4649 if (Op.getValueType().getSizeInBits() == 128) {
4650 IndexVT = MVT::v16i8;
4654 SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
4655 SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);
4658 if (V2.getNode()->getOpcode() == ISD::UNDEF) {
4660 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
4661 Shuffle = DAG.getNode(
4662 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4663 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4664 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4666 if (IndexLen == 8) {
4667 V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
4668 Shuffle = DAG.getNode(
4669 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4670 DAG.getConstant(Intrinsic::arm64_neon_tbl1, MVT::i32), V1Cst,
4671 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4673 // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
4674 // cannot currently represent the register constraints on the input
4676 // Shuffle = DAG.getNode(ARM64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
4677 // DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
4678 // &TBLMask[0], IndexLen));
4679 Shuffle = DAG.getNode(
4680 ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
4681 DAG.getConstant(Intrinsic::arm64_neon_tbl2, MVT::i32), V1Cst, V2Cst,
4682 DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT, &TBLMask[0], IndexLen));
4685 return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
4688 static unsigned getDUPLANEOp(EVT EltType) {
4689 if (EltType == MVT::i8)
4690 return ARM64ISD::DUPLANE8;
4691 if (EltType == MVT::i16)
4692 return ARM64ISD::DUPLANE16;
4693 if (EltType == MVT::i32 || EltType == MVT::f32)
4694 return ARM64ISD::DUPLANE32;
4695 if (EltType == MVT::i64 || EltType == MVT::f64)
4696 return ARM64ISD::DUPLANE64;
4698 llvm_unreachable("Invalid vector element type?");
4701 SDValue ARM64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4702 SelectionDAG &DAG) const {
4704 EVT VT = Op.getValueType();
4706 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4708 // Convert shuffles that are directly supported on NEON to target-specific
4709 // DAG nodes, instead of keeping them as shuffles and matching them again
4710 // during code selection. This is more efficient and avoids the possibility
4711 // of inconsistencies between legalization and selection.
4712 ArrayRef<int> ShuffleMask = SVN->getMask();
4714 SDValue V1 = Op.getOperand(0);
4715 SDValue V2 = Op.getOperand(1);
4717 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
4718 V1.getValueType().getSimpleVT())) {
4719 int Lane = SVN->getSplatIndex();
4720 // If this is undef splat, generate it via "just" vdup, if possible.
4724 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
4725 return DAG.getNode(ARM64ISD::DUP, dl, V1.getValueType(),
4727 // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
4728 // constant. If so, we can just reference the lane's definition directly.
4729 if (V1.getOpcode() == ISD::BUILD_VECTOR &&
4730 !isa<ConstantSDNode>(V1.getOperand(Lane)))
4731 return DAG.getNode(ARM64ISD::DUP, dl, VT, V1.getOperand(Lane));
4733 // Otherwise, duplicate from the lane of the input vector.
4734 unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
4736 // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
4737 // to make a vector of the same size as this SHUFFLE. We can ignore the
4738 // extract entirely, and canonicalise the concat using WidenVector.
4739 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4740 Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4741 V1 = V1.getOperand(0);
4742 } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
4743 unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
4744 Lane -= Idx * VT.getVectorNumElements() / 2;
4745 V1 = WidenVector(V1.getOperand(Idx), DAG);
4746 } else if (VT.getSizeInBits() == 64)
4747 V1 = WidenVector(V1, DAG);
4749 return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, MVT::i64));
4752 if (isREVMask(ShuffleMask, VT, 64))
4753 return DAG.getNode(ARM64ISD::REV64, dl, V1.getValueType(), V1, V2);
4754 if (isREVMask(ShuffleMask, VT, 32))
4755 return DAG.getNode(ARM64ISD::REV32, dl, V1.getValueType(), V1, V2);
4756 if (isREVMask(ShuffleMask, VT, 16))
4757 return DAG.getNode(ARM64ISD::REV16, dl, V1.getValueType(), V1, V2);
4759 bool ReverseEXT = false;
4761 if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
4764 Imm *= getExtFactor(V1);
4765 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V2,
4766 DAG.getConstant(Imm, MVT::i32));
4767 } else if (V2->getOpcode() == ISD::UNDEF &&
4768 isSingletonEXTMask(ShuffleMask, VT, Imm)) {
4769 Imm *= getExtFactor(V1);
4770 return DAG.getNode(ARM64ISD::EXT, dl, V1.getValueType(), V1, V1,
4771 DAG.getConstant(Imm, MVT::i32));
4774 unsigned WhichResult;
4775 if (isZIPMask(ShuffleMask, VT, WhichResult)) {
4776 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4777 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4779 if (isUZPMask(ShuffleMask, VT, WhichResult)) {
4780 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4781 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4783 if (isTRNMask(ShuffleMask, VT, WhichResult)) {
4784 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4785 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
4788 if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4789 unsigned Opc = (WhichResult == 0) ? ARM64ISD::ZIP1 : ARM64ISD::ZIP2;
4790 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4792 if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4793 unsigned Opc = (WhichResult == 0) ? ARM64ISD::UZP1 : ARM64ISD::UZP2;
4794 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4796 if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
4797 unsigned Opc = (WhichResult == 0) ? ARM64ISD::TRN1 : ARM64ISD::TRN2;
4798 return DAG.getNode(Opc, dl, V1.getValueType(), V1, V1);
4801 // If the shuffle is not directly supported and it has 4 elements, use
4802 // the PerfectShuffle-generated table to synthesize it from other shuffles.
4803 unsigned NumElts = VT.getVectorNumElements();
4805 unsigned PFIndexes[4];
4806 for (unsigned i = 0; i != 4; ++i) {
4807 if (ShuffleMask[i] < 0)
4810 PFIndexes[i] = ShuffleMask[i];
4813 // Compute the index in the perfect shuffle table.
4814 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
4815 PFIndexes[2] * 9 + PFIndexes[3];
4816 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4817 unsigned Cost = (PFEntry >> 30);
4820 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4823 return GenerateTBL(Op, ShuffleMask, DAG);
4826 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
4828 EVT VT = BVN->getValueType(0);
4829 APInt SplatBits, SplatUndef;
4830 unsigned SplatBitSize;
4832 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4833 unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;
4835 for (unsigned i = 0; i < NumSplats; ++i) {
4836 CnstBits <<= SplatBitSize;
4837 UndefBits <<= SplatBitSize;
4838 CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
4839 UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
4848 SDValue ARM64TargetLowering::LowerVectorAND(SDValue Op,
4849 SelectionDAG &DAG) const {
4850 BuildVectorSDNode *BVN =
4851 dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
4852 SDValue LHS = Op.getOperand(0);
4854 EVT VT = Op.getValueType();
4859 APInt CnstBits(VT.getSizeInBits(), 0);
4860 APInt UndefBits(VT.getSizeInBits(), 0);
4861 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
4862 // We only have BIC vector immediate instruction, which is and-not.
4863 CnstBits = ~CnstBits;
4865 // We make use of a little bit of goto ickiness in order to avoid having to
4866 // duplicate the immediate matching logic for the undef toggled case.
4867 bool SecondTry = false;
4870 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
4871 CnstBits = CnstBits.zextOrTrunc(64);
4872 uint64_t CnstVal = CnstBits.getZExtValue();
4874 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
4875 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
4876 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4877 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4878 DAG.getConstant(CnstVal, MVT::i32),
4879 DAG.getConstant(0, MVT::i32));
4880 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4883 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
4884 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
4885 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4886 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4887 DAG.getConstant(CnstVal, MVT::i32),
4888 DAG.getConstant(8, MVT::i32));
4889 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4892 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
4893 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
4894 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4895 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4896 DAG.getConstant(CnstVal, MVT::i32),
4897 DAG.getConstant(16, MVT::i32));
4898 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4901 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
4902 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
4903 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
4904 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4905 DAG.getConstant(CnstVal, MVT::i32),
4906 DAG.getConstant(24, MVT::i32));
4907 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4910 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
4911 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
4912 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4913 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4914 DAG.getConstant(CnstVal, MVT::i32),
4915 DAG.getConstant(0, MVT::i32));
4916 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4919 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
4920 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
4921 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
4922 SDValue Mov = DAG.getNode(ARM64ISD::BICi, dl, MovTy, LHS,
4923 DAG.getConstant(CnstVal, MVT::i32),
4924 DAG.getConstant(8, MVT::i32));
4925 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
4932 CnstBits = ~UndefBits;
4936 // We can always fall back to a non-immediate AND.
4941 // Specialized code to quickly find if PotentialBVec is a BuildVector that
4942 // consists of only the same constant int value, returned in reference arg
4944 static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
4945 uint64_t &ConstVal) {
4946 BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
4949 ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
4952 EVT VT = Bvec->getValueType(0);
4953 unsigned NumElts = VT.getVectorNumElements();
4954 for (unsigned i = 1; i < NumElts; ++i)
4955 if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
4957 ConstVal = FirstElt->getZExtValue();
4961 static unsigned getIntrinsicID(const SDNode *N) {
4962 unsigned Opcode = N->getOpcode();
4965 return Intrinsic::not_intrinsic;
4966 case ISD::INTRINSIC_WO_CHAIN: {
4967 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4968 if (IID < Intrinsic::num_intrinsics)
4970 return Intrinsic::not_intrinsic;
4975 // Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
4976 // to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
4977 // BUILD_VECTORs with constant element C1, C2 is a constant, and C1 == ~C2.
4978 // Also, logical shift right -> sri, with the same structure.
4979 static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
4980 EVT VT = N->getValueType(0);
4987 // Is the first op an AND?
4988 const SDValue And = N->getOperand(0);
4989 if (And.getOpcode() != ISD::AND)
4992 // Is the second op an shl or lshr?
4993 SDValue Shift = N->getOperand(1);
4994 // This will have been turned into: ARM64ISD::VSHL vector, #shift
4995 // or ARM64ISD::VLSHR vector, #shift
4996 unsigned ShiftOpc = Shift.getOpcode();
4997 if ((ShiftOpc != ARM64ISD::VSHL && ShiftOpc != ARM64ISD::VLSHR))
4999 bool IsShiftRight = ShiftOpc == ARM64ISD::VLSHR;
5001 // Is the shift amount constant?
5002 ConstantSDNode *C2node = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
5006 // Is the and mask vector all constant?
5008 if (!isAllConstantBuildVector(And.getOperand(1), C1))
5011 // Is C1 == ~C2, taking into account how much one can shift elements of a
5013 uint64_t C2 = C2node->getZExtValue();
5014 unsigned ElemSizeInBits = VT.getVectorElementType().getSizeInBits();
5015 if (C2 > ElemSizeInBits)
5017 unsigned ElemMask = (1 << ElemSizeInBits) - 1;
5018 if ((C1 & ElemMask) != (~C2 & ElemMask))
5021 SDValue X = And.getOperand(0);
5022 SDValue Y = Shift.getOperand(0);
5025 IsShiftRight ? Intrinsic::arm64_neon_vsri : Intrinsic::arm64_neon_vsli;
5027 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5028 DAG.getConstant(Intrin, MVT::i32), X, Y, Shift.getOperand(1));
5030 DEBUG(dbgs() << "arm64-lower: transformed: \n");
5031 DEBUG(N->dump(&DAG));
5032 DEBUG(dbgs() << "into: \n");
5033 DEBUG(ResultSLI->dump(&DAG));
5039 SDValue ARM64TargetLowering::LowerVectorOR(SDValue Op,
5040 SelectionDAG &DAG) const {
5041 // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
5042 if (EnableARM64SlrGeneration) {
5043 SDValue Res = tryLowerToSLI(Op.getNode(), DAG);
5048 BuildVectorSDNode *BVN =
5049 dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
5050 SDValue LHS = Op.getOperand(1);
5052 EVT VT = Op.getValueType();
5054 // OR commutes, so try swapping the operands.
5056 LHS = Op.getOperand(0);
5057 BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
5062 APInt CnstBits(VT.getSizeInBits(), 0);
5063 APInt UndefBits(VT.getSizeInBits(), 0);
5064 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5065 // We make use of a little bit of goto ickiness in order to avoid having to
5066 // duplicate the immediate matching logic for the undef toggled case.
5067 bool SecondTry = false;
5070 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5071 CnstBits = CnstBits.zextOrTrunc(64);
5072 uint64_t CnstVal = CnstBits.getZExtValue();
5074 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5075 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5076 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5077 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5078 DAG.getConstant(CnstVal, MVT::i32),
5079 DAG.getConstant(0, MVT::i32));
5080 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5083 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5084 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5085 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5086 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5087 DAG.getConstant(CnstVal, MVT::i32),
5088 DAG.getConstant(8, MVT::i32));
5089 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5092 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5093 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5094 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5095 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5096 DAG.getConstant(CnstVal, MVT::i32),
5097 DAG.getConstant(16, MVT::i32));
5098 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5101 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5102 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5103 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5104 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5105 DAG.getConstant(CnstVal, MVT::i32),
5106 DAG.getConstant(24, MVT::i32));
5107 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5110 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5111 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5112 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5113 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5114 DAG.getConstant(CnstVal, MVT::i32),
5115 DAG.getConstant(0, MVT::i32));
5116 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5119 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5120 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5121 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5122 SDValue Mov = DAG.getNode(ARM64ISD::ORRi, dl, MovTy, LHS,
5123 DAG.getConstant(CnstVal, MVT::i32),
5124 DAG.getConstant(8, MVT::i32));
5125 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5132 CnstBits = UndefBits;
5136 // We can always fall back to a non-immediate OR.
5141 SDValue ARM64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
5142 SelectionDAG &DAG) const {
5143 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5145 EVT VT = Op.getValueType();
5147 APInt CnstBits(VT.getSizeInBits(), 0);
5148 APInt UndefBits(VT.getSizeInBits(), 0);
5149 if (resolveBuildVector(BVN, CnstBits, UndefBits)) {
5150 // We make use of a little bit of goto ickiness in order to avoid having to
5151 // duplicate the immediate matching logic for the undef toggled case.
5152 bool SecondTry = false;
5155 if (CnstBits.getHiBits(64) == CnstBits.getLoBits(64)) {
5156 CnstBits = CnstBits.zextOrTrunc(64);
5157 uint64_t CnstVal = CnstBits.getZExtValue();
5159 // Certain magic vector constants (used to express things like NOT
5160 // and NEG) are passed through unmodified. This allows codegen patterns
5161 // for these operations to match. Special-purpose patterns will lower
5162 // these immediates to MOVIs if it proves necessary.
5163 if (VT.isInteger() && (CnstVal == 0 || CnstVal == ~0ULL))
5166 // The many faces of MOVI...
5167 if (ARM64_AM::isAdvSIMDModImmType10(CnstVal)) {
5168 CnstVal = ARM64_AM::encodeAdvSIMDModImmType10(CnstVal);
5169 if (VT.getSizeInBits() == 128) {
5170 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::v2i64,
5171 DAG.getConstant(CnstVal, MVT::i32));
5172 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5175 // Support the V64 version via subregister insertion.
5176 SDValue Mov = DAG.getNode(ARM64ISD::MOVIedit, dl, MVT::f64,
5177 DAG.getConstant(CnstVal, MVT::i32));
5178 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5181 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5182 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5183 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5184 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5185 DAG.getConstant(CnstVal, MVT::i32),
5186 DAG.getConstant(0, MVT::i32));
5187 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5190 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5191 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5192 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5193 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5194 DAG.getConstant(CnstVal, MVT::i32),
5195 DAG.getConstant(8, MVT::i32));
5196 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5199 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5200 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5201 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5202 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5203 DAG.getConstant(CnstVal, MVT::i32),
5204 DAG.getConstant(16, MVT::i32));
5205 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5208 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5209 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5210 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5211 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5212 DAG.getConstant(CnstVal, MVT::i32),
5213 DAG.getConstant(24, MVT::i32));
5214 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5217 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5218 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5219 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5220 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5221 DAG.getConstant(CnstVal, MVT::i32),
5222 DAG.getConstant(0, MVT::i32));
5223 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5226 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5227 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5228 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5229 SDValue Mov = DAG.getNode(ARM64ISD::MOVIshift, dl, MovTy,
5230 DAG.getConstant(CnstVal, MVT::i32),
5231 DAG.getConstant(8, MVT::i32));
5232 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5235 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5236 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5237 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5238 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5239 DAG.getConstant(CnstVal, MVT::i32),
5240 DAG.getConstant(264, MVT::i32));
5241 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5244 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5245 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5246 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5247 SDValue Mov = DAG.getNode(ARM64ISD::MOVImsl, dl, MovTy,
5248 DAG.getConstant(CnstVal, MVT::i32),
5249 DAG.getConstant(272, MVT::i32));
5250 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5253 if (ARM64_AM::isAdvSIMDModImmType9(CnstVal)) {
5254 CnstVal = ARM64_AM::encodeAdvSIMDModImmType9(CnstVal);
5255 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;
5256 SDValue Mov = DAG.getNode(ARM64ISD::MOVI, dl, MovTy,
5257 DAG.getConstant(CnstVal, MVT::i32));
5258 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5261 // The few faces of FMOV...
5262 if (ARM64_AM::isAdvSIMDModImmType11(CnstVal)) {
5263 CnstVal = ARM64_AM::encodeAdvSIMDModImmType11(CnstVal);
5264 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4f32 : MVT::v2f32;
5265 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MovTy,
5266 DAG.getConstant(CnstVal, MVT::i32));
5267 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5270 if (ARM64_AM::isAdvSIMDModImmType12(CnstVal) &&
5271 VT.getSizeInBits() == 128) {
5272 CnstVal = ARM64_AM::encodeAdvSIMDModImmType12(CnstVal);
5273 SDValue Mov = DAG.getNode(ARM64ISD::FMOV, dl, MVT::v2f64,
5274 DAG.getConstant(CnstVal, MVT::i32));
5275 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5278 // The many faces of MVNI...
5280 if (ARM64_AM::isAdvSIMDModImmType1(CnstVal)) {
5281 CnstVal = ARM64_AM::encodeAdvSIMDModImmType1(CnstVal);
5282 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5283 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5284 DAG.getConstant(CnstVal, MVT::i32),
5285 DAG.getConstant(0, MVT::i32));
5286 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5289 if (ARM64_AM::isAdvSIMDModImmType2(CnstVal)) {
5290 CnstVal = ARM64_AM::encodeAdvSIMDModImmType2(CnstVal);
5291 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5292 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5293 DAG.getConstant(CnstVal, MVT::i32),
5294 DAG.getConstant(8, MVT::i32));
5295 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5298 if (ARM64_AM::isAdvSIMDModImmType3(CnstVal)) {
5299 CnstVal = ARM64_AM::encodeAdvSIMDModImmType3(CnstVal);
5300 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5301 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5302 DAG.getConstant(CnstVal, MVT::i32),
5303 DAG.getConstant(16, MVT::i32));
5304 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5307 if (ARM64_AM::isAdvSIMDModImmType4(CnstVal)) {
5308 CnstVal = ARM64_AM::encodeAdvSIMDModImmType4(CnstVal);
5309 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5310 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5311 DAG.getConstant(CnstVal, MVT::i32),
5312 DAG.getConstant(24, MVT::i32));
5313 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5316 if (ARM64_AM::isAdvSIMDModImmType5(CnstVal)) {
5317 CnstVal = ARM64_AM::encodeAdvSIMDModImmType5(CnstVal);
5318 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5319 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5320 DAG.getConstant(CnstVal, MVT::i32),
5321 DAG.getConstant(0, MVT::i32));
5322 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5325 if (ARM64_AM::isAdvSIMDModImmType6(CnstVal)) {
5326 CnstVal = ARM64_AM::encodeAdvSIMDModImmType6(CnstVal);
5327 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
5328 SDValue Mov = DAG.getNode(ARM64ISD::MVNIshift, dl, MovTy,
5329 DAG.getConstant(CnstVal, MVT::i32),
5330 DAG.getConstant(8, MVT::i32));
5331 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5334 if (ARM64_AM::isAdvSIMDModImmType7(CnstVal)) {
5335 CnstVal = ARM64_AM::encodeAdvSIMDModImmType7(CnstVal);
5336 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5337 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5338 DAG.getConstant(CnstVal, MVT::i32),
5339 DAG.getConstant(264, MVT::i32));
5340 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5343 if (ARM64_AM::isAdvSIMDModImmType8(CnstVal)) {
5344 CnstVal = ARM64_AM::encodeAdvSIMDModImmType8(CnstVal);
5345 MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
5346 SDValue Mov = DAG.getNode(ARM64ISD::MVNImsl, dl, MovTy,
5347 DAG.getConstant(CnstVal, MVT::i32),
5348 DAG.getConstant(272, MVT::i32));
5349 return DAG.getNode(ISD::BITCAST, dl, VT, Mov);
5356 CnstBits = UndefBits;
5361 // Scan through the operands to find some interesting properties we can
5363 // 1) If only one value is used, we can use a DUP, or
5364 // 2) if only the low element is not undef, we can just insert that, or
5365 // 3) if only one constant value is used (w/ some non-constant lanes),
5366 // we can splat the constant value into the whole vector then fill
5367 // in the non-constant lanes.
5368 // 4) FIXME: If different constant values are used, but we can intelligently
5369 // select the values we'll be overwriting for the non-constant
5370 // lanes such that we can directly materialize the vector
5371 // some other way (MOVI, e.g.), we can be sneaky.
5372 unsigned NumElts = VT.getVectorNumElements();
5373 bool isOnlyLowElement = true;
5374 bool usesOnlyOneValue = true;
5375 bool usesOnlyOneConstantValue = true;
5376 bool isConstant = true;
5377 unsigned NumConstantLanes = 0;
5379 SDValue ConstantValue;
5380 for (unsigned i = 0; i < NumElts; ++i) {
5381 SDValue V = Op.getOperand(i);
5382 if (V.getOpcode() == ISD::UNDEF)
5385 isOnlyLowElement = false;
5386 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5389 if (isa<ConstantSDNode>(V)) {
5391 if (!ConstantValue.getNode())
5393 else if (ConstantValue != V)
5394 usesOnlyOneConstantValue = false;
5397 if (!Value.getNode())
5399 else if (V != Value)
5400 usesOnlyOneValue = false;
5403 if (!Value.getNode())
5404 return DAG.getUNDEF(VT);
5406 if (isOnlyLowElement)
5407 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5409 // Use DUP for non-constant splats. For f32 constant splats, reduce to
5410 // i32 and try again.
5411 if (usesOnlyOneValue) {
5413 if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5414 Value.getValueType() != VT)
5415 return DAG.getNode(ARM64ISD::DUP, dl, VT, Value);
5417 // This is actually a DUPLANExx operation, which keeps everything vectory.
5419 // DUPLANE works on 128-bit vectors, widen it if necessary.
5420 SDValue Lane = Value.getOperand(1);
5421 Value = Value.getOperand(0);
5422 if (Value.getValueType().getSizeInBits() == 64)
5423 Value = WidenVector(Value, DAG);
5425 unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
5426 return DAG.getNode(Opcode, dl, VT, Value, Lane);
5429 if (VT.getVectorElementType().isFloatingPoint()) {
5430 SmallVector<SDValue, 8> Ops;
5432 (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
5433 for (unsigned i = 0; i < NumElts; ++i)
5434 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, NewType, Op.getOperand(i)));
5435 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
5436 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
5437 Val = LowerBUILD_VECTOR(Val, DAG);
5439 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5443 // If there was only one constant value used and for more than one lane,
5444 // start by splatting that value, then replace the non-constant lanes. This
5445 // is better than the default, which will perform a separate initialization
5447 if (NumConstantLanes > 0 && usesOnlyOneConstantValue) {
5448 SDValue Val = DAG.getNode(ARM64ISD::DUP, dl, VT, ConstantValue);
5449 // Now insert the non-constant lanes.
5450 for (unsigned i = 0; i < NumElts; ++i) {
5451 SDValue V = Op.getOperand(i);
5452 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5453 if (!isa<ConstantSDNode>(V)) {
5454 // Note that type legalization likely mucked about with the VT of the
5455 // source operand, so we may have to convert it here before inserting.
5456 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, V, LaneIdx);
5462 // If all elements are constants and the case above didn't get hit, fall back
5463 // to the default expansion, which will generate a load from the constant
5468 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5470 SDValue shuffle = ReconstructShuffle(Op, DAG);
5471 if (shuffle != SDValue())
5475 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5476 // know the default expansion would otherwise fall back on something even
5477 // worse. For a vector with one or two non-undef values, that's
5478 // scalar_to_vector for the elements followed by a shuffle (provided the
5479 // shuffle is valid for the target) and materialization element by element
5480 // on the stack followed by a load for everything else.
5481 if (!isConstant && !usesOnlyOneValue) {
5482 SDValue Vec = DAG.getUNDEF(VT);
5483 SDValue Op0 = Op.getOperand(0);
5484 unsigned ElemSize = VT.getVectorElementType().getSizeInBits();
5486 // For 32 and 64 bit types, use INSERT_SUBREG for lane zero to
5487 // a) Avoid a RMW dependency on the full vector register, and
5488 // b) Allow the register coalescer to fold away the copy if the
5489 // value is already in an S or D register.
5490 if (Op0.getOpcode() != ISD::UNDEF && (ElemSize == 32 || ElemSize == 64)) {
5491 unsigned SubIdx = ElemSize == 32 ? ARM64::ssub : ARM64::dsub;
5493 DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl, VT, Vec, Op0,
5494 DAG.getTargetConstant(SubIdx, MVT::i32));
5495 Vec = SDValue(N, 0);
5498 for (; i < NumElts; ++i) {
5499 SDValue V = Op.getOperand(i);
5500 if (V.getOpcode() == ISD::UNDEF)
5502 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
5503 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5508 // Just use the default expansion. We failed to find a better alternative.
5512 SDValue ARM64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
5513 SelectionDAG &DAG) const {
5514 assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");
5516 // Check for non-constant lane.
5517 if (!isa<ConstantSDNode>(Op.getOperand(2)))
5520 EVT VT = Op.getOperand(0).getValueType();
5522 // Insertion/extraction are legal for V128 types.
5523 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5524 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5527 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5528 VT != MVT::v1i64 && VT != MVT::v2f32)
5531 // For V64 types, we perform insertion by expanding the value
5532 // to a V128 type and perform the insertion on that.
5534 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5535 EVT WideTy = WideVec.getValueType();
5537 SDValue Node = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, WideTy, WideVec,
5538 Op.getOperand(1), Op.getOperand(2));
5539 // Re-narrow the resultant vector.
5540 return NarrowVector(Node, DAG);
5543 SDValue ARM64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5544 SelectionDAG &DAG) const {
5545 assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
5547 // Check for non-constant lane.
5548 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5551 EVT VT = Op.getOperand(0).getValueType();
5553 // Insertion/extraction are legal for V128 types.
5554 if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
5555 VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64)
5558 if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
5559 VT != MVT::v1i64 && VT != MVT::v2f32)
5562 // For V64 types, we perform extraction by expanding the value
5563 // to a V128 type and perform the extraction on that.
5565 SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
5566 EVT WideTy = WideVec.getValueType();
5568 EVT ExtrTy = WideTy.getVectorElementType();
5569 if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
5572 // For extractions, we just return the result directly.
5573 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
5577 SDValue ARM64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
5578 SelectionDAG &DAG) const {
5579 EVT VT = Op.getOperand(0).getValueType();
5585 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5588 unsigned Val = Cst->getZExtValue();
5590 unsigned Size = Op.getValueType().getSizeInBits();
5594 return DAG.getTargetExtractSubreg(ARM64::bsub, dl, Op.getValueType(),
5597 return DAG.getTargetExtractSubreg(ARM64::hsub, dl, Op.getValueType(),
5600 return DAG.getTargetExtractSubreg(ARM64::ssub, dl, Op.getValueType(),
5603 return DAG.getTargetExtractSubreg(ARM64::dsub, dl, Op.getValueType(),
5606 llvm_unreachable("Unexpected vector type in extract_subvector!");
5609 // If this is extracting the upper 64-bits of a 128-bit vector, we match
5611 if (Size == 64 && Val * VT.getVectorElementType().getSizeInBits() == 64)
5617 bool ARM64TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5619 if (VT.getVectorNumElements() == 4 &&
5620 (VT.is128BitVector() || VT.is64BitVector())) {
5621 unsigned PFIndexes[4];
5622 for (unsigned i = 0; i != 4; ++i) {
5626 PFIndexes[i] = M[i];
5629 // Compute the index in the perfect shuffle table.
5630 unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
5631 PFIndexes[2] * 9 + PFIndexes[3];
5632 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5633 unsigned Cost = (PFEntry >> 30);
5640 unsigned Imm, WhichResult;
5642 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || isREVMask(M, VT, 64) ||
5643 isREVMask(M, VT, 32) || isREVMask(M, VT, 16) ||
5644 isEXTMask(M, VT, ReverseVEXT, Imm) ||
5645 // isTBLMask(M, VT) || // FIXME: Port TBL support from ARM.
5646 isTRNMask(M, VT, WhichResult) || isUZPMask(M, VT, WhichResult) ||
5647 isZIPMask(M, VT, WhichResult) ||
5648 isTRN_v_undef_Mask(M, VT, WhichResult) ||
5649 isUZP_v_undef_Mask(M, VT, WhichResult) ||
5650 isZIP_v_undef_Mask(M, VT, WhichResult));
5653 /// getVShiftImm - Check if this is a valid build_vector for the immediate
5654 /// operand of a vector shift operation, where all the elements of the
5655 /// build_vector must have the same constant integer value.
5656 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
5657 // Ignore bit_converts.
5658 while (Op.getOpcode() == ISD::BITCAST)
5659 Op = Op.getOperand(0);
5660 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
5661 APInt SplatBits, SplatUndef;
5662 unsigned SplatBitSize;
5664 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
5665 HasAnyUndefs, ElementBits) ||
5666 SplatBitSize > ElementBits)
5668 Cnt = SplatBits.getSExtValue();
5672 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
5673 /// operand of a vector shift left operation. That value must be in the range:
5674 /// 0 <= Value < ElementBits for a left shift; or
5675 /// 0 <= Value <= ElementBits for a long left shift.
5676 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
5677 assert(VT.isVector() && "vector shift count is not a vector type");
5678 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5679 if (!getVShiftImm(Op, ElementBits, Cnt))
5681 return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
5684 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
5685 /// operand of a vector shift right operation. For a shift opcode, the value
5686 /// is positive, but for an intrinsic the value count must be negative. The
5687 /// absolute value must be in the range:
5688 /// 1 <= |Value| <= ElementBits for a right shift; or
5689 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
5690 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
5692 assert(VT.isVector() && "vector shift count is not a vector type");
5693 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
5694 if (!getVShiftImm(Op, ElementBits, Cnt))
5698 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
5701 SDValue ARM64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
5702 SelectionDAG &DAG) const {
5703 EVT VT = Op.getValueType();
5707 if (!Op.getOperand(1).getValueType().isVector())
5709 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5711 switch (Op.getOpcode()) {
5713 llvm_unreachable("unexpected shift opcode");
5716 if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
5717 return DAG.getNode(ARM64ISD::VSHL, SDLoc(Op), VT, Op.getOperand(0),
5718 DAG.getConstant(Cnt, MVT::i32));
5719 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5720 DAG.getConstant(Intrinsic::arm64_neon_ushl, MVT::i32),
5721 Op.getOperand(0), Op.getOperand(1));
5724 // Right shift immediate
5725 if (isVShiftRImm(Op.getOperand(1), VT, false, false, Cnt) &&
5728 (Op.getOpcode() == ISD::SRA) ? ARM64ISD::VASHR : ARM64ISD::VLSHR;
5729 return DAG.getNode(Opc, SDLoc(Op), VT, Op.getOperand(0),
5730 DAG.getConstant(Cnt, MVT::i32));
5733 // Right shift register. Note, there is not a shift right register
5734 // instruction, but the shift left register instruction takes a signed
5735 // value, where negative numbers specify a right shift.
5736 unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::arm64_neon_sshl
5737 : Intrinsic::arm64_neon_ushl;
5738 // negate the shift amount
5739 SDValue NegShift = DAG.getNode(ARM64ISD::NEG, DL, VT, Op.getOperand(1));
5740 SDValue NegShiftLeft =
5741 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
5742 DAG.getConstant(Opc, MVT::i32), Op.getOperand(0), NegShift);
5743 return NegShiftLeft;
5749 static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
5750 ARM64CC::CondCode CC, bool NoNans, EVT VT,
5751 SDLoc dl, SelectionDAG &DAG) {
5752 EVT SrcVT = LHS.getValueType();
5754 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
5755 APInt CnstBits(VT.getSizeInBits(), 0);
5756 APInt UndefBits(VT.getSizeInBits(), 0);
5757 bool IsCnst = BVN && resolveBuildVector(BVN, CnstBits, UndefBits);
5758 bool IsZero = IsCnst && (CnstBits == 0);
5760 if (SrcVT.getVectorElementType().isFloatingPoint()) {
5767 Fcmeq = DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5769 Fcmeq = DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5770 return DAG.getNode(ARM64ISD::NOT, dl, VT, Fcmeq);
5774 return DAG.getNode(ARM64ISD::FCMEQz, dl, VT, LHS);
5775 return DAG.getNode(ARM64ISD::FCMEQ, dl, VT, LHS, RHS);
5778 return DAG.getNode(ARM64ISD::FCMGEz, dl, VT, LHS);
5779 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, LHS, RHS);
5782 return DAG.getNode(ARM64ISD::FCMGTz, dl, VT, LHS);
5783 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, LHS, RHS);
5786 return DAG.getNode(ARM64ISD::FCMLEz, dl, VT, LHS);
5787 return DAG.getNode(ARM64ISD::FCMGE, dl, VT, RHS, LHS);
5791 // If we ignore NaNs then we can use to the MI implementation.
5795 return DAG.getNode(ARM64ISD::FCMLTz, dl, VT, LHS);
5796 return DAG.getNode(ARM64ISD::FCMGT, dl, VT, RHS, LHS);
5806 Cmeq = DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5808 Cmeq = DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5809 return DAG.getNode(ARM64ISD::NOT, dl, VT, Cmeq);
5813 return DAG.getNode(ARM64ISD::CMEQz, dl, VT, LHS);
5814 return DAG.getNode(ARM64ISD::CMEQ, dl, VT, LHS, RHS);
5817 return DAG.getNode(ARM64ISD::CMGEz, dl, VT, LHS);
5818 return DAG.getNode(ARM64ISD::CMGE, dl, VT, LHS, RHS);
5821 return DAG.getNode(ARM64ISD::CMGTz, dl, VT, LHS);
5822 return DAG.getNode(ARM64ISD::CMGT, dl, VT, LHS, RHS);
5825 return DAG.getNode(ARM64ISD::CMLEz, dl, VT, LHS);
5826 return DAG.getNode(ARM64ISD::CMGE, dl, VT, RHS, LHS);
5828 return DAG.getNode(ARM64ISD::CMHS, dl, VT, RHS, LHS);
5830 return DAG.getNode(ARM64ISD::CMHI, dl, VT, RHS, LHS);
5833 return DAG.getNode(ARM64ISD::CMLTz, dl, VT, LHS);
5834 return DAG.getNode(ARM64ISD::CMGT, dl, VT, RHS, LHS);
5836 return DAG.getNode(ARM64ISD::CMHI, dl, VT, LHS, RHS);
5838 return DAG.getNode(ARM64ISD::CMHS, dl, VT, LHS, RHS);
5842 SDValue ARM64TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
5843 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5844 SDValue LHS = Op.getOperand(0);
5845 SDValue RHS = Op.getOperand(1);
5848 if (LHS.getValueType().getVectorElementType().isInteger()) {
5849 assert(LHS.getValueType() == RHS.getValueType());
5850 ARM64CC::CondCode ARM64CC = changeIntCCToARM64CC(CC);
5851 return EmitVectorComparison(LHS, RHS, ARM64CC, false, Op.getValueType(), dl,
5855 assert(LHS.getValueType().getVectorElementType() == MVT::f32 ||
5856 LHS.getValueType().getVectorElementType() == MVT::f64);
5858 // Unfortunately, the mapping of LLVM FP CC's onto ARM64 CC's isn't totally
5859 // clean. Some of them require two branches to implement.
5860 ARM64CC::CondCode CC1, CC2;
5861 changeFPCCToARM64CC(CC, CC1, CC2);
5863 bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
5865 EmitVectorComparison(LHS, RHS, CC1, NoNaNs, Op.getValueType(), dl, DAG);
5866 if (!Cmp1.getNode())
5869 if (CC2 != ARM64CC::AL) {
5871 EmitVectorComparison(LHS, RHS, CC2, NoNaNs, Op.getValueType(), dl, DAG);
5872 if (!Cmp2.getNode())
5875 return DAG.getNode(ISD::OR, dl, Cmp1.getValueType(), Cmp1, Cmp2);
5881 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
5882 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
5883 /// specified in the intrinsic calls.
5884 bool ARM64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5886 unsigned Intrinsic) const {
5887 switch (Intrinsic) {
5888 case Intrinsic::arm64_neon_ld2:
5889 case Intrinsic::arm64_neon_ld3:
5890 case Intrinsic::arm64_neon_ld4:
5891 case Intrinsic::arm64_neon_ld2lane:
5892 case Intrinsic::arm64_neon_ld3lane:
5893 case Intrinsic::arm64_neon_ld4lane:
5894 case Intrinsic::arm64_neon_ld2r:
5895 case Intrinsic::arm64_neon_ld3r:
5896 case Intrinsic::arm64_neon_ld4r: {
5897 Info.opc = ISD::INTRINSIC_W_CHAIN;
5898 // Conservatively set memVT to the entire set of vectors loaded.
5899 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5900 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5901 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5904 Info.vol = false; // volatile loads with NEON intrinsics not supported
5905 Info.readMem = true;
5906 Info.writeMem = false;
5909 case Intrinsic::arm64_neon_st2:
5910 case Intrinsic::arm64_neon_st3:
5911 case Intrinsic::arm64_neon_st4:
5912 case Intrinsic::arm64_neon_st2lane:
5913 case Intrinsic::arm64_neon_st3lane:
5914 case Intrinsic::arm64_neon_st4lane: {
5915 Info.opc = ISD::INTRINSIC_VOID;
5916 // Conservatively set memVT to the entire set of vectors stored.
5917 unsigned NumElts = 0;
5918 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5919 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5920 if (!ArgTy->isVectorTy())
5922 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5924 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5925 Info.ptrVal = I.getArgOperand(I.getNumArgOperands() - 1);
5928 Info.vol = false; // volatile stores with NEON intrinsics not supported
5929 Info.readMem = false;
5930 Info.writeMem = true;
5933 case Intrinsic::arm64_ldxr: {
5934 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
5935 Info.opc = ISD::INTRINSIC_W_CHAIN;
5936 Info.memVT = MVT::getVT(PtrTy->getElementType());
5937 Info.ptrVal = I.getArgOperand(0);
5939 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5941 Info.readMem = true;
5942 Info.writeMem = false;
5945 case Intrinsic::arm64_stxr: {
5946 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
5947 Info.opc = ISD::INTRINSIC_W_CHAIN;
5948 Info.memVT = MVT::getVT(PtrTy->getElementType());
5949 Info.ptrVal = I.getArgOperand(1);
5951 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
5953 Info.readMem = false;
5954 Info.writeMem = true;
5957 case Intrinsic::arm64_ldxp: {
5958 Info.opc = ISD::INTRINSIC_W_CHAIN;
5959 Info.memVT = MVT::i128;
5960 Info.ptrVal = I.getArgOperand(0);
5964 Info.readMem = true;
5965 Info.writeMem = false;
5968 case Intrinsic::arm64_stxp: {
5969 Info.opc = ISD::INTRINSIC_W_CHAIN;
5970 Info.memVT = MVT::i128;
5971 Info.ptrVal = I.getArgOperand(2);
5975 Info.readMem = false;
5976 Info.writeMem = true;
5986 // Truncations from 64-bit GPR to 32-bit GPR is free.
5987 bool ARM64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
5988 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5990 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5991 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5992 if (NumBits1 <= NumBits2)
5996 bool ARM64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
5997 if (!VT1.isInteger() || !VT2.isInteger())
5999 unsigned NumBits1 = VT1.getSizeInBits();
6000 unsigned NumBits2 = VT2.getSizeInBits();
6001 if (NumBits1 <= NumBits2)
6006 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
6008 bool ARM64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
6009 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
6011 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
6012 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
6013 if (NumBits1 == 32 && NumBits2 == 64)
6017 bool ARM64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
6018 if (!VT1.isInteger() || !VT2.isInteger())
6020 unsigned NumBits1 = VT1.getSizeInBits();
6021 unsigned NumBits2 = VT2.getSizeInBits();
6022 if (NumBits1 == 32 && NumBits2 == 64)
6027 bool ARM64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
6028 EVT VT1 = Val.getValueType();
6029 if (isZExtFree(VT1, VT2)) {
6033 if (Val.getOpcode() != ISD::LOAD)
6036 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
6037 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
6038 VT2.isInteger() && VT1.getSizeInBits() <= 32);
6041 bool ARM64TargetLowering::hasPairedLoad(Type *LoadedType,
6042 unsigned &RequiredAligment) const {
6043 if (!LoadedType->isIntegerTy() && !LoadedType->isFloatTy())
6045 // Cyclone supports unaligned accesses.
6046 RequiredAligment = 0;
6047 unsigned NumBits = LoadedType->getPrimitiveSizeInBits();
6048 return NumBits == 32 || NumBits == 64;
6051 bool ARM64TargetLowering::hasPairedLoad(EVT LoadedType,
6052 unsigned &RequiredAligment) const {
6053 if (!LoadedType.isSimple() ||
6054 (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
6056 // Cyclone supports unaligned accesses.
6057 RequiredAligment = 0;
6058 unsigned NumBits = LoadedType.getSizeInBits();
6059 return NumBits == 32 || NumBits == 64;
6062 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
6063 unsigned AlignCheck) {
6064 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
6065 (DstAlign == 0 || DstAlign % AlignCheck == 0));
6068 EVT ARM64TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
6069 unsigned SrcAlign, bool IsMemset,
6070 bool ZeroMemset, bool MemcpyStrSrc,
6071 MachineFunction &MF) const {
6072 // Don't use AdvSIMD to implement 16-byte memset. It would have taken one
6073 // instruction to materialize the v2i64 zero and one store (with restrictive
6074 // addressing mode). Just do two i64 store of zero-registers.
6076 const Function *F = MF.getFunction();
6077 if (!IsMemset && Size >= 16 &&
6078 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
6079 Attribute::NoImplicitFloat) &&
6080 (memOpAlign(SrcAlign, DstAlign, 16) ||
6081 (allowsUnalignedMemoryAccesses(MVT::v2i64, 0, &Fast) && Fast)))
6084 return Size >= 8 ? MVT::i64 : MVT::i32;
6087 // 12-bit optionally shifted immediates are legal for adds.
6088 bool ARM64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
6089 if ((Immed >> 12) == 0 || ((Immed & 0xfff) == 0 && Immed >> 24 == 0))
6094 // Integer comparisons are implemented with ADDS/SUBS, so the range of valid
6095 // immediates is the same as for an add or a sub.
6096 bool ARM64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
6099 return isLegalAddImmediate(Immed);
6102 /// isLegalAddressingMode - Return true if the addressing mode represented
6103 /// by AM is legal for this target, for a load/store of the specified type.
6104 bool ARM64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
6106 // ARM64 has five basic addressing modes:
6108 // reg + 9-bit signed offset
6109 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
6111 // reg + SIZE_IN_BYTES * reg
6113 // No global is ever allowed as a base.
6117 // No reg+reg+imm addressing.
6118 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
6121 // check reg + imm case:
6122 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
6123 uint64_t NumBytes = 0;
6124 if (Ty->isSized()) {
6125 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
6126 NumBytes = NumBits / 8;
6127 if (!isPowerOf2_64(NumBits))
6132 int64_t Offset = AM.BaseOffs;
6134 // 9-bit signed offset
6135 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
6138 // 12-bit unsigned offset
6139 unsigned shift = Log2_64(NumBytes);
6140 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
6141 // Must be a multiple of NumBytes (NumBytes is a power of 2)
6142 (Offset >> shift) << shift == Offset)
6147 // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2
6149 if (!AM.Scale || AM.Scale == 1 ||
6150 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
6155 int ARM64TargetLowering::getScalingFactorCost(const AddrMode &AM,
6157 // Scaling factors are not free at all.
6158 // Operands | Rt Latency
6159 // -------------------------------------------
6161 // -------------------------------------------
6162 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
6163 // Rt, [Xn, Wm, <extend> #imm] |
6164 if (isLegalAddressingMode(AM, Ty))
6165 // Scale represents reg2 * scale, thus account for 1 if
6166 // it is not equal to 0 or 1.
6167 return AM.Scale != 0 && AM.Scale != 1;
6171 bool ARM64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
6172 VT = VT.getScalarType();
6177 switch (VT.getSimpleVT().SimpleTy) {
6189 ARM64TargetLowering::getScratchRegisters(CallingConv::ID) const {
6190 // LR is a callee-save register, but we must treat it as clobbered by any call
6191 // site. Hence we include LR in the scratch registers, which are in turn added
6192 // as implicit-defs for stackmaps and patchpoints.
6193 static const MCPhysReg ScratchRegs[] = {
6194 ARM64::X16, ARM64::X17, ARM64::LR, 0
6199 bool ARM64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
6201 assert(Ty->isIntegerTy());
6203 unsigned BitSize = Ty->getPrimitiveSizeInBits();
6207 int64_t Val = Imm.getSExtValue();
6208 if (Val == 0 || ARM64_AM::isLogicalImmediate(Val, BitSize))
6211 if ((int64_t)Val < 0)
6214 Val &= (1LL << 32) - 1;
6216 unsigned LZ = countLeadingZeros((uint64_t)Val);
6217 unsigned Shift = (63 - LZ) / 16;
6218 // MOVZ is free so return true for one or fewer MOVK.
6219 return (Shift < 3) ? true : false;
6222 // Generate SUBS and CSEL for integer abs.
6223 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
6224 EVT VT = N->getValueType(0);
6226 SDValue N0 = N->getOperand(0);
6227 SDValue N1 = N->getOperand(1);
6230 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
6231 // and change it to SUB and CSEL.
6232 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
6233 N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1 &&
6234 N1.getOpcode() == ISD::SRA && N1.getOperand(0) == N0.getOperand(0))
6235 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
6236 if (Y1C->getAPIntValue() == VT.getSizeInBits() - 1) {
6237 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
6239 // Generate SUBS & CSEL.
6241 DAG.getNode(ARM64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
6242 N0.getOperand(0), DAG.getConstant(0, VT));
6243 return DAG.getNode(ARM64ISD::CSEL, DL, VT, N0.getOperand(0), Neg,
6244 DAG.getConstant(ARM64CC::PL, MVT::i32),
6245 SDValue(Cmp.getNode(), 1));
6250 // performXorCombine - Attempts to handle integer ABS.
6251 static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
6252 TargetLowering::DAGCombinerInfo &DCI,
6253 const ARM64Subtarget *Subtarget) {
6254 if (DCI.isBeforeLegalizeOps())
6257 return performIntegerAbsCombine(N, DAG);
6260 static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
6261 TargetLowering::DAGCombinerInfo &DCI,
6262 const ARM64Subtarget *Subtarget) {
6263 if (DCI.isBeforeLegalizeOps())
6266 // Multiplication of a power of two plus/minus one can be done more
6267 // cheaply as as shift+add/sub. For now, this is true unilaterally. If
6268 // future CPUs have a cheaper MADD instruction, this may need to be
6269 // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
6270 // 64-bit is 5 cycles, so this is always a win.
6271 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
6272 APInt Value = C->getAPIntValue();
6273 EVT VT = N->getValueType(0);
6274 APInt VP1 = Value + 1;
6275 if (VP1.isPowerOf2()) {
6276 // Multiplying by one less than a power of two, replace with a shift
6278 SDValue ShiftedVal =
6279 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6280 DAG.getConstant(VP1.logBase2(), MVT::i64));
6281 return DAG.getNode(ISD::SUB, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6283 APInt VM1 = Value - 1;
6284 if (VM1.isPowerOf2()) {
6285 // Multiplying by one more than a power of two, replace with a shift
6287 SDValue ShiftedVal =
6288 DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0),
6289 DAG.getConstant(VM1.logBase2(), MVT::i64));
6290 return DAG.getNode(ISD::ADD, SDLoc(N), VT, ShiftedVal, N->getOperand(0));
6296 static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG) {
6297 EVT VT = N->getValueType(0);
6298 if (VT != MVT::f32 && VT != MVT::f64)
6300 // Only optimize when the source and destination types have the same width.
6301 if (VT.getSizeInBits() != N->getOperand(0).getValueType().getSizeInBits())
6304 // If the result of an integer load is only used by an integer-to-float
6305 // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
6306 // This eliminates an "integer-to-vector-move UOP and improve throughput.
6307 SDValue N0 = N->getOperand(0);
6308 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6309 // Do not change the width of a volatile load.
6310 !cast<LoadSDNode>(N0)->isVolatile()) {
6311 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6312 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
6313 LN0->getPointerInfo(), LN0->isVolatile(),
6314 LN0->isNonTemporal(), LN0->isInvariant(),
6315 LN0->getAlignment());
6317 // Make sure successors of the original load stay after it by updating them
6318 // to use the new Chain.
6319 DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));
6322 (N->getOpcode() == ISD::SINT_TO_FP) ? ARM64ISD::SITOF : ARM64ISD::UITOF;
6323 return DAG.getNode(Opcode, SDLoc(N), VT, Load);
6329 /// An EXTR instruction is made up of two shifts, ORed together. This helper
6330 /// searches for and classifies those shifts.
6331 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
6333 if (N.getOpcode() == ISD::SHL)
6335 else if (N.getOpcode() == ISD::SRL)
6340 if (!isa<ConstantSDNode>(N.getOperand(1)))
6343 ShiftAmount = N->getConstantOperandVal(1);
6344 Src = N->getOperand(0);
6348 /// EXTR instruction extracts a contiguous chunk of bits from two existing
6349 /// registers viewed as a high/low pair. This function looks for the pattern:
6350 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
6351 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
6353 static SDValue tryCombineToEXTR(SDNode *N,
6354 TargetLowering::DAGCombinerInfo &DCI) {
6355 SelectionDAG &DAG = DCI.DAG;
6357 EVT VT = N->getValueType(0);
6359 assert(N->getOpcode() == ISD::OR && "Unexpected root");
6361 if (VT != MVT::i32 && VT != MVT::i64)
6365 uint32_t ShiftLHS = 0;
6367 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
6371 uint32_t ShiftRHS = 0;
6373 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
6376 // If they're both trying to come from the high part of the register, they're
6377 // not really an EXTR.
6378 if (LHSFromHi == RHSFromHi)
6381 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
6385 std::swap(LHS, RHS);
6386 std::swap(ShiftLHS, ShiftRHS);
6389 return DAG.getNode(ARM64ISD::EXTR, DL, VT, LHS, RHS,
6390 DAG.getConstant(ShiftRHS, MVT::i64));
6393 static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
6394 const ARM64Subtarget *Subtarget) {
6395 // Attempt to form an EXTR from (or (shl VAL1, #N), (srl VAL2, #RegWidth-N))
6396 if (!EnableARM64ExtrGeneration)
6398 SelectionDAG &DAG = DCI.DAG;
6399 EVT VT = N->getValueType(0);
6401 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6404 SDValue Res = tryCombineToEXTR(N, DCI);
6411 static SDValue performBitcastCombine(SDNode *N,
6412 TargetLowering::DAGCombinerInfo &DCI,
6413 SelectionDAG &DAG) {
6414 // Wait 'til after everything is legalized to try this. That way we have
6415 // legal vector types and such.
6416 if (DCI.isBeforeLegalizeOps())
6419 // Remove extraneous bitcasts around an extract_subvector.
6421 // (v4i16 (bitconvert
6422 // (extract_subvector (v2i64 (bitconvert (v8i16 ...)), (i64 1)))))
6424 // (extract_subvector ((v8i16 ...), (i64 4)))
6426 // Only interested in 64-bit vectors as the ultimate result.
6427 EVT VT = N->getValueType(0);
6430 if (VT.getSimpleVT().getSizeInBits() != 64)
6432 // Is the operand an extract_subvector starting at the beginning or halfway
6433 // point of the vector? A low half may also come through as an
6434 // EXTRACT_SUBREG, so look for that, too.
6435 SDValue Op0 = N->getOperand(0);
6436 if (Op0->getOpcode() != ISD::EXTRACT_SUBVECTOR &&
6437 !(Op0->isMachineOpcode() &&
6438 Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG))
6440 uint64_t idx = cast<ConstantSDNode>(Op0->getOperand(1))->getZExtValue();
6441 if (Op0->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
6442 if (Op0->getValueType(0).getVectorNumElements() != idx && idx != 0)
6444 } else if (Op0->getMachineOpcode() == ARM64::EXTRACT_SUBREG) {
6445 if (idx != ARM64::dsub)
6447 // The dsub reference is equivalent to a lane zero subvector reference.
6450 // Look through the bitcast of the input to the extract.
6451 if (Op0->getOperand(0)->getOpcode() != ISD::BITCAST)
6453 SDValue Source = Op0->getOperand(0)->getOperand(0);
6454 // If the source type has twice the number of elements as our destination
6455 // type, we know this is an extract of the high or low half of the vector.
6456 EVT SVT = Source->getValueType(0);
6457 if (SVT.getVectorNumElements() != VT.getVectorNumElements() * 2)
6460 DEBUG(dbgs() << "arm64-lower: bitcast extract_subvector simplification\n");
6462 // Create the simplified form to just extract the low or high half of the
6463 // vector directly rather than bothering with the bitcasts.
6465 unsigned NumElements = VT.getVectorNumElements();
6467 SDValue HalfIdx = DAG.getConstant(NumElements, MVT::i64);
6468 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Source, HalfIdx);
6470 SDValue SubReg = DAG.getTargetConstant(ARM64::dsub, MVT::i32);
6471 return SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl, VT,
6477 static SDValue performConcatVectorsCombine(SDNode *N,
6478 TargetLowering::DAGCombinerInfo &DCI,
6479 SelectionDAG &DAG) {
6480 // Wait 'til after everything is legalized to try this. That way we have
6481 // legal vector types and such.
6482 if (DCI.isBeforeLegalizeOps())
6486 EVT VT = N->getValueType(0);
6488 // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
6489 // splat. The indexed instructions are going to be expecting a DUPLANE64, so
6490 // canonicalise to that.
6491 if (N->getOperand(0) == N->getOperand(1) && VT.getVectorNumElements() == 2) {
6492 assert(VT.getVectorElementType().getSizeInBits() == 64);
6493 return DAG.getNode(ARM64ISD::DUPLANE64, dl, VT,
6494 WidenVector(N->getOperand(0), DAG),
6495 DAG.getConstant(0, MVT::i64));
6498 // Canonicalise concat_vectors so that the right-hand vector has as few
6499 // bit-casts as possible before its real operation. The primary matching
6500 // destination for these operations will be the narrowing "2" instructions,
6501 // which depend on the operation being performed on this right-hand vector.
6503 // (concat_vectors LHS, (v1i64 (bitconvert (v4i16 RHS))))
6505 // (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))
6507 SDValue Op1 = N->getOperand(1);
6508 if (Op1->getOpcode() != ISD::BITCAST)
6510 SDValue RHS = Op1->getOperand(0);
6511 MVT RHSTy = RHS.getValueType().getSimpleVT();
6512 // If the RHS is not a vector, this is not the pattern we're looking for.
6513 if (!RHSTy.isVector())
6516 DEBUG(dbgs() << "arm64-lower: concat_vectors bitcast simplification\n");
6518 MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
6519 RHSTy.getVectorNumElements() * 2);
6521 ISD::BITCAST, dl, VT,
6522 DAG.getNode(ISD::CONCAT_VECTORS, dl, ConcatTy,
6523 DAG.getNode(ISD::BITCAST, dl, RHSTy, N->getOperand(0)), RHS));
6526 static SDValue tryCombineFixedPointConvert(SDNode *N,
6527 TargetLowering::DAGCombinerInfo &DCI,
6528 SelectionDAG &DAG) {
6529 // Wait 'til after everything is legalized to try this. That way we have
6530 // legal vector types and such.
6531 if (DCI.isBeforeLegalizeOps())
6533 // Transform a scalar conversion of a value from a lane extract into a
6534 // lane extract of a vector conversion. E.g., from foo1 to foo2:
6535 // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
6536 // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
6538 // The second form interacts better with instruction selection and the
6539 // register allocator to avoid cross-class register copies that aren't
6540 // coalescable due to a lane reference.
6542 // Check the operand and see if it originates from a lane extract.
6543 SDValue Op1 = N->getOperand(1);
6544 if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6545 // Yep, no additional predication needed. Perform the transform.
6546 SDValue IID = N->getOperand(0);
6547 SDValue Shift = N->getOperand(2);
6548 SDValue Vec = Op1.getOperand(0);
6549 SDValue Lane = Op1.getOperand(1);
6550 EVT ResTy = N->getValueType(0);
6554 // The vector width should be 128 bits by the time we get here, even
6555 // if it started as 64 bits (the extract_vector handling will have
6557 assert(Vec.getValueType().getSizeInBits() == 128 &&
6558 "unexpected vector size on extract_vector_elt!");
6559 if (Vec.getValueType() == MVT::v4i32)
6560 VecResTy = MVT::v4f32;
6561 else if (Vec.getValueType() == MVT::v2i64)
6562 VecResTy = MVT::v2f64;
6564 assert(0 && "unexpected vector type!");
6567 DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
6568 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
6573 // AArch64 high-vector "long" operations are formed by performing the non-high
6574 // version on an extract_subvector of each operand which gets the high half:
6576 // (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
6578 // However, there are cases which don't have an extract_high explicitly, but
6579 // have another operation that can be made compatible with one for free. For
6582 // (dupv64 scalar) --> (extract_high (dup128 scalar))
6584 // This routine does the actual conversion of such DUPs, once outer routines
6585 // have determined that everything else is in order.
6586 static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
6587 // We can handle most types of duplicate, but the lane ones have an extra
6588 // operand saying *which* lane, so we need to know.
6590 switch (N.getOpcode()) {
6594 case ARM64ISD::DUPLANE8:
6595 case ARM64ISD::DUPLANE16:
6596 case ARM64ISD::DUPLANE32:
6597 case ARM64ISD::DUPLANE64:
6604 MVT NarrowTy = N.getSimpleValueType();
6605 if (!NarrowTy.is64BitVector())
6608 MVT ElementTy = NarrowTy.getVectorElementType();
6609 unsigned NumElems = NarrowTy.getVectorNumElements();
6610 MVT NewDUPVT = MVT::getVectorVT(ElementTy, NumElems * 2);
6614 NewDUP = DAG.getNode(N.getOpcode(), SDLoc(N), NewDUPVT, N.getOperand(0),
6617 NewDUP = DAG.getNode(ARM64ISD::DUP, SDLoc(N), NewDUPVT, N.getOperand(0));
6619 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N.getNode()), NarrowTy,
6620 NewDUP, DAG.getConstant(NumElems, MVT::i64));
6623 static bool isEssentiallyExtractSubvector(SDValue N) {
6624 if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR)
6627 return N.getOpcode() == ISD::BITCAST &&
6628 N.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR;
6631 /// \brief Helper structure to keep track of ISD::SET_CC operands.
6632 struct GenericSetCCInfo {
6633 const SDValue *Opnd0;
6634 const SDValue *Opnd1;
6638 /// \brief Helper structure to keep track of a SET_CC lowered into ARM64 code.
6639 struct ARM64SetCCInfo {
6641 ARM64CC::CondCode CC;
6644 /// \brief Helper structure to keep track of SetCC information.
6646 GenericSetCCInfo Generic;
6647 ARM64SetCCInfo ARM64;
6650 /// \brief Helper structure to be able to read SetCC information.
6651 /// If set to true, IsARM64 field, Info is a ARM64SetCCInfo, otherwise Info is
6652 /// a GenericSetCCInfo.
6653 struct SetCCInfoAndKind {
6658 /// \brief Check whether or not \p Op is a SET_CC operation, either a generic or
6660 /// ARM64 lowered one.
6661 /// \p SetCCInfo is filled accordingly.
6662 /// \post SetCCInfo is meanginfull only when this function returns true.
6663 /// \return True when Op is a kind of SET_CC operation.
6664 static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
6665 // If this is a setcc, this is straight forward.
6666 if (Op.getOpcode() == ISD::SETCC) {
6667 SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
6668 SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
6669 SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6670 SetCCInfo.IsARM64 = false;
6673 // Otherwise, check if this is a matching csel instruction.
6677 if (Op.getOpcode() != ARM64ISD::CSEL)
6679 // Set the information about the operands.
6680 // TODO: we want the operands of the Cmp not the csel
6681 SetCCInfo.Info.ARM64.Cmp = &Op.getOperand(3);
6682 SetCCInfo.IsARM64 = true;
6683 SetCCInfo.Info.ARM64.CC = static_cast<ARM64CC::CondCode>(
6684 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
6686 // Check that the operands matches the constraints:
6687 // (1) Both operands must be constants.
6688 // (2) One must be 1 and the other must be 0.
6689 ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
6690 ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6693 if (!TValue || !FValue)
6697 if (!TValue->isOne()) {
6698 // Update the comparison when we are interested in !cc.
6699 std::swap(TValue, FValue);
6700 SetCCInfo.Info.ARM64.CC =
6701 ARM64CC::getInvertedCondCode(SetCCInfo.Info.ARM64.CC);
6703 return TValue->isOne() && FValue->isNullValue();
6706 // The folding we want to perform is:
6707 // (add x, (setcc cc ...) )
6709 // (csel x, (add x, 1), !cc ...)
6711 // The latter will get matched to a CSINC instruction.
6712 static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
6713 assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
6714 SDValue LHS = Op->getOperand(0);
6715 SDValue RHS = Op->getOperand(1);
6716 SetCCInfoAndKind InfoAndKind;
6718 // If neither operand is a SET_CC, give up.
6719 if (!isSetCC(LHS, InfoAndKind)) {
6720 std::swap(LHS, RHS);
6721 if (!isSetCC(LHS, InfoAndKind))
6725 // FIXME: This could be generatized to work for FP comparisons.
6726 EVT CmpVT = InfoAndKind.IsARM64
6727 ? InfoAndKind.Info.ARM64.Cmp->getOperand(0).getValueType()
6728 : InfoAndKind.Info.Generic.Opnd0->getValueType();
6729 if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
6735 if (InfoAndKind.IsARM64) {
6736 CCVal = DAG.getConstant(
6737 ARM64CC::getInvertedCondCode(InfoAndKind.Info.ARM64.CC), MVT::i32);
6738 Cmp = *InfoAndKind.Info.ARM64.Cmp;
6740 Cmp = getARM64Cmp(*InfoAndKind.Info.Generic.Opnd0,
6741 *InfoAndKind.Info.Generic.Opnd1,
6742 ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, true),
6745 EVT VT = Op->getValueType(0);
6746 LHS = DAG.getNode(ISD::ADD, dl, VT, RHS, DAG.getConstant(1, VT));
6747 return DAG.getNode(ARM64ISD::CSEL, dl, VT, RHS, LHS, CCVal, Cmp);
6750 // The basic add/sub long vector instructions have variants with "2" on the end
6751 // which act on the high-half of their inputs. They are normally matched by
6754 // (add (zeroext (extract_high LHS)),
6755 // (zeroext (extract_high RHS)))
6756 // -> uaddl2 vD, vN, vM
6758 // However, if one of the extracts is something like a duplicate, this
6759 // instruction can still be used profitably. This function puts the DAG into a
6760 // more appropriate form for those patterns to trigger.
6761 static SDValue performAddSubLongCombine(SDNode *N,
6762 TargetLowering::DAGCombinerInfo &DCI,
6763 SelectionDAG &DAG) {
6764 if (DCI.isBeforeLegalizeOps())
6767 MVT VT = N->getSimpleValueType(0);
6768 if (!VT.is128BitVector()) {
6769 if (N->getOpcode() == ISD::ADD)
6770 return performSetccAddFolding(N, DAG);
6774 // Make sure both branches are extended in the same way.
6775 SDValue LHS = N->getOperand(0);
6776 SDValue RHS = N->getOperand(1);
6777 if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
6778 LHS.getOpcode() != ISD::SIGN_EXTEND) ||
6779 LHS.getOpcode() != RHS.getOpcode())
6782 unsigned ExtType = LHS.getOpcode();
6784 // It's not worth doing if at least one of the inputs isn't already an
6785 // extract, but we don't know which it'll be so we have to try both.
6786 if (isEssentiallyExtractSubvector(LHS.getOperand(0))) {
6787 RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
6791 RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
6792 } else if (isEssentiallyExtractSubvector(RHS.getOperand(0))) {
6793 LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
6797 LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
6800 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
6803 // Massage DAGs which we can use the high-half "long" operations on into
6804 // something isel will recognize better. E.g.
6806 // (arm64_neon_umull (extract_high vec) (dupv64 scalar)) -->
6807 // (arm64_neon_umull (extract_high (v2i64 vec)))
6808 // (extract_high (v2i64 (dup128 scalar)))))
6810 static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
6811 TargetLowering::DAGCombinerInfo &DCI,
6812 SelectionDAG &DAG) {
6813 if (DCI.isBeforeLegalizeOps())
6816 SDValue LHS = N->getOperand(1);
6817 SDValue RHS = N->getOperand(2);
6818 assert(LHS.getValueType().is64BitVector() &&
6819 RHS.getValueType().is64BitVector() &&
6820 "unexpected shape for long operation");
6822 // Either node could be a DUP, but it's not worth doing both of them (you'd
6823 // just as well use the non-high version) so look for a corresponding extract
6824 // operation on the other "wing".
6825 if (isEssentiallyExtractSubvector(LHS)) {
6826 RHS = tryExtendDUPToExtractHigh(RHS, DAG);
6829 } else if (isEssentiallyExtractSubvector(RHS)) {
6830 LHS = tryExtendDUPToExtractHigh(LHS, DAG);
6835 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
6836 N->getOperand(0), LHS, RHS);
6839 static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
6840 MVT ElemTy = N->getSimpleValueType(0).getScalarType();
6841 unsigned ElemBits = ElemTy.getSizeInBits();
6843 int64_t ShiftAmount;
6844 if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
6845 APInt SplatValue, SplatUndef;
6846 unsigned SplatBitSize;
6848 if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
6849 HasAnyUndefs, ElemBits) ||
6850 SplatBitSize != ElemBits)
6853 ShiftAmount = SplatValue.getSExtValue();
6854 } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
6855 ShiftAmount = CVN->getSExtValue();
6863 llvm_unreachable("Unknown shift intrinsic");
6864 case Intrinsic::arm64_neon_sqshl:
6865 Opcode = ARM64ISD::SQSHL_I;
6866 IsRightShift = false;
6868 case Intrinsic::arm64_neon_uqshl:
6869 Opcode = ARM64ISD::UQSHL_I;
6870 IsRightShift = false;
6872 case Intrinsic::arm64_neon_srshl:
6873 Opcode = ARM64ISD::SRSHR_I;
6874 IsRightShift = true;
6876 case Intrinsic::arm64_neon_urshl:
6877 Opcode = ARM64ISD::URSHR_I;
6878 IsRightShift = true;
6880 case Intrinsic::arm64_neon_sqshlu:
6881 Opcode = ARM64ISD::SQSHLU_I;
6882 IsRightShift = false;
6886 if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits)
6887 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6888 DAG.getConstant(-ShiftAmount, MVT::i32));
6889 else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount <= ElemBits)
6890 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), N->getOperand(1),
6891 DAG.getConstant(ShiftAmount, MVT::i32));
6896 // The CRC32[BH] instructions ignore the high bits of their data operand. Since
6897 // the intrinsics must be legal and take an i32, this means there's almost
6898 // certainly going to be a zext in the DAG which we can eliminate.
6899 static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
6900 SDValue AndN = N->getOperand(2);
6901 if (AndN.getOpcode() != ISD::AND)
6904 ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
6905 if (!CMask || CMask->getZExtValue() != Mask)
6908 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
6909 N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
6912 static SDValue performIntrinsicCombine(SDNode *N,
6913 TargetLowering::DAGCombinerInfo &DCI,
6914 const ARM64Subtarget *Subtarget) {
6915 SelectionDAG &DAG = DCI.DAG;
6916 unsigned IID = getIntrinsicID(N);
6920 case Intrinsic::arm64_neon_vcvtfxs2fp:
6921 case Intrinsic::arm64_neon_vcvtfxu2fp:
6922 return tryCombineFixedPointConvert(N, DCI, DAG);
6924 case Intrinsic::arm64_neon_fmax:
6925 return DAG.getNode(ARM64ISD::FMAX, SDLoc(N), N->getValueType(0),
6926 N->getOperand(1), N->getOperand(2));
6927 case Intrinsic::arm64_neon_fmin:
6928 return DAG.getNode(ARM64ISD::FMIN, SDLoc(N), N->getValueType(0),
6929 N->getOperand(1), N->getOperand(2));
6930 case Intrinsic::arm64_neon_smull:
6931 case Intrinsic::arm64_neon_umull:
6932 case Intrinsic::arm64_neon_pmull:
6933 case Intrinsic::arm64_neon_sqdmull:
6934 return tryCombineLongOpWithDup(IID, N, DCI, DAG);
6935 case Intrinsic::arm64_neon_sqshl:
6936 case Intrinsic::arm64_neon_uqshl:
6937 case Intrinsic::arm64_neon_sqshlu:
6938 case Intrinsic::arm64_neon_srshl:
6939 case Intrinsic::arm64_neon_urshl:
6940 return tryCombineShiftImm(IID, N, DAG);
6941 case Intrinsic::arm64_crc32b:
6942 case Intrinsic::arm64_crc32cb:
6943 return tryCombineCRC32(0xff, N, DAG);
6944 case Intrinsic::arm64_crc32h:
6945 case Intrinsic::arm64_crc32ch:
6946 return tryCombineCRC32(0xffff, N, DAG);
6951 static SDValue performExtendCombine(SDNode *N,
6952 TargetLowering::DAGCombinerInfo &DCI,
6953 SelectionDAG &DAG) {
6954 // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
6955 // we can convert that DUP into another extract_high (of a bigger DUP), which
6956 // helps the backend to decide that an sabdl2 would be useful, saving a real
6957 // extract_high operation.
6958 if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
6959 N->getOperand(0).getOpcode() == ISD::INTRINSIC_WO_CHAIN) {
6960 SDNode *ABDNode = N->getOperand(0).getNode();
6961 unsigned IID = getIntrinsicID(ABDNode);
6962 if (IID == Intrinsic::arm64_neon_sabd ||
6963 IID == Intrinsic::arm64_neon_uabd) {
6964 SDValue NewABD = tryCombineLongOpWithDup(IID, ABDNode, DCI, DAG);
6965 if (!NewABD.getNode())
6968 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0),
6973 // This is effectively a custom type legalization for ARM64.
6975 // Type legalization will split an extend of a small, legal, type to a larger
6976 // illegal type by first splitting the destination type, often creating
6977 // illegal source types, which then get legalized in isel-confusing ways,
6978 // leading to really terrible codegen. E.g.,
6979 // %result = v8i32 sext v8i8 %value
6981 // %losrc = extract_subreg %value, ...
6982 // %hisrc = extract_subreg %value, ...
6983 // %lo = v4i32 sext v4i8 %losrc
6984 // %hi = v4i32 sext v4i8 %hisrc
6985 // Things go rapidly downhill from there.
6987 // For ARM64, the [sz]ext vector instructions can only go up one element
6988 // size, so we can, e.g., extend from i8 to i16, but to go from i8 to i32
6989 // take two instructions.
6991 // This implies that the most efficient way to do the extend from v8i8
6992 // to two v4i32 values is to first extend the v8i8 to v8i16, then do
6993 // the normal splitting to happen for the v8i16->v8i32.
6995 // This is pre-legalization to catch some cases where the default
6996 // type legalization will create ill-tempered code.
6997 if (!DCI.isBeforeLegalizeOps())
7000 // We're only interested in cleaning things up for non-legal vector types
7001 // here. If both the source and destination are legal, things will just
7002 // work naturally without any fiddling.
7003 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7004 EVT ResVT = N->getValueType(0);
7005 if (!ResVT.isVector() || TLI.isTypeLegal(ResVT))
7007 // If the vector type isn't a simple VT, it's beyond the scope of what
7008 // we're worried about here. Let legalization do its thing and hope for
7010 if (!ResVT.isSimple())
7013 SDValue Src = N->getOperand(0);
7014 MVT SrcVT = Src->getValueType(0).getSimpleVT();
7015 // If the source VT is a 64-bit vector, we can play games and get the
7016 // better results we want.
7017 if (SrcVT.getSizeInBits() != 64)
7020 unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
7021 unsigned ElementCount = SrcVT.getVectorNumElements();
7022 SrcVT = MVT::getVectorVT(MVT::getIntegerVT(SrcEltSize * 2), ElementCount);
7024 Src = DAG.getNode(N->getOpcode(), DL, SrcVT, Src);
7026 // Now split the rest of the operation into two halves, each with a 64
7030 unsigned NumElements = ResVT.getVectorNumElements();
7031 assert(!(NumElements & 1) && "Splitting vector, but not in half!");
7032 LoVT = HiVT = EVT::getVectorVT(*DAG.getContext(),
7033 ResVT.getVectorElementType(), NumElements / 2);
7035 EVT InNVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
7036 LoVT.getVectorNumElements());
7037 Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7038 DAG.getIntPtrConstant(0));
7039 Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InNVT, Src,
7040 DAG.getIntPtrConstant(InNVT.getVectorNumElements()));
7041 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, Lo);
7042 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, Hi);
7044 // Now combine the parts back together so we still have a single result
7045 // like the combiner expects.
7046 return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Lo, Hi);
7049 /// Replace a splat of a scalar to a vector store by scalar stores of the scalar
7050 /// value. The load store optimizer pass will merge them to store pair stores.
7051 /// This has better performance than a splat of the scalar followed by a split
7052 /// vector store. Even if the stores are not merged it is four stores vs a dup,
7053 /// followed by an ext.b and two stores.
7054 static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode *St) {
7055 SDValue StVal = St->getValue();
7056 EVT VT = StVal.getValueType();
7058 // Don't replace floating point stores, they possibly won't be transformed to
7059 // stp because of the store pair suppress pass.
7060 if (VT.isFloatingPoint())
7063 // Check for insert vector elements.
7064 if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
7067 // We can express a splat as store pair(s) for 2 or 4 elements.
7068 unsigned NumVecElts = VT.getVectorNumElements();
7069 if (NumVecElts != 4 && NumVecElts != 2)
7071 SDValue SplatVal = StVal.getOperand(1);
7072 unsigned RemainInsertElts = NumVecElts - 1;
7074 // Check that this is a splat.
7075 while (--RemainInsertElts) {
7076 SDValue NextInsertElt = StVal.getOperand(0);
7077 if (NextInsertElt.getOpcode() != ISD::INSERT_VECTOR_ELT)
7079 if (NextInsertElt.getOperand(1) != SplatVal)
7081 StVal = NextInsertElt;
7083 unsigned OrigAlignment = St->getAlignment();
7084 unsigned EltOffset = NumVecElts == 4 ? 4 : 8;
7085 unsigned Alignment = std::min(OrigAlignment, EltOffset);
7087 // Create scalar stores. This is at least as good as the code sequence for a
7088 // split unaligned store wich is a dup.s, ext.b, and two stores.
7089 // Most of the time the three stores should be replaced by store pair
7090 // instructions (stp).
7092 SDValue BasePtr = St->getBasePtr();
7094 DAG.getStore(St->getChain(), DL, SplatVal, BasePtr, St->getPointerInfo(),
7095 St->isVolatile(), St->isNonTemporal(), St->getAlignment());
7097 unsigned Offset = EltOffset;
7098 while (--NumVecElts) {
7099 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7100 DAG.getConstant(Offset, MVT::i64));
7101 NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
7102 St->getPointerInfo(), St->isVolatile(),
7103 St->isNonTemporal(), Alignment);
7104 Offset += EltOffset;
7109 static SDValue performSTORECombine(SDNode *N,
7110 TargetLowering::DAGCombinerInfo &DCI,
7112 const ARM64Subtarget *Subtarget) {
7113 if (!DCI.isBeforeLegalize())
7116 StoreSDNode *S = cast<StoreSDNode>(N);
7117 if (S->isVolatile())
7120 // Cyclone has bad performance on unaligned 16B stores when crossing line and
7121 // page boundries. We want to split such stores.
7122 if (!Subtarget->isCyclone())
7125 // Don't split at Oz.
7126 MachineFunction &MF = DAG.getMachineFunction();
7127 bool IsMinSize = MF.getFunction()->getAttributes().hasAttribute(
7128 AttributeSet::FunctionIndex, Attribute::MinSize);
7132 SDValue StVal = S->getValue();
7133 EVT VT = StVal.getValueType();
7135 // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
7136 // those up regresses performance on micro-benchmarks and olden/bh.
7137 if (!VT.isVector() || VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
7140 // Split unaligned 16B stores. They are terrible for performance.
7141 // Don't split stores with alignment of 1 or 2. Code that uses clang vector
7142 // extensions can use this to mark that it does not want splitting to happen
7143 // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
7144 // eliminating alignment hazards is only 1 in 8 for alignment of 2.
7145 if (VT.getSizeInBits() != 128 || S->getAlignment() >= 16 ||
7146 S->getAlignment() <= 2)
7149 // If we get a splat of a scalar convert this vector store to a store of
7150 // scalars. They will be merged into store pairs thereby removing two
7152 SDValue ReplacedSplat = replaceSplatVectorStore(DAG, S);
7153 if (ReplacedSplat != SDValue())
7154 return ReplacedSplat;
7157 unsigned NumElts = VT.getVectorNumElements() / 2;
7158 // Split VT into two.
7160 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), NumElts);
7161 SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7162 DAG.getIntPtrConstant(0));
7163 SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
7164 DAG.getIntPtrConstant(NumElts));
7165 SDValue BasePtr = S->getBasePtr();
7167 DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
7168 S->isVolatile(), S->isNonTemporal(), S->getAlignment());
7169 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
7170 DAG.getConstant(8, MVT::i64));
7171 return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
7172 S->getPointerInfo(), S->isVolatile(), S->isNonTemporal(),
7176 // Optimize compare with zero and branch.
7177 static SDValue performBRCONDCombine(SDNode *N,
7178 TargetLowering::DAGCombinerInfo &DCI,
7179 SelectionDAG &DAG) {
7180 SDValue Chain = N->getOperand(0);
7181 SDValue Dest = N->getOperand(1);
7182 SDValue CCVal = N->getOperand(2);
7183 SDValue Cmp = N->getOperand(3);
7185 assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
7186 unsigned CC = cast<ConstantSDNode>(CCVal)->getZExtValue();
7187 if (CC != ARM64CC::EQ && CC != ARM64CC::NE)
7190 unsigned CmpOpc = Cmp.getOpcode();
7191 if (CmpOpc != ARM64ISD::ADDS && CmpOpc != ARM64ISD::SUBS)
7194 // Only attempt folding if there is only one use of the flag and no use of the
7196 if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
7199 SDValue LHS = Cmp.getOperand(0);
7200 SDValue RHS = Cmp.getOperand(1);
7202 assert(LHS.getValueType() == RHS.getValueType() &&
7203 "Expected the value type to be the same for both operands!");
7204 if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
7207 if (isa<ConstantSDNode>(LHS) && cast<ConstantSDNode>(LHS)->isNullValue())
7208 std::swap(LHS, RHS);
7210 if (!isa<ConstantSDNode>(RHS) || !cast<ConstantSDNode>(RHS)->isNullValue())
7213 if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
7214 LHS.getOpcode() == ISD::SRL)
7217 // Fold the compare into the branch instruction.
7219 if (CC == ARM64CC::EQ)
7220 BR = DAG.getNode(ARM64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7222 BR = DAG.getNode(ARM64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
7224 // Do not add new nodes to DAG combiner worklist.
7225 DCI.CombineTo(N, BR, false);
7230 SDValue ARM64TargetLowering::PerformDAGCombine(SDNode *N,
7231 DAGCombinerInfo &DCI) const {
7232 SelectionDAG &DAG = DCI.DAG;
7233 switch (N->getOpcode()) {
7238 return performAddSubLongCombine(N, DCI, DAG);
7240 return performXorCombine(N, DAG, DCI, Subtarget);
7242 return performMulCombine(N, DAG, DCI, Subtarget);
7243 case ISD::SINT_TO_FP:
7244 case ISD::UINT_TO_FP:
7245 return performIntToFpCombine(N, DAG);
7247 return performORCombine(N, DCI, Subtarget);
7248 case ISD::INTRINSIC_WO_CHAIN:
7249 return performIntrinsicCombine(N, DCI, Subtarget);
7250 case ISD::ANY_EXTEND:
7251 case ISD::ZERO_EXTEND:
7252 case ISD::SIGN_EXTEND:
7253 return performExtendCombine(N, DCI, DAG);
7255 return performBitcastCombine(N, DCI, DAG);
7256 case ISD::CONCAT_VECTORS:
7257 return performConcatVectorsCombine(N, DCI, DAG);
7259 return performSTORECombine(N, DCI, DAG, Subtarget);
7260 case ARM64ISD::BRCOND:
7261 return performBRCONDCombine(N, DCI, DAG);
7266 // Check if the return value is used as only a return value, as otherwise
7267 // we can't perform a tail-call. In particular, we need to check for
7268 // target ISD nodes that are returns and any other "odd" constructs
7269 // that the generic analysis code won't necessarily catch.
7270 bool ARM64TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
7271 if (N->getNumValues() != 1)
7273 if (!N->hasNUsesOfValue(1, 0))
7276 SDValue TCChain = Chain;
7277 SDNode *Copy = *N->use_begin();
7278 if (Copy->getOpcode() == ISD::CopyToReg) {
7279 // If the copy has a glue operand, we conservatively assume it isn't safe to
7280 // perform a tail call.
7281 if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
7284 TCChain = Copy->getOperand(0);
7285 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
7288 bool HasRet = false;
7289 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
7291 if (UI->getOpcode() != ARM64ISD::RET_FLAG)
7303 // Return whether the an instruction can potentially be optimized to a tail
7304 // call. This will cause the optimizers to attempt to move, or duplicate,
7305 // return instructions to help enable tail call optimizations for this
7307 bool ARM64TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
7308 if (!EnableARM64TailCalls)
7311 if (!CI->isTailCall())
7317 bool ARM64TargetLowering::getIndexedAddressParts(SDNode *Op, SDValue &Base,
7319 ISD::MemIndexedMode &AM,
7321 SelectionDAG &DAG) const {
7322 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
7325 Base = Op->getOperand(0);
7326 // All of the indexed addressing mode instructions take a signed
7327 // 9 bit immediate offset.
7328 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
7329 int64_t RHSC = (int64_t)RHS->getZExtValue();
7330 if (RHSC >= 256 || RHSC <= -256)
7332 IsInc = (Op->getOpcode() == ISD::ADD);
7333 Offset = Op->getOperand(1);
7339 bool ARM64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
7341 ISD::MemIndexedMode &AM,
7342 SelectionDAG &DAG) const {
7345 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7346 VT = LD->getMemoryVT();
7347 Ptr = LD->getBasePtr();
7348 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7349 VT = ST->getMemoryVT();
7350 Ptr = ST->getBasePtr();
7355 if (!getIndexedAddressParts(Ptr.getNode(), Base, Offset, AM, IsInc, DAG))
7357 AM = IsInc ? ISD::PRE_INC : ISD::PRE_DEC;
7361 bool ARM64TargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
7364 ISD::MemIndexedMode &AM,
7365 SelectionDAG &DAG) const {
7368 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7369 VT = LD->getMemoryVT();
7370 Ptr = LD->getBasePtr();
7371 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7372 VT = ST->getMemoryVT();
7373 Ptr = ST->getBasePtr();
7378 if (!getIndexedAddressParts(Op, Base, Offset, AM, IsInc, DAG))
7380 // Post-indexing updates the base, so it's not a valid transform
7381 // if that's not the same as the load's pointer.
7384 AM = IsInc ? ISD::POST_INC : ISD::POST_DEC;
7388 /// The only 128-bit atomic operation is an stxp that succeeds. In particular
7389 /// neither ldp nor ldxp are atomic. So the canonical sequence for an atomic
7392 /// ldxp x0, x1, [x8]
7393 /// stxp w2, x0, x1, [x8]
7395 /// If the stxp succeeds then the ldxp managed to get both halves without an
7396 /// intervening stxp from a different thread and the read was atomic.
7397 static void ReplaceATOMIC_LOAD_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7398 SelectionDAG &DAG) {
7400 AtomicSDNode *AN = cast<AtomicSDNode>(N);
7401 EVT VT = AN->getMemoryVT();
7402 SDValue Zero = DAG.getConstant(0, VT);
7404 // FIXME: Really want ATOMIC_LOAD_NOP but that doesn't fit into the existing
7405 // scheme very well. Given the complexity of what we're already generating, an
7406 // extra couple of ORRs probably won't make much difference.
7407 SDValue Result = DAG.getAtomic(ISD::ATOMIC_LOAD_OR, DL, AN->getMemoryVT(),
7408 N->getOperand(0), N->getOperand(1), Zero,
7409 AN->getMemOperand(), AN->getOrdering(),
7410 AN->getSynchScope());
7412 Results.push_back(Result.getValue(0)); // Value
7413 Results.push_back(Result.getValue(1)); // Chain
7416 static void ReplaceATOMIC_OP_128(SDNode *N, SmallVectorImpl<SDValue> &Results,
7417 SelectionDAG &DAG, unsigned NewOp) {
7419 AtomicOrdering Ordering = cast<AtomicSDNode>(N)->getOrdering();
7420 assert(N->getValueType(0) == MVT::i128 &&
7421 "Only know how to expand i128 atomics");
7423 SmallVector<SDValue, 6> Ops;
7424 Ops.push_back(N->getOperand(1)); // Ptr
7426 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7427 N->getOperand(2), DAG.getIntPtrConstant(0)));
7428 // High part of Val1
7429 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7430 N->getOperand(2), DAG.getIntPtrConstant(1)));
7431 if (NewOp == ARM64::ATOMIC_CMP_SWAP_I128) {
7433 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7434 N->getOperand(3), DAG.getIntPtrConstant(0)));
7435 // High part of Val2
7436 Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i64,
7437 N->getOperand(3), DAG.getIntPtrConstant(1)));
7440 Ops.push_back(DAG.getTargetConstant(Ordering, MVT::i32));
7441 Ops.push_back(N->getOperand(0)); // Chain
7443 SDVTList Tys = DAG.getVTList(MVT::i64, MVT::i64, MVT::Other);
7444 SDNode *Result = DAG.getMachineNode(NewOp, DL, Tys, Ops);
7445 SDValue OpsF[] = { SDValue(Result, 0), SDValue(Result, 1) };
7446 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128, OpsF, 2));
7447 Results.push_back(SDValue(Result, 2));
7450 void ARM64TargetLowering::ReplaceNodeResults(SDNode *N,
7451 SmallVectorImpl<SDValue> &Results,
7452 SelectionDAG &DAG) const {
7453 switch (N->getOpcode()) {
7455 llvm_unreachable("Don't know how to custom expand this");
7456 case ISD::ATOMIC_LOAD:
7457 ReplaceATOMIC_LOAD_128(N, Results, DAG);
7459 case ISD::ATOMIC_LOAD_ADD:
7460 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_ADD_I128);
7462 case ISD::ATOMIC_LOAD_SUB:
7463 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_SUB_I128);
7465 case ISD::ATOMIC_LOAD_AND:
7466 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_AND_I128);
7468 case ISD::ATOMIC_LOAD_OR:
7469 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_OR_I128);
7471 case ISD::ATOMIC_LOAD_XOR:
7472 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_XOR_I128);
7474 case ISD::ATOMIC_LOAD_NAND:
7475 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_NAND_I128);
7477 case ISD::ATOMIC_SWAP:
7478 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_SWAP_I128);
7480 case ISD::ATOMIC_LOAD_MIN:
7481 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MIN_I128);
7483 case ISD::ATOMIC_LOAD_MAX:
7484 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_MAX_I128);
7486 case ISD::ATOMIC_LOAD_UMIN:
7487 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMIN_I128);
7489 case ISD::ATOMIC_LOAD_UMAX:
7490 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_LOAD_UMAX_I128);
7492 case ISD::ATOMIC_CMP_SWAP:
7493 ReplaceATOMIC_OP_128(N, Results, DAG, ARM64::ATOMIC_CMP_SWAP_I128);
7495 case ISD::FP_TO_UINT:
7496 case ISD::FP_TO_SINT:
7497 assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
7498 // Let normal code take care of it by not adding anything to Results.