1 //===-- ARMISelLowering.cpp - ARM 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 defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "arm-isel"
17 #include "ARMAddressingModes.h"
18 #include "ARMConstantPoolValue.h"
19 #include "ARMISelLowering.h"
20 #include "ARMMachineFunctionInfo.h"
21 #include "ARMPerfectShuffle.h"
22 #include "ARMRegisterInfo.h"
23 #include "ARMSubtarget.h"
24 #include "ARMTargetMachine.h"
25 #include "ARMTargetObjectFile.h"
26 #include "llvm/CallingConv.h"
27 #include "llvm/Constants.h"
28 #include "llvm/Function.h"
29 #include "llvm/GlobalValue.h"
30 #include "llvm/Instruction.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Type.h"
33 #include "llvm/CodeGen/CallingConvLower.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineFrameInfo.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineInstrBuilder.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/CodeGen/PseudoSourceValue.h"
40 #include "llvm/CodeGen/SelectionDAG.h"
41 #include "llvm/MC/MCSectionMachO.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/ADT/VectorExtras.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
52 STATISTIC(NumTailCalls, "Number of tail calls");
54 // This option should go away when tail calls fully work.
56 EnableARMTailCalls("arm-tail-calls", cl::Hidden,
57 cl::desc("Generate tail calls (TEMPORARY OPTION)."),
61 EnableARMLongCalls("arm-long-calls", cl::Hidden,
62 cl::desc("Generate calls via indirect call instructions."),
66 ARMInterworking("arm-interworking", cl::Hidden,
67 cl::desc("Enable / disable ARM interworking (for debugging only)"),
70 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
71 CCValAssign::LocInfo &LocInfo,
72 ISD::ArgFlagsTy &ArgFlags,
74 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
75 CCValAssign::LocInfo &LocInfo,
76 ISD::ArgFlagsTy &ArgFlags,
78 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
79 CCValAssign::LocInfo &LocInfo,
80 ISD::ArgFlagsTy &ArgFlags,
82 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
83 CCValAssign::LocInfo &LocInfo,
84 ISD::ArgFlagsTy &ArgFlags,
87 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
88 EVT PromotedBitwiseVT) {
89 if (VT != PromotedLdStVT) {
90 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
91 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
92 PromotedLdStVT.getSimpleVT());
94 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
95 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
96 PromotedLdStVT.getSimpleVT());
99 EVT ElemTy = VT.getVectorElementType();
100 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
101 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
102 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
103 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
104 if (ElemTy != MVT::i32) {
105 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
108 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
110 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
111 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
112 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
113 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
116 if (VT.isInteger()) {
117 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
118 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
119 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
122 // Promote all bit-wise operations.
123 if (VT.isInteger() && VT != PromotedBitwiseVT) {
124 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
125 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
126 PromotedBitwiseVT.getSimpleVT());
127 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
128 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
129 PromotedBitwiseVT.getSimpleVT());
130 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
131 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
132 PromotedBitwiseVT.getSimpleVT());
135 // Neon does not support vector divide/remainder operations.
136 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
137 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
138 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
139 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
140 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
141 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
144 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
145 addRegisterClass(VT, ARM::DPRRegisterClass);
146 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
149 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
150 addRegisterClass(VT, ARM::QPRRegisterClass);
151 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
154 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
155 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
156 return new TargetLoweringObjectFileMachO();
158 return new ARMElfTargetObjectFile();
161 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
162 : TargetLowering(TM, createTLOF(TM)) {
163 Subtarget = &TM.getSubtarget<ARMSubtarget>();
165 if (Subtarget->isTargetDarwin()) {
166 // Uses VFP for Thumb libfuncs if available.
167 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
168 // Single-precision floating-point arithmetic.
169 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
170 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
171 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
172 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
174 // Double-precision floating-point arithmetic.
175 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
176 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
177 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
178 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
180 // Single-precision comparisons.
181 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
182 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
183 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
184 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
185 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
186 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
187 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
188 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
190 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
192 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
193 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
194 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
195 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
196 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
197 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
199 // Double-precision comparisons.
200 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
201 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
202 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
203 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
204 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
205 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
206 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
207 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
209 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
210 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
211 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
212 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
213 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
214 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
215 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
216 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
218 // Floating-point to integer conversions.
219 // i64 conversions are done via library routines even when generating VFP
220 // instructions, so use the same ones.
221 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
222 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
223 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
224 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
226 // Conversions between floating types.
227 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
228 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
230 // Integer to floating-point conversions.
231 // i64 conversions are done via library routines even when generating VFP
232 // instructions, so use the same ones.
233 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
234 // e.g., __floatunsidf vs. __floatunssidfvfp.
235 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
236 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
237 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
238 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
242 // These libcalls are not available in 32-bit.
243 setLibcallName(RTLIB::SHL_I128, 0);
244 setLibcallName(RTLIB::SRL_I128, 0);
245 setLibcallName(RTLIB::SRA_I128, 0);
247 // Libcalls should use the AAPCS base standard ABI, even if hard float
248 // is in effect, as per the ARM RTABI specification, section 4.1.2.
249 if (Subtarget->isAAPCS_ABI()) {
250 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
251 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
252 CallingConv::ARM_AAPCS);
256 if (Subtarget->isThumb1Only())
257 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
259 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
260 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
261 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
262 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
264 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
267 if (Subtarget->hasNEON()) {
268 addDRTypeForNEON(MVT::v2f32);
269 addDRTypeForNEON(MVT::v8i8);
270 addDRTypeForNEON(MVT::v4i16);
271 addDRTypeForNEON(MVT::v2i32);
272 addDRTypeForNEON(MVT::v1i64);
274 addQRTypeForNEON(MVT::v4f32);
275 addQRTypeForNEON(MVT::v2f64);
276 addQRTypeForNEON(MVT::v16i8);
277 addQRTypeForNEON(MVT::v8i16);
278 addQRTypeForNEON(MVT::v4i32);
279 addQRTypeForNEON(MVT::v2i64);
281 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
282 // neither Neon nor VFP support any arithmetic operations on it.
283 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
284 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
285 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
286 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
287 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
288 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
289 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
290 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
291 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
292 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
293 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
294 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
295 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
296 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
297 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
298 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
299 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
300 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
301 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
302 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
303 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
304 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
305 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
306 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
308 // Neon does not support some operations on v1i64 and v2i64 types.
309 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
310 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
311 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
312 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
314 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
315 setTargetDAGCombine(ISD::SHL);
316 setTargetDAGCombine(ISD::SRL);
317 setTargetDAGCombine(ISD::SRA);
318 setTargetDAGCombine(ISD::SIGN_EXTEND);
319 setTargetDAGCombine(ISD::ZERO_EXTEND);
320 setTargetDAGCombine(ISD::ANY_EXTEND);
321 setTargetDAGCombine(ISD::SELECT_CC);
324 computeRegisterProperties();
326 // ARM does not have f32 extending load.
327 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
329 // ARM does not have i1 sign extending load.
330 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
332 // ARM supports all 4 flavors of integer indexed load / store.
333 if (!Subtarget->isThumb1Only()) {
334 for (unsigned im = (unsigned)ISD::PRE_INC;
335 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
336 setIndexedLoadAction(im, MVT::i1, Legal);
337 setIndexedLoadAction(im, MVT::i8, Legal);
338 setIndexedLoadAction(im, MVT::i16, Legal);
339 setIndexedLoadAction(im, MVT::i32, Legal);
340 setIndexedStoreAction(im, MVT::i1, Legal);
341 setIndexedStoreAction(im, MVT::i8, Legal);
342 setIndexedStoreAction(im, MVT::i16, Legal);
343 setIndexedStoreAction(im, MVT::i32, Legal);
347 // i64 operation support.
348 if (Subtarget->isThumb1Only()) {
349 setOperationAction(ISD::MUL, MVT::i64, Expand);
350 setOperationAction(ISD::MULHU, MVT::i32, Expand);
351 setOperationAction(ISD::MULHS, MVT::i32, Expand);
352 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
353 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
355 setOperationAction(ISD::MUL, MVT::i64, Expand);
356 setOperationAction(ISD::MULHU, MVT::i32, Expand);
357 if (!Subtarget->hasV6Ops())
358 setOperationAction(ISD::MULHS, MVT::i32, Expand);
360 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
361 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
362 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
363 setOperationAction(ISD::SRL, MVT::i64, Custom);
364 setOperationAction(ISD::SRA, MVT::i64, Custom);
366 // ARM does not have ROTL.
367 setOperationAction(ISD::ROTL, MVT::i32, Expand);
368 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
369 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
370 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
371 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
373 // Only ARMv6 has BSWAP.
374 if (!Subtarget->hasV6Ops())
375 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
377 // These are expanded into libcalls.
378 if (!Subtarget->hasDivide()) {
379 // v7M has a hardware divider
380 setOperationAction(ISD::SDIV, MVT::i32, Expand);
381 setOperationAction(ISD::UDIV, MVT::i32, Expand);
383 setOperationAction(ISD::SREM, MVT::i32, Expand);
384 setOperationAction(ISD::UREM, MVT::i32, Expand);
385 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
386 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
388 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
389 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
390 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
391 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
392 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
394 setOperationAction(ISD::TRAP, MVT::Other, Legal);
396 // Use the default implementation.
397 setOperationAction(ISD::VASTART, MVT::Other, Custom);
398 setOperationAction(ISD::VAARG, MVT::Other, Expand);
399 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
400 setOperationAction(ISD::VAEND, MVT::Other, Expand);
401 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
402 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
403 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
404 // FIXME: Shouldn't need this, since no register is used, but the legalizer
405 // doesn't yet know how to not do that for SjLj.
406 setExceptionSelectorRegister(ARM::R0);
407 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
408 setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
410 // If the subtarget does not have extract instructions, sign_extend_inreg
411 // needs to be expanded. Extract is available in ARM mode on v6 and up,
412 // and on most Thumb2 implementations.
413 if ((!Subtarget->isThumb() && !Subtarget->hasV6Ops())
414 || (Subtarget->isThumb2() && !Subtarget->hasT2ExtractPack())) {
415 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
416 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
418 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
420 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
421 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
422 // iff target supports vfp2.
423 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
425 // We want to custom lower some of our intrinsics.
426 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
428 setOperationAction(ISD::SETCC, MVT::i32, Expand);
429 setOperationAction(ISD::SETCC, MVT::f32, Expand);
430 setOperationAction(ISD::SETCC, MVT::f64, Expand);
431 setOperationAction(ISD::SELECT, MVT::i32, Expand);
432 setOperationAction(ISD::SELECT, MVT::f32, Expand);
433 setOperationAction(ISD::SELECT, MVT::f64, Expand);
434 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
435 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
436 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
438 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
439 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
440 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
441 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
442 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
444 // We don't support sin/cos/fmod/copysign/pow
445 setOperationAction(ISD::FSIN, MVT::f64, Expand);
446 setOperationAction(ISD::FSIN, MVT::f32, Expand);
447 setOperationAction(ISD::FCOS, MVT::f32, Expand);
448 setOperationAction(ISD::FCOS, MVT::f64, Expand);
449 setOperationAction(ISD::FREM, MVT::f64, Expand);
450 setOperationAction(ISD::FREM, MVT::f32, Expand);
451 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
452 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
453 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
455 setOperationAction(ISD::FPOW, MVT::f64, Expand);
456 setOperationAction(ISD::FPOW, MVT::f32, Expand);
458 // Various VFP goodness
459 if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
460 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
461 if (Subtarget->hasVFP2()) {
462 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
463 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
464 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
465 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
467 // Special handling for half-precision FP.
468 if (!Subtarget->hasFP16()) {
469 setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
470 setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
474 // We have target-specific dag combine patterns for the following nodes:
475 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
476 setTargetDAGCombine(ISD::ADD);
477 setTargetDAGCombine(ISD::SUB);
478 setTargetDAGCombine(ISD::MUL);
480 setStackPointerRegisterToSaveRestore(ARM::SP);
482 if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
483 setSchedulingPreference(Sched::RegPressure);
485 setSchedulingPreference(Sched::Hybrid);
487 // FIXME: If-converter should use instruction latency to determine
488 // profitability rather than relying on fixed limits.
489 if (Subtarget->getCPUString() == "generic") {
490 // Generic (and overly aggressive) if-conversion limits.
491 setIfCvtBlockSizeLimit(10);
492 setIfCvtDupBlockSizeLimit(2);
493 } else if (Subtarget->hasV7Ops()) {
494 setIfCvtBlockSizeLimit(3);
495 setIfCvtDupBlockSizeLimit(1);
496 } else if (Subtarget->hasV6Ops()) {
497 setIfCvtBlockSizeLimit(2);
498 setIfCvtDupBlockSizeLimit(1);
500 setIfCvtBlockSizeLimit(3);
501 setIfCvtDupBlockSizeLimit(2);
504 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
505 // Do not enable CodePlacementOpt for now: it currently runs after the
506 // ARMConstantIslandPass and messes up branch relaxation and placement
507 // of constant islands.
508 // benefitFromCodePlacementOpt = true;
511 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
514 case ARMISD::Wrapper: return "ARMISD::Wrapper";
515 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
516 case ARMISD::CALL: return "ARMISD::CALL";
517 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
518 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
519 case ARMISD::tCALL: return "ARMISD::tCALL";
520 case ARMISD::BRCOND: return "ARMISD::BRCOND";
521 case ARMISD::BR_JT: return "ARMISD::BR_JT";
522 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
523 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
524 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
525 case ARMISD::CMP: return "ARMISD::CMP";
526 case ARMISD::CMPZ: return "ARMISD::CMPZ";
527 case ARMISD::CMPFP: return "ARMISD::CMPFP";
528 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
529 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
530 case ARMISD::CMOV: return "ARMISD::CMOV";
531 case ARMISD::CNEG: return "ARMISD::CNEG";
533 case ARMISD::RBIT: return "ARMISD::RBIT";
535 case ARMISD::FTOSI: return "ARMISD::FTOSI";
536 case ARMISD::FTOUI: return "ARMISD::FTOUI";
537 case ARMISD::SITOF: return "ARMISD::SITOF";
538 case ARMISD::UITOF: return "ARMISD::UITOF";
540 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
541 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
542 case ARMISD::RRX: return "ARMISD::RRX";
544 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
545 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
547 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
548 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
550 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
552 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
554 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
556 case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
557 case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
559 case ARMISD::VCEQ: return "ARMISD::VCEQ";
560 case ARMISD::VCGE: return "ARMISD::VCGE";
561 case ARMISD::VCGEU: return "ARMISD::VCGEU";
562 case ARMISD::VCGT: return "ARMISD::VCGT";
563 case ARMISD::VCGTU: return "ARMISD::VCGTU";
564 case ARMISD::VTST: return "ARMISD::VTST";
566 case ARMISD::VSHL: return "ARMISD::VSHL";
567 case ARMISD::VSHRs: return "ARMISD::VSHRs";
568 case ARMISD::VSHRu: return "ARMISD::VSHRu";
569 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
570 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
571 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
572 case ARMISD::VSHRN: return "ARMISD::VSHRN";
573 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
574 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
575 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
576 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
577 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
578 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
579 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
580 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
581 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
582 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
583 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
584 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
585 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
586 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
587 case ARMISD::VDUP: return "ARMISD::VDUP";
588 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
589 case ARMISD::VEXT: return "ARMISD::VEXT";
590 case ARMISD::VREV64: return "ARMISD::VREV64";
591 case ARMISD::VREV32: return "ARMISD::VREV32";
592 case ARMISD::VREV16: return "ARMISD::VREV16";
593 case ARMISD::VZIP: return "ARMISD::VZIP";
594 case ARMISD::VUZP: return "ARMISD::VUZP";
595 case ARMISD::VTRN: return "ARMISD::VTRN";
596 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
597 case ARMISD::FMAX: return "ARMISD::FMAX";
598 case ARMISD::FMIN: return "ARMISD::FMIN";
602 /// getRegClassFor - Return the register class that should be used for the
603 /// specified value type.
604 TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
605 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
606 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
607 // load / store 4 to 8 consecutive D registers.
608 if (Subtarget->hasNEON()) {
609 if (VT == MVT::v4i64)
610 return ARM::QQPRRegisterClass;
611 else if (VT == MVT::v8i64)
612 return ARM::QQQQPRRegisterClass;
614 return TargetLowering::getRegClassFor(VT);
617 /// getFunctionAlignment - Return the Log2 alignment of this function.
618 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
619 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
622 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
623 unsigned NumVals = N->getNumValues();
625 return Sched::RegPressure;
627 for (unsigned i = 0; i != NumVals; ++i) {
628 EVT VT = N->getValueType(i);
629 if (VT.isFloatingPoint() || VT.isVector())
630 return Sched::Latency;
633 if (!N->isMachineOpcode())
634 return Sched::RegPressure;
636 // Load are scheduled for latency even if there instruction itinerary
638 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
639 const TargetInstrDesc &TID = TII->get(N->getMachineOpcode());
641 return Sched::Latency;
643 const InstrItineraryData &Itins = getTargetMachine().getInstrItineraryData();
644 if (!Itins.isEmpty() && Itins.getStageLatency(TID.getSchedClass()) > 2)
645 return Sched::Latency;
646 return Sched::RegPressure;
649 //===----------------------------------------------------------------------===//
651 //===----------------------------------------------------------------------===//
653 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
654 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
656 default: llvm_unreachable("Unknown condition code!");
657 case ISD::SETNE: return ARMCC::NE;
658 case ISD::SETEQ: return ARMCC::EQ;
659 case ISD::SETGT: return ARMCC::GT;
660 case ISD::SETGE: return ARMCC::GE;
661 case ISD::SETLT: return ARMCC::LT;
662 case ISD::SETLE: return ARMCC::LE;
663 case ISD::SETUGT: return ARMCC::HI;
664 case ISD::SETUGE: return ARMCC::HS;
665 case ISD::SETULT: return ARMCC::LO;
666 case ISD::SETULE: return ARMCC::LS;
670 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
671 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
672 ARMCC::CondCodes &CondCode2) {
673 CondCode2 = ARMCC::AL;
675 default: llvm_unreachable("Unknown FP condition!");
677 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
679 case ISD::SETOGT: CondCode = ARMCC::GT; break;
681 case ISD::SETOGE: CondCode = ARMCC::GE; break;
682 case ISD::SETOLT: CondCode = ARMCC::MI; break;
683 case ISD::SETOLE: CondCode = ARMCC::LS; break;
684 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
685 case ISD::SETO: CondCode = ARMCC::VC; break;
686 case ISD::SETUO: CondCode = ARMCC::VS; break;
687 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
688 case ISD::SETUGT: CondCode = ARMCC::HI; break;
689 case ISD::SETUGE: CondCode = ARMCC::PL; break;
691 case ISD::SETULT: CondCode = ARMCC::LT; break;
693 case ISD::SETULE: CondCode = ARMCC::LE; break;
695 case ISD::SETUNE: CondCode = ARMCC::NE; break;
699 //===----------------------------------------------------------------------===//
700 // Calling Convention Implementation
701 //===----------------------------------------------------------------------===//
703 #include "ARMGenCallingConv.inc"
705 // APCS f64 is in register pairs, possibly split to stack
706 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
707 CCValAssign::LocInfo &LocInfo,
708 CCState &State, bool CanFail) {
709 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
711 // Try to get the first register.
712 if (unsigned Reg = State.AllocateReg(RegList, 4))
713 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
715 // For the 2nd half of a v2f64, do not fail.
719 // Put the whole thing on the stack.
720 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
721 State.AllocateStack(8, 4),
726 // Try to get the second register.
727 if (unsigned Reg = State.AllocateReg(RegList, 4))
728 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
730 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
731 State.AllocateStack(4, 4),
736 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
737 CCValAssign::LocInfo &LocInfo,
738 ISD::ArgFlagsTy &ArgFlags,
740 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
742 if (LocVT == MVT::v2f64 &&
743 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
745 return true; // we handled it
748 // AAPCS f64 is in aligned register pairs
749 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
750 CCValAssign::LocInfo &LocInfo,
751 CCState &State, bool CanFail) {
752 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
753 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
755 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
757 // For the 2nd half of a v2f64, do not just fail.
761 // Put the whole thing on the stack.
762 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
763 State.AllocateStack(8, 8),
769 for (i = 0; i < 2; ++i)
770 if (HiRegList[i] == Reg)
773 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
774 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
779 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
780 CCValAssign::LocInfo &LocInfo,
781 ISD::ArgFlagsTy &ArgFlags,
783 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
785 if (LocVT == MVT::v2f64 &&
786 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
788 return true; // we handled it
791 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
792 CCValAssign::LocInfo &LocInfo, CCState &State) {
793 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
794 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
796 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
798 return false; // we didn't handle it
801 for (i = 0; i < 2; ++i)
802 if (HiRegList[i] == Reg)
805 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
806 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
811 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
812 CCValAssign::LocInfo &LocInfo,
813 ISD::ArgFlagsTy &ArgFlags,
815 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
817 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
819 return true; // we handled it
822 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
823 CCValAssign::LocInfo &LocInfo,
824 ISD::ArgFlagsTy &ArgFlags,
826 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
830 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
831 /// given CallingConvention value.
832 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
834 bool isVarArg) const {
837 llvm_unreachable("Unsupported calling convention");
839 case CallingConv::Fast:
840 // Use target triple & subtarget features to do actual dispatch.
841 if (Subtarget->isAAPCS_ABI()) {
842 if (Subtarget->hasVFP2() &&
843 FloatABIType == FloatABI::Hard && !isVarArg)
844 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
846 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
848 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
849 case CallingConv::ARM_AAPCS_VFP:
850 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
851 case CallingConv::ARM_AAPCS:
852 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
853 case CallingConv::ARM_APCS:
854 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
858 /// LowerCallResult - Lower the result values of a call into the
859 /// appropriate copies out of appropriate physical registers.
861 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
862 CallingConv::ID CallConv, bool isVarArg,
863 const SmallVectorImpl<ISD::InputArg> &Ins,
864 DebugLoc dl, SelectionDAG &DAG,
865 SmallVectorImpl<SDValue> &InVals) const {
867 // Assign locations to each value returned by this call.
868 SmallVector<CCValAssign, 16> RVLocs;
869 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
870 RVLocs, *DAG.getContext());
871 CCInfo.AnalyzeCallResult(Ins,
872 CCAssignFnForNode(CallConv, /* Return*/ true,
875 // Copy all of the result registers out of their specified physreg.
876 for (unsigned i = 0; i != RVLocs.size(); ++i) {
877 CCValAssign VA = RVLocs[i];
880 if (VA.needsCustom()) {
881 // Handle f64 or half of a v2f64.
882 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
884 Chain = Lo.getValue(1);
885 InFlag = Lo.getValue(2);
886 VA = RVLocs[++i]; // skip ahead to next loc
887 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
889 Chain = Hi.getValue(1);
890 InFlag = Hi.getValue(2);
891 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
893 if (VA.getLocVT() == MVT::v2f64) {
894 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
895 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
896 DAG.getConstant(0, MVT::i32));
898 VA = RVLocs[++i]; // skip ahead to next loc
899 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
900 Chain = Lo.getValue(1);
901 InFlag = Lo.getValue(2);
902 VA = RVLocs[++i]; // skip ahead to next loc
903 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
904 Chain = Hi.getValue(1);
905 InFlag = Hi.getValue(2);
906 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
907 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
908 DAG.getConstant(1, MVT::i32));
911 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
913 Chain = Val.getValue(1);
914 InFlag = Val.getValue(2);
917 switch (VA.getLocInfo()) {
918 default: llvm_unreachable("Unknown loc info!");
919 case CCValAssign::Full: break;
920 case CCValAssign::BCvt:
921 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
925 InVals.push_back(Val);
931 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
932 /// by "Src" to address "Dst" of size "Size". Alignment information is
933 /// specified by the specific parameter attribute. The copy will be passed as
934 /// a byval function parameter.
935 /// Sometimes what we are copying is the end of a larger object, the part that
936 /// does not fit in registers.
938 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
939 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
941 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
942 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
943 /*isVolatile=*/false, /*AlwaysInline=*/false,
947 /// LowerMemOpCallTo - Store the argument to the stack.
949 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
950 SDValue StackPtr, SDValue Arg,
951 DebugLoc dl, SelectionDAG &DAG,
952 const CCValAssign &VA,
953 ISD::ArgFlagsTy Flags) const {
954 unsigned LocMemOffset = VA.getLocMemOffset();
955 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
956 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
957 if (Flags.isByVal()) {
958 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
960 return DAG.getStore(Chain, dl, Arg, PtrOff,
961 PseudoSourceValue::getStack(), LocMemOffset,
965 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
966 SDValue Chain, SDValue &Arg,
967 RegsToPassVector &RegsToPass,
968 CCValAssign &VA, CCValAssign &NextVA,
970 SmallVector<SDValue, 8> &MemOpChains,
971 ISD::ArgFlagsTy Flags) const {
973 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
974 DAG.getVTList(MVT::i32, MVT::i32), Arg);
975 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
977 if (NextVA.isRegLoc())
978 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
980 assert(NextVA.isMemLoc());
981 if (StackPtr.getNode() == 0)
982 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
984 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
990 /// LowerCall - Lowering a call into a callseq_start <-
991 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
994 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
995 CallingConv::ID CallConv, bool isVarArg,
997 const SmallVectorImpl<ISD::OutputArg> &Outs,
998 const SmallVectorImpl<ISD::InputArg> &Ins,
999 DebugLoc dl, SelectionDAG &DAG,
1000 SmallVectorImpl<SDValue> &InVals) const {
1001 MachineFunction &MF = DAG.getMachineFunction();
1002 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1003 bool IsSibCall = false;
1004 // Temporarily disable tail calls so things don't break.
1005 if (!EnableARMTailCalls)
1008 // Check if it's really possible to do a tail call.
1009 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1010 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1012 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1013 // detected sibcalls.
1020 // Analyze operands of the call, assigning locations to each operand.
1021 SmallVector<CCValAssign, 16> ArgLocs;
1022 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1024 CCInfo.AnalyzeCallOperands(Outs,
1025 CCAssignFnForNode(CallConv, /* Return*/ false,
1028 // Get a count of how many bytes are to be pushed on the stack.
1029 unsigned NumBytes = CCInfo.getNextStackOffset();
1031 // For tail calls, memory operands are available in our caller's stack.
1035 // Adjust the stack pointer for the new arguments...
1036 // These operations are automatically eliminated by the prolog/epilog pass
1038 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1040 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1042 RegsToPassVector RegsToPass;
1043 SmallVector<SDValue, 8> MemOpChains;
1045 // Walk the register/memloc assignments, inserting copies/loads. In the case
1046 // of tail call optimization, arguments are handled later.
1047 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1049 ++i, ++realArgIdx) {
1050 CCValAssign &VA = ArgLocs[i];
1051 SDValue Arg = Outs[realArgIdx].Val;
1052 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1054 // Promote the value if needed.
1055 switch (VA.getLocInfo()) {
1056 default: llvm_unreachable("Unknown loc info!");
1057 case CCValAssign::Full: break;
1058 case CCValAssign::SExt:
1059 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1061 case CCValAssign::ZExt:
1062 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1064 case CCValAssign::AExt:
1065 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1067 case CCValAssign::BCvt:
1068 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1072 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1073 if (VA.needsCustom()) {
1074 if (VA.getLocVT() == MVT::v2f64) {
1075 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1076 DAG.getConstant(0, MVT::i32));
1077 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1078 DAG.getConstant(1, MVT::i32));
1080 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1081 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1083 VA = ArgLocs[++i]; // skip ahead to next loc
1084 if (VA.isRegLoc()) {
1085 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1086 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1088 assert(VA.isMemLoc());
1090 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1091 dl, DAG, VA, Flags));
1094 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1095 StackPtr, MemOpChains, Flags);
1097 } else if (VA.isRegLoc()) {
1098 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1100 assert(VA.isMemLoc());
1102 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1103 dl, DAG, VA, Flags));
1107 if (!MemOpChains.empty())
1108 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1109 &MemOpChains[0], MemOpChains.size());
1111 // Build a sequence of copy-to-reg nodes chained together with token chain
1112 // and flag operands which copy the outgoing args into the appropriate regs.
1114 // Tail call byval lowering might overwrite argument registers so in case of
1115 // tail call optimization the copies to registers are lowered later.
1117 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1118 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1119 RegsToPass[i].second, InFlag);
1120 InFlag = Chain.getValue(1);
1123 // For tail calls lower the arguments to the 'real' stack slot.
1125 // Force all the incoming stack arguments to be loaded from the stack
1126 // before any new outgoing arguments are stored to the stack, because the
1127 // outgoing stack slots may alias the incoming argument stack slots, and
1128 // the alias isn't otherwise explicit. This is slightly more conservative
1129 // than necessary, because it means that each store effectively depends
1130 // on every argument instead of just those arguments it would clobber.
1132 // Do not flag preceeding copytoreg stuff together with the following stuff.
1134 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1135 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1136 RegsToPass[i].second, InFlag);
1137 InFlag = Chain.getValue(1);
1142 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1143 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1144 // node so that legalize doesn't hack it.
1145 bool isDirect = false;
1146 bool isARMFunc = false;
1147 bool isLocalARMFunc = false;
1148 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1150 if (EnableARMLongCalls) {
1151 assert (getTargetMachine().getRelocationModel() == Reloc::Static
1152 && "long-calls with non-static relocation model!");
1153 // Handle a global address or an external symbol. If it's not one of
1154 // those, the target's already in a register, so we don't need to do
1156 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1157 const GlobalValue *GV = G->getGlobal();
1158 // Create a constant pool entry for the callee address
1159 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1160 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1163 // Get the address of the callee into a register
1164 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1165 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1166 Callee = DAG.getLoad(getPointerTy(), dl,
1167 DAG.getEntryNode(), CPAddr,
1168 PseudoSourceValue::getConstantPool(), 0,
1170 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1171 const char *Sym = S->getSymbol();
1173 // Create a constant pool entry for the callee address
1174 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1175 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1176 Sym, ARMPCLabelIndex, 0);
1177 // Get the address of the callee into a register
1178 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1179 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1180 Callee = DAG.getLoad(getPointerTy(), dl,
1181 DAG.getEntryNode(), CPAddr,
1182 PseudoSourceValue::getConstantPool(), 0,
1185 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1186 const GlobalValue *GV = G->getGlobal();
1188 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1189 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1190 getTargetMachine().getRelocationModel() != Reloc::Static;
1191 isARMFunc = !Subtarget->isThumb() || isStub;
1192 // ARM call to a local ARM function is predicable.
1193 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1194 // tBX takes a register source operand.
1195 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1196 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1197 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1200 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1201 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1202 Callee = DAG.getLoad(getPointerTy(), dl,
1203 DAG.getEntryNode(), CPAddr,
1204 PseudoSourceValue::getConstantPool(), 0,
1206 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1207 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1208 getPointerTy(), Callee, PICLabel);
1210 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1211 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1213 bool isStub = Subtarget->isTargetDarwin() &&
1214 getTargetMachine().getRelocationModel() != Reloc::Static;
1215 isARMFunc = !Subtarget->isThumb() || isStub;
1216 // tBX takes a register source operand.
1217 const char *Sym = S->getSymbol();
1218 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1219 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1220 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1221 Sym, ARMPCLabelIndex, 4);
1222 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1223 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1224 Callee = DAG.getLoad(getPointerTy(), dl,
1225 DAG.getEntryNode(), CPAddr,
1226 PseudoSourceValue::getConstantPool(), 0,
1228 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1229 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1230 getPointerTy(), Callee, PICLabel);
1232 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1235 // FIXME: handle tail calls differently.
1237 if (Subtarget->isThumb()) {
1238 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1239 CallOpc = ARMISD::CALL_NOLINK;
1241 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1243 CallOpc = (isDirect || Subtarget->hasV5TOps())
1244 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1245 : ARMISD::CALL_NOLINK;
1247 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1248 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1249 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1250 InFlag = Chain.getValue(1);
1253 std::vector<SDValue> Ops;
1254 Ops.push_back(Chain);
1255 Ops.push_back(Callee);
1257 // Add argument registers to the end of the list so that they are known live
1259 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1260 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1261 RegsToPass[i].second.getValueType()));
1263 if (InFlag.getNode())
1264 Ops.push_back(InFlag);
1266 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1268 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1270 // Returns a chain and a flag for retval copy to use.
1271 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1272 InFlag = Chain.getValue(1);
1274 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1275 DAG.getIntPtrConstant(0, true), InFlag);
1277 InFlag = Chain.getValue(1);
1279 // Handle result values, copying them out of physregs into vregs that we
1281 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1285 /// MatchingStackOffset - Return true if the given stack call argument is
1286 /// already available in the same position (relatively) of the caller's
1287 /// incoming argument stack.
1289 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1290 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1291 const ARMInstrInfo *TII) {
1292 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1294 if (Arg.getOpcode() == ISD::CopyFromReg) {
1295 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1296 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
1298 MachineInstr *Def = MRI->getVRegDef(VR);
1301 if (!Flags.isByVal()) {
1302 if (!TII->isLoadFromStackSlot(Def, FI))
1305 // unsigned Opcode = Def->getOpcode();
1306 // if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
1307 // Def->getOperand(1).isFI()) {
1308 // FI = Def->getOperand(1).getIndex();
1309 // Bytes = Flags.getByValSize();
1313 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1314 if (Flags.isByVal())
1315 // ByVal argument is passed in as a pointer but it's now being
1316 // dereferenced. e.g.
1317 // define @foo(%struct.X* %A) {
1318 // tail call @bar(%struct.X* byval %A)
1321 SDValue Ptr = Ld->getBasePtr();
1322 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1325 FI = FINode->getIndex();
1329 assert(FI != INT_MAX);
1330 if (!MFI->isFixedObjectIndex(FI))
1332 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1335 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1336 /// for tail call optimization. Targets which want to do tail call
1337 /// optimization should implement this function.
1339 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1340 CallingConv::ID CalleeCC,
1342 bool isCalleeStructRet,
1343 bool isCallerStructRet,
1344 const SmallVectorImpl<ISD::OutputArg> &Outs,
1345 const SmallVectorImpl<ISD::InputArg> &Ins,
1346 SelectionDAG& DAG) const {
1348 const Function *CallerF = DAG.getMachineFunction().getFunction();
1349 CallingConv::ID CallerCC = CallerF->getCallingConv();
1350 bool CCMatch = CallerCC == CalleeCC;
1352 // Look for obvious safe cases to perform tail call optimization that do not
1353 // require ABI changes. This is what gcc calls sibcall.
1355 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
1356 // emit a special epilogue.
1357 // Not sure yet if this is true on ARM.
1358 //?? if (RegInfo->needsStackRealignment(MF))
1361 // Do not sibcall optimize vararg calls unless the call site is not passing
1363 if (isVarArg && !Outs.empty())
1366 // Also avoid sibcall optimization if either caller or callee uses struct
1367 // return semantics.
1368 if (isCalleeStructRet || isCallerStructRet)
1371 // If the calling conventions do not match, then we'd better make sure the
1372 // results are returned in the same way as what the caller expects.
1374 SmallVector<CCValAssign, 16> RVLocs1;
1375 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
1376 RVLocs1, *DAG.getContext());
1377 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1379 SmallVector<CCValAssign, 16> RVLocs2;
1380 CCState CCInfo2(CallerCC, false, getTargetMachine(),
1381 RVLocs2, *DAG.getContext());
1382 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1384 if (RVLocs1.size() != RVLocs2.size())
1386 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1387 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1389 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1391 if (RVLocs1[i].isRegLoc()) {
1392 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1395 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1401 // If the callee takes no arguments then go on to check the results of the
1403 if (!Outs.empty()) {
1404 // Check if stack adjustment is needed. For now, do not do this if any
1405 // argument is passed on the stack.
1406 SmallVector<CCValAssign, 16> ArgLocs;
1407 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
1408 ArgLocs, *DAG.getContext());
1409 CCInfo.AnalyzeCallOperands(Outs,
1410 CCAssignFnForNode(CalleeCC, false, isVarArg));
1411 if (CCInfo.getNextStackOffset()) {
1412 MachineFunction &MF = DAG.getMachineFunction();
1414 // Check if the arguments are already laid out in the right way as
1415 // the caller's fixed stack objects.
1416 MachineFrameInfo *MFI = MF.getFrameInfo();
1417 const MachineRegisterInfo *MRI = &MF.getRegInfo();
1418 const ARMInstrInfo *TII =
1419 ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1420 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1422 ++i, ++realArgIdx) {
1423 CCValAssign &VA = ArgLocs[i];
1424 EVT RegVT = VA.getLocVT();
1425 SDValue Arg = Outs[realArgIdx].Val;
1426 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1427 if (VA.getLocInfo() == CCValAssign::Indirect)
1429 if (VA.needsCustom()) {
1430 // f64 and vector types are split into multiple registers or
1431 // register/stack-slot combinations. The types will not match
1432 // the registers; give up on memory f64 refs until we figure
1433 // out what to do about this.
1436 if (!ArgLocs[++i].isRegLoc())
1438 if (RegVT == MVT::v2f64) {
1439 if (!ArgLocs[++i].isRegLoc())
1441 if (!ArgLocs[++i].isRegLoc())
1444 } else if (!VA.isRegLoc()) {
1445 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1457 ARMTargetLowering::LowerReturn(SDValue Chain,
1458 CallingConv::ID CallConv, bool isVarArg,
1459 const SmallVectorImpl<ISD::OutputArg> &Outs,
1460 DebugLoc dl, SelectionDAG &DAG) const {
1462 // CCValAssign - represent the assignment of the return value to a location.
1463 SmallVector<CCValAssign, 16> RVLocs;
1465 // CCState - Info about the registers and stack slots.
1466 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1469 // Analyze outgoing return values.
1470 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1473 // If this is the first return lowered for this function, add
1474 // the regs to the liveout set for the function.
1475 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1476 for (unsigned i = 0; i != RVLocs.size(); ++i)
1477 if (RVLocs[i].isRegLoc())
1478 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1483 // Copy the result values into the output registers.
1484 for (unsigned i = 0, realRVLocIdx = 0;
1486 ++i, ++realRVLocIdx) {
1487 CCValAssign &VA = RVLocs[i];
1488 assert(VA.isRegLoc() && "Can only return in registers!");
1490 SDValue Arg = Outs[realRVLocIdx].Val;
1492 switch (VA.getLocInfo()) {
1493 default: llvm_unreachable("Unknown loc info!");
1494 case CCValAssign::Full: break;
1495 case CCValAssign::BCvt:
1496 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1500 if (VA.needsCustom()) {
1501 if (VA.getLocVT() == MVT::v2f64) {
1502 // Extract the first half and return it in two registers.
1503 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1504 DAG.getConstant(0, MVT::i32));
1505 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1506 DAG.getVTList(MVT::i32, MVT::i32), Half);
1508 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1509 Flag = Chain.getValue(1);
1510 VA = RVLocs[++i]; // skip ahead to next loc
1511 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1512 HalfGPRs.getValue(1), Flag);
1513 Flag = Chain.getValue(1);
1514 VA = RVLocs[++i]; // skip ahead to next loc
1516 // Extract the 2nd half and fall through to handle it as an f64 value.
1517 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1518 DAG.getConstant(1, MVT::i32));
1520 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1522 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1523 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1524 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1525 Flag = Chain.getValue(1);
1526 VA = RVLocs[++i]; // skip ahead to next loc
1527 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1530 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1532 // Guarantee that all emitted copies are
1533 // stuck together, avoiding something bad.
1534 Flag = Chain.getValue(1);
1539 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1541 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1546 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1547 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1548 // one of the above mentioned nodes. It has to be wrapped because otherwise
1549 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1550 // be used to form addressing mode. These wrapped nodes will be selected
1552 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1553 EVT PtrVT = Op.getValueType();
1554 // FIXME there is no actual debug info here
1555 DebugLoc dl = Op.getDebugLoc();
1556 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1558 if (CP->isMachineConstantPoolEntry())
1559 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1560 CP->getAlignment());
1562 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1563 CP->getAlignment());
1564 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1567 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1568 SelectionDAG &DAG) const {
1569 MachineFunction &MF = DAG.getMachineFunction();
1570 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1571 unsigned ARMPCLabelIndex = 0;
1572 DebugLoc DL = Op.getDebugLoc();
1573 EVT PtrVT = getPointerTy();
1574 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1575 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1577 if (RelocM == Reloc::Static) {
1578 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1580 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1581 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1582 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1583 ARMCP::CPBlockAddress,
1585 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1587 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1588 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1589 PseudoSourceValue::getConstantPool(), 0,
1591 if (RelocM == Reloc::Static)
1593 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1594 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1597 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1599 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1600 SelectionDAG &DAG) const {
1601 DebugLoc dl = GA->getDebugLoc();
1602 EVT PtrVT = getPointerTy();
1603 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1604 MachineFunction &MF = DAG.getMachineFunction();
1605 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1606 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1607 ARMConstantPoolValue *CPV =
1608 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1609 ARMCP::CPValue, PCAdj, "tlsgd", true);
1610 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1611 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1612 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1613 PseudoSourceValue::getConstantPool(), 0,
1615 SDValue Chain = Argument.getValue(1);
1617 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1618 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1620 // call __tls_get_addr.
1623 Entry.Node = Argument;
1624 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1625 Args.push_back(Entry);
1626 // FIXME: is there useful debug info available here?
1627 std::pair<SDValue, SDValue> CallResult =
1628 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1629 false, false, false, false,
1630 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1631 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1632 return CallResult.first;
1635 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1636 // "local exec" model.
1638 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1639 SelectionDAG &DAG) const {
1640 const GlobalValue *GV = GA->getGlobal();
1641 DebugLoc dl = GA->getDebugLoc();
1643 SDValue Chain = DAG.getEntryNode();
1644 EVT PtrVT = getPointerTy();
1645 // Get the Thread Pointer
1646 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1648 if (GV->isDeclaration()) {
1649 MachineFunction &MF = DAG.getMachineFunction();
1650 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1651 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1652 // Initial exec model.
1653 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1654 ARMConstantPoolValue *CPV =
1655 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1656 ARMCP::CPValue, PCAdj, "gottpoff", true);
1657 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1658 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1659 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1660 PseudoSourceValue::getConstantPool(), 0,
1662 Chain = Offset.getValue(1);
1664 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1665 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1667 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1668 PseudoSourceValue::getConstantPool(), 0,
1672 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1673 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1674 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1675 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1676 PseudoSourceValue::getConstantPool(), 0,
1680 // The address of the thread local variable is the add of the thread
1681 // pointer with the offset of the variable.
1682 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1686 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
1687 // TODO: implement the "local dynamic" model
1688 assert(Subtarget->isTargetELF() &&
1689 "TLS not implemented for non-ELF targets");
1690 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1691 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1692 // otherwise use the "Local Exec" TLS Model
1693 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1694 return LowerToTLSGeneralDynamicModel(GA, DAG);
1696 return LowerToTLSExecModels(GA, DAG);
1699 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1700 SelectionDAG &DAG) const {
1701 EVT PtrVT = getPointerTy();
1702 DebugLoc dl = Op.getDebugLoc();
1703 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1704 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1705 if (RelocM == Reloc::PIC_) {
1706 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1707 ARMConstantPoolValue *CPV =
1708 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1709 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1710 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1711 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1713 PseudoSourceValue::getConstantPool(), 0,
1715 SDValue Chain = Result.getValue(1);
1716 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1717 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1719 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1720 PseudoSourceValue::getGOT(), 0,
1724 // If we have T2 ops, we can materialize the address directly via movt/movw
1725 // pair. This is always cheaper.
1726 if (Subtarget->useMovt()) {
1727 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1728 DAG.getTargetGlobalAddress(GV, PtrVT));
1730 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1731 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1732 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1733 PseudoSourceValue::getConstantPool(), 0,
1739 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1740 SelectionDAG &DAG) const {
1741 MachineFunction &MF = DAG.getMachineFunction();
1742 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1743 unsigned ARMPCLabelIndex = 0;
1744 EVT PtrVT = getPointerTy();
1745 DebugLoc dl = Op.getDebugLoc();
1746 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1747 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1749 if (RelocM == Reloc::Static)
1750 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1752 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1753 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1754 ARMConstantPoolValue *CPV =
1755 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1756 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1758 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1760 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1761 PseudoSourceValue::getConstantPool(), 0,
1763 SDValue Chain = Result.getValue(1);
1765 if (RelocM == Reloc::PIC_) {
1766 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1767 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1770 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1771 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1772 PseudoSourceValue::getGOT(), 0,
1778 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1779 SelectionDAG &DAG) const {
1780 assert(Subtarget->isTargetELF() &&
1781 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1782 MachineFunction &MF = DAG.getMachineFunction();
1783 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1784 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1785 EVT PtrVT = getPointerTy();
1786 DebugLoc dl = Op.getDebugLoc();
1787 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1788 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1789 "_GLOBAL_OFFSET_TABLE_",
1790 ARMPCLabelIndex, PCAdj);
1791 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1792 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1793 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1794 PseudoSourceValue::getConstantPool(), 0,
1796 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1797 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1801 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
1802 DebugLoc dl = Op.getDebugLoc();
1803 SDValue Val = DAG.getConstant(0, MVT::i32);
1804 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(0),
1805 Op.getOperand(1), Val);
1809 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
1810 DebugLoc dl = Op.getDebugLoc();
1811 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
1812 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
1816 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
1817 const ARMSubtarget *Subtarget) const {
1818 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1819 DebugLoc dl = Op.getDebugLoc();
1821 default: return SDValue(); // Don't custom lower most intrinsics.
1822 case Intrinsic::arm_thread_pointer: {
1823 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1824 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1826 case Intrinsic::eh_sjlj_lsda: {
1827 MachineFunction &MF = DAG.getMachineFunction();
1828 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1829 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1830 EVT PtrVT = getPointerTy();
1831 DebugLoc dl = Op.getDebugLoc();
1832 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1834 unsigned PCAdj = (RelocM != Reloc::PIC_)
1835 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1836 ARMConstantPoolValue *CPV =
1837 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1838 ARMCP::CPLSDA, PCAdj);
1839 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1840 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1842 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1843 PseudoSourceValue::getConstantPool(), 0,
1845 SDValue Chain = Result.getValue(1);
1847 if (RelocM == Reloc::PIC_) {
1848 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1849 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1856 static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
1857 const ARMSubtarget *Subtarget) {
1858 DebugLoc dl = Op.getDebugLoc();
1859 SDValue Op5 = Op.getOperand(5);
1861 unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
1862 if (isDeviceBarrier) {
1863 if (Subtarget->hasV7Ops())
1864 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
1866 Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
1867 DAG.getConstant(0, MVT::i32));
1869 if (Subtarget->hasV7Ops())
1870 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
1872 Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
1873 DAG.getConstant(0, MVT::i32));
1878 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
1879 MachineFunction &MF = DAG.getMachineFunction();
1880 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
1882 // vastart just stores the address of the VarArgsFrameIndex slot into the
1883 // memory location argument.
1884 DebugLoc dl = Op.getDebugLoc();
1885 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1886 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1887 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1888 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
1893 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1894 SelectionDAG &DAG) const {
1895 SDNode *Node = Op.getNode();
1896 DebugLoc dl = Node->getDebugLoc();
1897 EVT VT = Node->getValueType(0);
1898 SDValue Chain = Op.getOperand(0);
1899 SDValue Size = Op.getOperand(1);
1900 SDValue Align = Op.getOperand(2);
1902 // Chain the dynamic stack allocation so that it doesn't modify the stack
1903 // pointer when other instructions are using the stack.
1904 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1906 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1907 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1908 if (AlignVal > StackAlign)
1909 // Do this now since selection pass cannot introduce new target
1910 // independent node.
1911 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1913 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1914 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1915 // do even more horrible hack later.
1916 MachineFunction &MF = DAG.getMachineFunction();
1917 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1918 if (AFI->isThumb1OnlyFunction()) {
1920 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1922 uint32_t Val = C->getZExtValue();
1923 if (Val <= 508 && ((Val & 3) == 0))
1927 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1930 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1931 SDValue Ops1[] = { Chain, Size, Align };
1932 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1933 Chain = Res.getValue(1);
1934 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1935 DAG.getIntPtrConstant(0, true), SDValue());
1936 SDValue Ops2[] = { Res, Chain };
1937 return DAG.getMergeValues(Ops2, 2, dl);
1941 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1942 SDValue &Root, SelectionDAG &DAG,
1943 DebugLoc dl) const {
1944 MachineFunction &MF = DAG.getMachineFunction();
1945 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1947 TargetRegisterClass *RC;
1948 if (AFI->isThumb1OnlyFunction())
1949 RC = ARM::tGPRRegisterClass;
1951 RC = ARM::GPRRegisterClass;
1953 // Transform the arguments stored in physical registers into virtual ones.
1954 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1955 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1958 if (NextVA.isMemLoc()) {
1959 MachineFrameInfo *MFI = MF.getFrameInfo();
1960 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true, false);
1962 // Create load node to retrieve arguments from the stack.
1963 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1964 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1965 PseudoSourceValue::getFixedStack(FI), 0,
1968 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1969 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1972 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1976 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1977 CallingConv::ID CallConv, bool isVarArg,
1978 const SmallVectorImpl<ISD::InputArg>
1980 DebugLoc dl, SelectionDAG &DAG,
1981 SmallVectorImpl<SDValue> &InVals)
1984 MachineFunction &MF = DAG.getMachineFunction();
1985 MachineFrameInfo *MFI = MF.getFrameInfo();
1987 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1989 // Assign locations to all of the incoming arguments.
1990 SmallVector<CCValAssign, 16> ArgLocs;
1991 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1993 CCInfo.AnalyzeFormalArguments(Ins,
1994 CCAssignFnForNode(CallConv, /* Return*/ false,
1997 SmallVector<SDValue, 16> ArgValues;
1999 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2000 CCValAssign &VA = ArgLocs[i];
2002 // Arguments stored in registers.
2003 if (VA.isRegLoc()) {
2004 EVT RegVT = VA.getLocVT();
2007 if (VA.needsCustom()) {
2008 // f64 and vector types are split up into multiple registers or
2009 // combinations of registers and stack slots.
2010 if (VA.getLocVT() == MVT::v2f64) {
2011 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2013 VA = ArgLocs[++i]; // skip ahead to next loc
2015 if (VA.isMemLoc()) {
2016 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(),
2018 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2019 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2020 PseudoSourceValue::getFixedStack(FI), 0,
2023 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2026 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2027 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2028 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2029 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2030 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2032 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2035 TargetRegisterClass *RC;
2037 if (RegVT == MVT::f32)
2038 RC = ARM::SPRRegisterClass;
2039 else if (RegVT == MVT::f64)
2040 RC = ARM::DPRRegisterClass;
2041 else if (RegVT == MVT::v2f64)
2042 RC = ARM::QPRRegisterClass;
2043 else if (RegVT == MVT::i32)
2044 RC = (AFI->isThumb1OnlyFunction() ?
2045 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2047 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2049 // Transform the arguments in physical registers into virtual ones.
2050 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2051 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2054 // If this is an 8 or 16-bit value, it is really passed promoted
2055 // to 32 bits. Insert an assert[sz]ext to capture this, then
2056 // truncate to the right size.
2057 switch (VA.getLocInfo()) {
2058 default: llvm_unreachable("Unknown loc info!");
2059 case CCValAssign::Full: break;
2060 case CCValAssign::BCvt:
2061 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
2063 case CCValAssign::SExt:
2064 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2065 DAG.getValueType(VA.getValVT()));
2066 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2068 case CCValAssign::ZExt:
2069 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2070 DAG.getValueType(VA.getValVT()));
2071 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2075 InVals.push_back(ArgValue);
2077 } else { // VA.isRegLoc()
2080 assert(VA.isMemLoc());
2081 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2083 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
2084 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
2087 // Create load nodes to retrieve arguments from the stack.
2088 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2089 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2090 PseudoSourceValue::getFixedStack(FI), 0,
2097 static const unsigned GPRArgRegs[] = {
2098 ARM::R0, ARM::R1, ARM::R2, ARM::R3
2101 unsigned NumGPRs = CCInfo.getFirstUnallocated
2102 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2104 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
2105 unsigned VARegSize = (4 - NumGPRs) * 4;
2106 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2107 unsigned ArgOffset = CCInfo.getNextStackOffset();
2108 if (VARegSaveSize) {
2109 // If this function is vararg, store any remaining integer argument regs
2110 // to their spots on the stack so that they may be loaded by deferencing
2111 // the result of va_next.
2112 AFI->setVarArgsRegSaveSize(VARegSaveSize);
2113 AFI->setVarArgsFrameIndex(
2114 MFI->CreateFixedObject(VARegSaveSize,
2115 ArgOffset + VARegSaveSize - VARegSize,
2117 SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2120 SmallVector<SDValue, 4> MemOps;
2121 for (; NumGPRs < 4; ++NumGPRs) {
2122 TargetRegisterClass *RC;
2123 if (AFI->isThumb1OnlyFunction())
2124 RC = ARM::tGPRRegisterClass;
2126 RC = ARM::GPRRegisterClass;
2128 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
2129 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2131 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2132 PseudoSourceValue::getFixedStack(AFI->getVarArgsFrameIndex()),
2133 0, false, false, 0);
2134 MemOps.push_back(Store);
2135 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2136 DAG.getConstant(4, getPointerTy()));
2138 if (!MemOps.empty())
2139 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2140 &MemOps[0], MemOps.size());
2142 // This will point to the next argument passed via stack.
2143 AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset,
2150 /// isFloatingPointZero - Return true if this is +0.0.
2151 static bool isFloatingPointZero(SDValue Op) {
2152 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2153 return CFP->getValueAPF().isPosZero();
2154 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2155 // Maybe this has already been legalized into the constant pool?
2156 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2157 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2158 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2159 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2160 return CFP->getValueAPF().isPosZero();
2166 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2167 /// the given operands.
2169 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2170 SDValue &ARMCC, SelectionDAG &DAG,
2171 DebugLoc dl) const {
2172 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2173 unsigned C = RHSC->getZExtValue();
2174 if (!isLegalICmpImmediate(C)) {
2175 // Constant does not fit, try adjusting it by one?
2180 if (isLegalICmpImmediate(C-1)) {
2181 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2182 RHS = DAG.getConstant(C-1, MVT::i32);
2187 if (C > 0 && isLegalICmpImmediate(C-1)) {
2188 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2189 RHS = DAG.getConstant(C-1, MVT::i32);
2194 if (isLegalICmpImmediate(C+1)) {
2195 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2196 RHS = DAG.getConstant(C+1, MVT::i32);
2201 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
2202 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2203 RHS = DAG.getConstant(C+1, MVT::i32);
2210 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2211 ARMISD::NodeType CompareType;
2214 CompareType = ARMISD::CMP;
2219 CompareType = ARMISD::CMPZ;
2222 ARMCC = DAG.getConstant(CondCode, MVT::i32);
2223 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
2226 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2227 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2230 if (!isFloatingPointZero(RHS))
2231 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
2233 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
2234 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
2237 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2238 EVT VT = Op.getValueType();
2239 SDValue LHS = Op.getOperand(0);
2240 SDValue RHS = Op.getOperand(1);
2241 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2242 SDValue TrueVal = Op.getOperand(2);
2243 SDValue FalseVal = Op.getOperand(3);
2244 DebugLoc dl = Op.getDebugLoc();
2246 if (LHS.getValueType() == MVT::i32) {
2248 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2249 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2250 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
2253 ARMCC::CondCodes CondCode, CondCode2;
2254 FPCCToARMCC(CC, CondCode, CondCode2);
2256 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2257 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2258 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2259 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2261 if (CondCode2 != ARMCC::AL) {
2262 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
2263 // FIXME: Needs another CMP because flag can have but one use.
2264 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2265 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2266 Result, TrueVal, ARMCC2, CCR, Cmp2);
2271 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2272 SDValue Chain = Op.getOperand(0);
2273 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2274 SDValue LHS = Op.getOperand(2);
2275 SDValue RHS = Op.getOperand(3);
2276 SDValue Dest = Op.getOperand(4);
2277 DebugLoc dl = Op.getDebugLoc();
2279 if (LHS.getValueType() == MVT::i32) {
2281 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2282 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
2283 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2284 Chain, Dest, ARMCC, CCR,Cmp);
2287 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2288 ARMCC::CondCodes CondCode, CondCode2;
2289 FPCCToARMCC(CC, CondCode, CondCode2);
2291 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2292 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
2293 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2294 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
2295 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
2296 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2297 if (CondCode2 != ARMCC::AL) {
2298 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
2299 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
2300 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2305 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2306 SDValue Chain = Op.getOperand(0);
2307 SDValue Table = Op.getOperand(1);
2308 SDValue Index = Op.getOperand(2);
2309 DebugLoc dl = Op.getDebugLoc();
2311 EVT PTy = getPointerTy();
2312 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2313 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2314 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2315 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2316 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2317 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2318 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2319 if (Subtarget->isThumb2()) {
2320 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2321 // which does another jump to the destination. This also makes it easier
2322 // to translate it to TBB / TBH later.
2323 // FIXME: This might not work if the function is extremely large.
2324 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2325 Addr, Op.getOperand(2), JTI, UId);
2327 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2328 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
2329 PseudoSourceValue::getJumpTable(), 0,
2331 Chain = Addr.getValue(1);
2332 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
2333 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2335 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
2336 PseudoSourceValue::getJumpTable(), 0, false, false, 0);
2337 Chain = Addr.getValue(1);
2338 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
2342 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
2343 DebugLoc dl = Op.getDebugLoc();
2346 switch (Op.getOpcode()) {
2348 assert(0 && "Invalid opcode!");
2349 case ISD::FP_TO_SINT:
2350 Opc = ARMISD::FTOSI;
2352 case ISD::FP_TO_UINT:
2353 Opc = ARMISD::FTOUI;
2356 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
2357 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
2360 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2361 EVT VT = Op.getValueType();
2362 DebugLoc dl = Op.getDebugLoc();
2365 switch (Op.getOpcode()) {
2367 assert(0 && "Invalid opcode!");
2368 case ISD::SINT_TO_FP:
2369 Opc = ARMISD::SITOF;
2371 case ISD::UINT_TO_FP:
2372 Opc = ARMISD::UITOF;
2376 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
2377 return DAG.getNode(Opc, dl, VT, Op);
2380 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
2381 // Implement fcopysign with a fabs and a conditional fneg.
2382 SDValue Tmp0 = Op.getOperand(0);
2383 SDValue Tmp1 = Op.getOperand(1);
2384 DebugLoc dl = Op.getDebugLoc();
2385 EVT VT = Op.getValueType();
2386 EVT SrcVT = Tmp1.getValueType();
2387 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
2388 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
2389 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
2390 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2391 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
2394 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2395 MachineFunction &MF = DAG.getMachineFunction();
2396 MachineFrameInfo *MFI = MF.getFrameInfo();
2397 MFI->setReturnAddressIsTaken(true);
2399 EVT VT = Op.getValueType();
2400 DebugLoc dl = Op.getDebugLoc();
2401 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2403 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2404 SDValue Offset = DAG.getConstant(4, MVT::i32);
2405 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2406 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2407 NULL, 0, false, false, 0);
2410 // Return LR, which contains the return address. Mark it an implicit live-in.
2411 unsigned Reg = MF.addLiveIn(ARM::LR, ARM::GPRRegisterClass);
2412 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
2415 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2416 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2417 MFI->setFrameAddressIsTaken(true);
2419 EVT VT = Op.getValueType();
2420 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
2421 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2422 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
2423 ? ARM::R7 : ARM::R11;
2424 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2426 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
2431 /// ExpandBIT_CONVERT - If the target supports VFP, this function is called to
2432 /// expand a bit convert where either the source or destination type is i64 to
2433 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
2434 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
2435 /// vectors), since the legalizer won't know what to do with that.
2436 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2437 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2438 DebugLoc dl = N->getDebugLoc();
2439 SDValue Op = N->getOperand(0);
2441 // This function is only supposed to be called for i64 types, either as the
2442 // source or destination of the bit convert.
2443 EVT SrcVT = Op.getValueType();
2444 EVT DstVT = N->getValueType(0);
2445 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
2446 "ExpandBIT_CONVERT called for non-i64 type");
2448 // Turn i64->f64 into VMOVDRR.
2449 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
2450 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2451 DAG.getConstant(0, MVT::i32));
2452 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2453 DAG.getConstant(1, MVT::i32));
2454 return DAG.getNode(ISD::BIT_CONVERT, dl, DstVT,
2455 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
2458 // Turn f64->i64 into VMOVRRD.
2459 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
2460 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2461 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2462 // Merge the pieces into a single i64 value.
2463 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2469 /// getZeroVector - Returns a vector of specified type with all zero elements.
2471 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2472 assert(VT.isVector() && "Expected a vector type");
2474 // Zero vectors are used to represent vector negation and in those cases
2475 // will be implemented with the NEON VNEG instruction. However, VNEG does
2476 // not support i64 elements, so sometimes the zero vectors will need to be
2477 // explicitly constructed. For those cases, and potentially other uses in
2478 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2479 // to their dest type. This ensures they get CSE'd.
2481 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2482 SmallVector<SDValue, 8> Ops;
2485 if (VT.getSizeInBits() == 64) {
2486 Ops.assign(8, Cst); TVT = MVT::v8i8;
2488 Ops.assign(16, Cst); TVT = MVT::v16i8;
2490 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2492 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2495 /// getOnesVector - Returns a vector of specified type with all bits set.
2497 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2498 assert(VT.isVector() && "Expected a vector type");
2500 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2501 // dest type. This ensures they get CSE'd.
2503 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2504 SmallVector<SDValue, 8> Ops;
2507 if (VT.getSizeInBits() == 64) {
2508 Ops.assign(8, Cst); TVT = MVT::v8i8;
2510 Ops.assign(16, Cst); TVT = MVT::v16i8;
2512 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2514 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2517 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2518 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2519 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
2520 SelectionDAG &DAG) const {
2521 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2522 EVT VT = Op.getValueType();
2523 unsigned VTBits = VT.getSizeInBits();
2524 DebugLoc dl = Op.getDebugLoc();
2525 SDValue ShOpLo = Op.getOperand(0);
2526 SDValue ShOpHi = Op.getOperand(1);
2527 SDValue ShAmt = Op.getOperand(2);
2529 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2531 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2533 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2534 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2535 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2536 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2537 DAG.getConstant(VTBits, MVT::i32));
2538 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2539 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2540 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2542 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2543 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2545 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2546 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2549 SDValue Ops[2] = { Lo, Hi };
2550 return DAG.getMergeValues(Ops, 2, dl);
2553 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2554 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2555 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
2556 SelectionDAG &DAG) const {
2557 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2558 EVT VT = Op.getValueType();
2559 unsigned VTBits = VT.getSizeInBits();
2560 DebugLoc dl = Op.getDebugLoc();
2561 SDValue ShOpLo = Op.getOperand(0);
2562 SDValue ShOpHi = Op.getOperand(1);
2563 SDValue ShAmt = Op.getOperand(2);
2566 assert(Op.getOpcode() == ISD::SHL_PARTS);
2567 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2568 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2569 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2570 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2571 DAG.getConstant(VTBits, MVT::i32));
2572 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2573 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2575 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2576 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2577 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2579 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2580 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2583 SDValue Ops[2] = { Lo, Hi };
2584 return DAG.getMergeValues(Ops, 2, dl);
2587 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
2588 const ARMSubtarget *ST) {
2589 EVT VT = N->getValueType(0);
2590 DebugLoc dl = N->getDebugLoc();
2592 if (!ST->hasV6T2Ops())
2595 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
2596 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
2599 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2600 const ARMSubtarget *ST) {
2601 EVT VT = N->getValueType(0);
2602 DebugLoc dl = N->getDebugLoc();
2604 // Lower vector shifts on NEON to use VSHL.
2605 if (VT.isVector()) {
2606 assert(ST->hasNEON() && "unexpected vector shift");
2608 // Left shifts translate directly to the vshiftu intrinsic.
2609 if (N->getOpcode() == ISD::SHL)
2610 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2611 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2612 N->getOperand(0), N->getOperand(1));
2614 assert((N->getOpcode() == ISD::SRA ||
2615 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2617 // NEON uses the same intrinsics for both left and right shifts. For
2618 // right shifts, the shift amounts are negative, so negate the vector of
2620 EVT ShiftVT = N->getOperand(1).getValueType();
2621 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2622 getZeroVector(ShiftVT, DAG, dl),
2624 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2625 Intrinsic::arm_neon_vshifts :
2626 Intrinsic::arm_neon_vshiftu);
2627 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2628 DAG.getConstant(vshiftInt, MVT::i32),
2629 N->getOperand(0), NegatedCount);
2632 // We can get here for a node like i32 = ISD::SHL i32, i64
2636 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2637 "Unknown shift to lower!");
2639 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2640 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2641 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2644 // If we are in thumb mode, we don't have RRX.
2645 if (ST->isThumb1Only()) return SDValue();
2647 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2648 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2649 DAG.getConstant(0, MVT::i32));
2650 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2651 DAG.getConstant(1, MVT::i32));
2653 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2654 // captures the result into a carry flag.
2655 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2656 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2658 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2659 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2661 // Merge the pieces into a single i64 value.
2662 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2665 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2666 SDValue TmpOp0, TmpOp1;
2667 bool Invert = false;
2671 SDValue Op0 = Op.getOperand(0);
2672 SDValue Op1 = Op.getOperand(1);
2673 SDValue CC = Op.getOperand(2);
2674 EVT VT = Op.getValueType();
2675 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2676 DebugLoc dl = Op.getDebugLoc();
2678 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2679 switch (SetCCOpcode) {
2680 default: llvm_unreachable("Illegal FP comparison"); break;
2682 case ISD::SETNE: Invert = true; // Fallthrough
2684 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2686 case ISD::SETLT: Swap = true; // Fallthrough
2688 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2690 case ISD::SETLE: Swap = true; // Fallthrough
2692 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2693 case ISD::SETUGE: Swap = true; // Fallthrough
2694 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2695 case ISD::SETUGT: Swap = true; // Fallthrough
2696 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2697 case ISD::SETUEQ: Invert = true; // Fallthrough
2699 // Expand this to (OLT | OGT).
2703 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2704 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2706 case ISD::SETUO: Invert = true; // Fallthrough
2708 // Expand this to (OLT | OGE).
2712 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2713 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2717 // Integer comparisons.
2718 switch (SetCCOpcode) {
2719 default: llvm_unreachable("Illegal integer comparison"); break;
2720 case ISD::SETNE: Invert = true;
2721 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2722 case ISD::SETLT: Swap = true;
2723 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2724 case ISD::SETLE: Swap = true;
2725 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2726 case ISD::SETULT: Swap = true;
2727 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2728 case ISD::SETULE: Swap = true;
2729 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2732 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2733 if (Opc == ARMISD::VCEQ) {
2736 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2738 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2741 // Ignore bitconvert.
2742 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2743 AndOp = AndOp.getOperand(0);
2745 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2747 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2748 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2755 std::swap(Op0, Op1);
2757 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2760 Result = DAG.getNOT(dl, Result, VT);
2765 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
2766 /// valid vector constant for a NEON instruction with a "modified immediate"
2767 /// operand (e.g., VMOV). If so, return either the constant being
2768 /// splatted or the encoded value, depending on the DoEncode parameter. The
2769 /// format of the encoded value is: bit12=Op, bits11-8=Cmode,
2770 /// bits7-0=Immediate.
2771 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
2772 unsigned SplatBitSize, SelectionDAG &DAG,
2773 bool isVMOV, bool DoEncode) {
2774 unsigned Op, Cmode, Imm;
2777 // SplatBitSize is set to the smallest size that splats the vector, so a
2778 // zero vector will always have SplatBitSize == 8. However, NEON modified
2779 // immediate instructions others than VMOV do not support the 8-bit encoding
2780 // of a zero vector, and the default encoding of zero is supposed to be the
2786 switch (SplatBitSize) {
2788 // Any 1-byte value is OK. Op=0, Cmode=1110.
2789 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2796 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2798 if ((SplatBits & ~0xff) == 0) {
2799 // Value = 0x00nn: Op=x, Cmode=100x.
2804 if ((SplatBits & ~0xff00) == 0) {
2805 // Value = 0xnn00: Op=x, Cmode=101x.
2807 Imm = SplatBits >> 8;
2813 // NEON's 32-bit VMOV supports splat values where:
2814 // * only one byte is nonzero, or
2815 // * the least significant byte is 0xff and the second byte is nonzero, or
2816 // * the least significant 2 bytes are 0xff and the third is nonzero.
2818 if ((SplatBits & ~0xff) == 0) {
2819 // Value = 0x000000nn: Op=x, Cmode=000x.
2824 if ((SplatBits & ~0xff00) == 0) {
2825 // Value = 0x0000nn00: Op=x, Cmode=001x.
2827 Imm = SplatBits >> 8;
2830 if ((SplatBits & ~0xff0000) == 0) {
2831 // Value = 0x00nn0000: Op=x, Cmode=010x.
2833 Imm = SplatBits >> 16;
2836 if ((SplatBits & ~0xff000000) == 0) {
2837 // Value = 0xnn000000: Op=x, Cmode=011x.
2839 Imm = SplatBits >> 24;
2843 if ((SplatBits & ~0xffff) == 0 &&
2844 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
2845 // Value = 0x0000nnff: Op=x, Cmode=1100.
2847 Imm = SplatBits >> 8;
2852 if ((SplatBits & ~0xffffff) == 0 &&
2853 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
2854 // Value = 0x00nnffff: Op=x, Cmode=1101.
2856 Imm = SplatBits >> 16;
2857 SplatBits |= 0xffff;
2861 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2862 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2863 // VMOV.I32. A (very) minor optimization would be to replicate the value
2864 // and fall through here to test for a valid 64-bit splat. But, then the
2865 // caller would also need to check and handle the change in size.
2869 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2872 uint64_t BitMask = 0xff;
2874 unsigned ImmMask = 1;
2876 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2877 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
2880 } else if ((SplatBits & BitMask) != 0) {
2886 // Op=1, Cmode=1110.
2895 llvm_unreachable("unexpected size for EncodeNEONModImm");
2900 return DAG.getTargetConstant((Op << 12) | (Cmode << 8) | Imm, MVT::i32);
2901 return DAG.getTargetConstant(SplatBits, VT);
2905 /// getNEONModImm - If this is a valid vector constant for a NEON instruction
2906 /// with a "modified immediate" operand (e.g., VMOV) of the specified element
2907 /// size, return the encoded value for that immediate. The ByteSize field
2908 /// indicates the number of bytes of each element [1248].
2909 SDValue ARM::getNEONModImm(SDNode *N, unsigned ByteSize, bool isVMOV,
2910 SelectionDAG &DAG) {
2911 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2912 APInt SplatBits, SplatUndef;
2913 unsigned SplatBitSize;
2915 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2916 HasAnyUndefs, ByteSize * 8))
2919 if (SplatBitSize > ByteSize * 8)
2922 return isNEONModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2923 SplatBitSize, DAG, isVMOV, true);
2926 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2927 bool &ReverseVEXT, unsigned &Imm) {
2928 unsigned NumElts = VT.getVectorNumElements();
2929 ReverseVEXT = false;
2932 // If this is a VEXT shuffle, the immediate value is the index of the first
2933 // element. The other shuffle indices must be the successive elements after
2935 unsigned ExpectedElt = Imm;
2936 for (unsigned i = 1; i < NumElts; ++i) {
2937 // Increment the expected index. If it wraps around, it may still be
2938 // a VEXT but the source vectors must be swapped.
2940 if (ExpectedElt == NumElts * 2) {
2945 if (ExpectedElt != static_cast<unsigned>(M[i]))
2949 // Adjust the index value if the source operands will be swapped.
2956 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2957 /// instruction with the specified blocksize. (The order of the elements
2958 /// within each block of the vector is reversed.)
2959 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2960 unsigned BlockSize) {
2961 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2962 "Only possible block sizes for VREV are: 16, 32, 64");
2964 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2968 unsigned NumElts = VT.getVectorNumElements();
2969 unsigned BlockElts = M[0] + 1;
2971 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2974 for (unsigned i = 0; i < NumElts; ++i) {
2975 if ((unsigned) M[i] !=
2976 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2983 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2984 unsigned &WhichResult) {
2985 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2989 unsigned NumElts = VT.getVectorNumElements();
2990 WhichResult = (M[0] == 0 ? 0 : 1);
2991 for (unsigned i = 0; i < NumElts; i += 2) {
2992 if ((unsigned) M[i] != i + WhichResult ||
2993 (unsigned) M[i+1] != i + NumElts + WhichResult)
2999 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3000 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3001 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3002 static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3003 unsigned &WhichResult) {
3004 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3008 unsigned NumElts = VT.getVectorNumElements();
3009 WhichResult = (M[0] == 0 ? 0 : 1);
3010 for (unsigned i = 0; i < NumElts; i += 2) {
3011 if ((unsigned) M[i] != i + WhichResult ||
3012 (unsigned) M[i+1] != i + WhichResult)
3018 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3019 unsigned &WhichResult) {
3020 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3024 unsigned NumElts = VT.getVectorNumElements();
3025 WhichResult = (M[0] == 0 ? 0 : 1);
3026 for (unsigned i = 0; i != NumElts; ++i) {
3027 if ((unsigned) M[i] != 2 * i + WhichResult)
3031 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3032 if (VT.is64BitVector() && EltSz == 32)
3038 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3039 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3040 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3041 static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3042 unsigned &WhichResult) {
3043 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3047 unsigned Half = VT.getVectorNumElements() / 2;
3048 WhichResult = (M[0] == 0 ? 0 : 1);
3049 for (unsigned j = 0; j != 2; ++j) {
3050 unsigned Idx = WhichResult;
3051 for (unsigned i = 0; i != Half; ++i) {
3052 if ((unsigned) M[i + j * Half] != Idx)
3058 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3059 if (VT.is64BitVector() && EltSz == 32)
3065 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3066 unsigned &WhichResult) {
3067 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3071 unsigned NumElts = VT.getVectorNumElements();
3072 WhichResult = (M[0] == 0 ? 0 : 1);
3073 unsigned Idx = WhichResult * NumElts / 2;
3074 for (unsigned i = 0; i != NumElts; i += 2) {
3075 if ((unsigned) M[i] != Idx ||
3076 (unsigned) M[i+1] != Idx + NumElts)
3081 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3082 if (VT.is64BitVector() && EltSz == 32)
3088 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3089 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3090 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3091 static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3092 unsigned &WhichResult) {
3093 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3097 unsigned NumElts = VT.getVectorNumElements();
3098 WhichResult = (M[0] == 0 ? 0 : 1);
3099 unsigned Idx = WhichResult * NumElts / 2;
3100 for (unsigned i = 0; i != NumElts; i += 2) {
3101 if ((unsigned) M[i] != Idx ||
3102 (unsigned) M[i+1] != Idx)
3107 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3108 if (VT.is64BitVector() && EltSz == 32)
3115 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3116 // Canonicalize all-zeros and all-ones vectors.
3117 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
3118 if (ConstVal->isNullValue())
3119 return getZeroVector(VT, DAG, dl);
3120 if (ConstVal->isAllOnesValue())
3121 return getOnesVector(VT, DAG, dl);
3124 if (VT.is64BitVector()) {
3125 switch (Val.getValueType().getSizeInBits()) {
3126 case 8: CanonicalVT = MVT::v8i8; break;
3127 case 16: CanonicalVT = MVT::v4i16; break;
3128 case 32: CanonicalVT = MVT::v2i32; break;
3129 case 64: CanonicalVT = MVT::v1i64; break;
3130 default: llvm_unreachable("unexpected splat element type"); break;
3133 assert(VT.is128BitVector() && "unknown splat vector size");
3134 switch (Val.getValueType().getSizeInBits()) {
3135 case 8: CanonicalVT = MVT::v16i8; break;
3136 case 16: CanonicalVT = MVT::v8i16; break;
3137 case 32: CanonicalVT = MVT::v4i32; break;
3138 case 64: CanonicalVT = MVT::v2i64; break;
3139 default: llvm_unreachable("unexpected splat element type"); break;
3143 // Build a canonical splat for this value.
3144 SmallVector<SDValue, 8> Ops;
3145 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
3146 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
3148 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
3151 // If this is a case we can't handle, return null and let the default
3152 // expansion code take care of it.
3153 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3154 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3155 DebugLoc dl = Op.getDebugLoc();
3156 EVT VT = Op.getValueType();
3158 APInt SplatBits, SplatUndef;
3159 unsigned SplatBitSize;
3161 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3162 if (SplatBitSize <= 64) {
3163 // Check if an immediate VMOV works.
3164 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3165 SplatUndef.getZExtValue(),
3166 SplatBitSize, DAG, true, false);
3168 return BuildSplat(Val, VT, DAG, dl);
3172 // Scan through the operands to see if only one value is used.
3173 unsigned NumElts = VT.getVectorNumElements();
3174 bool isOnlyLowElement = true;
3175 bool usesOnlyOneValue = true;
3176 bool isConstant = true;
3178 for (unsigned i = 0; i < NumElts; ++i) {
3179 SDValue V = Op.getOperand(i);
3180 if (V.getOpcode() == ISD::UNDEF)
3183 isOnlyLowElement = false;
3184 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3187 if (!Value.getNode())
3189 else if (V != Value)
3190 usesOnlyOneValue = false;
3193 if (!Value.getNode())
3194 return DAG.getUNDEF(VT);
3196 if (isOnlyLowElement)
3197 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3199 // If all elements are constants, fall back to the default expansion, which
3200 // will generate a load from the constant pool.
3204 // Use VDUP for non-constant splats.
3205 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3206 if (usesOnlyOneValue && EltSize <= 32)
3207 return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3209 // Vectors with 32- or 64-bit elements can be built by directly assigning
3210 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
3211 // will be legalized.
3212 if (EltSize >= 32) {
3213 // Do the expansion with floating-point types, since that is what the VFP
3214 // registers are defined to use, and since i64 is not legal.
3215 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3216 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3217 SmallVector<SDValue, 8> Ops;
3218 for (unsigned i = 0; i < NumElts; ++i)
3219 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, EltVT, Op.getOperand(i)));
3220 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3221 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3227 /// isShuffleMaskLegal - Targets can use this to indicate that they only
3228 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
3229 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
3230 /// are assumed to be legal.
3232 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
3234 if (VT.getVectorNumElements() == 4 &&
3235 (VT.is128BitVector() || VT.is64BitVector())) {
3236 unsigned PFIndexes[4];
3237 for (unsigned i = 0; i != 4; ++i) {
3241 PFIndexes[i] = M[i];
3244 // Compute the index in the perfect shuffle table.
3245 unsigned PFTableIndex =
3246 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3247 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3248 unsigned Cost = (PFEntry >> 30);
3255 unsigned Imm, WhichResult;
3257 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3258 return (EltSize >= 32 ||
3259 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
3260 isVREVMask(M, VT, 64) ||
3261 isVREVMask(M, VT, 32) ||
3262 isVREVMask(M, VT, 16) ||
3263 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
3264 isVTRNMask(M, VT, WhichResult) ||
3265 isVUZPMask(M, VT, WhichResult) ||
3266 isVZIPMask(M, VT, WhichResult) ||
3267 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
3268 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
3269 isVZIP_v_undef_Mask(M, VT, WhichResult));
3272 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3273 /// the specified operations to build the shuffle.
3274 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3275 SDValue RHS, SelectionDAG &DAG,
3277 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3278 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3279 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3282 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3291 OP_VUZPL, // VUZP, left result
3292 OP_VUZPR, // VUZP, right result
3293 OP_VZIPL, // VZIP, left result
3294 OP_VZIPR, // VZIP, right result
3295 OP_VTRNL, // VTRN, left result
3296 OP_VTRNR // VTRN, right result
3299 if (OpNum == OP_COPY) {
3300 if (LHSID == (1*9+2)*9+3) return LHS;
3301 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3305 SDValue OpLHS, OpRHS;
3306 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3307 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3308 EVT VT = OpLHS.getValueType();
3311 default: llvm_unreachable("Unknown shuffle opcode!");
3313 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
3318 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
3319 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
3323 return DAG.getNode(ARMISD::VEXT, dl, VT,
3325 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
3328 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3329 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
3332 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3333 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
3336 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3337 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
3341 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
3342 SDValue V1 = Op.getOperand(0);
3343 SDValue V2 = Op.getOperand(1);
3344 DebugLoc dl = Op.getDebugLoc();
3345 EVT VT = Op.getValueType();
3346 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
3347 SmallVector<int, 8> ShuffleMask;
3349 // Convert shuffles that are directly supported on NEON to target-specific
3350 // DAG nodes, instead of keeping them as shuffles and matching them again
3351 // during code selection. This is more efficient and avoids the possibility
3352 // of inconsistencies between legalization and selection.
3353 // FIXME: floating-point vectors should be canonicalized to integer vectors
3354 // of the same time so that they get CSEd properly.
3355 SVN->getMask(ShuffleMask);
3357 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3358 if (EltSize <= 32) {
3359 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
3360 int Lane = SVN->getSplatIndex();
3361 // If this is undef splat, generate it via "just" vdup, if possible.
3362 if (Lane == -1) Lane = 0;
3364 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3365 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
3367 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
3368 DAG.getConstant(Lane, MVT::i32));
3373 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
3376 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
3377 DAG.getConstant(Imm, MVT::i32));
3380 if (isVREVMask(ShuffleMask, VT, 64))
3381 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
3382 if (isVREVMask(ShuffleMask, VT, 32))
3383 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
3384 if (isVREVMask(ShuffleMask, VT, 16))
3385 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
3387 // Check for Neon shuffles that modify both input vectors in place.
3388 // If both results are used, i.e., if there are two shuffles with the same
3389 // source operands and with masks corresponding to both results of one of
3390 // these operations, DAG memoization will ensure that a single node is
3391 // used for both shuffles.
3392 unsigned WhichResult;
3393 if (isVTRNMask(ShuffleMask, VT, WhichResult))
3394 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3395 V1, V2).getValue(WhichResult);
3396 if (isVUZPMask(ShuffleMask, VT, WhichResult))
3397 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3398 V1, V2).getValue(WhichResult);
3399 if (isVZIPMask(ShuffleMask, VT, WhichResult))
3400 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3401 V1, V2).getValue(WhichResult);
3403 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
3404 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
3405 V1, V1).getValue(WhichResult);
3406 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3407 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
3408 V1, V1).getValue(WhichResult);
3409 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
3410 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
3411 V1, V1).getValue(WhichResult);
3414 // If the shuffle is not directly supported and it has 4 elements, use
3415 // the PerfectShuffle-generated table to synthesize it from other shuffles.
3416 unsigned NumElts = VT.getVectorNumElements();
3418 unsigned PFIndexes[4];
3419 for (unsigned i = 0; i != 4; ++i) {
3420 if (ShuffleMask[i] < 0)
3423 PFIndexes[i] = ShuffleMask[i];
3426 // Compute the index in the perfect shuffle table.
3427 unsigned PFTableIndex =
3428 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3429 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3430 unsigned Cost = (PFEntry >> 30);
3433 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3436 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
3437 if (EltSize >= 32) {
3438 // Do the expansion with floating-point types, since that is what the VFP
3439 // registers are defined to use, and since i64 is not legal.
3440 EVT EltVT = EVT::getFloatingPointVT(EltSize);
3441 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
3442 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V1);
3443 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, VecVT, V2);
3444 SmallVector<SDValue, 8> Ops;
3445 for (unsigned i = 0; i < NumElts; ++i) {
3446 if (ShuffleMask[i] < 0)
3447 Ops.push_back(DAG.getUNDEF(EltVT));
3449 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3450 ShuffleMask[i] < (int)NumElts ? V1 : V2,
3451 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
3454 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
3455 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Val);
3461 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
3462 EVT VT = Op.getValueType();
3463 DebugLoc dl = Op.getDebugLoc();
3464 SDValue Vec = Op.getOperand(0);
3465 SDValue Lane = Op.getOperand(1);
3466 assert(VT == MVT::i32 &&
3467 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
3468 "unexpected type for custom-lowering vector extract");
3469 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
3472 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3473 // The only time a CONCAT_VECTORS operation can have legal types is when
3474 // two 64-bit vectors are concatenated to a 128-bit vector.
3475 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
3476 "unexpected CONCAT_VECTORS");
3477 DebugLoc dl = Op.getDebugLoc();
3478 SDValue Val = DAG.getUNDEF(MVT::v2f64);
3479 SDValue Op0 = Op.getOperand(0);
3480 SDValue Op1 = Op.getOperand(1);
3481 if (Op0.getOpcode() != ISD::UNDEF)
3482 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3483 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
3484 DAG.getIntPtrConstant(0));
3485 if (Op1.getOpcode() != ISD::UNDEF)
3486 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
3487 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
3488 DAG.getIntPtrConstant(1));
3489 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
3492 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3493 switch (Op.getOpcode()) {
3494 default: llvm_unreachable("Don't know how to custom lower this!");
3495 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3496 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3497 case ISD::GlobalAddress:
3498 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
3499 LowerGlobalAddressELF(Op, DAG);
3500 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3501 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3502 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3503 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
3504 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
3505 case ISD::VASTART: return LowerVASTART(Op, DAG);
3506 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
3507 case ISD::SINT_TO_FP:
3508 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
3509 case ISD::FP_TO_SINT:
3510 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
3511 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
3512 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3513 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3514 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
3515 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
3516 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
3517 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
3519 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
3522 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
3523 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3524 case ISD::SRL_PARTS:
3525 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3526 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
3527 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
3528 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3529 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3530 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
3531 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3536 /// ReplaceNodeResults - Replace the results of node with an illegal result
3537 /// type with new values built out of custom code.
3538 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
3539 SmallVectorImpl<SDValue>&Results,
3540 SelectionDAG &DAG) const {
3542 switch (N->getOpcode()) {
3544 llvm_unreachable("Don't know how to custom expand this!");
3546 case ISD::BIT_CONVERT:
3547 Res = ExpandBIT_CONVERT(N, DAG);
3551 Res = LowerShift(N, DAG, Subtarget);
3555 Results.push_back(Res);
3558 //===----------------------------------------------------------------------===//
3559 // ARM Scheduler Hooks
3560 //===----------------------------------------------------------------------===//
3563 ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
3564 MachineBasicBlock *BB,
3565 unsigned Size) const {
3566 unsigned dest = MI->getOperand(0).getReg();
3567 unsigned ptr = MI->getOperand(1).getReg();
3568 unsigned oldval = MI->getOperand(2).getReg();
3569 unsigned newval = MI->getOperand(3).getReg();
3570 unsigned scratch = BB->getParent()->getRegInfo()
3571 .createVirtualRegister(ARM::GPRRegisterClass);
3572 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3573 DebugLoc dl = MI->getDebugLoc();
3574 bool isThumb2 = Subtarget->isThumb2();
3576 unsigned ldrOpc, strOpc;
3578 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3580 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3581 strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
3584 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3585 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3588 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3589 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3593 MachineFunction *MF = BB->getParent();
3594 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3595 MachineFunction::iterator It = BB;
3596 ++It; // insert the new blocks after the current block
3598 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3599 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
3600 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3601 MF->insert(It, loop1MBB);
3602 MF->insert(It, loop2MBB);
3603 MF->insert(It, exitMBB);
3604 exitMBB->transferSuccessors(BB);
3608 // fallthrough --> loop1MBB
3609 BB->addSuccessor(loop1MBB);
3612 // ldrex dest, [ptr]
3616 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3617 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
3618 .addReg(dest).addReg(oldval));
3619 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3620 .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3621 BB->addSuccessor(loop2MBB);
3622 BB->addSuccessor(exitMBB);
3625 // strex scratch, newval, [ptr]
3629 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
3631 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3632 .addReg(scratch).addImm(0));
3633 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3634 .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3635 BB->addSuccessor(loop1MBB);
3636 BB->addSuccessor(exitMBB);
3642 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3648 ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3649 unsigned Size, unsigned BinOpcode) const {
3650 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3651 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3653 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3654 MachineFunction *MF = BB->getParent();
3655 MachineFunction::iterator It = BB;
3658 unsigned dest = MI->getOperand(0).getReg();
3659 unsigned ptr = MI->getOperand(1).getReg();
3660 unsigned incr = MI->getOperand(2).getReg();
3661 DebugLoc dl = MI->getDebugLoc();
3663 bool isThumb2 = Subtarget->isThumb2();
3664 unsigned ldrOpc, strOpc;
3666 default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
3668 ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
3669 strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
3672 ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
3673 strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
3676 ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
3677 strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
3681 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3682 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
3683 MF->insert(It, loopMBB);
3684 MF->insert(It, exitMBB);
3685 exitMBB->transferSuccessors(BB);
3687 MachineRegisterInfo &RegInfo = MF->getRegInfo();
3688 unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3689 unsigned scratch2 = (!BinOpcode) ? incr :
3690 RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
3694 // fallthrough --> loopMBB
3695 BB->addSuccessor(loopMBB);
3699 // <binop> scratch2, dest, incr
3700 // strex scratch, scratch2, ptr
3703 // fallthrough --> exitMBB
3705 AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
3707 // operand order needs to go the other way for NAND
3708 if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
3709 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3710 addReg(incr).addReg(dest)).addReg(0);
3712 AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
3713 addReg(dest).addReg(incr)).addReg(0);
3716 AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
3718 AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
3719 .addReg(scratch).addImm(0));
3720 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
3721 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
3723 BB->addSuccessor(loopMBB);
3724 BB->addSuccessor(exitMBB);
3730 MF->DeleteMachineInstr(MI); // The instruction is gone now.
3736 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3737 MachineBasicBlock *BB) const {
3738 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3739 DebugLoc dl = MI->getDebugLoc();
3740 bool isThumb2 = Subtarget->isThumb2();
3741 switch (MI->getOpcode()) {
3744 llvm_unreachable("Unexpected instr type to insert");
3746 case ARM::ATOMIC_LOAD_ADD_I8:
3747 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3748 case ARM::ATOMIC_LOAD_ADD_I16:
3749 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3750 case ARM::ATOMIC_LOAD_ADD_I32:
3751 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
3753 case ARM::ATOMIC_LOAD_AND_I8:
3754 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3755 case ARM::ATOMIC_LOAD_AND_I16:
3756 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3757 case ARM::ATOMIC_LOAD_AND_I32:
3758 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
3760 case ARM::ATOMIC_LOAD_OR_I8:
3761 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3762 case ARM::ATOMIC_LOAD_OR_I16:
3763 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3764 case ARM::ATOMIC_LOAD_OR_I32:
3765 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
3767 case ARM::ATOMIC_LOAD_XOR_I8:
3768 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3769 case ARM::ATOMIC_LOAD_XOR_I16:
3770 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3771 case ARM::ATOMIC_LOAD_XOR_I32:
3772 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
3774 case ARM::ATOMIC_LOAD_NAND_I8:
3775 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3776 case ARM::ATOMIC_LOAD_NAND_I16:
3777 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3778 case ARM::ATOMIC_LOAD_NAND_I32:
3779 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
3781 case ARM::ATOMIC_LOAD_SUB_I8:
3782 return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3783 case ARM::ATOMIC_LOAD_SUB_I16:
3784 return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3785 case ARM::ATOMIC_LOAD_SUB_I32:
3786 return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
3788 case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
3789 case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
3790 case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
3792 case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
3793 case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
3794 case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
3796 case ARM::tMOVCCr_pseudo: {
3797 // To "insert" a SELECT_CC instruction, we actually have to insert the
3798 // diamond control-flow pattern. The incoming instruction knows the
3799 // destination vreg to set, the condition code register to branch on, the
3800 // true/false values to select between, and a branch opcode to use.
3801 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3802 MachineFunction::iterator It = BB;
3808 // cmpTY ccX, r1, r2
3810 // fallthrough --> copy0MBB
3811 MachineBasicBlock *thisMBB = BB;
3812 MachineFunction *F = BB->getParent();
3813 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3814 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3815 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
3816 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
3817 F->insert(It, copy0MBB);
3818 F->insert(It, sinkMBB);
3819 // Update machine-CFG edges by first adding all successors of the current
3820 // block to the new block which will contain the Phi node for the select.
3821 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
3822 E = BB->succ_end(); I != E; ++I)
3823 sinkMBB->addSuccessor(*I);
3824 // Next, remove all successors of the current block, and add the true
3825 // and fallthrough blocks as its successors.
3826 while (!BB->succ_empty())
3827 BB->removeSuccessor(BB->succ_begin());
3828 BB->addSuccessor(copy0MBB);
3829 BB->addSuccessor(sinkMBB);
3832 // %FalseValue = ...
3833 // # fallthrough to sinkMBB
3836 // Update machine-CFG edges
3837 BB->addSuccessor(sinkMBB);
3840 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3843 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3844 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3845 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3847 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3854 case ARM::t2SUBrSPi_:
3855 case ARM::t2SUBrSPi12_:
3856 case ARM::t2SUBrSPs_: {
3857 MachineFunction *MF = BB->getParent();
3858 unsigned DstReg = MI->getOperand(0).getReg();
3859 unsigned SrcReg = MI->getOperand(1).getReg();
3860 bool DstIsDead = MI->getOperand(0).isDead();
3861 bool SrcIsKill = MI->getOperand(1).isKill();
3863 if (SrcReg != ARM::SP) {
3864 // Copy the source to SP from virtual register.
3865 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3866 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3867 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3868 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3869 .addReg(SrcReg, getKillRegState(SrcIsKill));
3873 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3874 switch (MI->getOpcode()) {
3876 llvm_unreachable("Unexpected pseudo instruction!");
3882 OpOpc = ARM::tADDspr;
3885 OpOpc = ARM::tSUBspi;
3887 case ARM::t2SUBrSPi_:
3888 OpOpc = ARM::t2SUBrSPi;
3889 NeedPred = true; NeedCC = true;
3891 case ARM::t2SUBrSPi12_:
3892 OpOpc = ARM::t2SUBrSPi12;
3895 case ARM::t2SUBrSPs_:
3896 OpOpc = ARM::t2SUBrSPs;
3897 NeedPred = true; NeedCC = true; NeedOp3 = true;
3900 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3901 if (OpOpc == ARM::tAND)
3902 AddDefaultT1CC(MIB);
3903 MIB.addReg(ARM::SP);
3904 MIB.addOperand(MI->getOperand(2));
3906 MIB.addOperand(MI->getOperand(3));
3908 AddDefaultPred(MIB);
3912 // Copy the result from SP to virtual register.
3913 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3914 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3915 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3916 BuildMI(BB, dl, TII->get(CopyOpc))
3917 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3919 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3925 //===----------------------------------------------------------------------===//
3926 // ARM Optimization Hooks
3927 //===----------------------------------------------------------------------===//
3930 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3931 TargetLowering::DAGCombinerInfo &DCI) {
3932 SelectionDAG &DAG = DCI.DAG;
3933 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3934 EVT VT = N->getValueType(0);
3935 unsigned Opc = N->getOpcode();
3936 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3937 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3938 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3939 ISD::CondCode CC = ISD::SETCC_INVALID;
3942 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3944 SDValue CCOp = Slct.getOperand(0);
3945 if (CCOp.getOpcode() == ISD::SETCC)
3946 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3949 bool DoXform = false;
3951 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3954 if (LHS.getOpcode() == ISD::Constant &&
3955 cast<ConstantSDNode>(LHS)->isNullValue()) {
3957 } else if (CC != ISD::SETCC_INVALID &&
3958 RHS.getOpcode() == ISD::Constant &&
3959 cast<ConstantSDNode>(RHS)->isNullValue()) {
3960 std::swap(LHS, RHS);
3961 SDValue Op0 = Slct.getOperand(0);
3962 EVT OpVT = isSlctCC ? Op0.getValueType() :
3963 Op0.getOperand(0).getValueType();
3964 bool isInt = OpVT.isInteger();
3965 CC = ISD::getSetCCInverse(CC, isInt);
3967 if (!TLI.isCondCodeLegal(CC, OpVT))
3968 return SDValue(); // Inverse operator isn't legal.
3975 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3977 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3978 Slct.getOperand(0), Slct.getOperand(1), CC);
3979 SDValue CCOp = Slct.getOperand(0);
3981 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3982 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3983 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3984 CCOp, OtherOp, Result);
3989 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3990 static SDValue PerformADDCombine(SDNode *N,
3991 TargetLowering::DAGCombinerInfo &DCI) {
3992 // added by evan in r37685 with no testcase.
3993 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3995 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3996 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3997 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3998 if (Result.getNode()) return Result;
4000 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4001 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4002 if (Result.getNode()) return Result;
4008 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
4009 static SDValue PerformSUBCombine(SDNode *N,
4010 TargetLowering::DAGCombinerInfo &DCI) {
4011 // added by evan in r37685 with no testcase.
4012 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
4014 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
4015 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
4016 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
4017 if (Result.getNode()) return Result;
4023 static SDValue PerformMULCombine(SDNode *N,
4024 TargetLowering::DAGCombinerInfo &DCI,
4025 const ARMSubtarget *Subtarget) {
4026 SelectionDAG &DAG = DCI.DAG;
4028 if (Subtarget->isThumb1Only())
4031 if (DAG.getMachineFunction().
4032 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
4035 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4038 EVT VT = N->getValueType(0);
4042 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
4046 uint64_t MulAmt = C->getZExtValue();
4047 unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
4048 ShiftAmt = ShiftAmt & (32 - 1);
4049 SDValue V = N->getOperand(0);
4050 DebugLoc DL = N->getDebugLoc();
4053 MulAmt >>= ShiftAmt;
4054 if (isPowerOf2_32(MulAmt - 1)) {
4055 // (mul x, 2^N + 1) => (add (shl x, N), x)
4056 Res = DAG.getNode(ISD::ADD, DL, VT,
4057 V, DAG.getNode(ISD::SHL, DL, VT,
4058 V, DAG.getConstant(Log2_32(MulAmt-1),
4060 } else if (isPowerOf2_32(MulAmt + 1)) {
4061 // (mul x, 2^N - 1) => (sub (shl x, N), x)
4062 Res = DAG.getNode(ISD::SUB, DL, VT,
4063 DAG.getNode(ISD::SHL, DL, VT,
4064 V, DAG.getConstant(Log2_32(MulAmt+1),
4071 Res = DAG.getNode(ISD::SHL, DL, VT, Res,
4072 DAG.getConstant(ShiftAmt, MVT::i32));
4074 // Do not add new nodes to DAG combiner worklist.
4075 DCI.CombineTo(N, Res, false);
4079 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
4080 /// ARMISD::VMOVRRD.
4081 static SDValue PerformVMOVRRDCombine(SDNode *N,
4082 TargetLowering::DAGCombinerInfo &DCI) {
4083 // fmrrd(fmdrr x, y) -> x,y
4084 SDValue InDouble = N->getOperand(0);
4085 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
4086 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
4090 /// getVShiftImm - Check if this is a valid build_vector for the immediate
4091 /// operand of a vector shift operation, where all the elements of the
4092 /// build_vector must have the same constant integer value.
4093 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
4094 // Ignore bit_converts.
4095 while (Op.getOpcode() == ISD::BIT_CONVERT)
4096 Op = Op.getOperand(0);
4097 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4098 APInt SplatBits, SplatUndef;
4099 unsigned SplatBitSize;
4101 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
4102 HasAnyUndefs, ElementBits) ||
4103 SplatBitSize > ElementBits)
4105 Cnt = SplatBits.getSExtValue();
4109 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
4110 /// operand of a vector shift left operation. That value must be in the range:
4111 /// 0 <= Value < ElementBits for a left shift; or
4112 /// 0 <= Value <= ElementBits for a long left shift.
4113 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
4114 assert(VT.isVector() && "vector shift count is not a vector type");
4115 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4116 if (! getVShiftImm(Op, ElementBits, Cnt))
4118 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
4121 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
4122 /// operand of a vector shift right operation. For a shift opcode, the value
4123 /// is positive, but for an intrinsic the value count must be negative. The
4124 /// absolute value must be in the range:
4125 /// 1 <= |Value| <= ElementBits for a right shift; or
4126 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
4127 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
4129 assert(VT.isVector() && "vector shift count is not a vector type");
4130 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
4131 if (! getVShiftImm(Op, ElementBits, Cnt))
4135 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
4138 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
4139 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4140 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4143 // Don't do anything for most intrinsics.
4146 // Vector shifts: check for immediate versions and lower them.
4147 // Note: This is done during DAG combining instead of DAG legalizing because
4148 // the build_vectors for 64-bit vector element shift counts are generally
4149 // not legal, and it is hard to see their values after they get legalized to
4150 // loads from a constant pool.
4151 case Intrinsic::arm_neon_vshifts:
4152 case Intrinsic::arm_neon_vshiftu:
4153 case Intrinsic::arm_neon_vshiftls:
4154 case Intrinsic::arm_neon_vshiftlu:
4155 case Intrinsic::arm_neon_vshiftn:
4156 case Intrinsic::arm_neon_vrshifts:
4157 case Intrinsic::arm_neon_vrshiftu:
4158 case Intrinsic::arm_neon_vrshiftn:
4159 case Intrinsic::arm_neon_vqshifts:
4160 case Intrinsic::arm_neon_vqshiftu:
4161 case Intrinsic::arm_neon_vqshiftsu:
4162 case Intrinsic::arm_neon_vqshiftns:
4163 case Intrinsic::arm_neon_vqshiftnu:
4164 case Intrinsic::arm_neon_vqshiftnsu:
4165 case Intrinsic::arm_neon_vqrshiftns:
4166 case Intrinsic::arm_neon_vqrshiftnu:
4167 case Intrinsic::arm_neon_vqrshiftnsu: {
4168 EVT VT = N->getOperand(1).getValueType();
4170 unsigned VShiftOpc = 0;
4173 case Intrinsic::arm_neon_vshifts:
4174 case Intrinsic::arm_neon_vshiftu:
4175 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
4176 VShiftOpc = ARMISD::VSHL;
4179 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
4180 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
4181 ARMISD::VSHRs : ARMISD::VSHRu);
4186 case Intrinsic::arm_neon_vshiftls:
4187 case Intrinsic::arm_neon_vshiftlu:
4188 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
4190 llvm_unreachable("invalid shift count for vshll intrinsic");
4192 case Intrinsic::arm_neon_vrshifts:
4193 case Intrinsic::arm_neon_vrshiftu:
4194 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
4198 case Intrinsic::arm_neon_vqshifts:
4199 case Intrinsic::arm_neon_vqshiftu:
4200 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4204 case Intrinsic::arm_neon_vqshiftsu:
4205 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
4207 llvm_unreachable("invalid shift count for vqshlu intrinsic");
4209 case Intrinsic::arm_neon_vshiftn:
4210 case Intrinsic::arm_neon_vrshiftn:
4211 case Intrinsic::arm_neon_vqshiftns:
4212 case Intrinsic::arm_neon_vqshiftnu:
4213 case Intrinsic::arm_neon_vqshiftnsu:
4214 case Intrinsic::arm_neon_vqrshiftns:
4215 case Intrinsic::arm_neon_vqrshiftnu:
4216 case Intrinsic::arm_neon_vqrshiftnsu:
4217 // Narrowing shifts require an immediate right shift.
4218 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
4220 llvm_unreachable("invalid shift count for narrowing vector shift "
4224 llvm_unreachable("unhandled vector shift");
4228 case Intrinsic::arm_neon_vshifts:
4229 case Intrinsic::arm_neon_vshiftu:
4230 // Opcode already set above.
4232 case Intrinsic::arm_neon_vshiftls:
4233 case Intrinsic::arm_neon_vshiftlu:
4234 if (Cnt == VT.getVectorElementType().getSizeInBits())
4235 VShiftOpc = ARMISD::VSHLLi;
4237 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
4238 ARMISD::VSHLLs : ARMISD::VSHLLu);
4240 case Intrinsic::arm_neon_vshiftn:
4241 VShiftOpc = ARMISD::VSHRN; break;
4242 case Intrinsic::arm_neon_vrshifts:
4243 VShiftOpc = ARMISD::VRSHRs; break;
4244 case Intrinsic::arm_neon_vrshiftu:
4245 VShiftOpc = ARMISD::VRSHRu; break;
4246 case Intrinsic::arm_neon_vrshiftn:
4247 VShiftOpc = ARMISD::VRSHRN; break;
4248 case Intrinsic::arm_neon_vqshifts:
4249 VShiftOpc = ARMISD::VQSHLs; break;
4250 case Intrinsic::arm_neon_vqshiftu:
4251 VShiftOpc = ARMISD::VQSHLu; break;
4252 case Intrinsic::arm_neon_vqshiftsu:
4253 VShiftOpc = ARMISD::VQSHLsu; break;
4254 case Intrinsic::arm_neon_vqshiftns:
4255 VShiftOpc = ARMISD::VQSHRNs; break;
4256 case Intrinsic::arm_neon_vqshiftnu:
4257 VShiftOpc = ARMISD::VQSHRNu; break;
4258 case Intrinsic::arm_neon_vqshiftnsu:
4259 VShiftOpc = ARMISD::VQSHRNsu; break;
4260 case Intrinsic::arm_neon_vqrshiftns:
4261 VShiftOpc = ARMISD::VQRSHRNs; break;
4262 case Intrinsic::arm_neon_vqrshiftnu:
4263 VShiftOpc = ARMISD::VQRSHRNu; break;
4264 case Intrinsic::arm_neon_vqrshiftnsu:
4265 VShiftOpc = ARMISD::VQRSHRNsu; break;
4268 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4269 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4272 case Intrinsic::arm_neon_vshiftins: {
4273 EVT VT = N->getOperand(1).getValueType();
4275 unsigned VShiftOpc = 0;
4277 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
4278 VShiftOpc = ARMISD::VSLI;
4279 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
4280 VShiftOpc = ARMISD::VSRI;
4282 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
4285 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
4286 N->getOperand(1), N->getOperand(2),
4287 DAG.getConstant(Cnt, MVT::i32));
4290 case Intrinsic::arm_neon_vqrshifts:
4291 case Intrinsic::arm_neon_vqrshiftu:
4292 // No immediate versions of these to check for.
4299 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
4300 /// lowers them. As with the vector shift intrinsics, this is done during DAG
4301 /// combining instead of DAG legalizing because the build_vectors for 64-bit
4302 /// vector element shift counts are generally not legal, and it is hard to see
4303 /// their values after they get legalized to loads from a constant pool.
4304 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
4305 const ARMSubtarget *ST) {
4306 EVT VT = N->getValueType(0);
4308 // Nothing to be done for scalar shifts.
4309 if (! VT.isVector())
4312 assert(ST->hasNEON() && "unexpected vector shift");
4315 switch (N->getOpcode()) {
4316 default: llvm_unreachable("unexpected shift opcode");
4319 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
4320 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
4321 DAG.getConstant(Cnt, MVT::i32));
4326 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
4327 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
4328 ARMISD::VSHRs : ARMISD::VSHRu);
4329 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
4330 DAG.getConstant(Cnt, MVT::i32));
4336 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
4337 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
4338 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
4339 const ARMSubtarget *ST) {
4340 SDValue N0 = N->getOperand(0);
4342 // Check for sign- and zero-extensions of vector extract operations of 8-
4343 // and 16-bit vector elements. NEON supports these directly. They are
4344 // handled during DAG combining because type legalization will promote them
4345 // to 32-bit types and it is messy to recognize the operations after that.
4346 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
4347 SDValue Vec = N0.getOperand(0);
4348 SDValue Lane = N0.getOperand(1);
4349 EVT VT = N->getValueType(0);
4350 EVT EltVT = N0.getValueType();
4351 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4353 if (VT == MVT::i32 &&
4354 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
4355 TLI.isTypeLegal(Vec.getValueType())) {
4358 switch (N->getOpcode()) {
4359 default: llvm_unreachable("unexpected opcode");
4360 case ISD::SIGN_EXTEND:
4361 Opc = ARMISD::VGETLANEs;
4363 case ISD::ZERO_EXTEND:
4364 case ISD::ANY_EXTEND:
4365 Opc = ARMISD::VGETLANEu;
4368 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
4375 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
4376 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
4377 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
4378 const ARMSubtarget *ST) {
4379 // If the target supports NEON, try to use vmax/vmin instructions for f32
4380 // selects like "x < y ? x : y". Unless the FiniteOnlyFPMath option is set,
4381 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
4382 // a NaN; only do the transformation when it matches that behavior.
4384 // For now only do this when using NEON for FP operations; if using VFP, it
4385 // is not obvious that the benefit outweighs the cost of switching to the
4387 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
4388 N->getValueType(0) != MVT::f32)
4391 SDValue CondLHS = N->getOperand(0);
4392 SDValue CondRHS = N->getOperand(1);
4393 SDValue LHS = N->getOperand(2);
4394 SDValue RHS = N->getOperand(3);
4395 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
4397 unsigned Opcode = 0;
4399 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
4400 IsReversed = false; // x CC y ? x : y
4401 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
4402 IsReversed = true ; // x CC y ? y : x
4416 // If LHS is NaN, an ordered comparison will be false and the result will
4417 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
4418 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4419 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
4420 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4422 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
4423 // will return -0, so vmin can only be used for unsafe math or if one of
4424 // the operands is known to be nonzero.
4425 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
4427 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4429 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
4438 // If LHS is NaN, an ordered comparison will be false and the result will
4439 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
4440 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
4441 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
4442 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
4444 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
4445 // will return +0, so vmax can only be used for unsafe math or if one of
4446 // the operands is known to be nonzero.
4447 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
4449 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
4451 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
4457 return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
4460 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
4461 DAGCombinerInfo &DCI) const {
4462 switch (N->getOpcode()) {
4464 case ISD::ADD: return PerformADDCombine(N, DCI);
4465 case ISD::SUB: return PerformSUBCombine(N, DCI);
4466 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
4467 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
4468 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
4471 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
4472 case ISD::SIGN_EXTEND:
4473 case ISD::ZERO_EXTEND:
4474 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
4475 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
4480 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
4481 if (!Subtarget->hasV6Ops())
4482 // Pre-v6 does not support unaligned mem access.
4485 // v6+ may or may not support unaligned mem access depending on the system
4487 // FIXME: This is pretty conservative. Should we provide cmdline option to
4488 // control the behaviour?
4489 if (!Subtarget->isTargetDarwin())
4493 switch (VT.getSimpleVT().SimpleTy) {
4500 // FIXME: VLD1 etc with standard alignment is legal.
4504 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
4509 switch (VT.getSimpleVT().SimpleTy) {
4510 default: return false;
4525 if ((V & (Scale - 1)) != 0)
4528 return V == (V & ((1LL << 5) - 1));
4531 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
4532 const ARMSubtarget *Subtarget) {
4539 switch (VT.getSimpleVT().SimpleTy) {
4540 default: return false;
4545 // + imm12 or - imm8
4547 return V == (V & ((1LL << 8) - 1));
4548 return V == (V & ((1LL << 12) - 1));
4551 // Same as ARM mode. FIXME: NEON?
4552 if (!Subtarget->hasVFP2())
4557 return V == (V & ((1LL << 8) - 1));
4561 /// isLegalAddressImmediate - Return true if the integer value can be used
4562 /// as the offset of the target addressing mode for load / store of the
4564 static bool isLegalAddressImmediate(int64_t V, EVT VT,
4565 const ARMSubtarget *Subtarget) {
4572 if (Subtarget->isThumb1Only())
4573 return isLegalT1AddressImmediate(V, VT);
4574 else if (Subtarget->isThumb2())
4575 return isLegalT2AddressImmediate(V, VT, Subtarget);
4580 switch (VT.getSimpleVT().SimpleTy) {
4581 default: return false;
4586 return V == (V & ((1LL << 12) - 1));
4589 return V == (V & ((1LL << 8) - 1));
4592 if (!Subtarget->hasVFP2()) // FIXME: NEON?
4597 return V == (V & ((1LL << 8) - 1));
4601 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
4603 int Scale = AM.Scale;
4607 switch (VT.getSimpleVT().SimpleTy) {
4608 default: return false;
4617 return Scale == 2 || Scale == 4 || Scale == 8;
4620 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4624 // Note, we allow "void" uses (basically, uses that aren't loads or
4625 // stores), because arm allows folding a scale into many arithmetic
4626 // operations. This should be made more precise and revisited later.
4628 // Allow r << imm, but the imm has to be a multiple of two.
4629 if (Scale & 1) return false;
4630 return isPowerOf2_32(Scale);
4634 /// isLegalAddressingMode - Return true if the addressing mode represented
4635 /// by AM is legal for this target, for a load/store of the specified type.
4636 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4637 const Type *Ty) const {
4638 EVT VT = getValueType(Ty, true);
4639 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
4642 // Can never fold addr of global into load/store.
4647 case 0: // no scale reg, must be "r+i" or "r", or "i".
4650 if (Subtarget->isThumb1Only())
4654 // ARM doesn't support any R+R*scale+imm addr modes.
4661 if (Subtarget->isThumb2())
4662 return isLegalT2ScaledAddressingMode(AM, VT);
4664 int Scale = AM.Scale;
4665 switch (VT.getSimpleVT().SimpleTy) {
4666 default: return false;
4670 if (Scale < 0) Scale = -Scale;
4674 return isPowerOf2_32(Scale & ~1);
4678 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
4683 // Note, we allow "void" uses (basically, uses that aren't loads or
4684 // stores), because arm allows folding a scale into many arithmetic
4685 // operations. This should be made more precise and revisited later.
4687 // Allow r << imm, but the imm has to be a multiple of two.
4688 if (Scale & 1) return false;
4689 return isPowerOf2_32(Scale);
4696 /// isLegalICmpImmediate - Return true if the specified immediate is legal
4697 /// icmp immediate, that is the target has icmp instructions which can compare
4698 /// a register against the immediate without having to materialize the
4699 /// immediate into a register.
4700 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
4701 if (!Subtarget->isThumb())
4702 return ARM_AM::getSOImmVal(Imm) != -1;
4703 if (Subtarget->isThumb2())
4704 return ARM_AM::getT2SOImmVal(Imm) != -1;
4705 return Imm >= 0 && Imm <= 255;
4708 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
4709 bool isSEXTLoad, SDValue &Base,
4710 SDValue &Offset, bool &isInc,
4711 SelectionDAG &DAG) {
4712 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4715 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
4717 Base = Ptr->getOperand(0);
4718 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4719 int RHSC = (int)RHS->getZExtValue();
4720 if (RHSC < 0 && RHSC > -256) {
4721 assert(Ptr->getOpcode() == ISD::ADD);
4723 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4727 isInc = (Ptr->getOpcode() == ISD::ADD);
4728 Offset = Ptr->getOperand(1);
4730 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
4732 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4733 int RHSC = (int)RHS->getZExtValue();
4734 if (RHSC < 0 && RHSC > -0x1000) {
4735 assert(Ptr->getOpcode() == ISD::ADD);
4737 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4738 Base = Ptr->getOperand(0);
4743 if (Ptr->getOpcode() == ISD::ADD) {
4745 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
4746 if (ShOpcVal != ARM_AM::no_shift) {
4747 Base = Ptr->getOperand(1);
4748 Offset = Ptr->getOperand(0);
4750 Base = Ptr->getOperand(0);
4751 Offset = Ptr->getOperand(1);
4756 isInc = (Ptr->getOpcode() == ISD::ADD);
4757 Base = Ptr->getOperand(0);
4758 Offset = Ptr->getOperand(1);
4762 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
4766 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
4767 bool isSEXTLoad, SDValue &Base,
4768 SDValue &Offset, bool &isInc,
4769 SelectionDAG &DAG) {
4770 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
4773 Base = Ptr->getOperand(0);
4774 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
4775 int RHSC = (int)RHS->getZExtValue();
4776 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
4777 assert(Ptr->getOpcode() == ISD::ADD);
4779 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
4781 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
4782 isInc = Ptr->getOpcode() == ISD::ADD;
4783 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
4791 /// getPreIndexedAddressParts - returns true by value, base pointer and
4792 /// offset pointer and addressing mode by reference if the node's address
4793 /// can be legally represented as pre-indexed load / store address.
4795 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
4797 ISD::MemIndexedMode &AM,
4798 SelectionDAG &DAG) const {
4799 if (Subtarget->isThumb1Only())
4804 bool isSEXTLoad = false;
4805 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4806 Ptr = LD->getBasePtr();
4807 VT = LD->getMemoryVT();
4808 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4809 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4810 Ptr = ST->getBasePtr();
4811 VT = ST->getMemoryVT();
4816 bool isLegal = false;
4817 if (Subtarget->isThumb2())
4818 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4819 Offset, isInc, DAG);
4821 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
4822 Offset, isInc, DAG);
4826 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
4830 /// getPostIndexedAddressParts - returns true by value, base pointer and
4831 /// offset pointer and addressing mode by reference if this node can be
4832 /// combined with a load / store to form a post-indexed load / store.
4833 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
4836 ISD::MemIndexedMode &AM,
4837 SelectionDAG &DAG) const {
4838 if (Subtarget->isThumb1Only())
4843 bool isSEXTLoad = false;
4844 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
4845 VT = LD->getMemoryVT();
4846 Ptr = LD->getBasePtr();
4847 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
4848 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
4849 VT = ST->getMemoryVT();
4850 Ptr = ST->getBasePtr();
4855 bool isLegal = false;
4856 if (Subtarget->isThumb2())
4857 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4860 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
4866 // Swap base ptr and offset to catch more post-index load / store when
4867 // it's legal. In Thumb2 mode, offset must be an immediate.
4868 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
4869 !Subtarget->isThumb2())
4870 std::swap(Base, Offset);
4872 // Post-indexed load / store update the base pointer.
4877 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
4881 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4885 const SelectionDAG &DAG,
4886 unsigned Depth) const {
4887 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4888 switch (Op.getOpcode()) {
4890 case ARMISD::CMOV: {
4891 // Bits are known zero/one if known on the LHS and RHS.
4892 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
4893 if (KnownZero == 0 && KnownOne == 0) return;
4895 APInt KnownZeroRHS, KnownOneRHS;
4896 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
4897 KnownZeroRHS, KnownOneRHS, Depth+1);
4898 KnownZero &= KnownZeroRHS;
4899 KnownOne &= KnownOneRHS;
4905 //===----------------------------------------------------------------------===//
4906 // ARM Inline Assembly Support
4907 //===----------------------------------------------------------------------===//
4909 /// getConstraintType - Given a constraint letter, return the type of
4910 /// constraint it is for this target.
4911 ARMTargetLowering::ConstraintType
4912 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
4913 if (Constraint.size() == 1) {
4914 switch (Constraint[0]) {
4916 case 'l': return C_RegisterClass;
4917 case 'w': return C_RegisterClass;
4920 return TargetLowering::getConstraintType(Constraint);
4923 std::pair<unsigned, const TargetRegisterClass*>
4924 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4926 if (Constraint.size() == 1) {
4927 // GCC ARM Constraint Letters
4928 switch (Constraint[0]) {
4930 if (Subtarget->isThumb())
4931 return std::make_pair(0U, ARM::tGPRRegisterClass);
4933 return std::make_pair(0U, ARM::GPRRegisterClass);
4935 return std::make_pair(0U, ARM::GPRRegisterClass);
4938 return std::make_pair(0U, ARM::SPRRegisterClass);
4939 if (VT.getSizeInBits() == 64)
4940 return std::make_pair(0U, ARM::DPRRegisterClass);
4941 if (VT.getSizeInBits() == 128)
4942 return std::make_pair(0U, ARM::QPRRegisterClass);
4946 if (StringRef("{cc}").equals_lower(Constraint))
4947 return std::make_pair(0U, ARM::CCRRegisterClass);
4949 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4952 std::vector<unsigned> ARMTargetLowering::
4953 getRegClassForInlineAsmConstraint(const std::string &Constraint,
4955 if (Constraint.size() != 1)
4956 return std::vector<unsigned>();
4958 switch (Constraint[0]) { // GCC ARM Constraint Letters
4961 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4962 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4965 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
4966 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
4967 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
4968 ARM::R12, ARM::LR, 0);
4971 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
4972 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
4973 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
4974 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
4975 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
4976 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
4977 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
4978 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
4979 if (VT.getSizeInBits() == 64)
4980 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
4981 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
4982 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
4983 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
4984 if (VT.getSizeInBits() == 128)
4985 return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
4986 ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
4990 return std::vector<unsigned>();
4993 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4994 /// vector. If it is invalid, don't add anything to Ops.
4995 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4998 std::vector<SDValue>&Ops,
4999 SelectionDAG &DAG) const {
5000 SDValue Result(0, 0);
5002 switch (Constraint) {
5004 case 'I': case 'J': case 'K': case 'L':
5005 case 'M': case 'N': case 'O':
5006 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5010 int64_t CVal64 = C->getSExtValue();
5011 int CVal = (int) CVal64;
5012 // None of these constraints allow values larger than 32 bits. Check
5013 // that the value fits in an int.
5017 switch (Constraint) {
5019 if (Subtarget->isThumb1Only()) {
5020 // This must be a constant between 0 and 255, for ADD
5022 if (CVal >= 0 && CVal <= 255)
5024 } else if (Subtarget->isThumb2()) {
5025 // A constant that can be used as an immediate value in a
5026 // data-processing instruction.
5027 if (ARM_AM::getT2SOImmVal(CVal) != -1)
5030 // A constant that can be used as an immediate value in a
5031 // data-processing instruction.
5032 if (ARM_AM::getSOImmVal(CVal) != -1)
5038 if (Subtarget->isThumb()) { // FIXME thumb2
5039 // This must be a constant between -255 and -1, for negated ADD
5040 // immediates. This can be used in GCC with an "n" modifier that
5041 // prints the negated value, for use with SUB instructions. It is
5042 // not useful otherwise but is implemented for compatibility.
5043 if (CVal >= -255 && CVal <= -1)
5046 // This must be a constant between -4095 and 4095. It is not clear
5047 // what this constraint is intended for. Implemented for
5048 // compatibility with GCC.
5049 if (CVal >= -4095 && CVal <= 4095)
5055 if (Subtarget->isThumb1Only()) {
5056 // A 32-bit value where only one byte has a nonzero value. Exclude
5057 // zero to match GCC. This constraint is used by GCC internally for
5058 // constants that can be loaded with a move/shift combination.
5059 // It is not useful otherwise but is implemented for compatibility.
5060 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
5062 } else if (Subtarget->isThumb2()) {
5063 // A constant whose bitwise inverse can be used as an immediate
5064 // value in a data-processing instruction. This can be used in GCC
5065 // with a "B" modifier that prints the inverted value, for use with
5066 // BIC and MVN instructions. It is not useful otherwise but is
5067 // implemented for compatibility.
5068 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
5071 // A constant whose bitwise inverse can be used as an immediate
5072 // value in a data-processing instruction. This can be used in GCC
5073 // with a "B" modifier that prints the inverted value, for use with
5074 // BIC and MVN instructions. It is not useful otherwise but is
5075 // implemented for compatibility.
5076 if (ARM_AM::getSOImmVal(~CVal) != -1)
5082 if (Subtarget->isThumb1Only()) {
5083 // This must be a constant between -7 and 7,
5084 // for 3-operand ADD/SUB immediate instructions.
5085 if (CVal >= -7 && CVal < 7)
5087 } else if (Subtarget->isThumb2()) {
5088 // A constant whose negation can be used as an immediate value in a
5089 // data-processing instruction. This can be used in GCC with an "n"
5090 // modifier that prints the negated value, for use with SUB
5091 // instructions. It is not useful otherwise but is implemented for
5093 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
5096 // A constant whose negation can be used as an immediate value in a
5097 // data-processing instruction. This can be used in GCC with an "n"
5098 // modifier that prints the negated value, for use with SUB
5099 // instructions. It is not useful otherwise but is implemented for
5101 if (ARM_AM::getSOImmVal(-CVal) != -1)
5107 if (Subtarget->isThumb()) { // FIXME thumb2
5108 // This must be a multiple of 4 between 0 and 1020, for
5109 // ADD sp + immediate.
5110 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
5113 // A power of two or a constant between 0 and 32. This is used in
5114 // GCC for the shift amount on shifted register operands, but it is
5115 // useful in general for any shift amounts.
5116 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
5122 if (Subtarget->isThumb()) { // FIXME thumb2
5123 // This must be a constant between 0 and 31, for shift amounts.
5124 if (CVal >= 0 && CVal <= 31)
5130 if (Subtarget->isThumb()) { // FIXME thumb2
5131 // This must be a multiple of 4 between -508 and 508, for
5132 // ADD/SUB sp = sp + immediate.
5133 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
5138 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5142 if (Result.getNode()) {
5143 Ops.push_back(Result);
5146 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
5151 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5152 // The ARM target isn't yet aware of offsets.
5156 int ARM::getVFPf32Imm(const APFloat &FPImm) {
5157 APInt Imm = FPImm.bitcastToAPInt();
5158 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
5159 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
5160 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
5162 // We can handle 4 bits of mantissa.
5163 // mantissa = (16+UInt(e:f:g:h))/16.
5164 if (Mantissa & 0x7ffff)
5167 if ((Mantissa & 0xf) != Mantissa)
5170 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5171 if (Exp < -3 || Exp > 4)
5173 Exp = ((Exp+3) & 0x7) ^ 4;
5175 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5178 int ARM::getVFPf64Imm(const APFloat &FPImm) {
5179 APInt Imm = FPImm.bitcastToAPInt();
5180 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
5181 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
5182 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
5184 // We can handle 4 bits of mantissa.
5185 // mantissa = (16+UInt(e:f:g:h))/16.
5186 if (Mantissa & 0xffffffffffffLL)
5189 if ((Mantissa & 0xf) != Mantissa)
5192 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
5193 if (Exp < -3 || Exp > 4)
5195 Exp = ((Exp+3) & 0x7) ^ 4;
5197 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
5200 /// isFPImmLegal - Returns true if the target can instruction select the
5201 /// specified FP immediate natively. If false, the legalizer will
5202 /// materialize the FP immediate as a load from a constant pool.
5203 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
5204 if (!Subtarget->hasVFP3())
5207 return ARM::getVFPf32Imm(Imm) != -1;
5209 return ARM::getVFPf64Imm(Imm) != -1;