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 #include "ARMISelLowering.h"
16 #include "ARMCallingConv.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMPerfectShuffle.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "ARMTargetObjectFile.h"
23 #include "MCTargetDesc/ARMAddressingModes.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/CodeGen/CallingConvLower.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GlobalValue.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/Type.h"
44 #include "llvm/MC/MCSectionMachO.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/ErrorHandling.h"
48 #include "llvm/Support/MathExtras.h"
49 #include "llvm/Target/TargetOptions.h"
53 #define DEBUG_TYPE "arm-isel"
55 STATISTIC(NumTailCalls, "Number of tail calls");
56 STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
57 STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");
60 EnableARMLongCalls("arm-long-calls", cl::Hidden,
61 cl::desc("Generate calls via indirect call instructions"),
65 ARMInterworking("arm-interworking", cl::Hidden,
66 cl::desc("Enable / disable ARM interworking (for debugging only)"),
70 class ARMCCState : public CCState {
72 ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
73 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
75 : CCState(CC, isVarArg, MF, locs, C) {
76 assert(((PC == Call) || (PC == Prologue)) &&
77 "ARMCCState users must specify whether their context is call"
78 "or prologue generation.");
84 // The APCS parameter registers.
85 static const MCPhysReg GPRArgRegs[] = {
86 ARM::R0, ARM::R1, ARM::R2, ARM::R3
89 void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
90 MVT PromotedBitwiseVT) {
91 if (VT != PromotedLdStVT) {
92 setOperationAction(ISD::LOAD, VT, Promote);
93 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);
95 setOperationAction(ISD::STORE, VT, Promote);
96 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
99 MVT ElemTy = VT.getVectorElementType();
100 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
101 setOperationAction(ISD::SETCC, VT, Custom);
102 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
103 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
104 if (ElemTy == MVT::i32) {
105 setOperationAction(ISD::SINT_TO_FP, VT, Custom);
106 setOperationAction(ISD::UINT_TO_FP, VT, Custom);
107 setOperationAction(ISD::FP_TO_SINT, VT, Custom);
108 setOperationAction(ISD::FP_TO_UINT, VT, Custom);
110 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
111 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
112 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
113 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
115 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
116 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
117 setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);
118 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
119 setOperationAction(ISD::SELECT, VT, Expand);
120 setOperationAction(ISD::SELECT_CC, VT, Expand);
121 setOperationAction(ISD::VSELECT, VT, Expand);
122 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
123 if (VT.isInteger()) {
124 setOperationAction(ISD::SHL, VT, Custom);
125 setOperationAction(ISD::SRA, VT, Custom);
126 setOperationAction(ISD::SRL, VT, Custom);
129 // Promote all bit-wise operations.
130 if (VT.isInteger() && VT != PromotedBitwiseVT) {
131 setOperationAction(ISD::AND, VT, Promote);
132 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
133 setOperationAction(ISD::OR, VT, Promote);
134 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT);
135 setOperationAction(ISD::XOR, VT, Promote);
136 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
139 // Neon does not support vector divide/remainder operations.
140 setOperationAction(ISD::SDIV, VT, Expand);
141 setOperationAction(ISD::UDIV, VT, Expand);
142 setOperationAction(ISD::FDIV, VT, Expand);
143 setOperationAction(ISD::SREM, VT, Expand);
144 setOperationAction(ISD::UREM, VT, Expand);
145 setOperationAction(ISD::FREM, VT, Expand);
148 void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
149 addRegisterClass(VT, &ARM::DPRRegClass);
150 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
153 void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
154 addRegisterClass(VT, &ARM::DPairRegClass);
155 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
158 static TargetLoweringObjectFile *createTLOF(const Triple &TT) {
159 if (TT.isOSBinFormatMachO())
160 return new TargetLoweringObjectFileMachO();
161 if (TT.isOSWindows())
162 return new TargetLoweringObjectFileCOFF();
163 return new ARMElfTargetObjectFile();
166 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
167 : TargetLowering(TM, createTLOF(Triple(TM.getTargetTriple()))) {
168 Subtarget = &TM.getSubtarget<ARMSubtarget>();
169 RegInfo = TM.getSubtargetImpl()->getRegisterInfo();
170 Itins = TM.getSubtargetImpl()->getInstrItineraryData();
172 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
174 if (Subtarget->isTargetMachO()) {
175 // Uses VFP for Thumb libfuncs if available.
176 if (Subtarget->isThumb() && Subtarget->hasVFP2() &&
177 Subtarget->hasARMOps() && !TM.Options.UseSoftFloat) {
178 // Single-precision floating-point arithmetic.
179 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
180 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
181 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
182 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
184 // Double-precision floating-point arithmetic.
185 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
186 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
187 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
188 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
190 // Single-precision comparisons.
191 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
192 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
193 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
194 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
195 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
196 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
197 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
198 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
200 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
201 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
202 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
203 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
204 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
205 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
206 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
207 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
209 // Double-precision comparisons.
210 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
211 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
212 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
213 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
214 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
215 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
216 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
217 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
219 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
220 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
221 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
222 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
223 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
224 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
225 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
226 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
228 // Floating-point to integer conversions.
229 // i64 conversions are done via library routines even when generating VFP
230 // instructions, so use the same ones.
231 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
232 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
233 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
234 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
236 // Conversions between floating types.
237 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
238 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
240 // Integer to floating-point conversions.
241 // i64 conversions are done via library routines even when generating VFP
242 // instructions, so use the same ones.
243 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
244 // e.g., __floatunsidf vs. __floatunssidfvfp.
245 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
246 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
247 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
248 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
252 // These libcalls are not available in 32-bit.
253 setLibcallName(RTLIB::SHL_I128, nullptr);
254 setLibcallName(RTLIB::SRL_I128, nullptr);
255 setLibcallName(RTLIB::SRA_I128, nullptr);
257 if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetMachO() &&
258 !Subtarget->isTargetWindows()) {
259 static const struct {
260 const RTLIB::Libcall Op;
261 const char * const Name;
262 const CallingConv::ID CC;
263 const ISD::CondCode Cond;
265 // Double-precision floating-point arithmetic helper functions
266 // RTABI chapter 4.1.2, Table 2
267 { RTLIB::ADD_F64, "__aeabi_dadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
268 { RTLIB::DIV_F64, "__aeabi_ddiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
269 { RTLIB::MUL_F64, "__aeabi_dmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
270 // FIXME: double __aeabi_drsub(double x, double y) (rsub)
271 { RTLIB::SUB_F64, "__aeabi_dsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
273 // Double-precision floating-point comparison helper functions
274 // RTABI chapter 4.1.2, Table 3
275 // FIXME: void __aeabi_cdcmpeq(double, double)
276 // FIXME: void __aeabi_cdcmple(double, double)
277 // FIXME: void __aeabi_cdrcmple(double, double)
278 { RTLIB::OEQ_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
279 { RTLIB::UNE_F64, "__aeabi_dcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
280 { RTLIB::OLT_F64, "__aeabi_dcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
281 { RTLIB::OLE_F64, "__aeabi_dcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
282 { RTLIB::OGE_F64, "__aeabi_dcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
283 { RTLIB::OGT_F64, "__aeabi_dcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
284 { RTLIB::UO_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
285 { RTLIB::O_F64, "__aeabi_dcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
287 // Single-precision floating-point arithmetic helper functions
288 // RTABI chapter 4.1.2, Table 4
289 { RTLIB::ADD_F32, "__aeabi_fadd", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
290 { RTLIB::DIV_F32, "__aeabi_fdiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
291 { RTLIB::MUL_F32, "__aeabi_fmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
292 // FIXME: void __aeabi_frsub(float x, float y)
293 { RTLIB::SUB_F32, "__aeabi_fsub", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
295 // Single-precision floating-point comparison helper functions
296 // RTABI chapter 4.1.2, Table 5
297 // FIXME: void __aeabi_cfcmpeq(float, float)
298 // FIXME: void __aeabi_cfcmple(float, float)
299 // FIXME: void __aeabi_cfrcmple(float, float)
300 { RTLIB::OEQ_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETNE },
301 { RTLIB::UNE_F32, "__aeabi_fcmpeq", CallingConv::ARM_AAPCS, ISD::SETEQ },
302 { RTLIB::OLT_F32, "__aeabi_fcmplt", CallingConv::ARM_AAPCS, ISD::SETNE },
303 { RTLIB::OLE_F32, "__aeabi_fcmple", CallingConv::ARM_AAPCS, ISD::SETNE },
304 { RTLIB::OGE_F32, "__aeabi_fcmpge", CallingConv::ARM_AAPCS, ISD::SETNE },
305 { RTLIB::OGT_F32, "__aeabi_fcmpgt", CallingConv::ARM_AAPCS, ISD::SETNE },
306 { RTLIB::UO_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETNE },
307 { RTLIB::O_F32, "__aeabi_fcmpun", CallingConv::ARM_AAPCS, ISD::SETEQ },
309 // Floating-point to integer conversions.
310 // RTABI chapter 4.1.2, Table 6
311 { RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
312 { RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
313 { RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
314 { RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
315 { RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
316 { RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
317 { RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
318 { RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
320 // Conversions between floating types.
321 // RTABI chapter 4.1.2, Table 7
322 { RTLIB::FPROUND_F64_F32, "__aeabi_d2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
323 { RTLIB::FPEXT_F32_F64, "__aeabi_f2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
324 // FIXME: float __aeabi_f2f(short)
325 // FIXME: float __aeabi_h2f_alt(short)
326 // FIXME: short __aeabi_f2h(float)
327 // FIXME: short __aeabi_f2h_alt(float)
328 { RTLIB::FPROUND_F64_F16, "__aeabi_d2h", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
329 // FIXME: short __aeabi_d2h_alt(double)
331 // Integer to floating-point conversions.
332 // RTABI chapter 4.1.2, Table 8
333 { RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
334 { RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
335 { RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
336 { RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
337 { RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
338 { RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
339 { RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
340 { RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
342 // Long long helper functions
343 // RTABI chapter 4.2, Table 9
344 { RTLIB::MUL_I64, "__aeabi_lmul", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
345 // FIXME: __aeabi_ldivmod is SDIVREM not SDIV; we should custom lower this
346 { RTLIB::SDIV_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
347 { RTLIB::SDIVREM_I64, "__aeabi_ldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
348 // FIXME: __aeabi_uldivmod is UDIVREM not UDIV; we should custom lower this
349 { RTLIB::UDIV_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
350 { RTLIB::UDIVREM_I64, "__aeabi_uldivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
351 { RTLIB::SHL_I64, "__aeabi_llsl", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
352 { RTLIB::SRL_I64, "__aeabi_llsr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
353 { RTLIB::SRA_I64, "__aeabi_lasr", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
354 // FIXME: int __aeabi_lcmp(long long, long long)
355 // FIXME: int __aeabi_ulcmp(unsigned long long, unsigned long long)
357 // Integer division functions
358 // RTABI chapter 4.3.1
359 { RTLIB::SDIV_I8, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
360 { RTLIB::SDIV_I16, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
361 { RTLIB::SDIV_I32, "__aeabi_idiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
362 { RTLIB::UDIV_I8, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
363 { RTLIB::UDIV_I16, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
364 { RTLIB::UDIV_I32, "__aeabi_uidiv", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
365 { RTLIB::SDIVREM_I8, "__aeabi_idivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
366 { RTLIB::SDIVREM_I16, "__aeabi_idivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
367 { RTLIB::SDIVREM_I32, "__aeabi_idivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
368 { RTLIB::UDIVREM_I8, "__aeabi_uidivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
369 { RTLIB::UDIVREM_I16, "__aeabi_uidivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
370 { RTLIB::UDIVREM_I32, "__aeabi_uidivmod", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
373 // RTABI chapter 4.3.4
374 // FIXME: void __aeabi_memcpy8(void *, const void *, size_t)
375 // FIXME: void __aeabi_memcpy4(void *, const void *, size_t)
376 { RTLIB::MEMCPY, "__aeabi_memcpy", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
377 // FIXME: void __aeabi_memmove8(void *, const void *, size_t)
378 // FIXME: void __aeabi_memmove4(void *, const void *, size_t)
379 { RTLIB::MEMMOVE, "__aeabi_memmove", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
380 // FIXME: void __aeabi_memset8(void *, size_t, int)
381 // FIXME: void __aeabi_memset4(void *, size_t, int)
382 { RTLIB::MEMSET, "__aeabi_memset", CallingConv::ARM_AAPCS, ISD::SETCC_INVALID },
383 // FIXME: void __aeabi_memclr8(void *, size_t)
384 // FIXME: void __aeabi_memclr4(void *, size_t)
385 // FIXME: void __aeabi_memclr(void *, size_t)
388 for (const auto &LC : LibraryCalls) {
389 setLibcallName(LC.Op, LC.Name);
390 setLibcallCallingConv(LC.Op, LC.CC);
391 if (LC.Cond != ISD::SETCC_INVALID)
392 setCmpLibcallCC(LC.Op, LC.Cond);
395 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
396 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
399 if (Subtarget->isTargetWindows()) {
400 static const struct {
401 const RTLIB::Libcall Op;
402 const char * const Name;
403 const CallingConv::ID CC;
405 { RTLIB::FPTOSINT_F32_I64, "__stoi64", CallingConv::ARM_AAPCS_VFP },
406 { RTLIB::FPTOSINT_F64_I64, "__dtoi64", CallingConv::ARM_AAPCS_VFP },
407 { RTLIB::FPTOUINT_F32_I64, "__stou64", CallingConv::ARM_AAPCS_VFP },
408 { RTLIB::FPTOUINT_F64_I64, "__dtou64", CallingConv::ARM_AAPCS_VFP },
409 { RTLIB::SINTTOFP_I64_F32, "__i64tos", CallingConv::ARM_AAPCS_VFP },
410 { RTLIB::SINTTOFP_I64_F64, "__i64tod", CallingConv::ARM_AAPCS_VFP },
411 { RTLIB::UINTTOFP_I64_F32, "__u64tos", CallingConv::ARM_AAPCS_VFP },
412 { RTLIB::UINTTOFP_I64_F64, "__u64tod", CallingConv::ARM_AAPCS_VFP },
415 for (const auto &LC : LibraryCalls) {
416 setLibcallName(LC.Op, LC.Name);
417 setLibcallCallingConv(LC.Op, LC.CC);
421 // Use divmod compiler-rt calls for iOS 5.0 and later.
422 if (Subtarget->getTargetTriple().isiOS() &&
423 !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
424 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
425 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
428 // The half <-> float conversion functions are always soft-float, but are
429 // needed for some targets which use a hard-float calling convention by
431 if (Subtarget->isAAPCS_ABI()) {
432 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_AAPCS);
433 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_AAPCS);
434 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_AAPCS);
436 setLibcallCallingConv(RTLIB::FPROUND_F32_F16, CallingConv::ARM_APCS);
437 setLibcallCallingConv(RTLIB::FPROUND_F64_F16, CallingConv::ARM_APCS);
438 setLibcallCallingConv(RTLIB::FPEXT_F16_F32, CallingConv::ARM_APCS);
441 if (Subtarget->isThumb1Only())
442 addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
444 addRegisterClass(MVT::i32, &ARM::GPRRegClass);
445 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
446 !Subtarget->isThumb1Only()) {
447 addRegisterClass(MVT::f32, &ARM::SPRRegClass);
448 addRegisterClass(MVT::f64, &ARM::DPRRegClass);
451 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
452 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
453 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
454 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
455 setTruncStoreAction((MVT::SimpleValueType)VT,
456 (MVT::SimpleValueType)InnerVT, Expand);
457 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
458 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
459 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
461 setOperationAction(ISD::MULHS, (MVT::SimpleValueType)VT, Expand);
462 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
463 setOperationAction(ISD::MULHU, (MVT::SimpleValueType)VT, Expand);
464 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
466 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
469 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
470 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
472 if (Subtarget->hasNEON()) {
473 addDRTypeForNEON(MVT::v2f32);
474 addDRTypeForNEON(MVT::v8i8);
475 addDRTypeForNEON(MVT::v4i16);
476 addDRTypeForNEON(MVT::v2i32);
477 addDRTypeForNEON(MVT::v1i64);
479 addQRTypeForNEON(MVT::v4f32);
480 addQRTypeForNEON(MVT::v2f64);
481 addQRTypeForNEON(MVT::v16i8);
482 addQRTypeForNEON(MVT::v8i16);
483 addQRTypeForNEON(MVT::v4i32);
484 addQRTypeForNEON(MVT::v2i64);
486 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
487 // neither Neon nor VFP support any arithmetic operations on it.
488 // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
489 // supported for v4f32.
490 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
491 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
492 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
493 // FIXME: Code duplication: FDIV and FREM are expanded always, see
494 // ARMTargetLowering::addTypeForNEON method for details.
495 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
496 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
497 // FIXME: Create unittest.
498 // In another words, find a way when "copysign" appears in DAG with vector
500 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
501 // FIXME: Code duplication: SETCC has custom operation action, see
502 // ARMTargetLowering::addTypeForNEON method for details.
503 setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
504 // FIXME: Create unittest for FNEG and for FABS.
505 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
506 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
507 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
508 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
509 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
510 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
511 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
512 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
513 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
514 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
515 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
516 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
517 // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
518 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
519 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
520 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
521 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
522 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
523 setOperationAction(ISD::FMA, MVT::v2f64, Expand);
525 setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
526 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
527 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
528 setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
529 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
530 setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
531 setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
532 setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
533 setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
534 setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
535 setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
536 setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
537 setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
538 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
539 setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
541 // Mark v2f32 intrinsics.
542 setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
543 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
544 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
545 setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
546 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
547 setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
548 setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
549 setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
550 setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
551 setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
552 setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
553 setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
554 setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
555 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
556 setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
558 // Neon does not support some operations on v1i64 and v2i64 types.
559 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
560 // Custom handling for some quad-vector types to detect VMULL.
561 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
562 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
563 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
564 // Custom handling for some vector types to avoid expensive expansions
565 setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
566 setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
567 setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
568 setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
569 setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
570 setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
571 // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
572 // a destination type that is wider than the source, and nor does
573 // it have a FP_TO_[SU]INT instruction with a narrower destination than
575 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
576 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
577 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
578 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
580 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand);
581 setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
583 // NEON does not have single instruction CTPOP for vectors with element
584 // types wider than 8-bits. However, custom lowering can leverage the
585 // v8i8/v16i8 vcnt instruction.
586 setOperationAction(ISD::CTPOP, MVT::v2i32, Custom);
587 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
588 setOperationAction(ISD::CTPOP, MVT::v4i16, Custom);
589 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
591 // NEON only has FMA instructions as of VFP4.
592 if (!Subtarget->hasVFP4()) {
593 setOperationAction(ISD::FMA, MVT::v2f32, Expand);
594 setOperationAction(ISD::FMA, MVT::v4f32, Expand);
597 setTargetDAGCombine(ISD::INTRINSIC_VOID);
598 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
599 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
600 setTargetDAGCombine(ISD::SHL);
601 setTargetDAGCombine(ISD::SRL);
602 setTargetDAGCombine(ISD::SRA);
603 setTargetDAGCombine(ISD::SIGN_EXTEND);
604 setTargetDAGCombine(ISD::ZERO_EXTEND);
605 setTargetDAGCombine(ISD::ANY_EXTEND);
606 setTargetDAGCombine(ISD::SELECT_CC);
607 setTargetDAGCombine(ISD::BUILD_VECTOR);
608 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
609 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
610 setTargetDAGCombine(ISD::STORE);
611 setTargetDAGCombine(ISD::FP_TO_SINT);
612 setTargetDAGCombine(ISD::FP_TO_UINT);
613 setTargetDAGCombine(ISD::FDIV);
615 // It is legal to extload from v4i8 to v4i16 or v4i32.
616 MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
617 MVT::v4i16, MVT::v2i16,
619 for (unsigned i = 0; i < 6; ++i) {
620 setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
621 setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
622 setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
626 // ARM and Thumb2 support UMLAL/SMLAL.
627 if (!Subtarget->isThumb1Only())
628 setTargetDAGCombine(ISD::ADDC);
630 if (Subtarget->isFPOnlySP()) {
631 // When targetting a floating-point unit with only single-precision
632 // operations, f64 is legal for the few double-precision instructions which
633 // are present However, no double-precision operations other than moves,
634 // loads and stores are provided by the hardware.
635 setOperationAction(ISD::FADD, MVT::f64, Expand);
636 setOperationAction(ISD::FSUB, MVT::f64, Expand);
637 setOperationAction(ISD::FMUL, MVT::f64, Expand);
638 setOperationAction(ISD::FMA, MVT::f64, Expand);
639 setOperationAction(ISD::FDIV, MVT::f64, Expand);
640 setOperationAction(ISD::FREM, MVT::f64, Expand);
641 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
642 setOperationAction(ISD::FGETSIGN, MVT::f64, Expand);
643 setOperationAction(ISD::FNEG, MVT::f64, Expand);
644 setOperationAction(ISD::FABS, MVT::f64, Expand);
645 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
646 setOperationAction(ISD::FSIN, MVT::f64, Expand);
647 setOperationAction(ISD::FCOS, MVT::f64, Expand);
648 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
649 setOperationAction(ISD::FPOW, MVT::f64, Expand);
650 setOperationAction(ISD::FLOG, MVT::f64, Expand);
651 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
652 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
653 setOperationAction(ISD::FEXP, MVT::f64, Expand);
654 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
655 setOperationAction(ISD::FCEIL, MVT::f64, Expand);
656 setOperationAction(ISD::FTRUNC, MVT::f64, Expand);
657 setOperationAction(ISD::FRINT, MVT::f64, Expand);
658 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand);
659 setOperationAction(ISD::FFLOOR, MVT::f64, Expand);
660 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
661 setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
664 computeRegisterProperties();
666 // ARM does not have floating-point extending loads.
667 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
668 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
670 // ... or truncating stores
671 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
672 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
673 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
675 // ARM does not have i1 sign extending load.
676 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
678 // ARM supports all 4 flavors of integer indexed load / store.
679 if (!Subtarget->isThumb1Only()) {
680 for (unsigned im = (unsigned)ISD::PRE_INC;
681 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
682 setIndexedLoadAction(im, MVT::i1, Legal);
683 setIndexedLoadAction(im, MVT::i8, Legal);
684 setIndexedLoadAction(im, MVT::i16, Legal);
685 setIndexedLoadAction(im, MVT::i32, Legal);
686 setIndexedStoreAction(im, MVT::i1, Legal);
687 setIndexedStoreAction(im, MVT::i8, Legal);
688 setIndexedStoreAction(im, MVT::i16, Legal);
689 setIndexedStoreAction(im, MVT::i32, Legal);
693 setOperationAction(ISD::SADDO, MVT::i32, Custom);
694 setOperationAction(ISD::UADDO, MVT::i32, Custom);
695 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
696 setOperationAction(ISD::USUBO, MVT::i32, Custom);
698 // i64 operation support.
699 setOperationAction(ISD::MUL, MVT::i64, Expand);
700 setOperationAction(ISD::MULHU, MVT::i32, Expand);
701 if (Subtarget->isThumb1Only()) {
702 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
703 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
705 if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
706 || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
707 setOperationAction(ISD::MULHS, MVT::i32, Expand);
709 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
710 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
711 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
712 setOperationAction(ISD::SRL, MVT::i64, Custom);
713 setOperationAction(ISD::SRA, MVT::i64, Custom);
715 if (!Subtarget->isThumb1Only()) {
716 // FIXME: We should do this for Thumb1 as well.
717 setOperationAction(ISD::ADDC, MVT::i32, Custom);
718 setOperationAction(ISD::ADDE, MVT::i32, Custom);
719 setOperationAction(ISD::SUBC, MVT::i32, Custom);
720 setOperationAction(ISD::SUBE, MVT::i32, Custom);
723 // ARM does not have ROTL.
724 setOperationAction(ISD::ROTL, MVT::i32, Expand);
725 setOperationAction(ISD::CTTZ, MVT::i32, Custom);
726 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
727 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
728 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
730 // These just redirect to CTTZ and CTLZ on ARM.
731 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
732 setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
734 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);
736 // Only ARMv6 has BSWAP.
737 if (!Subtarget->hasV6Ops())
738 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
740 if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
741 !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
742 // These are expanded into libcalls if the cpu doesn't have HW divider.
743 setOperationAction(ISD::SDIV, MVT::i32, Expand);
744 setOperationAction(ISD::UDIV, MVT::i32, Expand);
747 // FIXME: Also set divmod for SREM on EABI
748 setOperationAction(ISD::SREM, MVT::i32, Expand);
749 setOperationAction(ISD::UREM, MVT::i32, Expand);
750 if (!Subtarget->isTargetAEABI()) {
751 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
752 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
755 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
756 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
757 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
758 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
759 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
761 setOperationAction(ISD::TRAP, MVT::Other, Legal);
763 // Use the default implementation.
764 setOperationAction(ISD::VASTART, MVT::Other, Custom);
765 setOperationAction(ISD::VAARG, MVT::Other, Expand);
766 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
767 setOperationAction(ISD::VAEND, MVT::Other, Expand);
768 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
769 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
771 if (!Subtarget->isTargetMachO()) {
772 // Non-MachO platforms may return values in these registers via the
773 // personality function.
774 setExceptionPointerRegister(ARM::R0);
775 setExceptionSelectorRegister(ARM::R1);
778 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
779 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
781 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
783 // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
784 // the default expansion.
785 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only()) {
786 // ATOMIC_FENCE needs custom lowering; the others should have been expanded
787 // to ldrex/strex loops already.
788 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
790 // On v8, we have particularly efficient implementations of atomic fences
791 // if they can be combined with nearby atomic loads and stores.
792 if (!Subtarget->hasV8Ops()) {
793 // Automatically insert fences (dmb ish) around ATOMIC_SWAP etc.
794 setInsertFencesForAtomic(true);
797 // If there's anything we can use as a barrier, go through custom lowering
799 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other,
800 Subtarget->hasAnyDataBarrier() ? Custom : Expand);
802 // Set them all for expansion, which will force libcalls.
803 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
804 setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
805 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
806 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
807 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
808 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
809 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
810 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
811 setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
812 setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
813 setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
814 setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
815 // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
816 // Unordered/Monotonic case.
817 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
818 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
821 setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
823 // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
824 if (!Subtarget->hasV6Ops()) {
825 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
826 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
828 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
830 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
831 !Subtarget->isThumb1Only()) {
832 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
833 // iff target supports vfp2.
834 setOperationAction(ISD::BITCAST, MVT::i64, Custom);
835 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
838 // We want to custom lower some of our intrinsics.
839 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
840 if (Subtarget->isTargetDarwin()) {
841 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
842 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
843 setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
846 setOperationAction(ISD::SETCC, MVT::i32, Expand);
847 setOperationAction(ISD::SETCC, MVT::f32, Expand);
848 setOperationAction(ISD::SETCC, MVT::f64, Expand);
849 setOperationAction(ISD::SELECT, MVT::i32, Custom);
850 setOperationAction(ISD::SELECT, MVT::f32, Custom);
851 setOperationAction(ISD::SELECT, MVT::f64, Custom);
852 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
853 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
854 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
856 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
857 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
858 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
859 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
860 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
862 // We don't support sin/cos/fmod/copysign/pow
863 setOperationAction(ISD::FSIN, MVT::f64, Expand);
864 setOperationAction(ISD::FSIN, MVT::f32, Expand);
865 setOperationAction(ISD::FCOS, MVT::f32, Expand);
866 setOperationAction(ISD::FCOS, MVT::f64, Expand);
867 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
868 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
869 setOperationAction(ISD::FREM, MVT::f64, Expand);
870 setOperationAction(ISD::FREM, MVT::f32, Expand);
871 if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
872 !Subtarget->isThumb1Only()) {
873 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
874 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
876 setOperationAction(ISD::FPOW, MVT::f64, Expand);
877 setOperationAction(ISD::FPOW, MVT::f32, Expand);
879 if (!Subtarget->hasVFP4()) {
880 setOperationAction(ISD::FMA, MVT::f64, Expand);
881 setOperationAction(ISD::FMA, MVT::f32, Expand);
884 // Various VFP goodness
885 if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
886 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
887 if (Subtarget->hasVFP2()) {
888 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
889 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
890 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
891 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
894 // v8 adds f64 <-> f16 conversion. Before that it should be expanded.
895 if (!Subtarget->hasV8Ops()) {
896 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
897 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
900 // fp16 is a special v7 extension that adds f16 <-> f32 conversions.
901 if (!Subtarget->hasFP16()) {
902 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
903 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
907 // Combine sin / cos into one node or libcall if possible.
908 if (Subtarget->hasSinCos()) {
909 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
910 setLibcallName(RTLIB::SINCOS_F64, "sincos");
911 if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
912 // For iOS, we don't want to the normal expansion of a libcall to
913 // sincos. We want to issue a libcall to __sincos_stret.
914 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
915 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
919 // ARMv8 implements a lot of rounding-like FP operations.
920 if (Subtarget->hasV8Ops()) {
921 static MVT RoundingTypes[] = {MVT::f32, MVT::f64};
922 for (const auto Ty : RoundingTypes) {
923 setOperationAction(ISD::FFLOOR, Ty, Legal);
924 setOperationAction(ISD::FCEIL, Ty, Legal);
925 setOperationAction(ISD::FROUND, Ty, Legal);
926 setOperationAction(ISD::FTRUNC, Ty, Legal);
927 setOperationAction(ISD::FNEARBYINT, Ty, Legal);
928 setOperationAction(ISD::FRINT, Ty, Legal);
931 // We have target-specific dag combine patterns for the following nodes:
932 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
933 setTargetDAGCombine(ISD::ADD);
934 setTargetDAGCombine(ISD::SUB);
935 setTargetDAGCombine(ISD::MUL);
936 setTargetDAGCombine(ISD::AND);
937 setTargetDAGCombine(ISD::OR);
938 setTargetDAGCombine(ISD::XOR);
940 if (Subtarget->hasV6Ops())
941 setTargetDAGCombine(ISD::SRL);
943 setStackPointerRegisterToSaveRestore(ARM::SP);
945 if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
946 !Subtarget->hasVFP2())
947 setSchedulingPreference(Sched::RegPressure);
949 setSchedulingPreference(Sched::Hybrid);
951 //// temporary - rewrite interface to use type
952 MaxStoresPerMemset = 8;
953 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
954 MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
955 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
956 MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
957 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
959 // On ARM arguments smaller than 4 bytes are extended, so all arguments
960 // are at least 4 bytes aligned.
961 setMinStackArgumentAlignment(4);
963 // Prefer likely predicted branches to selects on out-of-order cores.
964 PredictableSelectIsExpensive = Subtarget->isLikeA9();
966 setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
969 // FIXME: It might make sense to define the representative register class as the
970 // nearest super-register that has a non-null superset. For example, DPR_VFP2 is
971 // a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
972 // SPR's representative would be DPR_VFP2. This should work well if register
973 // pressure tracking were modified such that a register use would increment the
974 // pressure of the register class's representative and all of it's super
975 // classes' representatives transitively. We have not implemented this because
976 // of the difficulty prior to coalescing of modeling operand register classes
977 // due to the common occurrence of cross class copies and subregister insertions
979 std::pair<const TargetRegisterClass*, uint8_t>
980 ARMTargetLowering::findRepresentativeClass(MVT VT) const{
981 const TargetRegisterClass *RRC = nullptr;
983 switch (VT.SimpleTy) {
985 return TargetLowering::findRepresentativeClass(VT);
986 // Use DPR as representative register class for all floating point
987 // and vector types. Since there are 32 SPR registers and 32 DPR registers so
988 // the cost is 1 for both f32 and f64.
989 case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
990 case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
991 RRC = &ARM::DPRRegClass;
992 // When NEON is used for SP, only half of the register file is available
993 // because operations that define both SP and DP results will be constrained
994 // to the VFP2 class (D0-D15). We currently model this constraint prior to
995 // coalescing by double-counting the SP regs. See the FIXME above.
996 if (Subtarget->useNEONForSinglePrecisionFP())
999 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1000 case MVT::v4f32: case MVT::v2f64:
1001 RRC = &ARM::DPRRegClass;
1005 RRC = &ARM::DPRRegClass;
1009 RRC = &ARM::DPRRegClass;
1013 return std::make_pair(RRC, Cost);
1016 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
1018 default: return nullptr;
1019 case ARMISD::Wrapper: return "ARMISD::Wrapper";
1020 case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
1021 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
1022 case ARMISD::CALL: return "ARMISD::CALL";
1023 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
1024 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
1025 case ARMISD::tCALL: return "ARMISD::tCALL";
1026 case ARMISD::BRCOND: return "ARMISD::BRCOND";
1027 case ARMISD::BR_JT: return "ARMISD::BR_JT";
1028 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
1029 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
1030 case ARMISD::INTRET_FLAG: return "ARMISD::INTRET_FLAG";
1031 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
1032 case ARMISD::CMP: return "ARMISD::CMP";
1033 case ARMISD::CMN: return "ARMISD::CMN";
1034 case ARMISD::CMPZ: return "ARMISD::CMPZ";
1035 case ARMISD::CMPFP: return "ARMISD::CMPFP";
1036 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
1037 case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
1038 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
1040 case ARMISD::CMOV: return "ARMISD::CMOV";
1042 case ARMISD::RBIT: return "ARMISD::RBIT";
1044 case ARMISD::FTOSI: return "ARMISD::FTOSI";
1045 case ARMISD::FTOUI: return "ARMISD::FTOUI";
1046 case ARMISD::SITOF: return "ARMISD::SITOF";
1047 case ARMISD::UITOF: return "ARMISD::UITOF";
1049 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
1050 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
1051 case ARMISD::RRX: return "ARMISD::RRX";
1053 case ARMISD::ADDC: return "ARMISD::ADDC";
1054 case ARMISD::ADDE: return "ARMISD::ADDE";
1055 case ARMISD::SUBC: return "ARMISD::SUBC";
1056 case ARMISD::SUBE: return "ARMISD::SUBE";
1058 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
1059 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
1061 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
1062 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
1064 case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
1066 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
1068 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
1070 case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
1072 case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
1074 case ARMISD::WIN__CHKSTK: return "ARMISD:::WIN__CHKSTK";
1076 case ARMISD::VCEQ: return "ARMISD::VCEQ";
1077 case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
1078 case ARMISD::VCGE: return "ARMISD::VCGE";
1079 case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
1080 case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
1081 case ARMISD::VCGEU: return "ARMISD::VCGEU";
1082 case ARMISD::VCGT: return "ARMISD::VCGT";
1083 case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
1084 case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
1085 case ARMISD::VCGTU: return "ARMISD::VCGTU";
1086 case ARMISD::VTST: return "ARMISD::VTST";
1088 case ARMISD::VSHL: return "ARMISD::VSHL";
1089 case ARMISD::VSHRs: return "ARMISD::VSHRs";
1090 case ARMISD::VSHRu: return "ARMISD::VSHRu";
1091 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
1092 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
1093 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
1094 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
1095 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
1096 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
1097 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
1098 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
1099 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
1100 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
1101 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
1102 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
1103 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
1104 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
1105 case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
1106 case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
1107 case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
1108 case ARMISD::VDUP: return "ARMISD::VDUP";
1109 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
1110 case ARMISD::VEXT: return "ARMISD::VEXT";
1111 case ARMISD::VREV64: return "ARMISD::VREV64";
1112 case ARMISD::VREV32: return "ARMISD::VREV32";
1113 case ARMISD::VREV16: return "ARMISD::VREV16";
1114 case ARMISD::VZIP: return "ARMISD::VZIP";
1115 case ARMISD::VUZP: return "ARMISD::VUZP";
1116 case ARMISD::VTRN: return "ARMISD::VTRN";
1117 case ARMISD::VTBL1: return "ARMISD::VTBL1";
1118 case ARMISD::VTBL2: return "ARMISD::VTBL2";
1119 case ARMISD::VMULLs: return "ARMISD::VMULLs";
1120 case ARMISD::VMULLu: return "ARMISD::VMULLu";
1121 case ARMISD::UMLAL: return "ARMISD::UMLAL";
1122 case ARMISD::SMLAL: return "ARMISD::SMLAL";
1123 case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
1124 case ARMISD::FMAX: return "ARMISD::FMAX";
1125 case ARMISD::FMIN: return "ARMISD::FMIN";
1126 case ARMISD::VMAXNM: return "ARMISD::VMAX";
1127 case ARMISD::VMINNM: return "ARMISD::VMIN";
1128 case ARMISD::BFI: return "ARMISD::BFI";
1129 case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
1130 case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
1131 case ARMISD::VBSL: return "ARMISD::VBSL";
1132 case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
1133 case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
1134 case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
1135 case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
1136 case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
1137 case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
1138 case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
1139 case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1140 case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1141 case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1142 case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1143 case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1144 case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1145 case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1146 case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1147 case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1148 case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1149 case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1150 case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1151 case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1155 EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1156 if (!VT.isVector()) return getPointerTy();
1157 return VT.changeVectorElementTypeToInteger();
1160 /// getRegClassFor - Return the register class that should be used for the
1161 /// specified value type.
1162 const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
1163 // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1164 // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1165 // load / store 4 to 8 consecutive D registers.
1166 if (Subtarget->hasNEON()) {
1167 if (VT == MVT::v4i64)
1168 return &ARM::QQPRRegClass;
1169 if (VT == MVT::v8i64)
1170 return &ARM::QQQQPRRegClass;
1172 return TargetLowering::getRegClassFor(VT);
1175 // Create a fast isel object.
1177 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
1178 const TargetLibraryInfo *libInfo) const {
1179 return ARM::createFastISel(funcInfo, libInfo);
1182 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
1183 /// be used for loads / stores from the global.
1184 unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1185 return (Subtarget->isThumb1Only() ? 127 : 4095);
1188 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1189 unsigned NumVals = N->getNumValues();
1191 return Sched::RegPressure;
1193 for (unsigned i = 0; i != NumVals; ++i) {
1194 EVT VT = N->getValueType(i);
1195 if (VT == MVT::Glue || VT == MVT::Other)
1197 if (VT.isFloatingPoint() || VT.isVector())
1201 if (!N->isMachineOpcode())
1202 return Sched::RegPressure;
1204 // Load are scheduled for latency even if there instruction itinerary
1205 // is not available.
1206 const TargetInstrInfo *TII =
1207 getTargetMachine().getSubtargetImpl()->getInstrInfo();
1208 const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1210 if (MCID.getNumDefs() == 0)
1211 return Sched::RegPressure;
1212 if (!Itins->isEmpty() &&
1213 Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1216 return Sched::RegPressure;
1219 //===----------------------------------------------------------------------===//
1221 //===----------------------------------------------------------------------===//
1223 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1224 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1226 default: llvm_unreachable("Unknown condition code!");
1227 case ISD::SETNE: return ARMCC::NE;
1228 case ISD::SETEQ: return ARMCC::EQ;
1229 case ISD::SETGT: return ARMCC::GT;
1230 case ISD::SETGE: return ARMCC::GE;
1231 case ISD::SETLT: return ARMCC::LT;
1232 case ISD::SETLE: return ARMCC::LE;
1233 case ISD::SETUGT: return ARMCC::HI;
1234 case ISD::SETUGE: return ARMCC::HS;
1235 case ISD::SETULT: return ARMCC::LO;
1236 case ISD::SETULE: return ARMCC::LS;
1240 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1241 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1242 ARMCC::CondCodes &CondCode2) {
1243 CondCode2 = ARMCC::AL;
1245 default: llvm_unreachable("Unknown FP condition!");
1247 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1249 case ISD::SETOGT: CondCode = ARMCC::GT; break;
1251 case ISD::SETOGE: CondCode = ARMCC::GE; break;
1252 case ISD::SETOLT: CondCode = ARMCC::MI; break;
1253 case ISD::SETOLE: CondCode = ARMCC::LS; break;
1254 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1255 case ISD::SETO: CondCode = ARMCC::VC; break;
1256 case ISD::SETUO: CondCode = ARMCC::VS; break;
1257 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1258 case ISD::SETUGT: CondCode = ARMCC::HI; break;
1259 case ISD::SETUGE: CondCode = ARMCC::PL; break;
1261 case ISD::SETULT: CondCode = ARMCC::LT; break;
1263 case ISD::SETULE: CondCode = ARMCC::LE; break;
1265 case ISD::SETUNE: CondCode = ARMCC::NE; break;
1269 //===----------------------------------------------------------------------===//
1270 // Calling Convention Implementation
1271 //===----------------------------------------------------------------------===//
1273 #include "ARMGenCallingConv.inc"
1275 /// getEffectiveCallingConv - Get the effective calling convention, taking into
1276 /// account presence of floating point hardware and calling convention
1277 /// limitations, such as support for variadic functions.
1279 ARMTargetLowering::getEffectiveCallingConv(CallingConv::ID CC,
1280 bool isVarArg) const {
1283 llvm_unreachable("Unsupported calling convention");
1284 case CallingConv::ARM_AAPCS:
1285 case CallingConv::ARM_APCS:
1286 case CallingConv::GHC:
1288 case CallingConv::ARM_AAPCS_VFP:
1289 return isVarArg ? CallingConv::ARM_AAPCS : CallingConv::ARM_AAPCS_VFP;
1290 case CallingConv::C:
1291 if (!Subtarget->isAAPCS_ABI())
1292 return CallingConv::ARM_APCS;
1293 else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() &&
1294 getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1296 return CallingConv::ARM_AAPCS_VFP;
1298 return CallingConv::ARM_AAPCS;
1299 case CallingConv::Fast:
1300 if (!Subtarget->isAAPCS_ABI()) {
1301 if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1302 return CallingConv::Fast;
1303 return CallingConv::ARM_APCS;
1304 } else if (Subtarget->hasVFP2() && !Subtarget->isThumb1Only() && !isVarArg)
1305 return CallingConv::ARM_AAPCS_VFP;
1307 return CallingConv::ARM_AAPCS;
1311 /// CCAssignFnForNode - Selects the correct CCAssignFn for the given
1312 /// CallingConvention.
1313 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1315 bool isVarArg) const {
1316 switch (getEffectiveCallingConv(CC, isVarArg)) {
1318 llvm_unreachable("Unsupported calling convention");
1319 case CallingConv::ARM_APCS:
1320 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1321 case CallingConv::ARM_AAPCS:
1322 return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1323 case CallingConv::ARM_AAPCS_VFP:
1324 return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1325 case CallingConv::Fast:
1326 return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1327 case CallingConv::GHC:
1328 return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
1332 /// LowerCallResult - Lower the result values of a call into the
1333 /// appropriate copies out of appropriate physical registers.
1335 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1336 CallingConv::ID CallConv, bool isVarArg,
1337 const SmallVectorImpl<ISD::InputArg> &Ins,
1338 SDLoc dl, SelectionDAG &DAG,
1339 SmallVectorImpl<SDValue> &InVals,
1340 bool isThisReturn, SDValue ThisVal) const {
1342 // Assign locations to each value returned by this call.
1343 SmallVector<CCValAssign, 16> RVLocs;
1344 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1345 *DAG.getContext(), Call);
1346 CCInfo.AnalyzeCallResult(Ins,
1347 CCAssignFnForNode(CallConv, /* Return*/ true,
1350 // Copy all of the result registers out of their specified physreg.
1351 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1352 CCValAssign VA = RVLocs[i];
1354 // Pass 'this' value directly from the argument to return value, to avoid
1355 // reg unit interference
1356 if (i == 0 && isThisReturn) {
1357 assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
1358 "unexpected return calling convention register assignment");
1359 InVals.push_back(ThisVal);
1364 if (VA.needsCustom()) {
1365 // Handle f64 or half of a v2f64.
1366 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1368 Chain = Lo.getValue(1);
1369 InFlag = Lo.getValue(2);
1370 VA = RVLocs[++i]; // skip ahead to next loc
1371 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1373 Chain = Hi.getValue(1);
1374 InFlag = Hi.getValue(2);
1375 if (!Subtarget->isLittle())
1377 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1379 if (VA.getLocVT() == MVT::v2f64) {
1380 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1381 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1382 DAG.getConstant(0, MVT::i32));
1384 VA = RVLocs[++i]; // skip ahead to next loc
1385 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1386 Chain = Lo.getValue(1);
1387 InFlag = Lo.getValue(2);
1388 VA = RVLocs[++i]; // skip ahead to next loc
1389 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1390 Chain = Hi.getValue(1);
1391 InFlag = Hi.getValue(2);
1392 if (!Subtarget->isLittle())
1394 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1395 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1396 DAG.getConstant(1, MVT::i32));
1399 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1401 Chain = Val.getValue(1);
1402 InFlag = Val.getValue(2);
1405 switch (VA.getLocInfo()) {
1406 default: llvm_unreachable("Unknown loc info!");
1407 case CCValAssign::Full: break;
1408 case CCValAssign::BCvt:
1409 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1413 InVals.push_back(Val);
1419 /// LowerMemOpCallTo - Store the argument to the stack.
1421 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1422 SDValue StackPtr, SDValue Arg,
1423 SDLoc dl, SelectionDAG &DAG,
1424 const CCValAssign &VA,
1425 ISD::ArgFlagsTy Flags) const {
1426 unsigned LocMemOffset = VA.getLocMemOffset();
1427 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1428 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1429 return DAG.getStore(Chain, dl, Arg, PtrOff,
1430 MachinePointerInfo::getStack(LocMemOffset),
1434 void ARMTargetLowering::PassF64ArgInRegs(SDLoc dl, SelectionDAG &DAG,
1435 SDValue Chain, SDValue &Arg,
1436 RegsToPassVector &RegsToPass,
1437 CCValAssign &VA, CCValAssign &NextVA,
1439 SmallVectorImpl<SDValue> &MemOpChains,
1440 ISD::ArgFlagsTy Flags) const {
1442 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1443 DAG.getVTList(MVT::i32, MVT::i32), Arg);
1444 unsigned id = Subtarget->isLittle() ? 0 : 1;
1445 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd.getValue(id)));
1447 if (NextVA.isRegLoc())
1448 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1-id)));
1450 assert(NextVA.isMemLoc());
1451 if (!StackPtr.getNode())
1452 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1454 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1-id),
1460 /// LowerCall - Lowering a call into a callseq_start <-
1461 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1464 ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1465 SmallVectorImpl<SDValue> &InVals) const {
1466 SelectionDAG &DAG = CLI.DAG;
1468 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1469 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1470 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1471 SDValue Chain = CLI.Chain;
1472 SDValue Callee = CLI.Callee;
1473 bool &isTailCall = CLI.IsTailCall;
1474 CallingConv::ID CallConv = CLI.CallConv;
1475 bool doesNotRet = CLI.DoesNotReturn;
1476 bool isVarArg = CLI.IsVarArg;
1478 MachineFunction &MF = DAG.getMachineFunction();
1479 bool isStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1480 bool isThisReturn = false;
1481 bool isSibCall = false;
1483 // Disable tail calls if they're not supported.
1484 if (!Subtarget->supportsTailCall() || MF.getTarget().Options.DisableTailCalls)
1488 // Check if it's really possible to do a tail call.
1489 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1490 isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
1491 Outs, OutVals, Ins, DAG);
1492 if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
1493 report_fatal_error("failed to perform tail call elimination on a call "
1494 "site marked musttail");
1495 // We don't support GuaranteedTailCallOpt for ARM, only automatically
1496 // detected sibcalls.
1503 // Analyze operands of the call, assigning locations to each operand.
1504 SmallVector<CCValAssign, 16> ArgLocs;
1505 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1506 *DAG.getContext(), Call);
1507 CCInfo.AnalyzeCallOperands(Outs,
1508 CCAssignFnForNode(CallConv, /* Return*/ false,
1511 // Get a count of how many bytes are to be pushed on the stack.
1512 unsigned NumBytes = CCInfo.getNextStackOffset();
1514 // For tail calls, memory operands are available in our caller's stack.
1518 // Adjust the stack pointer for the new arguments...
1519 // These operations are automatically eliminated by the prolog/epilog pass
1521 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1524 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1526 RegsToPassVector RegsToPass;
1527 SmallVector<SDValue, 8> MemOpChains;
1529 // Walk the register/memloc assignments, inserting copies/loads. In the case
1530 // of tail call optimization, arguments are handled later.
1531 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1533 ++i, ++realArgIdx) {
1534 CCValAssign &VA = ArgLocs[i];
1535 SDValue Arg = OutVals[realArgIdx];
1536 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1537 bool isByVal = Flags.isByVal();
1539 // Promote the value if needed.
1540 switch (VA.getLocInfo()) {
1541 default: llvm_unreachable("Unknown loc info!");
1542 case CCValAssign::Full: break;
1543 case CCValAssign::SExt:
1544 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1546 case CCValAssign::ZExt:
1547 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1549 case CCValAssign::AExt:
1550 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1552 case CCValAssign::BCvt:
1553 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1557 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1558 if (VA.needsCustom()) {
1559 if (VA.getLocVT() == MVT::v2f64) {
1560 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1561 DAG.getConstant(0, MVT::i32));
1562 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1563 DAG.getConstant(1, MVT::i32));
1565 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1566 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1568 VA = ArgLocs[++i]; // skip ahead to next loc
1569 if (VA.isRegLoc()) {
1570 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1571 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1573 assert(VA.isMemLoc());
1575 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1576 dl, DAG, VA, Flags));
1579 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1580 StackPtr, MemOpChains, Flags);
1582 } else if (VA.isRegLoc()) {
1583 if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
1584 assert(VA.getLocVT() == MVT::i32 &&
1585 "unexpected calling convention register assignment");
1586 assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
1587 "unexpected use of 'returned'");
1588 isThisReturn = true;
1590 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1591 } else if (isByVal) {
1592 assert(VA.isMemLoc());
1593 unsigned offset = 0;
1595 // True if this byval aggregate will be split between registers
1597 unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
1598 unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();
1600 if (CurByValIdx < ByValArgsCount) {
1602 unsigned RegBegin, RegEnd;
1603 CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);
1605 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1607 for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
1608 SDValue Const = DAG.getConstant(4*i, MVT::i32);
1609 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1610 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1611 MachinePointerInfo(),
1612 false, false, false,
1613 DAG.InferPtrAlignment(AddArg));
1614 MemOpChains.push_back(Load.getValue(1));
1615 RegsToPass.push_back(std::make_pair(j, Load));
1618 // If parameter size outsides register area, "offset" value
1619 // helps us to calculate stack slot for remained part properly.
1620 offset = RegEnd - RegBegin;
1622 CCInfo.nextInRegsParam();
1625 if (Flags.getByValSize() > 4*offset) {
1626 unsigned LocMemOffset = VA.getLocMemOffset();
1627 SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1628 SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1630 SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1631 SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1632 SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1634 SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1636 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
1637 SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
1638 MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
1641 } else if (!isSibCall) {
1642 assert(VA.isMemLoc());
1644 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1645 dl, DAG, VA, Flags));
1649 if (!MemOpChains.empty())
1650 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1652 // Build a sequence of copy-to-reg nodes chained together with token chain
1653 // and flag operands which copy the outgoing args into the appropriate regs.
1655 // Tail call byval lowering might overwrite argument registers so in case of
1656 // tail call optimization the copies to registers are lowered later.
1658 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1659 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1660 RegsToPass[i].second, InFlag);
1661 InFlag = Chain.getValue(1);
1664 // For tail calls lower the arguments to the 'real' stack slot.
1666 // Force all the incoming stack arguments to be loaded from the stack
1667 // before any new outgoing arguments are stored to the stack, because the
1668 // outgoing stack slots may alias the incoming argument stack slots, and
1669 // the alias isn't otherwise explicit. This is slightly more conservative
1670 // than necessary, because it means that each store effectively depends
1671 // on every argument instead of just those arguments it would clobber.
1673 // Do not flag preceding copytoreg stuff together with the following stuff.
1675 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1676 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1677 RegsToPass[i].second, InFlag);
1678 InFlag = Chain.getValue(1);
1683 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1684 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1685 // node so that legalize doesn't hack it.
1686 bool isDirect = false;
1687 bool isARMFunc = false;
1688 bool isLocalARMFunc = false;
1689 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1691 if (EnableARMLongCalls) {
1692 assert((Subtarget->isTargetWindows() ||
1693 getTargetMachine().getRelocationModel() == Reloc::Static) &&
1694 "long-calls with non-static relocation model!");
1695 // Handle a global address or an external symbol. If it's not one of
1696 // those, the target's already in a register, so we don't need to do
1698 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1699 const GlobalValue *GV = G->getGlobal();
1700 // Create a constant pool entry for the callee address
1701 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1702 ARMConstantPoolValue *CPV =
1703 ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1705 // Get the address of the callee into a register
1706 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1707 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1708 Callee = DAG.getLoad(getPointerTy(), dl,
1709 DAG.getEntryNode(), CPAddr,
1710 MachinePointerInfo::getConstantPool(),
1711 false, false, false, 0);
1712 } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1713 const char *Sym = S->getSymbol();
1715 // Create a constant pool entry for the callee address
1716 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1717 ARMConstantPoolValue *CPV =
1718 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1719 ARMPCLabelIndex, 0);
1720 // Get the address of the callee into a register
1721 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1722 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1723 Callee = DAG.getLoad(getPointerTy(), dl,
1724 DAG.getEntryNode(), CPAddr,
1725 MachinePointerInfo::getConstantPool(),
1726 false, false, false, 0);
1728 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1729 const GlobalValue *GV = G->getGlobal();
1731 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1732 bool isStub = (isExt && Subtarget->isTargetMachO()) &&
1733 getTargetMachine().getRelocationModel() != Reloc::Static;
1734 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1735 // ARM call to a local ARM function is predicable.
1736 isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1737 // tBX takes a register source operand.
1738 if (isStub && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1739 assert(Subtarget->isTargetMachO() && "WrapperPIC use on non-MachO?");
1740 Callee = DAG.getNode(ARMISD::WrapperPIC, dl, getPointerTy(),
1741 DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
1742 0, ARMII::MO_NONLAZY));
1743 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
1744 MachinePointerInfo::getGOT(), false, false, true, 0);
1745 } else if (Subtarget->isTargetCOFF()) {
1746 assert(Subtarget->isTargetWindows() &&
1747 "Windows is the only supported COFF target");
1748 unsigned TargetFlags = GV->hasDLLImportStorageClass()
1749 ? ARMII::MO_DLLIMPORT
1750 : ARMII::MO_NO_FLAG;
1751 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), /*Offset=*/0,
1753 if (GV->hasDLLImportStorageClass())
1754 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
1755 DAG.getNode(ARMISD::Wrapper, dl, getPointerTy(),
1756 Callee), MachinePointerInfo::getGOT(),
1757 false, false, false, 0);
1759 // On ELF targets for PIC code, direct calls should go through the PLT
1760 unsigned OpFlags = 0;
1761 if (Subtarget->isTargetELF() &&
1762 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1763 OpFlags = ARMII::MO_PLT;
1764 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1766 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1768 bool isStub = Subtarget->isTargetMachO() &&
1769 getTargetMachine().getRelocationModel() != Reloc::Static;
1770 isARMFunc = !Subtarget->isThumb() || (isStub && !Subtarget->isMClass());
1771 // tBX takes a register source operand.
1772 const char *Sym = S->getSymbol();
1773 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1774 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1775 ARMConstantPoolValue *CPV =
1776 ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1777 ARMPCLabelIndex, 4);
1778 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1779 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1780 Callee = DAG.getLoad(getPointerTy(), dl,
1781 DAG.getEntryNode(), CPAddr,
1782 MachinePointerInfo::getConstantPool(),
1783 false, false, false, 0);
1784 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1785 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1786 getPointerTy(), Callee, PICLabel);
1788 unsigned OpFlags = 0;
1789 // On ELF targets for PIC code, direct calls should go through the PLT
1790 if (Subtarget->isTargetELF() &&
1791 getTargetMachine().getRelocationModel() == Reloc::PIC_)
1792 OpFlags = ARMII::MO_PLT;
1793 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1797 // FIXME: handle tail calls differently.
1799 bool HasMinSizeAttr = MF.getFunction()->getAttributes().hasAttribute(
1800 AttributeSet::FunctionIndex, Attribute::MinSize);
1801 if (Subtarget->isThumb()) {
1802 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1803 CallOpc = ARMISD::CALL_NOLINK;
1805 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1807 if (!isDirect && !Subtarget->hasV5TOps())
1808 CallOpc = ARMISD::CALL_NOLINK;
1809 else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
1810 // Emit regular call when code size is the priority
1812 // "mov lr, pc; b _foo" to avoid confusing the RSP
1813 CallOpc = ARMISD::CALL_NOLINK;
1815 CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1818 std::vector<SDValue> Ops;
1819 Ops.push_back(Chain);
1820 Ops.push_back(Callee);
1822 // Add argument registers to the end of the list so that they are known live
1824 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1825 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1826 RegsToPass[i].second.getValueType()));
1828 // Add a register mask operand representing the call-preserved registers.
1830 const uint32_t *Mask;
1831 const TargetRegisterInfo *TRI =
1832 getTargetMachine().getSubtargetImpl()->getRegisterInfo();
1833 const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
1835 // For 'this' returns, use the R0-preserving mask if applicable
1836 Mask = ARI->getThisReturnPreservedMask(CallConv);
1838 // Set isThisReturn to false if the calling convention is not one that
1839 // allows 'returned' to be modeled in this way, so LowerCallResult does
1840 // not try to pass 'this' straight through
1841 isThisReturn = false;
1842 Mask = ARI->getCallPreservedMask(CallConv);
1845 Mask = ARI->getCallPreservedMask(CallConv);
1847 assert(Mask && "Missing call preserved mask for calling convention");
1848 Ops.push_back(DAG.getRegisterMask(Mask));
1851 if (InFlag.getNode())
1852 Ops.push_back(InFlag);
1854 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1856 return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, Ops);
1858 // Returns a chain and a flag for retval copy to use.
1859 Chain = DAG.getNode(CallOpc, dl, NodeTys, Ops);
1860 InFlag = Chain.getValue(1);
1862 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1863 DAG.getIntPtrConstant(0, true), InFlag, dl);
1865 InFlag = Chain.getValue(1);
1867 // Handle result values, copying them out of physregs into vregs that we
1869 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
1870 InVals, isThisReturn,
1871 isThisReturn ? OutVals[0] : SDValue());
1874 /// HandleByVal - Every parameter *after* a byval parameter is passed
1875 /// on the stack. Remember the next parameter register to allocate,
1876 /// and then confiscate the rest of the parameter registers to insure
1879 ARMTargetLowering::HandleByVal(
1880 CCState *State, unsigned &size, unsigned Align) const {
1881 unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1882 assert((State->getCallOrPrologue() == Prologue ||
1883 State->getCallOrPrologue() == Call) &&
1884 "unhandled ParmContext");
1886 if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
1887 if (Subtarget->isAAPCS_ABI() && Align > 4) {
1888 unsigned AlignInRegs = Align / 4;
1889 unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
1890 for (unsigned i = 0; i < Waste; ++i)
1891 reg = State->AllocateReg(GPRArgRegs, 4);
1894 unsigned excess = 4 * (ARM::R4 - reg);
1896 // Special case when NSAA != SP and parameter size greater than size of
1897 // all remained GPR regs. In that case we can't split parameter, we must
1898 // send it to stack. We also must set NCRN to R4, so waste all
1899 // remained registers.
1900 const unsigned NSAAOffset = State->getNextStackOffset();
1901 if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
1902 while (State->AllocateReg(GPRArgRegs, 4))
1907 // First register for byval parameter is the first register that wasn't
1908 // allocated before this method call, so it would be "reg".
1909 // If parameter is small enough to be saved in range [reg, r4), then
1910 // the end (first after last) register would be reg + param-size-in-regs,
1911 // else parameter would be splitted between registers and stack,
1912 // end register would be r4 in this case.
1913 unsigned ByValRegBegin = reg;
1914 unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
1915 State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
1916 // Note, first register is allocated in the beginning of function already,
1917 // allocate remained amount of registers we need.
1918 for (unsigned i = reg+1; i != ByValRegEnd; ++i)
1919 State->AllocateReg(GPRArgRegs, 4);
1920 // A byval parameter that is split between registers and memory needs its
1921 // size truncated here.
1922 // In the case where the entire structure fits in registers, we set the
1923 // size in memory to zero.
1932 /// MatchingStackOffset - Return true if the given stack call argument is
1933 /// already available in the same position (relatively) of the caller's
1934 /// incoming argument stack.
1936 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1937 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1938 const TargetInstrInfo *TII) {
1939 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1941 if (Arg.getOpcode() == ISD::CopyFromReg) {
1942 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1943 if (!TargetRegisterInfo::isVirtualRegister(VR))
1945 MachineInstr *Def = MRI->getVRegDef(VR);
1948 if (!Flags.isByVal()) {
1949 if (!TII->isLoadFromStackSlot(Def, FI))
1954 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1955 if (Flags.isByVal())
1956 // ByVal argument is passed in as a pointer but it's now being
1957 // dereferenced. e.g.
1958 // define @foo(%struct.X* %A) {
1959 // tail call @bar(%struct.X* byval %A)
1962 SDValue Ptr = Ld->getBasePtr();
1963 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1966 FI = FINode->getIndex();
1970 assert(FI != INT_MAX);
1971 if (!MFI->isFixedObjectIndex(FI))
1973 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1976 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
1977 /// for tail call optimization. Targets which want to do tail call
1978 /// optimization should implement this function.
1980 ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1981 CallingConv::ID CalleeCC,
1983 bool isCalleeStructRet,
1984 bool isCallerStructRet,
1985 const SmallVectorImpl<ISD::OutputArg> &Outs,
1986 const SmallVectorImpl<SDValue> &OutVals,
1987 const SmallVectorImpl<ISD::InputArg> &Ins,
1988 SelectionDAG& DAG) const {
1989 const Function *CallerF = DAG.getMachineFunction().getFunction();
1990 CallingConv::ID CallerCC = CallerF->getCallingConv();
1991 bool CCMatch = CallerCC == CalleeCC;
1993 // Look for obvious safe cases to perform tail call optimization that do not
1994 // require ABI changes. This is what gcc calls sibcall.
1996 // Do not sibcall optimize vararg calls unless the call site is not passing
1998 if (isVarArg && !Outs.empty())
2001 // Exception-handling functions need a special set of instructions to indicate
2002 // a return to the hardware. Tail-calling another function would probably
2004 if (CallerF->hasFnAttribute("interrupt"))
2007 // Also avoid sibcall optimization if either caller or callee uses struct
2008 // return semantics.
2009 if (isCalleeStructRet || isCallerStructRet)
2012 // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
2013 // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
2014 // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
2015 // support in the assembler and linker to be used. This would need to be
2016 // fixed to fully support tail calls in Thumb1.
2018 // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
2019 // LR. This means if we need to reload LR, it takes an extra instructions,
2020 // which outweighs the value of the tail call; but here we don't know yet
2021 // whether LR is going to be used. Probably the right approach is to
2022 // generate the tail call here and turn it back into CALL/RET in
2023 // emitEpilogue if LR is used.
2025 // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
2026 // but we need to make sure there are enough registers; the only valid
2027 // registers are the 4 used for parameters. We don't currently do this
2029 if (Subtarget->isThumb1Only())
2032 // Externally-defined functions with weak linkage should not be
2033 // tail-called on ARM when the OS does not support dynamic
2034 // pre-emption of symbols, as the AAELF spec requires normal calls
2035 // to undefined weak functions to be replaced with a NOP or jump to the
2036 // next instruction. The behaviour of branch instructions in this
2037 // situation (as used for tail calls) is implementation-defined, so we
2038 // cannot rely on the linker replacing the tail call with a return.
2039 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2040 const GlobalValue *GV = G->getGlobal();
2041 if (GV->hasExternalWeakLinkage())
2045 // If the calling conventions do not match, then we'd better make sure the
2046 // results are returned in the same way as what the caller expects.
2048 SmallVector<CCValAssign, 16> RVLocs1;
2049 ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
2050 *DAG.getContext(), Call);
2051 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
2053 SmallVector<CCValAssign, 16> RVLocs2;
2054 ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
2055 *DAG.getContext(), Call);
2056 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
2058 if (RVLocs1.size() != RVLocs2.size())
2060 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2061 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2063 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2065 if (RVLocs1[i].isRegLoc()) {
2066 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2069 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2075 // If Caller's vararg or byval argument has been split between registers and
2076 // stack, do not perform tail call, since part of the argument is in caller's
2078 const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
2079 getInfo<ARMFunctionInfo>();
2080 if (AFI_Caller->getArgRegsSaveSize())
2083 // If the callee takes no arguments then go on to check the results of the
2085 if (!Outs.empty()) {
2086 // Check if stack adjustment is needed. For now, do not do this if any
2087 // argument is passed on the stack.
2088 SmallVector<CCValAssign, 16> ArgLocs;
2089 ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
2090 *DAG.getContext(), Call);
2091 CCInfo.AnalyzeCallOperands(Outs,
2092 CCAssignFnForNode(CalleeCC, false, isVarArg));
2093 if (CCInfo.getNextStackOffset()) {
2094 MachineFunction &MF = DAG.getMachineFunction();
2096 // Check if the arguments are already laid out in the right way as
2097 // the caller's fixed stack objects.
2098 MachineFrameInfo *MFI = MF.getFrameInfo();
2099 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2100 const TargetInstrInfo *TII =
2101 getTargetMachine().getSubtargetImpl()->getInstrInfo();
2102 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
2104 ++i, ++realArgIdx) {
2105 CCValAssign &VA = ArgLocs[i];
2106 EVT RegVT = VA.getLocVT();
2107 SDValue Arg = OutVals[realArgIdx];
2108 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
2109 if (VA.getLocInfo() == CCValAssign::Indirect)
2111 if (VA.needsCustom()) {
2112 // f64 and vector types are split into multiple registers or
2113 // register/stack-slot combinations. The types will not match
2114 // the registers; give up on memory f64 refs until we figure
2115 // out what to do about this.
2118 if (!ArgLocs[++i].isRegLoc())
2120 if (RegVT == MVT::v2f64) {
2121 if (!ArgLocs[++i].isRegLoc())
2123 if (!ArgLocs[++i].isRegLoc())
2126 } else if (!VA.isRegLoc()) {
2127 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2139 ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2140 MachineFunction &MF, bool isVarArg,
2141 const SmallVectorImpl<ISD::OutputArg> &Outs,
2142 LLVMContext &Context) const {
2143 SmallVector<CCValAssign, 16> RVLocs;
2144 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2145 return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
2149 static SDValue LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
2150 SDLoc DL, SelectionDAG &DAG) {
2151 const MachineFunction &MF = DAG.getMachineFunction();
2152 const Function *F = MF.getFunction();
2154 StringRef IntKind = F->getFnAttribute("interrupt").getValueAsString();
2156 // See ARM ARM v7 B1.8.3. On exception entry LR is set to a possibly offset
2157 // version of the "preferred return address". These offsets affect the return
2158 // instruction if this is a return from PL1 without hypervisor extensions.
2159 // IRQ/FIQ: +4 "subs pc, lr, #4"
2160 // SWI: 0 "subs pc, lr, #0"
2161 // ABORT: +4 "subs pc, lr, #4"
2162 // UNDEF: +4/+2 "subs pc, lr, #0"
2163 // UNDEF varies depending on where the exception came from ARM or Thumb
2164 // mode. Alongside GCC, we throw our hands up in disgust and pretend it's 0.
2167 if (IntKind == "" || IntKind == "IRQ" || IntKind == "FIQ" ||
2170 else if (IntKind == "SWI" || IntKind == "UNDEF")
2173 report_fatal_error("Unsupported interrupt attribute. If present, value "
2174 "must be one of: IRQ, FIQ, SWI, ABORT or UNDEF");
2176 RetOps.insert(RetOps.begin() + 1, DAG.getConstant(LROffset, MVT::i32, false));
2178 return DAG.getNode(ARMISD::INTRET_FLAG, DL, MVT::Other, RetOps);
2182 ARMTargetLowering::LowerReturn(SDValue Chain,
2183 CallingConv::ID CallConv, bool isVarArg,
2184 const SmallVectorImpl<ISD::OutputArg> &Outs,
2185 const SmallVectorImpl<SDValue> &OutVals,
2186 SDLoc dl, SelectionDAG &DAG) const {
2188 // CCValAssign - represent the assignment of the return value to a location.
2189 SmallVector<CCValAssign, 16> RVLocs;
2191 // CCState - Info about the registers and stack slots.
2192 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2193 *DAG.getContext(), Call);
2195 // Analyze outgoing return values.
2196 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
2200 SmallVector<SDValue, 4> RetOps;
2201 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2202 bool isLittleEndian = Subtarget->isLittle();
2204 MachineFunction &MF = DAG.getMachineFunction();
2205 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2206 AFI->setReturnRegsCount(RVLocs.size());
2208 // Copy the result values into the output registers.
2209 for (unsigned i = 0, realRVLocIdx = 0;
2211 ++i, ++realRVLocIdx) {
2212 CCValAssign &VA = RVLocs[i];
2213 assert(VA.isRegLoc() && "Can only return in registers!");
2215 SDValue Arg = OutVals[realRVLocIdx];
2217 switch (VA.getLocInfo()) {
2218 default: llvm_unreachable("Unknown loc info!");
2219 case CCValAssign::Full: break;
2220 case CCValAssign::BCvt:
2221 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
2225 if (VA.needsCustom()) {
2226 if (VA.getLocVT() == MVT::v2f64) {
2227 // Extract the first half and return it in two registers.
2228 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2229 DAG.getConstant(0, MVT::i32));
2230 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
2231 DAG.getVTList(MVT::i32, MVT::i32), Half);
2233 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2234 HalfGPRs.getValue(isLittleEndian ? 0 : 1),
2236 Flag = Chain.getValue(1);
2237 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2238 VA = RVLocs[++i]; // skip ahead to next loc
2239 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2240 HalfGPRs.getValue(isLittleEndian ? 1 : 0),
2242 Flag = Chain.getValue(1);
2243 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2244 VA = RVLocs[++i]; // skip ahead to next loc
2246 // Extract the 2nd half and fall through to handle it as an f64 value.
2247 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
2248 DAG.getConstant(1, MVT::i32));
2250 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
2252 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
2253 DAG.getVTList(MVT::i32, MVT::i32), Arg);
2254 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2255 fmrrd.getValue(isLittleEndian ? 0 : 1),
2257 Flag = Chain.getValue(1);
2258 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2259 VA = RVLocs[++i]; // skip ahead to next loc
2260 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
2261 fmrrd.getValue(isLittleEndian ? 1 : 0),
2264 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
2266 // Guarantee that all emitted copies are
2267 // stuck together, avoiding something bad.
2268 Flag = Chain.getValue(1);
2269 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2272 // Update chain and glue.
2275 RetOps.push_back(Flag);
2277 // CPUs which aren't M-class use a special sequence to return from
2278 // exceptions (roughly, any instruction setting pc and cpsr simultaneously,
2279 // though we use "subs pc, lr, #N").
2281 // M-class CPUs actually use a normal return sequence with a special
2282 // (hardware-provided) value in LR, so the normal code path works.
2283 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt") &&
2284 !Subtarget->isMClass()) {
2285 if (Subtarget->isThumb1Only())
2286 report_fatal_error("interrupt attribute is not supported in Thumb1");
2287 return LowerInterruptReturn(RetOps, dl, DAG);
2290 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, RetOps);
2293 bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2294 if (N->getNumValues() != 1)
2296 if (!N->hasNUsesOfValue(1, 0))
2299 SDValue TCChain = Chain;
2300 SDNode *Copy = *N->use_begin();
2301 if (Copy->getOpcode() == ISD::CopyToReg) {
2302 // If the copy has a glue operand, we conservatively assume it isn't safe to
2303 // perform a tail call.
2304 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2306 TCChain = Copy->getOperand(0);
2307 } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
2308 SDNode *VMov = Copy;
2309 // f64 returned in a pair of GPRs.
2310 SmallPtrSet<SDNode*, 2> Copies;
2311 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2313 if (UI->getOpcode() != ISD::CopyToReg)
2317 if (Copies.size() > 2)
2320 for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
2322 SDValue UseChain = UI->getOperand(0);
2323 if (Copies.count(UseChain.getNode()))
2330 } else if (Copy->getOpcode() == ISD::BITCAST) {
2331 // f32 returned in a single GPR.
2332 if (!Copy->hasOneUse())
2334 Copy = *Copy->use_begin();
2335 if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
2337 TCChain = Copy->getOperand(0);
2342 bool HasRet = false;
2343 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2345 if (UI->getOpcode() != ARMISD::RET_FLAG &&
2346 UI->getOpcode() != ARMISD::INTRET_FLAG)
2358 bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2359 if (!Subtarget->supportsTailCall())
2362 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2365 return !Subtarget->isThumb1Only();
2368 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2369 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2370 // one of the above mentioned nodes. It has to be wrapped because otherwise
2371 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2372 // be used to form addressing mode. These wrapped nodes will be selected
2374 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2375 EVT PtrVT = Op.getValueType();
2376 // FIXME there is no actual debug info here
2378 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2380 if (CP->isMachineConstantPoolEntry())
2381 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2382 CP->getAlignment());
2384 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2385 CP->getAlignment());
2386 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2389 unsigned ARMTargetLowering::getJumpTableEncoding() const {
2390 return MachineJumpTableInfo::EK_Inline;
2393 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2394 SelectionDAG &DAG) const {
2395 MachineFunction &MF = DAG.getMachineFunction();
2396 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2397 unsigned ARMPCLabelIndex = 0;
2399 EVT PtrVT = getPointerTy();
2400 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2401 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2403 if (RelocM == Reloc::Static) {
2404 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2406 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2407 ARMPCLabelIndex = AFI->createPICLabelUId();
2408 ARMConstantPoolValue *CPV =
2409 ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2410 ARMCP::CPBlockAddress, PCAdj);
2411 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2413 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2414 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2415 MachinePointerInfo::getConstantPool(),
2416 false, false, false, 0);
2417 if (RelocM == Reloc::Static)
2419 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2420 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2423 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
2425 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2426 SelectionDAG &DAG) const {
2428 EVT PtrVT = getPointerTy();
2429 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2430 MachineFunction &MF = DAG.getMachineFunction();
2431 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2432 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2433 ARMConstantPoolValue *CPV =
2434 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2435 ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2436 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2437 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2438 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2439 MachinePointerInfo::getConstantPool(),
2440 false, false, false, 0);
2441 SDValue Chain = Argument.getValue(1);
2443 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2444 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2446 // call __tls_get_addr.
2449 Entry.Node = Argument;
2450 Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2451 Args.push_back(Entry);
2453 // FIXME: is there useful debug info available here?
2454 TargetLowering::CallLoweringInfo CLI(DAG);
2455 CLI.setDebugLoc(dl).setChain(Chain)
2456 .setCallee(CallingConv::C, Type::getInt32Ty(*DAG.getContext()),
2457 DAG.getExternalSymbol("__tls_get_addr", PtrVT), std::move(Args),
2460 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
2461 return CallResult.first;
2464 // Lower ISD::GlobalTLSAddress using the "initial exec" or
2465 // "local exec" model.
2467 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2469 TLSModel::Model model) const {
2470 const GlobalValue *GV = GA->getGlobal();
2473 SDValue Chain = DAG.getEntryNode();
2474 EVT PtrVT = getPointerTy();
2475 // Get the Thread Pointer
2476 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2478 if (model == TLSModel::InitialExec) {
2479 MachineFunction &MF = DAG.getMachineFunction();
2480 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2481 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2482 // Initial exec model.
2483 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2484 ARMConstantPoolValue *CPV =
2485 ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2486 ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2488 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2489 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2490 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2491 MachinePointerInfo::getConstantPool(),
2492 false, false, false, 0);
2493 Chain = Offset.getValue(1);
2495 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2496 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2498 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2499 MachinePointerInfo::getConstantPool(),
2500 false, false, false, 0);
2503 assert(model == TLSModel::LocalExec);
2504 ARMConstantPoolValue *CPV =
2505 ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2506 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2507 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2508 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2509 MachinePointerInfo::getConstantPool(),
2510 false, false, false, 0);
2513 // The address of the thread local variable is the add of the thread
2514 // pointer with the offset of the variable.
2515 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2519 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2520 // TODO: implement the "local dynamic" model
2521 assert(Subtarget->isTargetELF() &&
2522 "TLS not implemented for non-ELF targets");
2523 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2525 TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());
2528 case TLSModel::GeneralDynamic:
2529 case TLSModel::LocalDynamic:
2530 return LowerToTLSGeneralDynamicModel(GA, DAG);
2531 case TLSModel::InitialExec:
2532 case TLSModel::LocalExec:
2533 return LowerToTLSExecModels(GA, DAG, model);
2535 llvm_unreachable("bogus TLS model");
2538 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2539 SelectionDAG &DAG) const {
2540 EVT PtrVT = getPointerTy();
2542 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2543 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2544 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2545 ARMConstantPoolValue *CPV =
2546 ARMConstantPoolConstant::Create(GV,
2547 UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2548 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2549 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2550 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2552 MachinePointerInfo::getConstantPool(),
2553 false, false, false, 0);
2554 SDValue Chain = Result.getValue(1);
2555 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2556 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2558 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2559 MachinePointerInfo::getGOT(),
2560 false, false, false, 0);
2564 // If we have T2 ops, we can materialize the address directly via movt/movw
2565 // pair. This is always cheaper.
2566 if (Subtarget->useMovt(DAG.getMachineFunction())) {
2568 // FIXME: Once remat is capable of dealing with instructions with register
2569 // operands, expand this into two nodes.
2570 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2571 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2573 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2574 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2575 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2576 MachinePointerInfo::getConstantPool(),
2577 false, false, false, 0);
2581 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2582 SelectionDAG &DAG) const {
2583 EVT PtrVT = getPointerTy();
2585 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2586 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2588 if (Subtarget->useMovt(DAG.getMachineFunction()))
2591 // FIXME: Once remat is capable of dealing with instructions with register
2592 // operands, expand this into multiple nodes
2594 RelocM == Reloc::PIC_ ? ARMISD::WrapperPIC : ARMISD::Wrapper;
2596 SDValue G = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, ARMII::MO_NONLAZY);
2597 SDValue Result = DAG.getNode(Wrapper, dl, PtrVT, G);
2599 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2600 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2601 MachinePointerInfo::getGOT(), false, false, false, 0);
2605 SDValue ARMTargetLowering::LowerGlobalAddressWindows(SDValue Op,
2606 SelectionDAG &DAG) const {
2607 assert(Subtarget->isTargetWindows() && "non-Windows COFF is not supported");
2608 assert(Subtarget->useMovt(DAG.getMachineFunction()) &&
2609 "Windows on ARM expects to use movw/movt");
2611 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2612 const ARMII::TOF TargetFlags =
2613 (GV->hasDLLImportStorageClass() ? ARMII::MO_DLLIMPORT : ARMII::MO_NO_FLAG);
2614 EVT PtrVT = getPointerTy();
2620 // FIXME: Once remat is capable of dealing with instructions with register
2621 // operands, expand this into two nodes.
2622 Result = DAG.getNode(ARMISD::Wrapper, DL, PtrVT,
2623 DAG.getTargetGlobalAddress(GV, DL, PtrVT, /*Offset=*/0,
2625 if (GV->hasDLLImportStorageClass())
2626 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
2627 MachinePointerInfo::getGOT(), false, false, false, 0);
2631 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2632 SelectionDAG &DAG) const {
2633 assert(Subtarget->isTargetELF() &&
2634 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2635 MachineFunction &MF = DAG.getMachineFunction();
2636 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2637 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2638 EVT PtrVT = getPointerTy();
2640 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2641 ARMConstantPoolValue *CPV =
2642 ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2643 ARMPCLabelIndex, PCAdj);
2644 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2645 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2646 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2647 MachinePointerInfo::getConstantPool(),
2648 false, false, false, 0);
2649 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2650 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2654 ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2656 SDValue Val = DAG.getConstant(0, MVT::i32);
2657 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2658 DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2659 Op.getOperand(1), Val);
2663 ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2665 return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2666 Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2670 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2671 const ARMSubtarget *Subtarget) const {
2672 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2675 default: return SDValue(); // Don't custom lower most intrinsics.
2676 case Intrinsic::arm_rbit: {
2677 assert(Op.getOperand(1).getValueType() == MVT::i32 &&
2678 "RBIT intrinsic must have i32 type!");
2679 return DAG.getNode(ARMISD::RBIT, dl, MVT::i32, Op.getOperand(1));
2681 case Intrinsic::arm_thread_pointer: {
2682 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2683 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2685 case Intrinsic::eh_sjlj_lsda: {
2686 MachineFunction &MF = DAG.getMachineFunction();
2687 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2688 unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2689 EVT PtrVT = getPointerTy();
2690 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2692 unsigned PCAdj = (RelocM != Reloc::PIC_)
2693 ? 0 : (Subtarget->isThumb() ? 4 : 8);
2694 ARMConstantPoolValue *CPV =
2695 ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2696 ARMCP::CPLSDA, PCAdj);
2697 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2698 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2700 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2701 MachinePointerInfo::getConstantPool(),
2702 false, false, false, 0);
2704 if (RelocM == Reloc::PIC_) {
2705 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2706 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2710 case Intrinsic::arm_neon_vmulls:
2711 case Intrinsic::arm_neon_vmullu: {
2712 unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2713 ? ARMISD::VMULLs : ARMISD::VMULLu;
2714 return DAG.getNode(NewOpc, SDLoc(Op), Op.getValueType(),
2715 Op.getOperand(1), Op.getOperand(2));
2720 static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2721 const ARMSubtarget *Subtarget) {
2722 // FIXME: handle "fence singlethread" more efficiently.
2724 if (!Subtarget->hasDataBarrier()) {
2725 // Some ARMv6 cpus can support data barriers with an mcr instruction.
2726 // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2728 assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2729 "Unexpected ISD::ATOMIC_FENCE encountered. Should be libcall!");
2730 return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2731 DAG.getConstant(0, MVT::i32));
2734 ConstantSDNode *OrdN = cast<ConstantSDNode>(Op.getOperand(1));
2735 AtomicOrdering Ord = static_cast<AtomicOrdering>(OrdN->getZExtValue());
2736 unsigned Domain = ARM_MB::ISH;
2737 if (Subtarget->isMClass()) {
2738 // Only a full system barrier exists in the M-class architectures.
2739 Domain = ARM_MB::SY;
2740 } else if (Subtarget->isSwift() && Ord == Release) {
2741 // Swift happens to implement ISHST barriers in a way that's compatible with
2742 // Release semantics but weaker than ISH so we'd be fools not to use
2743 // it. Beware: other processors probably don't!
2744 Domain = ARM_MB::ISHST;
2747 return DAG.getNode(ISD::INTRINSIC_VOID, dl, MVT::Other, Op.getOperand(0),
2748 DAG.getConstant(Intrinsic::arm_dmb, MVT::i32),
2749 DAG.getConstant(Domain, MVT::i32));
2752 static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2753 const ARMSubtarget *Subtarget) {
2754 // ARM pre v5TE and Thumb1 does not have preload instructions.
2755 if (!(Subtarget->isThumb2() ||
2756 (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2757 // Just preserve the chain.
2758 return Op.getOperand(0);
2761 unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2763 (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2764 // ARMv7 with MP extension has PLDW.
2765 return Op.getOperand(0);
2767 unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2768 if (Subtarget->isThumb()) {
2770 isRead = ~isRead & 1;
2771 isData = ~isData & 1;
2774 return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2775 Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2776 DAG.getConstant(isData, MVT::i32));
2779 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2780 MachineFunction &MF = DAG.getMachineFunction();
2781 ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2783 // vastart just stores the address of the VarArgsFrameIndex slot into the
2784 // memory location argument.
2786 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2787 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2788 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2789 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2790 MachinePointerInfo(SV), false, false, 0);
2794 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2795 SDValue &Root, SelectionDAG &DAG,
2797 MachineFunction &MF = DAG.getMachineFunction();
2798 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2800 const TargetRegisterClass *RC;
2801 if (AFI->isThumb1OnlyFunction())
2802 RC = &ARM::tGPRRegClass;
2804 RC = &ARM::GPRRegClass;
2806 // Transform the arguments stored in physical registers into virtual ones.
2807 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2808 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2811 if (NextVA.isMemLoc()) {
2812 MachineFrameInfo *MFI = MF.getFrameInfo();
2813 int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2815 // Create load node to retrieve arguments from the stack.
2816 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2817 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2818 MachinePointerInfo::getFixedStack(FI),
2819 false, false, false, 0);
2821 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2822 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2824 if (!Subtarget->isLittle())
2825 std::swap (ArgValue, ArgValue2);
2826 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2830 ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2831 unsigned InRegsParamRecordIdx,
2833 unsigned &ArgRegsSize,
2834 unsigned &ArgRegsSaveSize)
2837 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2838 unsigned RBegin, REnd;
2839 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2840 NumGPRs = REnd - RBegin;
2842 unsigned int firstUnalloced;
2843 firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2844 sizeof(GPRArgRegs) /
2845 sizeof(GPRArgRegs[0]));
2846 NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2849 unsigned Align = MF.getTarget()
2851 ->getFrameLowering()
2852 ->getStackAlignment();
2853 ArgRegsSize = NumGPRs * 4;
2855 // If parameter is split between stack and GPRs...
2856 if (NumGPRs && Align > 4 &&
2857 (ArgRegsSize < ArgSize ||
2858 InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
2859 // Add padding for part of param recovered from GPRs. For example,
2860 // if Align == 8, its last byte must be at address K*8 - 1.
2861 // We need to do it, since remained (stack) part of parameter has
2862 // stack alignment, and we need to "attach" "GPRs head" without gaps
2865 // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
2866 // [ [padding] [GPRs head] ] [ Tail passed via stack ....
2868 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2870 OffsetToAlignment(ArgRegsSize + AFI->getArgRegsSaveSize(), Align);
2871 ArgRegsSaveSize = ArgRegsSize + Padding;
2873 // We don't need to extend regs save size for byval parameters if they
2874 // are passed via GPRs only.
2875 ArgRegsSaveSize = ArgRegsSize;
2878 // The remaining GPRs hold either the beginning of variable-argument
2879 // data, or the beginning of an aggregate passed by value (usually
2880 // byval). Either way, we allocate stack slots adjacent to the data
2881 // provided by our caller, and store the unallocated registers there.
2882 // If this is a variadic function, the va_list pointer will begin with
2883 // these values; otherwise, this reassembles a (byval) structure that
2884 // was split between registers and memory.
2885 // Return: The frame index registers were stored into.
2887 ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
2888 SDLoc dl, SDValue &Chain,
2889 const Value *OrigArg,
2890 unsigned InRegsParamRecordIdx,
2891 unsigned OffsetFromOrigArg,
2895 unsigned ByValStoreOffset,
2896 unsigned TotalArgRegsSaveSize) const {
2898 // Currently, two use-cases possible:
2899 // Case #1. Non-var-args function, and we meet first byval parameter.
2900 // Setup first unallocated register as first byval register;
2901 // eat all remained registers
2902 // (these two actions are performed by HandleByVal method).
2903 // Then, here, we initialize stack frame with
2904 // "store-reg" instructions.
2905 // Case #2. Var-args function, that doesn't contain byval parameters.
2906 // The same: eat all remained unallocated registers,
2907 // initialize stack frame.
2909 MachineFunction &MF = DAG.getMachineFunction();
2910 MachineFrameInfo *MFI = MF.getFrameInfo();
2911 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2912 unsigned firstRegToSaveIndex, lastRegToSaveIndex;
2913 unsigned RBegin, REnd;
2914 if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
2915 CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
2916 firstRegToSaveIndex = RBegin - ARM::R0;
2917 lastRegToSaveIndex = REnd - ARM::R0;
2919 firstRegToSaveIndex = CCInfo.getFirstUnallocated
2920 (GPRArgRegs, array_lengthof(GPRArgRegs));
2921 lastRegToSaveIndex = 4;
2924 unsigned ArgRegsSize, ArgRegsSaveSize;
2925 computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
2926 ArgRegsSize, ArgRegsSaveSize);
2928 // Store any by-val regs to their spots on the stack so that they may be
2929 // loaded by deferencing the result of formal parameter pointer or va_next.
2930 // Note: once stack area for byval/varargs registers
2931 // was initialized, it can't be initialized again.
2932 if (ArgRegsSaveSize) {
2933 unsigned Padding = ArgRegsSaveSize - ArgRegsSize;
2936 assert(AFI->getStoredByValParamsPadding() == 0 &&
2937 "The only parameter may be padded.");
2938 AFI->setStoredByValParamsPadding(Padding);
2941 int FrameIndex = MFI->CreateFixedObject(ArgRegsSaveSize,
2944 (int64_t)TotalArgRegsSaveSize,
2946 SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());
2948 MFI->CreateFixedObject(Padding,
2949 ArgOffset + ByValStoreOffset -
2950 (int64_t)ArgRegsSaveSize,
2954 SmallVector<SDValue, 4> MemOps;
2955 for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
2956 ++firstRegToSaveIndex, ++i) {
2957 const TargetRegisterClass *RC;
2958 if (AFI->isThumb1OnlyFunction())
2959 RC = &ARM::tGPRRegClass;
2961 RC = &ARM::GPRRegClass;
2963 unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2964 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2966 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2967 MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
2969 MemOps.push_back(Store);
2970 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2971 DAG.getConstant(4, getPointerTy()));
2974 AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());
2976 if (!MemOps.empty())
2977 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2981 // We cannot allocate a zero-byte object for the first variadic argument,
2982 // so just make up a size.
2985 // This will point to the next argument passed via stack.
2986 return MFI->CreateFixedObject(
2987 ArgSize, ArgOffset, !ForceMutable);
2991 // Setup stack frame, the va_list pointer will start from.
2993 ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2994 SDLoc dl, SDValue &Chain,
2996 unsigned TotalArgRegsSaveSize,
2997 bool ForceMutable) const {
2998 MachineFunction &MF = DAG.getMachineFunction();
2999 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3001 // Try to store any remaining integer argument regs
3002 // to their spots on the stack so that they may be loaded by deferencing
3003 // the result of va_next.
3004 // If there is no regs to be stored, just point address after last
3005 // argument passed via stack.
3007 StoreByValRegs(CCInfo, DAG, dl, Chain, nullptr,
3008 CCInfo.getInRegsParamsCount(), 0, ArgOffset, 0, ForceMutable,
3009 0, TotalArgRegsSaveSize);
3011 AFI->setVarArgsFrameIndex(FrameIndex);
3015 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
3016 CallingConv::ID CallConv, bool isVarArg,
3017 const SmallVectorImpl<ISD::InputArg>
3019 SDLoc dl, SelectionDAG &DAG,
3020 SmallVectorImpl<SDValue> &InVals)
3022 MachineFunction &MF = DAG.getMachineFunction();
3023 MachineFrameInfo *MFI = MF.getFrameInfo();
3025 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
3027 // Assign locations to all of the incoming arguments.
3028 SmallVector<CCValAssign, 16> ArgLocs;
3029 ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
3030 *DAG.getContext(), Prologue);
3031 CCInfo.AnalyzeFormalArguments(Ins,
3032 CCAssignFnForNode(CallConv, /* Return*/ false,
3035 SmallVector<SDValue, 16> ArgValues;
3036 int lastInsIndex = -1;
3038 Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
3039 unsigned CurArgIdx = 0;
3041 // Initially ArgRegsSaveSize is zero.
3042 // Then we increase this value each time we meet byval parameter.
3043 // We also increase this value in case of varargs function.
3044 AFI->setArgRegsSaveSize(0);
3046 unsigned ByValStoreOffset = 0;
3047 unsigned TotalArgRegsSaveSize = 0;
3048 unsigned ArgRegsSaveSizeMaxAlign = 4;
3050 // Calculate the amount of stack space that we need to allocate to store
3051 // byval and variadic arguments that are passed in registers.
3052 // We need to know this before we allocate the first byval or variadic
3053 // argument, as they will be allocated a stack slot below the CFA (Canonical
3054 // Frame Address, the stack pointer at entry to the function).
3055 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3056 CCValAssign &VA = ArgLocs[i];
3057 if (VA.isMemLoc()) {
3058 int index = VA.getValNo();
3059 if (index != lastInsIndex) {
3060 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3061 if (Flags.isByVal()) {
3062 unsigned ExtraArgRegsSize;
3063 unsigned ExtraArgRegsSaveSize;
3064 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsProceed(),
3065 Flags.getByValSize(),
3066 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3068 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3069 if (Flags.getByValAlign() > ArgRegsSaveSizeMaxAlign)
3070 ArgRegsSaveSizeMaxAlign = Flags.getByValAlign();
3071 CCInfo.nextInRegsParam();
3073 lastInsIndex = index;
3077 CCInfo.rewindByValRegsInfo();
3080 unsigned ExtraArgRegsSize;
3081 unsigned ExtraArgRegsSaveSize;
3082 computeRegArea(CCInfo, MF, CCInfo.getInRegsParamsCount(), 0,
3083 ExtraArgRegsSize, ExtraArgRegsSaveSize);
3084 TotalArgRegsSaveSize += ExtraArgRegsSaveSize;
3086 // If the arg regs save area contains N-byte aligned values, the
3087 // bottom of it must be at least N-byte aligned.
3088 TotalArgRegsSaveSize = RoundUpToAlignment(TotalArgRegsSaveSize, ArgRegsSaveSizeMaxAlign);
3089 TotalArgRegsSaveSize = std::min(TotalArgRegsSaveSize, 16U);
3091 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3092 CCValAssign &VA = ArgLocs[i];
3093 std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
3094 CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
3095 // Arguments stored in registers.
3096 if (VA.isRegLoc()) {
3097 EVT RegVT = VA.getLocVT();
3099 if (VA.needsCustom()) {
3100 // f64 and vector types are split up into multiple registers or
3101 // combinations of registers and stack slots.
3102 if (VA.getLocVT() == MVT::v2f64) {
3103 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
3105 VA = ArgLocs[++i]; // skip ahead to next loc
3107 if (VA.isMemLoc()) {
3108 int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
3109 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3110 ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
3111 MachinePointerInfo::getFixedStack(FI),
3112 false, false, false, 0);
3114 ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
3117 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
3118 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3119 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
3120 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
3121 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
3123 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
3126 const TargetRegisterClass *RC;
3128 if (RegVT == MVT::f32)
3129 RC = &ARM::SPRRegClass;
3130 else if (RegVT == MVT::f64)
3131 RC = &ARM::DPRRegClass;
3132 else if (RegVT == MVT::v2f64)
3133 RC = &ARM::QPRRegClass;
3134 else if (RegVT == MVT::i32)
3135 RC = AFI->isThumb1OnlyFunction() ?
3136 (const TargetRegisterClass*)&ARM::tGPRRegClass :
3137 (const TargetRegisterClass*)&ARM::GPRRegClass;
3139 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
3141 // Transform the arguments in physical registers into virtual ones.
3142 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
3143 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
3146 // If this is an 8 or 16-bit value, it is really passed promoted
3147 // to 32 bits. Insert an assert[sz]ext to capture this, then
3148 // truncate to the right size.
3149 switch (VA.getLocInfo()) {
3150 default: llvm_unreachable("Unknown loc info!");
3151 case CCValAssign::Full: break;
3152 case CCValAssign::BCvt:
3153 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
3155 case CCValAssign::SExt:
3156 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
3157 DAG.getValueType(VA.getValVT()));
3158 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3160 case CCValAssign::ZExt:
3161 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
3162 DAG.getValueType(VA.getValVT()));
3163 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
3167 InVals.push_back(ArgValue);
3169 } else { // VA.isRegLoc()
3172 assert(VA.isMemLoc());
3173 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
3175 int index = ArgLocs[i].getValNo();
3177 // Some Ins[] entries become multiple ArgLoc[] entries.
3178 // Process them only once.
3179 if (index != lastInsIndex)
3181 ISD::ArgFlagsTy Flags = Ins[index].Flags;
3182 // FIXME: For now, all byval parameter objects are marked mutable.
3183 // This can be changed with more analysis.
3184 // In case of tail call optimization mark all arguments mutable.
3185 // Since they could be overwritten by lowering of arguments in case of
3187 if (Flags.isByVal()) {
3188 unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
3190 ByValStoreOffset = RoundUpToAlignment(ByValStoreOffset, Flags.getByValAlign());
3191 int FrameIndex = StoreByValRegs(
3192 CCInfo, DAG, dl, Chain, CurOrigArg,
3194 Ins[VA.getValNo()].PartOffset,
3195 VA.getLocMemOffset(),
3196 Flags.getByValSize(),
3197 true /*force mutable frames*/,
3199 TotalArgRegsSaveSize);
3200 ByValStoreOffset += Flags.getByValSize();
3201 ByValStoreOffset = std::min(ByValStoreOffset, 16U);
3202 InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
3203 CCInfo.nextInRegsParam();
3205 unsigned FIOffset = VA.getLocMemOffset();
3206 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
3209 // Create load nodes to retrieve arguments from the stack.
3210 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
3211 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
3212 MachinePointerInfo::getFixedStack(FI),
3213 false, false, false, 0));
3215 lastInsIndex = index;
3222 VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
3223 CCInfo.getNextStackOffset(),
3224 TotalArgRegsSaveSize);
3226 AFI->setArgumentStackSize(CCInfo.getNextStackOffset());
3231 /// isFloatingPointZero - Return true if this is +0.0.
3232 static bool isFloatingPointZero(SDValue Op) {
3233 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
3234 return CFP->getValueAPF().isPosZero();
3235 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
3236 // Maybe this has already been legalized into the constant pool?
3237 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
3238 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
3239 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
3240 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
3241 return CFP->getValueAPF().isPosZero();
3247 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
3248 /// the given operands.
3250 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
3251 SDValue &ARMcc, SelectionDAG &DAG,
3253 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
3254 unsigned C = RHSC->getZExtValue();
3255 if (!isLegalICmpImmediate(C)) {
3256 // Constant does not fit, try adjusting it by one?
3261 if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
3262 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
3263 RHS = DAG.getConstant(C-1, MVT::i32);
3268 if (C != 0 && isLegalICmpImmediate(C-1)) {
3269 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
3270 RHS = DAG.getConstant(C-1, MVT::i32);
3275 if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
3276 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
3277 RHS = DAG.getConstant(C+1, MVT::i32);
3282 if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
3283 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
3284 RHS = DAG.getConstant(C+1, MVT::i32);
3291 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3292 ARMISD::NodeType CompareType;
3295 CompareType = ARMISD::CMP;
3300 CompareType = ARMISD::CMPZ;
3303 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3304 return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
3307 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
3309 ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
3311 assert(!Subtarget->isFPOnlySP() || RHS.getValueType() != MVT::f64);
3313 if (!isFloatingPointZero(RHS))
3314 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
3316 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
3317 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
3320 /// duplicateCmp - Glue values can have only one use, so this function
3321 /// duplicates a comparison node.
3323 ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
3324 unsigned Opc = Cmp.getOpcode();
3326 if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
3327 return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3329 assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
3330 Cmp = Cmp.getOperand(0);
3331 Opc = Cmp.getOpcode();
3332 if (Opc == ARMISD::CMPFP)
3333 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
3335 assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
3336 Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
3338 return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
3341 std::pair<SDValue, SDValue>
3342 ARMTargetLowering::getARMXALUOOp(SDValue Op, SelectionDAG &DAG,
3343 SDValue &ARMcc) const {
3344 assert(Op.getValueType() == MVT::i32 && "Unsupported value type");
3346 SDValue Value, OverflowCmp;
3347 SDValue LHS = Op.getOperand(0);
3348 SDValue RHS = Op.getOperand(1);
3351 // FIXME: We are currently always generating CMPs because we don't support
3352 // generating CMN through the backend. This is not as good as the natural
3353 // CMP case because it causes a register dependency and cannot be folded
3356 switch (Op.getOpcode()) {
3358 llvm_unreachable("Unknown overflow instruction!");
3360 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3361 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3362 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3365 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3366 Value = DAG.getNode(ISD::ADD, SDLoc(Op), Op.getValueType(), LHS, RHS);
3367 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, Value, LHS);
3370 ARMcc = DAG.getConstant(ARMCC::VC, MVT::i32);
3371 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3372 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3375 ARMcc = DAG.getConstant(ARMCC::HS, MVT::i32);
3376 Value = DAG.getNode(ISD::SUB, SDLoc(Op), Op.getValueType(), LHS, RHS);
3377 OverflowCmp = DAG.getNode(ARMISD::CMP, SDLoc(Op), MVT::Glue, LHS, RHS);
3381 return std::make_pair(Value, OverflowCmp);
3386 ARMTargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
3387 // Let legalize expand this if it isn't a legal type yet.
3388 if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
3391 SDValue Value, OverflowCmp;
3393 std::tie(Value, OverflowCmp) = getARMXALUOOp(Op, DAG, ARMcc);
3394 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3395 // We use 0 and 1 as false and true values.
3396 SDValue TVal = DAG.getConstant(1, MVT::i32);
3397 SDValue FVal = DAG.getConstant(0, MVT::i32);
3398 EVT VT = Op.getValueType();
3400 SDValue Overflow = DAG.getNode(ARMISD::CMOV, SDLoc(Op), VT, TVal, FVal,
3401 ARMcc, CCR, OverflowCmp);
3403 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
3404 return DAG.getNode(ISD::MERGE_VALUES, SDLoc(Op), VTs, Value, Overflow);
3408 SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
3409 SDValue Cond = Op.getOperand(0);
3410 SDValue SelectTrue = Op.getOperand(1);
3411 SDValue SelectFalse = Op.getOperand(2);
3413 unsigned Opc = Cond.getOpcode();
3415 if (Cond.getResNo() == 1 &&
3416 (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
3417 Opc == ISD::USUBO)) {
3418 if (!DAG.getTargetLoweringInfo().isTypeLegal(Cond->getValueType(0)))
3421 SDValue Value, OverflowCmp;
3423 std::tie(Value, OverflowCmp) = getARMXALUOOp(Cond, DAG, ARMcc);
3424 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3425 EVT VT = Op.getValueType();
3427 return getCMOV(SDLoc(Op), VT, SelectTrue, SelectFalse, ARMcc, CCR,
3433 // (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
3434 // (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
3436 if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
3437 const ConstantSDNode *CMOVTrue =
3438 dyn_cast<ConstantSDNode>(Cond.getOperand(0));
3439 const ConstantSDNode *CMOVFalse =
3440 dyn_cast<ConstantSDNode>(Cond.getOperand(1));
3442 if (CMOVTrue && CMOVFalse) {
3443 unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
3444 unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
3448 if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
3450 False = SelectFalse;
3451 } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
3456 if (True.getNode() && False.getNode()) {
3457 EVT VT = Op.getValueType();
3458 SDValue ARMcc = Cond.getOperand(2);
3459 SDValue CCR = Cond.getOperand(3);
3460 SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
3461 assert(True.getValueType() == VT);
3462 return getCMOV(dl, VT, True, False, ARMcc, CCR, Cmp, DAG);
3467 // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
3468 // undefined bits before doing a full-word comparison with zero.
3469 Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
3470 DAG.getConstant(1, Cond.getValueType()));
3472 return DAG.getSelectCC(dl, Cond,
3473 DAG.getConstant(0, Cond.getValueType()),
3474 SelectTrue, SelectFalse, ISD::SETNE);
3477 static ISD::CondCode getInverseCCForVSEL(ISD::CondCode CC) {
3478 if (CC == ISD::SETNE)
3480 return ISD::getSetCCInverse(CC, true);
3483 static void checkVSELConstraints(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
3484 bool &swpCmpOps, bool &swpVselOps) {
3485 // Start by selecting the GE condition code for opcodes that return true for
3487 if (CC == ISD::SETUGE || CC == ISD::SETOGE || CC == ISD::SETOLE ||
3489 CondCode = ARMCC::GE;
3491 // and GT for opcodes that return false for 'equality'.
3492 else if (CC == ISD::SETUGT || CC == ISD::SETOGT || CC == ISD::SETOLT ||
3494 CondCode = ARMCC::GT;
3496 // Since we are constrained to GE/GT, if the opcode contains 'less', we need
3497 // to swap the compare operands.
3498 if (CC == ISD::SETOLE || CC == ISD::SETULE || CC == ISD::SETOLT ||
3502 // Both GT and GE are ordered comparisons, and return false for 'unordered'.
3503 // If we have an unordered opcode, we need to swap the operands to the VSEL
3504 // instruction (effectively negating the condition).
3506 // This also has the effect of swapping which one of 'less' or 'greater'
3507 // returns true, so we also swap the compare operands. It also switches
3508 // whether we return true for 'equality', so we compensate by picking the
3509 // opposite condition code to our original choice.
3510 if (CC == ISD::SETULE || CC == ISD::SETULT || CC == ISD::SETUGE ||
3511 CC == ISD::SETUGT) {
3512 swpCmpOps = !swpCmpOps;
3513 swpVselOps = !swpVselOps;
3514 CondCode = CondCode == ARMCC::GT ? ARMCC::GE : ARMCC::GT;
3517 // 'ordered' is 'anything but unordered', so use the VS condition code and
3518 // swap the VSEL operands.
3519 if (CC == ISD::SETO) {
3520 CondCode = ARMCC::VS;
3524 // 'unordered or not equal' is 'anything but equal', so use the EQ condition
3525 // code and swap the VSEL operands.
3526 if (CC == ISD::SETUNE) {
3527 CondCode = ARMCC::EQ;
3532 SDValue ARMTargetLowering::getCMOV(SDLoc dl, EVT VT, SDValue FalseVal,
3533 SDValue TrueVal, SDValue ARMcc, SDValue CCR,
3534 SDValue Cmp, SelectionDAG &DAG) const {
3535 if (Subtarget->isFPOnlySP() && VT == MVT::f64) {
3536 FalseVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3537 DAG.getVTList(MVT::i32, MVT::i32), FalseVal);
3538 TrueVal = DAG.getNode(ARMISD::VMOVRRD, dl,
3539 DAG.getVTList(MVT::i32, MVT::i32), TrueVal);
3541 SDValue TrueLow = TrueVal.getValue(0);
3542 SDValue TrueHigh = TrueVal.getValue(1);
3543 SDValue FalseLow = FalseVal.getValue(0);
3544 SDValue FalseHigh = FalseVal.getValue(1);
3546 SDValue Low = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseLow, TrueLow,
3548 SDValue High = DAG.getNode(ARMISD::CMOV, dl, MVT::i32, FalseHigh, TrueHigh,
3549 ARMcc, CCR, duplicateCmp(Cmp, DAG));
3551 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Low, High);
3553 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,
3558 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3559 EVT VT = Op.getValueType();
3560 SDValue LHS = Op.getOperand(0);
3561 SDValue RHS = Op.getOperand(1);
3562 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3563 SDValue TrueVal = Op.getOperand(2);
3564 SDValue FalseVal = Op.getOperand(3);
3567 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3568 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3571 // If softenSetCCOperands only returned one value, we should compare it to
3573 if (!RHS.getNode()) {
3574 RHS = DAG.getConstant(0, LHS.getValueType());
3579 if (LHS.getValueType() == MVT::i32) {
3580 // Try to generate VSEL on ARMv8.
3581 // The VSEL instruction can't use all the usual ARM condition
3582 // codes: it only has two bits to select the condition code, so it's
3583 // constrained to use only GE, GT, VS and EQ.
3585 // To implement all the various ISD::SETXXX opcodes, we sometimes need to
3586 // swap the operands of the previous compare instruction (effectively
3587 // inverting the compare condition, swapping 'less' and 'greater') and
3588 // sometimes need to swap the operands to the VSEL (which inverts the
3589 // condition in the sense of firing whenever the previous condition didn't)
3590 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3591 TrueVal.getValueType() == MVT::f64)) {
3592 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3593 if (CondCode == ARMCC::LT || CondCode == ARMCC::LE ||
3594 CondCode == ARMCC::VC || CondCode == ARMCC::NE) {
3595 CC = getInverseCCForVSEL(CC);
3596 std::swap(TrueVal, FalseVal);
3601 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3602 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3603 return getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3606 ARMCC::CondCodes CondCode, CondCode2;
3607 FPCCToARMCC(CC, CondCode, CondCode2);
3609 // Try to generate VSEL on ARMv8.
3610 if (getSubtarget()->hasFPARMv8() && (TrueVal.getValueType() == MVT::f32 ||
3611 TrueVal.getValueType() == MVT::f64)) {
3612 // We can select VMAXNM/VMINNM from a compare followed by a select with the
3613 // same operands, as follows:
3614 // c = fcmp [ogt, olt, ugt, ult] a, b
3616 // We only do this in unsafe-fp-math, because signed zeros and NaNs are
3617 // handled differently than the original code sequence.
3618 if (getTargetMachine().Options.UnsafeFPMath && LHS == TrueVal &&
3620 if (CC == ISD::SETOGT || CC == ISD::SETUGT)
3621 return DAG.getNode(ARMISD::VMAXNM, dl, VT, TrueVal, FalseVal);
3622 if (CC == ISD::SETOLT || CC == ISD::SETULT)
3623 return DAG.getNode(ARMISD::VMINNM, dl, VT, TrueVal, FalseVal);
3626 bool swpCmpOps = false;
3627 bool swpVselOps = false;
3628 checkVSELConstraints(CC, CondCode, swpCmpOps, swpVselOps);
3630 if (CondCode == ARMCC::GT || CondCode == ARMCC::GE ||
3631 CondCode == ARMCC::VS || CondCode == ARMCC::EQ) {
3633 std::swap(LHS, RHS);
3635 std::swap(TrueVal, FalseVal);
3639 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3640 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3641 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3642 SDValue Result = getCMOV(dl, VT, FalseVal, TrueVal, ARMcc, CCR, Cmp, DAG);
3643 if (CondCode2 != ARMCC::AL) {
3644 SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
3645 // FIXME: Needs another CMP because flag can have but one use.
3646 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
3647 Result = getCMOV(dl, VT, Result, TrueVal, ARMcc2, CCR, Cmp2, DAG);
3652 /// canChangeToInt - Given the fp compare operand, return true if it is suitable
3653 /// to morph to an integer compare sequence.
3654 static bool canChangeToInt(SDValue Op, bool &SeenZero,
3655 const ARMSubtarget *Subtarget) {
3656 SDNode *N = Op.getNode();
3657 if (!N->hasOneUse())
3658 // Otherwise it requires moving the value from fp to integer registers.
3660 if (!N->getNumValues())
3662 EVT VT = Op.getValueType();
3663 if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
3664 // f32 case is generally profitable. f64 case only makes sense when vcmpe +
3665 // vmrs are very slow, e.g. cortex-a8.
3668 if (isFloatingPointZero(Op)) {
3672 return ISD::isNormalLoad(N);
3675 static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
3676 if (isFloatingPointZero(Op))
3677 return DAG.getConstant(0, MVT::i32);
3679 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
3680 return DAG.getLoad(MVT::i32, SDLoc(Op),
3681 Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
3682 Ld->isVolatile(), Ld->isNonTemporal(),
3683 Ld->isInvariant(), Ld->getAlignment());
3685 llvm_unreachable("Unknown VFP cmp argument!");
3688 static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
3689 SDValue &RetVal1, SDValue &RetVal2) {
3690 if (isFloatingPointZero(Op)) {
3691 RetVal1 = DAG.getConstant(0, MVT::i32);
3692 RetVal2 = DAG.getConstant(0, MVT::i32);
3696 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
3697 SDValue Ptr = Ld->getBasePtr();
3698 RetVal1 = DAG.getLoad(MVT::i32, SDLoc(Op),
3699 Ld->getChain(), Ptr,
3700 Ld->getPointerInfo(),
3701 Ld->isVolatile(), Ld->isNonTemporal(),
3702 Ld->isInvariant(), Ld->getAlignment());
3704 EVT PtrType = Ptr.getValueType();
3705 unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
3706 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(Op),
3707 PtrType, Ptr, DAG.getConstant(4, PtrType));
3708 RetVal2 = DAG.getLoad(MVT::i32, SDLoc(Op),
3709 Ld->getChain(), NewPtr,
3710 Ld->getPointerInfo().getWithOffset(4),
3711 Ld->isVolatile(), Ld->isNonTemporal(),
3712 Ld->isInvariant(), NewAlign);
3716 llvm_unreachable("Unknown VFP cmp argument!");
3719 /// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
3720 /// f32 and even f64 comparisons to integer ones.
3722 ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
3723 SDValue Chain = Op.getOperand(0);
3724 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3725 SDValue LHS = Op.getOperand(2);
3726 SDValue RHS = Op.getOperand(3);
3727 SDValue Dest = Op.getOperand(4);
3730 bool LHSSeenZero = false;
3731 bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
3732 bool RHSSeenZero = false;
3733 bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
3734 if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
3735 // If unsafe fp math optimization is enabled and there are no other uses of
3736 // the CMP operands, and the condition code is EQ or NE, we can optimize it
3737 // to an integer comparison.
3738 if (CC == ISD::SETOEQ)
3740 else if (CC == ISD::SETUNE)
3743 SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
3745 if (LHS.getValueType() == MVT::f32) {
3746 LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3747 bitcastf32Toi32(LHS, DAG), Mask);
3748 RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3749 bitcastf32Toi32(RHS, DAG), Mask);
3750 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3751 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3752 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3753 Chain, Dest, ARMcc, CCR, Cmp);
3758 expandf64Toi32(LHS, DAG, LHS1, LHS2);
3759 expandf64Toi32(RHS, DAG, RHS1, RHS2);
3760 LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3761 RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3762 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3763 ARMcc = DAG.getConstant(CondCode, MVT::i32);
3764 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3765 SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3766 return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops);
3772 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3773 SDValue Chain = Op.getOperand(0);
3774 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3775 SDValue LHS = Op.getOperand(2);
3776 SDValue RHS = Op.getOperand(3);
3777 SDValue Dest = Op.getOperand(4);
3780 if (Subtarget->isFPOnlySP() && LHS.getValueType() == MVT::f64) {
3781 DAG.getTargetLoweringInfo().softenSetCCOperands(DAG, MVT::f64, LHS, RHS, CC,
3784 // If softenSetCCOperands only returned one value, we should compare it to
3786 if (!RHS.getNode()) {
3787 RHS = DAG.getConstant(0, LHS.getValueType());
3792 if (LHS.getValueType() == MVT::i32) {
3794 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3795 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3796 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3797 Chain, Dest, ARMcc, CCR, Cmp);
3800 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3802 if (getTargetMachine().Options.UnsafeFPMath &&
3803 (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3804 CC == ISD::SETNE || CC == ISD::SETUNE)) {
3805 SDValue Result = OptimizeVFPBrcond(Op, DAG);
3806 if (Result.getNode())
3810 ARMCC::CondCodes CondCode, CondCode2;
3811 FPCCToARMCC(CC, CondCode, CondCode2);
3813 SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3814 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3815 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3816 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3817 SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3818 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3819 if (CondCode2 != ARMCC::AL) {
3820 ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3821 SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3822 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops);
3827 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3828 SDValue Chain = Op.getOperand(0);
3829 SDValue Table = Op.getOperand(1);
3830 SDValue Index = Op.getOperand(2);
3833 EVT PTy = getPointerTy();
3834 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3835 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3836 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3837 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3838 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3839 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3840 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3841 if (Subtarget->isThumb2()) {
3842 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3843 // which does another jump to the destination. This also makes it easier
3844 // to translate it to TBB / TBH later.
3845 // FIXME: This might not work if the function is extremely large.
3846 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3847 Addr, Op.getOperand(2), JTI, UId);
3849 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3850 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3851 MachinePointerInfo::getJumpTable(),
3852 false, false, false, 0);
3853 Chain = Addr.getValue(1);
3854 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3855 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3857 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3858 MachinePointerInfo::getJumpTable(),
3859 false, false, false, 0);
3860 Chain = Addr.getValue(1);
3861 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3865 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3866 EVT VT = Op.getValueType();
3869 if (Op.getValueType().getVectorElementType() == MVT::i32) {
3870 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3872 return DAG.UnrollVectorOp(Op.getNode());
3875 assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3876 "Invalid type for custom lowering!");
3877 if (VT != MVT::v4i16)
3878 return DAG.UnrollVectorOp(Op.getNode());
3880 Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3881 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3884 SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
3885 EVT VT = Op.getValueType();
3887 return LowerVectorFP_TO_INT(Op, DAG);
3889 if (Subtarget->isFPOnlySP() && Op.getOperand(0).getValueType() == MVT::f64) {
3891 if (Op.getOpcode() == ISD::FP_TO_SINT)
3892 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(),
3895 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(),
3897 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3898 /*isSigned*/ false, SDLoc(Op)).first;
3904 switch (Op.getOpcode()) {
3905 default: llvm_unreachable("Invalid opcode!");
3906 case ISD::FP_TO_SINT:
3907 Opc = ARMISD::FTOSI;
3909 case ISD::FP_TO_UINT:
3910 Opc = ARMISD::FTOUI;
3913 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3914 return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3917 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3918 EVT VT = Op.getValueType();
3921 if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3922 if (VT.getVectorElementType() == MVT::f32)
3924 return DAG.UnrollVectorOp(Op.getNode());
3927 assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3928 "Invalid type for custom lowering!");
3929 if (VT != MVT::v4f32)
3930 return DAG.UnrollVectorOp(Op.getNode());
3934 switch (Op.getOpcode()) {
3935 default: llvm_unreachable("Invalid opcode!");
3936 case ISD::SINT_TO_FP:
3937 CastOpc = ISD::SIGN_EXTEND;
3938 Opc = ISD::SINT_TO_FP;
3940 case ISD::UINT_TO_FP:
3941 CastOpc = ISD::ZERO_EXTEND;
3942 Opc = ISD::UINT_TO_FP;
3946 Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3947 return DAG.getNode(Opc, dl, VT, Op);
3950 SDValue ARMTargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const {
3951 EVT VT = Op.getValueType();
3953 return LowerVectorINT_TO_FP(Op, DAG);
3955 if (Subtarget->isFPOnlySP() && Op.getValueType() == MVT::f64) {
3957 if (Op.getOpcode() == ISD::SINT_TO_FP)
3958 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(),
3961 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(),
3963 return makeLibCall(DAG, LC, Op.getValueType(), &Op.getOperand(0), 1,
3964 /*isSigned*/ false, SDLoc(Op)).first;
3970 switch (Op.getOpcode()) {
3971 default: llvm_unreachable("Invalid opcode!");
3972 case ISD::SINT_TO_FP:
3973 Opc = ARMISD::SITOF;
3975 case ISD::UINT_TO_FP:
3976 Opc = ARMISD::UITOF;
3980 Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3981 return DAG.getNode(Opc, dl, VT, Op);
3984 SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3985 // Implement fcopysign with a fabs and a conditional fneg.
3986 SDValue Tmp0 = Op.getOperand(0);
3987 SDValue Tmp1 = Op.getOperand(1);
3989 EVT VT = Op.getValueType();
3990 EVT SrcVT = Tmp1.getValueType();
3991 bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3992 Tmp0.getOpcode() == ARMISD::VMOVDRR;
3993 bool UseNEON = !InGPR && Subtarget->hasNEON();
3996 // Use VBSL to copy the sign bit.
3997 unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3998 SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3999 DAG.getTargetConstant(EncodedVal, MVT::i32));
4000 EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
4002 Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4003 DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
4004 DAG.getConstant(32, MVT::i32));
4005 else /*if (VT == MVT::f32)*/
4006 Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
4007 if (SrcVT == MVT::f32) {
4008 Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
4010 Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
4011 DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
4012 DAG.getConstant(32, MVT::i32));
4013 } else if (VT == MVT::f32)
4014 Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
4015 DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
4016 DAG.getConstant(32, MVT::i32));
4017 Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
4018 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
4020 SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
4022 AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
4023 SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
4024 DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
4026 SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
4027 DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
4028 DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
4029 if (VT == MVT::f32) {
4030 Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
4031 Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
4032 DAG.getConstant(0, MVT::i32));
4034 Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
4040 // Bitcast operand 1 to i32.
4041 if (SrcVT == MVT::f64)
4042 Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4044 Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
4046 // Or in the signbit with integer operations.
4047 SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
4048 SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
4049 Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
4050 if (VT == MVT::f32) {
4051 Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
4052 DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
4053 return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
4054 DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
4057 // f64: Or the high part with signbit and then combine two parts.
4058 Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
4060 SDValue Lo = Tmp0.getValue(0);
4061 SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
4062 Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
4063 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
4066 SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
4067 MachineFunction &MF = DAG.getMachineFunction();
4068 MachineFrameInfo *MFI = MF.getFrameInfo();
4069 MFI->setReturnAddressIsTaken(true);
4071 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
4074 EVT VT = Op.getValueType();
4076 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4078 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
4079 SDValue Offset = DAG.getConstant(4, MVT::i32);
4080 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
4081 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
4082 MachinePointerInfo(), false, false, false, 0);
4085 // Return LR, which contains the return address. Mark it an implicit live-in.
4086 unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
4087 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
4090 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
4091 const ARMBaseRegisterInfo &ARI =
4092 *static_cast<const ARMBaseRegisterInfo*>(RegInfo);
4093 MachineFunction &MF = DAG.getMachineFunction();
4094 MachineFrameInfo *MFI = MF.getFrameInfo();
4095 MFI->setFrameAddressIsTaken(true);
4097 EVT VT = Op.getValueType();
4098 SDLoc dl(Op); // FIXME probably not meaningful
4099 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
4100 unsigned FrameReg = ARI.getFrameRegister(MF);
4101 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
4103 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
4104 MachinePointerInfo(),
4105 false, false, false, 0);
4109 // FIXME? Maybe this could be a TableGen attribute on some registers and
4110 // this table could be generated automatically from RegInfo.
4111 unsigned ARMTargetLowering::getRegisterByName(const char* RegName,
4113 unsigned Reg = StringSwitch<unsigned>(RegName)
4114 .Case("sp", ARM::SP)
4118 report_fatal_error("Invalid register name global variable");
4121 /// ExpandBITCAST - If the target supports VFP, this function is called to
4122 /// expand a bit convert where either the source or destination type is i64 to
4123 /// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
4124 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
4125 /// vectors), since the legalizer won't know what to do with that.
4126 static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
4127 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4129 SDValue Op = N->getOperand(0);
4131 // This function is only supposed to be called for i64 types, either as the
4132 // source or destination of the bit convert.
4133 EVT SrcVT = Op.getValueType();
4134 EVT DstVT = N->getValueType(0);
4135 assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
4136 "ExpandBITCAST called for non-i64 type");
4138 // Turn i64->f64 into VMOVDRR.
4139 if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
4140 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4141 DAG.getConstant(0, MVT::i32));
4142 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
4143 DAG.getConstant(1, MVT::i32));
4144 return DAG.getNode(ISD::BITCAST, dl, DstVT,
4145 DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
4148 // Turn f64->i64 into VMOVRRD.
4149 if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
4151 if (TLI.isBigEndian() && SrcVT.isVector() &&
4152 SrcVT.getVectorNumElements() > 1)
4153 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4154 DAG.getVTList(MVT::i32, MVT::i32),
4155 DAG.getNode(ARMISD::VREV64, dl, SrcVT, Op));
4157 Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
4158 DAG.getVTList(MVT::i32, MVT::i32), Op);
4159 // Merge the pieces into a single i64 value.
4160 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
4166 /// getZeroVector - Returns a vector of specified type with all zero elements.
4167 /// Zero vectors are used to represent vector negation and in those cases
4168 /// will be implemented with the NEON VNEG instruction. However, VNEG does
4169 /// not support i64 elements, so sometimes the zero vectors will need to be
4170 /// explicitly constructed. Regardless, use a canonical VMOV to create the
4172 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, SDLoc dl) {
4173 assert(VT.isVector() && "Expected a vector type");
4174 // The canonical modified immediate encoding of a zero vector is....0!
4175 SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
4176 EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
4177 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
4178 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4181 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
4182 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4183 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
4184 SelectionDAG &DAG) const {
4185 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4186 EVT VT = Op.getValueType();
4187 unsigned VTBits = VT.getSizeInBits();
4189 SDValue ShOpLo = Op.getOperand(0);
4190 SDValue ShOpHi = Op.getOperand(1);
4191 SDValue ShAmt = Op.getOperand(2);
4193 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4195 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4197 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4198 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4199 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4200 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4201 DAG.getConstant(VTBits, MVT::i32));
4202 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4203 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4204 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4206 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4207 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4209 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4210 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
4213 SDValue Ops[2] = { Lo, Hi };
4214 return DAG.getMergeValues(Ops, dl);
4217 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4218 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
4219 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
4220 SelectionDAG &DAG) const {
4221 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4222 EVT VT = Op.getValueType();
4223 unsigned VTBits = VT.getSizeInBits();
4225 SDValue ShOpLo = Op.getOperand(0);
4226 SDValue ShOpHi = Op.getOperand(1);
4227 SDValue ShAmt = Op.getOperand(2);
4230 assert(Op.getOpcode() == ISD::SHL_PARTS);
4231 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
4232 DAG.getConstant(VTBits, MVT::i32), ShAmt);
4233 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4234 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
4235 DAG.getConstant(VTBits, MVT::i32));
4236 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4237 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4239 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4240 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
4241 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
4243 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4244 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
4247 SDValue Ops[2] = { Lo, Hi };
4248 return DAG.getMergeValues(Ops, dl);
4251 SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4252 SelectionDAG &DAG) const {
4253 // The rounding mode is in bits 23:22 of the FPSCR.
4254 // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
4255 // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
4256 // so that the shift + and get folded into a bitfield extract.
4258 SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
4259 DAG.getConstant(Intrinsic::arm_get_fpscr,
4261 SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
4262 DAG.getConstant(1U << 22, MVT::i32));
4263 SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
4264 DAG.getConstant(22, MVT::i32));
4265 return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
4266 DAG.getConstant(3, MVT::i32));
4269 static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
4270 const ARMSubtarget *ST) {
4271 EVT VT = N->getValueType(0);
4274 if (!ST->hasV6T2Ops())
4277 SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
4278 return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
4281 /// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
4282 /// for each 16-bit element from operand, repeated. The basic idea is to
4283 /// leverage vcnt to get the 8-bit counts, gather and add the results.
4285 /// Trace for v4i16:
4286 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4287 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
4288 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
4289 /// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
4290 /// [b0 b1 b2 b3 b4 b5 b6 b7]
4291 /// +[b1 b0 b3 b2 b5 b4 b7 b6]
4292 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
4293 /// vuzp: = [k0 k1 k2 k3 k0 k1 k2 k3] each ki is 8-bits)
4294 static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
4295 EVT VT = N->getValueType(0);
4298 EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
4299 SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
4300 SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
4301 SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
4302 SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
4303 return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
4306 /// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
4307 /// bit-count for each 16-bit element from the operand. We need slightly
4308 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
4309 /// 64/128-bit registers.
4311 /// Trace for v4i16:
4312 /// input = [v0 v1 v2 v3 ] (vi 16-bit element)
4313 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
4314 /// v8i16:Extended = [k0 k1 k2 k3 k0 k1 k2 k3 ]
4315 /// v4i16:Extracted = [k0 k1 k2 k3 ]
4316 static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
4317 EVT VT = N->getValueType(0);
4320 SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
4321 if (VT.is64BitVector()) {
4322 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
4323 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
4324 DAG.getIntPtrConstant(0));
4326 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
4327 BitCounts, DAG.getIntPtrConstant(0));
4328 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
4332 /// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
4333 /// bit-count for each 32-bit element from the operand. The idea here is
4334 /// to split the vector into 16-bit elements, leverage the 16-bit count
4335 /// routine, and then combine the results.
4337 /// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
4338 /// input = [v0 v1 ] (vi: 32-bit elements)
4339 /// Bitcast = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
4340 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
4341 /// vrev: N0 = [k1 k0 k3 k2 ]
4343 /// N1 =+[k1 k0 k3 k2 ]
4345 /// N2 =+[k1 k3 k0 k2 ]
4347 /// Extended =+[k1 k3 k0 k2 ]
4349 /// Extracted=+[k1 k3 ]
4351 static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
4352 EVT VT = N->getValueType(0);
4355 EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;
4357 SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
4358 SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
4359 SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
4360 SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
4361 SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);
4363 if (VT.is64BitVector()) {
4364 SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
4365 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
4366 DAG.getIntPtrConstant(0));
4368 SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
4369 DAG.getIntPtrConstant(0));
4370 return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
4374 static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
4375 const ARMSubtarget *ST) {
4376 EVT VT = N->getValueType(0);
4378 assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
4379 assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
4380 VT == MVT::v4i16 || VT == MVT::v8i16) &&
4381 "Unexpected type for custom ctpop lowering");
4383 if (VT.getVectorElementType() == MVT::i32)
4384 return lowerCTPOP32BitElements(N, DAG);
4386 return lowerCTPOP16BitElements(N, DAG);
4389 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
4390 const ARMSubtarget *ST) {
4391 EVT VT = N->getValueType(0);
4397 // Lower vector shifts on NEON to use VSHL.
4398 assert(ST->hasNEON() && "unexpected vector shift");
4400 // Left shifts translate directly to the vshiftu intrinsic.
4401 if (N->getOpcode() == ISD::SHL)
4402 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4403 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
4404 N->getOperand(0), N->getOperand(1));
4406 assert((N->getOpcode() == ISD::SRA ||
4407 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
4409 // NEON uses the same intrinsics for both left and right shifts. For
4410 // right shifts, the shift amounts are negative, so negate the vector of
4412 EVT ShiftVT = N->getOperand(1).getValueType();
4413 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
4414 getZeroVector(ShiftVT, DAG, dl),
4416 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
4417 Intrinsic::arm_neon_vshifts :
4418 Intrinsic::arm_neon_vshiftu);
4419 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
4420 DAG.getConstant(vshiftInt, MVT::i32),
4421 N->getOperand(0), NegatedCount);
4424 static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
4425 const ARMSubtarget *ST) {
4426 EVT VT = N->getValueType(0);
4429 // We can get here for a node like i32 = ISD::SHL i32, i64
4433 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
4434 "Unknown shift to lower!");
4436 // We only lower SRA, SRL of 1 here, all others use generic lowering.
4437 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
4438 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
4441 // If we are in thumb mode, we don't have RRX.
4442 if (ST->isThumb1Only()) return SDValue();
4444 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
4445 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4446 DAG.getConstant(0, MVT::i32));
4447 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
4448 DAG.getConstant(1, MVT::i32));
4450 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
4451 // captures the result into a carry flag.
4452 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
4453 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), Hi);
4455 // The low part is an ARMISD::RRX operand, which shifts the carry in.
4456 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
4458 // Merge the pieces into a single i64 value.
4459 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
4462 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
4463 SDValue TmpOp0, TmpOp1;
4464 bool Invert = false;
4468 SDValue Op0 = Op.getOperand(0);
4469 SDValue Op1 = Op.getOperand(1);
4470 SDValue CC = Op.getOperand(2);
4471 EVT VT = Op.getValueType();
4472 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
4475 if (Op1.getValueType().isFloatingPoint()) {
4476 switch (SetCCOpcode) {
4477 default: llvm_unreachable("Illegal FP comparison");
4479 case ISD::SETNE: Invert = true; // Fallthrough
4481 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4483 case ISD::SETLT: Swap = true; // Fallthrough
4485 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4487 case ISD::SETLE: Swap = true; // Fallthrough
4489 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4490 case ISD::SETUGE: Swap = true; // Fallthrough
4491 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
4492 case ISD::SETUGT: Swap = true; // Fallthrough
4493 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
4494 case ISD::SETUEQ: Invert = true; // Fallthrough
4496 // Expand this to (OLT | OGT).
4500 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4501 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
4503 case ISD::SETUO: Invert = true; // Fallthrough
4505 // Expand this to (OLT | OGE).
4509 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
4510 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
4514 // Integer comparisons.
4515 switch (SetCCOpcode) {
4516 default: llvm_unreachable("Illegal integer comparison");
4517 case ISD::SETNE: Invert = true;
4518 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
4519 case ISD::SETLT: Swap = true;
4520 case ISD::SETGT: Opc = ARMISD::VCGT; break;
4521 case ISD::SETLE: Swap = true;
4522 case ISD::SETGE: Opc = ARMISD::VCGE; break;
4523 case ISD::SETULT: Swap = true;
4524 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
4525 case ISD::SETULE: Swap = true;
4526 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
4529 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
4530 if (Opc == ARMISD::VCEQ) {
4533 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4535 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
4538 // Ignore bitconvert.
4539 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
4540 AndOp = AndOp.getOperand(0);
4542 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
4544 Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
4545 Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
4552 std::swap(Op0, Op1);
4554 // If one of the operands is a constant vector zero, attempt to fold the
4555 // comparison to a specialized compare-against-zero form.
4557 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
4559 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
4560 if (Opc == ARMISD::VCGE)
4561 Opc = ARMISD::VCLEZ;
4562 else if (Opc == ARMISD::VCGT)
4563 Opc = ARMISD::VCLTZ;
4568 if (SingleOp.getNode()) {
4571 Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
4573 Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
4575 Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
4577 Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
4579 Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
4581 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4584 Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
4588 Result = DAG.getNOT(dl, Result, VT);
4593 /// isNEONModifiedImm - Check if the specified splat value corresponds to a
4594 /// valid vector constant for a NEON instruction with a "modified immediate"
4595 /// operand (e.g., VMOV). If so, return the encoded value.
4596 static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
4597 unsigned SplatBitSize, SelectionDAG &DAG,
4598 EVT &VT, bool is128Bits, NEONModImmType type) {
4599 unsigned OpCmode, Imm;
4601 // SplatBitSize is set to the smallest size that splats the vector, so a
4602 // zero vector will always have SplatBitSize == 8. However, NEON modified
4603 // immediate instructions others than VMOV do not support the 8-bit encoding
4604 // of a zero vector, and the default encoding of zero is supposed to be the
4609 switch (SplatBitSize) {
4611 if (type != VMOVModImm)
4613 // Any 1-byte value is OK. Op=0, Cmode=1110.
4614 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
4617 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
4621 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
4622 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
4623 if ((SplatBits & ~0xff) == 0) {
4624 // Value = 0x00nn: Op=x, Cmode=100x.
4629 if ((SplatBits & ~0xff00) == 0) {
4630 // Value = 0xnn00: Op=x, Cmode=101x.
4632 Imm = SplatBits >> 8;
4638 // NEON's 32-bit VMOV supports splat values where:
4639 // * only one byte is nonzero, or
4640 // * the least significant byte is 0xff and the second byte is nonzero, or
4641 // * the least significant 2 bytes are 0xff and the third is nonzero.
4642 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
4643 if ((SplatBits & ~0xff) == 0) {
4644 // Value = 0x000000nn: Op=x, Cmode=000x.
4649 if ((SplatBits & ~0xff00) == 0) {
4650 // Value = 0x0000nn00: Op=x, Cmode=001x.
4652 Imm = SplatBits >> 8;
4655 if ((SplatBits & ~0xff0000) == 0) {
4656 // Value = 0x00nn0000: Op=x, Cmode=010x.
4658 Imm = SplatBits >> 16;
4661 if ((SplatBits & ~0xff000000) == 0) {
4662 // Value = 0xnn000000: Op=x, Cmode=011x.
4664 Imm = SplatBits >> 24;
4668 // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
4669 if (type == OtherModImm) return SDValue();
4671 if ((SplatBits & ~0xffff) == 0 &&
4672 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
4673 // Value = 0x0000nnff: Op=x, Cmode=1100.
4675 Imm = SplatBits >> 8;
4679 if ((SplatBits & ~0xffffff) == 0 &&
4680 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
4681 // Value = 0x00nnffff: Op=x, Cmode=1101.
4683 Imm = SplatBits >> 16;
4687 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
4688 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
4689 // VMOV.I32. A (very) minor optimization would be to replicate the value
4690 // and fall through here to test for a valid 64-bit splat. But, then the
4691 // caller would also need to check and handle the change in size.
4695 if (type != VMOVModImm)
4697 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
4698 uint64_t BitMask = 0xff;
4700 unsigned ImmMask = 1;
4702 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
4703 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
4706 } else if ((SplatBits & BitMask) != 0) {
4713 if (DAG.getTargetLoweringInfo().isBigEndian())
4714 // swap higher and lower 32 bit word
4715 Imm = ((Imm & 0xf) << 4) | ((Imm & 0xf0) >> 4);
4717 // Op=1, Cmode=1110.
4719 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
4724 llvm_unreachable("unexpected size for isNEONModifiedImm");
4727 unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
4728 return DAG.getTargetConstant(EncodedVal, MVT::i32);
4731 SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
4732 const ARMSubtarget *ST) const {
4736 bool IsDouble = Op.getValueType() == MVT::f64;
4737 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
4739 // Use the default (constant pool) lowering for double constants when we have
4741 if (IsDouble && Subtarget->isFPOnlySP())
4744 // Try splatting with a VMOV.f32...
4745 APFloat FPVal = CFP->getValueAPF();
4746 int ImmVal = IsDouble ? ARM_AM::getFP64Imm(FPVal) : ARM_AM::getFP32Imm(FPVal);
4749 if (IsDouble || !ST->useNEONForSinglePrecisionFP()) {
4750 // We have code in place to select a valid ConstantFP already, no need to
4755 // It's a float and we are trying to use NEON operations where
4756 // possible. Lower it to a splat followed by an extract.
4758 SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
4759 SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
4761 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
4762 DAG.getConstant(0, MVT::i32));
4765 // The rest of our options are NEON only, make sure that's allowed before
4767 if (!ST->hasNEON() || (!IsDouble && !ST->useNEONForSinglePrecisionFP()))
4771 uint64_t iVal = FPVal.bitcastToAPInt().getZExtValue();
4773 // It wouldn't really be worth bothering for doubles except for one very
4774 // important value, which does happen to match: 0.0. So make sure we don't do
4776 if (IsDouble && (iVal & 0xffffffff) != (iVal >> 32))
4779 // Try a VMOV.i32 (FIXME: i8, i16, or i64 could work too).
4780 SDValue NewVal = isNEONModifiedImm(iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4782 if (NewVal != SDValue()) {
4784 SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
4787 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4789 // It's a float: cast and extract a vector element.
4790 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4792 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4793 DAG.getConstant(0, MVT::i32));
4796 // Finally, try a VMVN.i32
4797 NewVal = isNEONModifiedImm(~iVal & 0xffffffffU, 0, 32, DAG, VMovVT,
4799 if (NewVal != SDValue()) {
4801 SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
4804 return DAG.getNode(ISD::BITCAST, DL, MVT::f64, VecConstant);
4806 // It's a float: cast and extract a vector element.
4807 SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
4809 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
4810 DAG.getConstant(0, MVT::i32));
4816 // check if an VEXT instruction can handle the shuffle mask when the
4817 // vector sources of the shuffle are the same.
4818 static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
4819 unsigned NumElts = VT.getVectorNumElements();
4821 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4827 // If this is a VEXT shuffle, the immediate value is the index of the first
4828 // element. The other shuffle indices must be the successive elements after
4830 unsigned ExpectedElt = Imm;
4831 for (unsigned i = 1; i < NumElts; ++i) {
4832 // Increment the expected index. If it wraps around, just follow it
4833 // back to index zero and keep going.
4835 if (ExpectedElt == NumElts)
4838 if (M[i] < 0) continue; // ignore UNDEF indices
4839 if (ExpectedElt != static_cast<unsigned>(M[i]))
4847 static bool isVEXTMask(ArrayRef<int> M, EVT VT,
4848 bool &ReverseVEXT, unsigned &Imm) {
4849 unsigned NumElts = VT.getVectorNumElements();
4850 ReverseVEXT = false;
4852 // Assume that the first shuffle index is not UNDEF. Fail if it is.
4858 // If this is a VEXT shuffle, the immediate value is the index of the first
4859 // element. The other shuffle indices must be the successive elements after
4861 unsigned ExpectedElt = Imm;
4862 for (unsigned i = 1; i < NumElts; ++i) {
4863 // Increment the expected index. If it wraps around, it may still be
4864 // a VEXT but the source vectors must be swapped.
4866 if (ExpectedElt == NumElts * 2) {
4871 if (M[i] < 0) continue; // ignore UNDEF indices
4872 if (ExpectedElt != static_cast<unsigned>(M[i]))
4876 // Adjust the index value if the source operands will be swapped.
4883 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
4884 /// instruction with the specified blocksize. (The order of the elements
4885 /// within each block of the vector is reversed.)
4886 static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4887 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
4888 "Only possible block sizes for VREV are: 16, 32, 64");
4890 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4894 unsigned NumElts = VT.getVectorNumElements();
4895 unsigned BlockElts = M[0] + 1;
4896 // If the first shuffle index is UNDEF, be optimistic.
4898 BlockElts = BlockSize / EltSz;
4900 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4903 for (unsigned i = 0; i < NumElts; ++i) {
4904 if (M[i] < 0) continue; // ignore UNDEF indices
4905 if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
4912 static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
4913 // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
4914 // range, then 0 is placed into the resulting vector. So pretty much any mask
4915 // of 8 elements can work here.
4916 return VT == MVT::v8i8 && M.size() == 8;
4919 static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4920 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4924 unsigned NumElts = VT.getVectorNumElements();
4925 WhichResult = (M[0] == 0 ? 0 : 1);
4926 for (unsigned i = 0; i < NumElts; i += 2) {
4927 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4928 (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
4934 /// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
4935 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4936 /// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
4937 static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4938 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4942 unsigned NumElts = VT.getVectorNumElements();
4943 WhichResult = (M[0] == 0 ? 0 : 1);
4944 for (unsigned i = 0; i < NumElts; i += 2) {
4945 if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
4946 (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
4952 static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
4953 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4957 unsigned NumElts = VT.getVectorNumElements();
4958 WhichResult = (M[0] == 0 ? 0 : 1);
4959 for (unsigned i = 0; i != NumElts; ++i) {
4960 if (M[i] < 0) continue; // ignore UNDEF indices
4961 if ((unsigned) M[i] != 2 * i + WhichResult)
4965 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4966 if (VT.is64BitVector() && EltSz == 32)
4972 /// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
4973 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4974 /// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
4975 static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4976 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4980 unsigned Half = VT.getVectorNumElements() / 2;
4981 WhichResult = (M[0] == 0 ? 0 : 1);
4982 for (unsigned j = 0; j != 2; ++j) {
4983 unsigned Idx = WhichResult;
4984 for (unsigned i = 0; i != Half; ++i) {
4985 int MIdx = M[i + j * Half];
4986 if (MIdx >= 0 && (unsigned) MIdx != Idx)
4992 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4993 if (VT.is64BitVector() && EltSz == 32)
4999 static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
5000 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5004 unsigned NumElts = VT.getVectorNumElements();
5005 WhichResult = (M[0] == 0 ? 0 : 1);
5006 unsigned Idx = WhichResult * NumElts / 2;
5007 for (unsigned i = 0; i != NumElts; i += 2) {
5008 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5009 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
5014 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5015 if (VT.is64BitVector() && EltSz == 32)
5021 /// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
5022 /// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
5023 /// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
5024 static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
5025 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
5029 unsigned NumElts = VT.getVectorNumElements();
5030 WhichResult = (M[0] == 0 ? 0 : 1);
5031 unsigned Idx = WhichResult * NumElts / 2;
5032 for (unsigned i = 0; i != NumElts; i += 2) {
5033 if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
5034 (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
5039 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
5040 if (VT.is64BitVector() && EltSz == 32)
5046 /// \return true if this is a reverse operation on an vector.
5047 static bool isReverseMask(ArrayRef<int> M, EVT VT) {
5048 unsigned NumElts = VT.getVectorNumElements();
5049 // Make sure the mask has the right size.
5050 if (NumElts != M.size())
5053 // Look for <15, ..., 3, -1, 1, 0>.
5054 for (unsigned i = 0; i != NumElts; ++i)
5055 if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
5061 // If N is an integer constant that can be moved into a register in one
5062 // instruction, return an SDValue of such a constant (will become a MOV
5063 // instruction). Otherwise return null.
5064 static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
5065 const ARMSubtarget *ST, SDLoc dl) {
5067 if (!isa<ConstantSDNode>(N))
5069 Val = cast<ConstantSDNode>(N)->getZExtValue();
5071 if (ST->isThumb1Only()) {
5072 if (Val <= 255 || ~Val <= 255)
5073 return DAG.getConstant(Val, MVT::i32);
5075 if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
5076 return DAG.getConstant(Val, MVT::i32);
5081 // If this is a case we can't handle, return null and let the default
5082 // expansion code take care of it.
5083 SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
5084 const ARMSubtarget *ST) const {
5085 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
5087 EVT VT = Op.getValueType();
5089 APInt SplatBits, SplatUndef;
5090 unsigned SplatBitSize;
5092 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
5093 if (SplatBitSize <= 64) {
5094 // Check if an immediate VMOV works.
5096 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
5097 SplatUndef.getZExtValue(), SplatBitSize,
5098 DAG, VmovVT, VT.is128BitVector(),
5100 if (Val.getNode()) {
5101 SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
5102 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5105 // Try an immediate VMVN.
5106 uint64_t NegatedImm = (~SplatBits).getZExtValue();
5107 Val = isNEONModifiedImm(NegatedImm,
5108 SplatUndef.getZExtValue(), SplatBitSize,
5109 DAG, VmovVT, VT.is128BitVector(),
5111 if (Val.getNode()) {
5112 SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
5113 return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
5116 // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
5117 if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
5118 int ImmVal = ARM_AM::getFP32Imm(SplatBits);
5120 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
5121 return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
5127 // Scan through the operands to see if only one value is used.
5129 // As an optimisation, even if more than one value is used it may be more
5130 // profitable to splat with one value then change some lanes.
5132 // Heuristically we decide to do this if the vector has a "dominant" value,
5133 // defined as splatted to more than half of the lanes.
5134 unsigned NumElts = VT.getVectorNumElements();
5135 bool isOnlyLowElement = true;
5136 bool usesOnlyOneValue = true;
5137 bool hasDominantValue = false;
5138 bool isConstant = true;
5140 // Map of the number of times a particular SDValue appears in the
5142 DenseMap<SDValue, unsigned> ValueCounts;
5144 for (unsigned i = 0; i < NumElts; ++i) {
5145 SDValue V = Op.getOperand(i);
5146 if (V.getOpcode() == ISD::UNDEF)
5149 isOnlyLowElement = false;
5150 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
5153 ValueCounts.insert(std::make_pair(V, 0));
5154 unsigned &Count = ValueCounts[V];
5156 // Is this value dominant? (takes up more than half of the lanes)
5157 if (++Count > (NumElts / 2)) {
5158 hasDominantValue = true;
5162 if (ValueCounts.size() != 1)
5163 usesOnlyOneValue = false;
5164 if (!Value.getNode() && ValueCounts.size() > 0)
5165 Value = ValueCounts.begin()->first;
5167 if (ValueCounts.size() == 0)
5168 return DAG.getUNDEF(VT);
5170 // Loads are better lowered with insert_vector_elt/ARMISD::BUILD_VECTOR.
5171 // Keep going if we are hitting this case.
5172 if (isOnlyLowElement && !ISD::isNormalLoad(Value.getNode()))
5173 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
5175 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5177 // Use VDUP for non-constant splats. For f32 constant splats, reduce to
5178 // i32 and try again.
5179 if (hasDominantValue && EltSize <= 32) {
5183 // If we are VDUPing a value that comes directly from a vector, that will
5184 // cause an unnecessary move to and from a GPR, where instead we could
5185 // just use VDUPLANE. We can only do this if the lane being extracted
5186 // is at a constant index, as the VDUP from lane instructions only have
5187 // constant-index forms.
5188 if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5189 isa<ConstantSDNode>(Value->getOperand(1))) {
5190 // We need to create a new undef vector to use for the VDUPLANE if the
5191 // size of the vector from which we get the value is different than the
5192 // size of the vector that we need to create. We will insert the element
5193 // such that the register coalescer will remove unnecessary copies.
5194 if (VT != Value->getOperand(0).getValueType()) {
5195 ConstantSDNode *constIndex;
5196 constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
5197 assert(constIndex && "The index is not a constant!");
5198 unsigned index = constIndex->getAPIntValue().getLimitedValue() %
5199 VT.getVectorNumElements();
5200 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5201 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
5202 Value, DAG.getConstant(index, MVT::i32)),
5203 DAG.getConstant(index, MVT::i32));
5205 N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5206 Value->getOperand(0), Value->getOperand(1));
5208 N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
5210 if (!usesOnlyOneValue) {
5211 // The dominant value was splatted as 'N', but we now have to insert
5212 // all differing elements.
5213 for (unsigned I = 0; I < NumElts; ++I) {
5214 if (Op.getOperand(I) == Value)
5216 SmallVector<SDValue, 3> Ops;
5218 Ops.push_back(Op.getOperand(I));
5219 Ops.push_back(DAG.getConstant(I, MVT::i32));
5220 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ops);
5225 if (VT.getVectorElementType().isFloatingPoint()) {
5226 SmallVector<SDValue, 8> Ops;
5227 for (unsigned i = 0; i < NumElts; ++i)
5228 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
5230 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
5231 SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
5232 Val = LowerBUILD_VECTOR(Val, DAG, ST);
5234 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5236 if (usesOnlyOneValue) {
5237 SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
5238 if (isConstant && Val.getNode())
5239 return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
5243 // If all elements are constants and the case above didn't get hit, fall back
5244 // to the default expansion, which will generate a load from the constant
5249 // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
5251 SDValue shuffle = ReconstructShuffle(Op, DAG);
5252 if (shuffle != SDValue())
5256 // Vectors with 32- or 64-bit elements can be built by directly assigning
5257 // the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
5258 // will be legalized.
5259 if (EltSize >= 32) {
5260 // Do the expansion with floating-point types, since that is what the VFP
5261 // registers are defined to use, and since i64 is not legal.
5262 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5263 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5264 SmallVector<SDValue, 8> Ops;
5265 for (unsigned i = 0; i < NumElts; ++i)
5266 Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
5267 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5268 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5271 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
5272 // know the default expansion would otherwise fall back on something even
5273 // worse. For a vector with one or two non-undef values, that's
5274 // scalar_to_vector for the elements followed by a shuffle (provided the
5275 // shuffle is valid for the target) and materialization element by element
5276 // on the stack followed by a load for everything else.
5277 if (!isConstant && !usesOnlyOneValue) {
5278 SDValue Vec = DAG.getUNDEF(VT);
5279 for (unsigned i = 0 ; i < NumElts; ++i) {
5280 SDValue V = Op.getOperand(i);
5281 if (V.getOpcode() == ISD::UNDEF)
5283 SDValue LaneIdx = DAG.getConstant(i, MVT::i32);
5284 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Vec, V, LaneIdx);
5292 // Gather data to see if the operation can be modelled as a
5293 // shuffle in combination with VEXTs.
5294 SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
5295 SelectionDAG &DAG) const {
5297 EVT VT = Op.getValueType();
5298 unsigned NumElts = VT.getVectorNumElements();
5300 SmallVector<SDValue, 2> SourceVecs;
5301 SmallVector<unsigned, 2> MinElts;
5302 SmallVector<unsigned, 2> MaxElts;
5304 for (unsigned i = 0; i < NumElts; ++i) {
5305 SDValue V = Op.getOperand(i);
5306 if (V.getOpcode() == ISD::UNDEF)
5308 else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
5309 // A shuffle can only come from building a vector from various
5310 // elements of other vectors.
5312 } else if (V.getOperand(0).getValueType().getVectorElementType() !=
5313 VT.getVectorElementType()) {
5314 // This code doesn't know how to handle shuffles where the vector
5315 // element types do not match (this happens because type legalization
5316 // promotes the return type of EXTRACT_VECTOR_ELT).
5317 // FIXME: It might be appropriate to extend this code to handle
5318 // mismatched types.
5322 // Record this extraction against the appropriate vector if possible...
5323 SDValue SourceVec = V.getOperand(0);
5324 // If the element number isn't a constant, we can't effectively
5325 // analyze what's going on.
5326 if (!isa<ConstantSDNode>(V.getOperand(1)))
5328 unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
5329 bool FoundSource = false;
5330 for (unsigned j = 0; j < SourceVecs.size(); ++j) {
5331 if (SourceVecs[j] == SourceVec) {
5332 if (MinElts[j] > EltNo)
5334 if (MaxElts[j] < EltNo)
5341 // Or record a new source if not...
5343 SourceVecs.push_back(SourceVec);
5344 MinElts.push_back(EltNo);
5345 MaxElts.push_back(EltNo);
5349 // Currently only do something sane when at most two source vectors
5351 if (SourceVecs.size() > 2)
5354 SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
5355 int VEXTOffsets[2] = {0, 0};
5357 // This loop extracts the usage patterns of the source vectors
5358 // and prepares appropriate SDValues for a shuffle if possible.
5359 for (unsigned i = 0; i < SourceVecs.size(); ++i) {
5360 if (SourceVecs[i].getValueType() == VT) {
5361 // No VEXT necessary
5362 ShuffleSrcs[i] = SourceVecs[i];
5365 } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
5366 // It probably isn't worth padding out a smaller vector just to
5367 // break it down again in a shuffle.
5371 // Since only 64-bit and 128-bit vectors are legal on ARM and
5372 // we've eliminated the other cases...
5373 assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
5374 "unexpected vector sizes in ReconstructShuffle");
5376 if (MaxElts[i] - MinElts[i] >= NumElts) {
5377 // Span too large for a VEXT to cope
5381 if (MinElts[i] >= NumElts) {
5382 // The extraction can just take the second half
5383 VEXTOffsets[i] = NumElts;
5384 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5386 DAG.getIntPtrConstant(NumElts));
5387 } else if (MaxElts[i] < NumElts) {
5388 // The extraction can just take the first half
5390 ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5392 DAG.getIntPtrConstant(0));
5394 // An actual VEXT is needed
5395 VEXTOffsets[i] = MinElts[i];
5396 SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5398 DAG.getIntPtrConstant(0));
5399 SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
5401 DAG.getIntPtrConstant(NumElts));
5402 ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
5403 DAG.getConstant(VEXTOffsets[i], MVT::i32));
5407 SmallVector<int, 8> Mask;
5409 for (unsigned i = 0; i < NumElts; ++i) {
5410 SDValue Entry = Op.getOperand(i);
5411 if (Entry.getOpcode() == ISD::UNDEF) {
5416 SDValue ExtractVec = Entry.getOperand(0);
5417 int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
5418 .getOperand(1))->getSExtValue();
5419 if (ExtractVec == SourceVecs[0]) {
5420 Mask.push_back(ExtractElt - VEXTOffsets[0]);
5422 Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
5426 // Final check before we try to produce nonsense...
5427 if (isShuffleMaskLegal(Mask, VT))
5428 return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
5434 /// isShuffleMaskLegal - Targets can use this to indicate that they only
5435 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
5436 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
5437 /// are assumed to be legal.
5439 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
5441 if (VT.getVectorNumElements() == 4 &&
5442 (VT.is128BitVector() || VT.is64BitVector())) {
5443 unsigned PFIndexes[4];
5444 for (unsigned i = 0; i != 4; ++i) {
5448 PFIndexes[i] = M[i];
5451 // Compute the index in the perfect shuffle table.
5452 unsigned PFTableIndex =
5453 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5454 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5455 unsigned Cost = (PFEntry >> 30);
5462 unsigned Imm, WhichResult;
5464 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5465 return (EltSize >= 32 ||
5466 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
5467 isVREVMask(M, VT, 64) ||
5468 isVREVMask(M, VT, 32) ||
5469 isVREVMask(M, VT, 16) ||
5470 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
5471 isVTBLMask(M, VT) ||
5472 isVTRNMask(M, VT, WhichResult) ||
5473 isVUZPMask(M, VT, WhichResult) ||
5474 isVZIPMask(M, VT, WhichResult) ||
5475 isVTRN_v_undef_Mask(M, VT, WhichResult) ||
5476 isVUZP_v_undef_Mask(M, VT, WhichResult) ||
5477 isVZIP_v_undef_Mask(M, VT, WhichResult) ||
5478 ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
5481 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5482 /// the specified operations to build the shuffle.
5483 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5484 SDValue RHS, SelectionDAG &DAG,
5486 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5487 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5488 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5491 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5500 OP_VUZPL, // VUZP, left result
5501 OP_VUZPR, // VUZP, right result
5502 OP_VZIPL, // VZIP, left result
5503 OP_VZIPR, // VZIP, right result
5504 OP_VTRNL, // VTRN, left result
5505 OP_VTRNR // VTRN, right result
5508 if (OpNum == OP_COPY) {
5509 if (LHSID == (1*9+2)*9+3) return LHS;
5510 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5514 SDValue OpLHS, OpRHS;
5515 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5516 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5517 EVT VT = OpLHS.getValueType();
5520 default: llvm_unreachable("Unknown shuffle opcode!");
5522 // VREV divides the vector in half and swaps within the half.
5523 if (VT.getVectorElementType() == MVT::i32 ||
5524 VT.getVectorElementType() == MVT::f32)
5525 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
5526 // vrev <4 x i16> -> VREV32
5527 if (VT.getVectorElementType() == MVT::i16)
5528 return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
5529 // vrev <4 x i8> -> VREV16
5530 assert(VT.getVectorElementType() == MVT::i8);
5531 return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
5536 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
5537 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
5541 return DAG.getNode(ARMISD::VEXT, dl, VT,
5543 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
5546 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5547 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
5550 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5551 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
5554 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5555 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
5559 static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
5560 ArrayRef<int> ShuffleMask,
5561 SelectionDAG &DAG) {
5562 // Check to see if we can use the VTBL instruction.
5563 SDValue V1 = Op.getOperand(0);
5564 SDValue V2 = Op.getOperand(1);
5567 SmallVector<SDValue, 8> VTBLMask;
5568 for (ArrayRef<int>::iterator
5569 I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
5570 VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
5572 if (V2.getNode()->getOpcode() == ISD::UNDEF)
5573 return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
5574 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5576 return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
5577 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8, VTBLMask));
5580 static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
5581 SelectionDAG &DAG) {
5583 SDValue OpLHS = Op.getOperand(0);
5584 EVT VT = OpLHS.getValueType();
5586 assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
5587 "Expect an v8i16/v16i8 type");
5588 OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
5589 // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
5590 // extract the first 8 bytes into the top double word and the last 8 bytes
5591 // into the bottom double word. The v8i16 case is similar.
5592 unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
5593 return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
5594 DAG.getConstant(ExtractNum, MVT::i32));
5597 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
5598 SDValue V1 = Op.getOperand(0);
5599 SDValue V2 = Op.getOperand(1);
5601 EVT VT = Op.getValueType();
5602 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5604 // Convert shuffles that are directly supported on NEON to target-specific
5605 // DAG nodes, instead of keeping them as shuffles and matching them again
5606 // during code selection. This is more efficient and avoids the possibility
5607 // of inconsistencies between legalization and selection.
5608 // FIXME: floating-point vectors should be canonicalized to integer vectors
5609 // of the same time so that they get CSEd properly.
5610 ArrayRef<int> ShuffleMask = SVN->getMask();
5612 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5613 if (EltSize <= 32) {
5614 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5615 int Lane = SVN->getSplatIndex();
5616 // If this is undef splat, generate it via "just" vdup, if possible.
5617 if (Lane == -1) Lane = 0;
5619 // Test if V1 is a SCALAR_TO_VECTOR.
5620 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5621 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5623 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
5624 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
5626 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
5627 !isa<ConstantSDNode>(V1.getOperand(0))) {
5628 bool IsScalarToVector = true;
5629 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
5630 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
5631 IsScalarToVector = false;
5634 if (IsScalarToVector)
5635 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
5637 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
5638 DAG.getConstant(Lane, MVT::i32));
5643 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
5646 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
5647 DAG.getConstant(Imm, MVT::i32));
5650 if (isVREVMask(ShuffleMask, VT, 64))
5651 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
5652 if (isVREVMask(ShuffleMask, VT, 32))
5653 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
5654 if (isVREVMask(ShuffleMask, VT, 16))
5655 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
5657 if (V2->getOpcode() == ISD::UNDEF &&
5658 isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
5659 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
5660 DAG.getConstant(Imm, MVT::i32));
5663 // Check for Neon shuffles that modify both input vectors in place.
5664 // If both results are used, i.e., if there are two shuffles with the same
5665 // source operands and with masks corresponding to both results of one of
5666 // these operations, DAG memoization will ensure that a single node is
5667 // used for both shuffles.
5668 unsigned WhichResult;
5669 if (isVTRNMask(ShuffleMask, VT, WhichResult))
5670 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5671 V1, V2).getValue(WhichResult);
5672 if (isVUZPMask(ShuffleMask, VT, WhichResult))
5673 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5674 V1, V2).getValue(WhichResult);
5675 if (isVZIPMask(ShuffleMask, VT, WhichResult))
5676 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5677 V1, V2).getValue(WhichResult);
5679 if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
5680 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
5681 V1, V1).getValue(WhichResult);
5682 if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5683 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
5684 V1, V1).getValue(WhichResult);
5685 if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
5686 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
5687 V1, V1).getValue(WhichResult);
5690 // If the shuffle is not directly supported and it has 4 elements, use
5691 // the PerfectShuffle-generated table to synthesize it from other shuffles.
5692 unsigned NumElts = VT.getVectorNumElements();
5694 unsigned PFIndexes[4];
5695 for (unsigned i = 0; i != 4; ++i) {
5696 if (ShuffleMask[i] < 0)
5699 PFIndexes[i] = ShuffleMask[i];
5702 // Compute the index in the perfect shuffle table.
5703 unsigned PFTableIndex =
5704 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5705 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5706 unsigned Cost = (PFEntry >> 30);
5709 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5712 // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
5713 if (EltSize >= 32) {
5714 // Do the expansion with floating-point types, since that is what the VFP
5715 // registers are defined to use, and since i64 is not legal.
5716 EVT EltVT = EVT::getFloatingPointVT(EltSize);
5717 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
5718 V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
5719 V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
5720 SmallVector<SDValue, 8> Ops;
5721 for (unsigned i = 0; i < NumElts; ++i) {
5722 if (ShuffleMask[i] < 0)
5723 Ops.push_back(DAG.getUNDEF(EltVT));
5725 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
5726 ShuffleMask[i] < (int)NumElts ? V1 : V2,
5727 DAG.getConstant(ShuffleMask[i] & (NumElts-1),
5730 SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, Ops);
5731 return DAG.getNode(ISD::BITCAST, dl, VT, Val);
5734 if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
5735 return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
5737 if (VT == MVT::v8i8) {
5738 SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
5739 if (NewOp.getNode())
5746 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5747 // INSERT_VECTOR_ELT is legal only for immediate indexes.
5748 SDValue Lane = Op.getOperand(2);
5749 if (!isa<ConstantSDNode>(Lane))
5755 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
5756 // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
5757 SDValue Lane = Op.getOperand(1);
5758 if (!isa<ConstantSDNode>(Lane))
5761 SDValue Vec = Op.getOperand(0);
5762 if (Op.getValueType() == MVT::i32 &&
5763 Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
5765 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
5771 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5772 // The only time a CONCAT_VECTORS operation can have legal types is when
5773 // two 64-bit vectors are concatenated to a 128-bit vector.
5774 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
5775 "unexpected CONCAT_VECTORS");
5777 SDValue Val = DAG.getUNDEF(MVT::v2f64);
5778 SDValue Op0 = Op.getOperand(0);
5779 SDValue Op1 = Op.getOperand(1);
5780 if (Op0.getOpcode() != ISD::UNDEF)
5781 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5782 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
5783 DAG.getIntPtrConstant(0));
5784 if (Op1.getOpcode() != ISD::UNDEF)
5785 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
5786 DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
5787 DAG.getIntPtrConstant(1));
5788 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
5791 /// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
5792 /// element has been zero/sign-extended, depending on the isSigned parameter,
5793 /// from an integer type half its size.
5794 static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
5796 // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
5797 EVT VT = N->getValueType(0);
5798 if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
5799 SDNode *BVN = N->getOperand(0).getNode();
5800 if (BVN->getValueType(0) != MVT::v4i32 ||
5801 BVN->getOpcode() != ISD::BUILD_VECTOR)
5803 unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5804 unsigned HiElt = 1 - LoElt;
5805 ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
5806 ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
5807 ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
5808 ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
5809 if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
5812 if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
5813 Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
5816 if (Hi0->isNullValue() && Hi1->isNullValue())
5822 if (N->getOpcode() != ISD::BUILD_VECTOR)
5825 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
5826 SDNode *Elt = N->getOperand(i).getNode();
5827 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
5828 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5829 unsigned HalfSize = EltSize / 2;
5831 if (!isIntN(HalfSize, C->getSExtValue()))
5834 if (!isUIntN(HalfSize, C->getZExtValue()))
5845 /// isSignExtended - Check if a node is a vector value that is sign-extended
5846 /// or a constant BUILD_VECTOR with sign-extended elements.
5847 static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
5848 if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
5850 if (isExtendedBUILD_VECTOR(N, DAG, true))
5855 /// isZeroExtended - Check if a node is a vector value that is zero-extended
5856 /// or a constant BUILD_VECTOR with zero-extended elements.
5857 static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
5858 if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
5860 if (isExtendedBUILD_VECTOR(N, DAG, false))
5865 static EVT getExtensionTo64Bits(const EVT &OrigVT) {
5866 if (OrigVT.getSizeInBits() >= 64)
5869 assert(OrigVT.isSimple() && "Expecting a simple value type");
5871 MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
5872 switch (OrigSimpleTy) {
5873 default: llvm_unreachable("Unexpected Vector Type");
5882 /// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
5883 /// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
5884 /// We insert the required extension here to get the vector to fill a D register.
5885 static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
5888 unsigned ExtOpcode) {
5889 // The vector originally had a size of OrigTy. It was then extended to ExtTy.
5890 // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
5891 // 64-bits we need to insert a new extension so that it will be 64-bits.
5892 assert(ExtTy.is128BitVector() && "Unexpected extension size");
5893 if (OrigTy.getSizeInBits() >= 64)
5896 // Must extend size to at least 64 bits to be used as an operand for VMULL.
5897 EVT NewVT = getExtensionTo64Bits(OrigTy);
5899 return DAG.getNode(ExtOpcode, SDLoc(N), NewVT, N);
5902 /// SkipLoadExtensionForVMULL - return a load of the original vector size that
5903 /// does not do any sign/zero extension. If the original vector is less
5904 /// than 64 bits, an appropriate extension will be added after the load to
5905 /// reach a total size of 64 bits. We have to add the extension separately
5906 /// because ARM does not have a sign/zero extending load for vectors.
5907 static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
5908 EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());
5910 // The load already has the right type.
5911 if (ExtendedTy == LD->getMemoryVT())
5912 return DAG.getLoad(LD->getMemoryVT(), SDLoc(LD), LD->getChain(),
5913 LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
5914 LD->isNonTemporal(), LD->isInvariant(),
5915 LD->getAlignment());
5917 // We need to create a zextload/sextload. We cannot just create a load
5918 // followed by a zext/zext node because LowerMUL is also run during normal
5919 // operation legalization where we can't create illegal types.
5920 return DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), ExtendedTy,
5921 LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
5922 LD->getMemoryVT(), LD->isVolatile(), LD->isInvariant(),
5923 LD->isNonTemporal(), LD->getAlignment());
5926 /// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
5927 /// extending load, or BUILD_VECTOR with extended elements, return the
5928 /// unextended value. The unextended vector should be 64 bits so that it can
5929 /// be used as an operand to a VMULL instruction. If the original vector size
5930 /// before extension is less than 64 bits we add a an extension to resize
5931 /// the vector to 64 bits.
5932 static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
5933 if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
5934 return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
5935 N->getOperand(0)->getValueType(0),
5939 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
5940 return SkipLoadExtensionForVMULL(LD, DAG);
5942 // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
5943 // have been legalized as a BITCAST from v4i32.
5944 if (N->getOpcode() == ISD::BITCAST) {
5945 SDNode *BVN = N->getOperand(0).getNode();
5946 assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
5947 BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
5948 unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
5949 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), MVT::v2i32,
5950 BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
5952 // Construct a new BUILD_VECTOR with elements truncated to half the size.
5953 assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
5954 EVT VT = N->getValueType(0);
5955 unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
5956 unsigned NumElts = VT.getVectorNumElements();
5957 MVT TruncVT = MVT::getIntegerVT(EltSize);
5958 SmallVector<SDValue, 8> Ops;
5959 for (unsigned i = 0; i != NumElts; ++i) {
5960 ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
5961 const APInt &CInt = C->getAPIntValue();
5962 // Element types smaller than 32 bits are not legal, so use i32 elements.
5963 // The values are implicitly truncated so sext vs. zext doesn't matter.
5964 Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
5966 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
5967 MVT::getVectorVT(TruncVT, NumElts), Ops);
5970 static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
5971 unsigned Opcode = N->getOpcode();
5972 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5973 SDNode *N0 = N->getOperand(0).getNode();
5974 SDNode *N1 = N->getOperand(1).getNode();
5975 return N0->hasOneUse() && N1->hasOneUse() &&
5976 isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
5981 static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
5982 unsigned Opcode = N->getOpcode();
5983 if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
5984 SDNode *N0 = N->getOperand(0).getNode();
5985 SDNode *N1 = N->getOperand(1).getNode();
5986 return N0->hasOneUse() && N1->hasOneUse() &&
5987 isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
5992 static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
5993 // Multiplications are only custom-lowered for 128-bit vectors so that
5994 // VMULL can be detected. Otherwise v2i64 multiplications are not legal.
5995 EVT VT = Op.getValueType();
5996 assert(VT.is128BitVector() && VT.isInteger() &&
5997 "unexpected type for custom-lowering ISD::MUL");
5998 SDNode *N0 = Op.getOperand(0).getNode();
5999 SDNode *N1 = Op.getOperand(1).getNode();
6000 unsigned NewOpc = 0;
6002 bool isN0SExt = isSignExtended(N0, DAG);
6003 bool isN1SExt = isSignExtended(N1, DAG);
6004 if (isN0SExt && isN1SExt)
6005 NewOpc = ARMISD::VMULLs;
6007 bool isN0ZExt = isZeroExtended(N0, DAG);
6008 bool isN1ZExt = isZeroExtended(N1, DAG);
6009 if (isN0ZExt && isN1ZExt)
6010 NewOpc = ARMISD::VMULLu;
6011 else if (isN1SExt || isN1ZExt) {
6012 // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
6013 // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
6014 if (isN1SExt && isAddSubSExt(N0, DAG)) {
6015 NewOpc = ARMISD::VMULLs;
6017 } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
6018 NewOpc = ARMISD::VMULLu;
6020 } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
6022 NewOpc = ARMISD::VMULLu;
6028 if (VT == MVT::v2i64)
6029 // Fall through to expand this. It is not legal.
6032 // Other vector multiplications are legal.
6037 // Legalize to a VMULL instruction.
6040 SDValue Op1 = SkipExtensionForVMULL(N1, DAG);
6042 Op0 = SkipExtensionForVMULL(N0, DAG);
6043 assert(Op0.getValueType().is64BitVector() &&
6044 Op1.getValueType().is64BitVector() &&
6045 "unexpected types for extended operands to VMULL");
6046 return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
6049 // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
6050 // isel lowering to take advantage of no-stall back to back vmul + vmla.
6057 SDValue N00 = SkipExtensionForVMULL(N0->getOperand(0).getNode(), DAG);
6058 SDValue N01 = SkipExtensionForVMULL(N0->getOperand(1).getNode(), DAG);
6059 EVT Op1VT = Op1.getValueType();
6060 return DAG.getNode(N0->getOpcode(), DL, VT,
6061 DAG.getNode(NewOpc, DL, VT,
6062 DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
6063 DAG.getNode(NewOpc, DL, VT,
6064 DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
6068 LowerSDIV_v4i8(SDValue X, SDValue Y, SDLoc dl, SelectionDAG &DAG) {
6070 // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
6071 // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
6072 X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
6073 Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
6074 X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
6075 Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
6076 // Get reciprocal estimate.
6077 // float4 recip = vrecpeq_f32(yf);
6078 Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6079 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
6080 // Because char has a smaller range than uchar, we can actually get away
6081 // without any newton steps. This requires that we use a weird bias
6082 // of 0xb000, however (again, this has been exhaustively tested).
6083 // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
6084 X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
6085 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
6086 Y = DAG.getConstant(0xb000, MVT::i32);
6087 Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
6088 X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
6089 X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
6090 // Convert back to short.
6091 X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
6092 X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
6097 LowerSDIV_v4i16(SDValue N0, SDValue N1, SDLoc dl, SelectionDAG &DAG) {
6099 // Convert to float.
6100 // float4 yf = vcvt_f32_s32(vmovl_s16(y));
6101 // float4 xf = vcvt_f32_s32(vmovl_s16(x));
6102 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
6103 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
6104 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6105 N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6107 // Use reciprocal estimate and one refinement step.
6108 // float4 recip = vrecpeq_f32(yf);
6109 // recip *= vrecpsq_f32(yf, recip);
6110 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6111 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
6112 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6113 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6115 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6116 // Because short has a smaller range than ushort, we can actually get away
6117 // with only a single newton step. This requires that we use a weird bias
6118 // of 89, however (again, this has been exhaustively tested).
6119 // float4 result = as_float4(as_int4(xf*recip) + 0x89);
6120 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6121 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6122 N1 = DAG.getConstant(0x89, MVT::i32);
6123 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6124 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6125 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6126 // Convert back to integer and return.
6127 // return vmovn_s32(vcvt_s32_f32(result));
6128 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6129 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6133 static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
6134 EVT VT = Op.getValueType();
6135 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6136 "unexpected type for custom-lowering ISD::SDIV");
6139 SDValue N0 = Op.getOperand(0);
6140 SDValue N1 = Op.getOperand(1);
6143 if (VT == MVT::v8i8) {
6144 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
6145 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
6147 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6148 DAG.getIntPtrConstant(4));
6149 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6150 DAG.getIntPtrConstant(4));
6151 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6152 DAG.getIntPtrConstant(0));
6153 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6154 DAG.getIntPtrConstant(0));
6156 N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
6157 N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
6159 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6160 N0 = LowerCONCAT_VECTORS(N0, DAG);
6162 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
6165 return LowerSDIV_v4i16(N0, N1, dl, DAG);
6168 static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
6169 EVT VT = Op.getValueType();
6170 assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
6171 "unexpected type for custom-lowering ISD::UDIV");
6174 SDValue N0 = Op.getOperand(0);
6175 SDValue N1 = Op.getOperand(1);
6178 if (VT == MVT::v8i8) {
6179 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
6180 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
6182 N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6183 DAG.getIntPtrConstant(4));
6184 N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6185 DAG.getIntPtrConstant(4));
6186 N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
6187 DAG.getIntPtrConstant(0));
6188 N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
6189 DAG.getIntPtrConstant(0));
6191 N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
6192 N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
6194 N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
6195 N0 = LowerCONCAT_VECTORS(N0, DAG);
6197 N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
6198 DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
6203 // v4i16 sdiv ... Convert to float.
6204 // float4 yf = vcvt_f32_s32(vmovl_u16(y));
6205 // float4 xf = vcvt_f32_s32(vmovl_u16(x));
6206 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
6207 N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
6208 N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
6209 SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
6211 // Use reciprocal estimate and two refinement steps.
6212 // float4 recip = vrecpeq_f32(yf);
6213 // recip *= vrecpsq_f32(yf, recip);
6214 // recip *= vrecpsq_f32(yf, recip);
6215 N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6216 DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
6217 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6218 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6220 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6221 N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
6222 DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
6224 N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
6225 // Simply multiplying by the reciprocal estimate can leave us a few ulps
6226 // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
6227 // and that it will never cause us to return an answer too large).
6228 // float4 result = as_float4(as_int4(xf*recip) + 2);
6229 N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
6230 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
6231 N1 = DAG.getConstant(2, MVT::i32);
6232 N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
6233 N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
6234 N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
6235 // Convert back to integer and return.
6236 // return vmovn_u32(vcvt_s32_f32(result));
6237 N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
6238 N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
6242 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
6243 EVT VT = Op.getNode()->getValueType(0);
6244 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
6247 bool ExtraOp = false;
6248 switch (Op.getOpcode()) {
6249 default: llvm_unreachable("Invalid code");
6250 case ISD::ADDC: Opc = ARMISD::ADDC; break;
6251 case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
6252 case ISD::SUBC: Opc = ARMISD::SUBC; break;
6253 case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
6257 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6259 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
6260 Op.getOperand(1), Op.getOperand(2));
6263 SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
6264 assert(Subtarget->isTargetDarwin());
6266 // For iOS, we want to call an alternative entry point: __sincos_stret,
6267 // return values are passed via sret.
6269 SDValue Arg = Op.getOperand(0);
6270 EVT ArgVT = Arg.getValueType();
6271 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
6273 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
6274 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6276 // Pair of floats / doubles used to pass the result.
6277 StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
6279 // Create stack object for sret.
6280 const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
6281 const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
6282 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
6283 SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
6289 Entry.Ty = RetTy->getPointerTo();
6290 Entry.isSExt = false;
6291 Entry.isZExt = false;
6292 Entry.isSRet = true;
6293 Args.push_back(Entry);
6297 Entry.isSExt = false;
6298 Entry.isZExt = false;
6299 Args.push_back(Entry);
6301 const char *LibcallName = (ArgVT == MVT::f64)
6302 ? "__sincos_stret" : "__sincosf_stret";
6303 SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
6305 TargetLowering::CallLoweringInfo CLI(DAG);
6306 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
6307 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), Callee,
6309 .setDiscardResult();
6311 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
6313 SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
6314 MachinePointerInfo(), false, false, false, 0);
6316 // Address of cos field.
6317 SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
6318 DAG.getIntPtrConstant(ArgVT.getStoreSize()));
6319 SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
6320 MachinePointerInfo(), false, false, false, 0);
6322 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
6323 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
6324 LoadSin.getValue(0), LoadCos.getValue(0));
6327 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
6328 // Monotonic load/store is legal for all targets
6329 if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
6332 // Acquire/Release load/store is not legal for targets without a
6333 // dmb or equivalent available.
6337 static void ReplaceREADCYCLECOUNTER(SDNode *N,
6338 SmallVectorImpl<SDValue> &Results,
6340 const ARMSubtarget *Subtarget) {
6342 SDValue Cycles32, OutChain;
6344 if (Subtarget->hasPerfMon()) {
6345 // Under Power Management extensions, the cycle-count is:
6346 // mrc p15, #0, <Rt>, c9, c13, #0
6347 SDValue Ops[] = { N->getOperand(0), // Chain
6348 DAG.getConstant(Intrinsic::arm_mrc, MVT::i32),
6349 DAG.getConstant(15, MVT::i32),
6350 DAG.getConstant(0, MVT::i32),
6351 DAG.getConstant(9, MVT::i32),
6352 DAG.getConstant(13, MVT::i32),
6353 DAG.getConstant(0, MVT::i32)
6356 Cycles32 = DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL,
6357 DAG.getVTList(MVT::i32, MVT::Other), Ops);
6358 OutChain = Cycles32.getValue(1);
6360 // Intrinsic is defined to return 0 on unsupported platforms. Technically
6361 // there are older ARM CPUs that have implementation-specific ways of
6362 // obtaining this information (FIXME!).
6363 Cycles32 = DAG.getConstant(0, MVT::i32);
6364 OutChain = DAG.getEntryNode();
6368 SDValue Cycles64 = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64,
6369 Cycles32, DAG.getConstant(0, MVT::i32));
6370 Results.push_back(Cycles64);
6371 Results.push_back(OutChain);
6374 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
6375 switch (Op.getOpcode()) {
6376 default: llvm_unreachable("Don't know how to custom lower this!");
6377 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6378 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
6379 case ISD::GlobalAddress:
6380 switch (Subtarget->getTargetTriple().getObjectFormat()) {
6381 default: llvm_unreachable("unknown object format");
6383 return LowerGlobalAddressWindows(Op, DAG);
6385 return LowerGlobalAddressELF(Op, DAG);
6387 return LowerGlobalAddressDarwin(Op, DAG);
6389 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6390 case ISD::SELECT: return LowerSELECT(Op, DAG);
6391 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
6392 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
6393 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
6394 case ISD::VASTART: return LowerVASTART(Op, DAG);
6395 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
6396 case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
6397 case ISD::SINT_TO_FP:
6398 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
6399 case ISD::FP_TO_SINT:
6400 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
6401 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6402 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6403 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6404 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
6405 case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
6406 case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
6407 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
6409 case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
6412 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
6413 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
6414 case ISD::SRL_PARTS:
6415 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
6416 case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
6417 case ISD::CTPOP: return LowerCTPOP(Op.getNode(), DAG, Subtarget);
6418 case ISD::SETCC: return LowerVSETCC(Op, DAG);
6419 case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
6420 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
6421 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6422 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6423 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6424 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
6425 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6426 case ISD::MUL: return LowerMUL(Op, DAG);
6427 case ISD::SDIV: return LowerSDIV(Op, DAG);
6428 case ISD::UDIV: return LowerUDIV(Op, DAG);
6432 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
6437 return LowerXALUO(Op, DAG);
6438 case ISD::ATOMIC_LOAD:
6439 case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
6440 case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
6442 case ISD::UDIVREM: return LowerDivRem(Op, DAG);
6443 case ISD::DYNAMIC_STACKALLOC:
6444 if (Subtarget->getTargetTriple().isWindowsItaniumEnvironment())
6445 return LowerDYNAMIC_STACKALLOC(Op, DAG);
6446 llvm_unreachable("Don't know how to custom lower this!");
6447 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
6448 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
6452 /// ReplaceNodeResults - Replace the results of node with an illegal result
6453 /// type with new values built out of custom code.
6454 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
6455 SmallVectorImpl<SDValue>&Results,
6456 SelectionDAG &DAG) const {
6458 switch (N->getOpcode()) {
6460 llvm_unreachable("Don't know how to custom expand this!");
6462 Res = ExpandBITCAST(N, DAG);
6466 Res = Expand64BitShift(N, DAG, Subtarget);
6468 case ISD::READCYCLECOUNTER:
6469 ReplaceREADCYCLECOUNTER(N, Results, DAG, Subtarget);
6473 Results.push_back(Res);
6476 //===----------------------------------------------------------------------===//
6477 // ARM Scheduler Hooks
6478 //===----------------------------------------------------------------------===//
6480 /// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
6481 /// registers the function context.
6482 void ARMTargetLowering::
6483 SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
6484 MachineBasicBlock *DispatchBB, int FI) const {
6485 const TargetInstrInfo *TII =
6486 getTargetMachine().getSubtargetImpl()->getInstrInfo();
6487 DebugLoc dl = MI->getDebugLoc();
6488 MachineFunction *MF = MBB->getParent();
6489 MachineRegisterInfo *MRI = &MF->getRegInfo();
6490 MachineConstantPool *MCP = MF->getConstantPool();
6491 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6492 const Function *F = MF->getFunction();
6494 bool isThumb = Subtarget->isThumb();
6495 bool isThumb2 = Subtarget->isThumb2();
6497 unsigned PCLabelId = AFI->createPICLabelUId();
6498 unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
6499 ARMConstantPoolValue *CPV =
6500 ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
6501 unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
6503 const TargetRegisterClass *TRC = isThumb ?
6504 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6505 (const TargetRegisterClass*)&ARM::GPRRegClass;
6507 // Grab constant pool and fixed stack memory operands.
6508 MachineMemOperand *CPMMO =
6509 MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
6510 MachineMemOperand::MOLoad, 4, 4);
6512 MachineMemOperand *FIMMOSt =
6513 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6514 MachineMemOperand::MOStore, 4, 4);
6516 // Load the address of the dispatch MBB into the jump buffer.
6518 // Incoming value: jbuf
6519 // ldr.n r5, LCPI1_1
6522 // str r5, [$jbuf, #+4] ; &jbuf[1]
6523 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6524 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
6525 .addConstantPoolIndex(CPI)
6526 .addMemOperand(CPMMO));
6527 // Set the low bit because of thumb mode.
6528 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6530 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
6531 .addReg(NewVReg1, RegState::Kill)
6533 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6534 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
6535 .addReg(NewVReg2, RegState::Kill)
6537 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
6538 .addReg(NewVReg3, RegState::Kill)
6540 .addImm(36) // &jbuf[1] :: pc
6541 .addMemOperand(FIMMOSt));
6542 } else if (isThumb) {
6543 // Incoming value: jbuf
6544 // ldr.n r1, LCPI1_4
6548 // add r2, $jbuf, #+4 ; &jbuf[1]
6550 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6551 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
6552 .addConstantPoolIndex(CPI)
6553 .addMemOperand(CPMMO));
6554 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6555 BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
6556 .addReg(NewVReg1, RegState::Kill)
6558 // Set the low bit because of thumb mode.
6559 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6560 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
6561 .addReg(ARM::CPSR, RegState::Define)
6563 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6564 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
6565 .addReg(ARM::CPSR, RegState::Define)
6566 .addReg(NewVReg2, RegState::Kill)
6567 .addReg(NewVReg3, RegState::Kill));
6568 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6569 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
6571 .addImm(36)); // &jbuf[1] :: pc
6572 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
6573 .addReg(NewVReg4, RegState::Kill)
6574 .addReg(NewVReg5, RegState::Kill)
6576 .addMemOperand(FIMMOSt));
6578 // Incoming value: jbuf
6581 // str r1, [$jbuf, #+4] ; &jbuf[1]
6582 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6583 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
6584 .addConstantPoolIndex(CPI)
6586 .addMemOperand(CPMMO));
6587 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6588 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
6589 .addReg(NewVReg1, RegState::Kill)
6590 .addImm(PCLabelId));
6591 AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
6592 .addReg(NewVReg2, RegState::Kill)
6594 .addImm(36) // &jbuf[1] :: pc
6595 .addMemOperand(FIMMOSt));
6599 MachineBasicBlock *ARMTargetLowering::
6600 EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
6601 const TargetInstrInfo *TII =
6602 getTargetMachine().getSubtargetImpl()->getInstrInfo();
6603 DebugLoc dl = MI->getDebugLoc();
6604 MachineFunction *MF = MBB->getParent();
6605 MachineRegisterInfo *MRI = &MF->getRegInfo();
6606 ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
6607 MachineFrameInfo *MFI = MF->getFrameInfo();
6608 int FI = MFI->getFunctionContextIndex();
6610 const TargetRegisterClass *TRC = Subtarget->isThumb() ?
6611 (const TargetRegisterClass*)&ARM::tGPRRegClass :
6612 (const TargetRegisterClass*)&ARM::GPRnopcRegClass;
6614 // Get a mapping of the call site numbers to all of the landing pads they're
6616 DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
6617 unsigned MaxCSNum = 0;
6618 MachineModuleInfo &MMI = MF->getMMI();
6619 for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E;
6621 if (!BB->isLandingPad()) continue;
6623 // FIXME: We should assert that the EH_LABEL is the first MI in the landing
6625 for (MachineBasicBlock::iterator
6626 II = BB->begin(), IE = BB->end(); II != IE; ++II) {
6627 if (!II->isEHLabel()) continue;
6629 MCSymbol *Sym = II->getOperand(0).getMCSymbol();
6630 if (!MMI.hasCallSiteLandingPad(Sym)) continue;
6632 SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
6633 for (SmallVectorImpl<unsigned>::iterator
6634 CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
6635 CSI != CSE; ++CSI) {
6636 CallSiteNumToLPad[*CSI].push_back(BB);
6637 MaxCSNum = std::max(MaxCSNum, *CSI);
6643 // Get an ordered list of the machine basic blocks for the jump table.
6644 std::vector<MachineBasicBlock*> LPadList;
6645 SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
6646 LPadList.reserve(CallSiteNumToLPad.size());
6647 for (unsigned I = 1; I <= MaxCSNum; ++I) {
6648 SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
6649 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6650 II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
6651 LPadList.push_back(*II);
6652 InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
6656 assert(!LPadList.empty() &&
6657 "No landing pad destinations for the dispatch jump table!");
6659 // Create the jump table and associated information.
6660 MachineJumpTableInfo *JTI =
6661 MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
6662 unsigned MJTI = JTI->createJumpTableIndex(LPadList);
6663 unsigned UId = AFI->createJumpTableUId();
6664 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
6666 // Create the MBBs for the dispatch code.
6668 // Shove the dispatch's address into the return slot in the function context.
6669 MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
6670 DispatchBB->setIsLandingPad();
6672 MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
6673 unsigned trap_opcode;
6674 if (Subtarget->isThumb())
6675 trap_opcode = ARM::tTRAP;
6677 trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
6679 BuildMI(TrapBB, dl, TII->get(trap_opcode));
6680 DispatchBB->addSuccessor(TrapBB);
6682 MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
6683 DispatchBB->addSuccessor(DispContBB);
6686 MF->insert(MF->end(), DispatchBB);
6687 MF->insert(MF->end(), DispContBB);
6688 MF->insert(MF->end(), TrapBB);
6690 // Insert code into the entry block that creates and registers the function
6692 SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
6694 MachineMemOperand *FIMMOLd =
6695 MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
6696 MachineMemOperand::MOLoad |
6697 MachineMemOperand::MOVolatile, 4, 4);
6699 MachineInstrBuilder MIB;
6700 MIB = BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
6702 const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6703 const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6705 // Add a register mask with no preserved registers. This results in all
6706 // registers being marked as clobbered.
6707 MIB.addRegMask(RI.getNoPreservedMask());
6709 unsigned NumLPads = LPadList.size();
6710 if (Subtarget->isThumb2()) {
6711 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6712 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
6715 .addMemOperand(FIMMOLd));
6717 if (NumLPads < 256) {
6718 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
6720 .addImm(LPadList.size()));
6722 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6723 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
6724 .addImm(NumLPads & 0xFFFF));
6726 unsigned VReg2 = VReg1;
6727 if ((NumLPads & 0xFFFF0000) != 0) {
6728 VReg2 = MRI->createVirtualRegister(TRC);
6729 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
6731 .addImm(NumLPads >> 16));
6734 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
6739 BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
6744 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6745 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
6746 .addJumpTableIndex(MJTI)
6749 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6752 BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
6753 .addReg(NewVReg3, RegState::Kill)
6755 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6757 BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
6758 .addReg(NewVReg4, RegState::Kill)
6760 .addJumpTableIndex(MJTI)
6762 } else if (Subtarget->isThumb()) {
6763 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6764 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
6767 .addMemOperand(FIMMOLd));
6769 if (NumLPads < 256) {
6770 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
6774 MachineConstantPool *ConstantPool = MF->getConstantPool();
6775 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6776 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6778 // MachineConstantPool wants an explicit alignment.
6779 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6781 Align = getDataLayout()->getTypeAllocSize(C->getType());
6782 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6784 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6785 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
6786 .addReg(VReg1, RegState::Define)
6787 .addConstantPoolIndex(Idx));
6788 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
6793 BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
6798 unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
6799 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
6800 .addReg(ARM::CPSR, RegState::Define)
6804 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6805 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
6806 .addJumpTableIndex(MJTI)
6809 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6810 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
6811 .addReg(ARM::CPSR, RegState::Define)
6812 .addReg(NewVReg2, RegState::Kill)
6815 MachineMemOperand *JTMMOLd =
6816 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6817 MachineMemOperand::MOLoad, 4, 4);
6819 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6820 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
6821 .addReg(NewVReg4, RegState::Kill)
6823 .addMemOperand(JTMMOLd));
6825 unsigned NewVReg6 = NewVReg5;
6826 if (RelocM == Reloc::PIC_) {
6827 NewVReg6 = MRI->createVirtualRegister(TRC);
6828 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
6829 .addReg(ARM::CPSR, RegState::Define)
6830 .addReg(NewVReg5, RegState::Kill)
6834 BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
6835 .addReg(NewVReg6, RegState::Kill)
6836 .addJumpTableIndex(MJTI)
6839 unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6840 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6843 .addMemOperand(FIMMOLd));
6845 if (NumLPads < 256) {
6846 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6849 } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6850 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6851 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6852 .addImm(NumLPads & 0xFFFF));
6854 unsigned VReg2 = VReg1;
6855 if ((NumLPads & 0xFFFF0000) != 0) {
6856 VReg2 = MRI->createVirtualRegister(TRC);
6857 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6859 .addImm(NumLPads >> 16));
6862 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6866 MachineConstantPool *ConstantPool = MF->getConstantPool();
6867 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6868 const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6870 // MachineConstantPool wants an explicit alignment.
6871 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
6873 Align = getDataLayout()->getTypeAllocSize(C->getType());
6874 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6876 unsigned VReg1 = MRI->createVirtualRegister(TRC);
6877 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6878 .addReg(VReg1, RegState::Define)
6879 .addConstantPoolIndex(Idx)
6881 AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6883 .addReg(VReg1, RegState::Kill));
6886 BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6891 unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6893 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6895 .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6896 unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6897 AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6898 .addJumpTableIndex(MJTI)
6901 MachineMemOperand *JTMMOLd =
6902 MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6903 MachineMemOperand::MOLoad, 4, 4);
6904 unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6906 BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6907 .addReg(NewVReg3, RegState::Kill)
6910 .addMemOperand(JTMMOLd));
6912 if (RelocM == Reloc::PIC_) {
6913 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6914 .addReg(NewVReg5, RegState::Kill)
6916 .addJumpTableIndex(MJTI)
6919 BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
6920 .addReg(NewVReg5, RegState::Kill)
6921 .addJumpTableIndex(MJTI)
6926 // Add the jump table entries as successors to the MBB.
6927 SmallPtrSet<MachineBasicBlock*, 8> SeenMBBs;
6928 for (std::vector<MachineBasicBlock*>::iterator
6929 I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6930 MachineBasicBlock *CurMBB = *I;
6931 if (SeenMBBs.insert(CurMBB))
6932 DispContBB->addSuccessor(CurMBB);
6935 // N.B. the order the invoke BBs are processed in doesn't matter here.
6936 const MCPhysReg *SavedRegs = RI.getCalleeSavedRegs(MF);
6937 SmallVector<MachineBasicBlock*, 64> MBBLPads;
6938 for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6939 I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6940 MachineBasicBlock *BB = *I;
6942 // Remove the landing pad successor from the invoke block and replace it
6943 // with the new dispatch block.
6944 SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6946 while (!Successors.empty()) {
6947 MachineBasicBlock *SMBB = Successors.pop_back_val();
6948 if (SMBB->isLandingPad()) {
6949 BB->removeSuccessor(SMBB);
6950 MBBLPads.push_back(SMBB);
6954 BB->addSuccessor(DispatchBB);
6956 // Find the invoke call and mark all of the callee-saved registers as
6957 // 'implicit defined' so that they're spilled. This prevents code from
6958 // moving instructions to before the EH block, where they will never be
6960 for (MachineBasicBlock::reverse_iterator
6961 II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6962 if (!II->isCall()) continue;
6964 DenseMap<unsigned, bool> DefRegs;
6965 for (MachineInstr::mop_iterator
6966 OI = II->operands_begin(), OE = II->operands_end();
6968 if (!OI->isReg()) continue;
6969 DefRegs[OI->getReg()] = true;
6972 MachineInstrBuilder MIB(*MF, &*II);
6974 for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6975 unsigned Reg = SavedRegs[i];
6976 if (Subtarget->isThumb2() &&
6977 !ARM::tGPRRegClass.contains(Reg) &&
6978 !ARM::hGPRRegClass.contains(Reg))
6980 if (Subtarget->isThumb1Only() && !ARM::tGPRRegClass.contains(Reg))
6982 if (!Subtarget->isThumb() && !ARM::GPRRegClass.contains(Reg))
6985 MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6992 // Mark all former landing pads as non-landing pads. The dispatch is the only
6994 for (SmallVectorImpl<MachineBasicBlock*>::iterator
6995 I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6996 (*I)->setIsLandingPad(false);
6998 // The instruction is gone now.
6999 MI->eraseFromParent();
7005 MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
7006 for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
7007 E = MBB->succ_end(); I != E; ++I)
7010 llvm_unreachable("Expecting a BB with two successors!");
7013 /// Return the load opcode for a given load size. If load size >= 8,
7014 /// neon opcode will be returned.
7015 static unsigned getLdOpcode(unsigned LdSize, bool IsThumb1, bool IsThumb2) {
7017 return LdSize == 16 ? ARM::VLD1q32wb_fixed
7018 : LdSize == 8 ? ARM::VLD1d32wb_fixed : 0;
7020 return LdSize == 4 ? ARM::tLDRi
7021 : LdSize == 2 ? ARM::tLDRHi
7022 : LdSize == 1 ? ARM::tLDRBi : 0;
7024 return LdSize == 4 ? ARM::t2LDR_POST
7025 : LdSize == 2 ? ARM::t2LDRH_POST
7026 : LdSize == 1 ? ARM::t2LDRB_POST : 0;
7027 return LdSize == 4 ? ARM::LDR_POST_IMM
7028 : LdSize == 2 ? ARM::LDRH_POST
7029 : LdSize == 1 ? ARM::LDRB_POST_IMM : 0;
7032 /// Return the store opcode for a given store size. If store size >= 8,
7033 /// neon opcode will be returned.
7034 static unsigned getStOpcode(unsigned StSize, bool IsThumb1, bool IsThumb2) {
7036 return StSize == 16 ? ARM::VST1q32wb_fixed
7037 : StSize == 8 ? ARM::VST1d32wb_fixed : 0;
7039 return StSize == 4 ? ARM::tSTRi
7040 : StSize == 2 ? ARM::tSTRHi
7041 : StSize == 1 ? ARM::tSTRBi : 0;
7043 return StSize == 4 ? ARM::t2STR_POST
7044 : StSize == 2 ? ARM::t2STRH_POST
7045 : StSize == 1 ? ARM::t2STRB_POST : 0;
7046 return StSize == 4 ? ARM::STR_POST_IMM
7047 : StSize == 2 ? ARM::STRH_POST
7048 : StSize == 1 ? ARM::STRB_POST_IMM : 0;
7051 /// Emit a post-increment load operation with given size. The instructions
7052 /// will be added to BB at Pos.
7053 static void emitPostLd(MachineBasicBlock *BB, MachineInstr *Pos,
7054 const TargetInstrInfo *TII, DebugLoc dl,
7055 unsigned LdSize, unsigned Data, unsigned AddrIn,
7056 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7057 unsigned LdOpc = getLdOpcode(LdSize, IsThumb1, IsThumb2);
7058 assert(LdOpc != 0 && "Should have a load opcode");
7060 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7061 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7063 } else if (IsThumb1) {
7064 // load + update AddrIn
7065 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7066 .addReg(AddrIn).addImm(0));
7067 MachineInstrBuilder MIB =
7068 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7069 MIB = AddDefaultT1CC(MIB);
7070 MIB.addReg(AddrIn).addImm(LdSize);
7071 AddDefaultPred(MIB);
7072 } else if (IsThumb2) {
7073 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7074 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7077 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(LdOpc), Data)
7078 .addReg(AddrOut, RegState::Define).addReg(AddrIn)
7079 .addReg(0).addImm(LdSize));
7083 /// Emit a post-increment store operation with given size. The instructions
7084 /// will be added to BB at Pos.
7085 static void emitPostSt(MachineBasicBlock *BB, MachineInstr *Pos,
7086 const TargetInstrInfo *TII, DebugLoc dl,
7087 unsigned StSize, unsigned Data, unsigned AddrIn,
7088 unsigned AddrOut, bool IsThumb1, bool IsThumb2) {
7089 unsigned StOpc = getStOpcode(StSize, IsThumb1, IsThumb2);
7090 assert(StOpc != 0 && "Should have a store opcode");
7092 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7093 .addReg(AddrIn).addImm(0).addReg(Data));
7094 } else if (IsThumb1) {
7095 // store + update AddrIn
7096 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc)).addReg(Data)
7097 .addReg(AddrIn).addImm(0));
7098 MachineInstrBuilder MIB =
7099 BuildMI(*BB, Pos, dl, TII->get(ARM::tADDi8), AddrOut);
7100 MIB = AddDefaultT1CC(MIB);
7101 MIB.addReg(AddrIn).addImm(StSize);
7102 AddDefaultPred(MIB);
7103 } else if (IsThumb2) {
7104 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7105 .addReg(Data).addReg(AddrIn).addImm(StSize));
7107 AddDefaultPred(BuildMI(*BB, Pos, dl, TII->get(StOpc), AddrOut)
7108 .addReg(Data).addReg(AddrIn).addReg(0)
7114 ARMTargetLowering::EmitStructByval(MachineInstr *MI,
7115 MachineBasicBlock *BB) const {
7116 // This pseudo instruction has 3 operands: dst, src, size
7117 // We expand it to a loop if size > Subtarget->getMaxInlineSizeThreshold().
7118 // Otherwise, we will generate unrolled scalar copies.
7119 const TargetInstrInfo *TII =
7120 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7121 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7122 MachineFunction::iterator It = BB;
7125 unsigned dest = MI->getOperand(0).getReg();
7126 unsigned src = MI->getOperand(1).getReg();
7127 unsigned SizeVal = MI->getOperand(2).getImm();
7128 unsigned Align = MI->getOperand(3).getImm();
7129 DebugLoc dl = MI->getDebugLoc();
7131 MachineFunction *MF = BB->getParent();
7132 MachineRegisterInfo &MRI = MF->getRegInfo();
7133 unsigned UnitSize = 0;
7134 const TargetRegisterClass *TRC = nullptr;
7135 const TargetRegisterClass *VecTRC = nullptr;
7137 bool IsThumb1 = Subtarget->isThumb1Only();
7138 bool IsThumb2 = Subtarget->isThumb2();
7142 } else if (Align & 2) {
7145 // Check whether we can use NEON instructions.
7146 if (!MF->getFunction()->getAttributes().
7147 hasAttribute(AttributeSet::FunctionIndex,
7148 Attribute::NoImplicitFloat) &&
7149 Subtarget->hasNEON()) {
7150 if ((Align % 16 == 0) && SizeVal >= 16)
7152 else if ((Align % 8 == 0) && SizeVal >= 8)
7155 // Can't use NEON instructions.
7160 // Select the correct opcode and register class for unit size load/store
7161 bool IsNeon = UnitSize >= 8;
7162 TRC = (IsThumb1 || IsThumb2) ? (const TargetRegisterClass *)&ARM::tGPRRegClass
7163 : (const TargetRegisterClass *)&ARM::GPRRegClass;
7165 VecTRC = UnitSize == 16
7166 ? (const TargetRegisterClass *)&ARM::DPairRegClass
7168 ? (const TargetRegisterClass *)&ARM::DPRRegClass
7171 unsigned BytesLeft = SizeVal % UnitSize;
7172 unsigned LoopSize = SizeVal - BytesLeft;
7174 if (SizeVal <= Subtarget->getMaxInlineSizeThreshold()) {
7175 // Use LDR and STR to copy.
7176 // [scratch, srcOut] = LDR_POST(srcIn, UnitSize)
7177 // [destOut] = STR_POST(scratch, destIn, UnitSize)
7178 unsigned srcIn = src;
7179 unsigned destIn = dest;
7180 for (unsigned i = 0; i < LoopSize; i+=UnitSize) {
7181 unsigned srcOut = MRI.createVirtualRegister(TRC);
7182 unsigned destOut = MRI.createVirtualRegister(TRC);
7183 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7184 emitPostLd(BB, MI, TII, dl, UnitSize, scratch, srcIn, srcOut,
7185 IsThumb1, IsThumb2);
7186 emitPostSt(BB, MI, TII, dl, UnitSize, scratch, destIn, destOut,
7187 IsThumb1, IsThumb2);
7192 // Handle the leftover bytes with LDRB and STRB.
7193 // [scratch, srcOut] = LDRB_POST(srcIn, 1)
7194 // [destOut] = STRB_POST(scratch, destIn, 1)
7195 for (unsigned i = 0; i < BytesLeft; i++) {
7196 unsigned srcOut = MRI.createVirtualRegister(TRC);
7197 unsigned destOut = MRI.createVirtualRegister(TRC);
7198 unsigned scratch = MRI.createVirtualRegister(TRC);
7199 emitPostLd(BB, MI, TII, dl, 1, scratch, srcIn, srcOut,
7200 IsThumb1, IsThumb2);
7201 emitPostSt(BB, MI, TII, dl, 1, scratch, destIn, destOut,
7202 IsThumb1, IsThumb2);
7206 MI->eraseFromParent(); // The instruction is gone now.
7210 // Expand the pseudo op to a loop.
7213 // movw varEnd, # --> with thumb2
7215 // ldrcp varEnd, idx --> without thumb2
7216 // fallthrough --> loopMBB
7218 // PHI varPhi, varEnd, varLoop
7219 // PHI srcPhi, src, srcLoop
7220 // PHI destPhi, dst, destLoop
7221 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7222 // [destLoop] = STR_POST(scratch, destPhi, UnitSize)
7223 // subs varLoop, varPhi, #UnitSize
7225 // fallthrough --> exitMBB
7227 // epilogue to handle left-over bytes
7228 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7229 // [destOut] = STRB_POST(scratch, destLoop, 1)
7230 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7231 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
7232 MF->insert(It, loopMBB);
7233 MF->insert(It, exitMBB);
7235 // Transfer the remainder of BB and its successor edges to exitMBB.
7236 exitMBB->splice(exitMBB->begin(), BB,
7237 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7238 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
7240 // Load an immediate to varEnd.
7241 unsigned varEnd = MRI.createVirtualRegister(TRC);
7243 unsigned Vtmp = varEnd;
7244 if ((LoopSize & 0xFFFF0000) != 0)
7245 Vtmp = MRI.createVirtualRegister(TRC);
7246 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVi16), Vtmp)
7247 .addImm(LoopSize & 0xFFFF));
7249 if ((LoopSize & 0xFFFF0000) != 0)
7250 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2MOVTi16), varEnd)
7251 .addReg(Vtmp).addImm(LoopSize >> 16));
7253 MachineConstantPool *ConstantPool = MF->getConstantPool();
7254 Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
7255 const Constant *C = ConstantInt::get(Int32Ty, LoopSize);
7257 // MachineConstantPool wants an explicit alignment.
7258 unsigned Align = getDataLayout()->getPrefTypeAlignment(Int32Ty);
7260 Align = getDataLayout()->getTypeAllocSize(C->getType());
7261 unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
7264 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::tLDRpci)).addReg(
7265 varEnd, RegState::Define).addConstantPoolIndex(Idx));
7267 AddDefaultPred(BuildMI(*BB, MI, dl, TII->get(ARM::LDRcp)).addReg(
7268 varEnd, RegState::Define).addConstantPoolIndex(Idx).addImm(0));
7270 BB->addSuccessor(loopMBB);
7272 // Generate the loop body:
7273 // varPhi = PHI(varLoop, varEnd)
7274 // srcPhi = PHI(srcLoop, src)
7275 // destPhi = PHI(destLoop, dst)
7276 MachineBasicBlock *entryBB = BB;
7278 unsigned varLoop = MRI.createVirtualRegister(TRC);
7279 unsigned varPhi = MRI.createVirtualRegister(TRC);
7280 unsigned srcLoop = MRI.createVirtualRegister(TRC);
7281 unsigned srcPhi = MRI.createVirtualRegister(TRC);
7282 unsigned destLoop = MRI.createVirtualRegister(TRC);
7283 unsigned destPhi = MRI.createVirtualRegister(TRC);
7285 BuildMI(*BB, BB->begin(), dl, TII->get(ARM::PHI), varPhi)
7286 .addReg(varLoop).addMBB(loopMBB)
7287 .addReg(varEnd).addMBB(entryBB);
7288 BuildMI(BB, dl, TII->get(ARM::PHI), srcPhi)
7289 .addReg(srcLoop).addMBB(loopMBB)
7290 .addReg(src).addMBB(entryBB);
7291 BuildMI(BB, dl, TII->get(ARM::PHI), destPhi)
7292 .addReg(destLoop).addMBB(loopMBB)
7293 .addReg(dest).addMBB(entryBB);
7295 // [scratch, srcLoop] = LDR_POST(srcPhi, UnitSize)
7296 // [destLoop] = STR_POST(scratch, destPhi, UnitSiz)
7297 unsigned scratch = MRI.createVirtualRegister(IsNeon ? VecTRC : TRC);
7298 emitPostLd(BB, BB->end(), TII, dl, UnitSize, scratch, srcPhi, srcLoop,
7299 IsThumb1, IsThumb2);
7300 emitPostSt(BB, BB->end(), TII, dl, UnitSize, scratch, destPhi, destLoop,
7301 IsThumb1, IsThumb2);
7303 // Decrement loop variable by UnitSize.
7305 MachineInstrBuilder MIB =
7306 BuildMI(*BB, BB->end(), dl, TII->get(ARM::tSUBi8), varLoop);
7307 MIB = AddDefaultT1CC(MIB);
7308 MIB.addReg(varPhi).addImm(UnitSize);
7309 AddDefaultPred(MIB);
7311 MachineInstrBuilder MIB =
7312 BuildMI(*BB, BB->end(), dl,
7313 TII->get(IsThumb2 ? ARM::t2SUBri : ARM::SUBri), varLoop);
7314 AddDefaultCC(AddDefaultPred(MIB.addReg(varPhi).addImm(UnitSize)));
7315 MIB->getOperand(5).setReg(ARM::CPSR);
7316 MIB->getOperand(5).setIsDef(true);
7318 BuildMI(*BB, BB->end(), dl,
7319 TII->get(IsThumb1 ? ARM::tBcc : IsThumb2 ? ARM::t2Bcc : ARM::Bcc))
7320 .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
7322 // loopMBB can loop back to loopMBB or fall through to exitMBB.
7323 BB->addSuccessor(loopMBB);
7324 BB->addSuccessor(exitMBB);
7326 // Add epilogue to handle BytesLeft.
7328 MachineInstr *StartOfExit = exitMBB->begin();
7330 // [scratch, srcOut] = LDRB_POST(srcLoop, 1)
7331 // [destOut] = STRB_POST(scratch, destLoop, 1)
7332 unsigned srcIn = srcLoop;
7333 unsigned destIn = destLoop;
7334 for (unsigned i = 0; i < BytesLeft; i++) {
7335 unsigned srcOut = MRI.createVirtualRegister(TRC);
7336 unsigned destOut = MRI.createVirtualRegister(TRC);
7337 unsigned scratch = MRI.createVirtualRegister(TRC);
7338 emitPostLd(BB, StartOfExit, TII, dl, 1, scratch, srcIn, srcOut,
7339 IsThumb1, IsThumb2);
7340 emitPostSt(BB, StartOfExit, TII, dl, 1, scratch, destIn, destOut,
7341 IsThumb1, IsThumb2);
7346 MI->eraseFromParent(); // The instruction is gone now.
7351 ARMTargetLowering::EmitLowered__chkstk(MachineInstr *MI,
7352 MachineBasicBlock *MBB) const {
7353 const TargetMachine &TM = getTargetMachine();
7354 const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
7355 DebugLoc DL = MI->getDebugLoc();
7357 assert(Subtarget->isTargetWindows() &&
7358 "__chkstk is only supported on Windows");
7359 assert(Subtarget->isThumb2() && "Windows on ARM requires Thumb-2 mode");
7361 // __chkstk takes the number of words to allocate on the stack in R4, and
7362 // returns the stack adjustment in number of bytes in R4. This will not
7363 // clober any other registers (other than the obvious lr).
7365 // Although, technically, IP should be considered a register which may be
7366 // clobbered, the call itself will not touch it. Windows on ARM is a pure
7367 // thumb-2 environment, so there is no interworking required. As a result, we
7368 // do not expect a veneer to be emitted by the linker, clobbering IP.
7370 // Each module receives its own copy of __chkstk, so no import thunk is
7371 // required, again, ensuring that IP is not clobbered.
7373 // Finally, although some linkers may theoretically provide a trampoline for
7374 // out of range calls (which is quite common due to a 32M range limitation of
7375 // branches for Thumb), we can generate the long-call version via
7376 // -mcmodel=large, alleviating the need for the trampoline which may clobber
7379 switch (TM.getCodeModel()) {
7380 case CodeModel::Small:
7381 case CodeModel::Medium:
7382 case CodeModel::Default:
7383 case CodeModel::Kernel:
7384 BuildMI(*MBB, MI, DL, TII.get(ARM::tBL))
7385 .addImm((unsigned)ARMCC::AL).addReg(0)
7386 .addExternalSymbol("__chkstk")
7387 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7388 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7389 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7391 case CodeModel::Large:
7392 case CodeModel::JITDefault: {
7393 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
7394 unsigned Reg = MRI.createVirtualRegister(&ARM::rGPRRegClass);
7396 BuildMI(*MBB, MI, DL, TII.get(ARM::t2MOVi32imm), Reg)
7397 .addExternalSymbol("__chkstk");
7398 BuildMI(*MBB, MI, DL, TII.get(ARM::tBLXr))
7399 .addImm((unsigned)ARMCC::AL).addReg(0)
7400 .addReg(Reg, RegState::Kill)
7401 .addReg(ARM::R4, RegState::Implicit | RegState::Kill)
7402 .addReg(ARM::R4, RegState::Implicit | RegState::Define)
7403 .addReg(ARM::R12, RegState::Implicit | RegState::Define | RegState::Dead);
7408 AddDefaultCC(AddDefaultPred(BuildMI(*MBB, MI, DL, TII.get(ARM::t2SUBrr),
7410 .addReg(ARM::SP).addReg(ARM::R4)));
7412 MI->eraseFromParent();
7417 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7418 MachineBasicBlock *BB) const {
7419 const TargetInstrInfo *TII =
7420 getTargetMachine().getSubtargetImpl()->getInstrInfo();
7421 DebugLoc dl = MI->getDebugLoc();
7422 bool isThumb2 = Subtarget->isThumb2();
7423 switch (MI->getOpcode()) {
7426 llvm_unreachable("Unexpected instr type to insert");
7428 // The Thumb2 pre-indexed stores have the same MI operands, they just
7429 // define them differently in the .td files from the isel patterns, so
7430 // they need pseudos.
7431 case ARM::t2STR_preidx:
7432 MI->setDesc(TII->get(ARM::t2STR_PRE));
7434 case ARM::t2STRB_preidx:
7435 MI->setDesc(TII->get(ARM::t2STRB_PRE));
7437 case ARM::t2STRH_preidx:
7438 MI->setDesc(TII->get(ARM::t2STRH_PRE));
7441 case ARM::STRi_preidx:
7442 case ARM::STRBi_preidx: {
7443 unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
7444 ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
7445 // Decode the offset.
7446 unsigned Offset = MI->getOperand(4).getImm();
7447 bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
7448 Offset = ARM_AM::getAM2Offset(Offset);
7452 MachineMemOperand *MMO = *MI->memoperands_begin();
7453 BuildMI(*BB, MI, dl, TII->get(NewOpc))
7454 .addOperand(MI->getOperand(0)) // Rn_wb
7455 .addOperand(MI->getOperand(1)) // Rt
7456 .addOperand(MI->getOperand(2)) // Rn
7457 .addImm(Offset) // offset (skip GPR==zero_reg)
7458 .addOperand(MI->getOperand(5)) // pred
7459 .addOperand(MI->getOperand(6))
7460 .addMemOperand(MMO);
7461 MI->eraseFromParent();
7464 case ARM::STRr_preidx:
7465 case ARM::STRBr_preidx:
7466 case ARM::STRH_preidx: {
7468 switch (MI->getOpcode()) {
7469 default: llvm_unreachable("unexpected opcode!");
7470 case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
7471 case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
7472 case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
7474 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
7475 for (unsigned i = 0; i < MI->getNumOperands(); ++i)
7476 MIB.addOperand(MI->getOperand(i));
7477 MI->eraseFromParent();
7481 case ARM::tMOVCCr_pseudo: {
7482 // To "insert" a SELECT_CC instruction, we actually have to insert the
7483 // diamond control-flow pattern. The incoming instruction knows the
7484 // destination vreg to set, the condition code register to branch on, the
7485 // true/false values to select between, and a branch opcode to use.
7486 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7487 MachineFunction::iterator It = BB;
7493 // cmpTY ccX, r1, r2
7495 // fallthrough --> copy0MBB
7496 MachineBasicBlock *thisMBB = BB;
7497 MachineFunction *F = BB->getParent();
7498 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7499 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7500 F->insert(It, copy0MBB);
7501 F->insert(It, sinkMBB);
7503 // Transfer the remainder of BB and its successor edges to sinkMBB.
7504 sinkMBB->splice(sinkMBB->begin(), BB,
7505 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7506 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
7508 BB->addSuccessor(copy0MBB);
7509 BB->addSuccessor(sinkMBB);
7511 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
7512 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
7515 // %FalseValue = ...
7516 // # fallthrough to sinkMBB
7519 // Update machine-CFG edges
7520 BB->addSuccessor(sinkMBB);
7523 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7526 BuildMI(*BB, BB->begin(), dl,
7527 TII->get(ARM::PHI), MI->getOperand(0).getReg())
7528 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7529 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7531 MI->eraseFromParent(); // The pseudo instruction is gone now.
7536 case ARM::BCCZi64: {
7537 // If there is an unconditional branch to the other successor, remove it.
7538 BB->erase(std::next(MachineBasicBlock::iterator(MI)), BB->end());
7540 // Compare both parts that make up the double comparison separately for
7542 bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
7544 unsigned LHS1 = MI->getOperand(1).getReg();
7545 unsigned LHS2 = MI->getOperand(2).getReg();
7547 AddDefaultPred(BuildMI(BB, dl,
7548 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7549 .addReg(LHS1).addImm(0));
7550 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7551 .addReg(LHS2).addImm(0)
7552 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7554 unsigned RHS1 = MI->getOperand(3).getReg();
7555 unsigned RHS2 = MI->getOperand(4).getReg();
7556 AddDefaultPred(BuildMI(BB, dl,
7557 TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7558 .addReg(LHS1).addReg(RHS1));
7559 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
7560 .addReg(LHS2).addReg(RHS2)
7561 .addImm(ARMCC::EQ).addReg(ARM::CPSR);
7564 MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
7565 MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
7566 if (MI->getOperand(0).getImm() == ARMCC::NE)
7567 std::swap(destMBB, exitMBB);
7569 BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
7570 .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
7572 AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
7574 BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
7576 MI->eraseFromParent(); // The pseudo instruction is gone now.
7580 case ARM::Int_eh_sjlj_setjmp:
7581 case ARM::Int_eh_sjlj_setjmp_nofp:
7582 case ARM::tInt_eh_sjlj_setjmp:
7583 case ARM::t2Int_eh_sjlj_setjmp:
7584 case ARM::t2Int_eh_sjlj_setjmp_nofp:
7585 EmitSjLjDispatchBlock(MI, BB);
7590 // To insert an ABS instruction, we have to insert the
7591 // diamond control-flow pattern. The incoming instruction knows the
7592 // source vreg to test against 0, the destination vreg to set,
7593 // the condition code register to branch on, the
7594 // true/false values to select between, and a branch opcode to use.
7599 // BCC (branch to SinkBB if V0 >= 0)
7600 // RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
7601 // SinkBB: V1 = PHI(V2, V3)
7602 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7603 MachineFunction::iterator BBI = BB;
7605 MachineFunction *Fn = BB->getParent();
7606 MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7607 MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
7608 Fn->insert(BBI, RSBBB);
7609 Fn->insert(BBI, SinkBB);
7611 unsigned int ABSSrcReg = MI->getOperand(1).getReg();
7612 unsigned int ABSDstReg = MI->getOperand(0).getReg();
7613 bool isThumb2 = Subtarget->isThumb2();
7614 MachineRegisterInfo &MRI = Fn->getRegInfo();
7615 // In Thumb mode S must not be specified if source register is the SP or
7616 // PC and if destination register is the SP, so restrict register class
7617 unsigned NewRsbDstReg = MRI.createVirtualRegister(isThumb2 ?
7618 (const TargetRegisterClass*)&ARM::rGPRRegClass :
7619 (const TargetRegisterClass*)&ARM::GPRRegClass);
7621 // Transfer the remainder of BB and its successor edges to sinkMBB.
7622 SinkBB->splice(SinkBB->begin(), BB,
7623 std::next(MachineBasicBlock::iterator(MI)), BB->end());
7624 SinkBB->transferSuccessorsAndUpdatePHIs(BB);
7626 BB->addSuccessor(RSBBB);
7627 BB->addSuccessor(SinkBB);
7629 // fall through to SinkMBB
7630 RSBBB->addSuccessor(SinkBB);
7632 // insert a cmp at the end of BB
7633 AddDefaultPred(BuildMI(BB, dl,
7634 TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
7635 .addReg(ABSSrcReg).addImm(0));
7637 // insert a bcc with opposite CC to ARMCC::MI at the end of BB
7639 TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
7640 .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
7642 // insert rsbri in RSBBB
7643 // Note: BCC and rsbri will be converted into predicated rsbmi
7644 // by if-conversion pass
7645 BuildMI(*RSBBB, RSBBB->begin(), dl,
7646 TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
7647 .addReg(ABSSrcReg, RegState::Kill)
7648 .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
7650 // insert PHI in SinkBB,
7651 // reuse ABSDstReg to not change uses of ABS instruction
7652 BuildMI(*SinkBB, SinkBB->begin(), dl,
7653 TII->get(ARM::PHI), ABSDstReg)
7654 .addReg(NewRsbDstReg).addMBB(RSBBB)
7655 .addReg(ABSSrcReg).addMBB(BB);
7657 // remove ABS instruction
7658 MI->eraseFromParent();
7660 // return last added BB
7663 case ARM::COPY_STRUCT_BYVAL_I32:
7665 return EmitStructByval(MI, BB);
7666 case ARM::WIN__CHKSTK:
7667 return EmitLowered__chkstk(MI, BB);
7671 void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
7672 SDNode *Node) const {
7673 if (!MI->hasPostISelHook()) {
7674 assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
7675 "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
7679 const MCInstrDesc *MCID = &MI->getDesc();
7680 // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
7681 // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
7682 // operand is still set to noreg. If needed, set the optional operand's
7683 // register to CPSR, and remove the redundant implicit def.
7685 // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
7687 // Rename pseudo opcodes.
7688 unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
7690 const ARMBaseInstrInfo *TII = static_cast<const ARMBaseInstrInfo *>(
7691 getTargetMachine().getSubtargetImpl()->getInstrInfo());
7692 MCID = &TII->get(NewOpc);
7694 assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
7695 "converted opcode should be the same except for cc_out");
7699 // Add the optional cc_out operand
7700 MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
7702 unsigned ccOutIdx = MCID->getNumOperands() - 1;
7704 // Any ARM instruction that sets the 's' bit should specify an optional
7705 // "cc_out" operand in the last operand position.
7706 if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
7707 assert(!NewOpc && "Optional cc_out operand required");
7710 // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
7711 // since we already have an optional CPSR def.
7712 bool definesCPSR = false;
7713 bool deadCPSR = false;
7714 for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
7716 const MachineOperand &MO = MI->getOperand(i);
7717 if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
7721 MI->RemoveOperand(i);
7726 assert(!NewOpc && "Optional cc_out operand required");
7729 assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
7731 assert(!MI->getOperand(ccOutIdx).getReg() &&
7732 "expect uninitialized optional cc_out operand");
7736 // If this instruction was defined with an optional CPSR def and its dag node
7737 // had a live implicit CPSR def, then activate the optional CPSR def.
7738 MachineOperand &MO = MI->getOperand(ccOutIdx);
7739 MO.setReg(ARM::CPSR);
7743 //===----------------------------------------------------------------------===//
7744 // ARM Optimization Hooks
7745 //===----------------------------------------------------------------------===//
7747 // Helper function that checks if N is a null or all ones constant.
7748 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
7749 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
7752 return AllOnes ? C->isAllOnesValue() : C->isNullValue();
7755 // Return true if N is conditionally 0 or all ones.
7756 // Detects these expressions where cc is an i1 value:
7758 // (select cc 0, y) [AllOnes=0]
7759 // (select cc y, 0) [AllOnes=0]
7760 // (zext cc) [AllOnes=0]
7761 // (sext cc) [AllOnes=0/1]
7762 // (select cc -1, y) [AllOnes=1]
7763 // (select cc y, -1) [AllOnes=1]
7765 // Invert is set when N is the null/all ones constant when CC is false.
7766 // OtherOp is set to the alternative value of N.
7767 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes,
7768 SDValue &CC, bool &Invert,
7770 SelectionDAG &DAG) {
7771 switch (N->getOpcode()) {
7772 default: return false;
7774 CC = N->getOperand(0);
7775 SDValue N1 = N->getOperand(1);
7776 SDValue N2 = N->getOperand(2);
7777 if (isZeroOrAllOnes(N1, AllOnes)) {
7782 if (isZeroOrAllOnes(N2, AllOnes)) {
7789 case ISD::ZERO_EXTEND:
7790 // (zext cc) can never be the all ones value.
7794 case ISD::SIGN_EXTEND: {
7795 EVT VT = N->getValueType(0);
7796 CC = N->getOperand(0);
7797 if (CC.getValueType() != MVT::i1)
7801 // When looking for an AllOnes constant, N is an sext, and the 'other'
7803 OtherOp = DAG.getConstant(0, VT);
7804 else if (N->getOpcode() == ISD::ZERO_EXTEND)
7805 // When looking for a 0 constant, N can be zext or sext.
7806 OtherOp = DAG.getConstant(1, VT);
7808 OtherOp = DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT);
7814 // Combine a constant select operand into its use:
7816 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
7817 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
7818 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
7819 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
7820 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
7822 // The transform is rejected if the select doesn't have a constant operand that
7823 // is null, or all ones when AllOnes is set.
7825 // Also recognize sext/zext from i1:
7827 // (add (zext cc), x) -> (select cc (add x, 1), x)
7828 // (add (sext cc), x) -> (select cc (add x, -1), x)
7830 // These transformations eventually create predicated instructions.
7832 // @param N The node to transform.
7833 // @param Slct The N operand that is a select.
7834 // @param OtherOp The other N operand (x above).
7835 // @param DCI Context.
7836 // @param AllOnes Require the select constant to be all ones instead of null.
7837 // @returns The new node, or SDValue() on failure.
7839 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
7840 TargetLowering::DAGCombinerInfo &DCI,
7841 bool AllOnes = false) {
7842 SelectionDAG &DAG = DCI.DAG;
7843 EVT VT = N->getValueType(0);
7844 SDValue NonConstantVal;
7847 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
7848 NonConstantVal, DAG))
7851 // Slct is now know to be the desired identity constant when CC is true.
7852 SDValue TrueVal = OtherOp;
7853 SDValue FalseVal = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
7854 OtherOp, NonConstantVal);
7855 // Unless SwapSelectOps says CC should be false.
7857 std::swap(TrueVal, FalseVal);
7859 return DAG.getNode(ISD::SELECT, SDLoc(N), VT,
7860 CCOp, TrueVal, FalseVal);
7863 // Attempt combineSelectAndUse on each operand of a commutative operator N.
7865 SDValue combineSelectAndUseCommutative(SDNode *N, bool AllOnes,
7866 TargetLowering::DAGCombinerInfo &DCI) {
7867 SDValue N0 = N->getOperand(0);
7868 SDValue N1 = N->getOperand(1);
7869 if (N0.getNode()->hasOneUse()) {
7870 SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes);
7871 if (Result.getNode())
7874 if (N1.getNode()->hasOneUse()) {
7875 SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes);
7876 if (Result.getNode())
7882 // AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
7883 // (only after legalization).
7884 static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
7885 TargetLowering::DAGCombinerInfo &DCI,
7886 const ARMSubtarget *Subtarget) {
7888 // Only perform optimization if after legalize, and if NEON is available. We
7889 // also expected both operands to be BUILD_VECTORs.
7890 if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
7891 || N0.getOpcode() != ISD::BUILD_VECTOR
7892 || N1.getOpcode() != ISD::BUILD_VECTOR)
7895 // Check output type since VPADDL operand elements can only be 8, 16, or 32.
7896 EVT VT = N->getValueType(0);
7897 if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
7900 // Check that the vector operands are of the right form.
7901 // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
7902 // operands, where N is the size of the formed vector.
7903 // Each EXTRACT_VECTOR should have the same input vector and odd or even
7904 // index such that we have a pair wise add pattern.
7906 // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
7907 if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7909 SDValue Vec = N0->getOperand(0)->getOperand(0);
7910 SDNode *V = Vec.getNode();
7911 unsigned nextIndex = 0;
7913 // For each operands to the ADD which are BUILD_VECTORs,
7914 // check to see if each of their operands are an EXTRACT_VECTOR with
7915 // the same vector and appropriate index.
7916 for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
7917 if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
7918 && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7920 SDValue ExtVec0 = N0->getOperand(i);
7921 SDValue ExtVec1 = N1->getOperand(i);
7923 // First operand is the vector, verify its the same.
7924 if (V != ExtVec0->getOperand(0).getNode() ||
7925 V != ExtVec1->getOperand(0).getNode())
7928 // Second is the constant, verify its correct.
7929 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
7930 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
7932 // For the constant, we want to see all the even or all the odd.
7933 if (!C0 || !C1 || C0->getZExtValue() != nextIndex
7934 || C1->getZExtValue() != nextIndex+1)
7943 // Create VPADDL node.
7944 SelectionDAG &DAG = DCI.DAG;
7945 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7947 // Build operand list.
7948 SmallVector<SDValue, 8> Ops;
7949 Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
7950 TLI.getPointerTy()));
7952 // Input is the vector.
7955 // Get widened type and narrowed type.
7957 unsigned numElem = VT.getVectorNumElements();
7959 EVT inputLaneType = Vec.getValueType().getVectorElementType();
7960 switch (inputLaneType.getSimpleVT().SimpleTy) {
7961 case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
7962 case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
7963 case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
7965 llvm_unreachable("Invalid vector element type for padd optimization.");
7968 SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), widenType, Ops);
7969 unsigned ExtOp = VT.bitsGT(tmp.getValueType()) ? ISD::ANY_EXTEND : ISD::TRUNCATE;
7970 return DAG.getNode(ExtOp, SDLoc(N), VT, tmp);
7973 static SDValue findMUL_LOHI(SDValue V) {
7974 if (V->getOpcode() == ISD::UMUL_LOHI ||
7975 V->getOpcode() == ISD::SMUL_LOHI)
7980 static SDValue AddCombineTo64bitMLAL(SDNode *AddcNode,
7981 TargetLowering::DAGCombinerInfo &DCI,
7982 const ARMSubtarget *Subtarget) {
7984 if (Subtarget->isThumb1Only()) return SDValue();
7986 // Only perform the checks after legalize when the pattern is available.
7987 if (DCI.isBeforeLegalize()) return SDValue();
7989 // Look for multiply add opportunities.
7990 // The pattern is a ISD::UMUL_LOHI followed by two add nodes, where
7991 // each add nodes consumes a value from ISD::UMUL_LOHI and there is
7992 // a glue link from the first add to the second add.
7993 // If we find this pattern, we can replace the U/SMUL_LOHI, ADDC, and ADDE by
7994 // a S/UMLAL instruction.
7997 // \ / \ [no multiline comment]
8003 assert(AddcNode->getOpcode() == ISD::ADDC && "Expect an ADDC");
8004 SDValue AddcOp0 = AddcNode->getOperand(0);
8005 SDValue AddcOp1 = AddcNode->getOperand(1);
8007 // Check if the two operands are from the same mul_lohi node.
8008 if (AddcOp0.getNode() == AddcOp1.getNode())
8011 assert(AddcNode->getNumValues() == 2 &&
8012 AddcNode->getValueType(0) == MVT::i32 &&
8013 "Expect ADDC with two result values. First: i32");
8015 // Check that we have a glued ADDC node.
8016 if (AddcNode->getValueType(1) != MVT::Glue)
8019 // Check that the ADDC adds the low result of the S/UMUL_LOHI.
8020 if (AddcOp0->getOpcode() != ISD::UMUL_LOHI &&
8021 AddcOp0->getOpcode() != ISD::SMUL_LOHI &&
8022 AddcOp1->getOpcode() != ISD::UMUL_LOHI &&
8023 AddcOp1->getOpcode() != ISD::SMUL_LOHI)
8026 // Look for the glued ADDE.
8027 SDNode* AddeNode = AddcNode->getGluedUser();
8031 // Make sure it is really an ADDE.
8032 if (AddeNode->getOpcode() != ISD::ADDE)
8035 assert(AddeNode->getNumOperands() == 3 &&
8036 AddeNode->getOperand(2).getValueType() == MVT::Glue &&
8037 "ADDE node has the wrong inputs");
8039 // Check for the triangle shape.
8040 SDValue AddeOp0 = AddeNode->getOperand(0);
8041 SDValue AddeOp1 = AddeNode->getOperand(1);
8043 // Make sure that the ADDE operands are not coming from the same node.
8044 if (AddeOp0.getNode() == AddeOp1.getNode())
8047 // Find the MUL_LOHI node walking up ADDE's operands.
8048 bool IsLeftOperandMUL = false;
8049 SDValue MULOp = findMUL_LOHI(AddeOp0);
8050 if (MULOp == SDValue())
8051 MULOp = findMUL_LOHI(AddeOp1);
8053 IsLeftOperandMUL = true;
8054 if (MULOp == SDValue())
8057 // Figure out the right opcode.
8058 unsigned Opc = MULOp->getOpcode();
8059 unsigned FinalOpc = (Opc == ISD::SMUL_LOHI) ? ARMISD::SMLAL : ARMISD::UMLAL;
8061 // Figure out the high and low input values to the MLAL node.
8062 SDValue* HiMul = &MULOp;
8063 SDValue* HiAdd = nullptr;
8064 SDValue* LoMul = nullptr;
8065 SDValue* LowAdd = nullptr;
8067 if (IsLeftOperandMUL)
8073 if (AddcOp0->getOpcode() == Opc) {
8077 if (AddcOp1->getOpcode() == Opc) {
8085 if (LoMul->getNode() != HiMul->getNode())
8088 // Create the merged node.
8089 SelectionDAG &DAG = DCI.DAG;
8091 // Build operand list.
8092 SmallVector<SDValue, 8> Ops;
8093 Ops.push_back(LoMul->getOperand(0));
8094 Ops.push_back(LoMul->getOperand(1));
8095 Ops.push_back(*LowAdd);
8096 Ops.push_back(*HiAdd);
8098 SDValue MLALNode = DAG.getNode(FinalOpc, SDLoc(AddcNode),
8099 DAG.getVTList(MVT::i32, MVT::i32), Ops);
8101 // Replace the ADDs' nodes uses by the MLA node's values.
8102 SDValue HiMLALResult(MLALNode.getNode(), 1);
8103 DAG.ReplaceAllUsesOfValueWith(SDValue(AddeNode, 0), HiMLALResult);
8105 SDValue LoMLALResult(MLALNode.getNode(), 0);
8106 DAG.ReplaceAllUsesOfValueWith(SDValue(AddcNode, 0), LoMLALResult);
8108 // Return original node to notify the driver to stop replacing.
8109 SDValue resNode(AddcNode, 0);
8113 /// PerformADDCCombine - Target-specific dag combine transform from
8114 /// ISD::ADDC, ISD::ADDE, and ISD::MUL_LOHI to MLAL.
8115 static SDValue PerformADDCCombine(SDNode *N,
8116 TargetLowering::DAGCombinerInfo &DCI,
8117 const ARMSubtarget *Subtarget) {
8119 return AddCombineTo64bitMLAL(N, DCI, Subtarget);
8123 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
8124 /// operands N0 and N1. This is a helper for PerformADDCombine that is
8125 /// called with the default operands, and if that fails, with commuted
8127 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
8128 TargetLowering::DAGCombinerInfo &DCI,
8129 const ARMSubtarget *Subtarget){
8131 // Attempt to create vpaddl for this add.
8132 SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
8133 if (Result.getNode())
8136 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
8137 if (N0.getNode()->hasOneUse()) {
8138 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
8139 if (Result.getNode()) return Result;
8144 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
8146 static SDValue PerformADDCombine(SDNode *N,
8147 TargetLowering::DAGCombinerInfo &DCI,
8148 const ARMSubtarget *Subtarget) {
8149 SDValue N0 = N->getOperand(0);
8150 SDValue N1 = N->getOperand(1);
8152 // First try with the default operand order.
8153 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
8154 if (Result.getNode())
8157 // If that didn't work, try again with the operands commuted.
8158 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
8161 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
8163 static SDValue PerformSUBCombine(SDNode *N,
8164 TargetLowering::DAGCombinerInfo &DCI) {
8165 SDValue N0 = N->getOperand(0);
8166 SDValue N1 = N->getOperand(1);
8168 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
8169 if (N1.getNode()->hasOneUse()) {
8170 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
8171 if (Result.getNode()) return Result;
8177 /// PerformVMULCombine
8178 /// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
8179 /// special multiplier accumulator forwarding.
8185 // However, for (A + B) * (A + B),
8192 static SDValue PerformVMULCombine(SDNode *N,
8193 TargetLowering::DAGCombinerInfo &DCI,
8194 const ARMSubtarget *Subtarget) {
8195 if (!Subtarget->hasVMLxForwarding())
8198 SelectionDAG &DAG = DCI.DAG;
8199 SDValue N0 = N->getOperand(0);
8200 SDValue N1 = N->getOperand(1);
8201 unsigned Opcode = N0.getOpcode();
8202 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8203 Opcode != ISD::FADD && Opcode != ISD::FSUB) {
8204 Opcode = N1.getOpcode();
8205 if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
8206 Opcode != ISD::FADD && Opcode != ISD::FSUB)
8214 EVT VT = N->getValueType(0);
8216 SDValue N00 = N0->getOperand(0);
8217 SDValue N01 = N0->getOperand(1);
8218 return DAG.getNode(Opcode, DL, VT,
8219 DAG.getNode(ISD::MUL, DL, VT, N00, N1),
8220 DAG.getNode(ISD::MUL, DL, VT, N01, N1));
8223 static SDValue PerformMULCombine(SDNode *N,
8224 TargetLowering::DAGCombinerInfo &DCI,
8225 const ARMSubtarget *Subtarget) {
8226 SelectionDAG &DAG = DCI.DAG;
8228 if (Subtarget->isThumb1Only())
8231 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8234 EVT VT = N->getValueType(0);
8235 if (VT.is64BitVector() || VT.is128BitVector())
8236 return PerformVMULCombine(N, DCI, Subtarget);
8240 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8244 int64_t MulAmt = C->getSExtValue();
8245 unsigned ShiftAmt = countTrailingZeros<uint64_t>(MulAmt);
8247 ShiftAmt = ShiftAmt & (32 - 1);
8248 SDValue V = N->getOperand(0);
8252 MulAmt >>= ShiftAmt;
8255 if (isPowerOf2_32(MulAmt - 1)) {
8256 // (mul x, 2^N + 1) => (add (shl x, N), x)
8257 Res = DAG.getNode(ISD::ADD, DL, VT,
8259 DAG.getNode(ISD::SHL, DL, VT,
8261 DAG.getConstant(Log2_32(MulAmt - 1),
8263 } else if (isPowerOf2_32(MulAmt + 1)) {
8264 // (mul x, 2^N - 1) => (sub (shl x, N), x)
8265 Res = DAG.getNode(ISD::SUB, DL, VT,
8266 DAG.getNode(ISD::SHL, DL, VT,
8268 DAG.getConstant(Log2_32(MulAmt + 1),
8274 uint64_t MulAmtAbs = -MulAmt;
8275 if (isPowerOf2_32(MulAmtAbs + 1)) {
8276 // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
8277 Res = DAG.getNode(ISD::SUB, DL, VT,
8279 DAG.getNode(ISD::SHL, DL, VT,
8281 DAG.getConstant(Log2_32(MulAmtAbs + 1),
8283 } else if (isPowerOf2_32(MulAmtAbs - 1)) {
8284 // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
8285 Res = DAG.getNode(ISD::ADD, DL, VT,
8287 DAG.getNode(ISD::SHL, DL, VT,
8289 DAG.getConstant(Log2_32(MulAmtAbs-1),
8291 Res = DAG.getNode(ISD::SUB, DL, VT,
8292 DAG.getConstant(0, MVT::i32),Res);
8299 Res = DAG.getNode(ISD::SHL, DL, VT,
8300 Res, DAG.getConstant(ShiftAmt, MVT::i32));
8302 // Do not add new nodes to DAG combiner worklist.
8303 DCI.CombineTo(N, Res, false);
8307 static SDValue PerformANDCombine(SDNode *N,
8308 TargetLowering::DAGCombinerInfo &DCI,
8309 const ARMSubtarget *Subtarget) {
8311 // Attempt to use immediate-form VBIC
8312 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8314 EVT VT = N->getValueType(0);
8315 SelectionDAG &DAG = DCI.DAG;
8317 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8320 APInt SplatBits, SplatUndef;
8321 unsigned SplatBitSize;
8324 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8325 if (SplatBitSize <= 64) {
8327 SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
8328 SplatUndef.getZExtValue(), SplatBitSize,
8329 DAG, VbicVT, VT.is128BitVector(),
8331 if (Val.getNode()) {
8333 DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
8334 SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
8335 return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
8340 if (!Subtarget->isThumb1Only()) {
8341 // fold (and (select cc, -1, c), x) -> (select cc, x, (and, x, c))
8342 SDValue Result = combineSelectAndUseCommutative(N, true, DCI);
8343 if (Result.getNode())
8350 /// PerformORCombine - Target-specific dag combine xforms for ISD::OR
8351 static SDValue PerformORCombine(SDNode *N,
8352 TargetLowering::DAGCombinerInfo &DCI,
8353 const ARMSubtarget *Subtarget) {
8354 // Attempt to use immediate-form VORR
8355 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
8357 EVT VT = N->getValueType(0);
8358 SelectionDAG &DAG = DCI.DAG;
8360 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8363 APInt SplatBits, SplatUndef;
8364 unsigned SplatBitSize;
8366 if (BVN && Subtarget->hasNEON() &&
8367 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
8368 if (SplatBitSize <= 64) {
8370 SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
8371 SplatUndef.getZExtValue(), SplatBitSize,
8372 DAG, VorrVT, VT.is128BitVector(),
8374 if (Val.getNode()) {
8376 DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
8377 SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
8378 return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
8383 if (!Subtarget->isThumb1Only()) {
8384 // fold (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
8385 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8386 if (Result.getNode())
8390 // The code below optimizes (or (and X, Y), Z).
8391 // The AND operand needs to have a single user to make these optimizations
8393 SDValue N0 = N->getOperand(0);
8394 if (N0.getOpcode() != ISD::AND || !N0.hasOneUse())
8396 SDValue N1 = N->getOperand(1);
8398 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
8399 if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
8400 DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
8402 unsigned SplatBitSize;
8405 APInt SplatBits0, SplatBits1;
8406 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
8407 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
8408 // Ensure that the second operand of both ands are constants
8409 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
8410 HasAnyUndefs) && !HasAnyUndefs) {
8411 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
8412 HasAnyUndefs) && !HasAnyUndefs) {
8413 // Ensure that the bit width of the constants are the same and that
8414 // the splat arguments are logical inverses as per the pattern we
8415 // are trying to simplify.
8416 if (SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
8417 SplatBits0 == ~SplatBits1) {
8418 // Canonicalize the vector type to make instruction selection
8420 EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
8421 SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
8425 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
8431 // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
8434 // BFI is only available on V6T2+
8435 if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
8439 // 1) or (and A, mask), val => ARMbfi A, val, mask
8440 // iff (val & mask) == val
8442 // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8443 // 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
8444 // && mask == ~mask2
8445 // 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
8446 // && ~mask == mask2
8447 // (i.e., copy a bitfield value into another bitfield of the same width)
8452 SDValue N00 = N0.getOperand(0);
8454 // The value and the mask need to be constants so we can verify this is
8455 // actually a bitfield set. If the mask is 0xffff, we can do better
8456 // via a movt instruction, so don't use BFI in that case.
8457 SDValue MaskOp = N0.getOperand(1);
8458 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
8461 unsigned Mask = MaskC->getZExtValue();
8465 // Case (1): or (and A, mask), val => ARMbfi A, val, mask
8466 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8468 unsigned Val = N1C->getZExtValue();
8469 if ((Val & ~Mask) != Val)
8472 if (ARM::isBitFieldInvertedMask(Mask)) {
8473 Val >>= countTrailingZeros(~Mask);
8475 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
8476 DAG.getConstant(Val, MVT::i32),
8477 DAG.getConstant(Mask, MVT::i32));
8479 // Do not add new nodes to DAG combiner worklist.
8480 DCI.CombineTo(N, Res, false);
8483 } else if (N1.getOpcode() == ISD::AND) {
8484 // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
8485 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8488 unsigned Mask2 = N11C->getZExtValue();
8490 // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
8492 if (ARM::isBitFieldInvertedMask(Mask) &&
8494 // The pack halfword instruction works better for masks that fit it,
8495 // so use that when it's available.
8496 if (Subtarget->hasT2ExtractPack() &&
8497 (Mask == 0xffff || Mask == 0xffff0000))
8500 unsigned amt = countTrailingZeros(Mask2);
8501 Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
8502 DAG.getConstant(amt, MVT::i32));
8503 Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
8504 DAG.getConstant(Mask, MVT::i32));
8505 // Do not add new nodes to DAG combiner worklist.
8506 DCI.CombineTo(N, Res, false);
8508 } else if (ARM::isBitFieldInvertedMask(~Mask) &&
8510 // The pack halfword instruction works better for masks that fit it,
8511 // so use that when it's available.
8512 if (Subtarget->hasT2ExtractPack() &&
8513 (Mask2 == 0xffff || Mask2 == 0xffff0000))
8516 unsigned lsb = countTrailingZeros(Mask);
8517 Res = DAG.getNode(ISD::SRL, DL, VT, N00,
8518 DAG.getConstant(lsb, MVT::i32));
8519 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
8520 DAG.getConstant(Mask2, MVT::i32));
8521 // Do not add new nodes to DAG combiner worklist.
8522 DCI.CombineTo(N, Res, false);
8527 if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
8528 N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
8529 ARM::isBitFieldInvertedMask(~Mask)) {
8530 // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
8531 // where lsb(mask) == #shamt and masked bits of B are known zero.
8532 SDValue ShAmt = N00.getOperand(1);
8533 unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
8534 unsigned LSB = countTrailingZeros(Mask);
8538 Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
8539 DAG.getConstant(~Mask, MVT::i32));
8541 // Do not add new nodes to DAG combiner worklist.
8542 DCI.CombineTo(N, Res, false);
8548 static SDValue PerformXORCombine(SDNode *N,
8549 TargetLowering::DAGCombinerInfo &DCI,
8550 const ARMSubtarget *Subtarget) {
8551 EVT VT = N->getValueType(0);
8552 SelectionDAG &DAG = DCI.DAG;
8554 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8557 if (!Subtarget->isThumb1Only()) {
8558 // fold (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
8559 SDValue Result = combineSelectAndUseCommutative(N, false, DCI);
8560 if (Result.getNode())
8567 /// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
8568 /// the bits being cleared by the AND are not demanded by the BFI.
8569 static SDValue PerformBFICombine(SDNode *N,
8570 TargetLowering::DAGCombinerInfo &DCI) {
8571 SDValue N1 = N->getOperand(1);
8572 if (N1.getOpcode() == ISD::AND) {
8573 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
8576 unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
8577 unsigned LSB = countTrailingZeros(~InvMask);
8578 unsigned Width = (32 - countLeadingZeros(~InvMask)) - LSB;
8579 unsigned Mask = (1 << Width)-1;
8580 unsigned Mask2 = N11C->getZExtValue();
8581 if ((Mask & (~Mask2)) == 0)
8582 return DCI.DAG.getNode(ARMISD::BFI, SDLoc(N), N->getValueType(0),
8583 N->getOperand(0), N1.getOperand(0),
8589 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for
8590 /// ARMISD::VMOVRRD.
8591 static SDValue PerformVMOVRRDCombine(SDNode *N,
8592 TargetLowering::DAGCombinerInfo &DCI,
8593 const ARMSubtarget *Subtarget) {
8594 // vmovrrd(vmovdrr x, y) -> x,y
8595 SDValue InDouble = N->getOperand(0);
8596 if (InDouble.getOpcode() == ARMISD::VMOVDRR && !Subtarget->isFPOnlySP())
8597 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
8599 // vmovrrd(load f64) -> (load i32), (load i32)
8600 SDNode *InNode = InDouble.getNode();
8601 if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
8602 InNode->getValueType(0) == MVT::f64 &&
8603 InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
8604 !cast<LoadSDNode>(InNode)->isVolatile()) {
8605 // TODO: Should this be done for non-FrameIndex operands?
8606 LoadSDNode *LD = cast<LoadSDNode>(InNode);
8608 SelectionDAG &DAG = DCI.DAG;
8610 SDValue BasePtr = LD->getBasePtr();
8611 SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
8612 LD->getPointerInfo(), LD->isVolatile(),
8613 LD->isNonTemporal(), LD->isInvariant(),
8614 LD->getAlignment());
8616 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8617 DAG.getConstant(4, MVT::i32));
8618 SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
8619 LD->getPointerInfo(), LD->isVolatile(),
8620 LD->isNonTemporal(), LD->isInvariant(),
8621 std::min(4U, LD->getAlignment() / 2));
8623 DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
8624 if (DCI.DAG.getTargetLoweringInfo().isBigEndian())
8625 std::swap (NewLD1, NewLD2);
8626 SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
8633 /// PerformVMOVDRRCombine - Target-specific dag combine xforms for
8634 /// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
8635 static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
8636 // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
8637 SDValue Op0 = N->getOperand(0);
8638 SDValue Op1 = N->getOperand(1);
8639 if (Op0.getOpcode() == ISD::BITCAST)
8640 Op0 = Op0.getOperand(0);
8641 if (Op1.getOpcode() == ISD::BITCAST)
8642 Op1 = Op1.getOperand(0);
8643 if (Op0.getOpcode() == ARMISD::VMOVRRD &&
8644 Op0.getNode() == Op1.getNode() &&
8645 Op0.getResNo() == 0 && Op1.getResNo() == 1)
8646 return DAG.getNode(ISD::BITCAST, SDLoc(N),
8647 N->getValueType(0), Op0.getOperand(0));
8651 /// PerformSTORECombine - Target-specific dag combine xforms for
8653 static SDValue PerformSTORECombine(SDNode *N,
8654 TargetLowering::DAGCombinerInfo &DCI) {
8655 StoreSDNode *St = cast<StoreSDNode>(N);
8656 if (St->isVolatile())
8659 // Optimize trunc store (of multiple scalars) to shuffle and store. First,
8660 // pack all of the elements in one place. Next, store to memory in fewer
8662 SDValue StVal = St->getValue();
8663 EVT VT = StVal.getValueType();
8664 if (St->isTruncatingStore() && VT.isVector()) {
8665 SelectionDAG &DAG = DCI.DAG;
8666 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8667 EVT StVT = St->getMemoryVT();
8668 unsigned NumElems = VT.getVectorNumElements();
8669 assert(StVT != VT && "Cannot truncate to the same type");
8670 unsigned FromEltSz = VT.getVectorElementType().getSizeInBits();
8671 unsigned ToEltSz = StVT.getVectorElementType().getSizeInBits();
8673 // From, To sizes and ElemCount must be pow of two
8674 if (!isPowerOf2_32(NumElems * FromEltSz * ToEltSz)) return SDValue();
8676 // We are going to use the original vector elt for storing.
8677 // Accumulated smaller vector elements must be a multiple of the store size.
8678 if (0 != (NumElems * FromEltSz) % ToEltSz) return SDValue();
8680 unsigned SizeRatio = FromEltSz / ToEltSz;
8681 assert(SizeRatio * NumElems * ToEltSz == VT.getSizeInBits());
8683 // Create a type on which we perform the shuffle.
8684 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), StVT.getScalarType(),
8685 NumElems*SizeRatio);
8686 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
8689 SDValue WideVec = DAG.getNode(ISD::BITCAST, DL, WideVecVT, StVal);
8690 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
8691 for (unsigned i = 0; i < NumElems; ++i)
8692 ShuffleVec[i] = TLI.isBigEndian() ? (i+1) * SizeRatio - 1 : i * SizeRatio;
8694 // Can't shuffle using an illegal type.
8695 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
8697 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, DL, WideVec,
8698 DAG.getUNDEF(WideVec.getValueType()),
8700 // At this point all of the data is stored at the bottom of the
8701 // register. We now need to save it to mem.
8703 // Find the largest store unit
8704 MVT StoreType = MVT::i8;
8705 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
8706 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
8707 MVT Tp = (MVT::SimpleValueType)tp;
8708 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToEltSz)
8711 // Didn't find a legal store type.
8712 if (!TLI.isTypeLegal(StoreType))
8715 // Bitcast the original vector into a vector of store-size units
8716 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
8717 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
8718 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
8719 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, DL, StoreVecVT, Shuff);
8720 SmallVector<SDValue, 8> Chains;
8721 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
8722 TLI.getPointerTy());
8723 SDValue BasePtr = St->getBasePtr();
8725 // Perform one or more big stores into memory.
8726 unsigned E = (ToEltSz*NumElems)/StoreType.getSizeInBits();
8727 for (unsigned I = 0; I < E; I++) {
8728 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
8729 StoreType, ShuffWide,
8730 DAG.getIntPtrConstant(I));
8731 SDValue Ch = DAG.getStore(St->getChain(), DL, SubVec, BasePtr,
8732 St->getPointerInfo(), St->isVolatile(),
8733 St->isNonTemporal(), St->getAlignment());
8734 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
8736 Chains.push_back(Ch);
8738 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
8741 if (!ISD::isNormalStore(St))
8744 // Split a store of a VMOVDRR into two integer stores to avoid mixing NEON and
8745 // ARM stores of arguments in the same cache line.
8746 if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
8747 StVal.getNode()->hasOneUse()) {
8748 SelectionDAG &DAG = DCI.DAG;
8749 bool isBigEndian = DAG.getTargetLoweringInfo().isBigEndian();
8751 SDValue BasePtr = St->getBasePtr();
8752 SDValue NewST1 = DAG.getStore(St->getChain(), DL,
8753 StVal.getNode()->getOperand(isBigEndian ? 1 : 0 ),
8754 BasePtr, St->getPointerInfo(), St->isVolatile(),
8755 St->isNonTemporal(), St->getAlignment());
8757 SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
8758 DAG.getConstant(4, MVT::i32));
8759 return DAG.getStore(NewST1.getValue(0), DL,
8760 StVal.getNode()->getOperand(isBigEndian ? 0 : 1),
8761 OffsetPtr, St->getPointerInfo(), St->isVolatile(),
8762 St->isNonTemporal(),
8763 std::min(4U, St->getAlignment() / 2));
8766 if (StVal.getValueType() != MVT::i64 ||
8767 StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
8770 // Bitcast an i64 store extracted from a vector to f64.
8771 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8772 SelectionDAG &DAG = DCI.DAG;
8774 SDValue IntVec = StVal.getOperand(0);
8775 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8776 IntVec.getValueType().getVectorNumElements());
8777 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
8778 SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
8779 Vec, StVal.getOperand(1));
8781 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
8782 // Make the DAGCombiner fold the bitcasts.
8783 DCI.AddToWorklist(Vec.getNode());
8784 DCI.AddToWorklist(ExtElt.getNode());
8785 DCI.AddToWorklist(V.getNode());
8786 return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
8787 St->getPointerInfo(), St->isVolatile(),
8788 St->isNonTemporal(), St->getAlignment(),
8792 /// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
8793 /// are normal, non-volatile loads. If so, it is profitable to bitcast an
8794 /// i64 vector to have f64 elements, since the value can then be loaded
8795 /// directly into a VFP register.
8796 static bool hasNormalLoadOperand(SDNode *N) {
8797 unsigned NumElts = N->getValueType(0).getVectorNumElements();
8798 for (unsigned i = 0; i < NumElts; ++i) {
8799 SDNode *Elt = N->getOperand(i).getNode();
8800 if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
8806 /// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
8807 /// ISD::BUILD_VECTOR.
8808 static SDValue PerformBUILD_VECTORCombine(SDNode *N,
8809 TargetLowering::DAGCombinerInfo &DCI,
8810 const ARMSubtarget *Subtarget) {
8811 // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
8812 // VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
8813 // into a pair of GPRs, which is fine when the value is used as a scalar,
8814 // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
8815 SelectionDAG &DAG = DCI.DAG;
8816 if (N->getNumOperands() == 2) {
8817 SDValue RV = PerformVMOVDRRCombine(N, DAG);
8822 // Load i64 elements as f64 values so that type legalization does not split
8823 // them up into i32 values.
8824 EVT VT = N->getValueType(0);
8825 if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
8828 SmallVector<SDValue, 8> Ops;
8829 unsigned NumElts = VT.getVectorNumElements();
8830 for (unsigned i = 0; i < NumElts; ++i) {
8831 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
8833 // Make the DAGCombiner fold the bitcast.
8834 DCI.AddToWorklist(V.getNode());
8836 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
8837 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops);
8838 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8841 /// \brief Target-specific dag combine xforms for ARMISD::BUILD_VECTOR.
8843 PerformARMBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
8844 // ARMISD::BUILD_VECTOR is introduced when legalizing ISD::BUILD_VECTOR.
8845 // At that time, we may have inserted bitcasts from integer to float.
8846 // If these bitcasts have survived DAGCombine, change the lowering of this
8847 // BUILD_VECTOR in something more vector friendly, i.e., that does not
8848 // force to use floating point types.
8850 // Make sure we can change the type of the vector.
8851 // This is possible iff:
8852 // 1. The vector is only used in a bitcast to a integer type. I.e.,
8853 // 1.1. Vector is used only once.
8854 // 1.2. Use is a bit convert to an integer type.
8855 // 2. The size of its operands are 32-bits (64-bits are not legal).
8856 EVT VT = N->getValueType(0);
8857 EVT EltVT = VT.getVectorElementType();
8859 // Check 1.1. and 2.
8860 if (EltVT.getSizeInBits() != 32 || !N->hasOneUse())
8863 // By construction, the input type must be float.
8864 assert(EltVT == MVT::f32 && "Unexpected type!");
8867 SDNode *Use = *N->use_begin();
8868 if (Use->getOpcode() != ISD::BITCAST ||
8869 Use->getValueType(0).isFloatingPoint())
8872 // Check profitability.
8873 // Model is, if more than half of the relevant operands are bitcast from
8874 // i32, turn the build_vector into a sequence of insert_vector_elt.
8875 // Relevant operands are everything that is not statically
8876 // (i.e., at compile time) bitcasted.
8877 unsigned NumOfBitCastedElts = 0;
8878 unsigned NumElts = VT.getVectorNumElements();
8879 unsigned NumOfRelevantElts = NumElts;
8880 for (unsigned Idx = 0; Idx < NumElts; ++Idx) {
8881 SDValue Elt = N->getOperand(Idx);
8882 if (Elt->getOpcode() == ISD::BITCAST) {
8883 // Assume only bit cast to i32 will go away.
8884 if (Elt->getOperand(0).getValueType() == MVT::i32)
8885 ++NumOfBitCastedElts;
8886 } else if (Elt.getOpcode() == ISD::UNDEF || isa<ConstantSDNode>(Elt))
8887 // Constants are statically casted, thus do not count them as
8888 // relevant operands.
8889 --NumOfRelevantElts;
8892 // Check if more than half of the elements require a non-free bitcast.
8893 if (NumOfBitCastedElts <= NumOfRelevantElts / 2)
8896 SelectionDAG &DAG = DCI.DAG;
8897 // Create the new vector type.
8898 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
8899 // Check if the type is legal.
8900 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8901 if (!TLI.isTypeLegal(VecVT))
8905 // ARMISD::BUILD_VECTOR E1, E2, ..., EN.
8906 // => BITCAST INSERT_VECTOR_ELT
8907 // (INSERT_VECTOR_ELT (...), (BITCAST EN-1), N-1),
8909 SDValue Vec = DAG.getUNDEF(VecVT);
8911 for (unsigned Idx = 0 ; Idx < NumElts; ++Idx) {
8912 SDValue V = N->getOperand(Idx);
8913 if (V.getOpcode() == ISD::UNDEF)
8915 if (V.getOpcode() == ISD::BITCAST &&
8916 V->getOperand(0).getValueType() == MVT::i32)
8917 // Fold obvious case.
8918 V = V.getOperand(0);
8920 V = DAG.getNode(ISD::BITCAST, SDLoc(V), MVT::i32, V);
8921 // Make the DAGCombiner fold the bitcasts.
8922 DCI.AddToWorklist(V.getNode());
8924 SDValue LaneIdx = DAG.getConstant(Idx, MVT::i32);
8925 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VecVT, Vec, V, LaneIdx);
8927 Vec = DAG.getNode(ISD::BITCAST, dl, VT, Vec);
8928 // Make the DAGCombiner fold the bitcasts.
8929 DCI.AddToWorklist(Vec.getNode());
8933 /// PerformInsertEltCombine - Target-specific dag combine xforms for
8934 /// ISD::INSERT_VECTOR_ELT.
8935 static SDValue PerformInsertEltCombine(SDNode *N,
8936 TargetLowering::DAGCombinerInfo &DCI) {
8937 // Bitcast an i64 load inserted into a vector to f64.
8938 // Otherwise, the i64 value will be legalized to a pair of i32 values.
8939 EVT VT = N->getValueType(0);
8940 SDNode *Elt = N->getOperand(1).getNode();
8941 if (VT.getVectorElementType() != MVT::i64 ||
8942 !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
8945 SelectionDAG &DAG = DCI.DAG;
8947 EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
8948 VT.getVectorNumElements());
8949 SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
8950 SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
8951 // Make the DAGCombiner fold the bitcasts.
8952 DCI.AddToWorklist(Vec.getNode());
8953 DCI.AddToWorklist(V.getNode());
8954 SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
8955 Vec, V, N->getOperand(2));
8956 return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
8959 /// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
8960 /// ISD::VECTOR_SHUFFLE.
8961 static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
8962 // The LLVM shufflevector instruction does not require the shuffle mask
8963 // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
8964 // have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
8965 // operands do not match the mask length, they are extended by concatenating
8966 // them with undef vectors. That is probably the right thing for other
8967 // targets, but for NEON it is better to concatenate two double-register
8968 // size vector operands into a single quad-register size vector. Do that
8969 // transformation here:
8970 // shuffle(concat(v1, undef), concat(v2, undef)) ->
8971 // shuffle(concat(v1, v2), undef)
8972 SDValue Op0 = N->getOperand(0);
8973 SDValue Op1 = N->getOperand(1);
8974 if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
8975 Op1.getOpcode() != ISD::CONCAT_VECTORS ||
8976 Op0.getNumOperands() != 2 ||
8977 Op1.getNumOperands() != 2)
8979 SDValue Concat0Op1 = Op0.getOperand(1);
8980 SDValue Concat1Op1 = Op1.getOperand(1);
8981 if (Concat0Op1.getOpcode() != ISD::UNDEF ||
8982 Concat1Op1.getOpcode() != ISD::UNDEF)
8984 // Skip the transformation if any of the types are illegal.
8985 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8986 EVT VT = N->getValueType(0);
8987 if (!TLI.isTypeLegal(VT) ||
8988 !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
8989 !TLI.isTypeLegal(Concat1Op1.getValueType()))
8992 SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
8993 Op0.getOperand(0), Op1.getOperand(0));
8994 // Translate the shuffle mask.
8995 SmallVector<int, 16> NewMask;
8996 unsigned NumElts = VT.getVectorNumElements();
8997 unsigned HalfElts = NumElts/2;
8998 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8999 for (unsigned n = 0; n < NumElts; ++n) {
9000 int MaskElt = SVN->getMaskElt(n);
9002 if (MaskElt < (int)HalfElts)
9004 else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
9005 NewElt = HalfElts + MaskElt - NumElts;
9006 NewMask.push_back(NewElt);
9008 return DAG.getVectorShuffle(VT, SDLoc(N), NewConcat,
9009 DAG.getUNDEF(VT), NewMask.data());
9012 /// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
9013 /// NEON load/store intrinsics to merge base address updates.
9014 static SDValue CombineBaseUpdate(SDNode *N,
9015 TargetLowering::DAGCombinerInfo &DCI) {
9016 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9019 SelectionDAG &DAG = DCI.DAG;
9020 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
9021 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
9022 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
9023 SDValue Addr = N->getOperand(AddrOpIdx);
9025 // Search for a use of the address operand that is an increment.
9026 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
9027 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
9029 if (User->getOpcode() != ISD::ADD ||
9030 UI.getUse().getResNo() != Addr.getResNo())
9033 // Check that the add is independent of the load/store. Otherwise, folding
9034 // it would create a cycle.
9035 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
9038 // Find the new opcode for the updating load/store.
9040 bool isLaneOp = false;
9041 unsigned NewOpc = 0;
9042 unsigned NumVecs = 0;
9044 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
9046 default: llvm_unreachable("unexpected intrinsic for Neon base update");
9047 case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
9049 case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
9051 case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
9053 case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
9055 case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
9056 NumVecs = 2; isLaneOp = true; break;
9057 case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
9058 NumVecs = 3; isLaneOp = true; break;
9059 case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
9060 NumVecs = 4; isLaneOp = true; break;
9061 case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
9062 NumVecs = 1; isLoad = false; break;
9063 case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
9064 NumVecs = 2; isLoad = false; break;
9065 case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
9066 NumVecs = 3; isLoad = false; break;
9067 case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
9068 NumVecs = 4; isLoad = false; break;
9069 case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
9070 NumVecs = 2; isLoad = false; isLaneOp = true; break;
9071 case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
9072 NumVecs = 3; isLoad = false; isLaneOp = true; break;
9073 case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
9074 NumVecs = 4; isLoad = false; isLaneOp = true; break;
9078 switch (N->getOpcode()) {
9079 default: llvm_unreachable("unexpected opcode for Neon base update");
9080 case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
9081 case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
9082 case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
9086 // Find the size of memory referenced by the load/store.
9089 VecTy = N->getValueType(0);
9091 VecTy = N->getOperand(AddrOpIdx+1).getValueType();
9092 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
9094 NumBytes /= VecTy.getVectorNumElements();
9096 // If the increment is a constant, it must match the memory ref size.
9097 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
9098 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
9099 uint64_t IncVal = CInc->getZExtValue();
9100 if (IncVal != NumBytes)
9102 } else if (NumBytes >= 3 * 16) {
9103 // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
9104 // separate instructions that make it harder to use a non-constant update.
9108 // Create the new updating load/store node.
9110 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
9112 for (n = 0; n < NumResultVecs; ++n)
9114 Tys[n++] = MVT::i32;
9115 Tys[n] = MVT::Other;
9116 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs+2));
9117 SmallVector<SDValue, 8> Ops;
9118 Ops.push_back(N->getOperand(0)); // incoming chain
9119 Ops.push_back(N->getOperand(AddrOpIdx));
9121 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
9122 Ops.push_back(N->getOperand(i));
9124 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
9125 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
9126 Ops, MemInt->getMemoryVT(),
9127 MemInt->getMemOperand());
9130 std::vector<SDValue> NewResults;
9131 for (unsigned i = 0; i < NumResultVecs; ++i) {
9132 NewResults.push_back(SDValue(UpdN.getNode(), i));
9134 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
9135 DCI.CombineTo(N, NewResults);
9136 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
9143 /// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
9144 /// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
9145 /// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
9147 static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
9148 SelectionDAG &DAG = DCI.DAG;
9149 EVT VT = N->getValueType(0);
9150 // vldN-dup instructions only support 64-bit vectors for N > 1.
9151 if (!VT.is64BitVector())
9154 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
9155 SDNode *VLD = N->getOperand(0).getNode();
9156 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
9158 unsigned NumVecs = 0;
9159 unsigned NewOpc = 0;
9160 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
9161 if (IntNo == Intrinsic::arm_neon_vld2lane) {
9163 NewOpc = ARMISD::VLD2DUP;
9164 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
9166 NewOpc = ARMISD::VLD3DUP;
9167 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
9169 NewOpc = ARMISD::VLD4DUP;
9174 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
9175 // numbers match the load.
9176 unsigned VLDLaneNo =
9177 cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
9178 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9180 // Ignore uses of the chain result.
9181 if (UI.getUse().getResNo() == NumVecs)
9184 if (User->getOpcode() != ARMISD::VDUPLANE ||
9185 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
9189 // Create the vldN-dup node.
9192 for (n = 0; n < NumVecs; ++n)
9194 Tys[n] = MVT::Other;
9195 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs+1));
9196 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
9197 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
9198 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys,
9199 Ops, VLDMemInt->getMemoryVT(),
9200 VLDMemInt->getMemOperand());
9203 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
9205 unsigned ResNo = UI.getUse().getResNo();
9206 // Ignore uses of the chain result.
9207 if (ResNo == NumVecs)
9210 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
9213 // Now the vldN-lane intrinsic is dead except for its chain result.
9214 // Update uses of the chain.
9215 std::vector<SDValue> VLDDupResults;
9216 for (unsigned n = 0; n < NumVecs; ++n)
9217 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
9218 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
9219 DCI.CombineTo(VLD, VLDDupResults);
9224 /// PerformVDUPLANECombine - Target-specific dag combine xforms for
9225 /// ARMISD::VDUPLANE.
9226 static SDValue PerformVDUPLANECombine(SDNode *N,
9227 TargetLowering::DAGCombinerInfo &DCI) {
9228 SDValue Op = N->getOperand(0);
9230 // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
9231 // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
9232 if (CombineVLDDUP(N, DCI))
9233 return SDValue(N, 0);
9235 // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
9236 // redundant. Ignore bit_converts for now; element sizes are checked below.
9237 while (Op.getOpcode() == ISD::BITCAST)
9238 Op = Op.getOperand(0);
9239 if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
9242 // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
9243 unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
9244 // The canonical VMOV for a zero vector uses a 32-bit element size.
9245 unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9247 if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
9249 EVT VT = N->getValueType(0);
9250 if (EltSize > VT.getVectorElementType().getSizeInBits())
9253 return DCI.DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
9256 // isConstVecPow2 - Return true if each vector element is a power of 2, all
9257 // elements are the same constant, C, and Log2(C) ranges from 1 to 32.
9258 static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
9262 for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
9264 ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
9269 APFloat APF = C->getValueAPF();
9270 if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
9271 != APFloat::opOK || !isExact)
9274 c0 = (I == 0) ? cN : c0;
9275 if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
9282 /// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
9283 /// can replace combinations of VMUL and VCVT (floating-point to integer)
9284 /// when the VMUL has a constant operand that is a power of 2.
9286 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9287 /// vmul.f32 d16, d17, d16
9288 /// vcvt.s32.f32 d16, d16
9290 /// vcvt.s32.f32 d16, d16, #3
9291 static SDValue PerformVCVTCombine(SDNode *N,
9292 TargetLowering::DAGCombinerInfo &DCI,
9293 const ARMSubtarget *Subtarget) {
9294 SelectionDAG &DAG = DCI.DAG;
9295 SDValue Op = N->getOperand(0);
9297 if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
9298 Op.getOpcode() != ISD::FMUL)
9302 SDValue N0 = Op->getOperand(0);
9303 SDValue ConstVec = Op->getOperand(1);
9304 bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
9306 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9307 !isConstVecPow2(ConstVec, isSigned, C))
9310 MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
9311 MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
9312 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9313 // These instructions only exist converting from f32 to i32. We can handle
9314 // smaller integers by generating an extra truncate, but larger ones would
9319 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
9320 Intrinsic::arm_neon_vcvtfp2fxu;
9321 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9322 SDValue FixConv = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9323 NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9324 DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
9325 DAG.getConstant(Log2_64(C), MVT::i32));
9327 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9328 FixConv = DAG.getNode(ISD::TRUNCATE, SDLoc(N), N->getValueType(0), FixConv);
9333 /// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
9334 /// can replace combinations of VCVT (integer to floating-point) and VDIV
9335 /// when the VDIV has a constant operand that is a power of 2.
9337 /// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
9338 /// vcvt.f32.s32 d16, d16
9339 /// vdiv.f32 d16, d17, d16
9341 /// vcvt.f32.s32 d16, d16, #3
9342 static SDValue PerformVDIVCombine(SDNode *N,
9343 TargetLowering::DAGCombinerInfo &DCI,
9344 const ARMSubtarget *Subtarget) {
9345 SelectionDAG &DAG = DCI.DAG;
9346 SDValue Op = N->getOperand(0);
9347 unsigned OpOpcode = Op.getNode()->getOpcode();
9349 if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
9350 (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
9354 SDValue ConstVec = N->getOperand(1);
9355 bool isSigned = OpOpcode == ISD::SINT_TO_FP;
9357 if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
9358 !isConstVecPow2(ConstVec, isSigned, C))
9361 MVT FloatTy = N->getSimpleValueType(0).getVectorElementType();
9362 MVT IntTy = Op.getOperand(0).getSimpleValueType().getVectorElementType();
9363 if (FloatTy.getSizeInBits() != 32 || IntTy.getSizeInBits() > 32) {
9364 // These instructions only exist converting from i32 to f32. We can handle
9365 // smaller integers by generating an extra extend, but larger ones would
9370 SDValue ConvInput = Op.getOperand(0);
9371 unsigned NumLanes = Op.getValueType().getVectorNumElements();
9372 if (IntTy.getSizeInBits() < FloatTy.getSizeInBits())
9373 ConvInput = DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
9374 SDLoc(N), NumLanes == 2 ? MVT::v2i32 : MVT::v4i32,
9377 unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
9378 Intrinsic::arm_neon_vcvtfxu2fp;
9379 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N),
9381 DAG.getConstant(IntrinsicOpcode, MVT::i32),
9382 ConvInput, DAG.getConstant(Log2_64(C), MVT::i32));
9385 /// Getvshiftimm - Check if this is a valid build_vector for the immediate
9386 /// operand of a vector shift operation, where all the elements of the
9387 /// build_vector must have the same constant integer value.
9388 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
9389 // Ignore bit_converts.
9390 while (Op.getOpcode() == ISD::BITCAST)
9391 Op = Op.getOperand(0);
9392 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
9393 APInt SplatBits, SplatUndef;
9394 unsigned SplatBitSize;
9396 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
9397 HasAnyUndefs, ElementBits) ||
9398 SplatBitSize > ElementBits)
9400 Cnt = SplatBits.getSExtValue();
9404 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
9405 /// operand of a vector shift left operation. That value must be in the range:
9406 /// 0 <= Value < ElementBits for a left shift; or
9407 /// 0 <= Value <= ElementBits for a long left shift.
9408 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
9409 assert(VT.isVector() && "vector shift count is not a vector type");
9410 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9411 if (! getVShiftImm(Op, ElementBits, Cnt))
9413 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
9416 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
9417 /// operand of a vector shift right operation. For a shift opcode, the value
9418 /// is positive, but for an intrinsic the value count must be negative. The
9419 /// absolute value must be in the range:
9420 /// 1 <= |Value| <= ElementBits for a right shift; or
9421 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
9422 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
9424 assert(VT.isVector() && "vector shift count is not a vector type");
9425 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
9426 if (! getVShiftImm(Op, ElementBits, Cnt))
9430 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
9433 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
9434 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
9435 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
9438 // Don't do anything for most intrinsics.
9441 // Vector shifts: check for immediate versions and lower them.
9442 // Note: This is done during DAG combining instead of DAG legalizing because
9443 // the build_vectors for 64-bit vector element shift counts are generally
9444 // not legal, and it is hard to see their values after they get legalized to
9445 // loads from a constant pool.
9446 case Intrinsic::arm_neon_vshifts:
9447 case Intrinsic::arm_neon_vshiftu:
9448 case Intrinsic::arm_neon_vrshifts:
9449 case Intrinsic::arm_neon_vrshiftu:
9450 case Intrinsic::arm_neon_vrshiftn:
9451 case Intrinsic::arm_neon_vqshifts:
9452 case Intrinsic::arm_neon_vqshiftu:
9453 case Intrinsic::arm_neon_vqshiftsu:
9454 case Intrinsic::arm_neon_vqshiftns:
9455 case Intrinsic::arm_neon_vqshiftnu:
9456 case Intrinsic::arm_neon_vqshiftnsu:
9457 case Intrinsic::arm_neon_vqrshiftns:
9458 case Intrinsic::arm_neon_vqrshiftnu:
9459 case Intrinsic::arm_neon_vqrshiftnsu: {
9460 EVT VT = N->getOperand(1).getValueType();
9462 unsigned VShiftOpc = 0;
9465 case Intrinsic::arm_neon_vshifts:
9466 case Intrinsic::arm_neon_vshiftu:
9467 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
9468 VShiftOpc = ARMISD::VSHL;
9471 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
9472 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
9473 ARMISD::VSHRs : ARMISD::VSHRu);
9478 case Intrinsic::arm_neon_vrshifts:
9479 case Intrinsic::arm_neon_vrshiftu:
9480 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
9484 case Intrinsic::arm_neon_vqshifts:
9485 case Intrinsic::arm_neon_vqshiftu:
9486 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9490 case Intrinsic::arm_neon_vqshiftsu:
9491 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
9493 llvm_unreachable("invalid shift count for vqshlu intrinsic");
9495 case Intrinsic::arm_neon_vrshiftn:
9496 case Intrinsic::arm_neon_vqshiftns:
9497 case Intrinsic::arm_neon_vqshiftnu:
9498 case Intrinsic::arm_neon_vqshiftnsu:
9499 case Intrinsic::arm_neon_vqrshiftns:
9500 case Intrinsic::arm_neon_vqrshiftnu:
9501 case Intrinsic::arm_neon_vqrshiftnsu:
9502 // Narrowing shifts require an immediate right shift.
9503 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
9505 llvm_unreachable("invalid shift count for narrowing vector shift "
9509 llvm_unreachable("unhandled vector shift");
9513 case Intrinsic::arm_neon_vshifts:
9514 case Intrinsic::arm_neon_vshiftu:
9515 // Opcode already set above.
9517 case Intrinsic::arm_neon_vrshifts:
9518 VShiftOpc = ARMISD::VRSHRs; break;
9519 case Intrinsic::arm_neon_vrshiftu:
9520 VShiftOpc = ARMISD::VRSHRu; break;
9521 case Intrinsic::arm_neon_vrshiftn:
9522 VShiftOpc = ARMISD::VRSHRN; break;
9523 case Intrinsic::arm_neon_vqshifts:
9524 VShiftOpc = ARMISD::VQSHLs; break;
9525 case Intrinsic::arm_neon_vqshiftu:
9526 VShiftOpc = ARMISD::VQSHLu; break;
9527 case Intrinsic::arm_neon_vqshiftsu:
9528 VShiftOpc = ARMISD::VQSHLsu; break;
9529 case Intrinsic::arm_neon_vqshiftns:
9530 VShiftOpc = ARMISD::VQSHRNs; break;
9531 case Intrinsic::arm_neon_vqshiftnu:
9532 VShiftOpc = ARMISD::VQSHRNu; break;
9533 case Intrinsic::arm_neon_vqshiftnsu:
9534 VShiftOpc = ARMISD::VQSHRNsu; break;
9535 case Intrinsic::arm_neon_vqrshiftns:
9536 VShiftOpc = ARMISD::VQRSHRNs; break;
9537 case Intrinsic::arm_neon_vqrshiftnu:
9538 VShiftOpc = ARMISD::VQRSHRNu; break;
9539 case Intrinsic::arm_neon_vqrshiftnsu:
9540 VShiftOpc = ARMISD::VQRSHRNsu; break;
9543 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9544 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
9547 case Intrinsic::arm_neon_vshiftins: {
9548 EVT VT = N->getOperand(1).getValueType();
9550 unsigned VShiftOpc = 0;
9552 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
9553 VShiftOpc = ARMISD::VSLI;
9554 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
9555 VShiftOpc = ARMISD::VSRI;
9557 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
9560 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
9561 N->getOperand(1), N->getOperand(2),
9562 DAG.getConstant(Cnt, MVT::i32));
9565 case Intrinsic::arm_neon_vqrshifts:
9566 case Intrinsic::arm_neon_vqrshiftu:
9567 // No immediate versions of these to check for.
9574 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
9575 /// lowers them. As with the vector shift intrinsics, this is done during DAG
9576 /// combining instead of DAG legalizing because the build_vectors for 64-bit
9577 /// vector element shift counts are generally not legal, and it is hard to see
9578 /// their values after they get legalized to loads from a constant pool.
9579 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
9580 const ARMSubtarget *ST) {
9581 EVT VT = N->getValueType(0);
9582 if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
9583 // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
9584 // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
9585 SDValue N1 = N->getOperand(1);
9586 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
9587 SDValue N0 = N->getOperand(0);
9588 if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
9589 DAG.MaskedValueIsZero(N0.getOperand(0),
9590 APInt::getHighBitsSet(32, 16)))
9591 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, N0, N1);
9595 // Nothing to be done for scalar shifts.
9596 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9597 if (!VT.isVector() || !TLI.isTypeLegal(VT))
9600 assert(ST->hasNEON() && "unexpected vector shift");
9603 switch (N->getOpcode()) {
9604 default: llvm_unreachable("unexpected shift opcode");
9607 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
9608 return DAG.getNode(ARMISD::VSHL, SDLoc(N), VT, N->getOperand(0),
9609 DAG.getConstant(Cnt, MVT::i32));
9614 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
9615 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
9616 ARMISD::VSHRs : ARMISD::VSHRu);
9617 return DAG.getNode(VShiftOpc, SDLoc(N), VT, N->getOperand(0),
9618 DAG.getConstant(Cnt, MVT::i32));
9624 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
9625 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
9626 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
9627 const ARMSubtarget *ST) {
9628 SDValue N0 = N->getOperand(0);
9630 // Check for sign- and zero-extensions of vector extract operations of 8-
9631 // and 16-bit vector elements. NEON supports these directly. They are
9632 // handled during DAG combining because type legalization will promote them
9633 // to 32-bit types and it is messy to recognize the operations after that.
9634 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
9635 SDValue Vec = N0.getOperand(0);
9636 SDValue Lane = N0.getOperand(1);
9637 EVT VT = N->getValueType(0);
9638 EVT EltVT = N0.getValueType();
9639 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
9641 if (VT == MVT::i32 &&
9642 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
9643 TLI.isTypeLegal(Vec.getValueType()) &&
9644 isa<ConstantSDNode>(Lane)) {
9647 switch (N->getOpcode()) {
9648 default: llvm_unreachable("unexpected opcode");
9649 case ISD::SIGN_EXTEND:
9650 Opc = ARMISD::VGETLANEs;
9652 case ISD::ZERO_EXTEND:
9653 case ISD::ANY_EXTEND:
9654 Opc = ARMISD::VGETLANEu;
9657 return DAG.getNode(Opc, SDLoc(N), VT, Vec, Lane);
9664 /// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
9665 /// to match f32 max/min patterns to use NEON vmax/vmin instructions.
9666 static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
9667 const ARMSubtarget *ST) {
9668 // If the target supports NEON, try to use vmax/vmin instructions for f32
9669 // selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
9670 // be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
9671 // a NaN; only do the transformation when it matches that behavior.
9673 // For now only do this when using NEON for FP operations; if using VFP, it
9674 // is not obvious that the benefit outweighs the cost of switching to the
9676 if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
9677 N->getValueType(0) != MVT::f32)
9680 SDValue CondLHS = N->getOperand(0);
9681 SDValue CondRHS = N->getOperand(1);
9682 SDValue LHS = N->getOperand(2);
9683 SDValue RHS = N->getOperand(3);
9684 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
9686 unsigned Opcode = 0;
9688 if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
9689 IsReversed = false; // x CC y ? x : y
9690 } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
9691 IsReversed = true ; // x CC y ? y : x
9705 // If LHS is NaN, an ordered comparison will be false and the result will
9706 // be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
9707 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9708 IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
9709 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9711 // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
9712 // will return -0, so vmin can only be used for unsafe math or if one of
9713 // the operands is known to be nonzero.
9714 if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
9715 !DAG.getTarget().Options.UnsafeFPMath &&
9716 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9718 Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
9727 // If LHS is NaN, an ordered comparison will be false and the result will
9728 // be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
9729 // != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
9730 IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
9731 if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
9733 // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
9734 // will return +0, so vmax can only be used for unsafe math or if one of
9735 // the operands is known to be nonzero.
9736 if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
9737 !DAG.getTarget().Options.UnsafeFPMath &&
9738 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9740 Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
9746 return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), LHS, RHS);
9749 /// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
9751 ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
9752 SDValue Cmp = N->getOperand(4);
9753 if (Cmp.getOpcode() != ARMISD::CMPZ)
9754 // Only looking at EQ and NE cases.
9757 EVT VT = N->getValueType(0);
9759 SDValue LHS = Cmp.getOperand(0);
9760 SDValue RHS = Cmp.getOperand(1);
9761 SDValue FalseVal = N->getOperand(0);
9762 SDValue TrueVal = N->getOperand(1);
9763 SDValue ARMcc = N->getOperand(2);
9764 ARMCC::CondCodes CC =
9765 (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
9783 /// FIXME: Turn this into a target neutral optimization?
9785 if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
9786 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
9787 N->getOperand(3), Cmp);
9788 } else if (CC == ARMCC::EQ && TrueVal == RHS) {
9790 SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
9791 Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
9792 N->getOperand(3), NewCmp);
9795 if (Res.getNode()) {
9796 APInt KnownZero, KnownOne;
9797 DAG.computeKnownBits(SDValue(N,0), KnownZero, KnownOne);
9798 // Capture demanded bits information that would be otherwise lost.
9799 if (KnownZero == 0xfffffffe)
9800 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9801 DAG.getValueType(MVT::i1));
9802 else if (KnownZero == 0xffffff00)
9803 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9804 DAG.getValueType(MVT::i8));
9805 else if (KnownZero == 0xffff0000)
9806 Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
9807 DAG.getValueType(MVT::i16));
9813 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
9814 DAGCombinerInfo &DCI) const {
9815 switch (N->getOpcode()) {
9817 case ISD::ADDC: return PerformADDCCombine(N, DCI, Subtarget);
9818 case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
9819 case ISD::SUB: return PerformSUBCombine(N, DCI);
9820 case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
9821 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
9822 case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
9823 case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
9824 case ARMISD::BFI: return PerformBFICombine(N, DCI);
9825 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI, Subtarget);
9826 case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
9827 case ISD::STORE: return PerformSTORECombine(N, DCI);
9828 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI, Subtarget);
9829 case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
9830 case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
9831 case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
9832 case ISD::FP_TO_SINT:
9833 case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
9834 case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
9835 case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
9838 case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
9839 case ISD::SIGN_EXTEND:
9840 case ISD::ZERO_EXTEND:
9841 case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
9842 case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
9843 case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
9844 case ARMISD::VLD2DUP:
9845 case ARMISD::VLD3DUP:
9846 case ARMISD::VLD4DUP:
9847 return CombineBaseUpdate(N, DCI);
9848 case ARMISD::BUILD_VECTOR:
9849 return PerformARMBUILD_VECTORCombine(N, DCI);
9850 case ISD::INTRINSIC_VOID:
9851 case ISD::INTRINSIC_W_CHAIN:
9852 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
9853 case Intrinsic::arm_neon_vld1:
9854 case Intrinsic::arm_neon_vld2:
9855 case Intrinsic::arm_neon_vld3:
9856 case Intrinsic::arm_neon_vld4:
9857 case Intrinsic::arm_neon_vld2lane:
9858 case Intrinsic::arm_neon_vld3lane:
9859 case Intrinsic::arm_neon_vld4lane:
9860 case Intrinsic::arm_neon_vst1:
9861 case Intrinsic::arm_neon_vst2:
9862 case Intrinsic::arm_neon_vst3:
9863 case Intrinsic::arm_neon_vst4:
9864 case Intrinsic::arm_neon_vst2lane:
9865 case Intrinsic::arm_neon_vst3lane:
9866 case Intrinsic::arm_neon_vst4lane:
9867 return CombineBaseUpdate(N, DCI);
9875 bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
9877 return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
9880 bool ARMTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
9884 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
9885 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
9887 switch (VT.getSimpleVT().SimpleTy) {
9893 // Unaligned access can use (for example) LRDB, LRDH, LDR
9894 if (AllowsUnaligned) {
9896 *Fast = Subtarget->hasV7Ops();
9903 // For any little-endian targets with neon, we can support unaligned ld/st
9904 // of D and Q (e.g. {D0,D1}) registers by using vld1.i8/vst1.i8.
9905 // A big-endian target may also explicitly support unaligned accesses
9906 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
9916 static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
9917 unsigned AlignCheck) {
9918 return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
9919 (DstAlign == 0 || DstAlign % AlignCheck == 0));
9922 EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
9923 unsigned DstAlign, unsigned SrcAlign,
9924 bool IsMemset, bool ZeroMemset,
9926 MachineFunction &MF) const {
9927 const Function *F = MF.getFunction();
9929 // See if we can use NEON instructions for this...
9930 if ((!IsMemset || ZeroMemset) &&
9931 Subtarget->hasNEON() &&
9932 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
9933 Attribute::NoImplicitFloat)) {
9936 (memOpAlign(SrcAlign, DstAlign, 16) ||
9937 (allowsMisalignedMemoryAccesses(MVT::v2f64, 0, 1, &Fast) && Fast))) {
9939 } else if (Size >= 8 &&
9940 (memOpAlign(SrcAlign, DstAlign, 8) ||
9941 (allowsMisalignedMemoryAccesses(MVT::f64, 0, 1, &Fast) &&
9947 // Lowering to i32/i16 if the size permits.
9953 // Let the target-independent logic figure it out.
9957 bool ARMTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
9958 if (Val.getOpcode() != ISD::LOAD)
9961 EVT VT1 = Val.getValueType();
9962 if (!VT1.isSimple() || !VT1.isInteger() ||
9963 !VT2.isSimple() || !VT2.isInteger())
9966 switch (VT1.getSimpleVT().SimpleTy) {
9971 // 8-bit and 16-bit loads implicitly zero-extend to 32-bits.
9978 bool ARMTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
9979 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9982 if (!isTypeLegal(EVT::getEVT(Ty1)))
9985 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
9987 // Assuming the caller doesn't have a zeroext or signext return parameter,
9988 // truncation all the way down to i1 is valid.
9993 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
9998 switch (VT.getSimpleVT().SimpleTy) {
9999 default: return false;
10014 if ((V & (Scale - 1)) != 0)
10017 return V == (V & ((1LL << 5) - 1));
10020 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
10021 const ARMSubtarget *Subtarget) {
10022 bool isNeg = false;
10028 switch (VT.getSimpleVT().SimpleTy) {
10029 default: return false;
10034 // + imm12 or - imm8
10036 return V == (V & ((1LL << 8) - 1));
10037 return V == (V & ((1LL << 12) - 1));
10040 // Same as ARM mode. FIXME: NEON?
10041 if (!Subtarget->hasVFP2())
10046 return V == (V & ((1LL << 8) - 1));
10050 /// isLegalAddressImmediate - Return true if the integer value can be used
10051 /// as the offset of the target addressing mode for load / store of the
10053 static bool isLegalAddressImmediate(int64_t V, EVT VT,
10054 const ARMSubtarget *Subtarget) {
10058 if (!VT.isSimple())
10061 if (Subtarget->isThumb1Only())
10062 return isLegalT1AddressImmediate(V, VT);
10063 else if (Subtarget->isThumb2())
10064 return isLegalT2AddressImmediate(V, VT, Subtarget);
10069 switch (VT.getSimpleVT().SimpleTy) {
10070 default: return false;
10075 return V == (V & ((1LL << 12) - 1));
10078 return V == (V & ((1LL << 8) - 1));
10081 if (!Subtarget->hasVFP2()) // FIXME: NEON?
10086 return V == (V & ((1LL << 8) - 1));
10090 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
10092 int Scale = AM.Scale;
10096 switch (VT.getSimpleVT().SimpleTy) {
10097 default: return false;
10105 Scale = Scale & ~1;
10106 return Scale == 2 || Scale == 4 || Scale == 8;
10109 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10113 // Note, we allow "void" uses (basically, uses that aren't loads or
10114 // stores), because arm allows folding a scale into many arithmetic
10115 // operations. This should be made more precise and revisited later.
10117 // Allow r << imm, but the imm has to be a multiple of two.
10118 if (Scale & 1) return false;
10119 return isPowerOf2_32(Scale);
10123 /// isLegalAddressingMode - Return true if the addressing mode represented
10124 /// by AM is legal for this target, for a load/store of the specified type.
10125 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
10127 EVT VT = getValueType(Ty, true);
10128 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
10131 // Can never fold addr of global into load/store.
10135 switch (AM.Scale) {
10136 case 0: // no scale reg, must be "r+i" or "r", or "i".
10139 if (Subtarget->isThumb1Only())
10143 // ARM doesn't support any R+R*scale+imm addr modes.
10147 if (!VT.isSimple())
10150 if (Subtarget->isThumb2())
10151 return isLegalT2ScaledAddressingMode(AM, VT);
10153 int Scale = AM.Scale;
10154 switch (VT.getSimpleVT().SimpleTy) {
10155 default: return false;
10159 if (Scale < 0) Scale = -Scale;
10163 return isPowerOf2_32(Scale & ~1);
10167 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
10172 // Note, we allow "void" uses (basically, uses that aren't loads or
10173 // stores), because arm allows folding a scale into many arithmetic
10174 // operations. This should be made more precise and revisited later.
10176 // Allow r << imm, but the imm has to be a multiple of two.
10177 if (Scale & 1) return false;
10178 return isPowerOf2_32(Scale);
10184 /// isLegalICmpImmediate - Return true if the specified immediate is legal
10185 /// icmp immediate, that is the target has icmp instructions which can compare
10186 /// a register against the immediate without having to materialize the
10187 /// immediate into a register.
10188 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
10189 // Thumb2 and ARM modes can use cmn for negative immediates.
10190 if (!Subtarget->isThumb())
10191 return ARM_AM::getSOImmVal(llvm::abs64(Imm)) != -1;
10192 if (Subtarget->isThumb2())
10193 return ARM_AM::getT2SOImmVal(llvm::abs64(Imm)) != -1;
10194 // Thumb1 doesn't have cmn, and only 8-bit immediates.
10195 return Imm >= 0 && Imm <= 255;
10198 /// isLegalAddImmediate - Return true if the specified immediate is a legal add
10199 /// *or sub* immediate, that is the target has add or sub instructions which can
10200 /// add a register with the immediate without having to materialize the
10201 /// immediate into a register.
10202 bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
10203 // Same encoding for add/sub, just flip the sign.
10204 int64_t AbsImm = llvm::abs64(Imm);
10205 if (!Subtarget->isThumb())
10206 return ARM_AM::getSOImmVal(AbsImm) != -1;
10207 if (Subtarget->isThumb2())
10208 return ARM_AM::getT2SOImmVal(AbsImm) != -1;
10209 // Thumb1 only has 8-bit unsigned immediate.
10210 return AbsImm >= 0 && AbsImm <= 255;
10213 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
10214 bool isSEXTLoad, SDValue &Base,
10215 SDValue &Offset, bool &isInc,
10216 SelectionDAG &DAG) {
10217 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10220 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
10221 // AddressingMode 3
10222 Base = Ptr->getOperand(0);
10223 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10224 int RHSC = (int)RHS->getZExtValue();
10225 if (RHSC < 0 && RHSC > -256) {
10226 assert(Ptr->getOpcode() == ISD::ADD);
10228 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10232 isInc = (Ptr->getOpcode() == ISD::ADD);
10233 Offset = Ptr->getOperand(1);
10235 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
10236 // AddressingMode 2
10237 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10238 int RHSC = (int)RHS->getZExtValue();
10239 if (RHSC < 0 && RHSC > -0x1000) {
10240 assert(Ptr->getOpcode() == ISD::ADD);
10242 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10243 Base = Ptr->getOperand(0);
10248 if (Ptr->getOpcode() == ISD::ADD) {
10250 ARM_AM::ShiftOpc ShOpcVal=
10251 ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
10252 if (ShOpcVal != ARM_AM::no_shift) {
10253 Base = Ptr->getOperand(1);
10254 Offset = Ptr->getOperand(0);
10256 Base = Ptr->getOperand(0);
10257 Offset = Ptr->getOperand(1);
10262 isInc = (Ptr->getOpcode() == ISD::ADD);
10263 Base = Ptr->getOperand(0);
10264 Offset = Ptr->getOperand(1);
10268 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
10272 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
10273 bool isSEXTLoad, SDValue &Base,
10274 SDValue &Offset, bool &isInc,
10275 SelectionDAG &DAG) {
10276 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
10279 Base = Ptr->getOperand(0);
10280 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
10281 int RHSC = (int)RHS->getZExtValue();
10282 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
10283 assert(Ptr->getOpcode() == ISD::ADD);
10285 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
10287 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
10288 isInc = Ptr->getOpcode() == ISD::ADD;
10289 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
10297 /// getPreIndexedAddressParts - returns true by value, base pointer and
10298 /// offset pointer and addressing mode by reference if the node's address
10299 /// can be legally represented as pre-indexed load / store address.
10301 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
10303 ISD::MemIndexedMode &AM,
10304 SelectionDAG &DAG) const {
10305 if (Subtarget->isThumb1Only())
10310 bool isSEXTLoad = false;
10311 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10312 Ptr = LD->getBasePtr();
10313 VT = LD->getMemoryVT();
10314 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10315 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10316 Ptr = ST->getBasePtr();
10317 VT = ST->getMemoryVT();
10322 bool isLegal = false;
10323 if (Subtarget->isThumb2())
10324 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10325 Offset, isInc, DAG);
10327 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
10328 Offset, isInc, DAG);
10332 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
10336 /// getPostIndexedAddressParts - returns true by value, base pointer and
10337 /// offset pointer and addressing mode by reference if this node can be
10338 /// combined with a load / store to form a post-indexed load / store.
10339 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
10342 ISD::MemIndexedMode &AM,
10343 SelectionDAG &DAG) const {
10344 if (Subtarget->isThumb1Only())
10349 bool isSEXTLoad = false;
10350 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
10351 VT = LD->getMemoryVT();
10352 Ptr = LD->getBasePtr();
10353 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
10354 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
10355 VT = ST->getMemoryVT();
10356 Ptr = ST->getBasePtr();
10361 bool isLegal = false;
10362 if (Subtarget->isThumb2())
10363 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10366 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
10372 // Swap base ptr and offset to catch more post-index load / store when
10373 // it's legal. In Thumb2 mode, offset must be an immediate.
10374 if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
10375 !Subtarget->isThumb2())
10376 std::swap(Base, Offset);
10378 // Post-indexed load / store update the base pointer.
10383 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
10387 void ARMTargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
10390 const SelectionDAG &DAG,
10391 unsigned Depth) const {
10392 unsigned BitWidth = KnownOne.getBitWidth();
10393 KnownZero = KnownOne = APInt(BitWidth, 0);
10394 switch (Op.getOpcode()) {
10400 // These nodes' second result is a boolean
10401 if (Op.getResNo() == 0)
10403 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
10405 case ARMISD::CMOV: {
10406 // Bits are known zero/one if known on the LHS and RHS.
10407 DAG.computeKnownBits(Op.getOperand(0), KnownZero, KnownOne, Depth+1);
10408 if (KnownZero == 0 && KnownOne == 0) return;
10410 APInt KnownZeroRHS, KnownOneRHS;
10411 DAG.computeKnownBits(Op.getOperand(1), KnownZeroRHS, KnownOneRHS, Depth+1);
10412 KnownZero &= KnownZeroRHS;
10413 KnownOne &= KnownOneRHS;
10416 case ISD::INTRINSIC_W_CHAIN: {
10417 ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
10418 Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
10421 case Intrinsic::arm_ldaex:
10422 case Intrinsic::arm_ldrex: {
10423 EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
10424 unsigned MemBits = VT.getScalarType().getSizeInBits();
10425 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
10433 //===----------------------------------------------------------------------===//
10434 // ARM Inline Assembly Support
10435 //===----------------------------------------------------------------------===//
10437 bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
10438 // Looking for "rev" which is V6+.
10439 if (!Subtarget->hasV6Ops())
10442 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10443 std::string AsmStr = IA->getAsmString();
10444 SmallVector<StringRef, 4> AsmPieces;
10445 SplitString(AsmStr, AsmPieces, ";\n");
10447 switch (AsmPieces.size()) {
10448 default: return false;
10450 AsmStr = AsmPieces[0];
10452 SplitString(AsmStr, AsmPieces, " \t,");
10455 if (AsmPieces.size() == 3 &&
10456 AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
10457 IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
10458 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10459 if (Ty && Ty->getBitWidth() == 32)
10460 return IntrinsicLowering::LowerToByteSwap(CI);
10468 /// getConstraintType - Given a constraint letter, return the type of
10469 /// constraint it is for this target.
10470 ARMTargetLowering::ConstraintType
10471 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
10472 if (Constraint.size() == 1) {
10473 switch (Constraint[0]) {
10475 case 'l': return C_RegisterClass;
10476 case 'w': return C_RegisterClass;
10477 case 'h': return C_RegisterClass;
10478 case 'x': return C_RegisterClass;
10479 case 't': return C_RegisterClass;
10480 case 'j': return C_Other; // Constant for movw.
10481 // An address with a single base register. Due to the way we
10482 // currently handle addresses it is the same as an 'r' memory constraint.
10483 case 'Q': return C_Memory;
10485 } else if (Constraint.size() == 2) {
10486 switch (Constraint[0]) {
10488 // All 'U+' constraints are addresses.
10489 case 'U': return C_Memory;
10492 return TargetLowering::getConstraintType(Constraint);
10495 /// Examine constraint type and operand type and determine a weight value.
10496 /// This object must already have been set up with the operand type
10497 /// and the current alternative constraint selected.
10498 TargetLowering::ConstraintWeight
10499 ARMTargetLowering::getSingleConstraintMatchWeight(
10500 AsmOperandInfo &info, const char *constraint) const {
10501 ConstraintWeight weight = CW_Invalid;
10502 Value *CallOperandVal = info.CallOperandVal;
10503 // If we don't have a value, we can't do a match,
10504 // but allow it at the lowest weight.
10505 if (!CallOperandVal)
10507 Type *type = CallOperandVal->getType();
10508 // Look at the constraint type.
10509 switch (*constraint) {
10511 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
10514 if (type->isIntegerTy()) {
10515 if (Subtarget->isThumb())
10516 weight = CW_SpecificReg;
10518 weight = CW_Register;
10522 if (type->isFloatingPointTy())
10523 weight = CW_Register;
10529 typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
10531 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10533 if (Constraint.size() == 1) {
10534 // GCC ARM Constraint Letters
10535 switch (Constraint[0]) {
10536 case 'l': // Low regs or general regs.
10537 if (Subtarget->isThumb())
10538 return RCPair(0U, &ARM::tGPRRegClass);
10539 return RCPair(0U, &ARM::GPRRegClass);
10540 case 'h': // High regs or no regs.
10541 if (Subtarget->isThumb())
10542 return RCPair(0U, &ARM::hGPRRegClass);
10545 return RCPair(0U, &ARM::GPRRegClass);
10547 if (VT == MVT::Other)
10549 if (VT == MVT::f32)
10550 return RCPair(0U, &ARM::SPRRegClass);
10551 if (VT.getSizeInBits() == 64)
10552 return RCPair(0U, &ARM::DPRRegClass);
10553 if (VT.getSizeInBits() == 128)
10554 return RCPair(0U, &ARM::QPRRegClass);
10557 if (VT == MVT::Other)
10559 if (VT == MVT::f32)
10560 return RCPair(0U, &ARM::SPR_8RegClass);
10561 if (VT.getSizeInBits() == 64)
10562 return RCPair(0U, &ARM::DPR_8RegClass);
10563 if (VT.getSizeInBits() == 128)
10564 return RCPair(0U, &ARM::QPR_8RegClass);
10567 if (VT == MVT::f32)
10568 return RCPair(0U, &ARM::SPRRegClass);
10572 if (StringRef("{cc}").equals_lower(Constraint))
10573 return std::make_pair(unsigned(ARM::CPSR), &ARM::CCRRegClass);
10575 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10578 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10579 /// vector. If it is invalid, don't add anything to Ops.
10580 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10581 std::string &Constraint,
10582 std::vector<SDValue>&Ops,
10583 SelectionDAG &DAG) const {
10586 // Currently only support length 1 constraints.
10587 if (Constraint.length() != 1) return;
10589 char ConstraintLetter = Constraint[0];
10590 switch (ConstraintLetter) {
10593 case 'I': case 'J': case 'K': case 'L':
10594 case 'M': case 'N': case 'O':
10595 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
10599 int64_t CVal64 = C->getSExtValue();
10600 int CVal = (int) CVal64;
10601 // None of these constraints allow values larger than 32 bits. Check
10602 // that the value fits in an int.
10603 if (CVal != CVal64)
10606 switch (ConstraintLetter) {
10608 // Constant suitable for movw, must be between 0 and
10610 if (Subtarget->hasV6T2Ops())
10611 if (CVal >= 0 && CVal <= 65535)
10615 if (Subtarget->isThumb1Only()) {
10616 // This must be a constant between 0 and 255, for ADD
10618 if (CVal >= 0 && CVal <= 255)
10620 } else if (Subtarget->isThumb2()) {
10621 // A constant that can be used as an immediate value in a
10622 // data-processing instruction.
10623 if (ARM_AM::getT2SOImmVal(CVal) != -1)
10626 // A constant that can be used as an immediate value in a
10627 // data-processing instruction.
10628 if (ARM_AM::getSOImmVal(CVal) != -1)
10634 if (Subtarget->isThumb()) { // FIXME thumb2
10635 // This must be a constant between -255 and -1, for negated ADD
10636 // immediates. This can be used in GCC with an "n" modifier that
10637 // prints the negated value, for use with SUB instructions. It is
10638 // not useful otherwise but is implemented for compatibility.
10639 if (CVal >= -255 && CVal <= -1)
10642 // This must be a constant between -4095 and 4095. It is not clear
10643 // what this constraint is intended for. Implemented for
10644 // compatibility with GCC.
10645 if (CVal >= -4095 && CVal <= 4095)
10651 if (Subtarget->isThumb1Only()) {
10652 // A 32-bit value where only one byte has a nonzero value. Exclude
10653 // zero to match GCC. This constraint is used by GCC internally for
10654 // constants that can be loaded with a move/shift combination.
10655 // It is not useful otherwise but is implemented for compatibility.
10656 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
10658 } else if (Subtarget->isThumb2()) {
10659 // A constant whose bitwise inverse can be used as an immediate
10660 // value in a data-processing instruction. This can be used in GCC
10661 // with a "B" modifier that prints the inverted value, for use with
10662 // BIC and MVN instructions. It is not useful otherwise but is
10663 // implemented for compatibility.
10664 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
10667 // A constant whose bitwise inverse can be used as an immediate
10668 // value in a data-processing instruction. This can be used in GCC
10669 // with a "B" modifier that prints the inverted value, for use with
10670 // BIC and MVN instructions. It is not useful otherwise but is
10671 // implemented for compatibility.
10672 if (ARM_AM::getSOImmVal(~CVal) != -1)
10678 if (Subtarget->isThumb1Only()) {
10679 // This must be a constant between -7 and 7,
10680 // for 3-operand ADD/SUB immediate instructions.
10681 if (CVal >= -7 && CVal < 7)
10683 } else if (Subtarget->isThumb2()) {
10684 // A constant whose negation can be used as an immediate value in a
10685 // data-processing instruction. This can be used in GCC with an "n"
10686 // modifier that prints the negated value, for use with SUB
10687 // instructions. It is not useful otherwise but is implemented for
10689 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
10692 // A constant whose negation can be used as an immediate value in a
10693 // data-processing instruction. This can be used in GCC with an "n"
10694 // modifier that prints the negated value, for use with SUB
10695 // instructions. It is not useful otherwise but is implemented for
10697 if (ARM_AM::getSOImmVal(-CVal) != -1)
10703 if (Subtarget->isThumb()) { // FIXME thumb2
10704 // This must be a multiple of 4 between 0 and 1020, for
10705 // ADD sp + immediate.
10706 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
10709 // A power of two or a constant between 0 and 32. This is used in
10710 // GCC for the shift amount on shifted register operands, but it is
10711 // useful in general for any shift amounts.
10712 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
10718 if (Subtarget->isThumb()) { // FIXME thumb2
10719 // This must be a constant between 0 and 31, for shift amounts.
10720 if (CVal >= 0 && CVal <= 31)
10726 if (Subtarget->isThumb()) { // FIXME thumb2
10727 // This must be a multiple of 4 between -508 and 508, for
10728 // ADD/SUB sp = sp + immediate.
10729 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
10734 Result = DAG.getTargetConstant(CVal, Op.getValueType());
10738 if (Result.getNode()) {
10739 Ops.push_back(Result);
10742 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10745 SDValue ARMTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
10746 assert(Subtarget->isTargetAEABI() && "Register-based DivRem lowering only");
10747 unsigned Opcode = Op->getOpcode();
10748 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
10749 "Invalid opcode for Div/Rem lowering");
10750 bool isSigned = (Opcode == ISD::SDIVREM);
10751 EVT VT = Op->getValueType(0);
10752 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
10755 switch (VT.getSimpleVT().SimpleTy) {
10756 default: llvm_unreachable("Unexpected request for libcall!");
10757 case MVT::i8: LC = isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
10758 case MVT::i16: LC = isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
10759 case MVT::i32: LC = isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
10760 case MVT::i64: LC = isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
10763 SDValue InChain = DAG.getEntryNode();
10765 TargetLowering::ArgListTy Args;
10766 TargetLowering::ArgListEntry Entry;
10767 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
10768 EVT ArgVT = Op->getOperand(i).getValueType();
10769 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
10770 Entry.Node = Op->getOperand(i);
10772 Entry.isSExt = isSigned;
10773 Entry.isZExt = !isSigned;
10774 Args.push_back(Entry);
10777 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
10780 Type *RetTy = (Type*)StructType::get(Ty, Ty, NULL);
10783 TargetLowering::CallLoweringInfo CLI(DAG);
10784 CLI.setDebugLoc(dl).setChain(InChain)
10785 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
10786 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
10788 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
10789 return CallInfo.first;
10793 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const {
10794 assert(Subtarget->isTargetWindows() && "unsupported target platform");
10798 SDValue Chain = Op.getOperand(0);
10799 SDValue Size = Op.getOperand(1);
10801 SDValue Words = DAG.getNode(ISD::SRL, DL, MVT::i32, Size,
10802 DAG.getConstant(2, MVT::i32));
10805 Chain = DAG.getCopyToReg(Chain, DL, ARM::R4, Words, Flag);
10806 Flag = Chain.getValue(1);
10808 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10809 Chain = DAG.getNode(ARMISD::WIN__CHKSTK, DL, NodeTys, Chain, Flag);
10811 SDValue NewSP = DAG.getCopyFromReg(Chain, DL, ARM::SP, MVT::i32);
10812 Chain = NewSP.getValue(1);
10814 SDValue Ops[2] = { NewSP, Chain };
10815 return DAG.getMergeValues(Ops, DL);
10818 SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
10819 assert(Op.getValueType() == MVT::f64 && Subtarget->isFPOnlySP() &&
10820 "Unexpected type for custom-lowering FP_EXTEND");
10823 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
10825 SDValue SrcVal = Op.getOperand(0);
10826 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10827 /*isSigned*/ false, SDLoc(Op)).first;
10830 SDValue ARMTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
10831 assert(Op.getOperand(0).getValueType() == MVT::f64 &&
10832 Subtarget->isFPOnlySP() &&
10833 "Unexpected type for custom-lowering FP_ROUND");
10836 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
10838 SDValue SrcVal = Op.getOperand(0);
10839 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
10840 /*isSigned*/ false, SDLoc(Op)).first;
10844 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
10845 // The ARM target isn't yet aware of offsets.
10849 bool ARM::isBitFieldInvertedMask(unsigned v) {
10850 if (v == 0xffffffff)
10853 // there can be 1's on either or both "outsides", all the "inside"
10854 // bits must be 0's
10855 unsigned TO = CountTrailingOnes_32(v);
10856 unsigned LO = CountLeadingOnes_32(v);
10857 v = (v >> TO) << TO;
10858 v = (v << LO) >> LO;
10862 /// isFPImmLegal - Returns true if the target can instruction select the
10863 /// specified FP immediate natively. If false, the legalizer will
10864 /// materialize the FP immediate as a load from a constant pool.
10865 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
10866 if (!Subtarget->hasVFP3())
10868 if (VT == MVT::f32)
10869 return ARM_AM::getFP32Imm(Imm) != -1;
10870 if (VT == MVT::f64 && !Subtarget->isFPOnlySP())
10871 return ARM_AM::getFP64Imm(Imm) != -1;
10875 /// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
10876 /// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
10877 /// specified in the intrinsic calls.
10878 bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
10880 unsigned Intrinsic) const {
10881 switch (Intrinsic) {
10882 case Intrinsic::arm_neon_vld1:
10883 case Intrinsic::arm_neon_vld2:
10884 case Intrinsic::arm_neon_vld3:
10885 case Intrinsic::arm_neon_vld4:
10886 case Intrinsic::arm_neon_vld2lane:
10887 case Intrinsic::arm_neon_vld3lane:
10888 case Intrinsic::arm_neon_vld4lane: {
10889 Info.opc = ISD::INTRINSIC_W_CHAIN;
10890 // Conservatively set memVT to the entire set of vectors loaded.
10891 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
10892 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10893 Info.ptrVal = I.getArgOperand(0);
10895 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10896 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10897 Info.vol = false; // volatile loads with NEON intrinsics not supported
10898 Info.readMem = true;
10899 Info.writeMem = false;
10902 case Intrinsic::arm_neon_vst1:
10903 case Intrinsic::arm_neon_vst2:
10904 case Intrinsic::arm_neon_vst3:
10905 case Intrinsic::arm_neon_vst4:
10906 case Intrinsic::arm_neon_vst2lane:
10907 case Intrinsic::arm_neon_vst3lane:
10908 case Intrinsic::arm_neon_vst4lane: {
10909 Info.opc = ISD::INTRINSIC_VOID;
10910 // Conservatively set memVT to the entire set of vectors stored.
10911 unsigned NumElts = 0;
10912 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
10913 Type *ArgTy = I.getArgOperand(ArgI)->getType();
10914 if (!ArgTy->isVectorTy())
10916 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
10918 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
10919 Info.ptrVal = I.getArgOperand(0);
10921 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
10922 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
10923 Info.vol = false; // volatile stores with NEON intrinsics not supported
10924 Info.readMem = false;
10925 Info.writeMem = true;
10928 case Intrinsic::arm_ldaex:
10929 case Intrinsic::arm_ldrex: {
10930 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
10931 Info.opc = ISD::INTRINSIC_W_CHAIN;
10932 Info.memVT = MVT::getVT(PtrTy->getElementType());
10933 Info.ptrVal = I.getArgOperand(0);
10935 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10937 Info.readMem = true;
10938 Info.writeMem = false;
10941 case Intrinsic::arm_stlex:
10942 case Intrinsic::arm_strex: {
10943 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(1)->getType());
10944 Info.opc = ISD::INTRINSIC_W_CHAIN;
10945 Info.memVT = MVT::getVT(PtrTy->getElementType());
10946 Info.ptrVal = I.getArgOperand(1);
10948 Info.align = getDataLayout()->getABITypeAlignment(PtrTy->getElementType());
10950 Info.readMem = false;
10951 Info.writeMem = true;
10954 case Intrinsic::arm_stlexd:
10955 case Intrinsic::arm_strexd: {
10956 Info.opc = ISD::INTRINSIC_W_CHAIN;
10957 Info.memVT = MVT::i64;
10958 Info.ptrVal = I.getArgOperand(2);
10962 Info.readMem = false;
10963 Info.writeMem = true;
10966 case Intrinsic::arm_ldaexd:
10967 case Intrinsic::arm_ldrexd: {
10968 Info.opc = ISD::INTRINSIC_W_CHAIN;
10969 Info.memVT = MVT::i64;
10970 Info.ptrVal = I.getArgOperand(0);
10974 Info.readMem = true;
10975 Info.writeMem = false;
10985 /// \brief Returns true if it is beneficial to convert a load of a constant
10986 /// to just the constant itself.
10987 bool ARMTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
10989 assert(Ty->isIntegerTy());
10991 unsigned Bits = Ty->getPrimitiveSizeInBits();
10992 if (Bits == 0 || Bits > 32)
10997 bool ARMTargetLowering::shouldExpandAtomicInIR(Instruction *Inst) const {
10998 // Loads and stores less than 64-bits are already atomic; ones above that
10999 // are doomed anyway, so defer to the default libcall and blame the OS when
11000 // things go wrong. Cortex M doesn't have ldrexd/strexd though, so don't emit
11001 // anything for those.
11002 bool IsMClass = Subtarget->isMClass();
11003 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
11004 unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
11005 return Size == 64 && !IsMClass;
11006 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
11007 return LI->getType()->getPrimitiveSizeInBits() == 64 && !IsMClass;
11010 // For the real atomic operations, we have ldrex/strex up to 32 bits,
11011 // and up to 64 bits on the non-M profiles
11012 unsigned AtomicLimit = IsMClass ? 32 : 64;
11013 return Inst->getType()->getPrimitiveSizeInBits() <= AtomicLimit;
11016 // This has so far only been implemented for MachO.
11017 bool ARMTargetLowering::useLoadStackGuardNode() const {
11018 return Subtarget->getTargetTriple().getObjectFormat() == Triple::MachO;
11021 Value *ARMTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
11022 AtomicOrdering Ord) const {
11023 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11024 Type *ValTy = cast<PointerType>(Addr->getType())->getElementType();
11025 bool IsAcquire = isAtLeastAcquire(Ord);
11027 // Since i64 isn't legal and intrinsics don't get type-lowered, the ldrexd
11028 // intrinsic must return {i32, i32} and we have to recombine them into a
11029 // single i64 here.
11030 if (ValTy->getPrimitiveSizeInBits() == 64) {
11031 Intrinsic::ID Int =
11032 IsAcquire ? Intrinsic::arm_ldaexd : Intrinsic::arm_ldrexd;
11033 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int);
11035 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11036 Value *LoHi = Builder.CreateCall(Ldrex, Addr, "lohi");
11038 Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
11039 Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");
11040 if (!Subtarget->isLittle())
11041 std::swap (Lo, Hi);
11042 Lo = Builder.CreateZExt(Lo, ValTy, "lo64");
11043 Hi = Builder.CreateZExt(Hi, ValTy, "hi64");
11044 return Builder.CreateOr(
11045 Lo, Builder.CreateShl(Hi, ConstantInt::get(ValTy, 32)), "val64");
11048 Type *Tys[] = { Addr->getType() };
11049 Intrinsic::ID Int = IsAcquire ? Intrinsic::arm_ldaex : Intrinsic::arm_ldrex;
11050 Function *Ldrex = llvm::Intrinsic::getDeclaration(M, Int, Tys);
11052 return Builder.CreateTruncOrBitCast(
11053 Builder.CreateCall(Ldrex, Addr),
11054 cast<PointerType>(Addr->getType())->getElementType());
11057 Value *ARMTargetLowering::emitStoreConditional(IRBuilder<> &Builder, Value *Val,
11059 AtomicOrdering Ord) const {
11060 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
11061 bool IsRelease = isAtLeastRelease(Ord);
11063 // Since the intrinsics must have legal type, the i64 intrinsics take two
11064 // parameters: "i32, i32". We must marshal Val into the appropriate form
11065 // before the call.
11066 if (Val->getType()->getPrimitiveSizeInBits() == 64) {
11067 Intrinsic::ID Int =
11068 IsRelease ? Intrinsic::arm_stlexd : Intrinsic::arm_strexd;
11069 Function *Strex = Intrinsic::getDeclaration(M, Int);
11070 Type *Int32Ty = Type::getInt32Ty(M->getContext());
11072 Value *Lo = Builder.CreateTrunc(Val, Int32Ty, "lo");
11073 Value *Hi = Builder.CreateTrunc(Builder.CreateLShr(Val, 32), Int32Ty, "hi");
11074 if (!Subtarget->isLittle())
11075 std::swap (Lo, Hi);
11076 Addr = Builder.CreateBitCast(Addr, Type::getInt8PtrTy(M->getContext()));
11077 return Builder.CreateCall3(Strex, Lo, Hi, Addr);
11080 Intrinsic::ID Int = IsRelease ? Intrinsic::arm_stlex : Intrinsic::arm_strex;
11081 Type *Tys[] = { Addr->getType() };
11082 Function *Strex = Intrinsic::getDeclaration(M, Int, Tys);
11084 return Builder.CreateCall2(
11085 Strex, Builder.CreateZExtOrBitCast(
11086 Val, Strex->getFunctionType()->getParamType(0)),
11098 static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
11099 uint64_t &Members) {
11100 if (const StructType *ST = dyn_cast<StructType>(Ty)) {
11101 for (unsigned i = 0; i < ST->getNumElements(); ++i) {
11102 uint64_t SubMembers = 0;
11103 if (!isHomogeneousAggregate(ST->getElementType(i), Base, SubMembers))
11105 Members += SubMembers;
11107 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
11108 uint64_t SubMembers = 0;
11109 if (!isHomogeneousAggregate(AT->getElementType(), Base, SubMembers))
11111 Members += SubMembers * AT->getNumElements();
11112 } else if (Ty->isFloatTy()) {
11113 if (Base != HA_UNKNOWN && Base != HA_FLOAT)
11117 } else if (Ty->isDoubleTy()) {
11118 if (Base != HA_UNKNOWN && Base != HA_DOUBLE)
11122 } else if (const VectorType *VT = dyn_cast<VectorType>(Ty)) {
11129 return VT->getBitWidth() == 64;
11131 return VT->getBitWidth() == 128;
11133 switch (VT->getBitWidth()) {
11146 return (Members > 0 && Members <= 4);
11149 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
11150 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
11151 Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
11152 if (getEffectiveCallingConv(CallConv, isVarArg) !=
11153 CallingConv::ARM_AAPCS_VFP)
11156 HABaseType Base = HA_UNKNOWN;
11157 uint64_t Members = 0;
11158 bool result = isHomogeneousAggregate(Ty, Base, Members);
11159 DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump());