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 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instruction.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/CodeGen/CallingConvLower.h"
32 #include "llvm/CodeGen/MachineBasicBlock.h"
33 #include "llvm/CodeGen/MachineFrameInfo.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/CodeGen/SelectionDAG.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/ADT/VectorExtras.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/MathExtras.h"
46 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
47 CCValAssign::LocInfo &LocInfo,
48 ISD::ArgFlagsTy &ArgFlags,
50 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
51 CCValAssign::LocInfo &LocInfo,
52 ISD::ArgFlagsTy &ArgFlags,
54 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
55 CCValAssign::LocInfo &LocInfo,
56 ISD::ArgFlagsTy &ArgFlags,
58 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
59 CCValAssign::LocInfo &LocInfo,
60 ISD::ArgFlagsTy &ArgFlags,
63 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
64 EVT PromotedBitwiseVT) {
65 if (VT != PromotedLdStVT) {
66 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
67 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
68 PromotedLdStVT.getSimpleVT());
70 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
71 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
72 PromotedLdStVT.getSimpleVT());
75 EVT ElemTy = VT.getVectorElementType();
76 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
77 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
78 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
79 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
80 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
81 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
82 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
83 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
85 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
86 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
87 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
90 // Promote all bit-wise operations.
91 if (VT.isInteger() && VT != PromotedBitwiseVT) {
92 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
93 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
94 PromotedBitwiseVT.getSimpleVT());
95 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
96 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
97 PromotedBitwiseVT.getSimpleVT());
98 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
99 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
100 PromotedBitwiseVT.getSimpleVT());
103 // Neon does not support vector divide/remainder operations.
104 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
105 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
106 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
107 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
108 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
109 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
112 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
113 addRegisterClass(VT, ARM::DPRRegisterClass);
114 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
117 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
118 addRegisterClass(VT, ARM::QPRRegisterClass);
119 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
122 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
123 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
124 return new TargetLoweringObjectFileMachO();
125 return new ARMElfTargetObjectFile();
128 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
129 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
130 Subtarget = &TM.getSubtarget<ARMSubtarget>();
132 if (Subtarget->isTargetDarwin()) {
133 // Uses VFP for Thumb libfuncs if available.
134 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
135 // Single-precision floating-point arithmetic.
136 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
137 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
138 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
139 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
141 // Double-precision floating-point arithmetic.
142 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
143 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
144 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
145 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
147 // Single-precision comparisons.
148 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
149 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
150 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
151 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
152 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
153 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
154 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
155 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
157 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
158 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
159 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
160 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
161 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
162 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
163 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
164 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
166 // Double-precision comparisons.
167 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
168 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
169 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
170 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
171 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
172 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
173 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
174 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
176 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
177 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
178 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
179 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
180 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
181 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
182 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
183 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
185 // Floating-point to integer conversions.
186 // i64 conversions are done via library routines even when generating VFP
187 // instructions, so use the same ones.
188 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
189 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
190 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
191 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
193 // Conversions between floating types.
194 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
195 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
197 // Integer to floating-point conversions.
198 // i64 conversions are done via library routines even when generating VFP
199 // instructions, so use the same ones.
200 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
201 // e.g., __floatunsidf vs. __floatunssidfvfp.
202 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
203 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
204 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
205 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
209 // These libcalls are not available in 32-bit.
210 setLibcallName(RTLIB::SHL_I128, 0);
211 setLibcallName(RTLIB::SRL_I128, 0);
212 setLibcallName(RTLIB::SRA_I128, 0);
214 // Libcalls should use the AAPCS base standard ABI, even if hard float
215 // is in effect, as per the ARM RTABI specification, section 4.1.2.
216 if (Subtarget->isAAPCS_ABI()) {
217 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
218 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
219 CallingConv::ARM_AAPCS);
223 if (Subtarget->isThumb1Only())
224 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
226 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
227 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
228 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
229 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
231 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
234 if (Subtarget->hasNEON()) {
235 addDRTypeForNEON(MVT::v2f32);
236 addDRTypeForNEON(MVT::v8i8);
237 addDRTypeForNEON(MVT::v4i16);
238 addDRTypeForNEON(MVT::v2i32);
239 addDRTypeForNEON(MVT::v1i64);
241 addQRTypeForNEON(MVT::v4f32);
242 addQRTypeForNEON(MVT::v2f64);
243 addQRTypeForNEON(MVT::v16i8);
244 addQRTypeForNEON(MVT::v8i16);
245 addQRTypeForNEON(MVT::v4i32);
246 addQRTypeForNEON(MVT::v2i64);
248 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
249 // neither Neon nor VFP support any arithmetic operations on it.
250 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
251 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
252 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
253 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
254 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
255 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
256 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
257 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
258 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
259 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
260 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
261 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
262 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
263 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
264 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
265 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
266 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
267 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
268 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
269 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
270 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
271 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
272 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
273 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
275 // Neon does not support some operations on v1i64 and v2i64 types.
276 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
277 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
278 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
279 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
281 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
282 setTargetDAGCombine(ISD::SHL);
283 setTargetDAGCombine(ISD::SRL);
284 setTargetDAGCombine(ISD::SRA);
285 setTargetDAGCombine(ISD::SIGN_EXTEND);
286 setTargetDAGCombine(ISD::ZERO_EXTEND);
287 setTargetDAGCombine(ISD::ANY_EXTEND);
290 computeRegisterProperties();
292 // ARM does not have f32 extending load.
293 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
295 // ARM does not have i1 sign extending load.
296 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
298 // ARM supports all 4 flavors of integer indexed load / store.
299 if (!Subtarget->isThumb1Only()) {
300 for (unsigned im = (unsigned)ISD::PRE_INC;
301 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
302 setIndexedLoadAction(im, MVT::i1, Legal);
303 setIndexedLoadAction(im, MVT::i8, Legal);
304 setIndexedLoadAction(im, MVT::i16, Legal);
305 setIndexedLoadAction(im, MVT::i32, Legal);
306 setIndexedStoreAction(im, MVT::i1, Legal);
307 setIndexedStoreAction(im, MVT::i8, Legal);
308 setIndexedStoreAction(im, MVT::i16, Legal);
309 setIndexedStoreAction(im, MVT::i32, Legal);
313 // i64 operation support.
314 if (Subtarget->isThumb1Only()) {
315 setOperationAction(ISD::MUL, MVT::i64, Expand);
316 setOperationAction(ISD::MULHU, MVT::i32, Expand);
317 setOperationAction(ISD::MULHS, MVT::i32, Expand);
318 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
319 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
321 setOperationAction(ISD::MUL, MVT::i64, Expand);
322 setOperationAction(ISD::MULHU, MVT::i32, Expand);
323 if (!Subtarget->hasV6Ops())
324 setOperationAction(ISD::MULHS, MVT::i32, Expand);
326 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
327 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
328 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
329 setOperationAction(ISD::SRL, MVT::i64, Custom);
330 setOperationAction(ISD::SRA, MVT::i64, Custom);
332 // ARM does not have ROTL.
333 setOperationAction(ISD::ROTL, MVT::i32, Expand);
334 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
335 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
336 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
337 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
339 // Only ARMv6 has BSWAP.
340 if (!Subtarget->hasV6Ops())
341 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
343 // These are expanded into libcalls.
344 setOperationAction(ISD::SDIV, MVT::i32, Expand);
345 setOperationAction(ISD::UDIV, MVT::i32, Expand);
346 setOperationAction(ISD::SREM, MVT::i32, Expand);
347 setOperationAction(ISD::UREM, MVT::i32, Expand);
348 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
349 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
351 // Support label based line numbers.
352 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
353 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
355 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
356 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
357 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
358 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
360 // Use the default implementation.
361 setOperationAction(ISD::VASTART, MVT::Other, Custom);
362 setOperationAction(ISD::VAARG, MVT::Other, Expand);
363 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
364 setOperationAction(ISD::VAEND, MVT::Other, Expand);
365 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
366 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
367 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
368 // FIXME: Shouldn't need this, since no register is used, but the legalizer
369 // doesn't yet know how to not do that for SjLj.
370 setExceptionSelectorRegister(ARM::R0);
371 if (Subtarget->isThumb())
372 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
374 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
375 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
377 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
378 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
379 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
381 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
383 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
384 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
385 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
387 // We want to custom lower some of our intrinsics.
388 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
389 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
390 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
392 setOperationAction(ISD::SETCC, MVT::i32, Expand);
393 setOperationAction(ISD::SETCC, MVT::f32, Expand);
394 setOperationAction(ISD::SETCC, MVT::f64, Expand);
395 setOperationAction(ISD::SELECT, MVT::i32, Expand);
396 setOperationAction(ISD::SELECT, MVT::f32, Expand);
397 setOperationAction(ISD::SELECT, MVT::f64, Expand);
398 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
399 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
400 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
402 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
403 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
404 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
405 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
406 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
408 // We don't support sin/cos/fmod/copysign/pow
409 setOperationAction(ISD::FSIN, MVT::f64, Expand);
410 setOperationAction(ISD::FSIN, MVT::f32, Expand);
411 setOperationAction(ISD::FCOS, MVT::f32, Expand);
412 setOperationAction(ISD::FCOS, MVT::f64, Expand);
413 setOperationAction(ISD::FREM, MVT::f64, Expand);
414 setOperationAction(ISD::FREM, MVT::f32, Expand);
415 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
416 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
417 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
419 setOperationAction(ISD::FPOW, MVT::f64, Expand);
420 setOperationAction(ISD::FPOW, MVT::f32, Expand);
422 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
423 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
424 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
425 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
426 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
427 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
430 // We have target-specific dag combine patterns for the following nodes:
431 // ARMISD::FMRRD - No need to call setTargetDAGCombine
432 setTargetDAGCombine(ISD::ADD);
433 setTargetDAGCombine(ISD::SUB);
435 setStackPointerRegisterToSaveRestore(ARM::SP);
436 setSchedulingPreference(SchedulingForRegPressure);
438 // FIXME: If-converter should use instruction latency to determine
439 // profitability rather than relying on fixed limits.
440 if (Subtarget->getCPUString() == "generic") {
441 // Generic (and overly aggressive) if-conversion limits.
442 setIfCvtBlockSizeLimit(10);
443 setIfCvtDupBlockSizeLimit(2);
444 } else if (Subtarget->hasV6Ops()) {
445 setIfCvtBlockSizeLimit(2);
446 setIfCvtDupBlockSizeLimit(1);
448 setIfCvtBlockSizeLimit(3);
449 setIfCvtDupBlockSizeLimit(2);
452 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
453 // Do not enable CodePlacementOpt for now: it currently runs after the
454 // ARMConstantIslandPass and messes up branch relaxation and placement
455 // of constant islands.
456 // benefitFromCodePlacementOpt = true;
459 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
462 case ARMISD::Wrapper: return "ARMISD::Wrapper";
463 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
464 case ARMISD::CALL: return "ARMISD::CALL";
465 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
466 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
467 case ARMISD::tCALL: return "ARMISD::tCALL";
468 case ARMISD::BRCOND: return "ARMISD::BRCOND";
469 case ARMISD::BR_JT: return "ARMISD::BR_JT";
470 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
471 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
472 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
473 case ARMISD::CMP: return "ARMISD::CMP";
474 case ARMISD::CMPZ: return "ARMISD::CMPZ";
475 case ARMISD::CMPFP: return "ARMISD::CMPFP";
476 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
477 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
478 case ARMISD::CMOV: return "ARMISD::CMOV";
479 case ARMISD::CNEG: return "ARMISD::CNEG";
481 case ARMISD::FTOSI: return "ARMISD::FTOSI";
482 case ARMISD::FTOUI: return "ARMISD::FTOUI";
483 case ARMISD::SITOF: return "ARMISD::SITOF";
484 case ARMISD::UITOF: return "ARMISD::UITOF";
486 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
487 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
488 case ARMISD::RRX: return "ARMISD::RRX";
490 case ARMISD::FMRRD: return "ARMISD::FMRRD";
491 case ARMISD::FMDRR: return "ARMISD::FMDRR";
493 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
495 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
497 case ARMISD::VCEQ: return "ARMISD::VCEQ";
498 case ARMISD::VCGE: return "ARMISD::VCGE";
499 case ARMISD::VCGEU: return "ARMISD::VCGEU";
500 case ARMISD::VCGT: return "ARMISD::VCGT";
501 case ARMISD::VCGTU: return "ARMISD::VCGTU";
502 case ARMISD::VTST: return "ARMISD::VTST";
504 case ARMISD::VSHL: return "ARMISD::VSHL";
505 case ARMISD::VSHRs: return "ARMISD::VSHRs";
506 case ARMISD::VSHRu: return "ARMISD::VSHRu";
507 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
508 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
509 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
510 case ARMISD::VSHRN: return "ARMISD::VSHRN";
511 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
512 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
513 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
514 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
515 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
516 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
517 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
518 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
519 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
520 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
521 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
522 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
523 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
524 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
525 case ARMISD::VDUP: return "ARMISD::VDUP";
526 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
527 case ARMISD::VEXT: return "ARMISD::VEXT";
528 case ARMISD::VREV64: return "ARMISD::VREV64";
529 case ARMISD::VREV32: return "ARMISD::VREV32";
530 case ARMISD::VREV16: return "ARMISD::VREV16";
531 case ARMISD::VZIP: return "ARMISD::VZIP";
532 case ARMISD::VUZP: return "ARMISD::VUZP";
533 case ARMISD::VTRN: return "ARMISD::VTRN";
537 /// getFunctionAlignment - Return the Log2 alignment of this function.
538 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
539 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
542 //===----------------------------------------------------------------------===//
544 //===----------------------------------------------------------------------===//
546 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
547 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
549 default: llvm_unreachable("Unknown condition code!");
550 case ISD::SETNE: return ARMCC::NE;
551 case ISD::SETEQ: return ARMCC::EQ;
552 case ISD::SETGT: return ARMCC::GT;
553 case ISD::SETGE: return ARMCC::GE;
554 case ISD::SETLT: return ARMCC::LT;
555 case ISD::SETLE: return ARMCC::LE;
556 case ISD::SETUGT: return ARMCC::HI;
557 case ISD::SETUGE: return ARMCC::HS;
558 case ISD::SETULT: return ARMCC::LO;
559 case ISD::SETULE: return ARMCC::LS;
563 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
564 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
565 ARMCC::CondCodes &CondCode2) {
566 CondCode2 = ARMCC::AL;
568 default: llvm_unreachable("Unknown FP condition!");
570 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
572 case ISD::SETOGT: CondCode = ARMCC::GT; break;
574 case ISD::SETOGE: CondCode = ARMCC::GE; break;
575 case ISD::SETOLT: CondCode = ARMCC::MI; break;
576 case ISD::SETOLE: CondCode = ARMCC::LS; break;
577 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
578 case ISD::SETO: CondCode = ARMCC::VC; break;
579 case ISD::SETUO: CondCode = ARMCC::VS; break;
580 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
581 case ISD::SETUGT: CondCode = ARMCC::HI; break;
582 case ISD::SETUGE: CondCode = ARMCC::PL; break;
584 case ISD::SETULT: CondCode = ARMCC::LT; break;
586 case ISD::SETULE: CondCode = ARMCC::LE; break;
588 case ISD::SETUNE: CondCode = ARMCC::NE; break;
592 //===----------------------------------------------------------------------===//
593 // Calling Convention Implementation
594 //===----------------------------------------------------------------------===//
596 #include "ARMGenCallingConv.inc"
598 // APCS f64 is in register pairs, possibly split to stack
599 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
600 CCValAssign::LocInfo &LocInfo,
601 CCState &State, bool CanFail) {
602 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
604 // Try to get the first register.
605 if (unsigned Reg = State.AllocateReg(RegList, 4))
606 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
608 // For the 2nd half of a v2f64, do not fail.
612 // Put the whole thing on the stack.
613 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
614 State.AllocateStack(8, 4),
619 // Try to get the second register.
620 if (unsigned Reg = State.AllocateReg(RegList, 4))
621 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
623 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
624 State.AllocateStack(4, 4),
629 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
630 CCValAssign::LocInfo &LocInfo,
631 ISD::ArgFlagsTy &ArgFlags,
633 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
635 if (LocVT == MVT::v2f64 &&
636 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
638 return true; // we handled it
641 // AAPCS f64 is in aligned register pairs
642 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
643 CCValAssign::LocInfo &LocInfo,
644 CCState &State, bool CanFail) {
645 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
646 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
648 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
650 // For the 2nd half of a v2f64, do not just fail.
654 // Put the whole thing on the stack.
655 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
656 State.AllocateStack(8, 8),
662 for (i = 0; i < 2; ++i)
663 if (HiRegList[i] == Reg)
666 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
667 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
672 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
673 CCValAssign::LocInfo &LocInfo,
674 ISD::ArgFlagsTy &ArgFlags,
676 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
678 if (LocVT == MVT::v2f64 &&
679 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
681 return true; // we handled it
684 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
685 CCValAssign::LocInfo &LocInfo, CCState &State) {
686 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
687 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
689 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
691 return false; // we didn't handle it
694 for (i = 0; i < 2; ++i)
695 if (HiRegList[i] == Reg)
698 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
699 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
704 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
705 CCValAssign::LocInfo &LocInfo,
706 ISD::ArgFlagsTy &ArgFlags,
708 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
710 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
712 return true; // we handled it
715 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
716 CCValAssign::LocInfo &LocInfo,
717 ISD::ArgFlagsTy &ArgFlags,
719 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
723 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
724 /// given CallingConvention value.
725 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
727 bool isVarArg) const {
730 llvm_unreachable("Unsupported calling convention");
732 case CallingConv::Fast:
733 // Use target triple & subtarget features to do actual dispatch.
734 if (Subtarget->isAAPCS_ABI()) {
735 if (Subtarget->hasVFP2() &&
736 FloatABIType == FloatABI::Hard && !isVarArg)
737 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
739 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
741 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
742 case CallingConv::ARM_AAPCS_VFP:
743 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
744 case CallingConv::ARM_AAPCS:
745 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
746 case CallingConv::ARM_APCS:
747 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
751 /// LowerCallResult - Lower the result values of a call into the
752 /// appropriate copies out of appropriate physical registers.
754 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
755 CallingConv::ID CallConv, bool isVarArg,
756 const SmallVectorImpl<ISD::InputArg> &Ins,
757 DebugLoc dl, SelectionDAG &DAG,
758 SmallVectorImpl<SDValue> &InVals) {
760 // Assign locations to each value returned by this call.
761 SmallVector<CCValAssign, 16> RVLocs;
762 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
763 RVLocs, *DAG.getContext());
764 CCInfo.AnalyzeCallResult(Ins,
765 CCAssignFnForNode(CallConv, /* Return*/ true,
768 // Copy all of the result registers out of their specified physreg.
769 for (unsigned i = 0; i != RVLocs.size(); ++i) {
770 CCValAssign VA = RVLocs[i];
773 if (VA.needsCustom()) {
774 // Handle f64 or half of a v2f64.
775 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
777 Chain = Lo.getValue(1);
778 InFlag = Lo.getValue(2);
779 VA = RVLocs[++i]; // skip ahead to next loc
780 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
782 Chain = Hi.getValue(1);
783 InFlag = Hi.getValue(2);
784 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
786 if (VA.getLocVT() == MVT::v2f64) {
787 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
788 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
789 DAG.getConstant(0, MVT::i32));
791 VA = RVLocs[++i]; // skip ahead to next loc
792 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
793 Chain = Lo.getValue(1);
794 InFlag = Lo.getValue(2);
795 VA = RVLocs[++i]; // skip ahead to next loc
796 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
797 Chain = Hi.getValue(1);
798 InFlag = Hi.getValue(2);
799 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
800 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
801 DAG.getConstant(1, MVT::i32));
804 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
806 Chain = Val.getValue(1);
807 InFlag = Val.getValue(2);
810 switch (VA.getLocInfo()) {
811 default: llvm_unreachable("Unknown loc info!");
812 case CCValAssign::Full: break;
813 case CCValAssign::BCvt:
814 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
818 InVals.push_back(Val);
824 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
825 /// by "Src" to address "Dst" of size "Size". Alignment information is
826 /// specified by the specific parameter attribute. The copy will be passed as
827 /// a byval function parameter.
828 /// Sometimes what we are copying is the end of a larger object, the part that
829 /// does not fit in registers.
831 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
832 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
834 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
835 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
836 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
839 /// LowerMemOpCallTo - Store the argument to the stack.
841 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
842 SDValue StackPtr, SDValue Arg,
843 DebugLoc dl, SelectionDAG &DAG,
844 const CCValAssign &VA,
845 ISD::ArgFlagsTy Flags) {
846 unsigned LocMemOffset = VA.getLocMemOffset();
847 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
848 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
849 if (Flags.isByVal()) {
850 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
852 return DAG.getStore(Chain, dl, Arg, PtrOff,
853 PseudoSourceValue::getStack(), LocMemOffset);
856 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
857 SDValue Chain, SDValue &Arg,
858 RegsToPassVector &RegsToPass,
859 CCValAssign &VA, CCValAssign &NextVA,
861 SmallVector<SDValue, 8> &MemOpChains,
862 ISD::ArgFlagsTy Flags) {
864 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
865 DAG.getVTList(MVT::i32, MVT::i32), Arg);
866 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
868 if (NextVA.isRegLoc())
869 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
871 assert(NextVA.isMemLoc());
872 if (StackPtr.getNode() == 0)
873 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
875 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
881 /// LowerCall - Lowering a call into a callseq_start <-
882 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
885 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
886 CallingConv::ID CallConv, bool isVarArg,
888 const SmallVectorImpl<ISD::OutputArg> &Outs,
889 const SmallVectorImpl<ISD::InputArg> &Ins,
890 DebugLoc dl, SelectionDAG &DAG,
891 SmallVectorImpl<SDValue> &InVals) {
893 // Analyze operands of the call, assigning locations to each operand.
894 SmallVector<CCValAssign, 16> ArgLocs;
895 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
897 CCInfo.AnalyzeCallOperands(Outs,
898 CCAssignFnForNode(CallConv, /* Return*/ false,
901 // Get a count of how many bytes are to be pushed on the stack.
902 unsigned NumBytes = CCInfo.getNextStackOffset();
904 // Adjust the stack pointer for the new arguments...
905 // These operations are automatically eliminated by the prolog/epilog pass
906 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
908 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
910 RegsToPassVector RegsToPass;
911 SmallVector<SDValue, 8> MemOpChains;
913 // Walk the register/memloc assignments, inserting copies/loads. In the case
914 // of tail call optimization, arguments are handled later.
915 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
918 CCValAssign &VA = ArgLocs[i];
919 SDValue Arg = Outs[realArgIdx].Val;
920 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
922 // Promote the value if needed.
923 switch (VA.getLocInfo()) {
924 default: llvm_unreachable("Unknown loc info!");
925 case CCValAssign::Full: break;
926 case CCValAssign::SExt:
927 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
929 case CCValAssign::ZExt:
930 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
932 case CCValAssign::AExt:
933 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
935 case CCValAssign::BCvt:
936 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
940 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
941 if (VA.needsCustom()) {
942 if (VA.getLocVT() == MVT::v2f64) {
943 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
944 DAG.getConstant(0, MVT::i32));
945 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
946 DAG.getConstant(1, MVT::i32));
948 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
949 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
951 VA = ArgLocs[++i]; // skip ahead to next loc
953 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
954 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
956 assert(VA.isMemLoc());
957 if (StackPtr.getNode() == 0)
958 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
960 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
961 dl, DAG, VA, Flags));
964 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
965 StackPtr, MemOpChains, Flags);
967 } else if (VA.isRegLoc()) {
968 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
970 assert(VA.isMemLoc());
971 if (StackPtr.getNode() == 0)
972 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
974 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
975 dl, DAG, VA, Flags));
979 if (!MemOpChains.empty())
980 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
981 &MemOpChains[0], MemOpChains.size());
983 // Build a sequence of copy-to-reg nodes chained together with token chain
984 // and flag operands which copy the outgoing args into the appropriate regs.
986 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
987 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
988 RegsToPass[i].second, InFlag);
989 InFlag = Chain.getValue(1);
992 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
993 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
994 // node so that legalize doesn't hack it.
995 bool isDirect = false;
996 bool isARMFunc = false;
997 bool isLocalARMFunc = false;
998 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
999 GlobalValue *GV = G->getGlobal();
1001 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1002 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1003 getTargetMachine().getRelocationModel() != Reloc::Static;
1004 isARMFunc = !Subtarget->isThumb() || isStub;
1005 // ARM call to a local ARM function is predicable.
1006 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1007 // tBX takes a register source operand.
1008 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1009 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1012 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1013 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1014 Callee = DAG.getLoad(getPointerTy(), dl,
1015 DAG.getEntryNode(), CPAddr, NULL, 0);
1016 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1017 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1018 getPointerTy(), Callee, PICLabel);
1020 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1021 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1023 bool isStub = Subtarget->isTargetDarwin() &&
1024 getTargetMachine().getRelocationModel() != Reloc::Static;
1025 isARMFunc = !Subtarget->isThumb() || isStub;
1026 // tBX takes a register source operand.
1027 const char *Sym = S->getSymbol();
1028 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1029 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1030 Sym, ARMPCLabelIndex, 4);
1031 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1032 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1033 Callee = DAG.getLoad(getPointerTy(), dl,
1034 DAG.getEntryNode(), CPAddr, NULL, 0);
1035 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1036 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1037 getPointerTy(), Callee, PICLabel);
1039 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1042 // FIXME: handle tail calls differently.
1044 if (Subtarget->isThumb()) {
1045 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1046 CallOpc = ARMISD::CALL_NOLINK;
1048 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1050 CallOpc = (isDirect || Subtarget->hasV5TOps())
1051 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1052 : ARMISD::CALL_NOLINK;
1054 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1055 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1056 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1057 InFlag = Chain.getValue(1);
1060 std::vector<SDValue> Ops;
1061 Ops.push_back(Chain);
1062 Ops.push_back(Callee);
1064 // Add argument registers to the end of the list so that they are known live
1066 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1067 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1068 RegsToPass[i].second.getValueType()));
1070 if (InFlag.getNode())
1071 Ops.push_back(InFlag);
1072 // Returns a chain and a flag for retval copy to use.
1073 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1074 &Ops[0], Ops.size());
1075 InFlag = Chain.getValue(1);
1077 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1078 DAG.getIntPtrConstant(0, true), InFlag);
1080 InFlag = Chain.getValue(1);
1082 // Handle result values, copying them out of physregs into vregs that we
1084 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1089 ARMTargetLowering::LowerReturn(SDValue Chain,
1090 CallingConv::ID CallConv, bool isVarArg,
1091 const SmallVectorImpl<ISD::OutputArg> &Outs,
1092 DebugLoc dl, SelectionDAG &DAG) {
1094 // CCValAssign - represent the assignment of the return value to a location.
1095 SmallVector<CCValAssign, 16> RVLocs;
1097 // CCState - Info about the registers and stack slots.
1098 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1101 // Analyze outgoing return values.
1102 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1105 // If this is the first return lowered for this function, add
1106 // the regs to the liveout set for the function.
1107 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1108 for (unsigned i = 0; i != RVLocs.size(); ++i)
1109 if (RVLocs[i].isRegLoc())
1110 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1115 // Copy the result values into the output registers.
1116 for (unsigned i = 0, realRVLocIdx = 0;
1118 ++i, ++realRVLocIdx) {
1119 CCValAssign &VA = RVLocs[i];
1120 assert(VA.isRegLoc() && "Can only return in registers!");
1122 SDValue Arg = Outs[realRVLocIdx].Val;
1124 switch (VA.getLocInfo()) {
1125 default: llvm_unreachable("Unknown loc info!");
1126 case CCValAssign::Full: break;
1127 case CCValAssign::BCvt:
1128 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1132 if (VA.needsCustom()) {
1133 if (VA.getLocVT() == MVT::v2f64) {
1134 // Extract the first half and return it in two registers.
1135 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1136 DAG.getConstant(0, MVT::i32));
1137 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1138 DAG.getVTList(MVT::i32, MVT::i32), Half);
1140 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1141 Flag = Chain.getValue(1);
1142 VA = RVLocs[++i]; // skip ahead to next loc
1143 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1144 HalfGPRs.getValue(1), Flag);
1145 Flag = Chain.getValue(1);
1146 VA = RVLocs[++i]; // skip ahead to next loc
1148 // Extract the 2nd half and fall through to handle it as an f64 value.
1149 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1150 DAG.getConstant(1, MVT::i32));
1152 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1154 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1155 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1156 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1157 Flag = Chain.getValue(1);
1158 VA = RVLocs[++i]; // skip ahead to next loc
1159 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1162 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1164 // Guarantee that all emitted copies are
1165 // stuck together, avoiding something bad.
1166 Flag = Chain.getValue(1);
1171 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1173 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1178 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1179 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1180 // one of the above mentioned nodes. It has to be wrapped because otherwise
1181 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1182 // be used to form addressing mode. These wrapped nodes will be selected
1184 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1185 EVT PtrVT = Op.getValueType();
1186 // FIXME there is no actual debug info here
1187 DebugLoc dl = Op.getDebugLoc();
1188 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1190 if (CP->isMachineConstantPoolEntry())
1191 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1192 CP->getAlignment());
1194 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1195 CP->getAlignment());
1196 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1199 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1201 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1202 SelectionDAG &DAG) {
1203 DebugLoc dl = GA->getDebugLoc();
1204 EVT PtrVT = getPointerTy();
1205 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1206 ARMConstantPoolValue *CPV =
1207 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1208 ARMCP::CPValue, PCAdj, "tlsgd", true);
1209 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1210 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1211 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1212 SDValue Chain = Argument.getValue(1);
1214 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1215 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1217 // call __tls_get_addr.
1220 Entry.Node = Argument;
1221 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1222 Args.push_back(Entry);
1223 // FIXME: is there useful debug info available here?
1224 std::pair<SDValue, SDValue> CallResult =
1225 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1226 false, false, false, false,
1227 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1228 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1229 return CallResult.first;
1232 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1233 // "local exec" model.
1235 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1236 SelectionDAG &DAG) {
1237 GlobalValue *GV = GA->getGlobal();
1238 DebugLoc dl = GA->getDebugLoc();
1240 SDValue Chain = DAG.getEntryNode();
1241 EVT PtrVT = getPointerTy();
1242 // Get the Thread Pointer
1243 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1245 if (GV->isDeclaration()) {
1246 // initial exec model
1247 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1248 ARMConstantPoolValue *CPV =
1249 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1250 ARMCP::CPValue, PCAdj, "gottpoff", true);
1251 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1252 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1253 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1254 Chain = Offset.getValue(1);
1256 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1257 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1259 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1262 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1263 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1264 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1265 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1268 // The address of the thread local variable is the add of the thread
1269 // pointer with the offset of the variable.
1270 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1274 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1275 // TODO: implement the "local dynamic" model
1276 assert(Subtarget->isTargetELF() &&
1277 "TLS not implemented for non-ELF targets");
1278 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1279 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1280 // otherwise use the "Local Exec" TLS Model
1281 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1282 return LowerToTLSGeneralDynamicModel(GA, DAG);
1284 return LowerToTLSExecModels(GA, DAG);
1287 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1288 SelectionDAG &DAG) {
1289 EVT PtrVT = getPointerTy();
1290 DebugLoc dl = Op.getDebugLoc();
1291 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1292 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1293 if (RelocM == Reloc::PIC_) {
1294 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1295 ARMConstantPoolValue *CPV =
1296 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1297 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1298 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1299 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1301 SDValue Chain = Result.getValue(1);
1302 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1303 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1305 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1308 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1309 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1310 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1314 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1315 SelectionDAG &DAG) {
1316 EVT PtrVT = getPointerTy();
1317 DebugLoc dl = Op.getDebugLoc();
1318 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1319 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1321 if (RelocM == Reloc::Static)
1322 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1324 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1325 ARMConstantPoolValue *CPV =
1326 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1327 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1329 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1331 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1332 SDValue Chain = Result.getValue(1);
1334 if (RelocM == Reloc::PIC_) {
1335 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1336 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1339 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1340 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1345 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1347 assert(Subtarget->isTargetELF() &&
1348 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1349 EVT PtrVT = getPointerTy();
1350 DebugLoc dl = Op.getDebugLoc();
1351 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1352 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1353 "_GLOBAL_OFFSET_TABLE_",
1354 ARMPCLabelIndex, PCAdj);
1355 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1356 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1357 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1358 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1359 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1362 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1364 SDNode *Node = Op.getNode();
1365 EVT VT = Node->getValueType(0);
1367 // No expansion needed for 64-bit vectors.
1368 if (VT.is64BitVector())
1371 // FIXME: We need to expand VLD3 and VLD4 of 128-bit vectors into separate
1372 // operations to load the even and odd registers.
1376 static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
1378 SDNode *Node = Op.getNode();
1379 EVT VT = Node->getOperand(3).getValueType();
1381 // No expansion needed for 64-bit vectors.
1382 if (VT.is64BitVector())
1385 // FIXME: We need to expand VST3 and VST4 of 128-bit vectors into separate
1386 // operations to store the even and odd registers.
1390 static SDValue LowerNeonVLDLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
1392 SDNode *Node = Op.getNode();
1393 EVT VT = Node->getValueType(0);
1395 if (!VT.is64BitVector())
1396 return SDValue(); // unimplemented
1398 // Change the lane number operand to be a TargetConstant; otherwise it
1399 // will be legalized into a register.
1400 ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
1402 assert(false && "vld lane number must be a constant");
1405 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1406 Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
1407 return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
1410 static SDValue LowerNeonVSTLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
1412 SDNode *Node = Op.getNode();
1413 EVT VT = Node->getOperand(3).getValueType();
1415 if (!VT.is64BitVector())
1416 return SDValue(); // unimplemented
1418 // Change the lane number operand to be a TargetConstant; otherwise it
1419 // will be legalized into a register.
1420 ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
1422 assert(false && "vst lane number must be a constant");
1425 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1426 Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
1427 return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
1431 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1432 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1434 case Intrinsic::arm_neon_vld3:
1435 return LowerNeonVLDIntrinsic(Op, DAG, 3);
1436 case Intrinsic::arm_neon_vld4:
1437 return LowerNeonVLDIntrinsic(Op, DAG, 4);
1438 case Intrinsic::arm_neon_vld2lane:
1439 return LowerNeonVLDLaneIntrinsic(Op, DAG, 2);
1440 case Intrinsic::arm_neon_vld3lane:
1441 return LowerNeonVLDLaneIntrinsic(Op, DAG, 3);
1442 case Intrinsic::arm_neon_vld4lane:
1443 return LowerNeonVLDLaneIntrinsic(Op, DAG, 4);
1444 case Intrinsic::arm_neon_vst3:
1445 return LowerNeonVSTIntrinsic(Op, DAG, 3);
1446 case Intrinsic::arm_neon_vst4:
1447 return LowerNeonVSTIntrinsic(Op, DAG, 4);
1448 case Intrinsic::arm_neon_vst2lane:
1449 return LowerNeonVSTLaneIntrinsic(Op, DAG, 2);
1450 case Intrinsic::arm_neon_vst3lane:
1451 return LowerNeonVSTLaneIntrinsic(Op, DAG, 3);
1452 case Intrinsic::arm_neon_vst4lane:
1453 return LowerNeonVSTLaneIntrinsic(Op, DAG, 4);
1454 default: return SDValue(); // Don't custom lower most intrinsics.
1459 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1460 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1461 DebugLoc dl = Op.getDebugLoc();
1463 default: return SDValue(); // Don't custom lower most intrinsics.
1464 case Intrinsic::arm_thread_pointer: {
1465 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1466 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1468 case Intrinsic::eh_sjlj_lsda: {
1469 MachineFunction &MF = DAG.getMachineFunction();
1470 EVT PtrVT = getPointerTy();
1471 DebugLoc dl = Op.getDebugLoc();
1472 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1474 unsigned PCAdj = (RelocM != Reloc::PIC_)
1475 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1476 ARMConstantPoolValue *CPV =
1477 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1478 ARMCP::CPLSDA, PCAdj);
1479 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1480 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1482 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1483 SDValue Chain = Result.getValue(1);
1485 if (RelocM == Reloc::PIC_) {
1486 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1487 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1491 case Intrinsic::eh_sjlj_setjmp:
1492 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1496 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1497 unsigned VarArgsFrameIndex) {
1498 // vastart just stores the address of the VarArgsFrameIndex slot into the
1499 // memory location argument.
1500 DebugLoc dl = Op.getDebugLoc();
1501 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1502 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1503 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1504 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1508 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1509 SDNode *Node = Op.getNode();
1510 DebugLoc dl = Node->getDebugLoc();
1511 EVT VT = Node->getValueType(0);
1512 SDValue Chain = Op.getOperand(0);
1513 SDValue Size = Op.getOperand(1);
1514 SDValue Align = Op.getOperand(2);
1516 // Chain the dynamic stack allocation so that it doesn't modify the stack
1517 // pointer when other instructions are using the stack.
1518 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1520 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1521 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1522 if (AlignVal > StackAlign)
1523 // Do this now since selection pass cannot introduce new target
1524 // independent node.
1525 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1527 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1528 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1529 // do even more horrible hack later.
1530 MachineFunction &MF = DAG.getMachineFunction();
1531 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1532 if (AFI->isThumb1OnlyFunction()) {
1534 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1536 uint32_t Val = C->getZExtValue();
1537 if (Val <= 508 && ((Val & 3) == 0))
1541 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1544 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1545 SDValue Ops1[] = { Chain, Size, Align };
1546 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1547 Chain = Res.getValue(1);
1548 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1549 DAG.getIntPtrConstant(0, true), SDValue());
1550 SDValue Ops2[] = { Res, Chain };
1551 return DAG.getMergeValues(Ops2, 2, dl);
1555 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1556 SDValue &Root, SelectionDAG &DAG,
1558 MachineFunction &MF = DAG.getMachineFunction();
1559 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1561 TargetRegisterClass *RC;
1562 if (AFI->isThumb1OnlyFunction())
1563 RC = ARM::tGPRRegisterClass;
1565 RC = ARM::GPRRegisterClass;
1567 // Transform the arguments stored in physical registers into virtual ones.
1568 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1569 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1572 if (NextVA.isMemLoc()) {
1573 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1574 MachineFrameInfo *MFI = MF.getFrameInfo();
1575 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1577 // Create load node to retrieve arguments from the stack.
1578 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1579 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1581 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1582 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1585 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1589 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1590 CallingConv::ID CallConv, bool isVarArg,
1591 const SmallVectorImpl<ISD::InputArg>
1593 DebugLoc dl, SelectionDAG &DAG,
1594 SmallVectorImpl<SDValue> &InVals) {
1596 MachineFunction &MF = DAG.getMachineFunction();
1597 MachineFrameInfo *MFI = MF.getFrameInfo();
1599 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1601 // Assign locations to all of the incoming arguments.
1602 SmallVector<CCValAssign, 16> ArgLocs;
1603 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1605 CCInfo.AnalyzeFormalArguments(Ins,
1606 CCAssignFnForNode(CallConv, /* Return*/ false,
1609 SmallVector<SDValue, 16> ArgValues;
1611 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1612 CCValAssign &VA = ArgLocs[i];
1614 // Arguments stored in registers.
1615 if (VA.isRegLoc()) {
1616 EVT RegVT = VA.getLocVT();
1619 if (VA.needsCustom()) {
1620 // f64 and vector types are split up into multiple registers or
1621 // combinations of registers and stack slots.
1624 if (VA.getLocVT() == MVT::v2f64) {
1625 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1627 VA = ArgLocs[++i]; // skip ahead to next loc
1628 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1630 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1631 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1632 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1633 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1634 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1636 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1639 TargetRegisterClass *RC;
1641 if (RegVT == MVT::f32)
1642 RC = ARM::SPRRegisterClass;
1643 else if (RegVT == MVT::f64)
1644 RC = ARM::DPRRegisterClass;
1645 else if (RegVT == MVT::v2f64)
1646 RC = ARM::QPRRegisterClass;
1647 else if (RegVT == MVT::i32)
1648 RC = (AFI->isThumb1OnlyFunction() ?
1649 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1651 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1653 // Transform the arguments in physical registers into virtual ones.
1654 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1655 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1658 // If this is an 8 or 16-bit value, it is really passed promoted
1659 // to 32 bits. Insert an assert[sz]ext to capture this, then
1660 // truncate to the right size.
1661 switch (VA.getLocInfo()) {
1662 default: llvm_unreachable("Unknown loc info!");
1663 case CCValAssign::Full: break;
1664 case CCValAssign::BCvt:
1665 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1667 case CCValAssign::SExt:
1668 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1669 DAG.getValueType(VA.getValVT()));
1670 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1672 case CCValAssign::ZExt:
1673 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1674 DAG.getValueType(VA.getValVT()));
1675 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1679 InVals.push_back(ArgValue);
1681 } else { // VA.isRegLoc()
1684 assert(VA.isMemLoc());
1685 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1687 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1688 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1690 // Create load nodes to retrieve arguments from the stack.
1691 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1692 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1698 static const unsigned GPRArgRegs[] = {
1699 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1702 unsigned NumGPRs = CCInfo.getFirstUnallocated
1703 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1705 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1706 unsigned VARegSize = (4 - NumGPRs) * 4;
1707 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1708 unsigned ArgOffset = 0;
1709 if (VARegSaveSize) {
1710 // If this function is vararg, store any remaining integer argument regs
1711 // to their spots on the stack so that they may be loaded by deferencing
1712 // the result of va_next.
1713 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1714 ArgOffset = CCInfo.getNextStackOffset();
1715 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1716 VARegSaveSize - VARegSize);
1717 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1719 SmallVector<SDValue, 4> MemOps;
1720 for (; NumGPRs < 4; ++NumGPRs) {
1721 TargetRegisterClass *RC;
1722 if (AFI->isThumb1OnlyFunction())
1723 RC = ARM::tGPRRegisterClass;
1725 RC = ARM::GPRRegisterClass;
1727 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1728 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1729 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1730 MemOps.push_back(Store);
1731 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1732 DAG.getConstant(4, getPointerTy()));
1734 if (!MemOps.empty())
1735 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1736 &MemOps[0], MemOps.size());
1738 // This will point to the next argument passed via stack.
1739 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1745 /// isFloatingPointZero - Return true if this is +0.0.
1746 static bool isFloatingPointZero(SDValue Op) {
1747 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1748 return CFP->getValueAPF().isPosZero();
1749 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1750 // Maybe this has already been legalized into the constant pool?
1751 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1752 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1753 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1754 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1755 return CFP->getValueAPF().isPosZero();
1761 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1762 return ( isThumb1Only && (C & ~255U) == 0) ||
1763 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1766 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1767 /// the given operands.
1768 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1769 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1771 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1772 unsigned C = RHSC->getZExtValue();
1773 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1774 // Constant does not fit, try adjusting it by one?
1779 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1780 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1781 RHS = DAG.getConstant(C-1, MVT::i32);
1786 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1787 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1788 RHS = DAG.getConstant(C-1, MVT::i32);
1793 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1794 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1795 RHS = DAG.getConstant(C+1, MVT::i32);
1800 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1801 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1802 RHS = DAG.getConstant(C+1, MVT::i32);
1809 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1810 ARMISD::NodeType CompareType;
1813 CompareType = ARMISD::CMP;
1818 CompareType = ARMISD::CMPZ;
1821 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1822 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1825 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1826 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1829 if (!isFloatingPointZero(RHS))
1830 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1832 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1833 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1836 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1837 const ARMSubtarget *ST) {
1838 EVT VT = Op.getValueType();
1839 SDValue LHS = Op.getOperand(0);
1840 SDValue RHS = Op.getOperand(1);
1841 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1842 SDValue TrueVal = Op.getOperand(2);
1843 SDValue FalseVal = Op.getOperand(3);
1844 DebugLoc dl = Op.getDebugLoc();
1846 if (LHS.getValueType() == MVT::i32) {
1848 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1849 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1850 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1853 ARMCC::CondCodes CondCode, CondCode2;
1854 FPCCToARMCC(CC, CondCode, CondCode2);
1856 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1857 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1858 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1859 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1861 if (CondCode2 != ARMCC::AL) {
1862 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1863 // FIXME: Needs another CMP because flag can have but one use.
1864 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1865 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1866 Result, TrueVal, ARMCC2, CCR, Cmp2);
1871 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1872 const ARMSubtarget *ST) {
1873 SDValue Chain = Op.getOperand(0);
1874 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1875 SDValue LHS = Op.getOperand(2);
1876 SDValue RHS = Op.getOperand(3);
1877 SDValue Dest = Op.getOperand(4);
1878 DebugLoc dl = Op.getDebugLoc();
1880 if (LHS.getValueType() == MVT::i32) {
1882 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1883 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1884 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1885 Chain, Dest, ARMCC, CCR,Cmp);
1888 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1889 ARMCC::CondCodes CondCode, CondCode2;
1890 FPCCToARMCC(CC, CondCode, CondCode2);
1892 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1893 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1894 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1895 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1896 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1897 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1898 if (CondCode2 != ARMCC::AL) {
1899 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1900 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1901 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1906 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1907 SDValue Chain = Op.getOperand(0);
1908 SDValue Table = Op.getOperand(1);
1909 SDValue Index = Op.getOperand(2);
1910 DebugLoc dl = Op.getDebugLoc();
1912 EVT PTy = getPointerTy();
1913 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1914 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1915 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1916 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1917 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1918 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1919 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1920 if (Subtarget->isThumb2()) {
1921 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1922 // which does another jump to the destination. This also makes it easier
1923 // to translate it to TBB / TBH later.
1924 // FIXME: This might not work if the function is extremely large.
1925 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1926 Addr, Op.getOperand(2), JTI, UId);
1928 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1929 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1930 Chain = Addr.getValue(1);
1931 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1932 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1934 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1935 Chain = Addr.getValue(1);
1936 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1940 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1941 DebugLoc dl = Op.getDebugLoc();
1943 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1944 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1945 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1948 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1949 EVT VT = Op.getValueType();
1950 DebugLoc dl = Op.getDebugLoc();
1952 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1954 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1955 return DAG.getNode(Opc, dl, VT, Op);
1958 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1959 // Implement fcopysign with a fabs and a conditional fneg.
1960 SDValue Tmp0 = Op.getOperand(0);
1961 SDValue Tmp1 = Op.getOperand(1);
1962 DebugLoc dl = Op.getDebugLoc();
1963 EVT VT = Op.getValueType();
1964 EVT SrcVT = Tmp1.getValueType();
1965 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1966 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1967 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1968 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1969 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1972 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1973 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1974 MFI->setFrameAddressIsTaken(true);
1975 EVT VT = Op.getValueType();
1976 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1977 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1978 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1979 ? ARM::R7 : ARM::R11;
1980 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1982 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1987 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1989 SDValue Dst, SDValue Src,
1990 SDValue Size, unsigned Align,
1992 const Value *DstSV, uint64_t DstSVOff,
1993 const Value *SrcSV, uint64_t SrcSVOff){
1994 // Do repeated 4-byte loads and stores. To be improved.
1995 // This requires 4-byte alignment.
1996 if ((Align & 3) != 0)
1998 // This requires the copy size to be a constant, preferrably
1999 // within a subtarget-specific limit.
2000 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
2003 uint64_t SizeVal = ConstantSize->getZExtValue();
2004 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
2007 unsigned BytesLeft = SizeVal & 3;
2008 unsigned NumMemOps = SizeVal >> 2;
2009 unsigned EmittedNumMemOps = 0;
2011 unsigned VTSize = 4;
2013 const unsigned MAX_LOADS_IN_LDM = 6;
2014 SDValue TFOps[MAX_LOADS_IN_LDM];
2015 SDValue Loads[MAX_LOADS_IN_LDM];
2016 uint64_t SrcOff = 0, DstOff = 0;
2018 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2019 // same number of stores. The loads and stores will get combined into
2020 // ldm/stm later on.
2021 while (EmittedNumMemOps < NumMemOps) {
2023 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2024 Loads[i] = DAG.getLoad(VT, dl, Chain,
2025 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2026 DAG.getConstant(SrcOff, MVT::i32)),
2027 SrcSV, SrcSVOff + SrcOff);
2028 TFOps[i] = Loads[i].getValue(1);
2031 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2034 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2035 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2036 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2037 DAG.getConstant(DstOff, MVT::i32)),
2038 DstSV, DstSVOff + DstOff);
2041 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2043 EmittedNumMemOps += i;
2049 // Issue loads / stores for the trailing (1 - 3) bytes.
2050 unsigned BytesLeftSave = BytesLeft;
2053 if (BytesLeft >= 2) {
2061 Loads[i] = DAG.getLoad(VT, dl, Chain,
2062 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2063 DAG.getConstant(SrcOff, MVT::i32)),
2064 SrcSV, SrcSVOff + SrcOff);
2065 TFOps[i] = Loads[i].getValue(1);
2068 BytesLeft -= VTSize;
2070 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2073 BytesLeft = BytesLeftSave;
2075 if (BytesLeft >= 2) {
2083 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2084 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2085 DAG.getConstant(DstOff, MVT::i32)),
2086 DstSV, DstSVOff + DstOff);
2089 BytesLeft -= VTSize;
2091 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2094 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2095 SDValue Op = N->getOperand(0);
2096 DebugLoc dl = N->getDebugLoc();
2097 if (N->getValueType(0) == MVT::f64) {
2098 // Turn i64->f64 into FMDRR.
2099 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2100 DAG.getConstant(0, MVT::i32));
2101 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2102 DAG.getConstant(1, MVT::i32));
2103 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2106 // Turn f64->i64 into FMRRD.
2107 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2108 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2110 // Merge the pieces into a single i64 value.
2111 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2114 /// getZeroVector - Returns a vector of specified type with all zero elements.
2116 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2117 assert(VT.isVector() && "Expected a vector type");
2119 // Zero vectors are used to represent vector negation and in those cases
2120 // will be implemented with the NEON VNEG instruction. However, VNEG does
2121 // not support i64 elements, so sometimes the zero vectors will need to be
2122 // explicitly constructed. For those cases, and potentially other uses in
2123 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2124 // to their dest type. This ensures they get CSE'd.
2126 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2127 SmallVector<SDValue, 8> Ops;
2130 if (VT.getSizeInBits() == 64) {
2131 Ops.assign(8, Cst); TVT = MVT::v8i8;
2133 Ops.assign(16, Cst); TVT = MVT::v16i8;
2135 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2137 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2140 /// getOnesVector - Returns a vector of specified type with all bits set.
2142 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2143 assert(VT.isVector() && "Expected a vector type");
2145 // Always build ones vectors as <16 x i32> or <8 x i32> bitcasted to their
2146 // dest type. This ensures they get CSE'd.
2148 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2149 SmallVector<SDValue, 8> Ops;
2152 if (VT.getSizeInBits() == 64) {
2153 Ops.assign(8, Cst); TVT = MVT::v8i8;
2155 Ops.assign(16, Cst); TVT = MVT::v16i8;
2157 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2159 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2162 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2163 const ARMSubtarget *ST) {
2164 EVT VT = N->getValueType(0);
2165 DebugLoc dl = N->getDebugLoc();
2167 // Lower vector shifts on NEON to use VSHL.
2168 if (VT.isVector()) {
2169 assert(ST->hasNEON() && "unexpected vector shift");
2171 // Left shifts translate directly to the vshiftu intrinsic.
2172 if (N->getOpcode() == ISD::SHL)
2173 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2174 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2175 N->getOperand(0), N->getOperand(1));
2177 assert((N->getOpcode() == ISD::SRA ||
2178 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2180 // NEON uses the same intrinsics for both left and right shifts. For
2181 // right shifts, the shift amounts are negative, so negate the vector of
2183 EVT ShiftVT = N->getOperand(1).getValueType();
2184 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2185 getZeroVector(ShiftVT, DAG, dl),
2187 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2188 Intrinsic::arm_neon_vshifts :
2189 Intrinsic::arm_neon_vshiftu);
2190 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2191 DAG.getConstant(vshiftInt, MVT::i32),
2192 N->getOperand(0), NegatedCount);
2195 // We can get here for a node like i32 = ISD::SHL i32, i64
2199 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2200 "Unknown shift to lower!");
2202 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2203 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2204 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2207 // If we are in thumb mode, we don't have RRX.
2208 if (ST->isThumb1Only()) return SDValue();
2210 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2211 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2212 DAG.getConstant(0, MVT::i32));
2213 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2214 DAG.getConstant(1, MVT::i32));
2216 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2217 // captures the result into a carry flag.
2218 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2219 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2221 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2222 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2224 // Merge the pieces into a single i64 value.
2225 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2228 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2229 SDValue TmpOp0, TmpOp1;
2230 bool Invert = false;
2234 SDValue Op0 = Op.getOperand(0);
2235 SDValue Op1 = Op.getOperand(1);
2236 SDValue CC = Op.getOperand(2);
2237 EVT VT = Op.getValueType();
2238 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2239 DebugLoc dl = Op.getDebugLoc();
2241 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2242 switch (SetCCOpcode) {
2243 default: llvm_unreachable("Illegal FP comparison"); break;
2245 case ISD::SETNE: Invert = true; // Fallthrough
2247 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2249 case ISD::SETLT: Swap = true; // Fallthrough
2251 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2253 case ISD::SETLE: Swap = true; // Fallthrough
2255 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2256 case ISD::SETUGE: Swap = true; // Fallthrough
2257 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2258 case ISD::SETUGT: Swap = true; // Fallthrough
2259 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2260 case ISD::SETUEQ: Invert = true; // Fallthrough
2262 // Expand this to (OLT | OGT).
2266 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2267 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2269 case ISD::SETUO: Invert = true; // Fallthrough
2271 // Expand this to (OLT | OGE).
2275 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2276 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2280 // Integer comparisons.
2281 switch (SetCCOpcode) {
2282 default: llvm_unreachable("Illegal integer comparison"); break;
2283 case ISD::SETNE: Invert = true;
2284 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2285 case ISD::SETLT: Swap = true;
2286 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2287 case ISD::SETLE: Swap = true;
2288 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2289 case ISD::SETULT: Swap = true;
2290 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2291 case ISD::SETULE: Swap = true;
2292 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2295 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2296 if (Opc == ARMISD::VCEQ) {
2299 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2301 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2304 // Ignore bitconvert.
2305 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2306 AndOp = AndOp.getOperand(0);
2308 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2310 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2311 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2318 std::swap(Op0, Op1);
2320 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2323 Result = DAG.getNOT(dl, Result, VT);
2328 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2329 /// VMOV instruction, and if so, return the constant being splatted.
2330 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2331 unsigned SplatBitSize, SelectionDAG &DAG) {
2332 switch (SplatBitSize) {
2334 // Any 1-byte value is OK.
2335 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2336 return DAG.getTargetConstant(SplatBits, MVT::i8);
2339 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2340 if ((SplatBits & ~0xff) == 0 ||
2341 (SplatBits & ~0xff00) == 0)
2342 return DAG.getTargetConstant(SplatBits, MVT::i16);
2346 // NEON's 32-bit VMOV supports splat values where:
2347 // * only one byte is nonzero, or
2348 // * the least significant byte is 0xff and the second byte is nonzero, or
2349 // * the least significant 2 bytes are 0xff and the third is nonzero.
2350 if ((SplatBits & ~0xff) == 0 ||
2351 (SplatBits & ~0xff00) == 0 ||
2352 (SplatBits & ~0xff0000) == 0 ||
2353 (SplatBits & ~0xff000000) == 0)
2354 return DAG.getTargetConstant(SplatBits, MVT::i32);
2356 if ((SplatBits & ~0xffff) == 0 &&
2357 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2358 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2360 if ((SplatBits & ~0xffffff) == 0 &&
2361 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2362 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2364 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2365 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2366 // VMOV.I32. A (very) minor optimization would be to replicate the value
2367 // and fall through here to test for a valid 64-bit splat. But, then the
2368 // caller would also need to check and handle the change in size.
2372 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2373 uint64_t BitMask = 0xff;
2375 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2376 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2378 else if ((SplatBits & BitMask) != 0)
2382 return DAG.getTargetConstant(Val, MVT::i64);
2386 llvm_unreachable("unexpected size for isVMOVSplat");
2393 /// getVMOVImm - If this is a build_vector of constants which can be
2394 /// formed by using a VMOV instruction of the specified element size,
2395 /// return the constant being splatted. The ByteSize field indicates the
2396 /// number of bytes of each element [1248].
2397 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2398 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2399 APInt SplatBits, SplatUndef;
2400 unsigned SplatBitSize;
2402 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2403 HasAnyUndefs, ByteSize * 8))
2406 if (SplatBitSize > ByteSize * 8)
2409 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2413 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2414 bool &ReverseVEXT, unsigned &Imm) {
2415 unsigned NumElts = VT.getVectorNumElements();
2416 ReverseVEXT = false;
2419 // If this is a VEXT shuffle, the immediate value is the index of the first
2420 // element. The other shuffle indices must be the successive elements after
2422 unsigned ExpectedElt = Imm;
2423 for (unsigned i = 1; i < NumElts; ++i) {
2424 // Increment the expected index. If it wraps around, it may still be
2425 // a VEXT but the source vectors must be swapped.
2427 if (ExpectedElt == NumElts * 2) {
2432 if (ExpectedElt != static_cast<unsigned>(M[i]))
2436 // Adjust the index value if the source operands will be swapped.
2443 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2444 /// instruction with the specified blocksize. (The order of the elements
2445 /// within each block of the vector is reversed.)
2446 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2447 unsigned BlockSize) {
2448 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2449 "Only possible block sizes for VREV are: 16, 32, 64");
2451 unsigned NumElts = VT.getVectorNumElements();
2452 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2453 unsigned BlockElts = M[0] + 1;
2455 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2458 for (unsigned i = 0; i < NumElts; ++i) {
2459 if ((unsigned) M[i] !=
2460 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2467 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2468 unsigned &WhichResult) {
2469 unsigned NumElts = VT.getVectorNumElements();
2470 WhichResult = (M[0] == 0 ? 0 : 1);
2471 for (unsigned i = 0; i < NumElts; i += 2) {
2472 if ((unsigned) M[i] != i + WhichResult ||
2473 (unsigned) M[i+1] != i + NumElts + WhichResult)
2479 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2480 unsigned &WhichResult) {
2481 unsigned NumElts = VT.getVectorNumElements();
2482 WhichResult = (M[0] == 0 ? 0 : 1);
2483 for (unsigned i = 0; i != NumElts; ++i) {
2484 if ((unsigned) M[i] != 2 * i + WhichResult)
2488 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2489 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2495 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2496 unsigned &WhichResult) {
2497 unsigned NumElts = VT.getVectorNumElements();
2498 WhichResult = (M[0] == 0 ? 0 : 1);
2499 unsigned Idx = WhichResult * NumElts / 2;
2500 for (unsigned i = 0; i != NumElts; i += 2) {
2501 if ((unsigned) M[i] != Idx ||
2502 (unsigned) M[i+1] != Idx + NumElts)
2507 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2508 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2514 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2515 // Canonicalize all-zeros and all-ones vectors.
2516 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2517 if (ConstVal->isNullValue())
2518 return getZeroVector(VT, DAG, dl);
2519 if (ConstVal->isAllOnesValue())
2520 return getOnesVector(VT, DAG, dl);
2523 if (VT.is64BitVector()) {
2524 switch (Val.getValueType().getSizeInBits()) {
2525 case 8: CanonicalVT = MVT::v8i8; break;
2526 case 16: CanonicalVT = MVT::v4i16; break;
2527 case 32: CanonicalVT = MVT::v2i32; break;
2528 case 64: CanonicalVT = MVT::v1i64; break;
2529 default: llvm_unreachable("unexpected splat element type"); break;
2532 assert(VT.is128BitVector() && "unknown splat vector size");
2533 switch (Val.getValueType().getSizeInBits()) {
2534 case 8: CanonicalVT = MVT::v16i8; break;
2535 case 16: CanonicalVT = MVT::v8i16; break;
2536 case 32: CanonicalVT = MVT::v4i32; break;
2537 case 64: CanonicalVT = MVT::v2i64; break;
2538 default: llvm_unreachable("unexpected splat element type"); break;
2542 // Build a canonical splat for this value.
2543 SmallVector<SDValue, 8> Ops;
2544 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2545 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2547 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2550 // If this is a case we can't handle, return null and let the default
2551 // expansion code take care of it.
2552 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2553 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2554 DebugLoc dl = Op.getDebugLoc();
2555 EVT VT = Op.getValueType();
2557 APInt SplatBits, SplatUndef;
2558 unsigned SplatBitSize;
2560 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2561 if (SplatBitSize <= 64) {
2562 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2563 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2565 return BuildSplat(Val, VT, DAG, dl);
2569 // If there are only 2 elements in a 128-bit vector, insert them into an
2570 // undef vector. This handles the common case for 128-bit vector argument
2571 // passing, where the insertions should be translated to subreg accesses
2572 // with no real instructions.
2573 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2574 SDValue Val = DAG.getUNDEF(VT);
2575 SDValue Op0 = Op.getOperand(0);
2576 SDValue Op1 = Op.getOperand(1);
2577 if (Op0.getOpcode() != ISD::UNDEF)
2578 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2579 DAG.getIntPtrConstant(0));
2580 if (Op1.getOpcode() != ISD::UNDEF)
2581 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2582 DAG.getIntPtrConstant(1));
2589 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2590 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2591 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2592 /// are assumed to be legal.
2594 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2596 if (VT.getVectorNumElements() == 4 &&
2597 (VT.is128BitVector() || VT.is64BitVector())) {
2598 unsigned PFIndexes[4];
2599 for (unsigned i = 0; i != 4; ++i) {
2603 PFIndexes[i] = M[i];
2606 // Compute the index in the perfect shuffle table.
2607 unsigned PFTableIndex =
2608 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2609 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2610 unsigned Cost = (PFEntry >> 30);
2617 unsigned Imm, WhichResult;
2619 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2620 isVREVMask(M, VT, 64) ||
2621 isVREVMask(M, VT, 32) ||
2622 isVREVMask(M, VT, 16) ||
2623 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2624 isVTRNMask(M, VT, WhichResult) ||
2625 isVUZPMask(M, VT, WhichResult) ||
2626 isVZIPMask(M, VT, WhichResult));
2629 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2630 /// the specified operations to build the shuffle.
2631 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2632 SDValue RHS, SelectionDAG &DAG,
2634 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2635 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2636 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2639 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2648 OP_VUZPL, // VUZP, left result
2649 OP_VUZPR, // VUZP, right result
2650 OP_VZIPL, // VZIP, left result
2651 OP_VZIPR, // VZIP, right result
2652 OP_VTRNL, // VTRN, left result
2653 OP_VTRNR // VTRN, right result
2656 if (OpNum == OP_COPY) {
2657 if (LHSID == (1*9+2)*9+3) return LHS;
2658 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2662 SDValue OpLHS, OpRHS;
2663 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2664 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2665 EVT VT = OpLHS.getValueType();
2668 default: llvm_unreachable("Unknown shuffle opcode!");
2670 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2675 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2676 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2680 return DAG.getNode(ARMISD::VEXT, dl, VT,
2682 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2685 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2686 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2689 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2690 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2693 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2694 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2698 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2699 SDValue V1 = Op.getOperand(0);
2700 SDValue V2 = Op.getOperand(1);
2701 DebugLoc dl = Op.getDebugLoc();
2702 EVT VT = Op.getValueType();
2703 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2704 SmallVector<int, 8> ShuffleMask;
2706 // Convert shuffles that are directly supported on NEON to target-specific
2707 // DAG nodes, instead of keeping them as shuffles and matching them again
2708 // during code selection. This is more efficient and avoids the possibility
2709 // of inconsistencies between legalization and selection.
2710 // FIXME: floating-point vectors should be canonicalized to integer vectors
2711 // of the same time so that they get CSEd properly.
2712 SVN->getMask(ShuffleMask);
2714 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2715 int Lane = SVN->getSplatIndex();
2716 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2717 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2719 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2720 DAG.getConstant(Lane, MVT::i32));
2725 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2728 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2729 DAG.getConstant(Imm, MVT::i32));
2732 if (isVREVMask(ShuffleMask, VT, 64))
2733 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2734 if (isVREVMask(ShuffleMask, VT, 32))
2735 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2736 if (isVREVMask(ShuffleMask, VT, 16))
2737 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2739 // Check for Neon shuffles that modify both input vectors in place.
2740 // If both results are used, i.e., if there are two shuffles with the same
2741 // source operands and with masks corresponding to both results of one of
2742 // these operations, DAG memoization will ensure that a single node is
2743 // used for both shuffles.
2744 unsigned WhichResult;
2745 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2746 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2747 V1, V2).getValue(WhichResult);
2748 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2749 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2750 V1, V2).getValue(WhichResult);
2751 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2752 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2753 V1, V2).getValue(WhichResult);
2755 // If the shuffle is not directly supported and it has 4 elements, use
2756 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2757 if (VT.getVectorNumElements() == 4 &&
2758 (VT.is128BitVector() || VT.is64BitVector())) {
2759 unsigned PFIndexes[4];
2760 for (unsigned i = 0; i != 4; ++i) {
2761 if (ShuffleMask[i] < 0)
2764 PFIndexes[i] = ShuffleMask[i];
2767 // Compute the index in the perfect shuffle table.
2768 unsigned PFTableIndex =
2769 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2771 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2772 unsigned Cost = (PFEntry >> 30);
2775 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2781 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2782 EVT VT = Op.getValueType();
2783 DebugLoc dl = Op.getDebugLoc();
2784 SDValue Vec = Op.getOperand(0);
2785 SDValue Lane = Op.getOperand(1);
2787 // FIXME: This is invalid for 8 and 16-bit elements - the information about
2788 // sign / zero extension is lost!
2789 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2790 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2792 if (VT.bitsLT(MVT::i32))
2793 Op = DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2794 else if (VT.bitsGT(MVT::i32))
2795 Op = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op);
2800 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2801 // The only time a CONCAT_VECTORS operation can have legal types is when
2802 // two 64-bit vectors are concatenated to a 128-bit vector.
2803 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2804 "unexpected CONCAT_VECTORS");
2805 DebugLoc dl = Op.getDebugLoc();
2806 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2807 SDValue Op0 = Op.getOperand(0);
2808 SDValue Op1 = Op.getOperand(1);
2809 if (Op0.getOpcode() != ISD::UNDEF)
2810 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2811 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2812 DAG.getIntPtrConstant(0));
2813 if (Op1.getOpcode() != ISD::UNDEF)
2814 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2815 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2816 DAG.getIntPtrConstant(1));
2817 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2820 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2821 switch (Op.getOpcode()) {
2822 default: llvm_unreachable("Don't know how to custom lower this!");
2823 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2824 case ISD::GlobalAddress:
2825 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2826 LowerGlobalAddressELF(Op, DAG);
2827 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2828 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2829 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2830 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2831 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2832 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2833 case ISD::SINT_TO_FP:
2834 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2835 case ISD::FP_TO_SINT:
2836 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2837 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2838 case ISD::RETURNADDR: break;
2839 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2840 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2841 case ISD::INTRINSIC_VOID:
2842 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2843 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2844 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2847 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2848 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2849 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2850 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2851 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2852 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2857 /// ReplaceNodeResults - Replace the results of node with an illegal result
2858 /// type with new values built out of custom code.
2859 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2860 SmallVectorImpl<SDValue>&Results,
2861 SelectionDAG &DAG) {
2862 switch (N->getOpcode()) {
2864 llvm_unreachable("Don't know how to custom expand this!");
2866 case ISD::BIT_CONVERT:
2867 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2871 SDValue Res = LowerShift(N, DAG, Subtarget);
2873 Results.push_back(Res);
2879 //===----------------------------------------------------------------------===//
2880 // ARM Scheduler Hooks
2881 //===----------------------------------------------------------------------===//
2884 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2885 MachineBasicBlock *BB) const {
2886 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2887 DebugLoc dl = MI->getDebugLoc();
2888 switch (MI->getOpcode()) {
2890 llvm_unreachable("Unexpected instr type to insert");
2891 case ARM::tMOVCCr_pseudo: {
2892 // To "insert" a SELECT_CC instruction, we actually have to insert the
2893 // diamond control-flow pattern. The incoming instruction knows the
2894 // destination vreg to set, the condition code register to branch on, the
2895 // true/false values to select between, and a branch opcode to use.
2896 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2897 MachineFunction::iterator It = BB;
2903 // cmpTY ccX, r1, r2
2905 // fallthrough --> copy0MBB
2906 MachineBasicBlock *thisMBB = BB;
2907 MachineFunction *F = BB->getParent();
2908 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2909 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2910 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2911 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2912 F->insert(It, copy0MBB);
2913 F->insert(It, sinkMBB);
2914 // Update machine-CFG edges by first adding all successors of the current
2915 // block to the new block which will contain the Phi node for the select.
2916 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2917 e = BB->succ_end(); i != e; ++i)
2918 sinkMBB->addSuccessor(*i);
2919 // Next, remove all successors of the current block, and add the true
2920 // and fallthrough blocks as its successors.
2921 while(!BB->succ_empty())
2922 BB->removeSuccessor(BB->succ_begin());
2923 BB->addSuccessor(copy0MBB);
2924 BB->addSuccessor(sinkMBB);
2927 // %FalseValue = ...
2928 // # fallthrough to sinkMBB
2931 // Update machine-CFG edges
2932 BB->addSuccessor(sinkMBB);
2935 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2938 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2939 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2940 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2942 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2949 case ARM::t2SUBrSPi_:
2950 case ARM::t2SUBrSPi12_:
2951 case ARM::t2SUBrSPs_: {
2952 MachineFunction *MF = BB->getParent();
2953 unsigned DstReg = MI->getOperand(0).getReg();
2954 unsigned SrcReg = MI->getOperand(1).getReg();
2955 bool DstIsDead = MI->getOperand(0).isDead();
2956 bool SrcIsKill = MI->getOperand(1).isKill();
2958 if (SrcReg != ARM::SP) {
2959 // Copy the source to SP from virtual register.
2960 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2961 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2962 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2963 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2964 .addReg(SrcReg, getKillRegState(SrcIsKill));
2968 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2969 switch (MI->getOpcode()) {
2971 llvm_unreachable("Unexpected pseudo instruction!");
2977 OpOpc = ARM::tADDspr;
2980 OpOpc = ARM::tSUBspi;
2982 case ARM::t2SUBrSPi_:
2983 OpOpc = ARM::t2SUBrSPi;
2984 NeedPred = true; NeedCC = true;
2986 case ARM::t2SUBrSPi12_:
2987 OpOpc = ARM::t2SUBrSPi12;
2990 case ARM::t2SUBrSPs_:
2991 OpOpc = ARM::t2SUBrSPs;
2992 NeedPred = true; NeedCC = true; NeedOp3 = true;
2995 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2996 if (OpOpc == ARM::tAND)
2997 AddDefaultT1CC(MIB);
2998 MIB.addReg(ARM::SP);
2999 MIB.addOperand(MI->getOperand(2));
3001 MIB.addOperand(MI->getOperand(3));
3003 AddDefaultPred(MIB);
3007 // Copy the result from SP to virtual register.
3008 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3009 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3010 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3011 BuildMI(BB, dl, TII->get(CopyOpc))
3012 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3014 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3020 //===----------------------------------------------------------------------===//
3021 // ARM Optimization Hooks
3022 //===----------------------------------------------------------------------===//
3025 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3026 TargetLowering::DAGCombinerInfo &DCI) {
3027 SelectionDAG &DAG = DCI.DAG;
3028 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3029 EVT VT = N->getValueType(0);
3030 unsigned Opc = N->getOpcode();
3031 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3032 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3033 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3034 ISD::CondCode CC = ISD::SETCC_INVALID;
3037 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3039 SDValue CCOp = Slct.getOperand(0);
3040 if (CCOp.getOpcode() == ISD::SETCC)
3041 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3044 bool DoXform = false;
3046 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3049 if (LHS.getOpcode() == ISD::Constant &&
3050 cast<ConstantSDNode>(LHS)->isNullValue()) {
3052 } else if (CC != ISD::SETCC_INVALID &&
3053 RHS.getOpcode() == ISD::Constant &&
3054 cast<ConstantSDNode>(RHS)->isNullValue()) {
3055 std::swap(LHS, RHS);
3056 SDValue Op0 = Slct.getOperand(0);
3057 EVT OpVT = isSlctCC ? Op0.getValueType() :
3058 Op0.getOperand(0).getValueType();
3059 bool isInt = OpVT.isInteger();
3060 CC = ISD::getSetCCInverse(CC, isInt);
3062 if (!TLI.isCondCodeLegal(CC, OpVT))
3063 return SDValue(); // Inverse operator isn't legal.
3070 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3072 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3073 Slct.getOperand(0), Slct.getOperand(1), CC);
3074 SDValue CCOp = Slct.getOperand(0);
3076 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3077 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3078 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3079 CCOp, OtherOp, Result);
3084 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3085 static SDValue PerformADDCombine(SDNode *N,
3086 TargetLowering::DAGCombinerInfo &DCI) {
3087 // added by evan in r37685 with no testcase.
3088 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3090 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3091 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3092 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3093 if (Result.getNode()) return Result;
3095 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3096 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3097 if (Result.getNode()) return Result;
3103 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3104 static SDValue PerformSUBCombine(SDNode *N,
3105 TargetLowering::DAGCombinerInfo &DCI) {
3106 // added by evan in r37685 with no testcase.
3107 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3109 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3110 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3111 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3112 if (Result.getNode()) return Result;
3119 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3120 static SDValue PerformFMRRDCombine(SDNode *N,
3121 TargetLowering::DAGCombinerInfo &DCI) {
3122 // fmrrd(fmdrr x, y) -> x,y
3123 SDValue InDouble = N->getOperand(0);
3124 if (InDouble.getOpcode() == ARMISD::FMDRR)
3125 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3129 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3130 /// operand of a vector shift operation, where all the elements of the
3131 /// build_vector must have the same constant integer value.
3132 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3133 // Ignore bit_converts.
3134 while (Op.getOpcode() == ISD::BIT_CONVERT)
3135 Op = Op.getOperand(0);
3136 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3137 APInt SplatBits, SplatUndef;
3138 unsigned SplatBitSize;
3140 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3141 HasAnyUndefs, ElementBits) ||
3142 SplatBitSize > ElementBits)
3144 Cnt = SplatBits.getSExtValue();
3148 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3149 /// operand of a vector shift left operation. That value must be in the range:
3150 /// 0 <= Value < ElementBits for a left shift; or
3151 /// 0 <= Value <= ElementBits for a long left shift.
3152 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3153 assert(VT.isVector() && "vector shift count is not a vector type");
3154 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3155 if (! getVShiftImm(Op, ElementBits, Cnt))
3157 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3160 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3161 /// operand of a vector shift right operation. For a shift opcode, the value
3162 /// is positive, but for an intrinsic the value count must be negative. The
3163 /// absolute value must be in the range:
3164 /// 1 <= |Value| <= ElementBits for a right shift; or
3165 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3166 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3168 assert(VT.isVector() && "vector shift count is not a vector type");
3169 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3170 if (! getVShiftImm(Op, ElementBits, Cnt))
3174 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3177 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3178 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3179 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3182 // Don't do anything for most intrinsics.
3185 // Vector shifts: check for immediate versions and lower them.
3186 // Note: This is done during DAG combining instead of DAG legalizing because
3187 // the build_vectors for 64-bit vector element shift counts are generally
3188 // not legal, and it is hard to see their values after they get legalized to
3189 // loads from a constant pool.
3190 case Intrinsic::arm_neon_vshifts:
3191 case Intrinsic::arm_neon_vshiftu:
3192 case Intrinsic::arm_neon_vshiftls:
3193 case Intrinsic::arm_neon_vshiftlu:
3194 case Intrinsic::arm_neon_vshiftn:
3195 case Intrinsic::arm_neon_vrshifts:
3196 case Intrinsic::arm_neon_vrshiftu:
3197 case Intrinsic::arm_neon_vrshiftn:
3198 case Intrinsic::arm_neon_vqshifts:
3199 case Intrinsic::arm_neon_vqshiftu:
3200 case Intrinsic::arm_neon_vqshiftsu:
3201 case Intrinsic::arm_neon_vqshiftns:
3202 case Intrinsic::arm_neon_vqshiftnu:
3203 case Intrinsic::arm_neon_vqshiftnsu:
3204 case Intrinsic::arm_neon_vqrshiftns:
3205 case Intrinsic::arm_neon_vqrshiftnu:
3206 case Intrinsic::arm_neon_vqrshiftnsu: {
3207 EVT VT = N->getOperand(1).getValueType();
3209 unsigned VShiftOpc = 0;
3212 case Intrinsic::arm_neon_vshifts:
3213 case Intrinsic::arm_neon_vshiftu:
3214 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3215 VShiftOpc = ARMISD::VSHL;
3218 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3219 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3220 ARMISD::VSHRs : ARMISD::VSHRu);
3225 case Intrinsic::arm_neon_vshiftls:
3226 case Intrinsic::arm_neon_vshiftlu:
3227 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3229 llvm_unreachable("invalid shift count for vshll intrinsic");
3231 case Intrinsic::arm_neon_vrshifts:
3232 case Intrinsic::arm_neon_vrshiftu:
3233 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3237 case Intrinsic::arm_neon_vqshifts:
3238 case Intrinsic::arm_neon_vqshiftu:
3239 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3243 case Intrinsic::arm_neon_vqshiftsu:
3244 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3246 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3248 case Intrinsic::arm_neon_vshiftn:
3249 case Intrinsic::arm_neon_vrshiftn:
3250 case Intrinsic::arm_neon_vqshiftns:
3251 case Intrinsic::arm_neon_vqshiftnu:
3252 case Intrinsic::arm_neon_vqshiftnsu:
3253 case Intrinsic::arm_neon_vqrshiftns:
3254 case Intrinsic::arm_neon_vqrshiftnu:
3255 case Intrinsic::arm_neon_vqrshiftnsu:
3256 // Narrowing shifts require an immediate right shift.
3257 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3259 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3262 llvm_unreachable("unhandled vector shift");
3266 case Intrinsic::arm_neon_vshifts:
3267 case Intrinsic::arm_neon_vshiftu:
3268 // Opcode already set above.
3270 case Intrinsic::arm_neon_vshiftls:
3271 case Intrinsic::arm_neon_vshiftlu:
3272 if (Cnt == VT.getVectorElementType().getSizeInBits())
3273 VShiftOpc = ARMISD::VSHLLi;
3275 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3276 ARMISD::VSHLLs : ARMISD::VSHLLu);
3278 case Intrinsic::arm_neon_vshiftn:
3279 VShiftOpc = ARMISD::VSHRN; break;
3280 case Intrinsic::arm_neon_vrshifts:
3281 VShiftOpc = ARMISD::VRSHRs; break;
3282 case Intrinsic::arm_neon_vrshiftu:
3283 VShiftOpc = ARMISD::VRSHRu; break;
3284 case Intrinsic::arm_neon_vrshiftn:
3285 VShiftOpc = ARMISD::VRSHRN; break;
3286 case Intrinsic::arm_neon_vqshifts:
3287 VShiftOpc = ARMISD::VQSHLs; break;
3288 case Intrinsic::arm_neon_vqshiftu:
3289 VShiftOpc = ARMISD::VQSHLu; break;
3290 case Intrinsic::arm_neon_vqshiftsu:
3291 VShiftOpc = ARMISD::VQSHLsu; break;
3292 case Intrinsic::arm_neon_vqshiftns:
3293 VShiftOpc = ARMISD::VQSHRNs; break;
3294 case Intrinsic::arm_neon_vqshiftnu:
3295 VShiftOpc = ARMISD::VQSHRNu; break;
3296 case Intrinsic::arm_neon_vqshiftnsu:
3297 VShiftOpc = ARMISD::VQSHRNsu; break;
3298 case Intrinsic::arm_neon_vqrshiftns:
3299 VShiftOpc = ARMISD::VQRSHRNs; break;
3300 case Intrinsic::arm_neon_vqrshiftnu:
3301 VShiftOpc = ARMISD::VQRSHRNu; break;
3302 case Intrinsic::arm_neon_vqrshiftnsu:
3303 VShiftOpc = ARMISD::VQRSHRNsu; break;
3306 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3307 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3310 case Intrinsic::arm_neon_vshiftins: {
3311 EVT VT = N->getOperand(1).getValueType();
3313 unsigned VShiftOpc = 0;
3315 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3316 VShiftOpc = ARMISD::VSLI;
3317 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3318 VShiftOpc = ARMISD::VSRI;
3320 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3323 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3324 N->getOperand(1), N->getOperand(2),
3325 DAG.getConstant(Cnt, MVT::i32));
3328 case Intrinsic::arm_neon_vqrshifts:
3329 case Intrinsic::arm_neon_vqrshiftu:
3330 // No immediate versions of these to check for.
3337 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3338 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3339 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3340 /// vector element shift counts are generally not legal, and it is hard to see
3341 /// their values after they get legalized to loads from a constant pool.
3342 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3343 const ARMSubtarget *ST) {
3344 EVT VT = N->getValueType(0);
3346 // Nothing to be done for scalar shifts.
3347 if (! VT.isVector())
3350 assert(ST->hasNEON() && "unexpected vector shift");
3353 switch (N->getOpcode()) {
3354 default: llvm_unreachable("unexpected shift opcode");
3357 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3358 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3359 DAG.getConstant(Cnt, MVT::i32));
3364 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3365 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3366 ARMISD::VSHRs : ARMISD::VSHRu);
3367 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3368 DAG.getConstant(Cnt, MVT::i32));
3374 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3375 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3376 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3377 const ARMSubtarget *ST) {
3378 SDValue N0 = N->getOperand(0);
3380 // Check for sign- and zero-extensions of vector extract operations of 8-
3381 // and 16-bit vector elements. NEON supports these directly. They are
3382 // handled during DAG combining because type legalization will promote them
3383 // to 32-bit types and it is messy to recognize the operations after that.
3384 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3385 SDValue Vec = N0.getOperand(0);
3386 SDValue Lane = N0.getOperand(1);
3387 EVT VT = N->getValueType(0);
3388 EVT EltVT = N0.getValueType();
3389 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3391 if (VT == MVT::i32 &&
3392 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3393 TLI.isTypeLegal(Vec.getValueType())) {
3396 switch (N->getOpcode()) {
3397 default: llvm_unreachable("unexpected opcode");
3398 case ISD::SIGN_EXTEND:
3399 Opc = ARMISD::VGETLANEs;
3401 case ISD::ZERO_EXTEND:
3402 case ISD::ANY_EXTEND:
3403 Opc = ARMISD::VGETLANEu;
3406 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3413 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3414 DAGCombinerInfo &DCI) const {
3415 switch (N->getOpcode()) {
3417 case ISD::ADD: return PerformADDCombine(N, DCI);
3418 case ISD::SUB: return PerformSUBCombine(N, DCI);
3419 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3420 case ISD::INTRINSIC_WO_CHAIN:
3421 return PerformIntrinsicCombine(N, DCI.DAG);
3425 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3426 case ISD::SIGN_EXTEND:
3427 case ISD::ZERO_EXTEND:
3428 case ISD::ANY_EXTEND:
3429 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3434 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3435 if (!Subtarget->hasV6Ops())
3436 // Pre-v6 does not support unaligned mem access.
3438 else if (!Subtarget->hasV6Ops()) {
3439 // v6 may or may not support unaligned mem access.
3440 if (!Subtarget->isTargetDarwin())
3444 switch (VT.getSimpleVT().SimpleTy) {
3451 // FIXME: VLD1 etc with standard alignment is legal.
3455 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3460 switch (VT.getSimpleVT().SimpleTy) {
3461 default: return false;
3476 if ((V & (Scale - 1)) != 0)
3479 return V == (V & ((1LL << 5) - 1));
3482 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3483 const ARMSubtarget *Subtarget) {
3490 switch (VT.getSimpleVT().SimpleTy) {
3491 default: return false;
3496 // + imm12 or - imm8
3498 return V == (V & ((1LL << 8) - 1));
3499 return V == (V & ((1LL << 12) - 1));
3502 // Same as ARM mode. FIXME: NEON?
3503 if (!Subtarget->hasVFP2())
3508 return V == (V & ((1LL << 8) - 1));
3512 /// isLegalAddressImmediate - Return true if the integer value can be used
3513 /// as the offset of the target addressing mode for load / store of the
3515 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3516 const ARMSubtarget *Subtarget) {
3523 if (Subtarget->isThumb1Only())
3524 return isLegalT1AddressImmediate(V, VT);
3525 else if (Subtarget->isThumb2())
3526 return isLegalT2AddressImmediate(V, VT, Subtarget);
3531 switch (VT.getSimpleVT().SimpleTy) {
3532 default: return false;
3537 return V == (V & ((1LL << 12) - 1));
3540 return V == (V & ((1LL << 8) - 1));
3543 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3548 return V == (V & ((1LL << 8) - 1));
3552 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3554 int Scale = AM.Scale;
3558 switch (VT.getSimpleVT().SimpleTy) {
3559 default: return false;
3568 return Scale == 2 || Scale == 4 || Scale == 8;
3571 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3575 // Note, we allow "void" uses (basically, uses that aren't loads or
3576 // stores), because arm allows folding a scale into many arithmetic
3577 // operations. This should be made more precise and revisited later.
3579 // Allow r << imm, but the imm has to be a multiple of two.
3580 if (Scale & 1) return false;
3581 return isPowerOf2_32(Scale);
3585 /// isLegalAddressingMode - Return true if the addressing mode represented
3586 /// by AM is legal for this target, for a load/store of the specified type.
3587 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3588 const Type *Ty) const {
3589 EVT VT = getValueType(Ty, true);
3590 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3593 // Can never fold addr of global into load/store.
3598 case 0: // no scale reg, must be "r+i" or "r", or "i".
3601 if (Subtarget->isThumb1Only())
3605 // ARM doesn't support any R+R*scale+imm addr modes.
3612 if (Subtarget->isThumb2())
3613 return isLegalT2ScaledAddressingMode(AM, VT);
3615 int Scale = AM.Scale;
3616 switch (VT.getSimpleVT().SimpleTy) {
3617 default: return false;
3621 if (Scale < 0) Scale = -Scale;
3625 return isPowerOf2_32(Scale & ~1);
3629 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3634 // Note, we allow "void" uses (basically, uses that aren't loads or
3635 // stores), because arm allows folding a scale into many arithmetic
3636 // operations. This should be made more precise and revisited later.
3638 // Allow r << imm, but the imm has to be a multiple of two.
3639 if (Scale & 1) return false;
3640 return isPowerOf2_32(Scale);
3647 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3648 bool isSEXTLoad, SDValue &Base,
3649 SDValue &Offset, bool &isInc,
3650 SelectionDAG &DAG) {
3651 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3654 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3656 Base = Ptr->getOperand(0);
3657 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3658 int RHSC = (int)RHS->getZExtValue();
3659 if (RHSC < 0 && RHSC > -256) {
3660 assert(Ptr->getOpcode() == ISD::ADD);
3662 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3666 isInc = (Ptr->getOpcode() == ISD::ADD);
3667 Offset = Ptr->getOperand(1);
3669 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3671 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3672 int RHSC = (int)RHS->getZExtValue();
3673 if (RHSC < 0 && RHSC > -0x1000) {
3674 assert(Ptr->getOpcode() == ISD::ADD);
3676 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3677 Base = Ptr->getOperand(0);
3682 if (Ptr->getOpcode() == ISD::ADD) {
3684 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3685 if (ShOpcVal != ARM_AM::no_shift) {
3686 Base = Ptr->getOperand(1);
3687 Offset = Ptr->getOperand(0);
3689 Base = Ptr->getOperand(0);
3690 Offset = Ptr->getOperand(1);
3695 isInc = (Ptr->getOpcode() == ISD::ADD);
3696 Base = Ptr->getOperand(0);
3697 Offset = Ptr->getOperand(1);
3701 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3705 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3706 bool isSEXTLoad, SDValue &Base,
3707 SDValue &Offset, bool &isInc,
3708 SelectionDAG &DAG) {
3709 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3712 Base = Ptr->getOperand(0);
3713 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3714 int RHSC = (int)RHS->getZExtValue();
3715 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3716 assert(Ptr->getOpcode() == ISD::ADD);
3718 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3720 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3721 isInc = Ptr->getOpcode() == ISD::ADD;
3722 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3730 /// getPreIndexedAddressParts - returns true by value, base pointer and
3731 /// offset pointer and addressing mode by reference if the node's address
3732 /// can be legally represented as pre-indexed load / store address.
3734 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3736 ISD::MemIndexedMode &AM,
3737 SelectionDAG &DAG) const {
3738 if (Subtarget->isThumb1Only())
3743 bool isSEXTLoad = false;
3744 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3745 Ptr = LD->getBasePtr();
3746 VT = LD->getMemoryVT();
3747 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3748 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3749 Ptr = ST->getBasePtr();
3750 VT = ST->getMemoryVT();
3755 bool isLegal = false;
3756 if (Subtarget->isThumb2())
3757 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3758 Offset, isInc, DAG);
3760 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3761 Offset, isInc, DAG);
3765 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3769 /// getPostIndexedAddressParts - returns true by value, base pointer and
3770 /// offset pointer and addressing mode by reference if this node can be
3771 /// combined with a load / store to form a post-indexed load / store.
3772 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3775 ISD::MemIndexedMode &AM,
3776 SelectionDAG &DAG) const {
3777 if (Subtarget->isThumb1Only())
3782 bool isSEXTLoad = false;
3783 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3784 VT = LD->getMemoryVT();
3785 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3786 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3787 VT = ST->getMemoryVT();
3792 bool isLegal = false;
3793 if (Subtarget->isThumb2())
3794 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3797 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3802 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3806 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3810 const SelectionDAG &DAG,
3811 unsigned Depth) const {
3812 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3813 switch (Op.getOpcode()) {
3815 case ARMISD::CMOV: {
3816 // Bits are known zero/one if known on the LHS and RHS.
3817 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3818 if (KnownZero == 0 && KnownOne == 0) return;
3820 APInt KnownZeroRHS, KnownOneRHS;
3821 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3822 KnownZeroRHS, KnownOneRHS, Depth+1);
3823 KnownZero &= KnownZeroRHS;
3824 KnownOne &= KnownOneRHS;
3830 //===----------------------------------------------------------------------===//
3831 // ARM Inline Assembly Support
3832 //===----------------------------------------------------------------------===//
3834 /// getConstraintType - Given a constraint letter, return the type of
3835 /// constraint it is for this target.
3836 ARMTargetLowering::ConstraintType
3837 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3838 if (Constraint.size() == 1) {
3839 switch (Constraint[0]) {
3841 case 'l': return C_RegisterClass;
3842 case 'w': return C_RegisterClass;
3845 return TargetLowering::getConstraintType(Constraint);
3848 std::pair<unsigned, const TargetRegisterClass*>
3849 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3851 if (Constraint.size() == 1) {
3852 // GCC RS6000 Constraint Letters
3853 switch (Constraint[0]) {
3855 if (Subtarget->isThumb1Only())
3856 return std::make_pair(0U, ARM::tGPRRegisterClass);
3858 return std::make_pair(0U, ARM::GPRRegisterClass);
3860 return std::make_pair(0U, ARM::GPRRegisterClass);
3863 return std::make_pair(0U, ARM::SPRRegisterClass);
3865 return std::make_pair(0U, ARM::DPRRegisterClass);
3869 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3872 std::vector<unsigned> ARMTargetLowering::
3873 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3875 if (Constraint.size() != 1)
3876 return std::vector<unsigned>();
3878 switch (Constraint[0]) { // GCC ARM Constraint Letters
3881 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3882 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3885 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3886 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3887 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3888 ARM::R12, ARM::LR, 0);
3891 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3892 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3893 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3894 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3895 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3896 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3897 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3898 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3900 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3901 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3902 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3903 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3907 return std::vector<unsigned>();
3910 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3911 /// vector. If it is invalid, don't add anything to Ops.
3912 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3915 std::vector<SDValue>&Ops,
3916 SelectionDAG &DAG) const {
3917 SDValue Result(0, 0);
3919 switch (Constraint) {
3921 case 'I': case 'J': case 'K': case 'L':
3922 case 'M': case 'N': case 'O':
3923 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3927 int64_t CVal64 = C->getSExtValue();
3928 int CVal = (int) CVal64;
3929 // None of these constraints allow values larger than 32 bits. Check
3930 // that the value fits in an int.
3934 switch (Constraint) {
3936 if (Subtarget->isThumb1Only()) {
3937 // This must be a constant between 0 and 255, for ADD
3939 if (CVal >= 0 && CVal <= 255)
3941 } else if (Subtarget->isThumb2()) {
3942 // A constant that can be used as an immediate value in a
3943 // data-processing instruction.
3944 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3947 // A constant that can be used as an immediate value in a
3948 // data-processing instruction.
3949 if (ARM_AM::getSOImmVal(CVal) != -1)
3955 if (Subtarget->isThumb()) { // FIXME thumb2
3956 // This must be a constant between -255 and -1, for negated ADD
3957 // immediates. This can be used in GCC with an "n" modifier that
3958 // prints the negated value, for use with SUB instructions. It is
3959 // not useful otherwise but is implemented for compatibility.
3960 if (CVal >= -255 && CVal <= -1)
3963 // This must be a constant between -4095 and 4095. It is not clear
3964 // what this constraint is intended for. Implemented for
3965 // compatibility with GCC.
3966 if (CVal >= -4095 && CVal <= 4095)
3972 if (Subtarget->isThumb1Only()) {
3973 // A 32-bit value where only one byte has a nonzero value. Exclude
3974 // zero to match GCC. This constraint is used by GCC internally for
3975 // constants that can be loaded with a move/shift combination.
3976 // It is not useful otherwise but is implemented for compatibility.
3977 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3979 } else if (Subtarget->isThumb2()) {
3980 // A constant whose bitwise inverse can be used as an immediate
3981 // value in a data-processing instruction. This can be used in GCC
3982 // with a "B" modifier that prints the inverted value, for use with
3983 // BIC and MVN instructions. It is not useful otherwise but is
3984 // implemented for compatibility.
3985 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3988 // A constant whose bitwise inverse can be used as an immediate
3989 // value in a data-processing instruction. This can be used in GCC
3990 // with a "B" modifier that prints the inverted value, for use with
3991 // BIC and MVN instructions. It is not useful otherwise but is
3992 // implemented for compatibility.
3993 if (ARM_AM::getSOImmVal(~CVal) != -1)
3999 if (Subtarget->isThumb1Only()) {
4000 // This must be a constant between -7 and 7,
4001 // for 3-operand ADD/SUB immediate instructions.
4002 if (CVal >= -7 && CVal < 7)
4004 } else if (Subtarget->isThumb2()) {
4005 // A constant whose negation can be used as an immediate value in a
4006 // data-processing instruction. This can be used in GCC with an "n"
4007 // modifier that prints the negated value, for use with SUB
4008 // instructions. It is not useful otherwise but is implemented for
4010 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4013 // A constant whose negation can be used as an immediate value in a
4014 // data-processing instruction. This can be used in GCC with an "n"
4015 // modifier that prints the negated value, for use with SUB
4016 // instructions. It is not useful otherwise but is implemented for
4018 if (ARM_AM::getSOImmVal(-CVal) != -1)
4024 if (Subtarget->isThumb()) { // FIXME thumb2
4025 // This must be a multiple of 4 between 0 and 1020, for
4026 // ADD sp + immediate.
4027 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4030 // A power of two or a constant between 0 and 32. This is used in
4031 // GCC for the shift amount on shifted register operands, but it is
4032 // useful in general for any shift amounts.
4033 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4039 if (Subtarget->isThumb()) { // FIXME thumb2
4040 // This must be a constant between 0 and 31, for shift amounts.
4041 if (CVal >= 0 && CVal <= 31)
4047 if (Subtarget->isThumb()) { // FIXME thumb2
4048 // This must be a multiple of 4 between -508 and 508, for
4049 // ADD/SUB sp = sp + immediate.
4050 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4055 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4059 if (Result.getNode()) {
4060 Ops.push_back(Result);
4063 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,