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());
104 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
105 addRegisterClass(VT, ARM::DPRRegisterClass);
106 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
109 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
110 addRegisterClass(VT, ARM::QPRRegisterClass);
111 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
114 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
115 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
116 return new TargetLoweringObjectFileMachO();
117 return new ARMElfTargetObjectFile();
120 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
121 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
122 Subtarget = &TM.getSubtarget<ARMSubtarget>();
124 if (Subtarget->isTargetDarwin()) {
125 // Uses VFP for Thumb libfuncs if available.
126 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
127 // Single-precision floating-point arithmetic.
128 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
129 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
130 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
131 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
133 // Double-precision floating-point arithmetic.
134 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
135 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
136 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
137 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
139 // Single-precision comparisons.
140 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
141 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
142 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
143 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
144 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
145 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
146 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
147 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
149 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
150 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
151 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
152 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
153 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
154 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
155 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
156 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
158 // Double-precision comparisons.
159 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
160 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
161 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
162 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
163 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
164 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
165 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
166 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
168 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
172 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
173 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
174 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
175 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
177 // Floating-point to integer conversions.
178 // i64 conversions are done via library routines even when generating VFP
179 // instructions, so use the same ones.
180 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
181 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
182 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
183 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
185 // Conversions between floating types.
186 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
187 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
189 // Integer to floating-point conversions.
190 // i64 conversions are done via library routines even when generating VFP
191 // instructions, so use the same ones.
192 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
193 // e.g., __floatunsidf vs. __floatunssidfvfp.
194 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
195 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
196 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
197 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
201 // These libcalls are not available in 32-bit.
202 setLibcallName(RTLIB::SHL_I128, 0);
203 setLibcallName(RTLIB::SRL_I128, 0);
204 setLibcallName(RTLIB::SRA_I128, 0);
206 // Libcalls should use the AAPCS base standard ABI, even if hard float
207 // is in effect, as per the ARM RTABI specification, section 4.1.2.
208 if (Subtarget->isAAPCS_ABI()) {
209 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
210 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
211 CallingConv::ARM_AAPCS);
215 if (Subtarget->isThumb1Only())
216 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
218 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
219 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
220 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
221 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
223 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
226 if (Subtarget->hasNEON()) {
227 addDRTypeForNEON(MVT::v2f32);
228 addDRTypeForNEON(MVT::v8i8);
229 addDRTypeForNEON(MVT::v4i16);
230 addDRTypeForNEON(MVT::v2i32);
231 addDRTypeForNEON(MVT::v1i64);
233 addQRTypeForNEON(MVT::v4f32);
234 addQRTypeForNEON(MVT::v2f64);
235 addQRTypeForNEON(MVT::v16i8);
236 addQRTypeForNEON(MVT::v8i16);
237 addQRTypeForNEON(MVT::v4i32);
238 addQRTypeForNEON(MVT::v2i64);
240 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
241 // neither Neon nor VFP support any arithmetic operations on it.
242 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
243 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
244 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
245 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
246 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
247 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
248 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
249 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
250 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
251 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
252 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
253 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
254 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
255 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
256 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
257 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
258 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
259 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
260 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
261 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
262 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
263 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
264 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
265 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
267 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
268 setTargetDAGCombine(ISD::SHL);
269 setTargetDAGCombine(ISD::SRL);
270 setTargetDAGCombine(ISD::SRA);
271 setTargetDAGCombine(ISD::SIGN_EXTEND);
272 setTargetDAGCombine(ISD::ZERO_EXTEND);
273 setTargetDAGCombine(ISD::ANY_EXTEND);
276 computeRegisterProperties();
278 // ARM does not have f32 extending load.
279 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
281 // ARM does not have i1 sign extending load.
282 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
284 // ARM supports all 4 flavors of integer indexed load / store.
285 if (!Subtarget->isThumb1Only()) {
286 for (unsigned im = (unsigned)ISD::PRE_INC;
287 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
288 setIndexedLoadAction(im, MVT::i1, Legal);
289 setIndexedLoadAction(im, MVT::i8, Legal);
290 setIndexedLoadAction(im, MVT::i16, Legal);
291 setIndexedLoadAction(im, MVT::i32, Legal);
292 setIndexedStoreAction(im, MVT::i1, Legal);
293 setIndexedStoreAction(im, MVT::i8, Legal);
294 setIndexedStoreAction(im, MVT::i16, Legal);
295 setIndexedStoreAction(im, MVT::i32, Legal);
299 // i64 operation support.
300 if (Subtarget->isThumb1Only()) {
301 setOperationAction(ISD::MUL, MVT::i64, Expand);
302 setOperationAction(ISD::MULHU, MVT::i32, Expand);
303 setOperationAction(ISD::MULHS, MVT::i32, Expand);
304 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
305 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
307 setOperationAction(ISD::MUL, MVT::i64, Expand);
308 setOperationAction(ISD::MULHU, MVT::i32, Expand);
309 if (!Subtarget->hasV6Ops())
310 setOperationAction(ISD::MULHS, MVT::i32, Expand);
312 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
313 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
314 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
315 setOperationAction(ISD::SRL, MVT::i64, Custom);
316 setOperationAction(ISD::SRA, MVT::i64, Custom);
318 // ARM does not have ROTL.
319 setOperationAction(ISD::ROTL, MVT::i32, Expand);
320 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
321 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
322 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
323 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
325 // Only ARMv6 has BSWAP.
326 if (!Subtarget->hasV6Ops())
327 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
329 // These are expanded into libcalls.
330 setOperationAction(ISD::SDIV, MVT::i32, Expand);
331 setOperationAction(ISD::UDIV, MVT::i32, Expand);
332 setOperationAction(ISD::SREM, MVT::i32, Expand);
333 setOperationAction(ISD::UREM, MVT::i32, Expand);
334 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
335 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
337 // Support label based line numbers.
338 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
339 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
341 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
342 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
343 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
344 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
346 // Use the default implementation.
347 setOperationAction(ISD::VASTART, MVT::Other, Custom);
348 setOperationAction(ISD::VAARG, MVT::Other, Expand);
349 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
350 setOperationAction(ISD::VAEND, MVT::Other, Expand);
351 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
352 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
353 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
354 // FIXME: Shouldn't need this, since no register is used, but the legalizer
355 // doesn't yet know how to not do that for SjLj.
356 setExceptionSelectorRegister(ARM::R0);
357 if (Subtarget->isThumb())
358 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
360 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
361 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
363 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
364 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
365 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
367 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
369 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
370 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
371 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
373 // We want to custom lower some of our intrinsics.
374 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
375 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
376 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
378 setOperationAction(ISD::SETCC, MVT::i32, Expand);
379 setOperationAction(ISD::SETCC, MVT::f32, Expand);
380 setOperationAction(ISD::SETCC, MVT::f64, Expand);
381 setOperationAction(ISD::SELECT, MVT::i32, Expand);
382 setOperationAction(ISD::SELECT, MVT::f32, Expand);
383 setOperationAction(ISD::SELECT, MVT::f64, Expand);
384 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
385 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
386 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
388 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
389 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
390 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
391 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
392 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
394 // We don't support sin/cos/fmod/copysign/pow
395 setOperationAction(ISD::FSIN, MVT::f64, Expand);
396 setOperationAction(ISD::FSIN, MVT::f32, Expand);
397 setOperationAction(ISD::FCOS, MVT::f32, Expand);
398 setOperationAction(ISD::FCOS, MVT::f64, Expand);
399 setOperationAction(ISD::FREM, MVT::f64, Expand);
400 setOperationAction(ISD::FREM, MVT::f32, Expand);
401 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
402 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
403 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
405 setOperationAction(ISD::FPOW, MVT::f64, Expand);
406 setOperationAction(ISD::FPOW, MVT::f32, Expand);
408 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
409 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
410 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
411 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
412 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
413 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
416 // We have target-specific dag combine patterns for the following nodes:
417 // ARMISD::FMRRD - No need to call setTargetDAGCombine
418 setTargetDAGCombine(ISD::ADD);
419 setTargetDAGCombine(ISD::SUB);
421 setStackPointerRegisterToSaveRestore(ARM::SP);
422 setSchedulingPreference(SchedulingForRegPressure);
424 // FIXME: If-converter should use instruction latency to determine
425 // profitability rather than relying on fixed limits.
426 if (Subtarget->getCPUString() == "generic") {
427 // Generic (and overly aggressive) if-conversion limits.
428 setIfCvtBlockSizeLimit(10);
429 setIfCvtDupBlockSizeLimit(2);
430 } else if (Subtarget->hasV6Ops()) {
431 setIfCvtBlockSizeLimit(2);
432 setIfCvtDupBlockSizeLimit(1);
434 setIfCvtBlockSizeLimit(3);
435 setIfCvtDupBlockSizeLimit(2);
438 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
439 // Do not enable CodePlacementOpt for now: it currently runs after the
440 // ARMConstantIslandPass and messes up branch relaxation and placement
441 // of constant islands.
442 // benefitFromCodePlacementOpt = true;
445 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
448 case ARMISD::Wrapper: return "ARMISD::Wrapper";
449 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
450 case ARMISD::CALL: return "ARMISD::CALL";
451 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
452 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
453 case ARMISD::tCALL: return "ARMISD::tCALL";
454 case ARMISD::BRCOND: return "ARMISD::BRCOND";
455 case ARMISD::BR_JT: return "ARMISD::BR_JT";
456 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
457 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
458 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
459 case ARMISD::CMP: return "ARMISD::CMP";
460 case ARMISD::CMPZ: return "ARMISD::CMPZ";
461 case ARMISD::CMPFP: return "ARMISD::CMPFP";
462 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
463 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
464 case ARMISD::CMOV: return "ARMISD::CMOV";
465 case ARMISD::CNEG: return "ARMISD::CNEG";
467 case ARMISD::FTOSI: return "ARMISD::FTOSI";
468 case ARMISD::FTOUI: return "ARMISD::FTOUI";
469 case ARMISD::SITOF: return "ARMISD::SITOF";
470 case ARMISD::UITOF: return "ARMISD::UITOF";
472 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
473 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
474 case ARMISD::RRX: return "ARMISD::RRX";
476 case ARMISD::FMRRD: return "ARMISD::FMRRD";
477 case ARMISD::FMDRR: return "ARMISD::FMDRR";
479 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
481 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
483 case ARMISD::VCEQ: return "ARMISD::VCEQ";
484 case ARMISD::VCGE: return "ARMISD::VCGE";
485 case ARMISD::VCGEU: return "ARMISD::VCGEU";
486 case ARMISD::VCGT: return "ARMISD::VCGT";
487 case ARMISD::VCGTU: return "ARMISD::VCGTU";
488 case ARMISD::VTST: return "ARMISD::VTST";
490 case ARMISD::VSHL: return "ARMISD::VSHL";
491 case ARMISD::VSHRs: return "ARMISD::VSHRs";
492 case ARMISD::VSHRu: return "ARMISD::VSHRu";
493 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
494 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
495 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
496 case ARMISD::VSHRN: return "ARMISD::VSHRN";
497 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
498 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
499 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
500 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
501 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
502 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
503 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
504 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
505 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
506 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
507 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
508 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
509 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
510 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
511 case ARMISD::VDUP: return "ARMISD::VDUP";
512 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
513 case ARMISD::VEXT: return "ARMISD::VEXT";
514 case ARMISD::VREV64: return "ARMISD::VREV64";
515 case ARMISD::VREV32: return "ARMISD::VREV32";
516 case ARMISD::VREV16: return "ARMISD::VREV16";
517 case ARMISD::VZIP: return "ARMISD::VZIP";
518 case ARMISD::VUZP: return "ARMISD::VUZP";
519 case ARMISD::VTRN: return "ARMISD::VTRN";
523 /// getFunctionAlignment - Return the Log2 alignment of this function.
524 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
525 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
528 //===----------------------------------------------------------------------===//
530 //===----------------------------------------------------------------------===//
532 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
533 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
535 default: llvm_unreachable("Unknown condition code!");
536 case ISD::SETNE: return ARMCC::NE;
537 case ISD::SETEQ: return ARMCC::EQ;
538 case ISD::SETGT: return ARMCC::GT;
539 case ISD::SETGE: return ARMCC::GE;
540 case ISD::SETLT: return ARMCC::LT;
541 case ISD::SETLE: return ARMCC::LE;
542 case ISD::SETUGT: return ARMCC::HI;
543 case ISD::SETUGE: return ARMCC::HS;
544 case ISD::SETULT: return ARMCC::LO;
545 case ISD::SETULE: return ARMCC::LS;
549 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
550 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
551 ARMCC::CondCodes &CondCode2) {
552 CondCode2 = ARMCC::AL;
554 default: llvm_unreachable("Unknown FP condition!");
556 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
558 case ISD::SETOGT: CondCode = ARMCC::GT; break;
560 case ISD::SETOGE: CondCode = ARMCC::GE; break;
561 case ISD::SETOLT: CondCode = ARMCC::MI; break;
562 case ISD::SETOLE: CondCode = ARMCC::LS; break;
563 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
564 case ISD::SETO: CondCode = ARMCC::VC; break;
565 case ISD::SETUO: CondCode = ARMCC::VS; break;
566 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
567 case ISD::SETUGT: CondCode = ARMCC::HI; break;
568 case ISD::SETUGE: CondCode = ARMCC::PL; break;
570 case ISD::SETULT: CondCode = ARMCC::LT; break;
572 case ISD::SETULE: CondCode = ARMCC::LE; break;
574 case ISD::SETUNE: CondCode = ARMCC::NE; break;
578 //===----------------------------------------------------------------------===//
579 // Calling Convention Implementation
580 //===----------------------------------------------------------------------===//
582 #include "ARMGenCallingConv.inc"
584 // APCS f64 is in register pairs, possibly split to stack
585 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
586 CCValAssign::LocInfo &LocInfo,
587 CCState &State, bool CanFail) {
588 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
590 // Try to get the first register.
591 if (unsigned Reg = State.AllocateReg(RegList, 4))
592 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
594 // For the 2nd half of a v2f64, do not fail.
598 // Put the whole thing on the stack.
599 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
600 State.AllocateStack(8, 4),
605 // Try to get the second register.
606 if (unsigned Reg = State.AllocateReg(RegList, 4))
607 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
609 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
610 State.AllocateStack(4, 4),
615 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
616 CCValAssign::LocInfo &LocInfo,
617 ISD::ArgFlagsTy &ArgFlags,
619 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
621 if (LocVT == MVT::v2f64 &&
622 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
624 return true; // we handled it
627 // AAPCS f64 is in aligned register pairs
628 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
629 CCValAssign::LocInfo &LocInfo,
630 CCState &State, bool CanFail) {
631 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
632 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
634 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
636 // For the 2nd half of a v2f64, do not just fail.
640 // Put the whole thing on the stack.
641 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
642 State.AllocateStack(8, 8),
648 for (i = 0; i < 2; ++i)
649 if (HiRegList[i] == Reg)
652 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
653 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
658 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
659 CCValAssign::LocInfo &LocInfo,
660 ISD::ArgFlagsTy &ArgFlags,
662 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
664 if (LocVT == MVT::v2f64 &&
665 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
667 return true; // we handled it
670 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
671 CCValAssign::LocInfo &LocInfo, CCState &State) {
672 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
673 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
675 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
677 return false; // we didn't handle it
680 for (i = 0; i < 2; ++i)
681 if (HiRegList[i] == Reg)
684 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
685 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
690 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
691 CCValAssign::LocInfo &LocInfo,
692 ISD::ArgFlagsTy &ArgFlags,
694 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
696 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
698 return true; // we handled it
701 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
702 CCValAssign::LocInfo &LocInfo,
703 ISD::ArgFlagsTy &ArgFlags,
705 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
709 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
710 /// given CallingConvention value.
711 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
713 bool isVarArg) const {
716 llvm_unreachable("Unsupported calling convention");
718 case CallingConv::Fast:
719 // Use target triple & subtarget features to do actual dispatch.
720 if (Subtarget->isAAPCS_ABI()) {
721 if (Subtarget->hasVFP2() &&
722 FloatABIType == FloatABI::Hard && !isVarArg)
723 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
725 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
727 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
728 case CallingConv::ARM_AAPCS_VFP:
729 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
730 case CallingConv::ARM_AAPCS:
731 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
732 case CallingConv::ARM_APCS:
733 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
737 /// LowerCallResult - Lower the result values of a call into the
738 /// appropriate copies out of appropriate physical registers.
740 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
741 CallingConv::ID CallConv, bool isVarArg,
742 const SmallVectorImpl<ISD::InputArg> &Ins,
743 DebugLoc dl, SelectionDAG &DAG,
744 SmallVectorImpl<SDValue> &InVals) {
746 // Assign locations to each value returned by this call.
747 SmallVector<CCValAssign, 16> RVLocs;
748 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
749 RVLocs, *DAG.getContext());
750 CCInfo.AnalyzeCallResult(Ins,
751 CCAssignFnForNode(CallConv, /* Return*/ true,
754 // Copy all of the result registers out of their specified physreg.
755 for (unsigned i = 0; i != RVLocs.size(); ++i) {
756 CCValAssign VA = RVLocs[i];
759 if (VA.needsCustom()) {
760 // Handle f64 or half of a v2f64.
761 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
763 Chain = Lo.getValue(1);
764 InFlag = Lo.getValue(2);
765 VA = RVLocs[++i]; // skip ahead to next loc
766 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
768 Chain = Hi.getValue(1);
769 InFlag = Hi.getValue(2);
770 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
772 if (VA.getLocVT() == MVT::v2f64) {
773 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
774 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
775 DAG.getConstant(0, MVT::i32));
777 VA = RVLocs[++i]; // skip ahead to next loc
778 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
779 Chain = Lo.getValue(1);
780 InFlag = Lo.getValue(2);
781 VA = RVLocs[++i]; // skip ahead to next loc
782 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
783 Chain = Hi.getValue(1);
784 InFlag = Hi.getValue(2);
785 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
786 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
787 DAG.getConstant(1, MVT::i32));
790 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
792 Chain = Val.getValue(1);
793 InFlag = Val.getValue(2);
796 switch (VA.getLocInfo()) {
797 default: llvm_unreachable("Unknown loc info!");
798 case CCValAssign::Full: break;
799 case CCValAssign::BCvt:
800 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
804 InVals.push_back(Val);
810 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
811 /// by "Src" to address "Dst" of size "Size". Alignment information is
812 /// specified by the specific parameter attribute. The copy will be passed as
813 /// a byval function parameter.
814 /// Sometimes what we are copying is the end of a larger object, the part that
815 /// does not fit in registers.
817 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
818 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
820 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
821 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
822 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
825 /// LowerMemOpCallTo - Store the argument to the stack.
827 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
828 SDValue StackPtr, SDValue Arg,
829 DebugLoc dl, SelectionDAG &DAG,
830 const CCValAssign &VA,
831 ISD::ArgFlagsTy Flags) {
832 unsigned LocMemOffset = VA.getLocMemOffset();
833 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
834 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
835 if (Flags.isByVal()) {
836 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
838 return DAG.getStore(Chain, dl, Arg, PtrOff,
839 PseudoSourceValue::getStack(), LocMemOffset);
842 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
843 SDValue Chain, SDValue &Arg,
844 RegsToPassVector &RegsToPass,
845 CCValAssign &VA, CCValAssign &NextVA,
847 SmallVector<SDValue, 8> &MemOpChains,
848 ISD::ArgFlagsTy Flags) {
850 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
851 DAG.getVTList(MVT::i32, MVT::i32), Arg);
852 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
854 if (NextVA.isRegLoc())
855 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
857 assert(NextVA.isMemLoc());
858 if (StackPtr.getNode() == 0)
859 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
861 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
867 /// LowerCall - Lowering a call into a callseq_start <-
868 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
871 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
872 CallingConv::ID CallConv, bool isVarArg,
874 const SmallVectorImpl<ISD::OutputArg> &Outs,
875 const SmallVectorImpl<ISD::InputArg> &Ins,
876 DebugLoc dl, SelectionDAG &DAG,
877 SmallVectorImpl<SDValue> &InVals) {
879 // Analyze operands of the call, assigning locations to each operand.
880 SmallVector<CCValAssign, 16> ArgLocs;
881 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
883 CCInfo.AnalyzeCallOperands(Outs,
884 CCAssignFnForNode(CallConv, /* Return*/ false,
887 // Get a count of how many bytes are to be pushed on the stack.
888 unsigned NumBytes = CCInfo.getNextStackOffset();
890 // Adjust the stack pointer for the new arguments...
891 // These operations are automatically eliminated by the prolog/epilog pass
892 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
894 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
896 RegsToPassVector RegsToPass;
897 SmallVector<SDValue, 8> MemOpChains;
899 // Walk the register/memloc assignments, inserting copies/loads. In the case
900 // of tail call optimization, arguments are handled later.
901 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
904 CCValAssign &VA = ArgLocs[i];
905 SDValue Arg = Outs[realArgIdx].Val;
906 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
908 // Promote the value if needed.
909 switch (VA.getLocInfo()) {
910 default: llvm_unreachable("Unknown loc info!");
911 case CCValAssign::Full: break;
912 case CCValAssign::SExt:
913 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
915 case CCValAssign::ZExt:
916 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
918 case CCValAssign::AExt:
919 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
921 case CCValAssign::BCvt:
922 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
926 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
927 if (VA.needsCustom()) {
928 if (VA.getLocVT() == MVT::v2f64) {
929 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
930 DAG.getConstant(0, MVT::i32));
931 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
932 DAG.getConstant(1, MVT::i32));
934 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
935 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
937 VA = ArgLocs[++i]; // skip ahead to next loc
939 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
940 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
942 assert(VA.isMemLoc());
943 if (StackPtr.getNode() == 0)
944 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
946 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
947 dl, DAG, VA, Flags));
950 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
951 StackPtr, MemOpChains, Flags);
953 } else if (VA.isRegLoc()) {
954 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
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, Arg,
961 dl, DAG, VA, Flags));
965 if (!MemOpChains.empty())
966 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
967 &MemOpChains[0], MemOpChains.size());
969 // Build a sequence of copy-to-reg nodes chained together with token chain
970 // and flag operands which copy the outgoing args into the appropriate regs.
972 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
973 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
974 RegsToPass[i].second, InFlag);
975 InFlag = Chain.getValue(1);
978 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
979 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
980 // node so that legalize doesn't hack it.
981 bool isDirect = false;
982 bool isARMFunc = false;
983 bool isLocalARMFunc = false;
984 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
985 GlobalValue *GV = G->getGlobal();
987 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
988 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
989 getTargetMachine().getRelocationModel() != Reloc::Static;
990 isARMFunc = !Subtarget->isThumb() || isStub;
991 // ARM call to a local ARM function is predicable.
992 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
993 // tBX takes a register source operand.
994 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
995 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
998 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
999 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1000 Callee = DAG.getLoad(getPointerTy(), dl,
1001 DAG.getEntryNode(), CPAddr, NULL, 0);
1002 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1003 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1004 getPointerTy(), Callee, PICLabel);
1006 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1007 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1009 bool isStub = Subtarget->isTargetDarwin() &&
1010 getTargetMachine().getRelocationModel() != Reloc::Static;
1011 isARMFunc = !Subtarget->isThumb() || isStub;
1012 // tBX takes a register source operand.
1013 const char *Sym = S->getSymbol();
1014 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1015 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1016 Sym, ARMPCLabelIndex, 4);
1017 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1018 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1019 Callee = DAG.getLoad(getPointerTy(), dl,
1020 DAG.getEntryNode(), CPAddr, NULL, 0);
1021 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1022 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1023 getPointerTy(), Callee, PICLabel);
1025 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1028 // FIXME: handle tail calls differently.
1030 if (Subtarget->isThumb()) {
1031 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1032 CallOpc = ARMISD::CALL_NOLINK;
1034 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1036 CallOpc = (isDirect || Subtarget->hasV5TOps())
1037 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1038 : ARMISD::CALL_NOLINK;
1040 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1041 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1042 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1043 InFlag = Chain.getValue(1);
1046 std::vector<SDValue> Ops;
1047 Ops.push_back(Chain);
1048 Ops.push_back(Callee);
1050 // Add argument registers to the end of the list so that they are known live
1052 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1053 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1054 RegsToPass[i].second.getValueType()));
1056 if (InFlag.getNode())
1057 Ops.push_back(InFlag);
1058 // Returns a chain and a flag for retval copy to use.
1059 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1060 &Ops[0], Ops.size());
1061 InFlag = Chain.getValue(1);
1063 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1064 DAG.getIntPtrConstant(0, true), InFlag);
1066 InFlag = Chain.getValue(1);
1068 // Handle result values, copying them out of physregs into vregs that we
1070 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1075 ARMTargetLowering::LowerReturn(SDValue Chain,
1076 CallingConv::ID CallConv, bool isVarArg,
1077 const SmallVectorImpl<ISD::OutputArg> &Outs,
1078 DebugLoc dl, SelectionDAG &DAG) {
1080 // CCValAssign - represent the assignment of the return value to a location.
1081 SmallVector<CCValAssign, 16> RVLocs;
1083 // CCState - Info about the registers and stack slots.
1084 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1087 // Analyze outgoing return values.
1088 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1091 // If this is the first return lowered for this function, add
1092 // the regs to the liveout set for the function.
1093 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1094 for (unsigned i = 0; i != RVLocs.size(); ++i)
1095 if (RVLocs[i].isRegLoc())
1096 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1101 // Copy the result values into the output registers.
1102 for (unsigned i = 0, realRVLocIdx = 0;
1104 ++i, ++realRVLocIdx) {
1105 CCValAssign &VA = RVLocs[i];
1106 assert(VA.isRegLoc() && "Can only return in registers!");
1108 SDValue Arg = Outs[realRVLocIdx].Val;
1110 switch (VA.getLocInfo()) {
1111 default: llvm_unreachable("Unknown loc info!");
1112 case CCValAssign::Full: break;
1113 case CCValAssign::BCvt:
1114 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1118 if (VA.needsCustom()) {
1119 if (VA.getLocVT() == MVT::v2f64) {
1120 // Extract the first half and return it in two registers.
1121 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1122 DAG.getConstant(0, MVT::i32));
1123 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1124 DAG.getVTList(MVT::i32, MVT::i32), Half);
1126 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1127 Flag = Chain.getValue(1);
1128 VA = RVLocs[++i]; // skip ahead to next loc
1129 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1130 HalfGPRs.getValue(1), Flag);
1131 Flag = Chain.getValue(1);
1132 VA = RVLocs[++i]; // skip ahead to next loc
1134 // Extract the 2nd half and fall through to handle it as an f64 value.
1135 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1136 DAG.getConstant(1, MVT::i32));
1138 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1140 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1141 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1142 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1143 Flag = Chain.getValue(1);
1144 VA = RVLocs[++i]; // skip ahead to next loc
1145 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1148 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1150 // Guarantee that all emitted copies are
1151 // stuck together, avoiding something bad.
1152 Flag = Chain.getValue(1);
1157 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1159 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1164 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1165 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1166 // one of the above mentioned nodes. It has to be wrapped because otherwise
1167 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1168 // be used to form addressing mode. These wrapped nodes will be selected
1170 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1171 EVT PtrVT = Op.getValueType();
1172 // FIXME there is no actual debug info here
1173 DebugLoc dl = Op.getDebugLoc();
1174 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1176 if (CP->isMachineConstantPoolEntry())
1177 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1178 CP->getAlignment());
1180 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1181 CP->getAlignment());
1182 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1185 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1187 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1188 SelectionDAG &DAG) {
1189 DebugLoc dl = GA->getDebugLoc();
1190 EVT PtrVT = getPointerTy();
1191 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1192 ARMConstantPoolValue *CPV =
1193 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1194 ARMCP::CPValue, PCAdj, "tlsgd", true);
1195 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1196 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1197 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1198 SDValue Chain = Argument.getValue(1);
1200 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1201 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1203 // call __tls_get_addr.
1206 Entry.Node = Argument;
1207 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1208 Args.push_back(Entry);
1209 // FIXME: is there useful debug info available here?
1210 std::pair<SDValue, SDValue> CallResult =
1211 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1212 false, false, false, false,
1213 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1214 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1215 return CallResult.first;
1218 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1219 // "local exec" model.
1221 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1222 SelectionDAG &DAG) {
1223 GlobalValue *GV = GA->getGlobal();
1224 DebugLoc dl = GA->getDebugLoc();
1226 SDValue Chain = DAG.getEntryNode();
1227 EVT PtrVT = getPointerTy();
1228 // Get the Thread Pointer
1229 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1231 if (GV->isDeclaration()) {
1232 // initial exec model
1233 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1234 ARMConstantPoolValue *CPV =
1235 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1236 ARMCP::CPValue, PCAdj, "gottpoff", true);
1237 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1238 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1239 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1240 Chain = Offset.getValue(1);
1242 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1243 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1245 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1248 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1249 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1250 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1251 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1254 // The address of the thread local variable is the add of the thread
1255 // pointer with the offset of the variable.
1256 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1260 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1261 // TODO: implement the "local dynamic" model
1262 assert(Subtarget->isTargetELF() &&
1263 "TLS not implemented for non-ELF targets");
1264 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1265 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1266 // otherwise use the "Local Exec" TLS Model
1267 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1268 return LowerToTLSGeneralDynamicModel(GA, DAG);
1270 return LowerToTLSExecModels(GA, DAG);
1273 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1274 SelectionDAG &DAG) {
1275 EVT PtrVT = getPointerTy();
1276 DebugLoc dl = Op.getDebugLoc();
1277 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1278 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1279 if (RelocM == Reloc::PIC_) {
1280 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1281 ARMConstantPoolValue *CPV =
1282 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1283 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1284 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1285 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1287 SDValue Chain = Result.getValue(1);
1288 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1289 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1291 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1294 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1295 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1296 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1300 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1301 SelectionDAG &DAG) {
1302 EVT PtrVT = getPointerTy();
1303 DebugLoc dl = Op.getDebugLoc();
1304 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1305 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1307 if (RelocM == Reloc::Static)
1308 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1310 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1311 ARMConstantPoolValue *CPV =
1312 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1313 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1315 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1317 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1318 SDValue Chain = Result.getValue(1);
1320 if (RelocM == Reloc::PIC_) {
1321 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1322 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1325 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1326 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1331 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1333 assert(Subtarget->isTargetELF() &&
1334 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1335 EVT PtrVT = getPointerTy();
1336 DebugLoc dl = Op.getDebugLoc();
1337 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1338 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1339 "_GLOBAL_OFFSET_TABLE_",
1340 ARMPCLabelIndex, PCAdj);
1341 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1342 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1343 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1344 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1345 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1348 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1350 SDNode *Node = Op.getNode();
1351 EVT VT = Node->getValueType(0);
1353 // No expansion needed for 64-bit vectors.
1354 if (VT.is64BitVector())
1357 // FIXME: We need to expand VLD3 and VLD4 of 128-bit vectors into separate
1358 // operations to load the even and odd registers.
1362 static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
1364 SDNode *Node = Op.getNode();
1365 EVT VT = Node->getOperand(3).getValueType();
1367 // No expansion needed for 64-bit vectors.
1368 if (VT.is64BitVector())
1371 // FIXME: We need to expand VST3 and VST4 of 128-bit vectors into separate
1372 // operations to store the even and odd registers.
1376 static SDValue LowerNeonVLDLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
1378 SDNode *Node = Op.getNode();
1379 EVT VT = Node->getValueType(0);
1381 if (!VT.is64BitVector())
1382 return SDValue(); // unimplemented
1384 // Change the lane number operand to be a TargetConstant; otherwise it
1385 // will be legalized into a register.
1386 ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
1388 assert(false && "vld lane number must be a constant");
1391 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1392 Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
1393 return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
1396 static SDValue LowerNeonVSTLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
1398 SDNode *Node = Op.getNode();
1399 EVT VT = Node->getOperand(3).getValueType();
1401 if (!VT.is64BitVector())
1402 return SDValue(); // unimplemented
1404 // Change the lane number operand to be a TargetConstant; otherwise it
1405 // will be legalized into a register.
1406 ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
1408 assert(false && "vst lane number must be a constant");
1411 SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
1412 Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
1413 return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
1417 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1418 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1420 case Intrinsic::arm_neon_vld3:
1421 return LowerNeonVLDIntrinsic(Op, DAG, 3);
1422 case Intrinsic::arm_neon_vld4:
1423 return LowerNeonVLDIntrinsic(Op, DAG, 4);
1424 case Intrinsic::arm_neon_vld2lane:
1425 return LowerNeonVLDLaneIntrinsic(Op, DAG, 2);
1426 case Intrinsic::arm_neon_vld3lane:
1427 return LowerNeonVLDLaneIntrinsic(Op, DAG, 3);
1428 case Intrinsic::arm_neon_vld4lane:
1429 return LowerNeonVLDLaneIntrinsic(Op, DAG, 4);
1430 case Intrinsic::arm_neon_vst3:
1431 return LowerNeonVSTIntrinsic(Op, DAG, 3);
1432 case Intrinsic::arm_neon_vst4:
1433 return LowerNeonVSTIntrinsic(Op, DAG, 4);
1434 case Intrinsic::arm_neon_vst2lane:
1435 return LowerNeonVSTLaneIntrinsic(Op, DAG, 2);
1436 case Intrinsic::arm_neon_vst3lane:
1437 return LowerNeonVSTLaneIntrinsic(Op, DAG, 3);
1438 case Intrinsic::arm_neon_vst4lane:
1439 return LowerNeonVSTLaneIntrinsic(Op, DAG, 4);
1440 default: return SDValue(); // Don't custom lower most intrinsics.
1445 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1446 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1447 DebugLoc dl = Op.getDebugLoc();
1449 default: return SDValue(); // Don't custom lower most intrinsics.
1450 case Intrinsic::arm_thread_pointer: {
1451 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1452 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1454 case Intrinsic::eh_sjlj_lsda: {
1455 MachineFunction &MF = DAG.getMachineFunction();
1456 EVT PtrVT = getPointerTy();
1457 DebugLoc dl = Op.getDebugLoc();
1458 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1460 unsigned PCAdj = (RelocM != Reloc::PIC_)
1461 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1462 ARMConstantPoolValue *CPV =
1463 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1464 ARMCP::CPLSDA, PCAdj);
1465 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1466 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1468 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1469 SDValue Chain = Result.getValue(1);
1471 if (RelocM == Reloc::PIC_) {
1472 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1473 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1477 case Intrinsic::eh_sjlj_setjmp:
1478 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1482 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1483 unsigned VarArgsFrameIndex) {
1484 // vastart just stores the address of the VarArgsFrameIndex slot into the
1485 // memory location argument.
1486 DebugLoc dl = Op.getDebugLoc();
1487 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1488 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1489 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1490 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1494 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1495 SDNode *Node = Op.getNode();
1496 DebugLoc dl = Node->getDebugLoc();
1497 EVT VT = Node->getValueType(0);
1498 SDValue Chain = Op.getOperand(0);
1499 SDValue Size = Op.getOperand(1);
1500 SDValue Align = Op.getOperand(2);
1502 // Chain the dynamic stack allocation so that it doesn't modify the stack
1503 // pointer when other instructions are using the stack.
1504 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1506 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1507 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1508 if (AlignVal > StackAlign)
1509 // Do this now since selection pass cannot introduce new target
1510 // independent node.
1511 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1513 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1514 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1515 // do even more horrible hack later.
1516 MachineFunction &MF = DAG.getMachineFunction();
1517 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1518 if (AFI->isThumb1OnlyFunction()) {
1520 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1522 uint32_t Val = C->getZExtValue();
1523 if (Val <= 508 && ((Val & 3) == 0))
1527 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1530 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1531 SDValue Ops1[] = { Chain, Size, Align };
1532 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1533 Chain = Res.getValue(1);
1534 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1535 DAG.getIntPtrConstant(0, true), SDValue());
1536 SDValue Ops2[] = { Res, Chain };
1537 return DAG.getMergeValues(Ops2, 2, dl);
1541 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1542 SDValue &Root, SelectionDAG &DAG,
1544 MachineFunction &MF = DAG.getMachineFunction();
1545 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1547 TargetRegisterClass *RC;
1548 if (AFI->isThumb1OnlyFunction())
1549 RC = ARM::tGPRRegisterClass;
1551 RC = ARM::GPRRegisterClass;
1553 // Transform the arguments stored in physical registers into virtual ones.
1554 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1555 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1558 if (NextVA.isMemLoc()) {
1559 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1560 MachineFrameInfo *MFI = MF.getFrameInfo();
1561 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1563 // Create load node to retrieve arguments from the stack.
1564 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1565 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1567 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1568 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1571 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1575 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1576 CallingConv::ID CallConv, bool isVarArg,
1577 const SmallVectorImpl<ISD::InputArg>
1579 DebugLoc dl, SelectionDAG &DAG,
1580 SmallVectorImpl<SDValue> &InVals) {
1582 MachineFunction &MF = DAG.getMachineFunction();
1583 MachineFrameInfo *MFI = MF.getFrameInfo();
1585 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1587 // Assign locations to all of the incoming arguments.
1588 SmallVector<CCValAssign, 16> ArgLocs;
1589 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1591 CCInfo.AnalyzeFormalArguments(Ins,
1592 CCAssignFnForNode(CallConv, /* Return*/ false,
1595 SmallVector<SDValue, 16> ArgValues;
1597 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1598 CCValAssign &VA = ArgLocs[i];
1600 // Arguments stored in registers.
1601 if (VA.isRegLoc()) {
1602 EVT RegVT = VA.getLocVT();
1605 if (VA.needsCustom()) {
1606 // f64 and vector types are split up into multiple registers or
1607 // combinations of registers and stack slots.
1610 if (VA.getLocVT() == MVT::v2f64) {
1611 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1613 VA = ArgLocs[++i]; // skip ahead to next loc
1614 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1616 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1617 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1618 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1619 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1620 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1622 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1625 TargetRegisterClass *RC;
1627 if (RegVT == MVT::f32)
1628 RC = ARM::SPRRegisterClass;
1629 else if (RegVT == MVT::f64)
1630 RC = ARM::DPRRegisterClass;
1631 else if (RegVT == MVT::v2f64)
1632 RC = ARM::QPRRegisterClass;
1633 else if (RegVT == MVT::i32)
1634 RC = (AFI->isThumb1OnlyFunction() ?
1635 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1637 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1639 // Transform the arguments in physical registers into virtual ones.
1640 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1641 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1644 // If this is an 8 or 16-bit value, it is really passed promoted
1645 // to 32 bits. Insert an assert[sz]ext to capture this, then
1646 // truncate to the right size.
1647 switch (VA.getLocInfo()) {
1648 default: llvm_unreachable("Unknown loc info!");
1649 case CCValAssign::Full: break;
1650 case CCValAssign::BCvt:
1651 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1653 case CCValAssign::SExt:
1654 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1655 DAG.getValueType(VA.getValVT()));
1656 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1658 case CCValAssign::ZExt:
1659 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1660 DAG.getValueType(VA.getValVT()));
1661 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1665 InVals.push_back(ArgValue);
1667 } else { // VA.isRegLoc()
1670 assert(VA.isMemLoc());
1671 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1673 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1674 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1676 // Create load nodes to retrieve arguments from the stack.
1677 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1678 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1684 static const unsigned GPRArgRegs[] = {
1685 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1688 unsigned NumGPRs = CCInfo.getFirstUnallocated
1689 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1691 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1692 unsigned VARegSize = (4 - NumGPRs) * 4;
1693 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1694 unsigned ArgOffset = 0;
1695 if (VARegSaveSize) {
1696 // If this function is vararg, store any remaining integer argument regs
1697 // to their spots on the stack so that they may be loaded by deferencing
1698 // the result of va_next.
1699 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1700 ArgOffset = CCInfo.getNextStackOffset();
1701 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1702 VARegSaveSize - VARegSize);
1703 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1705 SmallVector<SDValue, 4> MemOps;
1706 for (; NumGPRs < 4; ++NumGPRs) {
1707 TargetRegisterClass *RC;
1708 if (AFI->isThumb1OnlyFunction())
1709 RC = ARM::tGPRRegisterClass;
1711 RC = ARM::GPRRegisterClass;
1713 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1714 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1715 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1716 MemOps.push_back(Store);
1717 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1718 DAG.getConstant(4, getPointerTy()));
1720 if (!MemOps.empty())
1721 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1722 &MemOps[0], MemOps.size());
1724 // This will point to the next argument passed via stack.
1725 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1731 /// isFloatingPointZero - Return true if this is +0.0.
1732 static bool isFloatingPointZero(SDValue Op) {
1733 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1734 return CFP->getValueAPF().isPosZero();
1735 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1736 // Maybe this has already been legalized into the constant pool?
1737 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1738 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1739 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1740 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1741 return CFP->getValueAPF().isPosZero();
1747 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1748 return ( isThumb1Only && (C & ~255U) == 0) ||
1749 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1752 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1753 /// the given operands.
1754 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1755 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1757 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1758 unsigned C = RHSC->getZExtValue();
1759 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1760 // Constant does not fit, try adjusting it by one?
1765 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1766 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1767 RHS = DAG.getConstant(C-1, MVT::i32);
1772 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1773 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1774 RHS = DAG.getConstant(C-1, MVT::i32);
1779 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1780 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1781 RHS = DAG.getConstant(C+1, MVT::i32);
1786 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1787 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1788 RHS = DAG.getConstant(C+1, MVT::i32);
1795 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1796 ARMISD::NodeType CompareType;
1799 CompareType = ARMISD::CMP;
1804 CompareType = ARMISD::CMPZ;
1807 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1808 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1811 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1812 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1815 if (!isFloatingPointZero(RHS))
1816 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1818 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1819 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1822 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1823 const ARMSubtarget *ST) {
1824 EVT VT = Op.getValueType();
1825 SDValue LHS = Op.getOperand(0);
1826 SDValue RHS = Op.getOperand(1);
1827 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1828 SDValue TrueVal = Op.getOperand(2);
1829 SDValue FalseVal = Op.getOperand(3);
1830 DebugLoc dl = Op.getDebugLoc();
1832 if (LHS.getValueType() == MVT::i32) {
1834 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1835 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1836 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1839 ARMCC::CondCodes CondCode, CondCode2;
1840 FPCCToARMCC(CC, CondCode, CondCode2);
1842 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1843 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1844 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1845 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1847 if (CondCode2 != ARMCC::AL) {
1848 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1849 // FIXME: Needs another CMP because flag can have but one use.
1850 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1851 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1852 Result, TrueVal, ARMCC2, CCR, Cmp2);
1857 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1858 const ARMSubtarget *ST) {
1859 SDValue Chain = Op.getOperand(0);
1860 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1861 SDValue LHS = Op.getOperand(2);
1862 SDValue RHS = Op.getOperand(3);
1863 SDValue Dest = Op.getOperand(4);
1864 DebugLoc dl = Op.getDebugLoc();
1866 if (LHS.getValueType() == MVT::i32) {
1868 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1869 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1870 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1871 Chain, Dest, ARMCC, CCR,Cmp);
1874 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1875 ARMCC::CondCodes CondCode, CondCode2;
1876 FPCCToARMCC(CC, CondCode, CondCode2);
1878 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1879 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1880 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1881 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1882 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1883 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1884 if (CondCode2 != ARMCC::AL) {
1885 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1886 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1887 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1892 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1893 SDValue Chain = Op.getOperand(0);
1894 SDValue Table = Op.getOperand(1);
1895 SDValue Index = Op.getOperand(2);
1896 DebugLoc dl = Op.getDebugLoc();
1898 EVT PTy = getPointerTy();
1899 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1900 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1901 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1902 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1903 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1904 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1905 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1906 if (Subtarget->isThumb2()) {
1907 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1908 // which does another jump to the destination. This also makes it easier
1909 // to translate it to TBB / TBH later.
1910 // FIXME: This might not work if the function is extremely large.
1911 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1912 Addr, Op.getOperand(2), JTI, UId);
1914 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1915 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1916 Chain = Addr.getValue(1);
1917 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1918 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1920 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1921 Chain = Addr.getValue(1);
1922 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1926 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1927 DebugLoc dl = Op.getDebugLoc();
1929 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1930 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1931 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1934 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1935 EVT VT = Op.getValueType();
1936 DebugLoc dl = Op.getDebugLoc();
1938 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1940 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1941 return DAG.getNode(Opc, dl, VT, Op);
1944 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1945 // Implement fcopysign with a fabs and a conditional fneg.
1946 SDValue Tmp0 = Op.getOperand(0);
1947 SDValue Tmp1 = Op.getOperand(1);
1948 DebugLoc dl = Op.getDebugLoc();
1949 EVT VT = Op.getValueType();
1950 EVT SrcVT = Tmp1.getValueType();
1951 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1952 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1953 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1954 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1955 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1958 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1959 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1960 MFI->setFrameAddressIsTaken(true);
1961 EVT VT = Op.getValueType();
1962 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1963 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1964 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1965 ? ARM::R7 : ARM::R11;
1966 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1968 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1973 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1975 SDValue Dst, SDValue Src,
1976 SDValue Size, unsigned Align,
1978 const Value *DstSV, uint64_t DstSVOff,
1979 const Value *SrcSV, uint64_t SrcSVOff){
1980 // Do repeated 4-byte loads and stores. To be improved.
1981 // This requires 4-byte alignment.
1982 if ((Align & 3) != 0)
1984 // This requires the copy size to be a constant, preferrably
1985 // within a subtarget-specific limit.
1986 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1989 uint64_t SizeVal = ConstantSize->getZExtValue();
1990 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1993 unsigned BytesLeft = SizeVal & 3;
1994 unsigned NumMemOps = SizeVal >> 2;
1995 unsigned EmittedNumMemOps = 0;
1997 unsigned VTSize = 4;
1999 const unsigned MAX_LOADS_IN_LDM = 6;
2000 SDValue TFOps[MAX_LOADS_IN_LDM];
2001 SDValue Loads[MAX_LOADS_IN_LDM];
2002 uint64_t SrcOff = 0, DstOff = 0;
2004 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
2005 // same number of stores. The loads and stores will get combined into
2006 // ldm/stm later on.
2007 while (EmittedNumMemOps < NumMemOps) {
2009 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2010 Loads[i] = DAG.getLoad(VT, dl, Chain,
2011 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2012 DAG.getConstant(SrcOff, MVT::i32)),
2013 SrcSV, SrcSVOff + SrcOff);
2014 TFOps[i] = Loads[i].getValue(1);
2017 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2020 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2021 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2022 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2023 DAG.getConstant(DstOff, MVT::i32)),
2024 DstSV, DstSVOff + DstOff);
2027 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2029 EmittedNumMemOps += i;
2035 // Issue loads / stores for the trailing (1 - 3) bytes.
2036 unsigned BytesLeftSave = BytesLeft;
2039 if (BytesLeft >= 2) {
2047 Loads[i] = DAG.getLoad(VT, dl, Chain,
2048 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2049 DAG.getConstant(SrcOff, MVT::i32)),
2050 SrcSV, SrcSVOff + SrcOff);
2051 TFOps[i] = Loads[i].getValue(1);
2054 BytesLeft -= VTSize;
2056 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2059 BytesLeft = BytesLeftSave;
2061 if (BytesLeft >= 2) {
2069 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2070 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2071 DAG.getConstant(DstOff, MVT::i32)),
2072 DstSV, DstSVOff + DstOff);
2075 BytesLeft -= VTSize;
2077 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2080 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2081 SDValue Op = N->getOperand(0);
2082 DebugLoc dl = N->getDebugLoc();
2083 if (N->getValueType(0) == MVT::f64) {
2084 // Turn i64->f64 into FMDRR.
2085 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2086 DAG.getConstant(0, MVT::i32));
2087 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2088 DAG.getConstant(1, MVT::i32));
2089 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2092 // Turn f64->i64 into FMRRD.
2093 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2094 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2096 // Merge the pieces into a single i64 value.
2097 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2100 /// getZeroVector - Returns a vector of specified type with all zero elements.
2102 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2103 assert(VT.isVector() && "Expected a vector type");
2105 // Zero vectors are used to represent vector negation and in those cases
2106 // will be implemented with the NEON VNEG instruction. However, VNEG does
2107 // not support i64 elements, so sometimes the zero vectors will need to be
2108 // explicitly constructed. For those cases, and potentially other uses in
2109 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2110 // to their dest type. This ensures they get CSE'd.
2112 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2113 SmallVector<SDValue, 8> Ops;
2116 if (VT.getSizeInBits() == 64) {
2117 Ops.assign(8, Cst); TVT = MVT::v8i8;
2119 Ops.assign(16, Cst); TVT = MVT::v16i8;
2121 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2123 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2126 /// getOnesVector - Returns a vector of specified type with all bits set.
2128 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2129 assert(VT.isVector() && "Expected a vector type");
2131 // Always build ones vectors as <16 x i32> or <8 x i32> bitcasted to their
2132 // dest type. This ensures they get CSE'd.
2134 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2135 SmallVector<SDValue, 8> Ops;
2138 if (VT.getSizeInBits() == 64) {
2139 Ops.assign(8, Cst); TVT = MVT::v8i8;
2141 Ops.assign(16, Cst); TVT = MVT::v16i8;
2143 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2145 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2148 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2149 const ARMSubtarget *ST) {
2150 EVT VT = N->getValueType(0);
2151 DebugLoc dl = N->getDebugLoc();
2153 // Lower vector shifts on NEON to use VSHL.
2154 if (VT.isVector()) {
2155 assert(ST->hasNEON() && "unexpected vector shift");
2157 // Left shifts translate directly to the vshiftu intrinsic.
2158 if (N->getOpcode() == ISD::SHL)
2159 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2160 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2161 N->getOperand(0), N->getOperand(1));
2163 assert((N->getOpcode() == ISD::SRA ||
2164 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2166 // NEON uses the same intrinsics for both left and right shifts. For
2167 // right shifts, the shift amounts are negative, so negate the vector of
2169 EVT ShiftVT = N->getOperand(1).getValueType();
2170 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2171 getZeroVector(ShiftVT, DAG, dl),
2173 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2174 Intrinsic::arm_neon_vshifts :
2175 Intrinsic::arm_neon_vshiftu);
2176 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2177 DAG.getConstant(vshiftInt, MVT::i32),
2178 N->getOperand(0), NegatedCount);
2181 // We can get here for a node like i32 = ISD::SHL i32, i64
2185 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2186 "Unknown shift to lower!");
2188 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2189 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2190 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2193 // If we are in thumb mode, we don't have RRX.
2194 if (ST->isThumb1Only()) return SDValue();
2196 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2197 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2198 DAG.getConstant(0, MVT::i32));
2199 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2200 DAG.getConstant(1, MVT::i32));
2202 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2203 // captures the result into a carry flag.
2204 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2205 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2207 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2208 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2210 // Merge the pieces into a single i64 value.
2211 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2214 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2215 SDValue TmpOp0, TmpOp1;
2216 bool Invert = false;
2220 SDValue Op0 = Op.getOperand(0);
2221 SDValue Op1 = Op.getOperand(1);
2222 SDValue CC = Op.getOperand(2);
2223 EVT VT = Op.getValueType();
2224 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2225 DebugLoc dl = Op.getDebugLoc();
2227 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2228 switch (SetCCOpcode) {
2229 default: llvm_unreachable("Illegal FP comparison"); break;
2231 case ISD::SETNE: Invert = true; // Fallthrough
2233 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2235 case ISD::SETLT: Swap = true; // Fallthrough
2237 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2239 case ISD::SETLE: Swap = true; // Fallthrough
2241 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2242 case ISD::SETUGE: Swap = true; // Fallthrough
2243 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2244 case ISD::SETUGT: Swap = true; // Fallthrough
2245 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2246 case ISD::SETUEQ: Invert = true; // Fallthrough
2248 // Expand this to (OLT | OGT).
2252 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2253 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2255 case ISD::SETUO: Invert = true; // Fallthrough
2257 // Expand this to (OLT | OGE).
2261 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2262 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2266 // Integer comparisons.
2267 switch (SetCCOpcode) {
2268 default: llvm_unreachable("Illegal integer comparison"); break;
2269 case ISD::SETNE: Invert = true;
2270 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2271 case ISD::SETLT: Swap = true;
2272 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2273 case ISD::SETLE: Swap = true;
2274 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2275 case ISD::SETULT: Swap = true;
2276 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2277 case ISD::SETULE: Swap = true;
2278 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2281 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2282 if (Opc == ARMISD::VCEQ) {
2285 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2287 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2290 // Ignore bitconvert.
2291 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2292 AndOp = AndOp.getOperand(0);
2294 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2296 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2297 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2304 std::swap(Op0, Op1);
2306 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2309 Result = DAG.getNOT(dl, Result, VT);
2314 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2315 /// VMOV instruction, and if so, return the constant being splatted.
2316 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2317 unsigned SplatBitSize, SelectionDAG &DAG) {
2318 switch (SplatBitSize) {
2320 // Any 1-byte value is OK.
2321 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2322 return DAG.getTargetConstant(SplatBits, MVT::i8);
2325 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2326 if ((SplatBits & ~0xff) == 0 ||
2327 (SplatBits & ~0xff00) == 0)
2328 return DAG.getTargetConstant(SplatBits, MVT::i16);
2332 // NEON's 32-bit VMOV supports splat values where:
2333 // * only one byte is nonzero, or
2334 // * the least significant byte is 0xff and the second byte is nonzero, or
2335 // * the least significant 2 bytes are 0xff and the third is nonzero.
2336 if ((SplatBits & ~0xff) == 0 ||
2337 (SplatBits & ~0xff00) == 0 ||
2338 (SplatBits & ~0xff0000) == 0 ||
2339 (SplatBits & ~0xff000000) == 0)
2340 return DAG.getTargetConstant(SplatBits, MVT::i32);
2342 if ((SplatBits & ~0xffff) == 0 &&
2343 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2344 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2346 if ((SplatBits & ~0xffffff) == 0 &&
2347 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2348 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2350 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2351 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2352 // VMOV.I32. A (very) minor optimization would be to replicate the value
2353 // and fall through here to test for a valid 64-bit splat. But, then the
2354 // caller would also need to check and handle the change in size.
2358 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2359 uint64_t BitMask = 0xff;
2361 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2362 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2364 else if ((SplatBits & BitMask) != 0)
2368 return DAG.getTargetConstant(Val, MVT::i64);
2372 llvm_unreachable("unexpected size for isVMOVSplat");
2379 /// getVMOVImm - If this is a build_vector of constants which can be
2380 /// formed by using a VMOV instruction of the specified element size,
2381 /// return the constant being splatted. The ByteSize field indicates the
2382 /// number of bytes of each element [1248].
2383 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2384 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2385 APInt SplatBits, SplatUndef;
2386 unsigned SplatBitSize;
2388 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2389 HasAnyUndefs, ByteSize * 8))
2392 if (SplatBitSize > ByteSize * 8)
2395 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2399 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2400 bool &ReverseVEXT, unsigned &Imm) {
2401 unsigned NumElts = VT.getVectorNumElements();
2402 ReverseVEXT = false;
2405 // If this is a VEXT shuffle, the immediate value is the index of the first
2406 // element. The other shuffle indices must be the successive elements after
2408 unsigned ExpectedElt = Imm;
2409 for (unsigned i = 1; i < NumElts; ++i) {
2410 // Increment the expected index. If it wraps around, it may still be
2411 // a VEXT but the source vectors must be swapped.
2413 if (ExpectedElt == NumElts * 2) {
2418 if (ExpectedElt != static_cast<unsigned>(M[i]))
2422 // Adjust the index value if the source operands will be swapped.
2429 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2430 /// instruction with the specified blocksize. (The order of the elements
2431 /// within each block of the vector is reversed.)
2432 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2433 unsigned BlockSize) {
2434 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2435 "Only possible block sizes for VREV are: 16, 32, 64");
2437 unsigned NumElts = VT.getVectorNumElements();
2438 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2439 unsigned BlockElts = M[0] + 1;
2441 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2444 for (unsigned i = 0; i < NumElts; ++i) {
2445 if ((unsigned) M[i] !=
2446 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2453 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2454 unsigned &WhichResult) {
2455 unsigned NumElts = VT.getVectorNumElements();
2456 WhichResult = (M[0] == 0 ? 0 : 1);
2457 for (unsigned i = 0; i < NumElts; i += 2) {
2458 if ((unsigned) M[i] != i + WhichResult ||
2459 (unsigned) M[i+1] != i + NumElts + WhichResult)
2465 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2466 unsigned &WhichResult) {
2467 unsigned NumElts = VT.getVectorNumElements();
2468 WhichResult = (M[0] == 0 ? 0 : 1);
2469 for (unsigned i = 0; i != NumElts; ++i) {
2470 if ((unsigned) M[i] != 2 * i + WhichResult)
2474 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2475 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2481 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2482 unsigned &WhichResult) {
2483 unsigned NumElts = VT.getVectorNumElements();
2484 WhichResult = (M[0] == 0 ? 0 : 1);
2485 unsigned Idx = WhichResult * NumElts / 2;
2486 for (unsigned i = 0; i != NumElts; i += 2) {
2487 if ((unsigned) M[i] != Idx ||
2488 (unsigned) M[i+1] != Idx + NumElts)
2493 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2494 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2500 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2501 // Canonicalize all-zeros and all-ones vectors.
2502 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2503 if (ConstVal->isNullValue())
2504 return getZeroVector(VT, DAG, dl);
2505 if (ConstVal->isAllOnesValue())
2506 return getOnesVector(VT, DAG, dl);
2509 if (VT.is64BitVector()) {
2510 switch (Val.getValueType().getSizeInBits()) {
2511 case 8: CanonicalVT = MVT::v8i8; break;
2512 case 16: CanonicalVT = MVT::v4i16; break;
2513 case 32: CanonicalVT = MVT::v2i32; break;
2514 case 64: CanonicalVT = MVT::v1i64; break;
2515 default: llvm_unreachable("unexpected splat element type"); break;
2518 assert(VT.is128BitVector() && "unknown splat vector size");
2519 switch (Val.getValueType().getSizeInBits()) {
2520 case 8: CanonicalVT = MVT::v16i8; break;
2521 case 16: CanonicalVT = MVT::v8i16; break;
2522 case 32: CanonicalVT = MVT::v4i32; break;
2523 case 64: CanonicalVT = MVT::v2i64; break;
2524 default: llvm_unreachable("unexpected splat element type"); break;
2528 // Build a canonical splat for this value.
2529 SmallVector<SDValue, 8> Ops;
2530 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2531 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2533 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2536 // If this is a case we can't handle, return null and let the default
2537 // expansion code take care of it.
2538 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2539 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2540 DebugLoc dl = Op.getDebugLoc();
2541 EVT VT = Op.getValueType();
2543 APInt SplatBits, SplatUndef;
2544 unsigned SplatBitSize;
2546 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2547 if (SplatBitSize <= 64) {
2548 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2549 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2551 return BuildSplat(Val, VT, DAG, dl);
2555 // If there are only 2 elements in a 128-bit vector, insert them into an
2556 // undef vector. This handles the common case for 128-bit vector argument
2557 // passing, where the insertions should be translated to subreg accesses
2558 // with no real instructions.
2559 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2560 SDValue Val = DAG.getUNDEF(VT);
2561 SDValue Op0 = Op.getOperand(0);
2562 SDValue Op1 = Op.getOperand(1);
2563 if (Op0.getOpcode() != ISD::UNDEF)
2564 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2565 DAG.getIntPtrConstant(0));
2566 if (Op1.getOpcode() != ISD::UNDEF)
2567 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2568 DAG.getIntPtrConstant(1));
2575 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2576 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2577 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2578 /// are assumed to be legal.
2580 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2582 if (VT.getVectorNumElements() == 4 &&
2583 (VT.is128BitVector() || VT.is64BitVector())) {
2584 unsigned PFIndexes[4];
2585 for (unsigned i = 0; i != 4; ++i) {
2589 PFIndexes[i] = M[i];
2592 // Compute the index in the perfect shuffle table.
2593 unsigned PFTableIndex =
2594 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2595 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2596 unsigned Cost = (PFEntry >> 30);
2603 unsigned Imm, WhichResult;
2605 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2606 isVREVMask(M, VT, 64) ||
2607 isVREVMask(M, VT, 32) ||
2608 isVREVMask(M, VT, 16) ||
2609 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2610 isVTRNMask(M, VT, WhichResult) ||
2611 isVUZPMask(M, VT, WhichResult) ||
2612 isVZIPMask(M, VT, WhichResult));
2615 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2616 /// the specified operations to build the shuffle.
2617 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2618 SDValue RHS, SelectionDAG &DAG,
2620 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2621 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2622 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2625 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2634 OP_VUZPL, // VUZP, left result
2635 OP_VUZPR, // VUZP, right result
2636 OP_VZIPL, // VZIP, left result
2637 OP_VZIPR, // VZIP, right result
2638 OP_VTRNL, // VTRN, left result
2639 OP_VTRNR // VTRN, right result
2642 if (OpNum == OP_COPY) {
2643 if (LHSID == (1*9+2)*9+3) return LHS;
2644 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2648 SDValue OpLHS, OpRHS;
2649 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2650 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2651 EVT VT = OpLHS.getValueType();
2654 default: llvm_unreachable("Unknown shuffle opcode!");
2656 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2661 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2662 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2666 return DAG.getNode(ARMISD::VEXT, dl, VT,
2668 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2671 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2672 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2675 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2676 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2679 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2680 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2684 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2685 SDValue V1 = Op.getOperand(0);
2686 SDValue V2 = Op.getOperand(1);
2687 DebugLoc dl = Op.getDebugLoc();
2688 EVT VT = Op.getValueType();
2689 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2690 SmallVector<int, 8> ShuffleMask;
2692 // Convert shuffles that are directly supported on NEON to target-specific
2693 // DAG nodes, instead of keeping them as shuffles and matching them again
2694 // during code selection. This is more efficient and avoids the possibility
2695 // of inconsistencies between legalization and selection.
2696 // FIXME: floating-point vectors should be canonicalized to integer vectors
2697 // of the same time so that they get CSEd properly.
2698 SVN->getMask(ShuffleMask);
2700 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2701 int Lane = SVN->getSplatIndex();
2702 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2703 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2705 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2706 DAG.getConstant(Lane, MVT::i32));
2711 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2714 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2715 DAG.getConstant(Imm, MVT::i32));
2718 if (isVREVMask(ShuffleMask, VT, 64))
2719 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2720 if (isVREVMask(ShuffleMask, VT, 32))
2721 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2722 if (isVREVMask(ShuffleMask, VT, 16))
2723 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2725 // Check for Neon shuffles that modify both input vectors in place.
2726 // If both results are used, i.e., if there are two shuffles with the same
2727 // source operands and with masks corresponding to both results of one of
2728 // these operations, DAG memoization will ensure that a single node is
2729 // used for both shuffles.
2730 unsigned WhichResult;
2731 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2732 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2733 V1, V2).getValue(WhichResult);
2734 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2735 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2736 V1, V2).getValue(WhichResult);
2737 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2738 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2739 V1, V2).getValue(WhichResult);
2741 // If the shuffle is not directly supported and it has 4 elements, use
2742 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2743 if (VT.getVectorNumElements() == 4 &&
2744 (VT.is128BitVector() || VT.is64BitVector())) {
2745 unsigned PFIndexes[4];
2746 for (unsigned i = 0; i != 4; ++i) {
2747 if (ShuffleMask[i] < 0)
2750 PFIndexes[i] = ShuffleMask[i];
2753 // Compute the index in the perfect shuffle table.
2754 unsigned PFTableIndex =
2755 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2757 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2758 unsigned Cost = (PFEntry >> 30);
2761 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2767 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2768 EVT VT = Op.getValueType();
2769 DebugLoc dl = Op.getDebugLoc();
2770 SDValue Vec = Op.getOperand(0);
2771 SDValue Lane = Op.getOperand(1);
2773 // FIXME: This is invalid for 8 and 16-bit elements - the information about
2774 // sign / zero extension is lost!
2775 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2776 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2778 if (VT.bitsLT(MVT::i32))
2779 Op = DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2780 else if (VT.bitsGT(MVT::i32))
2781 Op = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op);
2786 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2787 // The only time a CONCAT_VECTORS operation can have legal types is when
2788 // two 64-bit vectors are concatenated to a 128-bit vector.
2789 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2790 "unexpected CONCAT_VECTORS");
2791 DebugLoc dl = Op.getDebugLoc();
2792 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2793 SDValue Op0 = Op.getOperand(0);
2794 SDValue Op1 = Op.getOperand(1);
2795 if (Op0.getOpcode() != ISD::UNDEF)
2796 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2797 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2798 DAG.getIntPtrConstant(0));
2799 if (Op1.getOpcode() != ISD::UNDEF)
2800 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2801 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2802 DAG.getIntPtrConstant(1));
2803 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2806 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2807 switch (Op.getOpcode()) {
2808 default: llvm_unreachable("Don't know how to custom lower this!");
2809 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2810 case ISD::GlobalAddress:
2811 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2812 LowerGlobalAddressELF(Op, DAG);
2813 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2814 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2815 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2816 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2817 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2818 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2819 case ISD::SINT_TO_FP:
2820 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2821 case ISD::FP_TO_SINT:
2822 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2823 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2824 case ISD::RETURNADDR: break;
2825 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2826 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2827 case ISD::INTRINSIC_VOID:
2828 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2829 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2830 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2833 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2834 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2835 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2836 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2837 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2838 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2843 /// ReplaceNodeResults - Replace the results of node with an illegal result
2844 /// type with new values built out of custom code.
2845 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2846 SmallVectorImpl<SDValue>&Results,
2847 SelectionDAG &DAG) {
2848 switch (N->getOpcode()) {
2850 llvm_unreachable("Don't know how to custom expand this!");
2852 case ISD::BIT_CONVERT:
2853 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2857 SDValue Res = LowerShift(N, DAG, Subtarget);
2859 Results.push_back(Res);
2865 //===----------------------------------------------------------------------===//
2866 // ARM Scheduler Hooks
2867 //===----------------------------------------------------------------------===//
2870 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2871 MachineBasicBlock *BB) const {
2872 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2873 DebugLoc dl = MI->getDebugLoc();
2874 switch (MI->getOpcode()) {
2876 llvm_unreachable("Unexpected instr type to insert");
2877 case ARM::tMOVCCr_pseudo: {
2878 // To "insert" a SELECT_CC instruction, we actually have to insert the
2879 // diamond control-flow pattern. The incoming instruction knows the
2880 // destination vreg to set, the condition code register to branch on, the
2881 // true/false values to select between, and a branch opcode to use.
2882 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2883 MachineFunction::iterator It = BB;
2889 // cmpTY ccX, r1, r2
2891 // fallthrough --> copy0MBB
2892 MachineBasicBlock *thisMBB = BB;
2893 MachineFunction *F = BB->getParent();
2894 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2895 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2896 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2897 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2898 F->insert(It, copy0MBB);
2899 F->insert(It, sinkMBB);
2900 // Update machine-CFG edges by first adding all successors of the current
2901 // block to the new block which will contain the Phi node for the select.
2902 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2903 e = BB->succ_end(); i != e; ++i)
2904 sinkMBB->addSuccessor(*i);
2905 // Next, remove all successors of the current block, and add the true
2906 // and fallthrough blocks as its successors.
2907 while(!BB->succ_empty())
2908 BB->removeSuccessor(BB->succ_begin());
2909 BB->addSuccessor(copy0MBB);
2910 BB->addSuccessor(sinkMBB);
2913 // %FalseValue = ...
2914 // # fallthrough to sinkMBB
2917 // Update machine-CFG edges
2918 BB->addSuccessor(sinkMBB);
2921 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2924 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2925 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2926 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2928 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2935 case ARM::t2SUBrSPi_:
2936 case ARM::t2SUBrSPi12_:
2937 case ARM::t2SUBrSPs_: {
2938 MachineFunction *MF = BB->getParent();
2939 unsigned DstReg = MI->getOperand(0).getReg();
2940 unsigned SrcReg = MI->getOperand(1).getReg();
2941 bool DstIsDead = MI->getOperand(0).isDead();
2942 bool SrcIsKill = MI->getOperand(1).isKill();
2944 if (SrcReg != ARM::SP) {
2945 // Copy the source to SP from virtual register.
2946 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2947 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2948 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2949 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2950 .addReg(SrcReg, getKillRegState(SrcIsKill));
2954 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2955 switch (MI->getOpcode()) {
2957 llvm_unreachable("Unexpected pseudo instruction!");
2963 OpOpc = ARM::tADDspr;
2966 OpOpc = ARM::tSUBspi;
2968 case ARM::t2SUBrSPi_:
2969 OpOpc = ARM::t2SUBrSPi;
2970 NeedPred = true; NeedCC = true;
2972 case ARM::t2SUBrSPi12_:
2973 OpOpc = ARM::t2SUBrSPi12;
2976 case ARM::t2SUBrSPs_:
2977 OpOpc = ARM::t2SUBrSPs;
2978 NeedPred = true; NeedCC = true; NeedOp3 = true;
2981 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2982 if (OpOpc == ARM::tAND)
2983 AddDefaultT1CC(MIB);
2984 MIB.addReg(ARM::SP);
2985 MIB.addOperand(MI->getOperand(2));
2987 MIB.addOperand(MI->getOperand(3));
2989 AddDefaultPred(MIB);
2993 // Copy the result from SP to virtual register.
2994 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2995 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2996 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2997 BuildMI(BB, dl, TII->get(CopyOpc))
2998 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3000 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3006 //===----------------------------------------------------------------------===//
3007 // ARM Optimization Hooks
3008 //===----------------------------------------------------------------------===//
3011 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3012 TargetLowering::DAGCombinerInfo &DCI) {
3013 SelectionDAG &DAG = DCI.DAG;
3014 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3015 EVT VT = N->getValueType(0);
3016 unsigned Opc = N->getOpcode();
3017 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3018 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3019 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3020 ISD::CondCode CC = ISD::SETCC_INVALID;
3023 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3025 SDValue CCOp = Slct.getOperand(0);
3026 if (CCOp.getOpcode() == ISD::SETCC)
3027 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3030 bool DoXform = false;
3032 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3035 if (LHS.getOpcode() == ISD::Constant &&
3036 cast<ConstantSDNode>(LHS)->isNullValue()) {
3038 } else if (CC != ISD::SETCC_INVALID &&
3039 RHS.getOpcode() == ISD::Constant &&
3040 cast<ConstantSDNode>(RHS)->isNullValue()) {
3041 std::swap(LHS, RHS);
3042 SDValue Op0 = Slct.getOperand(0);
3043 EVT OpVT = isSlctCC ? Op0.getValueType() :
3044 Op0.getOperand(0).getValueType();
3045 bool isInt = OpVT.isInteger();
3046 CC = ISD::getSetCCInverse(CC, isInt);
3048 if (!TLI.isCondCodeLegal(CC, OpVT))
3049 return SDValue(); // Inverse operator isn't legal.
3056 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3058 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3059 Slct.getOperand(0), Slct.getOperand(1), CC);
3060 SDValue CCOp = Slct.getOperand(0);
3062 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3063 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3064 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3065 CCOp, OtherOp, Result);
3070 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3071 static SDValue PerformADDCombine(SDNode *N,
3072 TargetLowering::DAGCombinerInfo &DCI) {
3073 // added by evan in r37685 with no testcase.
3074 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3076 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3077 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3078 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3079 if (Result.getNode()) return Result;
3081 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3082 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3083 if (Result.getNode()) return Result;
3089 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3090 static SDValue PerformSUBCombine(SDNode *N,
3091 TargetLowering::DAGCombinerInfo &DCI) {
3092 // added by evan in r37685 with no testcase.
3093 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3095 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3096 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3097 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3098 if (Result.getNode()) return Result;
3105 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3106 static SDValue PerformFMRRDCombine(SDNode *N,
3107 TargetLowering::DAGCombinerInfo &DCI) {
3108 // fmrrd(fmdrr x, y) -> x,y
3109 SDValue InDouble = N->getOperand(0);
3110 if (InDouble.getOpcode() == ARMISD::FMDRR)
3111 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3115 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3116 /// operand of a vector shift operation, where all the elements of the
3117 /// build_vector must have the same constant integer value.
3118 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3119 // Ignore bit_converts.
3120 while (Op.getOpcode() == ISD::BIT_CONVERT)
3121 Op = Op.getOperand(0);
3122 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3123 APInt SplatBits, SplatUndef;
3124 unsigned SplatBitSize;
3126 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3127 HasAnyUndefs, ElementBits) ||
3128 SplatBitSize > ElementBits)
3130 Cnt = SplatBits.getSExtValue();
3134 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3135 /// operand of a vector shift left operation. That value must be in the range:
3136 /// 0 <= Value < ElementBits for a left shift; or
3137 /// 0 <= Value <= ElementBits for a long left shift.
3138 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3139 assert(VT.isVector() && "vector shift count is not a vector type");
3140 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3141 if (! getVShiftImm(Op, ElementBits, Cnt))
3143 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3146 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3147 /// operand of a vector shift right operation. For a shift opcode, the value
3148 /// is positive, but for an intrinsic the value count must be negative. The
3149 /// absolute value must be in the range:
3150 /// 1 <= |Value| <= ElementBits for a right shift; or
3151 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3152 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3154 assert(VT.isVector() && "vector shift count is not a vector type");
3155 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3156 if (! getVShiftImm(Op, ElementBits, Cnt))
3160 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3163 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3164 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3165 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3168 // Don't do anything for most intrinsics.
3171 // Vector shifts: check for immediate versions and lower them.
3172 // Note: This is done during DAG combining instead of DAG legalizing because
3173 // the build_vectors for 64-bit vector element shift counts are generally
3174 // not legal, and it is hard to see their values after they get legalized to
3175 // loads from a constant pool.
3176 case Intrinsic::arm_neon_vshifts:
3177 case Intrinsic::arm_neon_vshiftu:
3178 case Intrinsic::arm_neon_vshiftls:
3179 case Intrinsic::arm_neon_vshiftlu:
3180 case Intrinsic::arm_neon_vshiftn:
3181 case Intrinsic::arm_neon_vrshifts:
3182 case Intrinsic::arm_neon_vrshiftu:
3183 case Intrinsic::arm_neon_vrshiftn:
3184 case Intrinsic::arm_neon_vqshifts:
3185 case Intrinsic::arm_neon_vqshiftu:
3186 case Intrinsic::arm_neon_vqshiftsu:
3187 case Intrinsic::arm_neon_vqshiftns:
3188 case Intrinsic::arm_neon_vqshiftnu:
3189 case Intrinsic::arm_neon_vqshiftnsu:
3190 case Intrinsic::arm_neon_vqrshiftns:
3191 case Intrinsic::arm_neon_vqrshiftnu:
3192 case Intrinsic::arm_neon_vqrshiftnsu: {
3193 EVT VT = N->getOperand(1).getValueType();
3195 unsigned VShiftOpc = 0;
3198 case Intrinsic::arm_neon_vshifts:
3199 case Intrinsic::arm_neon_vshiftu:
3200 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3201 VShiftOpc = ARMISD::VSHL;
3204 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3205 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3206 ARMISD::VSHRs : ARMISD::VSHRu);
3211 case Intrinsic::arm_neon_vshiftls:
3212 case Intrinsic::arm_neon_vshiftlu:
3213 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3215 llvm_unreachable("invalid shift count for vshll intrinsic");
3217 case Intrinsic::arm_neon_vrshifts:
3218 case Intrinsic::arm_neon_vrshiftu:
3219 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3223 case Intrinsic::arm_neon_vqshifts:
3224 case Intrinsic::arm_neon_vqshiftu:
3225 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3229 case Intrinsic::arm_neon_vqshiftsu:
3230 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3232 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3234 case Intrinsic::arm_neon_vshiftn:
3235 case Intrinsic::arm_neon_vrshiftn:
3236 case Intrinsic::arm_neon_vqshiftns:
3237 case Intrinsic::arm_neon_vqshiftnu:
3238 case Intrinsic::arm_neon_vqshiftnsu:
3239 case Intrinsic::arm_neon_vqrshiftns:
3240 case Intrinsic::arm_neon_vqrshiftnu:
3241 case Intrinsic::arm_neon_vqrshiftnsu:
3242 // Narrowing shifts require an immediate right shift.
3243 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3245 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3248 llvm_unreachable("unhandled vector shift");
3252 case Intrinsic::arm_neon_vshifts:
3253 case Intrinsic::arm_neon_vshiftu:
3254 // Opcode already set above.
3256 case Intrinsic::arm_neon_vshiftls:
3257 case Intrinsic::arm_neon_vshiftlu:
3258 if (Cnt == VT.getVectorElementType().getSizeInBits())
3259 VShiftOpc = ARMISD::VSHLLi;
3261 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3262 ARMISD::VSHLLs : ARMISD::VSHLLu);
3264 case Intrinsic::arm_neon_vshiftn:
3265 VShiftOpc = ARMISD::VSHRN; break;
3266 case Intrinsic::arm_neon_vrshifts:
3267 VShiftOpc = ARMISD::VRSHRs; break;
3268 case Intrinsic::arm_neon_vrshiftu:
3269 VShiftOpc = ARMISD::VRSHRu; break;
3270 case Intrinsic::arm_neon_vrshiftn:
3271 VShiftOpc = ARMISD::VRSHRN; break;
3272 case Intrinsic::arm_neon_vqshifts:
3273 VShiftOpc = ARMISD::VQSHLs; break;
3274 case Intrinsic::arm_neon_vqshiftu:
3275 VShiftOpc = ARMISD::VQSHLu; break;
3276 case Intrinsic::arm_neon_vqshiftsu:
3277 VShiftOpc = ARMISD::VQSHLsu; break;
3278 case Intrinsic::arm_neon_vqshiftns:
3279 VShiftOpc = ARMISD::VQSHRNs; break;
3280 case Intrinsic::arm_neon_vqshiftnu:
3281 VShiftOpc = ARMISD::VQSHRNu; break;
3282 case Intrinsic::arm_neon_vqshiftnsu:
3283 VShiftOpc = ARMISD::VQSHRNsu; break;
3284 case Intrinsic::arm_neon_vqrshiftns:
3285 VShiftOpc = ARMISD::VQRSHRNs; break;
3286 case Intrinsic::arm_neon_vqrshiftnu:
3287 VShiftOpc = ARMISD::VQRSHRNu; break;
3288 case Intrinsic::arm_neon_vqrshiftnsu:
3289 VShiftOpc = ARMISD::VQRSHRNsu; break;
3292 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3293 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3296 case Intrinsic::arm_neon_vshiftins: {
3297 EVT VT = N->getOperand(1).getValueType();
3299 unsigned VShiftOpc = 0;
3301 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3302 VShiftOpc = ARMISD::VSLI;
3303 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3304 VShiftOpc = ARMISD::VSRI;
3306 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3309 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3310 N->getOperand(1), N->getOperand(2),
3311 DAG.getConstant(Cnt, MVT::i32));
3314 case Intrinsic::arm_neon_vqrshifts:
3315 case Intrinsic::arm_neon_vqrshiftu:
3316 // No immediate versions of these to check for.
3323 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3324 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3325 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3326 /// vector element shift counts are generally not legal, and it is hard to see
3327 /// their values after they get legalized to loads from a constant pool.
3328 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3329 const ARMSubtarget *ST) {
3330 EVT VT = N->getValueType(0);
3332 // Nothing to be done for scalar shifts.
3333 if (! VT.isVector())
3336 assert(ST->hasNEON() && "unexpected vector shift");
3339 switch (N->getOpcode()) {
3340 default: llvm_unreachable("unexpected shift opcode");
3343 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3344 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3345 DAG.getConstant(Cnt, MVT::i32));
3350 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3351 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3352 ARMISD::VSHRs : ARMISD::VSHRu);
3353 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3354 DAG.getConstant(Cnt, MVT::i32));
3360 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3361 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3362 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3363 const ARMSubtarget *ST) {
3364 SDValue N0 = N->getOperand(0);
3366 // Check for sign- and zero-extensions of vector extract operations of 8-
3367 // and 16-bit vector elements. NEON supports these directly. They are
3368 // handled during DAG combining because type legalization will promote them
3369 // to 32-bit types and it is messy to recognize the operations after that.
3370 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3371 SDValue Vec = N0.getOperand(0);
3372 SDValue Lane = N0.getOperand(1);
3373 EVT VT = N->getValueType(0);
3374 EVT EltVT = N0.getValueType();
3375 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3377 if (VT == MVT::i32 &&
3378 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3379 TLI.isTypeLegal(Vec.getValueType())) {
3382 switch (N->getOpcode()) {
3383 default: llvm_unreachable("unexpected opcode");
3384 case ISD::SIGN_EXTEND:
3385 Opc = ARMISD::VGETLANEs;
3387 case ISD::ZERO_EXTEND:
3388 case ISD::ANY_EXTEND:
3389 Opc = ARMISD::VGETLANEu;
3392 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3399 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3400 DAGCombinerInfo &DCI) const {
3401 switch (N->getOpcode()) {
3403 case ISD::ADD: return PerformADDCombine(N, DCI);
3404 case ISD::SUB: return PerformSUBCombine(N, DCI);
3405 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3406 case ISD::INTRINSIC_WO_CHAIN:
3407 return PerformIntrinsicCombine(N, DCI.DAG);
3411 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3412 case ISD::SIGN_EXTEND:
3413 case ISD::ZERO_EXTEND:
3414 case ISD::ANY_EXTEND:
3415 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3420 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3421 if (!Subtarget->hasV6Ops())
3422 // Pre-v6 does not support unaligned mem access.
3424 else if (!Subtarget->hasV6Ops()) {
3425 // v6 may or may not support unaligned mem access.
3426 if (!Subtarget->isTargetDarwin())
3430 switch (VT.getSimpleVT().SimpleTy) {
3437 // FIXME: VLD1 etc with standard alignment is legal.
3441 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3446 switch (VT.getSimpleVT().SimpleTy) {
3447 default: return false;
3462 if ((V & (Scale - 1)) != 0)
3465 return V == (V & ((1LL << 5) - 1));
3468 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3469 const ARMSubtarget *Subtarget) {
3476 switch (VT.getSimpleVT().SimpleTy) {
3477 default: return false;
3482 // + imm12 or - imm8
3484 return V == (V & ((1LL << 8) - 1));
3485 return V == (V & ((1LL << 12) - 1));
3488 // Same as ARM mode. FIXME: NEON?
3489 if (!Subtarget->hasVFP2())
3494 return V == (V & ((1LL << 8) - 1));
3498 /// isLegalAddressImmediate - Return true if the integer value can be used
3499 /// as the offset of the target addressing mode for load / store of the
3501 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3502 const ARMSubtarget *Subtarget) {
3509 if (Subtarget->isThumb1Only())
3510 return isLegalT1AddressImmediate(V, VT);
3511 else if (Subtarget->isThumb2())
3512 return isLegalT2AddressImmediate(V, VT, Subtarget);
3517 switch (VT.getSimpleVT().SimpleTy) {
3518 default: return false;
3523 return V == (V & ((1LL << 12) - 1));
3526 return V == (V & ((1LL << 8) - 1));
3529 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3534 return V == (V & ((1LL << 8) - 1));
3538 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3540 int Scale = AM.Scale;
3544 switch (VT.getSimpleVT().SimpleTy) {
3545 default: return false;
3554 return Scale == 2 || Scale == 4 || Scale == 8;
3557 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3561 // Note, we allow "void" uses (basically, uses that aren't loads or
3562 // stores), because arm allows folding a scale into many arithmetic
3563 // operations. This should be made more precise and revisited later.
3565 // Allow r << imm, but the imm has to be a multiple of two.
3566 if (Scale & 1) return false;
3567 return isPowerOf2_32(Scale);
3571 /// isLegalAddressingMode - Return true if the addressing mode represented
3572 /// by AM is legal for this target, for a load/store of the specified type.
3573 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3574 const Type *Ty) const {
3575 EVT VT = getValueType(Ty, true);
3576 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3579 // Can never fold addr of global into load/store.
3584 case 0: // no scale reg, must be "r+i" or "r", or "i".
3587 if (Subtarget->isThumb1Only())
3591 // ARM doesn't support any R+R*scale+imm addr modes.
3598 if (Subtarget->isThumb2())
3599 return isLegalT2ScaledAddressingMode(AM, VT);
3601 int Scale = AM.Scale;
3602 switch (VT.getSimpleVT().SimpleTy) {
3603 default: return false;
3607 if (Scale < 0) Scale = -Scale;
3611 return isPowerOf2_32(Scale & ~1);
3615 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3620 // Note, we allow "void" uses (basically, uses that aren't loads or
3621 // stores), because arm allows folding a scale into many arithmetic
3622 // operations. This should be made more precise and revisited later.
3624 // Allow r << imm, but the imm has to be a multiple of two.
3625 if (Scale & 1) return false;
3626 return isPowerOf2_32(Scale);
3633 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3634 bool isSEXTLoad, SDValue &Base,
3635 SDValue &Offset, bool &isInc,
3636 SelectionDAG &DAG) {
3637 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3640 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3642 Base = Ptr->getOperand(0);
3643 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3644 int RHSC = (int)RHS->getZExtValue();
3645 if (RHSC < 0 && RHSC > -256) {
3646 assert(Ptr->getOpcode() == ISD::ADD);
3648 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3652 isInc = (Ptr->getOpcode() == ISD::ADD);
3653 Offset = Ptr->getOperand(1);
3655 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3657 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3658 int RHSC = (int)RHS->getZExtValue();
3659 if (RHSC < 0 && RHSC > -0x1000) {
3660 assert(Ptr->getOpcode() == ISD::ADD);
3662 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3663 Base = Ptr->getOperand(0);
3668 if (Ptr->getOpcode() == ISD::ADD) {
3670 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3671 if (ShOpcVal != ARM_AM::no_shift) {
3672 Base = Ptr->getOperand(1);
3673 Offset = Ptr->getOperand(0);
3675 Base = Ptr->getOperand(0);
3676 Offset = Ptr->getOperand(1);
3681 isInc = (Ptr->getOpcode() == ISD::ADD);
3682 Base = Ptr->getOperand(0);
3683 Offset = Ptr->getOperand(1);
3687 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3691 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3692 bool isSEXTLoad, SDValue &Base,
3693 SDValue &Offset, bool &isInc,
3694 SelectionDAG &DAG) {
3695 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3698 Base = Ptr->getOperand(0);
3699 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3700 int RHSC = (int)RHS->getZExtValue();
3701 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3702 assert(Ptr->getOpcode() == ISD::ADD);
3704 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3706 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3707 isInc = Ptr->getOpcode() == ISD::ADD;
3708 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3716 /// getPreIndexedAddressParts - returns true by value, base pointer and
3717 /// offset pointer and addressing mode by reference if the node's address
3718 /// can be legally represented as pre-indexed load / store address.
3720 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3722 ISD::MemIndexedMode &AM,
3723 SelectionDAG &DAG) const {
3724 if (Subtarget->isThumb1Only())
3729 bool isSEXTLoad = false;
3730 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3731 Ptr = LD->getBasePtr();
3732 VT = LD->getMemoryVT();
3733 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3734 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3735 Ptr = ST->getBasePtr();
3736 VT = ST->getMemoryVT();
3741 bool isLegal = false;
3742 if (Subtarget->isThumb2())
3743 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3744 Offset, isInc, DAG);
3746 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3747 Offset, isInc, DAG);
3751 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3755 /// getPostIndexedAddressParts - returns true by value, base pointer and
3756 /// offset pointer and addressing mode by reference if this node can be
3757 /// combined with a load / store to form a post-indexed load / store.
3758 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3761 ISD::MemIndexedMode &AM,
3762 SelectionDAG &DAG) const {
3763 if (Subtarget->isThumb1Only())
3768 bool isSEXTLoad = false;
3769 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3770 VT = LD->getMemoryVT();
3771 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3772 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3773 VT = ST->getMemoryVT();
3778 bool isLegal = false;
3779 if (Subtarget->isThumb2())
3780 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3783 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3788 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3792 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3796 const SelectionDAG &DAG,
3797 unsigned Depth) const {
3798 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3799 switch (Op.getOpcode()) {
3801 case ARMISD::CMOV: {
3802 // Bits are known zero/one if known on the LHS and RHS.
3803 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3804 if (KnownZero == 0 && KnownOne == 0) return;
3806 APInt KnownZeroRHS, KnownOneRHS;
3807 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3808 KnownZeroRHS, KnownOneRHS, Depth+1);
3809 KnownZero &= KnownZeroRHS;
3810 KnownOne &= KnownOneRHS;
3816 //===----------------------------------------------------------------------===//
3817 // ARM Inline Assembly Support
3818 //===----------------------------------------------------------------------===//
3820 /// getConstraintType - Given a constraint letter, return the type of
3821 /// constraint it is for this target.
3822 ARMTargetLowering::ConstraintType
3823 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3824 if (Constraint.size() == 1) {
3825 switch (Constraint[0]) {
3827 case 'l': return C_RegisterClass;
3828 case 'w': return C_RegisterClass;
3831 return TargetLowering::getConstraintType(Constraint);
3834 std::pair<unsigned, const TargetRegisterClass*>
3835 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3837 if (Constraint.size() == 1) {
3838 // GCC RS6000 Constraint Letters
3839 switch (Constraint[0]) {
3841 if (Subtarget->isThumb1Only())
3842 return std::make_pair(0U, ARM::tGPRRegisterClass);
3844 return std::make_pair(0U, ARM::GPRRegisterClass);
3846 return std::make_pair(0U, ARM::GPRRegisterClass);
3849 return std::make_pair(0U, ARM::SPRRegisterClass);
3851 return std::make_pair(0U, ARM::DPRRegisterClass);
3855 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3858 std::vector<unsigned> ARMTargetLowering::
3859 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3861 if (Constraint.size() != 1)
3862 return std::vector<unsigned>();
3864 switch (Constraint[0]) { // GCC ARM Constraint Letters
3867 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3868 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3871 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3872 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3873 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3874 ARM::R12, ARM::LR, 0);
3877 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3878 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3879 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3880 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3881 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3882 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3883 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3884 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3886 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3887 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3888 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3889 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3893 return std::vector<unsigned>();
3896 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3897 /// vector. If it is invalid, don't add anything to Ops.
3898 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3901 std::vector<SDValue>&Ops,
3902 SelectionDAG &DAG) const {
3903 SDValue Result(0, 0);
3905 switch (Constraint) {
3907 case 'I': case 'J': case 'K': case 'L':
3908 case 'M': case 'N': case 'O':
3909 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3913 int64_t CVal64 = C->getSExtValue();
3914 int CVal = (int) CVal64;
3915 // None of these constraints allow values larger than 32 bits. Check
3916 // that the value fits in an int.
3920 switch (Constraint) {
3922 if (Subtarget->isThumb1Only()) {
3923 // This must be a constant between 0 and 255, for ADD
3925 if (CVal >= 0 && CVal <= 255)
3927 } else if (Subtarget->isThumb2()) {
3928 // A constant that can be used as an immediate value in a
3929 // data-processing instruction.
3930 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3933 // A constant that can be used as an immediate value in a
3934 // data-processing instruction.
3935 if (ARM_AM::getSOImmVal(CVal) != -1)
3941 if (Subtarget->isThumb()) { // FIXME thumb2
3942 // This must be a constant between -255 and -1, for negated ADD
3943 // immediates. This can be used in GCC with an "n" modifier that
3944 // prints the negated value, for use with SUB instructions. It is
3945 // not useful otherwise but is implemented for compatibility.
3946 if (CVal >= -255 && CVal <= -1)
3949 // This must be a constant between -4095 and 4095. It is not clear
3950 // what this constraint is intended for. Implemented for
3951 // compatibility with GCC.
3952 if (CVal >= -4095 && CVal <= 4095)
3958 if (Subtarget->isThumb1Only()) {
3959 // A 32-bit value where only one byte has a nonzero value. Exclude
3960 // zero to match GCC. This constraint is used by GCC internally for
3961 // constants that can be loaded with a move/shift combination.
3962 // It is not useful otherwise but is implemented for compatibility.
3963 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3965 } else if (Subtarget->isThumb2()) {
3966 // A constant whose bitwise inverse can be used as an immediate
3967 // value in a data-processing instruction. This can be used in GCC
3968 // with a "B" modifier that prints the inverted value, for use with
3969 // BIC and MVN instructions. It is not useful otherwise but is
3970 // implemented for compatibility.
3971 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3974 // A constant whose bitwise inverse can be used as an immediate
3975 // value in a data-processing instruction. This can be used in GCC
3976 // with a "B" modifier that prints the inverted value, for use with
3977 // BIC and MVN instructions. It is not useful otherwise but is
3978 // implemented for compatibility.
3979 if (ARM_AM::getSOImmVal(~CVal) != -1)
3985 if (Subtarget->isThumb1Only()) {
3986 // This must be a constant between -7 and 7,
3987 // for 3-operand ADD/SUB immediate instructions.
3988 if (CVal >= -7 && CVal < 7)
3990 } else if (Subtarget->isThumb2()) {
3991 // A constant whose negation can be used as an immediate value in a
3992 // data-processing instruction. This can be used in GCC with an "n"
3993 // modifier that prints the negated value, for use with SUB
3994 // instructions. It is not useful otherwise but is implemented for
3996 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3999 // A constant whose negation can be used as an immediate value in a
4000 // data-processing instruction. This can be used in GCC with an "n"
4001 // modifier that prints the negated value, for use with SUB
4002 // instructions. It is not useful otherwise but is implemented for
4004 if (ARM_AM::getSOImmVal(-CVal) != -1)
4010 if (Subtarget->isThumb()) { // FIXME thumb2
4011 // This must be a multiple of 4 between 0 and 1020, for
4012 // ADD sp + immediate.
4013 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4016 // A power of two or a constant between 0 and 32. This is used in
4017 // GCC for the shift amount on shifted register operands, but it is
4018 // useful in general for any shift amounts.
4019 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4025 if (Subtarget->isThumb()) { // FIXME thumb2
4026 // This must be a constant between 0 and 31, for shift amounts.
4027 if (CVal >= 0 && CVal <= 31)
4033 if (Subtarget->isThumb()) { // FIXME thumb2
4034 // This must be a multiple of 4 between -508 and 508, for
4035 // ADD/SUB sp = sp + immediate.
4036 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4041 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4045 if (Result.getNode()) {
4046 Ops.push_back(Result);
4049 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,