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"
45 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
46 CCValAssign::LocInfo &LocInfo,
47 ISD::ArgFlagsTy &ArgFlags,
49 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
53 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
54 CCValAssign::LocInfo &LocInfo,
55 ISD::ArgFlagsTy &ArgFlags,
57 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
58 CCValAssign::LocInfo &LocInfo,
59 ISD::ArgFlagsTy &ArgFlags,
62 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
63 EVT PromotedBitwiseVT) {
64 if (VT != PromotedLdStVT) {
65 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
66 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
67 PromotedLdStVT.getSimpleVT());
69 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
70 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
71 PromotedLdStVT.getSimpleVT());
74 EVT ElemTy = VT.getVectorElementType();
75 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
76 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
77 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
78 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
79 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
80 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
81 setOperationAction(ISD::SCALAR_TO_VECTOR, 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 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
241 setTargetDAGCombine(ISD::SHL);
242 setTargetDAGCombine(ISD::SRL);
243 setTargetDAGCombine(ISD::SRA);
244 setTargetDAGCombine(ISD::SIGN_EXTEND);
245 setTargetDAGCombine(ISD::ZERO_EXTEND);
246 setTargetDAGCombine(ISD::ANY_EXTEND);
249 computeRegisterProperties();
251 // ARM does not have f32 extending load.
252 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
254 // ARM does not have i1 sign extending load.
255 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
257 // ARM supports all 4 flavors of integer indexed load / store.
258 if (!Subtarget->isThumb1Only()) {
259 for (unsigned im = (unsigned)ISD::PRE_INC;
260 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
261 setIndexedLoadAction(im, MVT::i1, Legal);
262 setIndexedLoadAction(im, MVT::i8, Legal);
263 setIndexedLoadAction(im, MVT::i16, Legal);
264 setIndexedLoadAction(im, MVT::i32, Legal);
265 setIndexedStoreAction(im, MVT::i1, Legal);
266 setIndexedStoreAction(im, MVT::i8, Legal);
267 setIndexedStoreAction(im, MVT::i16, Legal);
268 setIndexedStoreAction(im, MVT::i32, Legal);
272 // i64 operation support.
273 if (Subtarget->isThumb1Only()) {
274 setOperationAction(ISD::MUL, MVT::i64, Expand);
275 setOperationAction(ISD::MULHU, MVT::i32, Expand);
276 setOperationAction(ISD::MULHS, MVT::i32, Expand);
277 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
278 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
280 setOperationAction(ISD::MUL, MVT::i64, Expand);
281 setOperationAction(ISD::MULHU, MVT::i32, Expand);
282 if (!Subtarget->hasV6Ops())
283 setOperationAction(ISD::MULHS, MVT::i32, Expand);
285 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
286 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
287 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
288 setOperationAction(ISD::SRL, MVT::i64, Custom);
289 setOperationAction(ISD::SRA, MVT::i64, Custom);
291 // ARM does not have ROTL.
292 setOperationAction(ISD::ROTL, MVT::i32, Expand);
293 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
294 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
295 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
296 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
298 // Only ARMv6 has BSWAP.
299 if (!Subtarget->hasV6Ops())
300 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
302 // These are expanded into libcalls.
303 setOperationAction(ISD::SDIV, MVT::i32, Expand);
304 setOperationAction(ISD::UDIV, MVT::i32, Expand);
305 setOperationAction(ISD::SREM, MVT::i32, Expand);
306 setOperationAction(ISD::UREM, MVT::i32, Expand);
307 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
308 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
310 // Support label based line numbers.
311 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
312 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
314 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
315 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
316 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
317 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
319 // Use the default implementation.
320 setOperationAction(ISD::VASTART, MVT::Other, Custom);
321 setOperationAction(ISD::VAARG, MVT::Other, Expand);
322 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
323 setOperationAction(ISD::VAEND, MVT::Other, Expand);
324 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
325 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
326 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
327 // FIXME: Shouldn't need this, since no register is used, but the legalizer
328 // doesn't yet know how to not do that for SjLj.
329 setExceptionSelectorRegister(ARM::R0);
330 if (Subtarget->isThumb())
331 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
333 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
334 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
336 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
337 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
338 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
340 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
342 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
343 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
344 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
346 // We want to custom lower some of our intrinsics.
347 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
348 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
349 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
351 setOperationAction(ISD::SETCC, MVT::i32, Expand);
352 setOperationAction(ISD::SETCC, MVT::f32, Expand);
353 setOperationAction(ISD::SETCC, MVT::f64, Expand);
354 setOperationAction(ISD::SELECT, MVT::i32, Expand);
355 setOperationAction(ISD::SELECT, MVT::f32, Expand);
356 setOperationAction(ISD::SELECT, MVT::f64, Expand);
357 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
358 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
359 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
361 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
362 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
363 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
364 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
365 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
367 // We don't support sin/cos/fmod/copysign/pow
368 setOperationAction(ISD::FSIN, MVT::f64, Expand);
369 setOperationAction(ISD::FSIN, MVT::f32, Expand);
370 setOperationAction(ISD::FCOS, MVT::f32, Expand);
371 setOperationAction(ISD::FCOS, MVT::f64, Expand);
372 setOperationAction(ISD::FREM, MVT::f64, Expand);
373 setOperationAction(ISD::FREM, MVT::f32, Expand);
374 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
375 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
376 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
378 setOperationAction(ISD::FPOW, MVT::f64, Expand);
379 setOperationAction(ISD::FPOW, MVT::f32, Expand);
381 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
382 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
383 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
384 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
385 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
386 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
389 // We have target-specific dag combine patterns for the following nodes:
390 // ARMISD::FMRRD - No need to call setTargetDAGCombine
391 setTargetDAGCombine(ISD::ADD);
392 setTargetDAGCombine(ISD::SUB);
394 setStackPointerRegisterToSaveRestore(ARM::SP);
395 setSchedulingPreference(SchedulingForRegPressure);
397 // FIXME: If-converter should use instruction latency to determine
398 // profitability rather than relying on fixed limits.
399 if (Subtarget->getCPUString() == "generic") {
400 // Generic (and overly aggressive) if-conversion limits.
401 setIfCvtBlockSizeLimit(10);
402 setIfCvtDupBlockSizeLimit(2);
403 } else if (Subtarget->hasV6Ops()) {
404 setIfCvtBlockSizeLimit(2);
405 setIfCvtDupBlockSizeLimit(1);
407 setIfCvtBlockSizeLimit(3);
408 setIfCvtDupBlockSizeLimit(2);
411 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
412 // Do not enable CodePlacementOpt for now: it currently runs after the
413 // ARMConstantIslandPass and messes up branch relaxation and placement
414 // of constant islands.
415 // benefitFromCodePlacementOpt = true;
418 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
421 case ARMISD::Wrapper: return "ARMISD::Wrapper";
422 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
423 case ARMISD::CALL: return "ARMISD::CALL";
424 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
425 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
426 case ARMISD::tCALL: return "ARMISD::tCALL";
427 case ARMISD::BRCOND: return "ARMISD::BRCOND";
428 case ARMISD::BR_JT: return "ARMISD::BR_JT";
429 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
430 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
431 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
432 case ARMISD::CMP: return "ARMISD::CMP";
433 case ARMISD::CMPZ: return "ARMISD::CMPZ";
434 case ARMISD::CMPFP: return "ARMISD::CMPFP";
435 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
436 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
437 case ARMISD::CMOV: return "ARMISD::CMOV";
438 case ARMISD::CNEG: return "ARMISD::CNEG";
440 case ARMISD::FTOSI: return "ARMISD::FTOSI";
441 case ARMISD::FTOUI: return "ARMISD::FTOUI";
442 case ARMISD::SITOF: return "ARMISD::SITOF";
443 case ARMISD::UITOF: return "ARMISD::UITOF";
445 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
446 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
447 case ARMISD::RRX: return "ARMISD::RRX";
449 case ARMISD::FMRRD: return "ARMISD::FMRRD";
450 case ARMISD::FMDRR: return "ARMISD::FMDRR";
452 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
454 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
456 case ARMISD::VCEQ: return "ARMISD::VCEQ";
457 case ARMISD::VCGE: return "ARMISD::VCGE";
458 case ARMISD::VCGEU: return "ARMISD::VCGEU";
459 case ARMISD::VCGT: return "ARMISD::VCGT";
460 case ARMISD::VCGTU: return "ARMISD::VCGTU";
461 case ARMISD::VTST: return "ARMISD::VTST";
463 case ARMISD::VSHL: return "ARMISD::VSHL";
464 case ARMISD::VSHRs: return "ARMISD::VSHRs";
465 case ARMISD::VSHRu: return "ARMISD::VSHRu";
466 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
467 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
468 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
469 case ARMISD::VSHRN: return "ARMISD::VSHRN";
470 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
471 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
472 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
473 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
474 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
475 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
476 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
477 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
478 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
479 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
480 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
481 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
482 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
483 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
484 case ARMISD::VDUP: return "ARMISD::VDUP";
485 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
486 case ARMISD::VLD2D: return "ARMISD::VLD2D";
487 case ARMISD::VLD3D: return "ARMISD::VLD3D";
488 case ARMISD::VLD4D: return "ARMISD::VLD4D";
489 case ARMISD::VST2D: return "ARMISD::VST2D";
490 case ARMISD::VST3D: return "ARMISD::VST3D";
491 case ARMISD::VST4D: return "ARMISD::VST4D";
492 case ARMISD::VEXT: return "ARMISD::VEXT";
493 case ARMISD::VREV64: return "ARMISD::VREV64";
494 case ARMISD::VREV32: return "ARMISD::VREV32";
495 case ARMISD::VREV16: return "ARMISD::VREV16";
496 case ARMISD::VZIP32: return "ARMISD::VZIP32";
497 case ARMISD::VZIP16: return "ARMISD::VZIP16";
498 case ARMISD::VZIP8: return "ARMISD::VZIP8";
499 case ARMISD::VUZP32: return "ARMISD::VUZP32";
500 case ARMISD::VUZP16: return "ARMISD::VUZP16";
501 case ARMISD::VUZP8: return "ARMISD::VUZP8";
502 case ARMISD::VTRN32: return "ARMISD::VTRN32";
503 case ARMISD::VTRN16: return "ARMISD::VTRN16";
504 case ARMISD::VTRN8: return "ARMISD::VTRN8";
508 /// getFunctionAlignment - Return the Log2 alignment of this function.
509 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
510 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
513 //===----------------------------------------------------------------------===//
515 //===----------------------------------------------------------------------===//
517 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
518 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
520 default: llvm_unreachable("Unknown condition code!");
521 case ISD::SETNE: return ARMCC::NE;
522 case ISD::SETEQ: return ARMCC::EQ;
523 case ISD::SETGT: return ARMCC::GT;
524 case ISD::SETGE: return ARMCC::GE;
525 case ISD::SETLT: return ARMCC::LT;
526 case ISD::SETLE: return ARMCC::LE;
527 case ISD::SETUGT: return ARMCC::HI;
528 case ISD::SETUGE: return ARMCC::HS;
529 case ISD::SETULT: return ARMCC::LO;
530 case ISD::SETULE: return ARMCC::LS;
534 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
535 /// returns true if the operands should be inverted to form the proper
537 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
538 ARMCC::CondCodes &CondCode2) {
540 CondCode2 = ARMCC::AL;
542 default: llvm_unreachable("Unknown FP condition!");
544 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
546 case ISD::SETOGT: CondCode = ARMCC::GT; break;
548 case ISD::SETOGE: CondCode = ARMCC::GE; break;
549 case ISD::SETOLT: CondCode = ARMCC::MI; break;
550 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
551 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
552 case ISD::SETO: CondCode = ARMCC::VC; break;
553 case ISD::SETUO: CondCode = ARMCC::VS; break;
554 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
555 case ISD::SETUGT: CondCode = ARMCC::HI; break;
556 case ISD::SETUGE: CondCode = ARMCC::PL; break;
558 case ISD::SETULT: CondCode = ARMCC::LT; break;
560 case ISD::SETULE: CondCode = ARMCC::LE; break;
562 case ISD::SETUNE: CondCode = ARMCC::NE; break;
567 //===----------------------------------------------------------------------===//
568 // Calling Convention Implementation
569 //===----------------------------------------------------------------------===//
571 #include "ARMGenCallingConv.inc"
573 // APCS f64 is in register pairs, possibly split to stack
574 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
575 CCValAssign::LocInfo &LocInfo,
576 CCState &State, bool CanFail) {
577 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
579 // Try to get the first register.
580 if (unsigned Reg = State.AllocateReg(RegList, 4))
581 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
583 // For the 2nd half of a v2f64, do not fail.
587 // Put the whole thing on the stack.
588 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
589 State.AllocateStack(8, 4),
594 // Try to get the second register.
595 if (unsigned Reg = State.AllocateReg(RegList, 4))
596 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
598 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
599 State.AllocateStack(4, 4),
604 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
605 CCValAssign::LocInfo &LocInfo,
606 ISD::ArgFlagsTy &ArgFlags,
608 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
610 if (LocVT == MVT::v2f64 &&
611 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
613 return true; // we handled it
616 // AAPCS f64 is in aligned register pairs
617 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
618 CCValAssign::LocInfo &LocInfo,
619 CCState &State, bool CanFail) {
620 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
621 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
623 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
625 // For the 2nd half of a v2f64, do not just fail.
629 // Put the whole thing on the stack.
630 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
631 State.AllocateStack(8, 8),
637 for (i = 0; i < 2; ++i)
638 if (HiRegList[i] == Reg)
641 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
642 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
647 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
648 CCValAssign::LocInfo &LocInfo,
649 ISD::ArgFlagsTy &ArgFlags,
651 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
653 if (LocVT == MVT::v2f64 &&
654 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
656 return true; // we handled it
659 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
660 CCValAssign::LocInfo &LocInfo, CCState &State) {
661 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
662 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
664 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
666 return false; // we didn't handle it
669 for (i = 0; i < 2; ++i)
670 if (HiRegList[i] == Reg)
673 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
674 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
679 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
680 CCValAssign::LocInfo &LocInfo,
681 ISD::ArgFlagsTy &ArgFlags,
683 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
685 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
687 return true; // we handled it
690 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
691 CCValAssign::LocInfo &LocInfo,
692 ISD::ArgFlagsTy &ArgFlags,
694 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
698 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
699 /// given CallingConvention value.
700 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC,
702 bool isVarArg) const {
705 llvm_unreachable("Unsupported calling convention");
707 case CallingConv::Fast:
708 // Use target triple & subtarget features to do actual dispatch.
709 if (Subtarget->isAAPCS_ABI()) {
710 if (Subtarget->hasVFP2() &&
711 FloatABIType == FloatABI::Hard && !isVarArg)
712 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
714 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
716 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
717 case CallingConv::ARM_AAPCS_VFP:
718 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
719 case CallingConv::ARM_AAPCS:
720 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
721 case CallingConv::ARM_APCS:
722 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
726 /// LowerCallResult - Lower the result values of a call into the
727 /// appropriate copies out of appropriate physical registers.
729 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
730 unsigned CallConv, bool isVarArg,
731 const SmallVectorImpl<ISD::InputArg> &Ins,
732 DebugLoc dl, SelectionDAG &DAG,
733 SmallVectorImpl<SDValue> &InVals) {
735 // Assign locations to each value returned by this call.
736 SmallVector<CCValAssign, 16> RVLocs;
737 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
738 RVLocs, *DAG.getContext());
739 CCInfo.AnalyzeCallResult(Ins,
740 CCAssignFnForNode(CallConv, /* Return*/ true,
743 // Copy all of the result registers out of their specified physreg.
744 for (unsigned i = 0; i != RVLocs.size(); ++i) {
745 CCValAssign VA = RVLocs[i];
748 if (VA.needsCustom()) {
749 // Handle f64 or half of a v2f64.
750 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
752 Chain = Lo.getValue(1);
753 InFlag = Lo.getValue(2);
754 VA = RVLocs[++i]; // skip ahead to next loc
755 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
757 Chain = Hi.getValue(1);
758 InFlag = Hi.getValue(2);
759 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
761 if (VA.getLocVT() == MVT::v2f64) {
762 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
763 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
764 DAG.getConstant(0, MVT::i32));
766 VA = RVLocs[++i]; // skip ahead to next loc
767 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
768 Chain = Lo.getValue(1);
769 InFlag = Lo.getValue(2);
770 VA = RVLocs[++i]; // skip ahead to next loc
771 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
772 Chain = Hi.getValue(1);
773 InFlag = Hi.getValue(2);
774 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
775 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
776 DAG.getConstant(1, MVT::i32));
779 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
781 Chain = Val.getValue(1);
782 InFlag = Val.getValue(2);
785 switch (VA.getLocInfo()) {
786 default: llvm_unreachable("Unknown loc info!");
787 case CCValAssign::Full: break;
788 case CCValAssign::BCvt:
789 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
793 InVals.push_back(Val);
799 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
800 /// by "Src" to address "Dst" of size "Size". Alignment information is
801 /// specified by the specific parameter attribute. The copy will be passed as
802 /// a byval function parameter.
803 /// Sometimes what we are copying is the end of a larger object, the part that
804 /// does not fit in registers.
806 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
807 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
809 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
810 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
811 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
814 /// LowerMemOpCallTo - Store the argument to the stack.
816 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
817 SDValue StackPtr, SDValue Arg,
818 DebugLoc dl, SelectionDAG &DAG,
819 const CCValAssign &VA,
820 ISD::ArgFlagsTy Flags) {
821 unsigned LocMemOffset = VA.getLocMemOffset();
822 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
823 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
824 if (Flags.isByVal()) {
825 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
827 return DAG.getStore(Chain, dl, Arg, PtrOff,
828 PseudoSourceValue::getStack(), LocMemOffset);
831 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
832 SDValue Chain, SDValue &Arg,
833 RegsToPassVector &RegsToPass,
834 CCValAssign &VA, CCValAssign &NextVA,
836 SmallVector<SDValue, 8> &MemOpChains,
837 ISD::ArgFlagsTy Flags) {
839 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
840 DAG.getVTList(MVT::i32, MVT::i32), Arg);
841 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
843 if (NextVA.isRegLoc())
844 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
846 assert(NextVA.isMemLoc());
847 if (StackPtr.getNode() == 0)
848 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
850 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
856 /// LowerCall - Lowering a call into a callseq_start <-
857 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
860 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
861 unsigned CallConv, bool isVarArg,
863 const SmallVectorImpl<ISD::OutputArg> &Outs,
864 const SmallVectorImpl<ISD::InputArg> &Ins,
865 DebugLoc dl, SelectionDAG &DAG,
866 SmallVectorImpl<SDValue> &InVals) {
868 // Analyze operands of the call, assigning locations to each operand.
869 SmallVector<CCValAssign, 16> ArgLocs;
870 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
872 CCInfo.AnalyzeCallOperands(Outs,
873 CCAssignFnForNode(CallConv, /* Return*/ false,
876 // Get a count of how many bytes are to be pushed on the stack.
877 unsigned NumBytes = CCInfo.getNextStackOffset();
879 // Adjust the stack pointer for the new arguments...
880 // These operations are automatically eliminated by the prolog/epilog pass
881 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
883 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
885 RegsToPassVector RegsToPass;
886 SmallVector<SDValue, 8> MemOpChains;
888 // Walk the register/memloc assignments, inserting copies/loads. In the case
889 // of tail call optimization, arguments are handled later.
890 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
893 CCValAssign &VA = ArgLocs[i];
894 SDValue Arg = Outs[realArgIdx].Val;
895 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
897 // Promote the value if needed.
898 switch (VA.getLocInfo()) {
899 default: llvm_unreachable("Unknown loc info!");
900 case CCValAssign::Full: break;
901 case CCValAssign::SExt:
902 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
904 case CCValAssign::ZExt:
905 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
907 case CCValAssign::AExt:
908 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
910 case CCValAssign::BCvt:
911 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
915 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
916 if (VA.needsCustom()) {
917 if (VA.getLocVT() == MVT::v2f64) {
918 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
919 DAG.getConstant(0, MVT::i32));
920 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
921 DAG.getConstant(1, MVT::i32));
923 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
924 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
926 VA = ArgLocs[++i]; // skip ahead to next loc
928 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
929 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
931 assert(VA.isMemLoc());
932 if (StackPtr.getNode() == 0)
933 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
935 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
936 dl, DAG, VA, Flags));
939 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
940 StackPtr, MemOpChains, Flags);
942 } else if (VA.isRegLoc()) {
943 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
945 assert(VA.isMemLoc());
946 if (StackPtr.getNode() == 0)
947 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
949 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
950 dl, DAG, VA, Flags));
954 if (!MemOpChains.empty())
955 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
956 &MemOpChains[0], MemOpChains.size());
958 // Build a sequence of copy-to-reg nodes chained together with token chain
959 // and flag operands which copy the outgoing args into the appropriate regs.
961 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
962 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
963 RegsToPass[i].second, InFlag);
964 InFlag = Chain.getValue(1);
967 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
968 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
969 // node so that legalize doesn't hack it.
970 bool isDirect = false;
971 bool isARMFunc = false;
972 bool isLocalARMFunc = false;
973 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
974 GlobalValue *GV = G->getGlobal();
976 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
977 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
978 getTargetMachine().getRelocationModel() != Reloc::Static;
979 isARMFunc = !Subtarget->isThumb() || isStub;
980 // ARM call to a local ARM function is predicable.
981 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
982 // tBX takes a register source operand.
983 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
984 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
986 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
987 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
988 Callee = DAG.getLoad(getPointerTy(), dl,
989 DAG.getEntryNode(), CPAddr, NULL, 0);
990 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
991 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
992 getPointerTy(), Callee, PICLabel);
994 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
995 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
997 bool isStub = Subtarget->isTargetDarwin() &&
998 getTargetMachine().getRelocationModel() != Reloc::Static;
999 isARMFunc = !Subtarget->isThumb() || isStub;
1000 // tBX takes a register source operand.
1001 const char *Sym = S->getSymbol();
1002 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1003 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1004 Sym, ARMPCLabelIndex,
1006 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1007 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1008 Callee = DAG.getLoad(getPointerTy(), dl,
1009 DAG.getEntryNode(), CPAddr, NULL, 0);
1010 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1011 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1012 getPointerTy(), Callee, PICLabel);
1014 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1017 // FIXME: handle tail calls differently.
1019 if (Subtarget->isThumb()) {
1020 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1021 CallOpc = ARMISD::CALL_NOLINK;
1023 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1025 CallOpc = (isDirect || Subtarget->hasV5TOps())
1026 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1027 : ARMISD::CALL_NOLINK;
1029 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1030 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1031 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1032 InFlag = Chain.getValue(1);
1035 std::vector<SDValue> Ops;
1036 Ops.push_back(Chain);
1037 Ops.push_back(Callee);
1039 // Add argument registers to the end of the list so that they are known live
1041 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1042 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1043 RegsToPass[i].second.getValueType()));
1045 if (InFlag.getNode())
1046 Ops.push_back(InFlag);
1047 // Returns a chain and a flag for retval copy to use.
1048 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1049 &Ops[0], Ops.size());
1050 InFlag = Chain.getValue(1);
1052 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1053 DAG.getIntPtrConstant(0, true), InFlag);
1055 InFlag = Chain.getValue(1);
1057 // Handle result values, copying them out of physregs into vregs that we
1059 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1064 ARMTargetLowering::LowerReturn(SDValue Chain,
1065 unsigned CallConv, bool isVarArg,
1066 const SmallVectorImpl<ISD::OutputArg> &Outs,
1067 DebugLoc dl, SelectionDAG &DAG) {
1069 // CCValAssign - represent the assignment of the return value to a location.
1070 SmallVector<CCValAssign, 16> RVLocs;
1072 // CCState - Info about the registers and stack slots.
1073 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1076 // Analyze outgoing return values.
1077 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1080 // If this is the first return lowered for this function, add
1081 // the regs to the liveout set for the function.
1082 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1083 for (unsigned i = 0; i != RVLocs.size(); ++i)
1084 if (RVLocs[i].isRegLoc())
1085 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1090 // Copy the result values into the output registers.
1091 for (unsigned i = 0, realRVLocIdx = 0;
1093 ++i, ++realRVLocIdx) {
1094 CCValAssign &VA = RVLocs[i];
1095 assert(VA.isRegLoc() && "Can only return in registers!");
1097 SDValue Arg = Outs[realRVLocIdx].Val;
1099 switch (VA.getLocInfo()) {
1100 default: llvm_unreachable("Unknown loc info!");
1101 case CCValAssign::Full: break;
1102 case CCValAssign::BCvt:
1103 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1107 if (VA.needsCustom()) {
1108 if (VA.getLocVT() == MVT::v2f64) {
1109 // Extract the first half and return it in two registers.
1110 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1111 DAG.getConstant(0, MVT::i32));
1112 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1113 DAG.getVTList(MVT::i32, MVT::i32), Half);
1115 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1116 Flag = Chain.getValue(1);
1117 VA = RVLocs[++i]; // skip ahead to next loc
1118 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1119 HalfGPRs.getValue(1), Flag);
1120 Flag = Chain.getValue(1);
1121 VA = RVLocs[++i]; // skip ahead to next loc
1123 // Extract the 2nd half and fall through to handle it as an f64 value.
1124 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1125 DAG.getConstant(1, MVT::i32));
1127 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1129 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1130 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1131 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1132 Flag = Chain.getValue(1);
1133 VA = RVLocs[++i]; // skip ahead to next loc
1134 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1137 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1139 // Guarantee that all emitted copies are
1140 // stuck together, avoiding something bad.
1141 Flag = Chain.getValue(1);
1146 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1148 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1153 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1154 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1155 // one of the above mentioned nodes. It has to be wrapped because otherwise
1156 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1157 // be used to form addressing mode. These wrapped nodes will be selected
1159 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1160 EVT PtrVT = Op.getValueType();
1161 // FIXME there is no actual debug info here
1162 DebugLoc dl = Op.getDebugLoc();
1163 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1165 if (CP->isMachineConstantPoolEntry())
1166 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1167 CP->getAlignment());
1169 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1170 CP->getAlignment());
1171 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1174 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1176 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1177 SelectionDAG &DAG) {
1178 DebugLoc dl = GA->getDebugLoc();
1179 EVT PtrVT = getPointerTy();
1180 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1181 ARMConstantPoolValue *CPV =
1182 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1183 PCAdj, "tlsgd", true);
1184 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1185 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1186 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1187 SDValue Chain = Argument.getValue(1);
1189 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1190 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1192 // call __tls_get_addr.
1195 Entry.Node = Argument;
1196 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1197 Args.push_back(Entry);
1198 // FIXME: is there useful debug info available here?
1199 std::pair<SDValue, SDValue> CallResult =
1200 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1201 false, false, false, false,
1202 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1203 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1204 return CallResult.first;
1207 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1208 // "local exec" model.
1210 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1211 SelectionDAG &DAG) {
1212 GlobalValue *GV = GA->getGlobal();
1213 DebugLoc dl = GA->getDebugLoc();
1215 SDValue Chain = DAG.getEntryNode();
1216 EVT PtrVT = getPointerTy();
1217 // Get the Thread Pointer
1218 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1220 if (GV->isDeclaration()) {
1221 // initial exec model
1222 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1223 ARMConstantPoolValue *CPV =
1224 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1225 PCAdj, "gottpoff", true);
1226 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1227 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1228 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1229 Chain = Offset.getValue(1);
1231 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1232 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1234 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1237 ARMConstantPoolValue *CPV =
1238 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
1239 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1240 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1241 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1244 // The address of the thread local variable is the add of the thread
1245 // pointer with the offset of the variable.
1246 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1250 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1251 // TODO: implement the "local dynamic" model
1252 assert(Subtarget->isTargetELF() &&
1253 "TLS not implemented for non-ELF targets");
1254 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1255 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1256 // otherwise use the "Local Exec" TLS Model
1257 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1258 return LowerToTLSGeneralDynamicModel(GA, DAG);
1260 return LowerToTLSExecModels(GA, DAG);
1263 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1264 SelectionDAG &DAG) {
1265 EVT PtrVT = getPointerTy();
1266 DebugLoc dl = Op.getDebugLoc();
1267 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1268 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1269 if (RelocM == Reloc::PIC_) {
1270 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1271 ARMConstantPoolValue *CPV =
1272 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
1273 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1274 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1275 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1277 SDValue Chain = Result.getValue(1);
1278 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1279 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1281 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1284 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1285 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1286 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1290 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
1291 /// even in non-static mode.
1292 static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
1293 // If symbol visibility is hidden, the extra load is not needed if
1294 // the symbol is definitely defined in the current translation unit.
1295 bool isDecl = GV->isDeclaration() || GV->hasAvailableExternallyLinkage();
1296 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage()))
1298 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker());
1301 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1302 SelectionDAG &DAG) {
1303 EVT PtrVT = getPointerTy();
1304 DebugLoc dl = Op.getDebugLoc();
1305 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1306 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1307 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
1309 if (RelocM == Reloc::Static)
1310 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1312 unsigned PCAdj = (RelocM != Reloc::PIC_)
1313 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1314 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
1316 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
1318 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1320 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1322 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1323 SDValue Chain = Result.getValue(1);
1325 if (RelocM == Reloc::PIC_) {
1326 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1327 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1330 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1335 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1337 assert(Subtarget->isTargetELF() &&
1338 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1339 EVT PtrVT = getPointerTy();
1340 DebugLoc dl = Op.getDebugLoc();
1341 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1342 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1343 "_GLOBAL_OFFSET_TABLE_",
1345 ARMCP::CPValue, PCAdj);
1346 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1347 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1348 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1349 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1350 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1353 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1355 SDNode *Node = Op.getNode();
1356 EVT VT = Node->getValueType(0);
1357 DebugLoc dl = Op.getDebugLoc();
1359 if (!VT.is64BitVector())
1360 return SDValue(); // unimplemented
1362 SDValue Ops[] = { Node->getOperand(0),
1363 Node->getOperand(2) };
1364 return DAG.getNode(Opcode, dl, Node->getVTList(), Ops, 2);
1367 static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
1368 unsigned Opcode, unsigned NumVecs) {
1369 SDNode *Node = Op.getNode();
1370 EVT VT = Node->getOperand(3).getValueType();
1371 DebugLoc dl = Op.getDebugLoc();
1373 if (!VT.is64BitVector())
1374 return SDValue(); // unimplemented
1376 SmallVector<SDValue, 6> Ops;
1377 Ops.push_back(Node->getOperand(0));
1378 Ops.push_back(Node->getOperand(2));
1379 for (unsigned N = 0; N < NumVecs; ++N)
1380 Ops.push_back(Node->getOperand(N + 3));
1381 return DAG.getNode(Opcode, dl, MVT::Other, Ops.data(), Ops.size());
1385 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1386 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1388 case Intrinsic::arm_neon_vld2:
1389 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD2D);
1390 case Intrinsic::arm_neon_vld3:
1391 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD3D);
1392 case Intrinsic::arm_neon_vld4:
1393 return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD4D);
1394 case Intrinsic::arm_neon_vst2:
1395 return LowerNeonVSTIntrinsic(Op, DAG, ARMISD::VST2D, 2);
1396 case Intrinsic::arm_neon_vst3:
1397 return LowerNeonVSTIntrinsic(Op, DAG, ARMISD::VST3D, 3);
1398 case Intrinsic::arm_neon_vst4:
1399 return LowerNeonVSTIntrinsic(Op, DAG, ARMISD::VST4D, 4);
1400 default: return SDValue(); // Don't custom lower most intrinsics.
1405 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1406 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1407 DebugLoc dl = Op.getDebugLoc();
1409 default: return SDValue(); // Don't custom lower most intrinsics.
1410 case Intrinsic::arm_thread_pointer: {
1411 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1412 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1414 case Intrinsic::eh_sjlj_lsda: {
1415 // blah. horrible, horrible hack with the forced magic name.
1416 // really need to clean this up. It belongs in the target-independent
1417 // layer somehow that doesn't require the coupling with the asm
1419 MachineFunction &MF = DAG.getMachineFunction();
1420 EVT PtrVT = getPointerTy();
1421 DebugLoc dl = Op.getDebugLoc();
1422 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1424 unsigned PCAdj = (RelocM != Reloc::PIC_)
1425 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1426 ARMCP::ARMCPKind Kind = ARMCP::CPValue;
1427 // Save off the LSDA name for the AsmPrinter to use when it's time
1428 // to emit the table
1429 std::string LSDAName = "L_lsda_";
1430 LSDAName += MF.getFunction()->getName();
1431 ARMConstantPoolValue *CPV =
1432 new ARMConstantPoolValue(*DAG.getContext(), LSDAName.c_str(),
1433 ARMPCLabelIndex, Kind, PCAdj);
1434 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1435 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1437 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1438 SDValue Chain = Result.getValue(1);
1440 if (RelocM == Reloc::PIC_) {
1441 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1442 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1446 case Intrinsic::eh_sjlj_setjmp:
1447 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1451 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1452 unsigned VarArgsFrameIndex) {
1453 // vastart just stores the address of the VarArgsFrameIndex slot into the
1454 // memory location argument.
1455 DebugLoc dl = Op.getDebugLoc();
1456 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1457 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1458 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1459 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1463 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1464 SDNode *Node = Op.getNode();
1465 DebugLoc dl = Node->getDebugLoc();
1466 EVT VT = Node->getValueType(0);
1467 SDValue Chain = Op.getOperand(0);
1468 SDValue Size = Op.getOperand(1);
1469 SDValue Align = Op.getOperand(2);
1471 // Chain the dynamic stack allocation so that it doesn't modify the stack
1472 // pointer when other instructions are using the stack.
1473 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1475 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1476 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1477 if (AlignVal > StackAlign)
1478 // Do this now since selection pass cannot introduce new target
1479 // independent node.
1480 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1482 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1483 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1484 // do even more horrible hack later.
1485 MachineFunction &MF = DAG.getMachineFunction();
1486 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1487 if (AFI->isThumb1OnlyFunction()) {
1489 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1491 uint32_t Val = C->getZExtValue();
1492 if (Val <= 508 && ((Val & 3) == 0))
1496 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1499 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1500 SDValue Ops1[] = { Chain, Size, Align };
1501 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1502 Chain = Res.getValue(1);
1503 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1504 DAG.getIntPtrConstant(0, true), SDValue());
1505 SDValue Ops2[] = { Res, Chain };
1506 return DAG.getMergeValues(Ops2, 2, dl);
1510 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1511 SDValue &Root, SelectionDAG &DAG,
1513 MachineFunction &MF = DAG.getMachineFunction();
1514 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1516 TargetRegisterClass *RC;
1517 if (AFI->isThumb1OnlyFunction())
1518 RC = ARM::tGPRRegisterClass;
1520 RC = ARM::GPRRegisterClass;
1522 // Transform the arguments stored in physical registers into virtual ones.
1523 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1524 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1527 if (NextVA.isMemLoc()) {
1528 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1529 MachineFrameInfo *MFI = MF.getFrameInfo();
1530 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1532 // Create load node to retrieve arguments from the stack.
1533 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1534 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1536 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1537 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1540 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1544 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1545 unsigned CallConv, bool isVarArg,
1546 const SmallVectorImpl<ISD::InputArg>
1548 DebugLoc dl, SelectionDAG &DAG,
1549 SmallVectorImpl<SDValue> &InVals) {
1551 MachineFunction &MF = DAG.getMachineFunction();
1552 MachineFrameInfo *MFI = MF.getFrameInfo();
1554 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1556 // Assign locations to all of the incoming arguments.
1557 SmallVector<CCValAssign, 16> ArgLocs;
1558 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1560 CCInfo.AnalyzeFormalArguments(Ins,
1561 CCAssignFnForNode(CallConv, /* Return*/ false,
1564 SmallVector<SDValue, 16> ArgValues;
1566 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1567 CCValAssign &VA = ArgLocs[i];
1569 // Arguments stored in registers.
1570 if (VA.isRegLoc()) {
1571 EVT RegVT = VA.getLocVT();
1574 if (VA.needsCustom()) {
1575 // f64 and vector types are split up into multiple registers or
1576 // combinations of registers and stack slots.
1579 if (VA.getLocVT() == MVT::v2f64) {
1580 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1582 VA = ArgLocs[++i]; // skip ahead to next loc
1583 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1585 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1586 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1587 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1588 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1589 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1591 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1594 TargetRegisterClass *RC;
1596 if (RegVT == MVT::f32)
1597 RC = ARM::SPRRegisterClass;
1598 else if (RegVT == MVT::f64)
1599 RC = ARM::DPRRegisterClass;
1600 else if (RegVT == MVT::v2f64)
1601 RC = ARM::QPRRegisterClass;
1602 else if (RegVT == MVT::i32)
1603 RC = (AFI->isThumb1OnlyFunction() ?
1604 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1606 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1608 // Transform the arguments in physical registers into virtual ones.
1609 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1610 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1613 // If this is an 8 or 16-bit value, it is really passed promoted
1614 // to 32 bits. Insert an assert[sz]ext to capture this, then
1615 // truncate to the right size.
1616 switch (VA.getLocInfo()) {
1617 default: llvm_unreachable("Unknown loc info!");
1618 case CCValAssign::Full: break;
1619 case CCValAssign::BCvt:
1620 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1622 case CCValAssign::SExt:
1623 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1624 DAG.getValueType(VA.getValVT()));
1625 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1627 case CCValAssign::ZExt:
1628 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1629 DAG.getValueType(VA.getValVT()));
1630 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1634 InVals.push_back(ArgValue);
1636 } else { // VA.isRegLoc()
1639 assert(VA.isMemLoc());
1640 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1642 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1643 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1645 // Create load nodes to retrieve arguments from the stack.
1646 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1647 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1653 static const unsigned GPRArgRegs[] = {
1654 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1657 unsigned NumGPRs = CCInfo.getFirstUnallocated
1658 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1660 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1661 unsigned VARegSize = (4 - NumGPRs) * 4;
1662 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1663 unsigned ArgOffset = 0;
1664 if (VARegSaveSize) {
1665 // If this function is vararg, store any remaining integer argument regs
1666 // to their spots on the stack so that they may be loaded by deferencing
1667 // the result of va_next.
1668 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1669 ArgOffset = CCInfo.getNextStackOffset();
1670 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1671 VARegSaveSize - VARegSize);
1672 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1674 SmallVector<SDValue, 4> MemOps;
1675 for (; NumGPRs < 4; ++NumGPRs) {
1676 TargetRegisterClass *RC;
1677 if (AFI->isThumb1OnlyFunction())
1678 RC = ARM::tGPRRegisterClass;
1680 RC = ARM::GPRRegisterClass;
1682 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1683 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1684 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1685 MemOps.push_back(Store);
1686 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1687 DAG.getConstant(4, getPointerTy()));
1689 if (!MemOps.empty())
1690 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1691 &MemOps[0], MemOps.size());
1693 // This will point to the next argument passed via stack.
1694 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1700 /// isFloatingPointZero - Return true if this is +0.0.
1701 static bool isFloatingPointZero(SDValue Op) {
1702 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1703 return CFP->getValueAPF().isPosZero();
1704 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1705 // Maybe this has already been legalized into the constant pool?
1706 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1707 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1708 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1709 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1710 return CFP->getValueAPF().isPosZero();
1716 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1717 return ( isThumb1Only && (C & ~255U) == 0) ||
1718 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1721 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1722 /// the given operands.
1723 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1724 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1726 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1727 unsigned C = RHSC->getZExtValue();
1728 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1729 // Constant does not fit, try adjusting it by one?
1734 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1735 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1736 RHS = DAG.getConstant(C-1, MVT::i32);
1741 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1742 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1743 RHS = DAG.getConstant(C-1, MVT::i32);
1748 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1749 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1750 RHS = DAG.getConstant(C+1, MVT::i32);
1755 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1756 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1757 RHS = DAG.getConstant(C+1, MVT::i32);
1764 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1765 ARMISD::NodeType CompareType;
1768 CompareType = ARMISD::CMP;
1773 CompareType = ARMISD::CMPZ;
1776 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1777 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1780 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1781 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1784 if (!isFloatingPointZero(RHS))
1785 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1787 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1788 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1791 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1792 const ARMSubtarget *ST) {
1793 EVT VT = Op.getValueType();
1794 SDValue LHS = Op.getOperand(0);
1795 SDValue RHS = Op.getOperand(1);
1796 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1797 SDValue TrueVal = Op.getOperand(2);
1798 SDValue FalseVal = Op.getOperand(3);
1799 DebugLoc dl = Op.getDebugLoc();
1801 if (LHS.getValueType() == MVT::i32) {
1803 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1804 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1805 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1808 ARMCC::CondCodes CondCode, CondCode2;
1809 if (FPCCToARMCC(CC, CondCode, CondCode2))
1810 std::swap(TrueVal, FalseVal);
1812 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1813 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1814 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1815 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1817 if (CondCode2 != ARMCC::AL) {
1818 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1819 // FIXME: Needs another CMP because flag can have but one use.
1820 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1821 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1822 Result, TrueVal, ARMCC2, CCR, Cmp2);
1827 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1828 const ARMSubtarget *ST) {
1829 SDValue Chain = Op.getOperand(0);
1830 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1831 SDValue LHS = Op.getOperand(2);
1832 SDValue RHS = Op.getOperand(3);
1833 SDValue Dest = Op.getOperand(4);
1834 DebugLoc dl = Op.getDebugLoc();
1836 if (LHS.getValueType() == MVT::i32) {
1838 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1839 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1840 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1841 Chain, Dest, ARMCC, CCR,Cmp);
1844 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1845 ARMCC::CondCodes CondCode, CondCode2;
1846 if (FPCCToARMCC(CC, CondCode, CondCode2))
1847 // Swap the LHS/RHS of the comparison if needed.
1848 std::swap(LHS, RHS);
1850 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1851 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1852 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1853 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1854 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1855 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1856 if (CondCode2 != ARMCC::AL) {
1857 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1858 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1859 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1864 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1865 SDValue Chain = Op.getOperand(0);
1866 SDValue Table = Op.getOperand(1);
1867 SDValue Index = Op.getOperand(2);
1868 DebugLoc dl = Op.getDebugLoc();
1870 EVT PTy = getPointerTy();
1871 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1872 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1873 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1874 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1875 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1876 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1877 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1878 if (Subtarget->isThumb2()) {
1879 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1880 // which does another jump to the destination. This also makes it easier
1881 // to translate it to TBB / TBH later.
1882 // FIXME: This might not work if the function is extremely large.
1883 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1884 Addr, Op.getOperand(2), JTI, UId);
1886 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1887 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1888 Chain = Addr.getValue(1);
1889 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1890 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1892 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1893 Chain = Addr.getValue(1);
1894 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1898 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1899 DebugLoc dl = Op.getDebugLoc();
1901 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1902 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1903 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1906 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1907 EVT VT = Op.getValueType();
1908 DebugLoc dl = Op.getDebugLoc();
1910 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1912 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1913 return DAG.getNode(Opc, dl, VT, Op);
1916 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1917 // Implement fcopysign with a fabs and a conditional fneg.
1918 SDValue Tmp0 = Op.getOperand(0);
1919 SDValue Tmp1 = Op.getOperand(1);
1920 DebugLoc dl = Op.getDebugLoc();
1921 EVT VT = Op.getValueType();
1922 EVT SrcVT = Tmp1.getValueType();
1923 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1924 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1925 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1926 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1927 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1930 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1931 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1932 MFI->setFrameAddressIsTaken(true);
1933 EVT VT = Op.getValueType();
1934 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1935 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1936 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1937 ? ARM::R7 : ARM::R11;
1938 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1940 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1945 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1947 SDValue Dst, SDValue Src,
1948 SDValue Size, unsigned Align,
1950 const Value *DstSV, uint64_t DstSVOff,
1951 const Value *SrcSV, uint64_t SrcSVOff){
1952 // Do repeated 4-byte loads and stores. To be improved.
1953 // This requires 4-byte alignment.
1954 if ((Align & 3) != 0)
1956 // This requires the copy size to be a constant, preferrably
1957 // within a subtarget-specific limit.
1958 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1961 uint64_t SizeVal = ConstantSize->getZExtValue();
1962 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1965 unsigned BytesLeft = SizeVal & 3;
1966 unsigned NumMemOps = SizeVal >> 2;
1967 unsigned EmittedNumMemOps = 0;
1969 unsigned VTSize = 4;
1971 const unsigned MAX_LOADS_IN_LDM = 6;
1972 SDValue TFOps[MAX_LOADS_IN_LDM];
1973 SDValue Loads[MAX_LOADS_IN_LDM];
1974 uint64_t SrcOff = 0, DstOff = 0;
1976 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1977 // same number of stores. The loads and stores will get combined into
1978 // ldm/stm later on.
1979 while (EmittedNumMemOps < NumMemOps) {
1981 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1982 Loads[i] = DAG.getLoad(VT, dl, Chain,
1983 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1984 DAG.getConstant(SrcOff, MVT::i32)),
1985 SrcSV, SrcSVOff + SrcOff);
1986 TFOps[i] = Loads[i].getValue(1);
1989 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1992 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1993 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1994 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1995 DAG.getConstant(DstOff, MVT::i32)),
1996 DstSV, DstSVOff + DstOff);
1999 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2001 EmittedNumMemOps += i;
2007 // Issue loads / stores for the trailing (1 - 3) bytes.
2008 unsigned BytesLeftSave = BytesLeft;
2011 if (BytesLeft >= 2) {
2019 Loads[i] = DAG.getLoad(VT, dl, Chain,
2020 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2021 DAG.getConstant(SrcOff, MVT::i32)),
2022 SrcSV, SrcSVOff + SrcOff);
2023 TFOps[i] = Loads[i].getValue(1);
2026 BytesLeft -= VTSize;
2028 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2031 BytesLeft = BytesLeftSave;
2033 if (BytesLeft >= 2) {
2041 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2042 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2043 DAG.getConstant(DstOff, MVT::i32)),
2044 DstSV, DstSVOff + DstOff);
2047 BytesLeft -= VTSize;
2049 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2052 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2053 SDValue Op = N->getOperand(0);
2054 DebugLoc dl = N->getDebugLoc();
2055 if (N->getValueType(0) == MVT::f64) {
2056 // Turn i64->f64 into FMDRR.
2057 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2058 DAG.getConstant(0, MVT::i32));
2059 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2060 DAG.getConstant(1, MVT::i32));
2061 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2064 // Turn f64->i64 into FMRRD.
2065 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2066 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2068 // Merge the pieces into a single i64 value.
2069 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2072 /// getZeroVector - Returns a vector of specified type with all zero elements.
2074 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2075 assert(VT.isVector() && "Expected a vector type");
2077 // Zero vectors are used to represent vector negation and in those cases
2078 // will be implemented with the NEON VNEG instruction. However, VNEG does
2079 // not support i64 elements, so sometimes the zero vectors will need to be
2080 // explicitly constructed. For those cases, and potentially other uses in
2081 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
2082 // to their dest type. This ensures they get CSE'd.
2084 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2085 if (VT.getSizeInBits() == 64)
2086 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2088 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2090 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2093 /// getOnesVector - Returns a vector of specified type with all bits set.
2095 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2096 assert(VT.isVector() && "Expected a vector type");
2098 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2099 // type. This ensures they get CSE'd.
2101 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2102 if (VT.getSizeInBits() == 64)
2103 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2105 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2107 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2110 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2111 const ARMSubtarget *ST) {
2112 EVT VT = N->getValueType(0);
2113 DebugLoc dl = N->getDebugLoc();
2115 // Lower vector shifts on NEON to use VSHL.
2116 if (VT.isVector()) {
2117 assert(ST->hasNEON() && "unexpected vector shift");
2119 // Left shifts translate directly to the vshiftu intrinsic.
2120 if (N->getOpcode() == ISD::SHL)
2121 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2122 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2123 N->getOperand(0), N->getOperand(1));
2125 assert((N->getOpcode() == ISD::SRA ||
2126 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2128 // NEON uses the same intrinsics for both left and right shifts. For
2129 // right shifts, the shift amounts are negative, so negate the vector of
2131 EVT ShiftVT = N->getOperand(1).getValueType();
2132 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2133 getZeroVector(ShiftVT, DAG, dl),
2135 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2136 Intrinsic::arm_neon_vshifts :
2137 Intrinsic::arm_neon_vshiftu);
2138 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2139 DAG.getConstant(vshiftInt, MVT::i32),
2140 N->getOperand(0), NegatedCount);
2143 assert(VT == MVT::i64 &&
2144 (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2145 "Unknown shift to lower!");
2147 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2148 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2149 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2152 // If we are in thumb mode, we don't have RRX.
2153 if (ST->isThumb1Only()) return SDValue();
2155 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2156 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2157 DAG.getConstant(0, MVT::i32));
2158 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2159 DAG.getConstant(1, MVT::i32));
2161 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2162 // captures the result into a carry flag.
2163 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2164 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2166 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2167 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2169 // Merge the pieces into a single i64 value.
2170 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2173 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2174 SDValue TmpOp0, TmpOp1;
2175 bool Invert = false;
2179 SDValue Op0 = Op.getOperand(0);
2180 SDValue Op1 = Op.getOperand(1);
2181 SDValue CC = Op.getOperand(2);
2182 EVT VT = Op.getValueType();
2183 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2184 DebugLoc dl = Op.getDebugLoc();
2186 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2187 switch (SetCCOpcode) {
2188 default: llvm_unreachable("Illegal FP comparison"); break;
2190 case ISD::SETNE: Invert = true; // Fallthrough
2192 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2194 case ISD::SETLT: Swap = true; // Fallthrough
2196 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2198 case ISD::SETLE: Swap = true; // Fallthrough
2200 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2201 case ISD::SETUGE: Swap = true; // Fallthrough
2202 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2203 case ISD::SETUGT: Swap = true; // Fallthrough
2204 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2205 case ISD::SETUEQ: Invert = true; // Fallthrough
2207 // Expand this to (OLT | OGT).
2211 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2212 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2214 case ISD::SETUO: Invert = true; // Fallthrough
2216 // Expand this to (OLT | OGE).
2220 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2221 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2225 // Integer comparisons.
2226 switch (SetCCOpcode) {
2227 default: llvm_unreachable("Illegal integer comparison"); break;
2228 case ISD::SETNE: Invert = true;
2229 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2230 case ISD::SETLT: Swap = true;
2231 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2232 case ISD::SETLE: Swap = true;
2233 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2234 case ISD::SETULT: Swap = true;
2235 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2236 case ISD::SETULE: Swap = true;
2237 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2240 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2241 if (Opc == ARMISD::VCEQ) {
2244 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2246 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2249 // Ignore bitconvert.
2250 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2251 AndOp = AndOp.getOperand(0);
2253 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2255 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2256 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2263 std::swap(Op0, Op1);
2265 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2268 Result = DAG.getNOT(dl, Result, VT);
2273 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2274 /// VMOV instruction, and if so, return the constant being splatted.
2275 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2276 unsigned SplatBitSize, SelectionDAG &DAG) {
2277 switch (SplatBitSize) {
2279 // Any 1-byte value is OK.
2280 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2281 return DAG.getTargetConstant(SplatBits, MVT::i8);
2284 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2285 if ((SplatBits & ~0xff) == 0 ||
2286 (SplatBits & ~0xff00) == 0)
2287 return DAG.getTargetConstant(SplatBits, MVT::i16);
2291 // NEON's 32-bit VMOV supports splat values where:
2292 // * only one byte is nonzero, or
2293 // * the least significant byte is 0xff and the second byte is nonzero, or
2294 // * the least significant 2 bytes are 0xff and the third is nonzero.
2295 if ((SplatBits & ~0xff) == 0 ||
2296 (SplatBits & ~0xff00) == 0 ||
2297 (SplatBits & ~0xff0000) == 0 ||
2298 (SplatBits & ~0xff000000) == 0)
2299 return DAG.getTargetConstant(SplatBits, MVT::i32);
2301 if ((SplatBits & ~0xffff) == 0 &&
2302 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2303 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2305 if ((SplatBits & ~0xffffff) == 0 &&
2306 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2307 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2309 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2310 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2311 // VMOV.I32. A (very) minor optimization would be to replicate the value
2312 // and fall through here to test for a valid 64-bit splat. But, then the
2313 // caller would also need to check and handle the change in size.
2317 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2318 uint64_t BitMask = 0xff;
2320 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2321 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2323 else if ((SplatBits & BitMask) != 0)
2327 return DAG.getTargetConstant(Val, MVT::i64);
2331 llvm_unreachable("unexpected size for isVMOVSplat");
2338 /// getVMOVImm - If this is a build_vector of constants which can be
2339 /// formed by using a VMOV instruction of the specified element size,
2340 /// return the constant being splatted. The ByteSize field indicates the
2341 /// number of bytes of each element [1248].
2342 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2343 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2344 APInt SplatBits, SplatUndef;
2345 unsigned SplatBitSize;
2347 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2348 HasAnyUndefs, ByteSize * 8))
2351 if (SplatBitSize > ByteSize * 8)
2354 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2358 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2359 bool &ReverseVEXT, unsigned &Imm) {
2360 unsigned NumElts = VT.getVectorNumElements();
2361 ReverseVEXT = false;
2364 // If this is a VEXT shuffle, the immediate value is the index of the first
2365 // element. The other shuffle indices must be the successive elements after
2367 unsigned ExpectedElt = Imm;
2368 for (unsigned i = 1; i < NumElts; ++i) {
2369 // Increment the expected index. If it wraps around, it may still be
2370 // a VEXT but the source vectors must be swapped.
2372 if (ExpectedElt == NumElts * 2) {
2377 if (ExpectedElt != static_cast<unsigned>(M[i]))
2381 // Adjust the index value if the source operands will be swapped.
2388 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2389 /// instruction with the specified blocksize. (The order of the elements
2390 /// within each block of the vector is reversed.)
2391 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2392 unsigned BlockSize) {
2393 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2394 "Only possible block sizes for VREV are: 16, 32, 64");
2396 unsigned NumElts = VT.getVectorNumElements();
2397 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2398 unsigned BlockElts = M[0] + 1;
2400 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2403 for (unsigned i = 0; i < NumElts; ++i) {
2404 if ((unsigned) M[i] !=
2405 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2412 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2413 // Canonicalize all-zeros and all-ones vectors.
2414 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2415 if (ConstVal->isNullValue())
2416 return getZeroVector(VT, DAG, dl);
2417 if (ConstVal->isAllOnesValue())
2418 return getOnesVector(VT, DAG, dl);
2421 if (VT.is64BitVector()) {
2422 switch (Val.getValueType().getSizeInBits()) {
2423 case 8: CanonicalVT = MVT::v8i8; break;
2424 case 16: CanonicalVT = MVT::v4i16; break;
2425 case 32: CanonicalVT = MVT::v2i32; break;
2426 case 64: CanonicalVT = MVT::v1i64; break;
2427 default: llvm_unreachable("unexpected splat element type"); break;
2430 assert(VT.is128BitVector() && "unknown splat vector size");
2431 switch (Val.getValueType().getSizeInBits()) {
2432 case 8: CanonicalVT = MVT::v16i8; break;
2433 case 16: CanonicalVT = MVT::v8i16; break;
2434 case 32: CanonicalVT = MVT::v4i32; break;
2435 case 64: CanonicalVT = MVT::v2i64; break;
2436 default: llvm_unreachable("unexpected splat element type"); break;
2440 // Build a canonical splat for this value.
2441 SmallVector<SDValue, 8> Ops;
2442 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2443 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2445 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2448 // If this is a case we can't handle, return null and let the default
2449 // expansion code take care of it.
2450 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2451 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2452 DebugLoc dl = Op.getDebugLoc();
2453 EVT VT = Op.getValueType();
2455 APInt SplatBits, SplatUndef;
2456 unsigned SplatBitSize;
2458 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2459 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2460 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2462 return BuildSplat(Val, VT, DAG, dl);
2465 // If there are only 2 elements in a 128-bit vector, insert them into an
2466 // undef vector. This handles the common case for 128-bit vector argument
2467 // passing, where the insertions should be translated to subreg accesses
2468 // with no real instructions.
2469 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2470 SDValue Val = DAG.getUNDEF(VT);
2471 SDValue Op0 = Op.getOperand(0);
2472 SDValue Op1 = Op.getOperand(1);
2473 if (Op0.getOpcode() != ISD::UNDEF)
2474 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2475 DAG.getIntPtrConstant(0));
2476 if (Op1.getOpcode() != ISD::UNDEF)
2477 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2478 DAG.getIntPtrConstant(1));
2485 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2486 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2487 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2488 /// are assumed to be legal.
2490 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2492 if (VT.getVectorNumElements() == 4 &&
2493 (VT.is128BitVector() || VT.is64BitVector())) {
2494 unsigned PFIndexes[4];
2495 for (unsigned i = 0; i != 4; ++i) {
2499 PFIndexes[i] = M[i];
2502 // Compute the index in the perfect shuffle table.
2503 unsigned PFTableIndex =
2504 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2505 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2506 unsigned Cost = (PFEntry >> 30);
2515 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2516 isVREVMask(M, VT, 64) ||
2517 isVREVMask(M, VT, 32) ||
2518 isVREVMask(M, VT, 16) ||
2519 isVEXTMask(M, VT, ReverseVEXT, Imm));
2522 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2523 /// the specified operations to build the shuffle.
2524 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2525 SDValue RHS, SelectionDAG &DAG,
2527 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2528 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2529 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2532 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2541 OP_VUZPL, // VUZP, left result
2542 OP_VUZPR, // VUZP, right result
2543 OP_VZIPL, // VZIP, left result
2544 OP_VZIPR, // VZIP, right result
2545 OP_VTRNL, // VTRN, left result
2546 OP_VTRNR // VTRN, right result
2549 if (OpNum == OP_COPY) {
2550 if (LHSID == (1*9+2)*9+3) return LHS;
2551 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2555 SDValue OpLHS, OpRHS;
2556 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2557 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2558 EVT VT = OpLHS.getValueType();
2561 default: llvm_unreachable("Unknown shuffle opcode!");
2563 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2568 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2569 OpLHS, DAG.getConstant(OpNum-OP_VDUP0+1, MVT::i32));
2573 return DAG.getNode(ARMISD::VEXT, dl, VT,
2575 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2578 return DAG.getNode(VT.is64BitVector() ? ARMISD::VUZP16 : ARMISD::VUZP32,
2579 dl, DAG.getVTList(VT, VT),
2580 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2583 return DAG.getNode(VT.is64BitVector() ? ARMISD::VZIP16 : ARMISD::VZIP32,
2584 dl, DAG.getVTList(VT, VT),
2585 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2588 return DAG.getNode(VT.is64BitVector() ? ARMISD::VTRN16 : ARMISD::VTRN32,
2589 dl, DAG.getVTList(VT, VT),
2590 OpLHS, OpRHS).getValue(0);
2594 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2595 SDValue V1 = Op.getOperand(0);
2596 SDValue V2 = Op.getOperand(1);
2597 DebugLoc dl = Op.getDebugLoc();
2598 EVT VT = Op.getValueType();
2599 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2600 SmallVector<int, 8> ShuffleMask;
2602 // Convert shuffles that are directly supported on NEON to target-specific
2603 // DAG nodes, instead of keeping them as shuffles and matching them again
2604 // during code selection. This is more efficient and avoids the possibility
2605 // of inconsistencies between legalization and selection.
2606 // FIXME: floating-point vectors should be canonicalized to integer vectors
2607 // of the same time so that they get CSEd properly.
2608 SVN->getMask(ShuffleMask);
2610 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2611 int Lane = SVN->getSplatIndex();
2612 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2613 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2615 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2616 DAG.getConstant(Lane, MVT::i32));
2621 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2622 SDValue Op0 = SVN->getOperand(0);
2623 SDValue Op1 = SVN->getOperand(1);
2625 std::swap(Op0, Op1);
2626 return DAG.getNode(ARMISD::VEXT, dl, VT, Op0, Op1,
2627 DAG.getConstant(Imm, MVT::i32));
2630 if (isVREVMask(ShuffleMask, VT, 64))
2631 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2632 if (isVREVMask(ShuffleMask, VT, 32))
2633 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2634 if (isVREVMask(ShuffleMask, VT, 16))
2635 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2637 if (VT.getVectorNumElements() == 4 &&
2638 (VT.is128BitVector() || VT.is64BitVector())) {
2639 unsigned PFIndexes[4];
2640 for (unsigned i = 0; i != 4; ++i) {
2641 if (ShuffleMask[i] < 0)
2644 PFIndexes[i] = ShuffleMask[i];
2647 // Compute the index in the perfect shuffle table.
2648 unsigned PFTableIndex =
2649 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2651 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2652 unsigned Cost = (PFEntry >> 30);
2655 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2661 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
2665 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2666 EVT VT = Op.getValueType();
2667 DebugLoc dl = Op.getDebugLoc();
2668 assert((VT == MVT::i8 || VT == MVT::i16) &&
2669 "unexpected type for custom-lowering vector extract");
2670 SDValue Vec = Op.getOperand(0);
2671 SDValue Lane = Op.getOperand(1);
2672 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2673 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2674 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2677 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2678 // The only time a CONCAT_VECTORS operation can have legal types is when
2679 // two 64-bit vectors are concatenated to a 128-bit vector.
2680 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2681 "unexpected CONCAT_VECTORS");
2682 DebugLoc dl = Op.getDebugLoc();
2683 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2684 SDValue Op0 = Op.getOperand(0);
2685 SDValue Op1 = Op.getOperand(1);
2686 if (Op0.getOpcode() != ISD::UNDEF)
2687 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2688 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2689 DAG.getIntPtrConstant(0));
2690 if (Op1.getOpcode() != ISD::UNDEF)
2691 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2692 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2693 DAG.getIntPtrConstant(1));
2694 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2697 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2698 switch (Op.getOpcode()) {
2699 default: llvm_unreachable("Don't know how to custom lower this!");
2700 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2701 case ISD::GlobalAddress:
2702 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2703 LowerGlobalAddressELF(Op, DAG);
2704 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2705 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2706 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2707 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2708 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2709 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2710 case ISD::SINT_TO_FP:
2711 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2712 case ISD::FP_TO_SINT:
2713 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2714 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2715 case ISD::RETURNADDR: break;
2716 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2717 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2718 case ISD::INTRINSIC_VOID:
2719 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2720 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2721 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2724 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2725 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2726 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2727 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2728 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
2729 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2730 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2735 /// ReplaceNodeResults - Replace the results of node with an illegal result
2736 /// type with new values built out of custom code.
2737 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2738 SmallVectorImpl<SDValue>&Results,
2739 SelectionDAG &DAG) {
2740 switch (N->getOpcode()) {
2742 llvm_unreachable("Don't know how to custom expand this!");
2744 case ISD::BIT_CONVERT:
2745 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2749 SDValue Res = LowerShift(N, DAG, Subtarget);
2751 Results.push_back(Res);
2757 //===----------------------------------------------------------------------===//
2758 // ARM Scheduler Hooks
2759 //===----------------------------------------------------------------------===//
2762 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2763 MachineBasicBlock *BB) const {
2764 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2765 DebugLoc dl = MI->getDebugLoc();
2766 switch (MI->getOpcode()) {
2768 llvm_unreachable("Unexpected instr type to insert");
2769 case ARM::tMOVCCr_pseudo: {
2770 // To "insert" a SELECT_CC instruction, we actually have to insert the
2771 // diamond control-flow pattern. The incoming instruction knows the
2772 // destination vreg to set, the condition code register to branch on, the
2773 // true/false values to select between, and a branch opcode to use.
2774 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2775 MachineFunction::iterator It = BB;
2781 // cmpTY ccX, r1, r2
2783 // fallthrough --> copy0MBB
2784 MachineBasicBlock *thisMBB = BB;
2785 MachineFunction *F = BB->getParent();
2786 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2787 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2788 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2789 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2790 F->insert(It, copy0MBB);
2791 F->insert(It, sinkMBB);
2792 // Update machine-CFG edges by first adding all successors of the current
2793 // block to the new block which will contain the Phi node for the select.
2794 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2795 e = BB->succ_end(); i != e; ++i)
2796 sinkMBB->addSuccessor(*i);
2797 // Next, remove all successors of the current block, and add the true
2798 // and fallthrough blocks as its successors.
2799 while(!BB->succ_empty())
2800 BB->removeSuccessor(BB->succ_begin());
2801 BB->addSuccessor(copy0MBB);
2802 BB->addSuccessor(sinkMBB);
2805 // %FalseValue = ...
2806 // # fallthrough to sinkMBB
2809 // Update machine-CFG edges
2810 BB->addSuccessor(sinkMBB);
2813 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2816 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2817 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2818 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2820 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2827 case ARM::t2SUBrSPi_:
2828 case ARM::t2SUBrSPi12_:
2829 case ARM::t2SUBrSPs_: {
2830 MachineFunction *MF = BB->getParent();
2831 unsigned DstReg = MI->getOperand(0).getReg();
2832 unsigned SrcReg = MI->getOperand(1).getReg();
2833 bool DstIsDead = MI->getOperand(0).isDead();
2834 bool SrcIsKill = MI->getOperand(1).isKill();
2836 if (SrcReg != ARM::SP) {
2837 // Copy the source to SP from virtual register.
2838 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2839 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2840 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2841 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2842 .addReg(SrcReg, getKillRegState(SrcIsKill));
2846 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2847 switch (MI->getOpcode()) {
2849 llvm_unreachable("Unexpected pseudo instruction!");
2855 OpOpc = ARM::tADDspr;
2858 OpOpc = ARM::tSUBspi;
2860 case ARM::t2SUBrSPi_:
2861 OpOpc = ARM::t2SUBrSPi;
2862 NeedPred = true; NeedCC = true;
2864 case ARM::t2SUBrSPi12_:
2865 OpOpc = ARM::t2SUBrSPi12;
2868 case ARM::t2SUBrSPs_:
2869 OpOpc = ARM::t2SUBrSPs;
2870 NeedPred = true; NeedCC = true; NeedOp3 = true;
2873 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2874 if (OpOpc == ARM::tAND)
2875 AddDefaultT1CC(MIB);
2876 MIB.addReg(ARM::SP);
2877 MIB.addOperand(MI->getOperand(2));
2879 MIB.addOperand(MI->getOperand(3));
2881 AddDefaultPred(MIB);
2885 // Copy the result from SP to virtual register.
2886 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2887 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2888 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2889 BuildMI(BB, dl, TII->get(CopyOpc))
2890 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2892 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2898 //===----------------------------------------------------------------------===//
2899 // ARM Optimization Hooks
2900 //===----------------------------------------------------------------------===//
2903 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2904 TargetLowering::DAGCombinerInfo &DCI) {
2905 SelectionDAG &DAG = DCI.DAG;
2906 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2907 EVT VT = N->getValueType(0);
2908 unsigned Opc = N->getOpcode();
2909 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2910 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2911 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2912 ISD::CondCode CC = ISD::SETCC_INVALID;
2915 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2917 SDValue CCOp = Slct.getOperand(0);
2918 if (CCOp.getOpcode() == ISD::SETCC)
2919 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2922 bool DoXform = false;
2924 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2927 if (LHS.getOpcode() == ISD::Constant &&
2928 cast<ConstantSDNode>(LHS)->isNullValue()) {
2930 } else if (CC != ISD::SETCC_INVALID &&
2931 RHS.getOpcode() == ISD::Constant &&
2932 cast<ConstantSDNode>(RHS)->isNullValue()) {
2933 std::swap(LHS, RHS);
2934 SDValue Op0 = Slct.getOperand(0);
2935 EVT OpVT = isSlctCC ? Op0.getValueType() :
2936 Op0.getOperand(0).getValueType();
2937 bool isInt = OpVT.isInteger();
2938 CC = ISD::getSetCCInverse(CC, isInt);
2940 if (!TLI.isCondCodeLegal(CC, OpVT))
2941 return SDValue(); // Inverse operator isn't legal.
2948 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
2950 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
2951 Slct.getOperand(0), Slct.getOperand(1), CC);
2952 SDValue CCOp = Slct.getOperand(0);
2954 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
2955 CCOp.getOperand(0), CCOp.getOperand(1), CC);
2956 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
2957 CCOp, OtherOp, Result);
2962 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
2963 static SDValue PerformADDCombine(SDNode *N,
2964 TargetLowering::DAGCombinerInfo &DCI) {
2965 // added by evan in r37685 with no testcase.
2966 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2968 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
2969 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
2970 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
2971 if (Result.getNode()) return Result;
2973 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2974 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2975 if (Result.getNode()) return Result;
2981 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
2982 static SDValue PerformSUBCombine(SDNode *N,
2983 TargetLowering::DAGCombinerInfo &DCI) {
2984 // added by evan in r37685 with no testcase.
2985 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2987 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
2988 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
2989 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
2990 if (Result.getNode()) return Result;
2997 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
2998 static SDValue PerformFMRRDCombine(SDNode *N,
2999 TargetLowering::DAGCombinerInfo &DCI) {
3000 // fmrrd(fmdrr x, y) -> x,y
3001 SDValue InDouble = N->getOperand(0);
3002 if (InDouble.getOpcode() == ARMISD::FMDRR)
3003 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3007 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3008 /// operand of a vector shift operation, where all the elements of the
3009 /// build_vector must have the same constant integer value.
3010 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3011 // Ignore bit_converts.
3012 while (Op.getOpcode() == ISD::BIT_CONVERT)
3013 Op = Op.getOperand(0);
3014 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3015 APInt SplatBits, SplatUndef;
3016 unsigned SplatBitSize;
3018 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3019 HasAnyUndefs, ElementBits) ||
3020 SplatBitSize > ElementBits)
3022 Cnt = SplatBits.getSExtValue();
3026 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3027 /// operand of a vector shift left operation. That value must be in the range:
3028 /// 0 <= Value < ElementBits for a left shift; or
3029 /// 0 <= Value <= ElementBits for a long left shift.
3030 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3031 assert(VT.isVector() && "vector shift count is not a vector type");
3032 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3033 if (! getVShiftImm(Op, ElementBits, Cnt))
3035 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3038 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3039 /// operand of a vector shift right operation. For a shift opcode, the value
3040 /// is positive, but for an intrinsic the value count must be negative. The
3041 /// absolute value must be in the range:
3042 /// 1 <= |Value| <= ElementBits for a right shift; or
3043 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3044 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3046 assert(VT.isVector() && "vector shift count is not a vector type");
3047 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3048 if (! getVShiftImm(Op, ElementBits, Cnt))
3052 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3055 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3056 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3057 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3060 // Don't do anything for most intrinsics.
3063 // Vector shifts: check for immediate versions and lower them.
3064 // Note: This is done during DAG combining instead of DAG legalizing because
3065 // the build_vectors for 64-bit vector element shift counts are generally
3066 // not legal, and it is hard to see their values after they get legalized to
3067 // loads from a constant pool.
3068 case Intrinsic::arm_neon_vshifts:
3069 case Intrinsic::arm_neon_vshiftu:
3070 case Intrinsic::arm_neon_vshiftls:
3071 case Intrinsic::arm_neon_vshiftlu:
3072 case Intrinsic::arm_neon_vshiftn:
3073 case Intrinsic::arm_neon_vrshifts:
3074 case Intrinsic::arm_neon_vrshiftu:
3075 case Intrinsic::arm_neon_vrshiftn:
3076 case Intrinsic::arm_neon_vqshifts:
3077 case Intrinsic::arm_neon_vqshiftu:
3078 case Intrinsic::arm_neon_vqshiftsu:
3079 case Intrinsic::arm_neon_vqshiftns:
3080 case Intrinsic::arm_neon_vqshiftnu:
3081 case Intrinsic::arm_neon_vqshiftnsu:
3082 case Intrinsic::arm_neon_vqrshiftns:
3083 case Intrinsic::arm_neon_vqrshiftnu:
3084 case Intrinsic::arm_neon_vqrshiftnsu: {
3085 EVT VT = N->getOperand(1).getValueType();
3087 unsigned VShiftOpc = 0;
3090 case Intrinsic::arm_neon_vshifts:
3091 case Intrinsic::arm_neon_vshiftu:
3092 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3093 VShiftOpc = ARMISD::VSHL;
3096 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3097 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3098 ARMISD::VSHRs : ARMISD::VSHRu);
3103 case Intrinsic::arm_neon_vshiftls:
3104 case Intrinsic::arm_neon_vshiftlu:
3105 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3107 llvm_unreachable("invalid shift count for vshll intrinsic");
3109 case Intrinsic::arm_neon_vrshifts:
3110 case Intrinsic::arm_neon_vrshiftu:
3111 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3115 case Intrinsic::arm_neon_vqshifts:
3116 case Intrinsic::arm_neon_vqshiftu:
3117 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3121 case Intrinsic::arm_neon_vqshiftsu:
3122 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3124 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3126 case Intrinsic::arm_neon_vshiftn:
3127 case Intrinsic::arm_neon_vrshiftn:
3128 case Intrinsic::arm_neon_vqshiftns:
3129 case Intrinsic::arm_neon_vqshiftnu:
3130 case Intrinsic::arm_neon_vqshiftnsu:
3131 case Intrinsic::arm_neon_vqrshiftns:
3132 case Intrinsic::arm_neon_vqrshiftnu:
3133 case Intrinsic::arm_neon_vqrshiftnsu:
3134 // Narrowing shifts require an immediate right shift.
3135 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3137 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3140 llvm_unreachable("unhandled vector shift");
3144 case Intrinsic::arm_neon_vshifts:
3145 case Intrinsic::arm_neon_vshiftu:
3146 // Opcode already set above.
3148 case Intrinsic::arm_neon_vshiftls:
3149 case Intrinsic::arm_neon_vshiftlu:
3150 if (Cnt == VT.getVectorElementType().getSizeInBits())
3151 VShiftOpc = ARMISD::VSHLLi;
3153 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3154 ARMISD::VSHLLs : ARMISD::VSHLLu);
3156 case Intrinsic::arm_neon_vshiftn:
3157 VShiftOpc = ARMISD::VSHRN; break;
3158 case Intrinsic::arm_neon_vrshifts:
3159 VShiftOpc = ARMISD::VRSHRs; break;
3160 case Intrinsic::arm_neon_vrshiftu:
3161 VShiftOpc = ARMISD::VRSHRu; break;
3162 case Intrinsic::arm_neon_vrshiftn:
3163 VShiftOpc = ARMISD::VRSHRN; break;
3164 case Intrinsic::arm_neon_vqshifts:
3165 VShiftOpc = ARMISD::VQSHLs; break;
3166 case Intrinsic::arm_neon_vqshiftu:
3167 VShiftOpc = ARMISD::VQSHLu; break;
3168 case Intrinsic::arm_neon_vqshiftsu:
3169 VShiftOpc = ARMISD::VQSHLsu; break;
3170 case Intrinsic::arm_neon_vqshiftns:
3171 VShiftOpc = ARMISD::VQSHRNs; break;
3172 case Intrinsic::arm_neon_vqshiftnu:
3173 VShiftOpc = ARMISD::VQSHRNu; break;
3174 case Intrinsic::arm_neon_vqshiftnsu:
3175 VShiftOpc = ARMISD::VQSHRNsu; break;
3176 case Intrinsic::arm_neon_vqrshiftns:
3177 VShiftOpc = ARMISD::VQRSHRNs; break;
3178 case Intrinsic::arm_neon_vqrshiftnu:
3179 VShiftOpc = ARMISD::VQRSHRNu; break;
3180 case Intrinsic::arm_neon_vqrshiftnsu:
3181 VShiftOpc = ARMISD::VQRSHRNsu; break;
3184 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3185 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3188 case Intrinsic::arm_neon_vshiftins: {
3189 EVT VT = N->getOperand(1).getValueType();
3191 unsigned VShiftOpc = 0;
3193 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3194 VShiftOpc = ARMISD::VSLI;
3195 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3196 VShiftOpc = ARMISD::VSRI;
3198 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3201 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3202 N->getOperand(1), N->getOperand(2),
3203 DAG.getConstant(Cnt, MVT::i32));
3206 case Intrinsic::arm_neon_vqrshifts:
3207 case Intrinsic::arm_neon_vqrshiftu:
3208 // No immediate versions of these to check for.
3215 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3216 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3217 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3218 /// vector element shift counts are generally not legal, and it is hard to see
3219 /// their values after they get legalized to loads from a constant pool.
3220 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3221 const ARMSubtarget *ST) {
3222 EVT VT = N->getValueType(0);
3224 // Nothing to be done for scalar shifts.
3225 if (! VT.isVector())
3228 assert(ST->hasNEON() && "unexpected vector shift");
3231 switch (N->getOpcode()) {
3232 default: llvm_unreachable("unexpected shift opcode");
3235 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3236 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3237 DAG.getConstant(Cnt, MVT::i32));
3242 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3243 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3244 ARMISD::VSHRs : ARMISD::VSHRu);
3245 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3246 DAG.getConstant(Cnt, MVT::i32));
3252 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3253 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3254 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3255 const ARMSubtarget *ST) {
3256 SDValue N0 = N->getOperand(0);
3258 // Check for sign- and zero-extensions of vector extract operations of 8-
3259 // and 16-bit vector elements. NEON supports these directly. They are
3260 // handled during DAG combining because type legalization will promote them
3261 // to 32-bit types and it is messy to recognize the operations after that.
3262 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3263 SDValue Vec = N0.getOperand(0);
3264 SDValue Lane = N0.getOperand(1);
3265 EVT VT = N->getValueType(0);
3266 EVT EltVT = N0.getValueType();
3267 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3269 if (VT == MVT::i32 &&
3270 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3271 TLI.isTypeLegal(Vec.getValueType())) {
3274 switch (N->getOpcode()) {
3275 default: llvm_unreachable("unexpected opcode");
3276 case ISD::SIGN_EXTEND:
3277 Opc = ARMISD::VGETLANEs;
3279 case ISD::ZERO_EXTEND:
3280 case ISD::ANY_EXTEND:
3281 Opc = ARMISD::VGETLANEu;
3284 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3291 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3292 DAGCombinerInfo &DCI) const {
3293 switch (N->getOpcode()) {
3295 case ISD::ADD: return PerformADDCombine(N, DCI);
3296 case ISD::SUB: return PerformSUBCombine(N, DCI);
3297 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3298 case ISD::INTRINSIC_WO_CHAIN:
3299 return PerformIntrinsicCombine(N, DCI.DAG);
3303 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3304 case ISD::SIGN_EXTEND:
3305 case ISD::ZERO_EXTEND:
3306 case ISD::ANY_EXTEND:
3307 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3312 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3313 if (!Subtarget->hasV6Ops())
3314 // Pre-v6 does not support unaligned mem access.
3316 else if (!Subtarget->hasV6Ops()) {
3317 // v6 may or may not support unaligned mem access.
3318 if (!Subtarget->isTargetDarwin())
3322 switch (VT.getSimpleVT().SimpleTy) {
3329 // FIXME: VLD1 etc with standard alignment is legal.
3333 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3338 switch (VT.getSimpleVT().SimpleTy) {
3339 default: return false;
3354 if ((V & (Scale - 1)) != 0)
3357 return V == (V & ((1LL << 5) - 1));
3360 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3361 const ARMSubtarget *Subtarget) {
3368 switch (VT.getSimpleVT().SimpleTy) {
3369 default: return false;
3374 // + imm12 or - imm8
3376 return V == (V & ((1LL << 8) - 1));
3377 return V == (V & ((1LL << 12) - 1));
3380 // Same as ARM mode. FIXME: NEON?
3381 if (!Subtarget->hasVFP2())
3386 return V == (V & ((1LL << 8) - 1));
3390 /// isLegalAddressImmediate - Return true if the integer value can be used
3391 /// as the offset of the target addressing mode for load / store of the
3393 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3394 const ARMSubtarget *Subtarget) {
3401 if (Subtarget->isThumb1Only())
3402 return isLegalT1AddressImmediate(V, VT);
3403 else if (Subtarget->isThumb2())
3404 return isLegalT2AddressImmediate(V, VT, Subtarget);
3409 switch (VT.getSimpleVT().SimpleTy) {
3410 default: return false;
3415 return V == (V & ((1LL << 12) - 1));
3418 return V == (V & ((1LL << 8) - 1));
3421 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3426 return V == (V & ((1LL << 8) - 1));
3430 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3432 int Scale = AM.Scale;
3436 switch (VT.getSimpleVT().SimpleTy) {
3437 default: return false;
3446 return Scale == 2 || Scale == 4 || Scale == 8;
3449 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3453 // Note, we allow "void" uses (basically, uses that aren't loads or
3454 // stores), because arm allows folding a scale into many arithmetic
3455 // operations. This should be made more precise and revisited later.
3457 // Allow r << imm, but the imm has to be a multiple of two.
3458 if (Scale & 1) return false;
3459 return isPowerOf2_32(Scale);
3463 /// isLegalAddressingMode - Return true if the addressing mode represented
3464 /// by AM is legal for this target, for a load/store of the specified type.
3465 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3466 const Type *Ty) const {
3467 EVT VT = getValueType(Ty, true);
3468 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3471 // Can never fold addr of global into load/store.
3476 case 0: // no scale reg, must be "r+i" or "r", or "i".
3479 if (Subtarget->isThumb1Only())
3483 // ARM doesn't support any R+R*scale+imm addr modes.
3490 if (Subtarget->isThumb2())
3491 return isLegalT2ScaledAddressingMode(AM, VT);
3493 int Scale = AM.Scale;
3494 switch (VT.getSimpleVT().SimpleTy) {
3495 default: return false;
3499 if (Scale < 0) Scale = -Scale;
3503 return isPowerOf2_32(Scale & ~1);
3507 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3512 // Note, we allow "void" uses (basically, uses that aren't loads or
3513 // stores), because arm allows folding a scale into many arithmetic
3514 // operations. This should be made more precise and revisited later.
3516 // Allow r << imm, but the imm has to be a multiple of two.
3517 if (Scale & 1) return false;
3518 return isPowerOf2_32(Scale);
3525 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3526 bool isSEXTLoad, SDValue &Base,
3527 SDValue &Offset, bool &isInc,
3528 SelectionDAG &DAG) {
3529 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3532 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3534 Base = Ptr->getOperand(0);
3535 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3536 int RHSC = (int)RHS->getZExtValue();
3537 if (RHSC < 0 && RHSC > -256) {
3538 assert(Ptr->getOpcode() == ISD::ADD);
3540 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3544 isInc = (Ptr->getOpcode() == ISD::ADD);
3545 Offset = Ptr->getOperand(1);
3547 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3549 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3550 int RHSC = (int)RHS->getZExtValue();
3551 if (RHSC < 0 && RHSC > -0x1000) {
3552 assert(Ptr->getOpcode() == ISD::ADD);
3554 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3555 Base = Ptr->getOperand(0);
3560 if (Ptr->getOpcode() == ISD::ADD) {
3562 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3563 if (ShOpcVal != ARM_AM::no_shift) {
3564 Base = Ptr->getOperand(1);
3565 Offset = Ptr->getOperand(0);
3567 Base = Ptr->getOperand(0);
3568 Offset = Ptr->getOperand(1);
3573 isInc = (Ptr->getOpcode() == ISD::ADD);
3574 Base = Ptr->getOperand(0);
3575 Offset = Ptr->getOperand(1);
3579 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3583 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3584 bool isSEXTLoad, SDValue &Base,
3585 SDValue &Offset, bool &isInc,
3586 SelectionDAG &DAG) {
3587 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3590 Base = Ptr->getOperand(0);
3591 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3592 int RHSC = (int)RHS->getZExtValue();
3593 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3594 assert(Ptr->getOpcode() == ISD::ADD);
3596 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3598 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3599 isInc = Ptr->getOpcode() == ISD::ADD;
3600 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3608 /// getPreIndexedAddressParts - returns true by value, base pointer and
3609 /// offset pointer and addressing mode by reference if the node's address
3610 /// can be legally represented as pre-indexed load / store address.
3612 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3614 ISD::MemIndexedMode &AM,
3615 SelectionDAG &DAG) const {
3616 if (Subtarget->isThumb1Only())
3621 bool isSEXTLoad = false;
3622 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3623 Ptr = LD->getBasePtr();
3624 VT = LD->getMemoryVT();
3625 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3626 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3627 Ptr = ST->getBasePtr();
3628 VT = ST->getMemoryVT();
3633 bool isLegal = false;
3634 if (Subtarget->isThumb2())
3635 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3636 Offset, isInc, DAG);
3638 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3639 Offset, isInc, DAG);
3643 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3647 /// getPostIndexedAddressParts - returns true by value, base pointer and
3648 /// offset pointer and addressing mode by reference if this node can be
3649 /// combined with a load / store to form a post-indexed load / store.
3650 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3653 ISD::MemIndexedMode &AM,
3654 SelectionDAG &DAG) const {
3655 if (Subtarget->isThumb1Only())
3660 bool isSEXTLoad = false;
3661 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3662 VT = LD->getMemoryVT();
3663 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3664 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3665 VT = ST->getMemoryVT();
3670 bool isLegal = false;
3671 if (Subtarget->isThumb2())
3672 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3675 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3680 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3684 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3688 const SelectionDAG &DAG,
3689 unsigned Depth) const {
3690 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3691 switch (Op.getOpcode()) {
3693 case ARMISD::CMOV: {
3694 // Bits are known zero/one if known on the LHS and RHS.
3695 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3696 if (KnownZero == 0 && KnownOne == 0) return;
3698 APInt KnownZeroRHS, KnownOneRHS;
3699 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3700 KnownZeroRHS, KnownOneRHS, Depth+1);
3701 KnownZero &= KnownZeroRHS;
3702 KnownOne &= KnownOneRHS;
3708 //===----------------------------------------------------------------------===//
3709 // ARM Inline Assembly Support
3710 //===----------------------------------------------------------------------===//
3712 /// getConstraintType - Given a constraint letter, return the type of
3713 /// constraint it is for this target.
3714 ARMTargetLowering::ConstraintType
3715 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3716 if (Constraint.size() == 1) {
3717 switch (Constraint[0]) {
3719 case 'l': return C_RegisterClass;
3720 case 'w': return C_RegisterClass;
3723 return TargetLowering::getConstraintType(Constraint);
3726 std::pair<unsigned, const TargetRegisterClass*>
3727 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3729 if (Constraint.size() == 1) {
3730 // GCC RS6000 Constraint Letters
3731 switch (Constraint[0]) {
3733 if (Subtarget->isThumb1Only())
3734 return std::make_pair(0U, ARM::tGPRRegisterClass);
3736 return std::make_pair(0U, ARM::GPRRegisterClass);
3738 return std::make_pair(0U, ARM::GPRRegisterClass);
3741 return std::make_pair(0U, ARM::SPRRegisterClass);
3743 return std::make_pair(0U, ARM::DPRRegisterClass);
3747 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3750 std::vector<unsigned> ARMTargetLowering::
3751 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3753 if (Constraint.size() != 1)
3754 return std::vector<unsigned>();
3756 switch (Constraint[0]) { // GCC ARM Constraint Letters
3759 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3760 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3763 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3764 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3765 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3766 ARM::R12, ARM::LR, 0);
3769 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3770 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3771 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3772 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3773 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3774 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3775 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3776 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3778 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3779 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3780 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3781 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3785 return std::vector<unsigned>();
3788 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3789 /// vector. If it is invalid, don't add anything to Ops.
3790 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3793 std::vector<SDValue>&Ops,
3794 SelectionDAG &DAG) const {
3795 SDValue Result(0, 0);
3797 switch (Constraint) {
3799 case 'I': case 'J': case 'K': case 'L':
3800 case 'M': case 'N': case 'O':
3801 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3805 int64_t CVal64 = C->getSExtValue();
3806 int CVal = (int) CVal64;
3807 // None of these constraints allow values larger than 32 bits. Check
3808 // that the value fits in an int.
3812 switch (Constraint) {
3814 if (Subtarget->isThumb1Only()) {
3815 // This must be a constant between 0 and 255, for ADD
3817 if (CVal >= 0 && CVal <= 255)
3819 } else if (Subtarget->isThumb2()) {
3820 // A constant that can be used as an immediate value in a
3821 // data-processing instruction.
3822 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3825 // A constant that can be used as an immediate value in a
3826 // data-processing instruction.
3827 if (ARM_AM::getSOImmVal(CVal) != -1)
3833 if (Subtarget->isThumb()) { // FIXME thumb2
3834 // This must be a constant between -255 and -1, for negated ADD
3835 // immediates. This can be used in GCC with an "n" modifier that
3836 // prints the negated value, for use with SUB instructions. It is
3837 // not useful otherwise but is implemented for compatibility.
3838 if (CVal >= -255 && CVal <= -1)
3841 // This must be a constant between -4095 and 4095. It is not clear
3842 // what this constraint is intended for. Implemented for
3843 // compatibility with GCC.
3844 if (CVal >= -4095 && CVal <= 4095)
3850 if (Subtarget->isThumb1Only()) {
3851 // A 32-bit value where only one byte has a nonzero value. Exclude
3852 // zero to match GCC. This constraint is used by GCC internally for
3853 // constants that can be loaded with a move/shift combination.
3854 // It is not useful otherwise but is implemented for compatibility.
3855 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3857 } else if (Subtarget->isThumb2()) {
3858 // A constant whose bitwise inverse can be used as an immediate
3859 // value in a data-processing instruction. This can be used in GCC
3860 // with a "B" modifier that prints the inverted value, for use with
3861 // BIC and MVN instructions. It is not useful otherwise but is
3862 // implemented for compatibility.
3863 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3866 // A constant whose bitwise inverse can be used as an immediate
3867 // value in a data-processing instruction. This can be used in GCC
3868 // with a "B" modifier that prints the inverted value, for use with
3869 // BIC and MVN instructions. It is not useful otherwise but is
3870 // implemented for compatibility.
3871 if (ARM_AM::getSOImmVal(~CVal) != -1)
3877 if (Subtarget->isThumb1Only()) {
3878 // This must be a constant between -7 and 7,
3879 // for 3-operand ADD/SUB immediate instructions.
3880 if (CVal >= -7 && CVal < 7)
3882 } else if (Subtarget->isThumb2()) {
3883 // A constant whose negation can be used as an immediate value in a
3884 // data-processing instruction. This can be used in GCC with an "n"
3885 // modifier that prints the negated value, for use with SUB
3886 // instructions. It is not useful otherwise but is implemented for
3888 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3891 // A constant whose negation can be used as an immediate value in a
3892 // data-processing instruction. This can be used in GCC with an "n"
3893 // modifier that prints the negated value, for use with SUB
3894 // instructions. It is not useful otherwise but is implemented for
3896 if (ARM_AM::getSOImmVal(-CVal) != -1)
3902 if (Subtarget->isThumb()) { // FIXME thumb2
3903 // This must be a multiple of 4 between 0 and 1020, for
3904 // ADD sp + immediate.
3905 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3908 // A power of two or a constant between 0 and 32. This is used in
3909 // GCC for the shift amount on shifted register operands, but it is
3910 // useful in general for any shift amounts.
3911 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3917 if (Subtarget->isThumb()) { // FIXME thumb2
3918 // This must be a constant between 0 and 31, for shift amounts.
3919 if (CVal >= 0 && CVal <= 31)
3925 if (Subtarget->isThumb()) { // FIXME thumb2
3926 // This must be a multiple of 4 between -508 and 508, for
3927 // ADD/SUB sp = sp + immediate.
3928 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3933 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3937 if (Result.getNode()) {
3938 Ops.push_back(Result);
3941 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,