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::VEXT: return "ARMISD::VEXT";
487 case ARMISD::VREV64: return "ARMISD::VREV64";
488 case ARMISD::VREV32: return "ARMISD::VREV32";
489 case ARMISD::VREV16: return "ARMISD::VREV16";
490 case ARMISD::VZIP: return "ARMISD::VZIP";
491 case ARMISD::VUZP: return "ARMISD::VUZP";
492 case ARMISD::VTRN: return "ARMISD::VTRN";
496 /// getFunctionAlignment - Return the Log2 alignment of this function.
497 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
498 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
501 //===----------------------------------------------------------------------===//
503 //===----------------------------------------------------------------------===//
505 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
506 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
508 default: llvm_unreachable("Unknown condition code!");
509 case ISD::SETNE: return ARMCC::NE;
510 case ISD::SETEQ: return ARMCC::EQ;
511 case ISD::SETGT: return ARMCC::GT;
512 case ISD::SETGE: return ARMCC::GE;
513 case ISD::SETLT: return ARMCC::LT;
514 case ISD::SETLE: return ARMCC::LE;
515 case ISD::SETUGT: return ARMCC::HI;
516 case ISD::SETUGE: return ARMCC::HS;
517 case ISD::SETULT: return ARMCC::LO;
518 case ISD::SETULE: return ARMCC::LS;
522 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
523 /// returns true if the operands should be inverted to form the proper
525 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
526 ARMCC::CondCodes &CondCode2) {
528 CondCode2 = ARMCC::AL;
530 default: llvm_unreachable("Unknown FP condition!");
532 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
534 case ISD::SETOGT: CondCode = ARMCC::GT; break;
536 case ISD::SETOGE: CondCode = ARMCC::GE; break;
537 case ISD::SETOLT: CondCode = ARMCC::MI; break;
538 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
539 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
540 case ISD::SETO: CondCode = ARMCC::VC; break;
541 case ISD::SETUO: CondCode = ARMCC::VS; break;
542 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
543 case ISD::SETUGT: CondCode = ARMCC::HI; break;
544 case ISD::SETUGE: CondCode = ARMCC::PL; break;
546 case ISD::SETULT: CondCode = ARMCC::LT; break;
548 case ISD::SETULE: CondCode = ARMCC::LE; break;
550 case ISD::SETUNE: CondCode = ARMCC::NE; break;
555 //===----------------------------------------------------------------------===//
556 // Calling Convention Implementation
557 //===----------------------------------------------------------------------===//
559 #include "ARMGenCallingConv.inc"
561 // APCS f64 is in register pairs, possibly split to stack
562 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
563 CCValAssign::LocInfo &LocInfo,
564 CCState &State, bool CanFail) {
565 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
567 // Try to get the first register.
568 if (unsigned Reg = State.AllocateReg(RegList, 4))
569 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
571 // For the 2nd half of a v2f64, do not fail.
575 // Put the whole thing on the stack.
576 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
577 State.AllocateStack(8, 4),
582 // Try to get the second register.
583 if (unsigned Reg = State.AllocateReg(RegList, 4))
584 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
586 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
587 State.AllocateStack(4, 4),
592 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
593 CCValAssign::LocInfo &LocInfo,
594 ISD::ArgFlagsTy &ArgFlags,
596 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
598 if (LocVT == MVT::v2f64 &&
599 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
601 return true; // we handled it
604 // AAPCS f64 is in aligned register pairs
605 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
606 CCValAssign::LocInfo &LocInfo,
607 CCState &State, bool CanFail) {
608 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
609 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
611 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
613 // For the 2nd half of a v2f64, do not just fail.
617 // Put the whole thing on the stack.
618 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
619 State.AllocateStack(8, 8),
625 for (i = 0; i < 2; ++i)
626 if (HiRegList[i] == Reg)
629 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
630 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
635 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
636 CCValAssign::LocInfo &LocInfo,
637 ISD::ArgFlagsTy &ArgFlags,
639 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
641 if (LocVT == MVT::v2f64 &&
642 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
644 return true; // we handled it
647 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
648 CCValAssign::LocInfo &LocInfo, CCState &State) {
649 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
650 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
652 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
654 return false; // we didn't handle it
657 for (i = 0; i < 2; ++i)
658 if (HiRegList[i] == Reg)
661 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
662 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
667 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
668 CCValAssign::LocInfo &LocInfo,
669 ISD::ArgFlagsTy &ArgFlags,
671 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
673 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
675 return true; // we handled it
678 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
679 CCValAssign::LocInfo &LocInfo,
680 ISD::ArgFlagsTy &ArgFlags,
682 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
686 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
687 /// given CallingConvention value.
688 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC,
690 bool isVarArg) const {
693 llvm_unreachable("Unsupported calling convention");
695 case CallingConv::Fast:
696 // Use target triple & subtarget features to do actual dispatch.
697 if (Subtarget->isAAPCS_ABI()) {
698 if (Subtarget->hasVFP2() &&
699 FloatABIType == FloatABI::Hard && !isVarArg)
700 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
702 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
704 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
705 case CallingConv::ARM_AAPCS_VFP:
706 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
707 case CallingConv::ARM_AAPCS:
708 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
709 case CallingConv::ARM_APCS:
710 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
714 /// LowerCallResult - Lower the result values of a call into the
715 /// appropriate copies out of appropriate physical registers.
717 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
718 unsigned CallConv, bool isVarArg,
719 const SmallVectorImpl<ISD::InputArg> &Ins,
720 DebugLoc dl, SelectionDAG &DAG,
721 SmallVectorImpl<SDValue> &InVals) {
723 // Assign locations to each value returned by this call.
724 SmallVector<CCValAssign, 16> RVLocs;
725 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
726 RVLocs, *DAG.getContext());
727 CCInfo.AnalyzeCallResult(Ins,
728 CCAssignFnForNode(CallConv, /* Return*/ true,
731 // Copy all of the result registers out of their specified physreg.
732 for (unsigned i = 0; i != RVLocs.size(); ++i) {
733 CCValAssign VA = RVLocs[i];
736 if (VA.needsCustom()) {
737 // Handle f64 or half of a v2f64.
738 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
740 Chain = Lo.getValue(1);
741 InFlag = Lo.getValue(2);
742 VA = RVLocs[++i]; // skip ahead to next loc
743 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
745 Chain = Hi.getValue(1);
746 InFlag = Hi.getValue(2);
747 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
749 if (VA.getLocVT() == MVT::v2f64) {
750 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
751 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
752 DAG.getConstant(0, MVT::i32));
754 VA = RVLocs[++i]; // skip ahead to next loc
755 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
756 Chain = Lo.getValue(1);
757 InFlag = Lo.getValue(2);
758 VA = RVLocs[++i]; // skip ahead to next loc
759 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
760 Chain = Hi.getValue(1);
761 InFlag = Hi.getValue(2);
762 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
763 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
764 DAG.getConstant(1, MVT::i32));
767 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
769 Chain = Val.getValue(1);
770 InFlag = Val.getValue(2);
773 switch (VA.getLocInfo()) {
774 default: llvm_unreachable("Unknown loc info!");
775 case CCValAssign::Full: break;
776 case CCValAssign::BCvt:
777 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
781 InVals.push_back(Val);
787 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
788 /// by "Src" to address "Dst" of size "Size". Alignment information is
789 /// specified by the specific parameter attribute. The copy will be passed as
790 /// a byval function parameter.
791 /// Sometimes what we are copying is the end of a larger object, the part that
792 /// does not fit in registers.
794 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
795 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
797 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
798 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
799 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
802 /// LowerMemOpCallTo - Store the argument to the stack.
804 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
805 SDValue StackPtr, SDValue Arg,
806 DebugLoc dl, SelectionDAG &DAG,
807 const CCValAssign &VA,
808 ISD::ArgFlagsTy Flags) {
809 unsigned LocMemOffset = VA.getLocMemOffset();
810 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
811 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
812 if (Flags.isByVal()) {
813 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
815 return DAG.getStore(Chain, dl, Arg, PtrOff,
816 PseudoSourceValue::getStack(), LocMemOffset);
819 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
820 SDValue Chain, SDValue &Arg,
821 RegsToPassVector &RegsToPass,
822 CCValAssign &VA, CCValAssign &NextVA,
824 SmallVector<SDValue, 8> &MemOpChains,
825 ISD::ArgFlagsTy Flags) {
827 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
828 DAG.getVTList(MVT::i32, MVT::i32), Arg);
829 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
831 if (NextVA.isRegLoc())
832 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
834 assert(NextVA.isMemLoc());
835 if (StackPtr.getNode() == 0)
836 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
838 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
844 /// LowerCall - Lowering a call into a callseq_start <-
845 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
848 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
849 unsigned CallConv, bool isVarArg,
851 const SmallVectorImpl<ISD::OutputArg> &Outs,
852 const SmallVectorImpl<ISD::InputArg> &Ins,
853 DebugLoc dl, SelectionDAG &DAG,
854 SmallVectorImpl<SDValue> &InVals) {
856 // Analyze operands of the call, assigning locations to each operand.
857 SmallVector<CCValAssign, 16> ArgLocs;
858 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
860 CCInfo.AnalyzeCallOperands(Outs,
861 CCAssignFnForNode(CallConv, /* Return*/ false,
864 // Get a count of how many bytes are to be pushed on the stack.
865 unsigned NumBytes = CCInfo.getNextStackOffset();
867 // Adjust the stack pointer for the new arguments...
868 // These operations are automatically eliminated by the prolog/epilog pass
869 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
871 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
873 RegsToPassVector RegsToPass;
874 SmallVector<SDValue, 8> MemOpChains;
876 // Walk the register/memloc assignments, inserting copies/loads. In the case
877 // of tail call optimization, arguments are handled later.
878 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
881 CCValAssign &VA = ArgLocs[i];
882 SDValue Arg = Outs[realArgIdx].Val;
883 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
885 // Promote the value if needed.
886 switch (VA.getLocInfo()) {
887 default: llvm_unreachable("Unknown loc info!");
888 case CCValAssign::Full: break;
889 case CCValAssign::SExt:
890 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
892 case CCValAssign::ZExt:
893 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
895 case CCValAssign::AExt:
896 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
898 case CCValAssign::BCvt:
899 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
903 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
904 if (VA.needsCustom()) {
905 if (VA.getLocVT() == MVT::v2f64) {
906 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
907 DAG.getConstant(0, MVT::i32));
908 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
909 DAG.getConstant(1, MVT::i32));
911 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
912 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
914 VA = ArgLocs[++i]; // skip ahead to next loc
916 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
917 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
919 assert(VA.isMemLoc());
920 if (StackPtr.getNode() == 0)
921 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
923 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
924 dl, DAG, VA, Flags));
927 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
928 StackPtr, MemOpChains, Flags);
930 } else if (VA.isRegLoc()) {
931 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
933 assert(VA.isMemLoc());
934 if (StackPtr.getNode() == 0)
935 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
937 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
938 dl, DAG, VA, Flags));
942 if (!MemOpChains.empty())
943 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
944 &MemOpChains[0], MemOpChains.size());
946 // Build a sequence of copy-to-reg nodes chained together with token chain
947 // and flag operands which copy the outgoing args into the appropriate regs.
949 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
950 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
951 RegsToPass[i].second, InFlag);
952 InFlag = Chain.getValue(1);
955 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
956 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
957 // node so that legalize doesn't hack it.
958 bool isDirect = false;
959 bool isARMFunc = false;
960 bool isLocalARMFunc = false;
961 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
962 GlobalValue *GV = G->getGlobal();
964 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
965 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
966 getTargetMachine().getRelocationModel() != Reloc::Static;
967 isARMFunc = !Subtarget->isThumb() || isStub;
968 // ARM call to a local ARM function is predicable.
969 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
970 // tBX takes a register source operand.
971 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
972 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
974 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
975 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
976 Callee = DAG.getLoad(getPointerTy(), dl,
977 DAG.getEntryNode(), CPAddr, NULL, 0);
978 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
979 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
980 getPointerTy(), Callee, PICLabel);
982 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
983 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
985 bool isStub = Subtarget->isTargetDarwin() &&
986 getTargetMachine().getRelocationModel() != Reloc::Static;
987 isARMFunc = !Subtarget->isThumb() || isStub;
988 // tBX takes a register source operand.
989 const char *Sym = S->getSymbol();
990 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
991 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
992 Sym, ARMPCLabelIndex,
994 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
995 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
996 Callee = DAG.getLoad(getPointerTy(), dl,
997 DAG.getEntryNode(), CPAddr, NULL, 0);
998 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
999 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1000 getPointerTy(), Callee, PICLabel);
1002 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1005 // FIXME: handle tail calls differently.
1007 if (Subtarget->isThumb()) {
1008 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1009 CallOpc = ARMISD::CALL_NOLINK;
1011 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1013 CallOpc = (isDirect || Subtarget->hasV5TOps())
1014 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1015 : ARMISD::CALL_NOLINK;
1017 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1018 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1019 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1020 InFlag = Chain.getValue(1);
1023 std::vector<SDValue> Ops;
1024 Ops.push_back(Chain);
1025 Ops.push_back(Callee);
1027 // Add argument registers to the end of the list so that they are known live
1029 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1030 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1031 RegsToPass[i].second.getValueType()));
1033 if (InFlag.getNode())
1034 Ops.push_back(InFlag);
1035 // Returns a chain and a flag for retval copy to use.
1036 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1037 &Ops[0], Ops.size());
1038 InFlag = Chain.getValue(1);
1040 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1041 DAG.getIntPtrConstant(0, true), InFlag);
1043 InFlag = Chain.getValue(1);
1045 // Handle result values, copying them out of physregs into vregs that we
1047 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1052 ARMTargetLowering::LowerReturn(SDValue Chain,
1053 unsigned CallConv, bool isVarArg,
1054 const SmallVectorImpl<ISD::OutputArg> &Outs,
1055 DebugLoc dl, SelectionDAG &DAG) {
1057 // CCValAssign - represent the assignment of the return value to a location.
1058 SmallVector<CCValAssign, 16> RVLocs;
1060 // CCState - Info about the registers and stack slots.
1061 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1064 // Analyze outgoing return values.
1065 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1068 // If this is the first return lowered for this function, add
1069 // the regs to the liveout set for the function.
1070 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1071 for (unsigned i = 0; i != RVLocs.size(); ++i)
1072 if (RVLocs[i].isRegLoc())
1073 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1078 // Copy the result values into the output registers.
1079 for (unsigned i = 0, realRVLocIdx = 0;
1081 ++i, ++realRVLocIdx) {
1082 CCValAssign &VA = RVLocs[i];
1083 assert(VA.isRegLoc() && "Can only return in registers!");
1085 SDValue Arg = Outs[realRVLocIdx].Val;
1087 switch (VA.getLocInfo()) {
1088 default: llvm_unreachable("Unknown loc info!");
1089 case CCValAssign::Full: break;
1090 case CCValAssign::BCvt:
1091 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1095 if (VA.needsCustom()) {
1096 if (VA.getLocVT() == MVT::v2f64) {
1097 // Extract the first half and return it in two registers.
1098 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1099 DAG.getConstant(0, MVT::i32));
1100 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
1101 DAG.getVTList(MVT::i32, MVT::i32), Half);
1103 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1104 Flag = Chain.getValue(1);
1105 VA = RVLocs[++i]; // skip ahead to next loc
1106 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1107 HalfGPRs.getValue(1), Flag);
1108 Flag = Chain.getValue(1);
1109 VA = RVLocs[++i]; // skip ahead to next loc
1111 // Extract the 2nd half and fall through to handle it as an f64 value.
1112 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1113 DAG.getConstant(1, MVT::i32));
1115 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1117 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
1118 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1119 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1120 Flag = Chain.getValue(1);
1121 VA = RVLocs[++i]; // skip ahead to next loc
1122 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1125 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1127 // Guarantee that all emitted copies are
1128 // stuck together, avoiding something bad.
1129 Flag = Chain.getValue(1);
1134 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1136 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1141 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1142 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1143 // one of the above mentioned nodes. It has to be wrapped because otherwise
1144 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1145 // be used to form addressing mode. These wrapped nodes will be selected
1147 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1148 EVT PtrVT = Op.getValueType();
1149 // FIXME there is no actual debug info here
1150 DebugLoc dl = Op.getDebugLoc();
1151 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1153 if (CP->isMachineConstantPoolEntry())
1154 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1155 CP->getAlignment());
1157 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1158 CP->getAlignment());
1159 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1162 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1164 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1165 SelectionDAG &DAG) {
1166 DebugLoc dl = GA->getDebugLoc();
1167 EVT PtrVT = getPointerTy();
1168 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1169 ARMConstantPoolValue *CPV =
1170 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1171 PCAdj, "tlsgd", true);
1172 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1173 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1174 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
1175 SDValue Chain = Argument.getValue(1);
1177 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1178 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1180 // call __tls_get_addr.
1183 Entry.Node = Argument;
1184 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1185 Args.push_back(Entry);
1186 // FIXME: is there useful debug info available here?
1187 std::pair<SDValue, SDValue> CallResult =
1188 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1189 false, false, false, false,
1190 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1191 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1192 return CallResult.first;
1195 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1196 // "local exec" model.
1198 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1199 SelectionDAG &DAG) {
1200 GlobalValue *GV = GA->getGlobal();
1201 DebugLoc dl = GA->getDebugLoc();
1203 SDValue Chain = DAG.getEntryNode();
1204 EVT PtrVT = getPointerTy();
1205 // Get the Thread Pointer
1206 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1208 if (GV->isDeclaration()) {
1209 // initial exec model
1210 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1211 ARMConstantPoolValue *CPV =
1212 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
1213 PCAdj, "gottpoff", true);
1214 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1215 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1216 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1217 Chain = Offset.getValue(1);
1219 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1220 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1222 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1225 ARMConstantPoolValue *CPV =
1226 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
1227 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1228 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1229 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
1232 // The address of the thread local variable is the add of the thread
1233 // pointer with the offset of the variable.
1234 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1238 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1239 // TODO: implement the "local dynamic" model
1240 assert(Subtarget->isTargetELF() &&
1241 "TLS not implemented for non-ELF targets");
1242 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1243 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1244 // otherwise use the "Local Exec" TLS Model
1245 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1246 return LowerToTLSGeneralDynamicModel(GA, DAG);
1248 return LowerToTLSExecModels(GA, DAG);
1251 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1252 SelectionDAG &DAG) {
1253 EVT PtrVT = getPointerTy();
1254 DebugLoc dl = Op.getDebugLoc();
1255 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1256 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1257 if (RelocM == Reloc::PIC_) {
1258 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1259 ARMConstantPoolValue *CPV =
1260 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
1261 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1262 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1263 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1265 SDValue Chain = Result.getValue(1);
1266 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1267 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1269 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1272 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1273 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1274 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1278 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
1279 /// even in non-static mode.
1280 static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
1281 // If symbol visibility is hidden, the extra load is not needed if
1282 // the symbol is definitely defined in the current translation unit.
1283 bool isDecl = GV->isDeclaration() || GV->hasAvailableExternallyLinkage();
1284 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage()))
1286 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker());
1289 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1290 SelectionDAG &DAG) {
1291 EVT PtrVT = getPointerTy();
1292 DebugLoc dl = Op.getDebugLoc();
1293 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1294 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1295 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
1297 if (RelocM == Reloc::Static)
1298 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1300 unsigned PCAdj = (RelocM != Reloc::PIC_)
1301 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1302 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
1304 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
1306 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1308 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1310 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1311 SDValue Chain = Result.getValue(1);
1313 if (RelocM == Reloc::PIC_) {
1314 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1315 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1318 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
1323 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1325 assert(Subtarget->isTargetELF() &&
1326 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1327 EVT PtrVT = getPointerTy();
1328 DebugLoc dl = Op.getDebugLoc();
1329 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1330 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1331 "_GLOBAL_OFFSET_TABLE_",
1333 ARMCP::CPValue, PCAdj);
1334 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1335 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1336 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1337 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1338 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1341 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
1343 SDNode *Node = Op.getNode();
1344 EVT VT = Node->getValueType(0);
1346 // No expansion needed for 64-bit vectors.
1347 if (VT.is64BitVector())
1350 // FIXME: We need to expand VLD3 and VLD4 of 128-bit vectors into separate
1351 // operations to load the even and odd registers.
1355 static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
1357 SDNode *Node = Op.getNode();
1358 EVT VT = Node->getOperand(3).getValueType();
1360 // No expansion needed for 64-bit vectors.
1361 if (VT.is64BitVector())
1364 // FIXME: We need to expand VST3 and VST4 of 128-bit vectors into separate
1365 // operations to store the even and odd registers.
1370 ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
1371 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1373 case Intrinsic::arm_neon_vld3:
1374 return LowerNeonVLDIntrinsic(Op, DAG, 3);
1375 case Intrinsic::arm_neon_vld4:
1376 return LowerNeonVLDIntrinsic(Op, DAG, 4);
1377 case Intrinsic::arm_neon_vst3:
1378 return LowerNeonVSTIntrinsic(Op, DAG, 3);
1379 case Intrinsic::arm_neon_vst4:
1380 return LowerNeonVSTIntrinsic(Op, DAG, 4);
1381 default: return SDValue(); // Don't custom lower most intrinsics.
1386 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1387 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1388 DebugLoc dl = Op.getDebugLoc();
1390 default: return SDValue(); // Don't custom lower most intrinsics.
1391 case Intrinsic::arm_thread_pointer: {
1392 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1393 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1395 case Intrinsic::eh_sjlj_lsda: {
1396 // blah. horrible, horrible hack with the forced magic name.
1397 // really need to clean this up. It belongs in the target-independent
1398 // layer somehow that doesn't require the coupling with the asm
1400 MachineFunction &MF = DAG.getMachineFunction();
1401 EVT PtrVT = getPointerTy();
1402 DebugLoc dl = Op.getDebugLoc();
1403 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1405 unsigned PCAdj = (RelocM != Reloc::PIC_)
1406 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1407 ARMCP::ARMCPKind Kind = ARMCP::CPValue;
1408 // Save off the LSDA name for the AsmPrinter to use when it's time
1409 // to emit the table
1410 std::string LSDAName = "L_lsda_";
1411 LSDAName += MF.getFunction()->getName();
1412 ARMConstantPoolValue *CPV =
1413 new ARMConstantPoolValue(*DAG.getContext(), LSDAName.c_str(),
1414 ARMPCLabelIndex, Kind, PCAdj);
1415 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1416 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1418 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
1419 SDValue Chain = Result.getValue(1);
1421 if (RelocM == Reloc::PIC_) {
1422 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
1423 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1427 case Intrinsic::eh_sjlj_setjmp:
1428 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1432 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1433 unsigned VarArgsFrameIndex) {
1434 // vastart just stores the address of the VarArgsFrameIndex slot into the
1435 // memory location argument.
1436 DebugLoc dl = Op.getDebugLoc();
1437 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1438 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1439 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1440 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1444 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1445 SDNode *Node = Op.getNode();
1446 DebugLoc dl = Node->getDebugLoc();
1447 EVT VT = Node->getValueType(0);
1448 SDValue Chain = Op.getOperand(0);
1449 SDValue Size = Op.getOperand(1);
1450 SDValue Align = Op.getOperand(2);
1452 // Chain the dynamic stack allocation so that it doesn't modify the stack
1453 // pointer when other instructions are using the stack.
1454 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1456 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1457 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1458 if (AlignVal > StackAlign)
1459 // Do this now since selection pass cannot introduce new target
1460 // independent node.
1461 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1463 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1464 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1465 // do even more horrible hack later.
1466 MachineFunction &MF = DAG.getMachineFunction();
1467 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1468 if (AFI->isThumb1OnlyFunction()) {
1470 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1472 uint32_t Val = C->getZExtValue();
1473 if (Val <= 508 && ((Val & 3) == 0))
1477 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1480 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1481 SDValue Ops1[] = { Chain, Size, Align };
1482 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1483 Chain = Res.getValue(1);
1484 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1485 DAG.getIntPtrConstant(0, true), SDValue());
1486 SDValue Ops2[] = { Res, Chain };
1487 return DAG.getMergeValues(Ops2, 2, dl);
1491 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1492 SDValue &Root, SelectionDAG &DAG,
1494 MachineFunction &MF = DAG.getMachineFunction();
1495 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1497 TargetRegisterClass *RC;
1498 if (AFI->isThumb1OnlyFunction())
1499 RC = ARM::tGPRRegisterClass;
1501 RC = ARM::GPRRegisterClass;
1503 // Transform the arguments stored in physical registers into virtual ones.
1504 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1505 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1508 if (NextVA.isMemLoc()) {
1509 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1510 MachineFrameInfo *MFI = MF.getFrameInfo();
1511 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
1513 // Create load node to retrieve arguments from the stack.
1514 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1515 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
1517 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1518 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1521 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
1525 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1526 unsigned CallConv, bool isVarArg,
1527 const SmallVectorImpl<ISD::InputArg>
1529 DebugLoc dl, SelectionDAG &DAG,
1530 SmallVectorImpl<SDValue> &InVals) {
1532 MachineFunction &MF = DAG.getMachineFunction();
1533 MachineFrameInfo *MFI = MF.getFrameInfo();
1535 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1537 // Assign locations to all of the incoming arguments.
1538 SmallVector<CCValAssign, 16> ArgLocs;
1539 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1541 CCInfo.AnalyzeFormalArguments(Ins,
1542 CCAssignFnForNode(CallConv, /* Return*/ false,
1545 SmallVector<SDValue, 16> ArgValues;
1547 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1548 CCValAssign &VA = ArgLocs[i];
1550 // Arguments stored in registers.
1551 if (VA.isRegLoc()) {
1552 EVT RegVT = VA.getLocVT();
1555 if (VA.needsCustom()) {
1556 // f64 and vector types are split up into multiple registers or
1557 // combinations of registers and stack slots.
1560 if (VA.getLocVT() == MVT::v2f64) {
1561 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1563 VA = ArgLocs[++i]; // skip ahead to next loc
1564 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1566 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1567 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1568 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1569 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1570 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1572 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1575 TargetRegisterClass *RC;
1577 if (RegVT == MVT::f32)
1578 RC = ARM::SPRRegisterClass;
1579 else if (RegVT == MVT::f64)
1580 RC = ARM::DPRRegisterClass;
1581 else if (RegVT == MVT::v2f64)
1582 RC = ARM::QPRRegisterClass;
1583 else if (RegVT == MVT::i32)
1584 RC = (AFI->isThumb1OnlyFunction() ?
1585 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1587 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1589 // Transform the arguments in physical registers into virtual ones.
1590 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1591 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1594 // If this is an 8 or 16-bit value, it is really passed promoted
1595 // to 32 bits. Insert an assert[sz]ext to capture this, then
1596 // truncate to the right size.
1597 switch (VA.getLocInfo()) {
1598 default: llvm_unreachable("Unknown loc info!");
1599 case CCValAssign::Full: break;
1600 case CCValAssign::BCvt:
1601 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1603 case CCValAssign::SExt:
1604 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1605 DAG.getValueType(VA.getValVT()));
1606 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1608 case CCValAssign::ZExt:
1609 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1610 DAG.getValueType(VA.getValVT()));
1611 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1615 InVals.push_back(ArgValue);
1617 } else { // VA.isRegLoc()
1620 assert(VA.isMemLoc());
1621 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1623 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1624 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
1626 // Create load nodes to retrieve arguments from the stack.
1627 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1628 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1634 static const unsigned GPRArgRegs[] = {
1635 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1638 unsigned NumGPRs = CCInfo.getFirstUnallocated
1639 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1641 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1642 unsigned VARegSize = (4 - NumGPRs) * 4;
1643 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1644 unsigned ArgOffset = 0;
1645 if (VARegSaveSize) {
1646 // If this function is vararg, store any remaining integer argument regs
1647 // to their spots on the stack so that they may be loaded by deferencing
1648 // the result of va_next.
1649 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1650 ArgOffset = CCInfo.getNextStackOffset();
1651 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1652 VARegSaveSize - VARegSize);
1653 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1655 SmallVector<SDValue, 4> MemOps;
1656 for (; NumGPRs < 4; ++NumGPRs) {
1657 TargetRegisterClass *RC;
1658 if (AFI->isThumb1OnlyFunction())
1659 RC = ARM::tGPRRegisterClass;
1661 RC = ARM::GPRRegisterClass;
1663 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1664 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1665 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1666 MemOps.push_back(Store);
1667 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1668 DAG.getConstant(4, getPointerTy()));
1670 if (!MemOps.empty())
1671 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1672 &MemOps[0], MemOps.size());
1674 // This will point to the next argument passed via stack.
1675 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1681 /// isFloatingPointZero - Return true if this is +0.0.
1682 static bool isFloatingPointZero(SDValue Op) {
1683 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1684 return CFP->getValueAPF().isPosZero();
1685 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1686 // Maybe this has already been legalized into the constant pool?
1687 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1688 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1689 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1690 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1691 return CFP->getValueAPF().isPosZero();
1697 static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
1698 return ( isThumb1Only && (C & ~255U) == 0) ||
1699 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
1702 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1703 /// the given operands.
1704 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1705 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
1707 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1708 unsigned C = RHSC->getZExtValue();
1709 if (!isLegalCmpImmediate(C, isThumb1Only)) {
1710 // Constant does not fit, try adjusting it by one?
1715 if (isLegalCmpImmediate(C-1, isThumb1Only)) {
1716 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1717 RHS = DAG.getConstant(C-1, MVT::i32);
1722 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
1723 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1724 RHS = DAG.getConstant(C-1, MVT::i32);
1729 if (isLegalCmpImmediate(C+1, isThumb1Only)) {
1730 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1731 RHS = DAG.getConstant(C+1, MVT::i32);
1736 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
1737 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1738 RHS = DAG.getConstant(C+1, MVT::i32);
1745 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1746 ARMISD::NodeType CompareType;
1749 CompareType = ARMISD::CMP;
1754 CompareType = ARMISD::CMPZ;
1757 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1758 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1761 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1762 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1765 if (!isFloatingPointZero(RHS))
1766 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1768 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1769 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1772 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1773 const ARMSubtarget *ST) {
1774 EVT VT = Op.getValueType();
1775 SDValue LHS = Op.getOperand(0);
1776 SDValue RHS = Op.getOperand(1);
1777 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1778 SDValue TrueVal = Op.getOperand(2);
1779 SDValue FalseVal = Op.getOperand(3);
1780 DebugLoc dl = Op.getDebugLoc();
1782 if (LHS.getValueType() == MVT::i32) {
1784 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1785 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1786 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1789 ARMCC::CondCodes CondCode, CondCode2;
1790 if (FPCCToARMCC(CC, CondCode, CondCode2))
1791 std::swap(TrueVal, FalseVal);
1793 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1794 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1795 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1796 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1798 if (CondCode2 != ARMCC::AL) {
1799 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1800 // FIXME: Needs another CMP because flag can have but one use.
1801 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1802 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1803 Result, TrueVal, ARMCC2, CCR, Cmp2);
1808 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1809 const ARMSubtarget *ST) {
1810 SDValue Chain = Op.getOperand(0);
1811 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1812 SDValue LHS = Op.getOperand(2);
1813 SDValue RHS = Op.getOperand(3);
1814 SDValue Dest = Op.getOperand(4);
1815 DebugLoc dl = Op.getDebugLoc();
1817 if (LHS.getValueType() == MVT::i32) {
1819 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1820 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
1821 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1822 Chain, Dest, ARMCC, CCR,Cmp);
1825 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1826 ARMCC::CondCodes CondCode, CondCode2;
1827 if (FPCCToARMCC(CC, CondCode, CondCode2))
1828 // Swap the LHS/RHS of the comparison if needed.
1829 std::swap(LHS, RHS);
1831 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1832 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1833 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1834 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1835 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1836 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1837 if (CondCode2 != ARMCC::AL) {
1838 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1839 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1840 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1845 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1846 SDValue Chain = Op.getOperand(0);
1847 SDValue Table = Op.getOperand(1);
1848 SDValue Index = Op.getOperand(2);
1849 DebugLoc dl = Op.getDebugLoc();
1851 EVT PTy = getPointerTy();
1852 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1853 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1854 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1855 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1856 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1857 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1858 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1859 if (Subtarget->isThumb2()) {
1860 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1861 // which does another jump to the destination. This also makes it easier
1862 // to translate it to TBB / TBH later.
1863 // FIXME: This might not work if the function is extremely large.
1864 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1865 Addr, Op.getOperand(2), JTI, UId);
1867 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1868 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
1869 Chain = Addr.getValue(1);
1870 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1871 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1873 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
1874 Chain = Addr.getValue(1);
1875 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1879 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1880 DebugLoc dl = Op.getDebugLoc();
1882 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1883 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1884 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1887 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1888 EVT VT = Op.getValueType();
1889 DebugLoc dl = Op.getDebugLoc();
1891 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1893 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1894 return DAG.getNode(Opc, dl, VT, Op);
1897 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1898 // Implement fcopysign with a fabs and a conditional fneg.
1899 SDValue Tmp0 = Op.getOperand(0);
1900 SDValue Tmp1 = Op.getOperand(1);
1901 DebugLoc dl = Op.getDebugLoc();
1902 EVT VT = Op.getValueType();
1903 EVT SrcVT = Tmp1.getValueType();
1904 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1905 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1906 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1907 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1908 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1911 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1912 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1913 MFI->setFrameAddressIsTaken(true);
1914 EVT VT = Op.getValueType();
1915 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1916 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1917 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1918 ? ARM::R7 : ARM::R11;
1919 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1921 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1926 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1928 SDValue Dst, SDValue Src,
1929 SDValue Size, unsigned Align,
1931 const Value *DstSV, uint64_t DstSVOff,
1932 const Value *SrcSV, uint64_t SrcSVOff){
1933 // Do repeated 4-byte loads and stores. To be improved.
1934 // This requires 4-byte alignment.
1935 if ((Align & 3) != 0)
1937 // This requires the copy size to be a constant, preferrably
1938 // within a subtarget-specific limit.
1939 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1942 uint64_t SizeVal = ConstantSize->getZExtValue();
1943 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1946 unsigned BytesLeft = SizeVal & 3;
1947 unsigned NumMemOps = SizeVal >> 2;
1948 unsigned EmittedNumMemOps = 0;
1950 unsigned VTSize = 4;
1952 const unsigned MAX_LOADS_IN_LDM = 6;
1953 SDValue TFOps[MAX_LOADS_IN_LDM];
1954 SDValue Loads[MAX_LOADS_IN_LDM];
1955 uint64_t SrcOff = 0, DstOff = 0;
1957 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1958 // same number of stores. The loads and stores will get combined into
1959 // ldm/stm later on.
1960 while (EmittedNumMemOps < NumMemOps) {
1962 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1963 Loads[i] = DAG.getLoad(VT, dl, Chain,
1964 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1965 DAG.getConstant(SrcOff, MVT::i32)),
1966 SrcSV, SrcSVOff + SrcOff);
1967 TFOps[i] = Loads[i].getValue(1);
1970 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1973 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1974 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1975 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1976 DAG.getConstant(DstOff, MVT::i32)),
1977 DstSV, DstSVOff + DstOff);
1980 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1982 EmittedNumMemOps += i;
1988 // Issue loads / stores for the trailing (1 - 3) bytes.
1989 unsigned BytesLeftSave = BytesLeft;
1992 if (BytesLeft >= 2) {
2000 Loads[i] = DAG.getLoad(VT, dl, Chain,
2001 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2002 DAG.getConstant(SrcOff, MVT::i32)),
2003 SrcSV, SrcSVOff + SrcOff);
2004 TFOps[i] = Loads[i].getValue(1);
2007 BytesLeft -= VTSize;
2009 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2012 BytesLeft = BytesLeftSave;
2014 if (BytesLeft >= 2) {
2022 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2023 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2024 DAG.getConstant(DstOff, MVT::i32)),
2025 DstSV, DstSVOff + DstOff);
2028 BytesLeft -= VTSize;
2030 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2033 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2034 SDValue Op = N->getOperand(0);
2035 DebugLoc dl = N->getDebugLoc();
2036 if (N->getValueType(0) == MVT::f64) {
2037 // Turn i64->f64 into FMDRR.
2038 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2039 DAG.getConstant(0, MVT::i32));
2040 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2041 DAG.getConstant(1, MVT::i32));
2042 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
2045 // Turn f64->i64 into FMRRD.
2046 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
2047 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2049 // Merge the pieces into a single i64 value.
2050 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2053 /// getZeroVector - Returns a vector of specified type with all zero elements.
2055 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2056 assert(VT.isVector() && "Expected a vector type");
2058 // Zero vectors are used to represent vector negation and in those cases
2059 // will be implemented with the NEON VNEG instruction. However, VNEG does
2060 // not support i64 elements, so sometimes the zero vectors will need to be
2061 // explicitly constructed. For those cases, and potentially other uses in
2062 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
2063 // to their dest type. This ensures they get CSE'd.
2065 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2066 if (VT.getSizeInBits() == 64)
2067 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2069 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2071 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2074 /// getOnesVector - Returns a vector of specified type with all bits set.
2076 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2077 assert(VT.isVector() && "Expected a vector type");
2079 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2080 // type. This ensures they get CSE'd.
2082 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2083 if (VT.getSizeInBits() == 64)
2084 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2086 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2088 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2091 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2092 const ARMSubtarget *ST) {
2093 EVT VT = N->getValueType(0);
2094 DebugLoc dl = N->getDebugLoc();
2096 // Lower vector shifts on NEON to use VSHL.
2097 if (VT.isVector()) {
2098 assert(ST->hasNEON() && "unexpected vector shift");
2100 // Left shifts translate directly to the vshiftu intrinsic.
2101 if (N->getOpcode() == ISD::SHL)
2102 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2103 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2104 N->getOperand(0), N->getOperand(1));
2106 assert((N->getOpcode() == ISD::SRA ||
2107 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2109 // NEON uses the same intrinsics for both left and right shifts. For
2110 // right shifts, the shift amounts are negative, so negate the vector of
2112 EVT ShiftVT = N->getOperand(1).getValueType();
2113 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2114 getZeroVector(ShiftVT, DAG, dl),
2116 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2117 Intrinsic::arm_neon_vshifts :
2118 Intrinsic::arm_neon_vshiftu);
2119 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2120 DAG.getConstant(vshiftInt, MVT::i32),
2121 N->getOperand(0), NegatedCount);
2124 // We can get here for a node like i32 = ISD::SHL i32, i64
2128 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2129 "Unknown shift to lower!");
2131 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2132 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2133 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2136 // If we are in thumb mode, we don't have RRX.
2137 if (ST->isThumb1Only()) return SDValue();
2139 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2140 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2141 DAG.getConstant(0, MVT::i32));
2142 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2143 DAG.getConstant(1, MVT::i32));
2145 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2146 // captures the result into a carry flag.
2147 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2148 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2150 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2151 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2153 // Merge the pieces into a single i64 value.
2154 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2157 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2158 SDValue TmpOp0, TmpOp1;
2159 bool Invert = false;
2163 SDValue Op0 = Op.getOperand(0);
2164 SDValue Op1 = Op.getOperand(1);
2165 SDValue CC = Op.getOperand(2);
2166 EVT VT = Op.getValueType();
2167 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2168 DebugLoc dl = Op.getDebugLoc();
2170 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2171 switch (SetCCOpcode) {
2172 default: llvm_unreachable("Illegal FP comparison"); break;
2174 case ISD::SETNE: Invert = true; // Fallthrough
2176 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2178 case ISD::SETLT: Swap = true; // Fallthrough
2180 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2182 case ISD::SETLE: Swap = true; // Fallthrough
2184 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2185 case ISD::SETUGE: Swap = true; // Fallthrough
2186 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2187 case ISD::SETUGT: Swap = true; // Fallthrough
2188 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2189 case ISD::SETUEQ: Invert = true; // Fallthrough
2191 // Expand this to (OLT | OGT).
2195 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2196 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2198 case ISD::SETUO: Invert = true; // Fallthrough
2200 // Expand this to (OLT | OGE).
2204 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2205 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2209 // Integer comparisons.
2210 switch (SetCCOpcode) {
2211 default: llvm_unreachable("Illegal integer comparison"); break;
2212 case ISD::SETNE: Invert = true;
2213 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2214 case ISD::SETLT: Swap = true;
2215 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2216 case ISD::SETLE: Swap = true;
2217 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2218 case ISD::SETULT: Swap = true;
2219 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2220 case ISD::SETULE: Swap = true;
2221 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2224 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2225 if (Opc == ARMISD::VCEQ) {
2228 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2230 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2233 // Ignore bitconvert.
2234 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2235 AndOp = AndOp.getOperand(0);
2237 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2239 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2240 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2247 std::swap(Op0, Op1);
2249 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2252 Result = DAG.getNOT(dl, Result, VT);
2257 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2258 /// VMOV instruction, and if so, return the constant being splatted.
2259 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2260 unsigned SplatBitSize, SelectionDAG &DAG) {
2261 switch (SplatBitSize) {
2263 // Any 1-byte value is OK.
2264 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2265 return DAG.getTargetConstant(SplatBits, MVT::i8);
2268 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2269 if ((SplatBits & ~0xff) == 0 ||
2270 (SplatBits & ~0xff00) == 0)
2271 return DAG.getTargetConstant(SplatBits, MVT::i16);
2275 // NEON's 32-bit VMOV supports splat values where:
2276 // * only one byte is nonzero, or
2277 // * the least significant byte is 0xff and the second byte is nonzero, or
2278 // * the least significant 2 bytes are 0xff and the third is nonzero.
2279 if ((SplatBits & ~0xff) == 0 ||
2280 (SplatBits & ~0xff00) == 0 ||
2281 (SplatBits & ~0xff0000) == 0 ||
2282 (SplatBits & ~0xff000000) == 0)
2283 return DAG.getTargetConstant(SplatBits, MVT::i32);
2285 if ((SplatBits & ~0xffff) == 0 &&
2286 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2287 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2289 if ((SplatBits & ~0xffffff) == 0 &&
2290 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2291 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2293 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2294 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2295 // VMOV.I32. A (very) minor optimization would be to replicate the value
2296 // and fall through here to test for a valid 64-bit splat. But, then the
2297 // caller would also need to check and handle the change in size.
2301 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2302 uint64_t BitMask = 0xff;
2304 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2305 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2307 else if ((SplatBits & BitMask) != 0)
2311 return DAG.getTargetConstant(Val, MVT::i64);
2315 llvm_unreachable("unexpected size for isVMOVSplat");
2322 /// getVMOVImm - If this is a build_vector of constants which can be
2323 /// formed by using a VMOV instruction of the specified element size,
2324 /// return the constant being splatted. The ByteSize field indicates the
2325 /// number of bytes of each element [1248].
2326 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2327 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2328 APInt SplatBits, SplatUndef;
2329 unsigned SplatBitSize;
2331 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2332 HasAnyUndefs, ByteSize * 8))
2335 if (SplatBitSize > ByteSize * 8)
2338 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2342 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2343 bool &ReverseVEXT, unsigned &Imm) {
2344 unsigned NumElts = VT.getVectorNumElements();
2345 ReverseVEXT = false;
2348 // If this is a VEXT shuffle, the immediate value is the index of the first
2349 // element. The other shuffle indices must be the successive elements after
2351 unsigned ExpectedElt = Imm;
2352 for (unsigned i = 1; i < NumElts; ++i) {
2353 // Increment the expected index. If it wraps around, it may still be
2354 // a VEXT but the source vectors must be swapped.
2356 if (ExpectedElt == NumElts * 2) {
2361 if (ExpectedElt != static_cast<unsigned>(M[i]))
2365 // Adjust the index value if the source operands will be swapped.
2372 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2373 /// instruction with the specified blocksize. (The order of the elements
2374 /// within each block of the vector is reversed.)
2375 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2376 unsigned BlockSize) {
2377 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2378 "Only possible block sizes for VREV are: 16, 32, 64");
2380 unsigned NumElts = VT.getVectorNumElements();
2381 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2382 unsigned BlockElts = M[0] + 1;
2384 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2387 for (unsigned i = 0; i < NumElts; ++i) {
2388 if ((unsigned) M[i] !=
2389 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2396 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2397 unsigned &WhichResult) {
2398 unsigned NumElts = VT.getVectorNumElements();
2399 WhichResult = (M[0] == 0 ? 0 : 1);
2400 for (unsigned i = 0; i < NumElts; i += 2) {
2401 if ((unsigned) M[i] != i + WhichResult ||
2402 (unsigned) M[i+1] != i + NumElts + WhichResult)
2408 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2409 unsigned &WhichResult) {
2410 unsigned NumElts = VT.getVectorNumElements();
2411 WhichResult = (M[0] == 0 ? 0 : 1);
2412 for (unsigned i = 0; i != NumElts; ++i) {
2413 if ((unsigned) M[i] != 2 * i + WhichResult)
2417 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2418 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2424 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2425 unsigned &WhichResult) {
2426 unsigned NumElts = VT.getVectorNumElements();
2427 WhichResult = (M[0] == 0 ? 0 : 1);
2428 unsigned Idx = WhichResult * NumElts / 2;
2429 for (unsigned i = 0; i != NumElts; i += 2) {
2430 if ((unsigned) M[i] != Idx ||
2431 (unsigned) M[i+1] != Idx + NumElts)
2436 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2437 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
2443 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2444 // Canonicalize all-zeros and all-ones vectors.
2445 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2446 if (ConstVal->isNullValue())
2447 return getZeroVector(VT, DAG, dl);
2448 if (ConstVal->isAllOnesValue())
2449 return getOnesVector(VT, DAG, dl);
2452 if (VT.is64BitVector()) {
2453 switch (Val.getValueType().getSizeInBits()) {
2454 case 8: CanonicalVT = MVT::v8i8; break;
2455 case 16: CanonicalVT = MVT::v4i16; break;
2456 case 32: CanonicalVT = MVT::v2i32; break;
2457 case 64: CanonicalVT = MVT::v1i64; break;
2458 default: llvm_unreachable("unexpected splat element type"); break;
2461 assert(VT.is128BitVector() && "unknown splat vector size");
2462 switch (Val.getValueType().getSizeInBits()) {
2463 case 8: CanonicalVT = MVT::v16i8; break;
2464 case 16: CanonicalVT = MVT::v8i16; break;
2465 case 32: CanonicalVT = MVT::v4i32; break;
2466 case 64: CanonicalVT = MVT::v2i64; break;
2467 default: llvm_unreachable("unexpected splat element type"); break;
2471 // Build a canonical splat for this value.
2472 SmallVector<SDValue, 8> Ops;
2473 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2474 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2476 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2479 // If this is a case we can't handle, return null and let the default
2480 // expansion code take care of it.
2481 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2482 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2483 DebugLoc dl = Op.getDebugLoc();
2484 EVT VT = Op.getValueType();
2486 APInt SplatBits, SplatUndef;
2487 unsigned SplatBitSize;
2489 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2490 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2491 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2493 return BuildSplat(Val, VT, DAG, dl);
2496 // If there are only 2 elements in a 128-bit vector, insert them into an
2497 // undef vector. This handles the common case for 128-bit vector argument
2498 // passing, where the insertions should be translated to subreg accesses
2499 // with no real instructions.
2500 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2501 SDValue Val = DAG.getUNDEF(VT);
2502 SDValue Op0 = Op.getOperand(0);
2503 SDValue Op1 = Op.getOperand(1);
2504 if (Op0.getOpcode() != ISD::UNDEF)
2505 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2506 DAG.getIntPtrConstant(0));
2507 if (Op1.getOpcode() != ISD::UNDEF)
2508 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2509 DAG.getIntPtrConstant(1));
2516 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2517 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2518 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2519 /// are assumed to be legal.
2521 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2523 if (VT.getVectorNumElements() == 4 &&
2524 (VT.is128BitVector() || VT.is64BitVector())) {
2525 unsigned PFIndexes[4];
2526 for (unsigned i = 0; i != 4; ++i) {
2530 PFIndexes[i] = M[i];
2533 // Compute the index in the perfect shuffle table.
2534 unsigned PFTableIndex =
2535 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2536 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2537 unsigned Cost = (PFEntry >> 30);
2544 unsigned Imm, WhichResult;
2546 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2547 isVREVMask(M, VT, 64) ||
2548 isVREVMask(M, VT, 32) ||
2549 isVREVMask(M, VT, 16) ||
2550 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2551 isVTRNMask(M, VT, WhichResult) ||
2552 isVUZPMask(M, VT, WhichResult) ||
2553 isVZIPMask(M, VT, WhichResult));
2556 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2557 /// the specified operations to build the shuffle.
2558 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2559 SDValue RHS, SelectionDAG &DAG,
2561 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2562 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2563 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2566 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2575 OP_VUZPL, // VUZP, left result
2576 OP_VUZPR, // VUZP, right result
2577 OP_VZIPL, // VZIP, left result
2578 OP_VZIPR, // VZIP, right result
2579 OP_VTRNL, // VTRN, left result
2580 OP_VTRNR // VTRN, right result
2583 if (OpNum == OP_COPY) {
2584 if (LHSID == (1*9+2)*9+3) return LHS;
2585 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2589 SDValue OpLHS, OpRHS;
2590 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2591 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2592 EVT VT = OpLHS.getValueType();
2595 default: llvm_unreachable("Unknown shuffle opcode!");
2597 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2602 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2603 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2607 return DAG.getNode(ARMISD::VEXT, dl, VT,
2609 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2612 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2613 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2616 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2617 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2620 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2621 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2625 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2626 SDValue V1 = Op.getOperand(0);
2627 SDValue V2 = Op.getOperand(1);
2628 DebugLoc dl = Op.getDebugLoc();
2629 EVT VT = Op.getValueType();
2630 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2631 SmallVector<int, 8> ShuffleMask;
2633 // Convert shuffles that are directly supported on NEON to target-specific
2634 // DAG nodes, instead of keeping them as shuffles and matching them again
2635 // during code selection. This is more efficient and avoids the possibility
2636 // of inconsistencies between legalization and selection.
2637 // FIXME: floating-point vectors should be canonicalized to integer vectors
2638 // of the same time so that they get CSEd properly.
2639 SVN->getMask(ShuffleMask);
2641 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2642 int Lane = SVN->getSplatIndex();
2643 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2644 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2646 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2647 DAG.getConstant(Lane, MVT::i32));
2652 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2655 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2656 DAG.getConstant(Imm, MVT::i32));
2659 if (isVREVMask(ShuffleMask, VT, 64))
2660 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2661 if (isVREVMask(ShuffleMask, VT, 32))
2662 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2663 if (isVREVMask(ShuffleMask, VT, 16))
2664 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2666 // Check for Neon shuffles that modify both input vectors in place.
2667 // If both results are used, i.e., if there are two shuffles with the same
2668 // source operands and with masks corresponding to both results of one of
2669 // these operations, DAG memoization will ensure that a single node is
2670 // used for both shuffles.
2671 unsigned WhichResult;
2672 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2673 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2674 V1, V2).getValue(WhichResult);
2675 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2676 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2677 V1, V2).getValue(WhichResult);
2678 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2679 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2680 V1, V2).getValue(WhichResult);
2682 // If the shuffle is not directly supported and it has 4 elements, use
2683 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2684 if (VT.getVectorNumElements() == 4 &&
2685 (VT.is128BitVector() || VT.is64BitVector())) {
2686 unsigned PFIndexes[4];
2687 for (unsigned i = 0; i != 4; ++i) {
2688 if (ShuffleMask[i] < 0)
2691 PFIndexes[i] = ShuffleMask[i];
2694 // Compute the index in the perfect shuffle table.
2695 unsigned PFTableIndex =
2696 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2698 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2699 unsigned Cost = (PFEntry >> 30);
2702 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2708 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
2709 EVT VT = Op.getValueType();
2710 EVT EltVT = VT.getVectorElementType();
2712 if (EltVT.isInteger())
2718 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2719 EVT VT = Op.getValueType();
2720 DebugLoc dl = Op.getDebugLoc();
2721 assert((VT == MVT::i8 || VT == MVT::i16) &&
2722 "unexpected type for custom-lowering vector extract");
2723 SDValue Vec = Op.getOperand(0);
2724 SDValue Lane = Op.getOperand(1);
2725 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2726 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
2727 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
2730 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2731 // The only time a CONCAT_VECTORS operation can have legal types is when
2732 // two 64-bit vectors are concatenated to a 128-bit vector.
2733 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2734 "unexpected CONCAT_VECTORS");
2735 DebugLoc dl = Op.getDebugLoc();
2736 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2737 SDValue Op0 = Op.getOperand(0);
2738 SDValue Op1 = Op.getOperand(1);
2739 if (Op0.getOpcode() != ISD::UNDEF)
2740 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2741 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2742 DAG.getIntPtrConstant(0));
2743 if (Op1.getOpcode() != ISD::UNDEF)
2744 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2745 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2746 DAG.getIntPtrConstant(1));
2747 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2750 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2751 switch (Op.getOpcode()) {
2752 default: llvm_unreachable("Don't know how to custom lower this!");
2753 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2754 case ISD::GlobalAddress:
2755 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2756 LowerGlobalAddressELF(Op, DAG);
2757 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2758 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
2759 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
2760 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2761 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2762 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2763 case ISD::SINT_TO_FP:
2764 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2765 case ISD::FP_TO_SINT:
2766 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2767 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2768 case ISD::RETURNADDR: break;
2769 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2770 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2771 case ISD::INTRINSIC_VOID:
2772 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
2773 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2774 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2777 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2778 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2779 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2780 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2781 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
2782 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2783 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2788 /// ReplaceNodeResults - Replace the results of node with an illegal result
2789 /// type with new values built out of custom code.
2790 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2791 SmallVectorImpl<SDValue>&Results,
2792 SelectionDAG &DAG) {
2793 switch (N->getOpcode()) {
2795 llvm_unreachable("Don't know how to custom expand this!");
2797 case ISD::BIT_CONVERT:
2798 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2802 SDValue Res = LowerShift(N, DAG, Subtarget);
2804 Results.push_back(Res);
2810 //===----------------------------------------------------------------------===//
2811 // ARM Scheduler Hooks
2812 //===----------------------------------------------------------------------===//
2815 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2816 MachineBasicBlock *BB) const {
2817 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2818 DebugLoc dl = MI->getDebugLoc();
2819 switch (MI->getOpcode()) {
2821 llvm_unreachable("Unexpected instr type to insert");
2822 case ARM::tMOVCCr_pseudo: {
2823 // To "insert" a SELECT_CC instruction, we actually have to insert the
2824 // diamond control-flow pattern. The incoming instruction knows the
2825 // destination vreg to set, the condition code register to branch on, the
2826 // true/false values to select between, and a branch opcode to use.
2827 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2828 MachineFunction::iterator It = BB;
2834 // cmpTY ccX, r1, r2
2836 // fallthrough --> copy0MBB
2837 MachineBasicBlock *thisMBB = BB;
2838 MachineFunction *F = BB->getParent();
2839 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2840 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2841 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2842 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2843 F->insert(It, copy0MBB);
2844 F->insert(It, sinkMBB);
2845 // Update machine-CFG edges by first adding all successors of the current
2846 // block to the new block which will contain the Phi node for the select.
2847 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2848 e = BB->succ_end(); i != e; ++i)
2849 sinkMBB->addSuccessor(*i);
2850 // Next, remove all successors of the current block, and add the true
2851 // and fallthrough blocks as its successors.
2852 while(!BB->succ_empty())
2853 BB->removeSuccessor(BB->succ_begin());
2854 BB->addSuccessor(copy0MBB);
2855 BB->addSuccessor(sinkMBB);
2858 // %FalseValue = ...
2859 // # fallthrough to sinkMBB
2862 // Update machine-CFG edges
2863 BB->addSuccessor(sinkMBB);
2866 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2869 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
2870 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
2871 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2873 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2880 case ARM::t2SUBrSPi_:
2881 case ARM::t2SUBrSPi12_:
2882 case ARM::t2SUBrSPs_: {
2883 MachineFunction *MF = BB->getParent();
2884 unsigned DstReg = MI->getOperand(0).getReg();
2885 unsigned SrcReg = MI->getOperand(1).getReg();
2886 bool DstIsDead = MI->getOperand(0).isDead();
2887 bool SrcIsKill = MI->getOperand(1).isKill();
2889 if (SrcReg != ARM::SP) {
2890 // Copy the source to SP from virtual register.
2891 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
2892 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2893 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
2894 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
2895 .addReg(SrcReg, getKillRegState(SrcIsKill));
2899 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
2900 switch (MI->getOpcode()) {
2902 llvm_unreachable("Unexpected pseudo instruction!");
2908 OpOpc = ARM::tADDspr;
2911 OpOpc = ARM::tSUBspi;
2913 case ARM::t2SUBrSPi_:
2914 OpOpc = ARM::t2SUBrSPi;
2915 NeedPred = true; NeedCC = true;
2917 case ARM::t2SUBrSPi12_:
2918 OpOpc = ARM::t2SUBrSPi12;
2921 case ARM::t2SUBrSPs_:
2922 OpOpc = ARM::t2SUBrSPs;
2923 NeedPred = true; NeedCC = true; NeedOp3 = true;
2926 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
2927 if (OpOpc == ARM::tAND)
2928 AddDefaultT1CC(MIB);
2929 MIB.addReg(ARM::SP);
2930 MIB.addOperand(MI->getOperand(2));
2932 MIB.addOperand(MI->getOperand(3));
2934 AddDefaultPred(MIB);
2938 // Copy the result from SP to virtual register.
2939 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
2940 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
2941 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
2942 BuildMI(BB, dl, TII->get(CopyOpc))
2943 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
2945 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
2951 //===----------------------------------------------------------------------===//
2952 // ARM Optimization Hooks
2953 //===----------------------------------------------------------------------===//
2956 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
2957 TargetLowering::DAGCombinerInfo &DCI) {
2958 SelectionDAG &DAG = DCI.DAG;
2959 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2960 EVT VT = N->getValueType(0);
2961 unsigned Opc = N->getOpcode();
2962 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
2963 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
2964 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
2965 ISD::CondCode CC = ISD::SETCC_INVALID;
2968 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
2970 SDValue CCOp = Slct.getOperand(0);
2971 if (CCOp.getOpcode() == ISD::SETCC)
2972 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
2975 bool DoXform = false;
2977 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
2980 if (LHS.getOpcode() == ISD::Constant &&
2981 cast<ConstantSDNode>(LHS)->isNullValue()) {
2983 } else if (CC != ISD::SETCC_INVALID &&
2984 RHS.getOpcode() == ISD::Constant &&
2985 cast<ConstantSDNode>(RHS)->isNullValue()) {
2986 std::swap(LHS, RHS);
2987 SDValue Op0 = Slct.getOperand(0);
2988 EVT OpVT = isSlctCC ? Op0.getValueType() :
2989 Op0.getOperand(0).getValueType();
2990 bool isInt = OpVT.isInteger();
2991 CC = ISD::getSetCCInverse(CC, isInt);
2993 if (!TLI.isCondCodeLegal(CC, OpVT))
2994 return SDValue(); // Inverse operator isn't legal.
3001 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3003 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3004 Slct.getOperand(0), Slct.getOperand(1), CC);
3005 SDValue CCOp = Slct.getOperand(0);
3007 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3008 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3009 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3010 CCOp, OtherOp, Result);
3015 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3016 static SDValue PerformADDCombine(SDNode *N,
3017 TargetLowering::DAGCombinerInfo &DCI) {
3018 // added by evan in r37685 with no testcase.
3019 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3021 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3022 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3023 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3024 if (Result.getNode()) return Result;
3026 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3027 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3028 if (Result.getNode()) return Result;
3034 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3035 static SDValue PerformSUBCombine(SDNode *N,
3036 TargetLowering::DAGCombinerInfo &DCI) {
3037 // added by evan in r37685 with no testcase.
3038 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3040 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3041 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3042 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3043 if (Result.getNode()) return Result;
3050 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
3051 static SDValue PerformFMRRDCombine(SDNode *N,
3052 TargetLowering::DAGCombinerInfo &DCI) {
3053 // fmrrd(fmdrr x, y) -> x,y
3054 SDValue InDouble = N->getOperand(0);
3055 if (InDouble.getOpcode() == ARMISD::FMDRR)
3056 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3060 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3061 /// operand of a vector shift operation, where all the elements of the
3062 /// build_vector must have the same constant integer value.
3063 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3064 // Ignore bit_converts.
3065 while (Op.getOpcode() == ISD::BIT_CONVERT)
3066 Op = Op.getOperand(0);
3067 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3068 APInt SplatBits, SplatUndef;
3069 unsigned SplatBitSize;
3071 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3072 HasAnyUndefs, ElementBits) ||
3073 SplatBitSize > ElementBits)
3075 Cnt = SplatBits.getSExtValue();
3079 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3080 /// operand of a vector shift left operation. That value must be in the range:
3081 /// 0 <= Value < ElementBits for a left shift; or
3082 /// 0 <= Value <= ElementBits for a long left shift.
3083 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3084 assert(VT.isVector() && "vector shift count is not a vector type");
3085 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3086 if (! getVShiftImm(Op, ElementBits, Cnt))
3088 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3091 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3092 /// operand of a vector shift right operation. For a shift opcode, the value
3093 /// is positive, but for an intrinsic the value count must be negative. The
3094 /// absolute value must be in the range:
3095 /// 1 <= |Value| <= ElementBits for a right shift; or
3096 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3097 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3099 assert(VT.isVector() && "vector shift count is not a vector type");
3100 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3101 if (! getVShiftImm(Op, ElementBits, Cnt))
3105 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3108 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3109 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3110 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3113 // Don't do anything for most intrinsics.
3116 // Vector shifts: check for immediate versions and lower them.
3117 // Note: This is done during DAG combining instead of DAG legalizing because
3118 // the build_vectors for 64-bit vector element shift counts are generally
3119 // not legal, and it is hard to see their values after they get legalized to
3120 // loads from a constant pool.
3121 case Intrinsic::arm_neon_vshifts:
3122 case Intrinsic::arm_neon_vshiftu:
3123 case Intrinsic::arm_neon_vshiftls:
3124 case Intrinsic::arm_neon_vshiftlu:
3125 case Intrinsic::arm_neon_vshiftn:
3126 case Intrinsic::arm_neon_vrshifts:
3127 case Intrinsic::arm_neon_vrshiftu:
3128 case Intrinsic::arm_neon_vrshiftn:
3129 case Intrinsic::arm_neon_vqshifts:
3130 case Intrinsic::arm_neon_vqshiftu:
3131 case Intrinsic::arm_neon_vqshiftsu:
3132 case Intrinsic::arm_neon_vqshiftns:
3133 case Intrinsic::arm_neon_vqshiftnu:
3134 case Intrinsic::arm_neon_vqshiftnsu:
3135 case Intrinsic::arm_neon_vqrshiftns:
3136 case Intrinsic::arm_neon_vqrshiftnu:
3137 case Intrinsic::arm_neon_vqrshiftnsu: {
3138 EVT VT = N->getOperand(1).getValueType();
3140 unsigned VShiftOpc = 0;
3143 case Intrinsic::arm_neon_vshifts:
3144 case Intrinsic::arm_neon_vshiftu:
3145 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3146 VShiftOpc = ARMISD::VSHL;
3149 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3150 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3151 ARMISD::VSHRs : ARMISD::VSHRu);
3156 case Intrinsic::arm_neon_vshiftls:
3157 case Intrinsic::arm_neon_vshiftlu:
3158 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3160 llvm_unreachable("invalid shift count for vshll intrinsic");
3162 case Intrinsic::arm_neon_vrshifts:
3163 case Intrinsic::arm_neon_vrshiftu:
3164 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3168 case Intrinsic::arm_neon_vqshifts:
3169 case Intrinsic::arm_neon_vqshiftu:
3170 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3174 case Intrinsic::arm_neon_vqshiftsu:
3175 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3177 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3179 case Intrinsic::arm_neon_vshiftn:
3180 case Intrinsic::arm_neon_vrshiftn:
3181 case Intrinsic::arm_neon_vqshiftns:
3182 case Intrinsic::arm_neon_vqshiftnu:
3183 case Intrinsic::arm_neon_vqshiftnsu:
3184 case Intrinsic::arm_neon_vqrshiftns:
3185 case Intrinsic::arm_neon_vqrshiftnu:
3186 case Intrinsic::arm_neon_vqrshiftnsu:
3187 // Narrowing shifts require an immediate right shift.
3188 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3190 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3193 llvm_unreachable("unhandled vector shift");
3197 case Intrinsic::arm_neon_vshifts:
3198 case Intrinsic::arm_neon_vshiftu:
3199 // Opcode already set above.
3201 case Intrinsic::arm_neon_vshiftls:
3202 case Intrinsic::arm_neon_vshiftlu:
3203 if (Cnt == VT.getVectorElementType().getSizeInBits())
3204 VShiftOpc = ARMISD::VSHLLi;
3206 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3207 ARMISD::VSHLLs : ARMISD::VSHLLu);
3209 case Intrinsic::arm_neon_vshiftn:
3210 VShiftOpc = ARMISD::VSHRN; break;
3211 case Intrinsic::arm_neon_vrshifts:
3212 VShiftOpc = ARMISD::VRSHRs; break;
3213 case Intrinsic::arm_neon_vrshiftu:
3214 VShiftOpc = ARMISD::VRSHRu; break;
3215 case Intrinsic::arm_neon_vrshiftn:
3216 VShiftOpc = ARMISD::VRSHRN; break;
3217 case Intrinsic::arm_neon_vqshifts:
3218 VShiftOpc = ARMISD::VQSHLs; break;
3219 case Intrinsic::arm_neon_vqshiftu:
3220 VShiftOpc = ARMISD::VQSHLu; break;
3221 case Intrinsic::arm_neon_vqshiftsu:
3222 VShiftOpc = ARMISD::VQSHLsu; break;
3223 case Intrinsic::arm_neon_vqshiftns:
3224 VShiftOpc = ARMISD::VQSHRNs; break;
3225 case Intrinsic::arm_neon_vqshiftnu:
3226 VShiftOpc = ARMISD::VQSHRNu; break;
3227 case Intrinsic::arm_neon_vqshiftnsu:
3228 VShiftOpc = ARMISD::VQSHRNsu; break;
3229 case Intrinsic::arm_neon_vqrshiftns:
3230 VShiftOpc = ARMISD::VQRSHRNs; break;
3231 case Intrinsic::arm_neon_vqrshiftnu:
3232 VShiftOpc = ARMISD::VQRSHRNu; break;
3233 case Intrinsic::arm_neon_vqrshiftnsu:
3234 VShiftOpc = ARMISD::VQRSHRNsu; break;
3237 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3238 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3241 case Intrinsic::arm_neon_vshiftins: {
3242 EVT VT = N->getOperand(1).getValueType();
3244 unsigned VShiftOpc = 0;
3246 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3247 VShiftOpc = ARMISD::VSLI;
3248 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3249 VShiftOpc = ARMISD::VSRI;
3251 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3254 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3255 N->getOperand(1), N->getOperand(2),
3256 DAG.getConstant(Cnt, MVT::i32));
3259 case Intrinsic::arm_neon_vqrshifts:
3260 case Intrinsic::arm_neon_vqrshiftu:
3261 // No immediate versions of these to check for.
3268 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3269 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3270 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3271 /// vector element shift counts are generally not legal, and it is hard to see
3272 /// their values after they get legalized to loads from a constant pool.
3273 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3274 const ARMSubtarget *ST) {
3275 EVT VT = N->getValueType(0);
3277 // Nothing to be done for scalar shifts.
3278 if (! VT.isVector())
3281 assert(ST->hasNEON() && "unexpected vector shift");
3284 switch (N->getOpcode()) {
3285 default: llvm_unreachable("unexpected shift opcode");
3288 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3289 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3290 DAG.getConstant(Cnt, MVT::i32));
3295 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3296 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3297 ARMISD::VSHRs : ARMISD::VSHRu);
3298 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3299 DAG.getConstant(Cnt, MVT::i32));
3305 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3306 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3307 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3308 const ARMSubtarget *ST) {
3309 SDValue N0 = N->getOperand(0);
3311 // Check for sign- and zero-extensions of vector extract operations of 8-
3312 // and 16-bit vector elements. NEON supports these directly. They are
3313 // handled during DAG combining because type legalization will promote them
3314 // to 32-bit types and it is messy to recognize the operations after that.
3315 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3316 SDValue Vec = N0.getOperand(0);
3317 SDValue Lane = N0.getOperand(1);
3318 EVT VT = N->getValueType(0);
3319 EVT EltVT = N0.getValueType();
3320 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3322 if (VT == MVT::i32 &&
3323 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3324 TLI.isTypeLegal(Vec.getValueType())) {
3327 switch (N->getOpcode()) {
3328 default: llvm_unreachable("unexpected opcode");
3329 case ISD::SIGN_EXTEND:
3330 Opc = ARMISD::VGETLANEs;
3332 case ISD::ZERO_EXTEND:
3333 case ISD::ANY_EXTEND:
3334 Opc = ARMISD::VGETLANEu;
3337 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3344 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3345 DAGCombinerInfo &DCI) const {
3346 switch (N->getOpcode()) {
3348 case ISD::ADD: return PerformADDCombine(N, DCI);
3349 case ISD::SUB: return PerformSUBCombine(N, DCI);
3350 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
3351 case ISD::INTRINSIC_WO_CHAIN:
3352 return PerformIntrinsicCombine(N, DCI.DAG);
3356 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3357 case ISD::SIGN_EXTEND:
3358 case ISD::ZERO_EXTEND:
3359 case ISD::ANY_EXTEND:
3360 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3365 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3366 if (!Subtarget->hasV6Ops())
3367 // Pre-v6 does not support unaligned mem access.
3369 else if (!Subtarget->hasV6Ops()) {
3370 // v6 may or may not support unaligned mem access.
3371 if (!Subtarget->isTargetDarwin())
3375 switch (VT.getSimpleVT().SimpleTy) {
3382 // FIXME: VLD1 etc with standard alignment is legal.
3386 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3391 switch (VT.getSimpleVT().SimpleTy) {
3392 default: return false;
3407 if ((V & (Scale - 1)) != 0)
3410 return V == (V & ((1LL << 5) - 1));
3413 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3414 const ARMSubtarget *Subtarget) {
3421 switch (VT.getSimpleVT().SimpleTy) {
3422 default: return false;
3427 // + imm12 or - imm8
3429 return V == (V & ((1LL << 8) - 1));
3430 return V == (V & ((1LL << 12) - 1));
3433 // Same as ARM mode. FIXME: NEON?
3434 if (!Subtarget->hasVFP2())
3439 return V == (V & ((1LL << 8) - 1));
3443 /// isLegalAddressImmediate - Return true if the integer value can be used
3444 /// as the offset of the target addressing mode for load / store of the
3446 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3447 const ARMSubtarget *Subtarget) {
3454 if (Subtarget->isThumb1Only())
3455 return isLegalT1AddressImmediate(V, VT);
3456 else if (Subtarget->isThumb2())
3457 return isLegalT2AddressImmediate(V, VT, Subtarget);
3462 switch (VT.getSimpleVT().SimpleTy) {
3463 default: return false;
3468 return V == (V & ((1LL << 12) - 1));
3471 return V == (V & ((1LL << 8) - 1));
3474 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3479 return V == (V & ((1LL << 8) - 1));
3483 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3485 int Scale = AM.Scale;
3489 switch (VT.getSimpleVT().SimpleTy) {
3490 default: return false;
3499 return Scale == 2 || Scale == 4 || Scale == 8;
3502 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3506 // Note, we allow "void" uses (basically, uses that aren't loads or
3507 // stores), because arm allows folding a scale into many arithmetic
3508 // operations. This should be made more precise and revisited later.
3510 // Allow r << imm, but the imm has to be a multiple of two.
3511 if (Scale & 1) return false;
3512 return isPowerOf2_32(Scale);
3516 /// isLegalAddressingMode - Return true if the addressing mode represented
3517 /// by AM is legal for this target, for a load/store of the specified type.
3518 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3519 const Type *Ty) const {
3520 EVT VT = getValueType(Ty, true);
3521 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3524 // Can never fold addr of global into load/store.
3529 case 0: // no scale reg, must be "r+i" or "r", or "i".
3532 if (Subtarget->isThumb1Only())
3536 // ARM doesn't support any R+R*scale+imm addr modes.
3543 if (Subtarget->isThumb2())
3544 return isLegalT2ScaledAddressingMode(AM, VT);
3546 int Scale = AM.Scale;
3547 switch (VT.getSimpleVT().SimpleTy) {
3548 default: return false;
3552 if (Scale < 0) Scale = -Scale;
3556 return isPowerOf2_32(Scale & ~1);
3560 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3565 // Note, we allow "void" uses (basically, uses that aren't loads or
3566 // stores), because arm allows folding a scale into many arithmetic
3567 // operations. This should be made more precise and revisited later.
3569 // Allow r << imm, but the imm has to be a multiple of two.
3570 if (Scale & 1) return false;
3571 return isPowerOf2_32(Scale);
3578 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3579 bool isSEXTLoad, SDValue &Base,
3580 SDValue &Offset, bool &isInc,
3581 SelectionDAG &DAG) {
3582 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3585 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3587 Base = Ptr->getOperand(0);
3588 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3589 int RHSC = (int)RHS->getZExtValue();
3590 if (RHSC < 0 && RHSC > -256) {
3591 assert(Ptr->getOpcode() == ISD::ADD);
3593 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3597 isInc = (Ptr->getOpcode() == ISD::ADD);
3598 Offset = Ptr->getOperand(1);
3600 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3602 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3603 int RHSC = (int)RHS->getZExtValue();
3604 if (RHSC < 0 && RHSC > -0x1000) {
3605 assert(Ptr->getOpcode() == ISD::ADD);
3607 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3608 Base = Ptr->getOperand(0);
3613 if (Ptr->getOpcode() == ISD::ADD) {
3615 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3616 if (ShOpcVal != ARM_AM::no_shift) {
3617 Base = Ptr->getOperand(1);
3618 Offset = Ptr->getOperand(0);
3620 Base = Ptr->getOperand(0);
3621 Offset = Ptr->getOperand(1);
3626 isInc = (Ptr->getOpcode() == ISD::ADD);
3627 Base = Ptr->getOperand(0);
3628 Offset = Ptr->getOperand(1);
3632 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
3636 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3637 bool isSEXTLoad, SDValue &Base,
3638 SDValue &Offset, bool &isInc,
3639 SelectionDAG &DAG) {
3640 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3643 Base = Ptr->getOperand(0);
3644 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3645 int RHSC = (int)RHS->getZExtValue();
3646 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3647 assert(Ptr->getOpcode() == ISD::ADD);
3649 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3651 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3652 isInc = Ptr->getOpcode() == ISD::ADD;
3653 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3661 /// getPreIndexedAddressParts - returns true by value, base pointer and
3662 /// offset pointer and addressing mode by reference if the node's address
3663 /// can be legally represented as pre-indexed load / store address.
3665 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3667 ISD::MemIndexedMode &AM,
3668 SelectionDAG &DAG) const {
3669 if (Subtarget->isThumb1Only())
3674 bool isSEXTLoad = false;
3675 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3676 Ptr = LD->getBasePtr();
3677 VT = LD->getMemoryVT();
3678 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3679 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3680 Ptr = ST->getBasePtr();
3681 VT = ST->getMemoryVT();
3686 bool isLegal = false;
3687 if (Subtarget->isThumb2())
3688 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3689 Offset, isInc, DAG);
3691 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3692 Offset, isInc, DAG);
3696 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3700 /// getPostIndexedAddressParts - returns true by value, base pointer and
3701 /// offset pointer and addressing mode by reference if this node can be
3702 /// combined with a load / store to form a post-indexed load / store.
3703 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3706 ISD::MemIndexedMode &AM,
3707 SelectionDAG &DAG) const {
3708 if (Subtarget->isThumb1Only())
3713 bool isSEXTLoad = false;
3714 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3715 VT = LD->getMemoryVT();
3716 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3717 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3718 VT = ST->getMemoryVT();
3723 bool isLegal = false;
3724 if (Subtarget->isThumb2())
3725 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3728 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3733 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3737 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3741 const SelectionDAG &DAG,
3742 unsigned Depth) const {
3743 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3744 switch (Op.getOpcode()) {
3746 case ARMISD::CMOV: {
3747 // Bits are known zero/one if known on the LHS and RHS.
3748 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3749 if (KnownZero == 0 && KnownOne == 0) return;
3751 APInt KnownZeroRHS, KnownOneRHS;
3752 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3753 KnownZeroRHS, KnownOneRHS, Depth+1);
3754 KnownZero &= KnownZeroRHS;
3755 KnownOne &= KnownOneRHS;
3761 //===----------------------------------------------------------------------===//
3762 // ARM Inline Assembly Support
3763 //===----------------------------------------------------------------------===//
3765 /// getConstraintType - Given a constraint letter, return the type of
3766 /// constraint it is for this target.
3767 ARMTargetLowering::ConstraintType
3768 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3769 if (Constraint.size() == 1) {
3770 switch (Constraint[0]) {
3772 case 'l': return C_RegisterClass;
3773 case 'w': return C_RegisterClass;
3776 return TargetLowering::getConstraintType(Constraint);
3779 std::pair<unsigned, const TargetRegisterClass*>
3780 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3782 if (Constraint.size() == 1) {
3783 // GCC RS6000 Constraint Letters
3784 switch (Constraint[0]) {
3786 if (Subtarget->isThumb1Only())
3787 return std::make_pair(0U, ARM::tGPRRegisterClass);
3789 return std::make_pair(0U, ARM::GPRRegisterClass);
3791 return std::make_pair(0U, ARM::GPRRegisterClass);
3794 return std::make_pair(0U, ARM::SPRRegisterClass);
3796 return std::make_pair(0U, ARM::DPRRegisterClass);
3800 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3803 std::vector<unsigned> ARMTargetLowering::
3804 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3806 if (Constraint.size() != 1)
3807 return std::vector<unsigned>();
3809 switch (Constraint[0]) { // GCC ARM Constraint Letters
3812 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3813 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3816 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3817 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3818 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3819 ARM::R12, ARM::LR, 0);
3822 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3823 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3824 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3825 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3826 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3827 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3828 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3829 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3831 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3832 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3833 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3834 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3838 return std::vector<unsigned>();
3841 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3842 /// vector. If it is invalid, don't add anything to Ops.
3843 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3846 std::vector<SDValue>&Ops,
3847 SelectionDAG &DAG) const {
3848 SDValue Result(0, 0);
3850 switch (Constraint) {
3852 case 'I': case 'J': case 'K': case 'L':
3853 case 'M': case 'N': case 'O':
3854 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3858 int64_t CVal64 = C->getSExtValue();
3859 int CVal = (int) CVal64;
3860 // None of these constraints allow values larger than 32 bits. Check
3861 // that the value fits in an int.
3865 switch (Constraint) {
3867 if (Subtarget->isThumb1Only()) {
3868 // This must be a constant between 0 and 255, for ADD
3870 if (CVal >= 0 && CVal <= 255)
3872 } else if (Subtarget->isThumb2()) {
3873 // A constant that can be used as an immediate value in a
3874 // data-processing instruction.
3875 if (ARM_AM::getT2SOImmVal(CVal) != -1)
3878 // A constant that can be used as an immediate value in a
3879 // data-processing instruction.
3880 if (ARM_AM::getSOImmVal(CVal) != -1)
3886 if (Subtarget->isThumb()) { // FIXME thumb2
3887 // This must be a constant between -255 and -1, for negated ADD
3888 // immediates. This can be used in GCC with an "n" modifier that
3889 // prints the negated value, for use with SUB instructions. It is
3890 // not useful otherwise but is implemented for compatibility.
3891 if (CVal >= -255 && CVal <= -1)
3894 // This must be a constant between -4095 and 4095. It is not clear
3895 // what this constraint is intended for. Implemented for
3896 // compatibility with GCC.
3897 if (CVal >= -4095 && CVal <= 4095)
3903 if (Subtarget->isThumb1Only()) {
3904 // A 32-bit value where only one byte has a nonzero value. Exclude
3905 // zero to match GCC. This constraint is used by GCC internally for
3906 // constants that can be loaded with a move/shift combination.
3907 // It is not useful otherwise but is implemented for compatibility.
3908 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
3910 } else if (Subtarget->isThumb2()) {
3911 // A constant whose bitwise inverse can be used as an immediate
3912 // value in a data-processing instruction. This can be used in GCC
3913 // with a "B" modifier that prints the inverted value, for use with
3914 // BIC and MVN instructions. It is not useful otherwise but is
3915 // implemented for compatibility.
3916 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
3919 // A constant whose bitwise inverse can be used as an immediate
3920 // value in a data-processing instruction. This can be used in GCC
3921 // with a "B" modifier that prints the inverted value, for use with
3922 // BIC and MVN instructions. It is not useful otherwise but is
3923 // implemented for compatibility.
3924 if (ARM_AM::getSOImmVal(~CVal) != -1)
3930 if (Subtarget->isThumb1Only()) {
3931 // This must be a constant between -7 and 7,
3932 // for 3-operand ADD/SUB immediate instructions.
3933 if (CVal >= -7 && CVal < 7)
3935 } else if (Subtarget->isThumb2()) {
3936 // A constant whose negation can be used as an immediate value in a
3937 // data-processing instruction. This can be used in GCC with an "n"
3938 // modifier that prints the negated value, for use with SUB
3939 // instructions. It is not useful otherwise but is implemented for
3941 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
3944 // A constant whose negation can be used as an immediate value in a
3945 // data-processing instruction. This can be used in GCC with an "n"
3946 // modifier that prints the negated value, for use with SUB
3947 // instructions. It is not useful otherwise but is implemented for
3949 if (ARM_AM::getSOImmVal(-CVal) != -1)
3955 if (Subtarget->isThumb()) { // FIXME thumb2
3956 // This must be a multiple of 4 between 0 and 1020, for
3957 // ADD sp + immediate.
3958 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
3961 // A power of two or a constant between 0 and 32. This is used in
3962 // GCC for the shift amount on shifted register operands, but it is
3963 // useful in general for any shift amounts.
3964 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
3970 if (Subtarget->isThumb()) { // FIXME thumb2
3971 // This must be a constant between 0 and 31, for shift amounts.
3972 if (CVal >= 0 && CVal <= 31)
3978 if (Subtarget->isThumb()) { // FIXME thumb2
3979 // This must be a multiple of 4 between -508 and 508, for
3980 // ADD/SUB sp = sp + immediate.
3981 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
3986 Result = DAG.getTargetConstant(CVal, Op.getValueType());
3990 if (Result.getNode()) {
3991 Ops.push_back(Result);
3994 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,