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 "ARMRegisterInfo.h"
21 #include "ARMSubtarget.h"
22 #include "ARMTargetMachine.h"
23 #include "llvm/CallingConv.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Instruction.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/GlobalValue.h"
28 #include "llvm/CodeGen/MachineBasicBlock.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/SelectionDAG.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/ADT/VectorExtras.h"
36 #include "llvm/Support/MathExtras.h"
39 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
40 : TargetLowering(TM), ARMPCLabelIndex(0) {
41 Subtarget = &TM.getSubtarget<ARMSubtarget>();
43 if (Subtarget->isTargetDarwin()) {
44 // Uses VFP for Thumb libfuncs if available.
45 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
46 // Single-precision floating-point arithmetic.
47 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
48 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
49 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
50 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
52 // Double-precision floating-point arithmetic.
53 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
54 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
55 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
56 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
58 // Single-precision comparisons.
59 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
60 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
61 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
62 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
63 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
64 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
65 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
66 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
68 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
69 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
70 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
71 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
72 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
73 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
74 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
75 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
77 // Double-precision comparisons.
78 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
79 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
80 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
81 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
82 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
83 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
84 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
85 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
87 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
88 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
89 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
90 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
91 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
92 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
93 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
94 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
96 // Floating-point to integer conversions.
97 // i64 conversions are done via library routines even when generating VFP
98 // instructions, so use the same ones.
99 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
100 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
101 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
102 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
104 // Conversions between floating types.
105 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
106 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
108 // Integer to floating-point conversions.
109 // i64 conversions are done via library routines even when generating VFP
110 // instructions, so use the same ones.
111 // FIXME: There appears to be some naming inconsistency in ARM libgcc: e.g.
112 // __floatunsidf vs. __floatunssidfvfp.
113 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
114 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
115 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
116 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
120 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
121 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
122 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
123 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
125 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
127 computeRegisterProperties();
129 // ARM does not have f32 extending load.
130 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
132 // ARM does not have i1 sign extending load.
133 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
135 // ARM supports all 4 flavors of integer indexed load / store.
136 for (unsigned im = (unsigned)ISD::PRE_INC;
137 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
138 setIndexedLoadAction(im, MVT::i1, Legal);
139 setIndexedLoadAction(im, MVT::i8, Legal);
140 setIndexedLoadAction(im, MVT::i16, Legal);
141 setIndexedLoadAction(im, MVT::i32, Legal);
142 setIndexedStoreAction(im, MVT::i1, Legal);
143 setIndexedStoreAction(im, MVT::i8, Legal);
144 setIndexedStoreAction(im, MVT::i16, Legal);
145 setIndexedStoreAction(im, MVT::i32, Legal);
148 // i64 operation support.
149 if (Subtarget->isThumb()) {
150 setOperationAction(ISD::MUL, MVT::i64, Expand);
151 setOperationAction(ISD::MULHU, MVT::i32, Expand);
152 setOperationAction(ISD::MULHS, MVT::i32, Expand);
153 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
154 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
156 setOperationAction(ISD::MUL, MVT::i64, Expand);
157 setOperationAction(ISD::MULHU, MVT::i32, Expand);
158 if (!Subtarget->hasV6Ops())
159 setOperationAction(ISD::MULHS, MVT::i32, Expand);
161 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
162 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
163 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
164 setOperationAction(ISD::SRL, MVT::i64, Custom);
165 setOperationAction(ISD::SRA, MVT::i64, Custom);
167 // ARM does not have ROTL.
168 setOperationAction(ISD::ROTL, MVT::i32, Expand);
169 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
170 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
171 if (!Subtarget->hasV5TOps() || Subtarget->isThumb())
172 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
174 // Only ARMv6 has BSWAP.
175 if (!Subtarget->hasV6Ops())
176 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
178 // These are expanded into libcalls.
179 setOperationAction(ISD::SDIV, MVT::i32, Expand);
180 setOperationAction(ISD::UDIV, MVT::i32, Expand);
181 setOperationAction(ISD::SREM, MVT::i32, Expand);
182 setOperationAction(ISD::UREM, MVT::i32, Expand);
183 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
184 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
186 // Support label based line numbers.
187 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
188 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
190 setOperationAction(ISD::RET, MVT::Other, Custom);
191 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
192 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
193 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
194 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
196 // Use the default implementation.
197 setOperationAction(ISD::VASTART , MVT::Other, Custom);
198 setOperationAction(ISD::VAARG , MVT::Other, Expand);
199 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
200 setOperationAction(ISD::VAEND , MVT::Other, Expand);
201 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
202 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
203 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
204 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
206 if (!Subtarget->hasV6Ops()) {
207 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
208 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
210 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
212 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb())
213 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
214 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
216 // We want to custom lower some of our intrinsics.
217 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
219 setOperationAction(ISD::SETCC , MVT::i32, Expand);
220 setOperationAction(ISD::SETCC , MVT::f32, Expand);
221 setOperationAction(ISD::SETCC , MVT::f64, Expand);
222 setOperationAction(ISD::SELECT , MVT::i32, Expand);
223 setOperationAction(ISD::SELECT , MVT::f32, Expand);
224 setOperationAction(ISD::SELECT , MVT::f64, Expand);
225 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
226 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
227 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
229 setOperationAction(ISD::BRCOND , MVT::Other, Expand);
230 setOperationAction(ISD::BR_CC , MVT::i32, Custom);
231 setOperationAction(ISD::BR_CC , MVT::f32, Custom);
232 setOperationAction(ISD::BR_CC , MVT::f64, Custom);
233 setOperationAction(ISD::BR_JT , MVT::Other, Custom);
235 // We don't support sin/cos/fmod/copysign/pow
236 setOperationAction(ISD::FSIN , MVT::f64, Expand);
237 setOperationAction(ISD::FSIN , MVT::f32, Expand);
238 setOperationAction(ISD::FCOS , MVT::f32, Expand);
239 setOperationAction(ISD::FCOS , MVT::f64, Expand);
240 setOperationAction(ISD::FREM , MVT::f64, Expand);
241 setOperationAction(ISD::FREM , MVT::f32, Expand);
242 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
243 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
244 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
246 setOperationAction(ISD::FPOW , MVT::f64, Expand);
247 setOperationAction(ISD::FPOW , MVT::f32, Expand);
249 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
250 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
251 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
252 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
253 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
254 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
257 // We have target-specific dag combine patterns for the following nodes:
258 // ARMISD::FMRRD - No need to call setTargetDAGCombine
260 setStackPointerRegisterToSaveRestore(ARM::SP);
261 setSchedulingPreference(SchedulingForRegPressure);
262 setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
263 setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
265 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
269 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
272 case ARMISD::Wrapper: return "ARMISD::Wrapper";
273 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
274 case ARMISD::CALL: return "ARMISD::CALL";
275 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
276 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
277 case ARMISD::tCALL: return "ARMISD::tCALL";
278 case ARMISD::BRCOND: return "ARMISD::BRCOND";
279 case ARMISD::BR_JT: return "ARMISD::BR_JT";
280 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
281 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
282 case ARMISD::CMP: return "ARMISD::CMP";
283 case ARMISD::CMPNZ: return "ARMISD::CMPNZ";
284 case ARMISD::CMPFP: return "ARMISD::CMPFP";
285 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
286 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
287 case ARMISD::CMOV: return "ARMISD::CMOV";
288 case ARMISD::CNEG: return "ARMISD::CNEG";
290 case ARMISD::FTOSI: return "ARMISD::FTOSI";
291 case ARMISD::FTOUI: return "ARMISD::FTOUI";
292 case ARMISD::SITOF: return "ARMISD::SITOF";
293 case ARMISD::UITOF: return "ARMISD::UITOF";
295 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
296 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
297 case ARMISD::RRX: return "ARMISD::RRX";
299 case ARMISD::FMRRD: return "ARMISD::FMRRD";
300 case ARMISD::FMDRR: return "ARMISD::FMDRR";
302 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
306 //===----------------------------------------------------------------------===//
308 //===----------------------------------------------------------------------===//
311 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
312 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
314 default: assert(0 && "Unknown condition code!");
315 case ISD::SETNE: return ARMCC::NE;
316 case ISD::SETEQ: return ARMCC::EQ;
317 case ISD::SETGT: return ARMCC::GT;
318 case ISD::SETGE: return ARMCC::GE;
319 case ISD::SETLT: return ARMCC::LT;
320 case ISD::SETLE: return ARMCC::LE;
321 case ISD::SETUGT: return ARMCC::HI;
322 case ISD::SETUGE: return ARMCC::HS;
323 case ISD::SETULT: return ARMCC::LO;
324 case ISD::SETULE: return ARMCC::LS;
328 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
329 /// returns true if the operands should be inverted to form the proper
331 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
332 ARMCC::CondCodes &CondCode2) {
334 CondCode2 = ARMCC::AL;
336 default: assert(0 && "Unknown FP condition!");
338 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
340 case ISD::SETOGT: CondCode = ARMCC::GT; break;
342 case ISD::SETOGE: CondCode = ARMCC::GE; break;
343 case ISD::SETOLT: CondCode = ARMCC::MI; break;
344 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
345 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
346 case ISD::SETO: CondCode = ARMCC::VC; break;
347 case ISD::SETUO: CondCode = ARMCC::VS; break;
348 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
349 case ISD::SETUGT: CondCode = ARMCC::HI; break;
350 case ISD::SETUGE: CondCode = ARMCC::PL; break;
352 case ISD::SETULT: CondCode = ARMCC::LT; break;
354 case ISD::SETULE: CondCode = ARMCC::LE; break;
356 case ISD::SETUNE: CondCode = ARMCC::NE; break;
362 HowToPassArgument(MVT ObjectVT, unsigned NumGPRs,
363 unsigned StackOffset, unsigned &NeededGPRs,
364 unsigned &NeededStackSize, unsigned &GPRPad,
365 unsigned &StackPad, ISD::ArgFlagsTy Flags) {
370 unsigned align = Flags.getOrigAlign();
371 GPRPad = NumGPRs % ((align + 3)/4);
372 StackPad = StackOffset % align;
373 unsigned firstGPR = NumGPRs + GPRPad;
374 switch (ObjectVT.getSimpleVT()) {
375 default: assert(0 && "Unhandled argument type!");
387 else if (firstGPR == 3) {
395 /// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
396 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
398 SDValue ARMTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
399 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
400 MVT RetVT = TheCall->getRetValType(0);
401 SDValue Chain = TheCall->getChain();
402 assert((TheCall->getCallingConv() == CallingConv::C ||
403 TheCall->getCallingConv() == CallingConv::Fast) &&
404 "unknown calling convention");
405 SDValue Callee = TheCall->getCallee();
406 unsigned NumOps = TheCall->getNumArgs();
407 DebugLoc dl = TheCall->getDebugLoc();
408 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
409 unsigned NumGPRs = 0; // GPRs used for parameter passing.
411 // Count how many bytes are to be pushed on the stack.
412 unsigned NumBytes = 0;
414 // Add up all the space actually used.
415 for (unsigned i = 0; i < NumOps; ++i) {
420 MVT ObjectVT = TheCall->getArg(i).getValueType();
421 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
422 HowToPassArgument(ObjectVT, NumGPRs, NumBytes, ObjGPRs, ObjSize,
423 GPRPad, StackPad, Flags);
424 NumBytes += ObjSize + StackPad;
425 NumGPRs += ObjGPRs + GPRPad;
428 // Adjust the stack pointer for the new arguments...
429 // These operations are automatically eliminated by the prolog/epilog pass
430 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
432 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
434 static const unsigned GPRArgRegs[] = {
435 ARM::R0, ARM::R1, ARM::R2, ARM::R3
439 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
440 std::vector<SDValue> MemOpChains;
441 for (unsigned i = 0; i != NumOps; ++i) {
442 SDValue Arg = TheCall->getArg(i);
443 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
444 MVT ArgVT = Arg.getValueType();
450 HowToPassArgument(ArgVT, NumGPRs, ArgOffset, ObjGPRs,
451 ObjSize, GPRPad, StackPad, Flags);
453 ArgOffset += StackPad;
455 switch (ArgVT.getSimpleVT()) {
456 default: assert(0 && "Unexpected ValueType for argument!");
458 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Arg));
461 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs],
462 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Arg)));
465 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Arg,
466 DAG.getConstant(0, getPointerTy()));
467 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Arg,
468 DAG.getConstant(1, getPointerTy()));
469 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Lo));
471 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1], Hi));
473 SDValue PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
474 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
475 MemOpChains.push_back(DAG.getStore(Chain, dl, Hi, PtrOff, NULL, 0));
480 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
481 DAG.getVTList(MVT::i32, MVT::i32),
483 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Cvt));
485 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1],
488 SDValue PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
489 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
490 MemOpChains.push_back(DAG.getStore(Chain, dl, Cvt.getValue(1), PtrOff,
497 assert(ObjSize != 0);
498 SDValue PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
499 PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
500 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0));
504 ArgOffset += ObjSize;
507 if (!MemOpChains.empty())
508 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
509 &MemOpChains[0], MemOpChains.size());
511 // Build a sequence of copy-to-reg nodes chained together with token chain
512 // and flag operands which copy the outgoing args into the appropriate regs.
514 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
515 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
516 RegsToPass[i].second, InFlag);
517 InFlag = Chain.getValue(1);
520 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
521 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
522 // node so that legalize doesn't hack it.
523 bool isDirect = false;
524 bool isARMFunc = false;
525 bool isLocalARMFunc = false;
526 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
527 GlobalValue *GV = G->getGlobal();
529 bool isExt = (GV->isDeclaration() || GV->hasWeakLinkage() ||
530 GV->hasLinkOnceLinkage());
531 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
532 getTargetMachine().getRelocationModel() != Reloc::Static;
533 isARMFunc = !Subtarget->isThumb() || isStub;
534 // ARM call to a local ARM function is predicable.
535 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
536 // tBX takes a register source operand.
537 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
538 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
540 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
541 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
542 Callee = DAG.getLoad(getPointerTy(), dl,
543 DAG.getEntryNode(), CPAddr, NULL, 0);
544 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
545 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
546 getPointerTy(), Callee, PICLabel);
548 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
549 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
551 bool isStub = Subtarget->isTargetDarwin() &&
552 getTargetMachine().getRelocationModel() != Reloc::Static;
553 isARMFunc = !Subtarget->isThumb() || isStub;
554 // tBX takes a register source operand.
555 const char *Sym = S->getSymbol();
556 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
557 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
559 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
560 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
561 Callee = DAG.getLoad(getPointerTy(), dl,
562 DAG.getEntryNode(), CPAddr, NULL, 0);
563 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
564 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
565 getPointerTy(), Callee, PICLabel);
567 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
570 // FIXME: handle tail calls differently.
572 if (Subtarget->isThumb()) {
573 if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc))
574 CallOpc = ARMISD::CALL_NOLINK;
576 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
578 CallOpc = (isDirect || Subtarget->hasV5TOps())
579 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
580 : ARMISD::CALL_NOLINK;
582 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb()) {
583 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
584 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
585 InFlag = Chain.getValue(1);
588 std::vector<SDValue> Ops;
589 Ops.push_back(Chain);
590 Ops.push_back(Callee);
592 // Add argument registers to the end of the list so that they are known live
594 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
595 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
596 RegsToPass[i].second.getValueType()));
598 if (InFlag.getNode())
599 Ops.push_back(InFlag);
600 // Returns a chain and a flag for retval copy to use.
601 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
602 &Ops[0], Ops.size());
603 InFlag = Chain.getValue(1);
605 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
606 DAG.getIntPtrConstant(0, true), InFlag);
607 if (RetVT != MVT::Other)
608 InFlag = Chain.getValue(1);
610 std::vector<SDValue> ResultVals;
612 // If the call has results, copy the values out of the ret val registers.
613 switch (RetVT.getSimpleVT()) {
614 default: assert(0 && "Unexpected ret value!");
618 Chain = DAG.getCopyFromReg(Chain, dl, ARM::R0,
619 MVT::i32, InFlag).getValue(1);
620 ResultVals.push_back(Chain.getValue(0));
621 if (TheCall->getNumRetVals() > 1 &&
622 TheCall->getRetValType(1) == MVT::i32) {
623 // Returns a i64 value.
624 Chain = DAG.getCopyFromReg(Chain, dl, ARM::R1, MVT::i32,
625 Chain.getValue(2)).getValue(1);
626 ResultVals.push_back(Chain.getValue(0));
630 Chain = DAG.getCopyFromReg(Chain, dl, ARM::R0,
631 MVT::i32, InFlag).getValue(1);
632 ResultVals.push_back(DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32,
636 SDValue Lo = DAG.getCopyFromReg(Chain, dl, ARM::R0, MVT::i32, InFlag);
637 SDValue Hi = DAG.getCopyFromReg(Lo, dl, ARM::R1, MVT::i32, Lo.getValue(2));
638 ResultVals.push_back(DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi));
643 if (ResultVals.empty())
646 ResultVals.push_back(Chain);
647 SDValue Res = DAG.getMergeValues(&ResultVals[0], ResultVals.size(), dl);
648 return Res.getValue(Op.getResNo());
651 static SDValue LowerRET(SDValue Op, SelectionDAG &DAG) {
653 SDValue Chain = Op.getOperand(0);
654 DebugLoc dl = Op.getDebugLoc();
655 switch(Op.getNumOperands()) {
657 assert(0 && "Do not know how to return this many arguments!");
660 SDValue LR = DAG.getRegister(ARM::LR, MVT::i32);
661 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
664 Op = Op.getOperand(1);
665 if (Op.getValueType() == MVT::f32) {
666 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
667 } else if (Op.getValueType() == MVT::f64) {
668 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
670 Op = DAG.getNode(ARMISD::FMRRD, dl,
671 DAG.getVTList(MVT::i32, MVT::i32), &Op,1);
672 SDValue Sign = DAG.getConstant(0, MVT::i32);
673 return DAG.getNode(ISD::RET, dl, MVT::Other, Chain, Op, Sign,
674 Op.getValue(1), Sign);
676 Copy = DAG.getCopyToReg(Chain, dl, ARM::R0, Op, SDValue());
677 if (DAG.getMachineFunction().getRegInfo().liveout_empty())
678 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R0);
681 Copy = DAG.getCopyToReg(Chain, dl, ARM::R1, Op.getOperand(3), SDValue());
682 Copy = DAG.getCopyToReg(Copy, dl, ARM::R0, Op.getOperand(1),
684 // If we haven't noted the R0+R1 are live out, do so now.
685 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
686 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R0);
687 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R1);
690 case 9: // i128 -> 4 regs
691 Copy = DAG.getCopyToReg(Chain, dl, ARM::R3, Op.getOperand(7), SDValue());
692 Copy = DAG.getCopyToReg(Copy , dl, ARM::R2, Op.getOperand(5),
694 Copy = DAG.getCopyToReg(Copy , dl, ARM::R1, Op.getOperand(3),
696 Copy = DAG.getCopyToReg(Copy , dl, ARM::R0, Op.getOperand(1),
698 // If we haven't noted the R0+R1 are live out, do so now.
699 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
700 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R0);
701 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R1);
702 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R2);
703 DAG.getMachineFunction().getRegInfo().addLiveOut(ARM::R3);
709 //We must use RET_FLAG instead of BRIND because BRIND doesn't have a flag
710 return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Copy, Copy.getValue(1));
713 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
714 // their target countpart wrapped in the ARMISD::Wrapper node. Suppose N is
715 // one of the above mentioned nodes. It has to be wrapped because otherwise
716 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
717 // be used to form addressing mode. These wrapped nodes will be selected
719 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
720 MVT PtrVT = Op.getValueType();
721 // FIXME there is no actual debug info here
722 DebugLoc dl = Op.getDebugLoc();
723 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
725 if (CP->isMachineConstantPoolEntry())
726 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
729 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
731 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
734 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
736 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
738 DebugLoc dl = GA->getDebugLoc();
739 MVT PtrVT = getPointerTy();
740 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
741 ARMConstantPoolValue *CPV =
742 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
743 PCAdj, "tlsgd", true);
744 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 2);
745 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
746 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
747 SDValue Chain = Argument.getValue(1);
749 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
750 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
752 // call __tls_get_addr.
755 Entry.Node = Argument;
756 Entry.Ty = (const Type *) Type::Int32Ty;
757 Args.push_back(Entry);
758 // FIXME: is there useful debug info available here?
759 std::pair<SDValue, SDValue> CallResult =
760 LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false, false, false,
761 CallingConv::C, false,
762 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
763 return CallResult.first;
766 // Lower ISD::GlobalTLSAddress using the "initial exec" or
767 // "local exec" model.
769 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
771 GlobalValue *GV = GA->getGlobal();
772 DebugLoc dl = GA->getDebugLoc();
774 SDValue Chain = DAG.getEntryNode();
775 MVT PtrVT = getPointerTy();
776 // Get the Thread Pointer
777 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
779 if (GV->isDeclaration()){
780 // initial exec model
781 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
782 ARMConstantPoolValue *CPV =
783 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue,
784 PCAdj, "gottpoff", true);
785 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 2);
786 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
787 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
788 Chain = Offset.getValue(1);
790 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
791 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
793 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
796 ARMConstantPoolValue *CPV =
797 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff");
798 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 2);
799 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
800 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
803 // The address of the thread local variable is the add of the thread
804 // pointer with the offset of the variable.
805 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
809 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
810 // TODO: implement the "local dynamic" model
811 assert(Subtarget->isTargetELF() &&
812 "TLS not implemented for non-ELF targets");
813 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
814 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
815 // otherwise use the "Local Exec" TLS Model
816 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
817 return LowerToTLSGeneralDynamicModel(GA, DAG);
819 return LowerToTLSExecModels(GA, DAG);
822 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
824 MVT PtrVT = getPointerTy();
825 DebugLoc dl = Op.getDebugLoc();
826 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
827 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
828 if (RelocM == Reloc::PIC_) {
829 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
830 ARMConstantPoolValue *CPV =
831 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT");
832 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
833 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
834 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
836 SDValue Chain = Result.getValue(1);
837 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
838 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
840 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
843 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
844 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
845 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
849 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
850 /// even in non-static mode.
851 static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) {
852 // If symbol visibility is hidden, the extra load is not needed if
853 // the symbol is definitely defined in the current translation unit.
854 bool isDecl = GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode();
855 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage()))
857 return RelocM != Reloc::Static && (isDecl || GV->mayBeOverridden());
860 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
862 MVT PtrVT = getPointerTy();
863 DebugLoc dl = Op.getDebugLoc();
864 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
865 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
866 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM);
868 if (RelocM == Reloc::Static)
869 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
871 unsigned PCAdj = (RelocM != Reloc::PIC_)
872 ? 0 : (Subtarget->isThumb() ? 4 : 8);
873 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
875 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
877 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
879 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
881 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
882 SDValue Chain = Result.getValue(1);
884 if (RelocM == Reloc::PIC_) {
885 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
886 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
889 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
894 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
896 assert(Subtarget->isTargetELF() &&
897 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
898 MVT PtrVT = getPointerTy();
899 DebugLoc dl = Op.getDebugLoc();
900 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
901 ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
903 ARMCP::CPValue, PCAdj);
904 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
905 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
906 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
907 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
908 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
911 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
912 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
913 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
915 default: return SDValue(); // Don't custom lower most intrinsics.
916 case Intrinsic::arm_thread_pointer:
917 return DAG.getNode(ARMISD::THREAD_POINTER, DebugLoc::getUnknownLoc(),
922 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
923 unsigned VarArgsFrameIndex) {
924 // vastart just stores the address of the VarArgsFrameIndex slot into the
925 // memory location argument.
926 DebugLoc dl = Op.getDebugLoc();
927 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
928 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
929 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
930 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
933 static SDValue LowerFORMAL_ARGUMENT(SDValue Op, SelectionDAG &DAG,
934 unsigned ArgNo, unsigned &NumGPRs,
935 unsigned &ArgOffset, DebugLoc dl) {
936 MachineFunction &MF = DAG.getMachineFunction();
937 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
938 SDValue Root = Op.getOperand(0);
939 MachineRegisterInfo &RegInfo = MF.getRegInfo();
941 static const unsigned GPRArgRegs[] = {
942 ARM::R0, ARM::R1, ARM::R2, ARM::R3
949 ISD::ArgFlagsTy Flags =
950 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo + 3))->getArgFlags();
951 HowToPassArgument(ObjectVT, NumGPRs, ArgOffset, ObjGPRs,
952 ObjSize, GPRPad, StackPad, Flags);
954 ArgOffset += StackPad;
958 unsigned VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
959 RegInfo.addLiveIn(GPRArgRegs[NumGPRs], VReg);
960 ArgValue = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
961 if (ObjectVT == MVT::f32)
962 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, ArgValue);
963 } else if (ObjGPRs == 2) {
964 unsigned VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
965 RegInfo.addLiveIn(GPRArgRegs[NumGPRs], VReg);
966 ArgValue = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
968 VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
969 RegInfo.addLiveIn(GPRArgRegs[NumGPRs+1], VReg);
970 SDValue ArgValue2 = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
972 assert(ObjectVT != MVT::i64 && "i64 should already be lowered");
973 ArgValue = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
978 MachineFrameInfo *MFI = MF.getFrameInfo();
979 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
980 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
982 ArgValue = DAG.getLoad(ObjectVT, dl, Root, FIN, NULL, 0);
984 SDValue ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
985 assert(ObjectVT != MVT::i64 && "i64 should already be lowered");
986 ArgValue = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
989 ArgOffset += ObjSize; // Move on to the next argument.
996 ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
997 std::vector<SDValue> ArgValues;
998 SDValue Root = Op.getOperand(0);
999 DebugLoc dl = Op.getDebugLoc();
1000 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
1001 unsigned NumGPRs = 0; // GPRs used for parameter passing.
1003 unsigned NumArgs = Op.getNode()->getNumValues()-1;
1004 for (unsigned ArgNo = 0; ArgNo < NumArgs; ++ArgNo)
1005 ArgValues.push_back(LowerFORMAL_ARGUMENT(Op, DAG, ArgNo,
1006 NumGPRs, ArgOffset, dl));
1008 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1010 static const unsigned GPRArgRegs[] = {
1011 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1014 MachineFunction &MF = DAG.getMachineFunction();
1015 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1016 MachineFrameInfo *MFI = MF.getFrameInfo();
1017 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1018 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1019 unsigned VARegSize = (4 - NumGPRs) * 4;
1020 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1021 if (VARegSaveSize) {
1022 // If this function is vararg, store any remaining integer argument regs
1023 // to their spots on the stack so that they may be loaded by deferencing
1024 // the result of va_next.
1025 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1026 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1027 VARegSaveSize - VARegSize);
1028 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1030 SmallVector<SDValue, 4> MemOps;
1031 for (; NumGPRs < 4; ++NumGPRs) {
1032 unsigned VReg = RegInfo.createVirtualRegister(&ARM::GPRRegClass);
1033 RegInfo.addLiveIn(GPRArgRegs[NumGPRs], VReg);
1034 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
1035 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1036 MemOps.push_back(Store);
1037 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1038 DAG.getConstant(4, getPointerTy()));
1040 if (!MemOps.empty())
1041 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1042 &MemOps[0], MemOps.size());
1044 // This will point to the next argument passed via stack.
1045 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1048 ArgValues.push_back(Root);
1050 // Return the new list of results.
1051 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1052 &ArgValues[0], ArgValues.size());
1055 /// isFloatingPointZero - Return true if this is +0.0.
1056 static bool isFloatingPointZero(SDValue Op) {
1057 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1058 return CFP->getValueAPF().isPosZero();
1059 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1060 // Maybe this has already been legalized into the constant pool?
1061 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1062 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1063 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1064 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1065 return CFP->getValueAPF().isPosZero();
1071 static bool isLegalCmpImmediate(unsigned C, bool isThumb) {
1072 return ( isThumb && (C & ~255U) == 0) ||
1073 (!isThumb && ARM_AM::getSOImmVal(C) != -1);
1076 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1077 /// the given operands.
1078 static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1079 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb,
1081 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1082 unsigned C = RHSC->getZExtValue();
1083 if (!isLegalCmpImmediate(C, isThumb)) {
1084 // Constant does not fit, try adjusting it by one?
1089 if (isLegalCmpImmediate(C-1, isThumb)) {
1090 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1091 RHS = DAG.getConstant(C-1, MVT::i32);
1096 if (C > 0 && isLegalCmpImmediate(C-1, isThumb)) {
1097 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1098 RHS = DAG.getConstant(C-1, MVT::i32);
1103 if (isLegalCmpImmediate(C+1, isThumb)) {
1104 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1105 RHS = DAG.getConstant(C+1, MVT::i32);
1110 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb)) {
1111 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1112 RHS = DAG.getConstant(C+1, MVT::i32);
1119 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1120 ARMISD::NodeType CompareType;
1123 CompareType = ARMISD::CMP;
1129 // Uses only N and Z Flags
1130 CompareType = ARMISD::CMPNZ;
1133 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1134 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1137 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1138 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1141 if (!isFloatingPointZero(RHS))
1142 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1144 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1145 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1148 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
1149 const ARMSubtarget *ST) {
1150 MVT VT = Op.getValueType();
1151 SDValue LHS = Op.getOperand(0);
1152 SDValue RHS = Op.getOperand(1);
1153 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1154 SDValue TrueVal = Op.getOperand(2);
1155 SDValue FalseVal = Op.getOperand(3);
1156 DebugLoc dl = Op.getDebugLoc();
1158 if (LHS.getValueType() == MVT::i32) {
1160 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1161 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb(), dl);
1162 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1165 ARMCC::CondCodes CondCode, CondCode2;
1166 if (FPCCToARMCC(CC, CondCode, CondCode2))
1167 std::swap(TrueVal, FalseVal);
1169 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1170 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1171 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1172 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1174 if (CondCode2 != ARMCC::AL) {
1175 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1176 // FIXME: Needs another CMP because flag can have but one use.
1177 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1178 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1179 Result, TrueVal, ARMCC2, CCR, Cmp2);
1184 static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
1185 const ARMSubtarget *ST) {
1186 SDValue Chain = Op.getOperand(0);
1187 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1188 SDValue LHS = Op.getOperand(2);
1189 SDValue RHS = Op.getOperand(3);
1190 SDValue Dest = Op.getOperand(4);
1191 DebugLoc dl = Op.getDebugLoc();
1193 if (LHS.getValueType() == MVT::i32) {
1195 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1196 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb(), dl);
1197 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1198 Chain, Dest, ARMCC, CCR,Cmp);
1201 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1202 ARMCC::CondCodes CondCode, CondCode2;
1203 if (FPCCToARMCC(CC, CondCode, CondCode2))
1204 // Swap the LHS/RHS of the comparison if needed.
1205 std::swap(LHS, RHS);
1207 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1208 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1209 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1210 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1211 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1212 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1213 if (CondCode2 != ARMCC::AL) {
1214 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1215 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1216 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1221 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1222 SDValue Chain = Op.getOperand(0);
1223 SDValue Table = Op.getOperand(1);
1224 SDValue Index = Op.getOperand(2);
1225 DebugLoc dl = Op.getDebugLoc();
1227 MVT PTy = getPointerTy();
1228 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1229 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1230 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1231 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1232 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1233 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1234 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1235 bool isPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1236 Addr = DAG.getLoad(isPIC ? (MVT)MVT::i32 : PTy, dl,
1237 Chain, Addr, NULL, 0);
1238 Chain = Addr.getValue(1);
1240 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1241 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1244 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1245 DebugLoc dl = Op.getDebugLoc();
1247 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1248 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1249 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1252 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1253 MVT VT = Op.getValueType();
1254 DebugLoc dl = Op.getDebugLoc();
1256 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1258 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1259 return DAG.getNode(Opc, dl, VT, Op);
1262 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1263 // Implement fcopysign with a fabs and a conditional fneg.
1264 SDValue Tmp0 = Op.getOperand(0);
1265 SDValue Tmp1 = Op.getOperand(1);
1266 DebugLoc dl = Op.getDebugLoc();
1267 MVT VT = Op.getValueType();
1268 MVT SrcVT = Tmp1.getValueType();
1269 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1270 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1271 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1272 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1273 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1277 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1279 SDValue Dst, SDValue Src,
1280 SDValue Size, unsigned Align,
1282 const Value *DstSV, uint64_t DstSVOff,
1283 const Value *SrcSV, uint64_t SrcSVOff){
1284 // Do repeated 4-byte loads and stores. To be improved.
1285 // This requires 4-byte alignment.
1286 if ((Align & 3) != 0)
1288 // This requires the copy size to be a constant, preferrably
1289 // within a subtarget-specific limit.
1290 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1293 uint64_t SizeVal = ConstantSize->getZExtValue();
1294 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1297 unsigned BytesLeft = SizeVal & 3;
1298 unsigned NumMemOps = SizeVal >> 2;
1299 unsigned EmittedNumMemOps = 0;
1301 unsigned VTSize = 4;
1303 const unsigned MAX_LOADS_IN_LDM = 6;
1304 SDValue TFOps[MAX_LOADS_IN_LDM];
1305 SDValue Loads[MAX_LOADS_IN_LDM];
1306 uint64_t SrcOff = 0, DstOff = 0;
1308 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1309 // same number of stores. The loads and stores will get combined into
1310 // ldm/stm later on.
1311 while (EmittedNumMemOps < NumMemOps) {
1313 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1314 Loads[i] = DAG.getLoad(VT, dl, Chain,
1315 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1316 DAG.getConstant(SrcOff, MVT::i32)),
1317 SrcSV, SrcSVOff + SrcOff);
1318 TFOps[i] = Loads[i].getValue(1);
1321 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1324 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
1325 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1326 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1327 DAG.getConstant(DstOff, MVT::i32)),
1328 DstSV, DstSVOff + DstOff);
1331 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1333 EmittedNumMemOps += i;
1339 // Issue loads / stores for the trailing (1 - 3) bytes.
1340 unsigned BytesLeftSave = BytesLeft;
1343 if (BytesLeft >= 2) {
1351 Loads[i] = DAG.getLoad(VT, dl, Chain,
1352 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
1353 DAG.getConstant(SrcOff, MVT::i32)),
1354 SrcSV, SrcSVOff + SrcOff);
1355 TFOps[i] = Loads[i].getValue(1);
1358 BytesLeft -= VTSize;
1360 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1363 BytesLeft = BytesLeftSave;
1365 if (BytesLeft >= 2) {
1373 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
1374 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
1375 DAG.getConstant(DstOff, MVT::i32)),
1376 DstSV, DstSVOff + DstOff);
1379 BytesLeft -= VTSize;
1381 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
1384 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
1385 SDValue Op = N->getOperand(0);
1386 DebugLoc dl = N->getDebugLoc();
1387 if (N->getValueType(0) == MVT::f64) {
1388 // Turn i64->f64 into FMDRR.
1389 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1390 DAG.getConstant(0, MVT::i32));
1391 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
1392 DAG.getConstant(1, MVT::i32));
1393 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
1396 // Turn f64->i64 into FMRRD.
1397 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
1398 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
1400 // Merge the pieces into a single i64 value.
1401 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
1404 static SDValue ExpandSRx(SDNode *N, SelectionDAG &DAG, const ARMSubtarget *ST) {
1405 assert(N->getValueType(0) == MVT::i64 &&
1406 (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
1407 "Unknown shift to lower!");
1409 // We only lower SRA, SRL of 1 here, all others use generic lowering.
1410 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
1411 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
1414 // If we are in thumb mode, we don't have RRX.
1415 if (ST->isThumb()) return SDValue();
1417 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
1418 DebugLoc dl = N->getDebugLoc();
1419 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
1420 DAG.getConstant(0, MVT::i32));
1421 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
1422 DAG.getConstant(1, MVT::i32));
1424 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
1425 // captures the result into a carry flag.
1426 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
1427 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
1429 // The low part is an ARMISD::RRX operand, which shifts the carry in.
1430 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
1432 // Merge the pieces into a single i64 value.
1433 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
1437 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
1438 switch (Op.getOpcode()) {
1439 default: assert(0 && "Don't know how to custom lower this!"); abort();
1440 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1441 case ISD::GlobalAddress:
1442 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
1443 LowerGlobalAddressELF(Op, DAG);
1444 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
1445 case ISD::CALL: return LowerCALL(Op, DAG);
1446 case ISD::RET: return LowerRET(Op, DAG);
1447 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
1448 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
1449 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
1450 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1451 case ISD::SINT_TO_FP:
1452 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
1453 case ISD::FP_TO_SINT:
1454 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
1455 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
1456 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
1457 case ISD::RETURNADDR: break;
1458 case ISD::FRAMEADDR: break;
1459 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
1460 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
1461 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
1463 case ISD::SRA: return ExpandSRx(Op.getNode(), DAG,Subtarget);
1469 /// ReplaceNodeResults - Replace the results of node with an illegal result
1470 /// type with new values built out of custom code.
1472 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
1473 SmallVectorImpl<SDValue>&Results,
1474 SelectionDAG &DAG) {
1475 switch (N->getOpcode()) {
1477 assert(0 && "Don't know how to custom expand this!");
1479 case ISD::BIT_CONVERT:
1480 Results.push_back(ExpandBIT_CONVERT(N, DAG));
1484 SDValue Res = ExpandSRx(N, DAG, Subtarget);
1486 Results.push_back(Res);
1493 //===----------------------------------------------------------------------===//
1494 // ARM Scheduler Hooks
1495 //===----------------------------------------------------------------------===//
1498 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
1499 MachineBasicBlock *BB) {
1500 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1501 switch (MI->getOpcode()) {
1502 default: assert(false && "Unexpected instr type to insert");
1503 case ARM::tMOVCCr: {
1504 // To "insert" a SELECT_CC instruction, we actually have to insert the
1505 // diamond control-flow pattern. The incoming instruction knows the
1506 // destination vreg to set, the condition code register to branch on, the
1507 // true/false values to select between, and a branch opcode to use.
1508 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1509 MachineFunction::iterator It = BB;
1515 // cmpTY ccX, r1, r2
1517 // fallthrough --> copy0MBB
1518 MachineBasicBlock *thisMBB = BB;
1519 MachineFunction *F = BB->getParent();
1520 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
1521 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
1522 BuildMI(BB, TII->get(ARM::tBcc)).addMBB(sinkMBB)
1523 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
1524 F->insert(It, copy0MBB);
1525 F->insert(It, sinkMBB);
1526 // Update machine-CFG edges by first adding all successors of the current
1527 // block to the new block which will contain the Phi node for the select.
1528 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1529 e = BB->succ_end(); i != e; ++i)
1530 sinkMBB->addSuccessor(*i);
1531 // Next, remove all successors of the current block, and add the true
1532 // and fallthrough blocks as its successors.
1533 while(!BB->succ_empty())
1534 BB->removeSuccessor(BB->succ_begin());
1535 BB->addSuccessor(copy0MBB);
1536 BB->addSuccessor(sinkMBB);
1539 // %FalseValue = ...
1540 // # fallthrough to sinkMBB
1543 // Update machine-CFG edges
1544 BB->addSuccessor(sinkMBB);
1547 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1550 BuildMI(BB, TII->get(ARM::PHI), MI->getOperand(0).getReg())
1551 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1552 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1554 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
1560 //===----------------------------------------------------------------------===//
1561 // ARM Optimization Hooks
1562 //===----------------------------------------------------------------------===//
1564 /// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
1565 static SDValue PerformFMRRDCombine(SDNode *N,
1566 TargetLowering::DAGCombinerInfo &DCI) {
1567 // fmrrd(fmdrr x, y) -> x,y
1568 SDValue InDouble = N->getOperand(0);
1569 if (InDouble.getOpcode() == ARMISD::FMDRR)
1570 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
1574 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
1575 DAGCombinerInfo &DCI) const {
1576 switch (N->getOpcode()) {
1578 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
1585 /// isLegalAddressImmediate - Return true if the integer value can be used
1586 /// as the offset of the target addressing mode for load / store of the
1588 static bool isLegalAddressImmediate(int64_t V, MVT VT,
1589 const ARMSubtarget *Subtarget) {
1593 if (Subtarget->isThumb()) {
1598 switch (VT.getSimpleVT()) {
1599 default: return false;
1614 if ((V & (Scale - 1)) != 0)
1617 return V == (V & ((1LL << 5) - 1));
1622 switch (VT.getSimpleVT()) {
1623 default: return false;
1628 return V == (V & ((1LL << 12) - 1));
1631 return V == (V & ((1LL << 8) - 1));
1634 if (!Subtarget->hasVFP2())
1639 return V == (V & ((1LL << 8) - 1));
1643 /// isLegalAddressingMode - Return true if the addressing mode represented
1644 /// by AM is legal for this target, for a load/store of the specified type.
1645 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
1646 const Type *Ty) const {
1647 if (!isLegalAddressImmediate(AM.BaseOffs, getValueType(Ty, true), Subtarget))
1650 // Can never fold addr of global into load/store.
1655 case 0: // no scale reg, must be "r+i" or "r", or "i".
1658 if (Subtarget->isThumb())
1662 // ARM doesn't support any R+R*scale+imm addr modes.
1666 int Scale = AM.Scale;
1667 switch (getValueType(Ty).getSimpleVT()) {
1668 default: return false;
1673 // This assumes i64 is legalized to a pair of i32. If not (i.e.
1674 // ldrd / strd are used, then its address mode is same as i16.
1676 if (Scale < 0) Scale = -Scale;
1680 return isPowerOf2_32(Scale & ~1);
1683 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
1688 // Note, we allow "void" uses (basically, uses that aren't loads or
1689 // stores), because arm allows folding a scale into many arithmetic
1690 // operations. This should be made more precise and revisited later.
1692 // Allow r << imm, but the imm has to be a multiple of two.
1693 if (AM.Scale & 1) return false;
1694 return isPowerOf2_32(AM.Scale);
1702 static bool getIndexedAddressParts(SDNode *Ptr, MVT VT,
1703 bool isSEXTLoad, SDValue &Base,
1704 SDValue &Offset, bool &isInc,
1705 SelectionDAG &DAG) {
1706 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
1709 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
1711 Base = Ptr->getOperand(0);
1712 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1713 int RHSC = (int)RHS->getZExtValue();
1714 if (RHSC < 0 && RHSC > -256) {
1716 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1720 isInc = (Ptr->getOpcode() == ISD::ADD);
1721 Offset = Ptr->getOperand(1);
1723 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
1725 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1726 int RHSC = (int)RHS->getZExtValue();
1727 if (RHSC < 0 && RHSC > -0x1000) {
1729 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1730 Base = Ptr->getOperand(0);
1735 if (Ptr->getOpcode() == ISD::ADD) {
1737 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
1738 if (ShOpcVal != ARM_AM::no_shift) {
1739 Base = Ptr->getOperand(1);
1740 Offset = Ptr->getOperand(0);
1742 Base = Ptr->getOperand(0);
1743 Offset = Ptr->getOperand(1);
1748 isInc = (Ptr->getOpcode() == ISD::ADD);
1749 Base = Ptr->getOperand(0);
1750 Offset = Ptr->getOperand(1);
1754 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
1758 /// getPreIndexedAddressParts - returns true by value, base pointer and
1759 /// offset pointer and addressing mode by reference if the node's address
1760 /// can be legally represented as pre-indexed load / store address.
1762 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1764 ISD::MemIndexedMode &AM,
1765 SelectionDAG &DAG) const {
1766 if (Subtarget->isThumb())
1771 bool isSEXTLoad = false;
1772 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1773 Ptr = LD->getBasePtr();
1774 VT = LD->getMemoryVT();
1775 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1776 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1777 Ptr = ST->getBasePtr();
1778 VT = ST->getMemoryVT();
1783 bool isLegal = getIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, Offset,
1786 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
1792 /// getPostIndexedAddressParts - returns true by value, base pointer and
1793 /// offset pointer and addressing mode by reference if this node can be
1794 /// combined with a load / store to form a post-indexed load / store.
1795 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
1798 ISD::MemIndexedMode &AM,
1799 SelectionDAG &DAG) const {
1800 if (Subtarget->isThumb())
1805 bool isSEXTLoad = false;
1806 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1807 VT = LD->getMemoryVT();
1808 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1809 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1810 VT = ST->getMemoryVT();
1815 bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
1818 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
1824 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
1828 const SelectionDAG &DAG,
1829 unsigned Depth) const {
1830 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
1831 switch (Op.getOpcode()) {
1833 case ARMISD::CMOV: {
1834 // Bits are known zero/one if known on the LHS and RHS.
1835 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
1836 if (KnownZero == 0 && KnownOne == 0) return;
1838 APInt KnownZeroRHS, KnownOneRHS;
1839 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
1840 KnownZeroRHS, KnownOneRHS, Depth+1);
1841 KnownZero &= KnownZeroRHS;
1842 KnownOne &= KnownOneRHS;
1848 //===----------------------------------------------------------------------===//
1849 // ARM Inline Assembly Support
1850 //===----------------------------------------------------------------------===//
1852 /// getConstraintType - Given a constraint letter, return the type of
1853 /// constraint it is for this target.
1854 ARMTargetLowering::ConstraintType
1855 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
1856 if (Constraint.size() == 1) {
1857 switch (Constraint[0]) {
1859 case 'l': return C_RegisterClass;
1860 case 'w': return C_RegisterClass;
1863 return TargetLowering::getConstraintType(Constraint);
1866 std::pair<unsigned, const TargetRegisterClass*>
1867 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
1869 if (Constraint.size() == 1) {
1870 // GCC RS6000 Constraint Letters
1871 switch (Constraint[0]) {
1873 // FIXME: in thumb mode, 'l' is only low-regs.
1876 return std::make_pair(0U, ARM::GPRRegisterClass);
1879 return std::make_pair(0U, ARM::SPRRegisterClass);
1881 return std::make_pair(0U, ARM::DPRRegisterClass);
1885 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
1888 std::vector<unsigned> ARMTargetLowering::
1889 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1891 if (Constraint.size() != 1)
1892 return std::vector<unsigned>();
1894 switch (Constraint[0]) { // GCC ARM Constraint Letters
1898 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
1899 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
1900 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
1901 ARM::R12, ARM::LR, 0);
1904 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
1905 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
1906 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
1907 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
1908 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
1909 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
1910 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
1911 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
1913 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
1914 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
1915 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
1916 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
1920 return std::vector<unsigned>();