1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Evan Cheng and is distributed under
6 // the University of Illinois Open Source 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/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/SelectionDAG.h"
30 #include "llvm/CodeGen/SSARegMap.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/ADT/VectorExtras.h"
35 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
36 : TargetLowering(TM), ARMPCLabelIndex(0) {
37 Subtarget = &TM.getSubtarget<ARMSubtarget>();
39 // Uses VFP for Thumb libfuncs if available.
40 if (!UseSoftFloat && Subtarget->isThumb() && Subtarget->hasVFP2()) {
41 // Single-precision floating-point arithmetic.
42 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
43 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
44 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
45 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
47 // Double-precision floating-point arithmetic.
48 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
49 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
50 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
51 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
53 // Single-precision comparisons.
54 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
55 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
56 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
57 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
58 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
59 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
60 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
61 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
63 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
64 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
65 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
66 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
67 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
68 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
69 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
70 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
72 // Double-precision comparisons.
73 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
74 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
75 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
76 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
77 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
78 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
79 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
80 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
82 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
83 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
84 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
85 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
86 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
87 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
88 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
89 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
91 // Floating-point to integer conversions.
92 // i64 conversions are done via library routines even when generating VFP
93 // instructions, so use the same ones.
94 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
95 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
96 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
97 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
99 // Conversions between floating types.
100 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
101 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
103 // Integer to floating-point conversions.
104 // i64 conversions are done via library routines even when generating VFP
105 // instructions, so use the same ones.
106 // FIXME: There appears to be some naming inconsistency in ARM libgcc: e.g.
107 // __floatunsidf vs. __floatunssidfvfp.
108 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
109 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
110 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
111 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
114 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
115 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) {
116 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
117 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
120 // ARM does not have f32 extending load.
121 setLoadXAction(ISD::EXTLOAD, MVT::f32, Expand);
123 // ARM supports all 4 flavors of integer indexed load / store.
124 for (unsigned im = (unsigned)ISD::PRE_INC;
125 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
126 setIndexedLoadAction(im, MVT::i1, Legal);
127 setIndexedLoadAction(im, MVT::i8, Legal);
128 setIndexedLoadAction(im, MVT::i16, Legal);
129 setIndexedLoadAction(im, MVT::i32, Legal);
130 setIndexedStoreAction(im, MVT::i1, Legal);
131 setIndexedStoreAction(im, MVT::i8, Legal);
132 setIndexedStoreAction(im, MVT::i16, Legal);
133 setIndexedStoreAction(im, MVT::i32, Legal);
136 // i64 operation support.
137 if (Subtarget->isThumb()) {
138 setOperationAction(ISD::MUL, MVT::i64, Expand);
139 setOperationAction(ISD::MULHU, MVT::i32, Expand);
140 setOperationAction(ISD::MULHS, MVT::i32, Expand);
142 setOperationAction(ISD::MUL, MVT::i64, Custom);
143 setOperationAction(ISD::MULHU, MVT::i32, Custom);
144 if (!Subtarget->hasV6Ops())
145 setOperationAction(ISD::MULHS, MVT::i32, Custom);
147 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
148 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
149 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
150 setOperationAction(ISD::SRL, MVT::i64, Custom);
151 setOperationAction(ISD::SRA, MVT::i64, Custom);
153 // ARM does not have ROTL.
154 setOperationAction(ISD::ROTL, MVT::i32, Expand);
155 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
156 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
157 if (!Subtarget->hasV5TOps() || Subtarget->isThumb())
158 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
160 // These are expanded into libcalls.
161 setOperationAction(ISD::SDIV, MVT::i32, Expand);
162 setOperationAction(ISD::UDIV, MVT::i32, Expand);
163 setOperationAction(ISD::SREM, MVT::i32, Expand);
164 setOperationAction(ISD::UREM, MVT::i32, Expand);
166 // Support label based line numbers.
167 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
168 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
169 // FIXME - use subtarget debug flags
170 if (Subtarget->isTargetDarwin())
171 setOperationAction(ISD::LABEL, MVT::Other, Expand);
173 setOperationAction(ISD::RET, MVT::Other, Custom);
174 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
175 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
177 // Expand mem operations genericly.
178 setOperationAction(ISD::MEMSET , MVT::Other, Expand);
179 setOperationAction(ISD::MEMCPY , MVT::Other, Expand);
180 setOperationAction(ISD::MEMMOVE , MVT::Other, Expand);
182 // Use the default implementation.
183 setOperationAction(ISD::VASTART , MVT::Other, Expand);
184 setOperationAction(ISD::VAARG , MVT::Other, Expand);
185 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
186 setOperationAction(ISD::VAEND , MVT::Other, Expand);
187 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
188 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
189 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
191 if (!Subtarget->hasV6Ops()) {
192 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
193 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
195 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
197 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb())
198 // Turn f64->i64 into FMRRD iff target supports vfp2.
199 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
201 setOperationAction(ISD::SETCC , MVT::i32, Expand);
202 setOperationAction(ISD::SETCC , MVT::f32, Expand);
203 setOperationAction(ISD::SETCC , MVT::f64, Expand);
204 setOperationAction(ISD::SELECT , MVT::i32, Expand);
205 setOperationAction(ISD::SELECT , MVT::f32, Expand);
206 setOperationAction(ISD::SELECT , MVT::f64, Expand);
207 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
208 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
209 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
211 setOperationAction(ISD::BRCOND , MVT::Other, Expand);
212 setOperationAction(ISD::BR_CC , MVT::i32, Custom);
213 setOperationAction(ISD::BR_CC , MVT::f32, Custom);
214 setOperationAction(ISD::BR_CC , MVT::f64, Custom);
215 setOperationAction(ISD::BR_JT , MVT::Other, Custom);
217 setOperationAction(ISD::VASTART, MVT::Other, Custom);
218 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
219 setOperationAction(ISD::VAEND, MVT::Other, Expand);
220 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
221 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
223 // FP Constants can't be immediates.
224 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
225 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
227 // We don't support sin/cos/fmod/copysign
228 setOperationAction(ISD::FSIN , MVT::f64, Expand);
229 setOperationAction(ISD::FSIN , MVT::f32, Expand);
230 setOperationAction(ISD::FCOS , MVT::f32, Expand);
231 setOperationAction(ISD::FCOS , MVT::f64, Expand);
232 setOperationAction(ISD::FREM , MVT::f64, Expand);
233 setOperationAction(ISD::FREM , MVT::f32, Expand);
234 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
235 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
237 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
238 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
239 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
240 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
241 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
243 setStackPointerRegisterToSaveRestore(ARM::SP);
245 setSchedulingPreference(SchedulingForRegPressure);
246 computeRegisterProperties();
250 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
253 case ARMISD::Wrapper: return "ARMISD::Wrapper";
254 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
255 case ARMISD::CALL: return "ARMISD::CALL";
256 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
257 case ARMISD::tCALL: return "ARMISD::tCALL";
258 case ARMISD::BRCOND: return "ARMISD::BRCOND";
259 case ARMISD::BR_JT: return "ARMISD::BR_JT";
260 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
261 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
262 case ARMISD::CMP: return "ARMISD::CMP";
263 case ARMISD::CMPFP: return "ARMISD::CMPFP";
264 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
265 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
266 case ARMISD::CMOV: return "ARMISD::CMOV";
267 case ARMISD::CNEG: return "ARMISD::CNEG";
269 case ARMISD::FTOSI: return "ARMISD::FTOSI";
270 case ARMISD::FTOUI: return "ARMISD::FTOUI";
271 case ARMISD::SITOF: return "ARMISD::SITOF";
272 case ARMISD::UITOF: return "ARMISD::UITOF";
273 case ARMISD::MULHILOU: return "ARMISD::MULHILOU";
274 case ARMISD::MULHILOS: return "ARMISD::MULHILOS";
276 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
277 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
278 case ARMISD::RRX: return "ARMISD::RRX";
280 case ARMISD::FMRRD: return "ARMISD::FMRRD";
281 case ARMISD::FMDRR: return "ARMISD::FMDRR";
285 //===----------------------------------------------------------------------===//
287 //===----------------------------------------------------------------------===//
290 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
291 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
293 default: assert(0 && "Unknown condition code!");
294 case ISD::SETNE: return ARMCC::NE;
295 case ISD::SETEQ: return ARMCC::EQ;
296 case ISD::SETGT: return ARMCC::GT;
297 case ISD::SETGE: return ARMCC::GE;
298 case ISD::SETLT: return ARMCC::LT;
299 case ISD::SETLE: return ARMCC::LE;
300 case ISD::SETUGT: return ARMCC::HI;
301 case ISD::SETUGE: return ARMCC::HS;
302 case ISD::SETULT: return ARMCC::LO;
303 case ISD::SETULE: return ARMCC::LS;
307 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
308 /// returns true if the operands should be inverted to form the proper
310 static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
311 ARMCC::CondCodes &CondCode2) {
313 CondCode2 = ARMCC::AL;
315 default: assert(0 && "Unknown FP condition!");
317 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
319 case ISD::SETOGT: CondCode = ARMCC::GT; break;
321 case ISD::SETOGE: CondCode = ARMCC::GE; break;
322 case ISD::SETOLT: CondCode = ARMCC::MI; break;
323 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
324 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
325 case ISD::SETO: CondCode = ARMCC::VC; break;
326 case ISD::SETUO: CondCode = ARMCC::VS; break;
327 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
328 case ISD::SETUGT: CondCode = ARMCC::HI; break;
329 case ISD::SETUGE: CondCode = ARMCC::PL; break;
331 case ISD::SETULT: CondCode = ARMCC::LT; break;
333 case ISD::SETULE: CondCode = ARMCC::LE; break;
335 case ISD::SETUNE: CondCode = ARMCC::NE; break;
341 HowToPassArgument(MVT::ValueType ObjectVT, unsigned NumGPRs,
342 unsigned StackOffset, unsigned &NeededGPRs,
343 unsigned &NeededStackSize, unsigned &GPRPad,
344 unsigned &StackPad, unsigned Flags) {
349 unsigned align = (Flags >> 27);
350 GPRPad = NumGPRs % ((align + 3)/4);
351 StackPad = StackOffset % align;
352 unsigned firstGPR = NumGPRs + GPRPad;
354 default: assert(0 && "Unhandled argument type!");
366 else if (firstGPR == 3) {
374 /// LowerCALL - Lowering a ISD::CALL node into a callseq_start <-
375 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
377 SDOperand ARMTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG) {
378 MVT::ValueType RetVT= Op.Val->getValueType(0);
379 SDOperand Chain = Op.getOperand(0);
380 unsigned CallConv = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
381 assert((CallConv == CallingConv::C ||
382 CallConv == CallingConv::Fast) && "unknown calling convention");
383 SDOperand Callee = Op.getOperand(4);
384 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
385 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
386 unsigned NumGPRs = 0; // GPRs used for parameter passing.
388 // Count how many bytes are to be pushed on the stack.
389 unsigned NumBytes = 0;
391 // Add up all the space actually used.
392 for (unsigned i = 0; i < NumOps; ++i) {
397 MVT::ValueType ObjectVT = Op.getOperand(5+2*i).getValueType();
398 unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
399 HowToPassArgument(ObjectVT, NumGPRs, NumBytes, ObjGPRs, ObjSize,
400 GPRPad, StackPad, Flags);
401 NumBytes += ObjSize + StackPad;
402 NumGPRs += ObjGPRs + GPRPad;
405 // Adjust the stack pointer for the new arguments...
406 // These operations are automatically eliminated by the prolog/epilog pass
407 Chain = DAG.getCALLSEQ_START(Chain,
408 DAG.getConstant(NumBytes, MVT::i32));
410 SDOperand StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
412 static const unsigned GPRArgRegs[] = {
413 ARM::R0, ARM::R1, ARM::R2, ARM::R3
417 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
418 std::vector<SDOperand> MemOpChains;
419 for (unsigned i = 0; i != NumOps; ++i) {
420 SDOperand Arg = Op.getOperand(5+2*i);
421 unsigned Flags = Op.getConstantOperandVal(5+2*i+1);
422 MVT::ValueType ArgVT = Arg.getValueType();
428 HowToPassArgument(ArgVT, NumGPRs, ArgOffset, ObjGPRs,
429 ObjSize, GPRPad, StackPad, Flags);
431 ArgOffset += StackPad;
434 default: assert(0 && "Unexpected ValueType for argument!");
436 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Arg));
439 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs],
440 DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Arg)));
443 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
444 DAG.getConstant(0, getPointerTy()));
445 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Arg,
446 DAG.getConstant(1, getPointerTy()));
447 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Lo));
449 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1], Hi));
451 SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
452 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
453 MemOpChains.push_back(DAG.getStore(Chain, Hi, PtrOff, NULL, 0));
458 SDOperand Cvt = DAG.getNode(ARMISD::FMRRD,
459 DAG.getVTList(MVT::i32, MVT::i32),
461 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs], Cvt));
463 RegsToPass.push_back(std::make_pair(GPRArgRegs[NumGPRs+1],
466 SDOperand PtrOff= DAG.getConstant(ArgOffset, StackPtr.getValueType());
467 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
468 MemOpChains.push_back(DAG.getStore(Chain, Cvt.getValue(1), PtrOff,
475 assert(ObjSize != 0);
476 SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
477 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
478 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
482 ArgOffset += ObjSize;
485 if (!MemOpChains.empty())
486 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
487 &MemOpChains[0], MemOpChains.size());
489 // Build a sequence of copy-to-reg nodes chained together with token chain
490 // and flag operands which copy the outgoing args into the appropriate regs.
492 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
493 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
495 InFlag = Chain.getValue(1);
498 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
499 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
500 // node so that legalize doesn't hack it.
501 bool isDirect = false;
502 bool isARMFunc = false;
503 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
504 GlobalValue *GV = G->getGlobal();
506 bool isExt = (GV->isDeclaration() || GV->hasWeakLinkage() ||
507 GV->hasLinkOnceLinkage());
508 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
509 getTargetMachine().getRelocationModel() != Reloc::Static;
510 isARMFunc = !Subtarget->isThumb() || isStub;
511 // tBX takes a register source operand.
512 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
513 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
515 SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
516 CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
517 Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
518 SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
519 Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
521 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
522 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
524 bool isStub = Subtarget->isTargetDarwin() &&
525 getTargetMachine().getRelocationModel() != Reloc::Static;
526 isARMFunc = !Subtarget->isThumb() || isStub;
527 // tBX takes a register source operand.
528 const char *Sym = S->getSymbol();
529 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) {
530 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
532 SDOperand CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 2);
533 CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
534 Callee = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), CPAddr, NULL, 0);
535 SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
536 Callee = DAG.getNode(ARMISD::PIC_ADD, getPointerTy(), Callee, PICLabel);
538 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
541 std::vector<MVT::ValueType> NodeTys;
542 NodeTys.push_back(MVT::Other); // Returns a chain
543 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
545 std::vector<SDOperand> Ops;
546 Ops.push_back(Chain);
547 Ops.push_back(Callee);
549 // Add argument registers to the end of the list so that they are known live
551 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
552 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
553 RegsToPass[i].second.getValueType()));
555 // FIXME: handle tail calls differently.
557 if (Subtarget->isThumb()) {
558 if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc))
559 CallOpc = ARMISD::CALL_NOLINK;
561 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
563 CallOpc = (isDirect || Subtarget->hasV5TOps())
564 ? ARMISD::CALL : ARMISD::CALL_NOLINK;
567 Ops.push_back(InFlag);
568 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
569 InFlag = Chain.getValue(1);
571 SDOperand CSOps[] = { Chain, DAG.getConstant(NumBytes, MVT::i32), InFlag };
572 Chain = DAG.getNode(ISD::CALLSEQ_END,
573 DAG.getNodeValueTypes(MVT::Other, MVT::Flag),
574 ((RetVT != MVT::Other) ? 2 : 1), CSOps, 3);
575 if (RetVT != MVT::Other)
576 InFlag = Chain.getValue(1);
578 std::vector<SDOperand> ResultVals;
581 // If the call has results, copy the values out of the ret val registers.
583 default: assert(0 && "Unexpected ret value!");
587 Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
588 ResultVals.push_back(Chain.getValue(0));
589 if (Op.Val->getValueType(1) == MVT::i32) {
590 // Returns a i64 value.
591 Chain = DAG.getCopyFromReg(Chain, ARM::R1, MVT::i32,
592 Chain.getValue(2)).getValue(1);
593 ResultVals.push_back(Chain.getValue(0));
594 NodeTys.push_back(MVT::i32);
596 NodeTys.push_back(MVT::i32);
599 Chain = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag).getValue(1);
600 ResultVals.push_back(DAG.getNode(ISD::BIT_CONVERT, MVT::f32,
602 NodeTys.push_back(MVT::f32);
605 SDOperand Lo = DAG.getCopyFromReg(Chain, ARM::R0, MVT::i32, InFlag);
606 SDOperand Hi = DAG.getCopyFromReg(Lo, ARM::R1, MVT::i32, Lo.getValue(2));
607 ResultVals.push_back(DAG.getNode(ARMISD::FMDRR, MVT::f64, Lo, Hi));
608 NodeTys.push_back(MVT::f64);
613 NodeTys.push_back(MVT::Other);
615 if (ResultVals.empty())
618 ResultVals.push_back(Chain);
619 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, &ResultVals[0],
621 return Res.getValue(Op.ResNo);
624 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG) {
626 SDOperand Chain = Op.getOperand(0);
627 switch(Op.getNumOperands()) {
629 assert(0 && "Do not know how to return this many arguments!");
632 SDOperand LR = DAG.getRegister(ARM::LR, MVT::i32);
633 return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Chain);
636 Op = Op.getOperand(1);
637 if (Op.getValueType() == MVT::f32) {
638 Op = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
639 } else if (Op.getValueType() == MVT::f64) {
640 // Recursively legalize f64 -> i64.
641 Op = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Op);
642 return DAG.getNode(ISD::RET, MVT::Other, Chain, Op,
643 DAG.getConstant(0, MVT::i32));
645 Copy = DAG.getCopyToReg(Chain, ARM::R0, Op, SDOperand());
646 if (DAG.getMachineFunction().liveout_empty())
647 DAG.getMachineFunction().addLiveOut(ARM::R0);
650 Copy = DAG.getCopyToReg(Chain, ARM::R1, Op.getOperand(3), SDOperand());
651 Copy = DAG.getCopyToReg(Copy, ARM::R0, Op.getOperand(1), Copy.getValue(1));
652 // If we haven't noted the R0+R1 are live out, do so now.
653 if (DAG.getMachineFunction().liveout_empty()) {
654 DAG.getMachineFunction().addLiveOut(ARM::R0);
655 DAG.getMachineFunction().addLiveOut(ARM::R1);
660 //We must use RET_FLAG instead of BRIND because BRIND doesn't have a flag
661 return DAG.getNode(ARMISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
664 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
665 // their target countpart wrapped in the ARMISD::Wrapper node. Suppose N is
666 // one of the above mentioned nodes. It has to be wrapped because otherwise
667 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
668 // be used to form addressing mode. These wrapped nodes will be selected
670 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
671 MVT::ValueType PtrVT = Op.getValueType();
672 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
674 if (CP->isMachineConstantPoolEntry())
675 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
678 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
680 return DAG.getNode(ARMISD::Wrapper, MVT::i32, Res);
683 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol
684 /// even in dynamic-no-pic mode.
685 static bool GVIsIndirectSymbol(GlobalValue *GV) {
686 return (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
687 (GV->isDeclaration() && !GV->hasNotBeenReadFromBytecode()));
690 SDOperand ARMTargetLowering::LowerGlobalAddress(SDOperand Op,
692 MVT::ValueType PtrVT = getPointerTy();
693 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
694 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
695 bool IsIndirect = Subtarget->isTargetDarwin() && GVIsIndirectSymbol(GV);
697 if (RelocM == Reloc::Static)
698 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 2);
700 unsigned PCAdj = (RelocM != Reloc::PIC_)
701 ? 0 : (Subtarget->isThumb() ? 4 : 8);
702 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr
704 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex,
706 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 2);
708 CPAddr = DAG.getNode(ARMISD::Wrapper, MVT::i32, CPAddr);
710 SDOperand Result = DAG.getLoad(PtrVT, DAG.getEntryNode(), CPAddr, NULL, 0);
711 SDOperand Chain = Result.getValue(1);
713 if (RelocM == Reloc::PIC_) {
714 SDOperand PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
715 Result = DAG.getNode(ARMISD::PIC_ADD, PtrVT, Result, PICLabel);
718 Result = DAG.getLoad(PtrVT, Chain, Result, NULL, 0);
723 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
724 unsigned VarArgsFrameIndex) {
725 // vastart just stores the address of the VarArgsFrameIndex slot into the
726 // memory location argument.
727 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
728 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
729 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
730 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV->getValue(),
734 static SDOperand LowerFORMAL_ARGUMENT(SDOperand Op, SelectionDAG &DAG,
735 unsigned *vRegs, unsigned ArgNo,
736 unsigned &NumGPRs, unsigned &ArgOffset) {
737 MachineFunction &MF = DAG.getMachineFunction();
738 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
739 SDOperand Root = Op.getOperand(0);
740 std::vector<SDOperand> ArgValues;
741 SSARegMap *RegMap = MF.getSSARegMap();
743 static const unsigned GPRArgRegs[] = {
744 ARM::R0, ARM::R1, ARM::R2, ARM::R3
751 unsigned Flags = Op.getConstantOperandVal(ArgNo + 3);
752 HowToPassArgument(ObjectVT, NumGPRs, ArgOffset, ObjGPRs,
753 ObjSize, GPRPad, StackPad, Flags);
755 ArgOffset += StackPad;
759 unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
760 MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
761 vRegs[NumGPRs] = VReg;
762 ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
763 if (ObjectVT == MVT::f32)
764 ArgValue = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, ArgValue);
765 } else if (ObjGPRs == 2) {
766 unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
767 MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
768 vRegs[NumGPRs] = VReg;
769 ArgValue = DAG.getCopyFromReg(Root, VReg, MVT::i32);
771 VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
772 MF.addLiveIn(GPRArgRegs[NumGPRs+1], VReg);
773 vRegs[NumGPRs+1] = VReg;
774 SDOperand ArgValue2 = DAG.getCopyFromReg(Root, VReg, MVT::i32);
776 if (ObjectVT == MVT::i64)
777 ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
779 ArgValue = DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
784 // If the argument is actually used, emit a load from the right stack
786 if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
787 MachineFrameInfo *MFI = MF.getFrameInfo();
788 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
789 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
791 ArgValue = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
793 SDOperand ArgValue2 =
794 DAG.getLoad(MVT::i32, Root, FIN, NULL, 0);
795 if (ObjectVT == MVT::i64)
796 ArgValue= DAG.getNode(ISD::BUILD_PAIR, MVT::i64, ArgValue, ArgValue2);
798 ArgValue= DAG.getNode(ARMISD::FMDRR, MVT::f64, ArgValue, ArgValue2);
801 // Don't emit a dead load.
802 ArgValue = DAG.getNode(ISD::UNDEF, ObjectVT);
805 ArgOffset += ObjSize; // Move on to the next argument.
812 ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG) {
813 std::vector<SDOperand> ArgValues;
814 SDOperand Root = Op.getOperand(0);
815 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
816 unsigned NumGPRs = 0; // GPRs used for parameter passing.
819 unsigned NumArgs = Op.Val->getNumValues()-1;
820 for (unsigned ArgNo = 0; ArgNo < NumArgs; ++ArgNo)
821 ArgValues.push_back(LowerFORMAL_ARGUMENT(Op, DAG, VRegs, ArgNo,
822 NumGPRs, ArgOffset));
824 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
826 static const unsigned GPRArgRegs[] = {
827 ARM::R0, ARM::R1, ARM::R2, ARM::R3
830 MachineFunction &MF = DAG.getMachineFunction();
831 SSARegMap *RegMap = MF.getSSARegMap();
832 MachineFrameInfo *MFI = MF.getFrameInfo();
833 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
834 unsigned VARegSaveSize = (4 - NumGPRs) * 4;
836 // If this function is vararg, store any remaining integer argument regs
837 // to their spots on the stack so that they may be loaded by deferencing
838 // the result of va_next.
839 AFI->setVarArgsRegSaveSize(VARegSaveSize);
840 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset);
841 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
843 SmallVector<SDOperand, 4> MemOps;
844 for (; NumGPRs < 4; ++NumGPRs) {
845 unsigned VReg = RegMap->createVirtualRegister(&ARM::GPRRegClass);
846 MF.addLiveIn(GPRArgRegs[NumGPRs], VReg);
847 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::i32);
848 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
849 MemOps.push_back(Store);
850 FIN = DAG.getNode(ISD::ADD, getPointerTy(), FIN,
851 DAG.getConstant(4, getPointerTy()));
854 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
855 &MemOps[0], MemOps.size());
857 // This will point to the next argument passed via stack.
858 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
861 ArgValues.push_back(Root);
863 // Return the new list of results.
864 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
865 Op.Val->value_end());
866 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
869 /// isFloatingPointZero - Return true if this is +0.0.
870 static bool isFloatingPointZero(SDOperand Op) {
871 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
872 return CFP->isExactlyValue(0.0);
873 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
874 // Maybe this has already been legalized into the constant pool?
875 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
876 SDOperand WrapperOp = Op.getOperand(1).getOperand(0);
877 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
878 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
879 return CFP->isExactlyValue(0.0);
885 static bool isLegalCmpImmediate(unsigned C, bool isThumb) {
886 return ( isThumb && (C & ~255U) == 0) ||
887 (!isThumb && ARM_AM::getSOImmVal(C) != -1);
890 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
891 /// the given operands.
892 static SDOperand getARMCmp(SDOperand LHS, SDOperand RHS, ISD::CondCode CC,
893 SDOperand &ARMCC, SelectionDAG &DAG, bool isThumb) {
894 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.Val)) {
895 unsigned C = RHSC->getValue();
896 if (!isLegalCmpImmediate(C, isThumb)) {
897 // Constant does not fit, try adjusting it by one?
902 if (isLegalCmpImmediate(C-1, isThumb)) {
903 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
904 RHS = DAG.getConstant(C-1, MVT::i32);
909 if (C > 0 && isLegalCmpImmediate(C-1, isThumb)) {
910 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
911 RHS = DAG.getConstant(C-1, MVT::i32);
916 if (isLegalCmpImmediate(C+1, isThumb)) {
917 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
918 RHS = DAG.getConstant(C+1, MVT::i32);
923 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb)) {
924 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
925 RHS = DAG.getConstant(C+1, MVT::i32);
932 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
933 ARMCC = DAG.getConstant(CondCode, MVT::i32);
934 return DAG.getNode(ARMISD::CMP, MVT::Flag, LHS, RHS);
937 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
938 static SDOperand getVFPCmp(SDOperand LHS, SDOperand RHS, SelectionDAG &DAG) {
940 if (!isFloatingPointZero(RHS))
941 Cmp = DAG.getNode(ARMISD::CMPFP, MVT::Flag, LHS, RHS);
943 Cmp = DAG.getNode(ARMISD::CMPFPw0, MVT::Flag, LHS);
944 return DAG.getNode(ARMISD::FMSTAT, MVT::Flag, Cmp);
947 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG,
948 const ARMSubtarget *ST) {
949 MVT::ValueType VT = Op.getValueType();
950 SDOperand LHS = Op.getOperand(0);
951 SDOperand RHS = Op.getOperand(1);
952 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
953 SDOperand TrueVal = Op.getOperand(2);
954 SDOperand FalseVal = Op.getOperand(3);
956 if (LHS.getValueType() == MVT::i32) {
958 SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
959 return DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal, ARMCC, Cmp);
962 ARMCC::CondCodes CondCode, CondCode2;
963 if (FPCCToARMCC(CC, CondCode, CondCode2))
964 std::swap(TrueVal, FalseVal);
966 SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
967 SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
968 SDOperand Result = DAG.getNode(ARMISD::CMOV, VT, FalseVal, TrueVal,
970 if (CondCode2 != ARMCC::AL) {
971 SDOperand ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
972 // FIXME: Needs another CMP because flag can have but one use.
973 SDOperand Cmp2 = getVFPCmp(LHS, RHS, DAG);
974 Result = DAG.getNode(ARMISD::CMOV, VT, Result, TrueVal, ARMCC2, Cmp2);
979 static SDOperand LowerBR_CC(SDOperand Op, SelectionDAG &DAG,
980 const ARMSubtarget *ST) {
981 SDOperand Chain = Op.getOperand(0);
982 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
983 SDOperand LHS = Op.getOperand(2);
984 SDOperand RHS = Op.getOperand(3);
985 SDOperand Dest = Op.getOperand(4);
987 if (LHS.getValueType() == MVT::i32) {
989 SDOperand Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb());
990 return DAG.getNode(ARMISD::BRCOND, MVT::Other, Chain, Dest, ARMCC, Cmp);
993 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
994 ARMCC::CondCodes CondCode, CondCode2;
995 if (FPCCToARMCC(CC, CondCode, CondCode2))
996 // Swap the LHS/RHS of the comparison if needed.
999 SDOperand Cmp = getVFPCmp(LHS, RHS, DAG);
1000 SDOperand ARMCC = DAG.getConstant(CondCode, MVT::i32);
1001 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1002 SDOperand Ops[] = { Chain, Dest, ARMCC, Cmp };
1003 SDOperand Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 4);
1004 if (CondCode2 != ARMCC::AL) {
1005 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1006 SDOperand Ops[] = { Res, Dest, ARMCC, Res.getValue(1) };
1007 Res = DAG.getNode(ARMISD::BRCOND, VTList, Ops, 4);
1012 SDOperand ARMTargetLowering::LowerBR_JT(SDOperand Op, SelectionDAG &DAG) {
1013 SDOperand Chain = Op.getOperand(0);
1014 SDOperand Table = Op.getOperand(1);
1015 SDOperand Index = Op.getOperand(2);
1017 MVT::ValueType PTy = getPointerTy();
1018 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1019 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1020 SDOperand UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1021 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1022 Table = DAG.getNode(ARMISD::WrapperJT, MVT::i32, JTI, UId);
1023 Index = DAG.getNode(ISD::MUL, PTy, Index, DAG.getConstant(4, PTy));
1024 SDOperand Addr = DAG.getNode(ISD::ADD, PTy, Index, Table);
1025 bool isPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
1026 Addr = DAG.getLoad(isPIC ? MVT::i32 : PTy, Chain, Addr, NULL, 0);
1027 Chain = Addr.getValue(1);
1029 Addr = DAG.getNode(ISD::ADD, PTy, Addr, Table);
1030 return DAG.getNode(ARMISD::BR_JT, MVT::Other, Chain, Addr, JTI, UId);
1033 static SDOperand LowerFP_TO_INT(SDOperand Op, SelectionDAG &DAG) {
1035 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1036 Op = DAG.getNode(Opc, MVT::f32, Op.getOperand(0));
1037 return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
1040 static SDOperand LowerINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
1041 MVT::ValueType VT = Op.getValueType();
1043 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1045 Op = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Op.getOperand(0));
1046 return DAG.getNode(Opc, VT, Op);
1049 static SDOperand LowerFCOPYSIGN(SDOperand Op, SelectionDAG &DAG) {
1050 // Implement fcopysign with a fabs and a conditional fneg.
1051 SDOperand Tmp0 = Op.getOperand(0);
1052 SDOperand Tmp1 = Op.getOperand(1);
1053 MVT::ValueType VT = Op.getValueType();
1054 MVT::ValueType SrcVT = Tmp1.getValueType();
1055 SDOperand AbsVal = DAG.getNode(ISD::FABS, VT, Tmp0);
1056 SDOperand Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG);
1057 SDOperand ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1058 return DAG.getNode(ARMISD::CNEG, VT, AbsVal, AbsVal, ARMCC, Cmp);
1061 static SDOperand LowerBIT_CONVERT(SDOperand Op, SelectionDAG &DAG) {
1062 // Turn f64->i64 into FMRRD.
1063 assert(Op.getValueType() == MVT::i64 &&
1064 Op.getOperand(0).getValueType() == MVT::f64);
1066 Op = Op.getOperand(0);
1067 SDOperand Cvt = DAG.getNode(ARMISD::FMRRD, DAG.getVTList(MVT::i32, MVT::i32),
1070 // Merge the pieces into a single i64 value.
1071 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Cvt, Cvt.getValue(1));
1074 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
1075 // FIXME: All this code is target-independent. Create a new target-indep
1076 // MULHILO node and move this code to the legalizer.
1078 assert(Op.getValueType() == MVT::i64 && "Only handles i64 expand right now!");
1080 SDOperand LL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1081 DAG.getConstant(0, MVT::i32));
1082 SDOperand RL = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(1),
1083 DAG.getConstant(0, MVT::i32));
1085 const TargetLowering &TL = DAG.getTargetLoweringInfo();
1086 unsigned LHSSB = TL.ComputeNumSignBits(Op.getOperand(0));
1087 unsigned RHSSB = TL.ComputeNumSignBits(Op.getOperand(1));
1090 // Figure out how to lower this multiply.
1091 if (LHSSB >= 33 && RHSSB >= 33) {
1092 // If the input values are both sign extended, we can emit a mulhs+mul.
1093 Lo = DAG.getNode(ISD::MUL, MVT::i32, LL, RL);
1094 Hi = DAG.getNode(ISD::MULHS, MVT::i32, LL, RL);
1095 } else if (LHSSB == 32 && RHSSB == 32 &&
1096 TL.MaskedValueIsZero(Op.getOperand(0), 0xFFFFFFFF00000000ULL) &&
1097 TL.MaskedValueIsZero(Op.getOperand(1), 0xFFFFFFFF00000000ULL)) {
1098 // If the inputs are zero extended, use mulhu.
1099 Lo = DAG.getNode(ISD::MUL, MVT::i32, LL, RL);
1100 Hi = DAG.getNode(ISD::MULHU, MVT::i32, LL, RL);
1102 SDOperand LH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1103 DAG.getConstant(1, MVT::i32));
1104 SDOperand RH = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(1),
1105 DAG.getConstant(1, MVT::i32));
1107 // Lo,Hi = umul LHS, RHS.
1108 SDOperand Ops[] = { LL, RL };
1109 SDOperand UMul64 = DAG.getNode(ARMISD::MULHILOU,
1110 DAG.getVTList(MVT::i32, MVT::i32), Ops, 2);
1112 Hi = UMul64.getValue(1);
1113 RH = DAG.getNode(ISD::MUL, MVT::i32, LL, RH);
1114 LH = DAG.getNode(ISD::MUL, MVT::i32, LH, RL);
1115 Hi = DAG.getNode(ISD::ADD, MVT::i32, Hi, RH);
1116 Hi = DAG.getNode(ISD::ADD, MVT::i32, Hi, LH);
1119 // Merge the pieces into a single i64 value.
1120 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
1123 static SDOperand LowerMULHU(SDOperand Op, SelectionDAG &DAG) {
1124 SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
1125 return DAG.getNode(ARMISD::MULHILOU,
1126 DAG.getVTList(MVT::i32, MVT::i32), Ops, 2).getValue(1);
1129 static SDOperand LowerMULHS(SDOperand Op, SelectionDAG &DAG) {
1130 SDOperand Ops[] = { Op.getOperand(0), Op.getOperand(1) };
1131 return DAG.getNode(ARMISD::MULHILOS,
1132 DAG.getVTList(MVT::i32, MVT::i32), Ops, 2).getValue(1);
1135 static SDOperand LowerSRx(SDOperand Op, SelectionDAG &DAG,
1136 const ARMSubtarget *ST) {
1137 assert(Op.getValueType() == MVT::i64 &&
1138 (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) &&
1139 "Unknown shift to lower!");
1141 // We only lower SRA, SRL of 1 here, all others use generic lowering.
1142 if (!isa<ConstantSDNode>(Op.getOperand(1)) ||
1143 cast<ConstantSDNode>(Op.getOperand(1))->getValue() != 1)
1146 // If we are in thumb mode, we don't have RRX.
1147 if (ST->isThumb()) return SDOperand();
1149 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
1150 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1151 DAG.getConstant(0, MVT::i32));
1152 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
1153 DAG.getConstant(1, MVT::i32));
1155 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
1156 // captures the result into a carry flag.
1157 unsigned Opc = Op.getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
1158 Hi = DAG.getNode(Opc, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
1160 // The low part is an ARMISD::RRX operand, which shifts the carry in.
1161 Lo = DAG.getNode(ARMISD::RRX, MVT::i32, Lo, Hi.getValue(1));
1163 // Merge the pieces into a single i64 value.
1164 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
1167 SDOperand ARMTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
1168 switch (Op.getOpcode()) {
1169 default: assert(0 && "Don't know how to custom lower this!"); abort();
1170 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1171 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
1172 case ISD::CALL: return LowerCALL(Op, DAG);
1173 case ISD::RET: return LowerRET(Op, DAG);
1174 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
1175 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
1176 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
1177 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1178 case ISD::SINT_TO_FP:
1179 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
1180 case ISD::FP_TO_SINT:
1181 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
1182 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
1183 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
1184 case ISD::MUL: return LowerMUL(Op, DAG);
1185 case ISD::MULHU: return LowerMULHU(Op, DAG);
1186 case ISD::MULHS: return LowerMULHS(Op, DAG);
1188 case ISD::SRA: return LowerSRx(Op, DAG, Subtarget);
1189 case ISD::FORMAL_ARGUMENTS:
1190 return LowerFORMAL_ARGUMENTS(Op, DAG);
1191 case ISD::RETURNADDR: break;
1192 case ISD::FRAMEADDR: break;
1197 //===----------------------------------------------------------------------===//
1198 // ARM Scheduler Hooks
1199 //===----------------------------------------------------------------------===//
1202 ARMTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1203 MachineBasicBlock *BB) {
1204 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1205 switch (MI->getOpcode()) {
1206 default: assert(false && "Unexpected instr type to insert");
1207 case ARM::tMOVCCr: {
1208 // To "insert" a SELECT_CC instruction, we actually have to insert the
1209 // diamond control-flow pattern. The incoming instruction knows the
1210 // destination vreg to set, the condition code register to branch on, the
1211 // true/false values to select between, and a branch opcode to use.
1212 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1213 ilist<MachineBasicBlock>::iterator It = BB;
1219 // cmpTY ccX, r1, r2
1221 // fallthrough --> copy0MBB
1222 MachineBasicBlock *thisMBB = BB;
1223 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1224 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1225 BuildMI(BB, TII->get(ARM::tBcc)).addMBB(sinkMBB)
1226 .addImm(MI->getOperand(3).getImm());
1227 MachineFunction *F = BB->getParent();
1228 F->getBasicBlockList().insert(It, copy0MBB);
1229 F->getBasicBlockList().insert(It, sinkMBB);
1230 // Update machine-CFG edges by first adding all successors of the current
1231 // block to the new block which will contain the Phi node for the select.
1232 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1233 e = BB->succ_end(); i != e; ++i)
1234 sinkMBB->addSuccessor(*i);
1235 // Next, remove all successors of the current block, and add the true
1236 // and fallthrough blocks as its successors.
1237 while(!BB->succ_empty())
1238 BB->removeSuccessor(BB->succ_begin());
1239 BB->addSuccessor(copy0MBB);
1240 BB->addSuccessor(sinkMBB);
1243 // %FalseValue = ...
1244 // # fallthrough to sinkMBB
1247 // Update machine-CFG edges
1248 BB->addSuccessor(sinkMBB);
1251 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1254 BuildMI(BB, TII->get(ARM::PHI), MI->getOperand(0).getReg())
1255 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
1256 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1258 delete MI; // The pseudo instruction is gone now.
1264 //===----------------------------------------------------------------------===//
1265 // ARM Optimization Hooks
1266 //===----------------------------------------------------------------------===//
1268 /// isLegalAddressImmediate - Return true if the integer value or
1269 /// GlobalValue can be used as the offset of the target addressing mode.
1270 bool ARMTargetLowering::isLegalAddressImmediate(int64_t V) const {
1271 // ARM allows a 12-bit immediate field.
1272 return V == V & ((1LL << 12) - 1);
1275 bool ARMTargetLowering::isLegalAddressImmediate(GlobalValue *GV) const {
1279 static bool getIndexedAddressParts(SDNode *Ptr, MVT::ValueType VT,
1280 bool isSEXTLoad, SDOperand &Base,
1281 SDOperand &Offset, bool &isInc,
1282 SelectionDAG &DAG) {
1283 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
1286 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
1288 Base = Ptr->getOperand(0);
1289 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1290 int RHSC = (int)RHS->getValue();
1291 if (RHSC < 0 && RHSC > -256) {
1293 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1297 isInc = (Ptr->getOpcode() == ISD::ADD);
1298 Offset = Ptr->getOperand(1);
1300 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
1302 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
1303 int RHSC = (int)RHS->getValue();
1304 if (RHSC < 0 && RHSC > -0x1000) {
1306 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
1307 Base = Ptr->getOperand(0);
1312 if (Ptr->getOpcode() == ISD::ADD) {
1314 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
1315 if (ShOpcVal != ARM_AM::no_shift) {
1316 Base = Ptr->getOperand(1);
1317 Offset = Ptr->getOperand(0);
1319 Base = Ptr->getOperand(0);
1320 Offset = Ptr->getOperand(1);
1325 isInc = (Ptr->getOpcode() == ISD::ADD);
1326 Base = Ptr->getOperand(0);
1327 Offset = Ptr->getOperand(1);
1331 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
1335 /// getPreIndexedAddressParts - returns true by value, base pointer and
1336 /// offset pointer and addressing mode by reference if the node's address
1337 /// can be legally represented as pre-indexed load / store address.
1339 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
1341 ISD::MemIndexedMode &AM,
1342 SelectionDAG &DAG) {
1343 if (Subtarget->isThumb())
1348 bool isSEXTLoad = false;
1349 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1350 Ptr = LD->getBasePtr();
1351 VT = LD->getLoadedVT();
1352 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1353 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1354 Ptr = ST->getBasePtr();
1355 VT = ST->getStoredVT();
1360 bool isLegal = getIndexedAddressParts(Ptr.Val, VT, isSEXTLoad, Base, Offset,
1363 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
1369 /// getPostIndexedAddressParts - returns true by value, base pointer and
1370 /// offset pointer and addressing mode by reference if this node can be
1371 /// combined with a load / store to form a post-indexed load / store.
1372 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
1375 ISD::MemIndexedMode &AM,
1376 SelectionDAG &DAG) {
1377 if (Subtarget->isThumb())
1382 bool isSEXTLoad = false;
1383 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1384 VT = LD->getLoadedVT();
1385 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
1386 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1387 VT = ST->getStoredVT();
1392 bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
1395 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
1401 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1403 uint64_t &KnownZero,
1405 unsigned Depth) const {
1408 switch (Op.getOpcode()) {
1410 case ARMISD::CMOV: {
1411 // Bits are known zero/one if known on the LHS and RHS.
1412 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
1413 if (KnownZero == 0 && KnownOne == 0) return;
1415 uint64_t KnownZeroRHS, KnownOneRHS;
1416 ComputeMaskedBits(Op.getOperand(1), Mask,
1417 KnownZeroRHS, KnownOneRHS, Depth+1);
1418 KnownZero &= KnownZeroRHS;
1419 KnownOne &= KnownOneRHS;
1425 //===----------------------------------------------------------------------===//
1426 // ARM Inline Assembly Support
1427 //===----------------------------------------------------------------------===//
1429 /// getConstraintType - Given a constraint letter, return the type of
1430 /// constraint it is for this target.
1431 ARMTargetLowering::ConstraintType
1432 ARMTargetLowering::getConstraintType(char ConstraintLetter) const {
1433 switch (ConstraintLetter) {
1435 return C_RegisterClass;
1436 default: return TargetLowering::getConstraintType(ConstraintLetter);
1440 std::pair<unsigned, const TargetRegisterClass*>
1441 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
1442 MVT::ValueType VT) const {
1443 if (Constraint.size() == 1) {
1444 // GCC RS6000 Constraint Letters
1445 switch (Constraint[0]) {
1447 // FIXME: in thumb mode, 'l' is only low-regs.
1450 return std::make_pair(0U, ARM::GPRRegisterClass);
1454 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
1457 std::vector<unsigned> ARMTargetLowering::
1458 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1459 MVT::ValueType VT) const {
1460 if (Constraint.size() != 1)
1461 return std::vector<unsigned>();
1463 switch (Constraint[0]) { // GCC ARM Constraint Letters
1467 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
1468 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
1469 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
1470 ARM::R12, ARM::LR, 0);
1473 return std::vector<unsigned>();