1 //===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that ARM uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "ARMAddressingModes.h"
17 #include "ARMConstantPoolValue.h"
18 #include "ARMISelLowering.h"
19 #include "ARMMachineFunctionInfo.h"
20 #include "ARMPerfectShuffle.h"
21 #include "ARMRegisterInfo.h"
22 #include "ARMSubtarget.h"
23 #include "ARMTargetMachine.h"
24 #include "ARMTargetObjectFile.h"
25 #include "llvm/CallingConv.h"
26 #include "llvm/Constants.h"
27 #include "llvm/Function.h"
28 #include "llvm/GlobalValue.h"
29 #include "llvm/Instruction.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Type.h"
32 #include "llvm/CodeGen/CallingConvLower.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineFrameInfo.h"
35 #include "llvm/CodeGen/MachineFunction.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/VectorExtras.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
48 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
49 CCValAssign::LocInfo &LocInfo,
50 ISD::ArgFlagsTy &ArgFlags,
52 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
53 CCValAssign::LocInfo &LocInfo,
54 ISD::ArgFlagsTy &ArgFlags,
56 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
57 CCValAssign::LocInfo &LocInfo,
58 ISD::ArgFlagsTy &ArgFlags,
60 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
61 CCValAssign::LocInfo &LocInfo,
62 ISD::ArgFlagsTy &ArgFlags,
65 void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
66 EVT PromotedBitwiseVT) {
67 if (VT != PromotedLdStVT) {
68 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
69 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
70 PromotedLdStVT.getSimpleVT());
72 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
73 AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
74 PromotedLdStVT.getSimpleVT());
77 EVT ElemTy = VT.getVectorElementType();
78 if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
79 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
80 if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
81 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
82 if (ElemTy != MVT::i32) {
83 setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
84 setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
85 setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
86 setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
88 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
89 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
90 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
91 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
93 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
94 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
95 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
98 // Promote all bit-wise operations.
99 if (VT.isInteger() && VT != PromotedBitwiseVT) {
100 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
101 AddPromotedToType (ISD::AND, VT.getSimpleVT(),
102 PromotedBitwiseVT.getSimpleVT());
103 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
104 AddPromotedToType (ISD::OR, VT.getSimpleVT(),
105 PromotedBitwiseVT.getSimpleVT());
106 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
107 AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
108 PromotedBitwiseVT.getSimpleVT());
111 // Neon does not support vector divide/remainder operations.
112 setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
113 setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
114 setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
115 setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
116 setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
117 setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
120 void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
121 addRegisterClass(VT, ARM::DPRRegisterClass);
122 addTypeForNEON(VT, MVT::f64, MVT::v2i32);
125 void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
126 addRegisterClass(VT, ARM::QPRRegisterClass);
127 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
130 static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
131 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
132 return new TargetLoweringObjectFileMachO();
133 return new ARMElfTargetObjectFile();
136 ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
137 : TargetLowering(TM, createTLOF(TM)) {
138 Subtarget = &TM.getSubtarget<ARMSubtarget>();
140 if (Subtarget->isTargetDarwin()) {
141 // Uses VFP for Thumb libfuncs if available.
142 if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
143 // Single-precision floating-point arithmetic.
144 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
145 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
146 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
147 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
149 // Double-precision floating-point arithmetic.
150 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
151 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
152 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
153 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
155 // Single-precision comparisons.
156 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
157 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
158 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
159 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
160 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
161 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
162 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
163 setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
165 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
166 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
167 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
168 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
169 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
170 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
171 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
172 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
174 // Double-precision comparisons.
175 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
176 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
177 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
178 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
179 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
180 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
181 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
182 setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
184 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
185 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
186 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
187 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
188 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
189 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
190 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
191 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
193 // Floating-point to integer conversions.
194 // i64 conversions are done via library routines even when generating VFP
195 // instructions, so use the same ones.
196 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
197 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
198 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
199 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
201 // Conversions between floating types.
202 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
203 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
205 // Integer to floating-point conversions.
206 // i64 conversions are done via library routines even when generating VFP
207 // instructions, so use the same ones.
208 // FIXME: There appears to be some naming inconsistency in ARM libgcc:
209 // e.g., __floatunsidf vs. __floatunssidfvfp.
210 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
211 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
212 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
213 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
217 // These libcalls are not available in 32-bit.
218 setLibcallName(RTLIB::SHL_I128, 0);
219 setLibcallName(RTLIB::SRL_I128, 0);
220 setLibcallName(RTLIB::SRA_I128, 0);
222 // Libcalls should use the AAPCS base standard ABI, even if hard float
223 // is in effect, as per the ARM RTABI specification, section 4.1.2.
224 if (Subtarget->isAAPCS_ABI()) {
225 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
226 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
227 CallingConv::ARM_AAPCS);
231 if (Subtarget->isThumb1Only())
232 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
234 addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
235 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
236 addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
237 addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
239 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
242 if (Subtarget->hasNEON()) {
243 addDRTypeForNEON(MVT::v2f32);
244 addDRTypeForNEON(MVT::v8i8);
245 addDRTypeForNEON(MVT::v4i16);
246 addDRTypeForNEON(MVT::v2i32);
247 addDRTypeForNEON(MVT::v1i64);
249 addQRTypeForNEON(MVT::v4f32);
250 addQRTypeForNEON(MVT::v2f64);
251 addQRTypeForNEON(MVT::v16i8);
252 addQRTypeForNEON(MVT::v8i16);
253 addQRTypeForNEON(MVT::v4i32);
254 addQRTypeForNEON(MVT::v2i64);
256 // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
257 // neither Neon nor VFP support any arithmetic operations on it.
258 setOperationAction(ISD::FADD, MVT::v2f64, Expand);
259 setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
260 setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
261 setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
262 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
263 setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
264 setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
265 setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
266 setOperationAction(ISD::FABS, MVT::v2f64, Expand);
267 setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
268 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
269 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
270 setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
271 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
272 setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
273 setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
274 setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
275 setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
276 setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
277 setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
278 setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
279 setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
280 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
281 setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
283 // Neon does not support some operations on v1i64 and v2i64 types.
284 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
285 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
286 setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
287 setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
289 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
290 setTargetDAGCombine(ISD::SHL);
291 setTargetDAGCombine(ISD::SRL);
292 setTargetDAGCombine(ISD::SRA);
293 setTargetDAGCombine(ISD::SIGN_EXTEND);
294 setTargetDAGCombine(ISD::ZERO_EXTEND);
295 setTargetDAGCombine(ISD::ANY_EXTEND);
298 computeRegisterProperties();
300 // ARM does not have f32 extending load.
301 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
303 // ARM does not have i1 sign extending load.
304 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
306 // ARM supports all 4 flavors of integer indexed load / store.
307 if (!Subtarget->isThumb1Only()) {
308 for (unsigned im = (unsigned)ISD::PRE_INC;
309 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
310 setIndexedLoadAction(im, MVT::i1, Legal);
311 setIndexedLoadAction(im, MVT::i8, Legal);
312 setIndexedLoadAction(im, MVT::i16, Legal);
313 setIndexedLoadAction(im, MVT::i32, Legal);
314 setIndexedStoreAction(im, MVT::i1, Legal);
315 setIndexedStoreAction(im, MVT::i8, Legal);
316 setIndexedStoreAction(im, MVT::i16, Legal);
317 setIndexedStoreAction(im, MVT::i32, Legal);
321 // i64 operation support.
322 if (Subtarget->isThumb1Only()) {
323 setOperationAction(ISD::MUL, MVT::i64, Expand);
324 setOperationAction(ISD::MULHU, MVT::i32, Expand);
325 setOperationAction(ISD::MULHS, MVT::i32, Expand);
326 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
327 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
329 setOperationAction(ISD::MUL, MVT::i64, Expand);
330 setOperationAction(ISD::MULHU, MVT::i32, Expand);
331 if (!Subtarget->hasV6Ops())
332 setOperationAction(ISD::MULHS, MVT::i32, Expand);
334 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
335 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
336 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
337 setOperationAction(ISD::SRL, MVT::i64, Custom);
338 setOperationAction(ISD::SRA, MVT::i64, Custom);
340 // ARM does not have ROTL.
341 setOperationAction(ISD::ROTL, MVT::i32, Expand);
342 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
343 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
344 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
345 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
347 // Only ARMv6 has BSWAP.
348 if (!Subtarget->hasV6Ops())
349 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
351 // These are expanded into libcalls.
352 setOperationAction(ISD::SDIV, MVT::i32, Expand);
353 setOperationAction(ISD::UDIV, MVT::i32, Expand);
354 setOperationAction(ISD::SREM, MVT::i32, Expand);
355 setOperationAction(ISD::UREM, MVT::i32, Expand);
356 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
357 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
359 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
360 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
361 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
362 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
363 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
365 // Use the default implementation.
366 setOperationAction(ISD::VASTART, MVT::Other, Custom);
367 setOperationAction(ISD::VAARG, MVT::Other, Expand);
368 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
369 setOperationAction(ISD::VAEND, MVT::Other, Expand);
370 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
371 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
372 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
373 // FIXME: Shouldn't need this, since no register is used, but the legalizer
374 // doesn't yet know how to not do that for SjLj.
375 setExceptionSelectorRegister(ARM::R0);
376 if (Subtarget->isThumb())
377 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
379 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
380 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
382 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
383 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
384 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
386 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
388 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
389 // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR iff target supports vfp2.
390 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
392 // We want to custom lower some of our intrinsics.
393 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
395 setOperationAction(ISD::SETCC, MVT::i32, Expand);
396 setOperationAction(ISD::SETCC, MVT::f32, Expand);
397 setOperationAction(ISD::SETCC, MVT::f64, Expand);
398 setOperationAction(ISD::SELECT, MVT::i32, Expand);
399 setOperationAction(ISD::SELECT, MVT::f32, Expand);
400 setOperationAction(ISD::SELECT, MVT::f64, Expand);
401 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
402 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
403 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
405 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
406 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
407 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
408 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
409 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
411 // We don't support sin/cos/fmod/copysign/pow
412 setOperationAction(ISD::FSIN, MVT::f64, Expand);
413 setOperationAction(ISD::FSIN, MVT::f32, Expand);
414 setOperationAction(ISD::FCOS, MVT::f32, Expand);
415 setOperationAction(ISD::FCOS, MVT::f64, Expand);
416 setOperationAction(ISD::FREM, MVT::f64, Expand);
417 setOperationAction(ISD::FREM, MVT::f32, Expand);
418 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
419 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
420 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
422 setOperationAction(ISD::FPOW, MVT::f64, Expand);
423 setOperationAction(ISD::FPOW, MVT::f32, Expand);
425 // int <-> fp are custom expanded into bit_convert + ARMISD ops.
426 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
427 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
428 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
429 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
430 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
433 // We have target-specific dag combine patterns for the following nodes:
434 // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
435 setTargetDAGCombine(ISD::ADD);
436 setTargetDAGCombine(ISD::SUB);
438 setStackPointerRegisterToSaveRestore(ARM::SP);
439 setSchedulingPreference(SchedulingForRegPressure);
441 // FIXME: If-converter should use instruction latency to determine
442 // profitability rather than relying on fixed limits.
443 if (Subtarget->getCPUString() == "generic") {
444 // Generic (and overly aggressive) if-conversion limits.
445 setIfCvtBlockSizeLimit(10);
446 setIfCvtDupBlockSizeLimit(2);
447 } else if (Subtarget->hasV6Ops()) {
448 setIfCvtBlockSizeLimit(2);
449 setIfCvtDupBlockSizeLimit(1);
451 setIfCvtBlockSizeLimit(3);
452 setIfCvtDupBlockSizeLimit(2);
455 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
456 // Do not enable CodePlacementOpt for now: it currently runs after the
457 // ARMConstantIslandPass and messes up branch relaxation and placement
458 // of constant islands.
459 // benefitFromCodePlacementOpt = true;
462 const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
465 case ARMISD::Wrapper: return "ARMISD::Wrapper";
466 case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
467 case ARMISD::CALL: return "ARMISD::CALL";
468 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
469 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
470 case ARMISD::tCALL: return "ARMISD::tCALL";
471 case ARMISD::BRCOND: return "ARMISD::BRCOND";
472 case ARMISD::BR_JT: return "ARMISD::BR_JT";
473 case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
474 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
475 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
476 case ARMISD::CMP: return "ARMISD::CMP";
477 case ARMISD::CMPZ: return "ARMISD::CMPZ";
478 case ARMISD::CMPFP: return "ARMISD::CMPFP";
479 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
480 case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
481 case ARMISD::CMOV: return "ARMISD::CMOV";
482 case ARMISD::CNEG: return "ARMISD::CNEG";
484 case ARMISD::FTOSI: return "ARMISD::FTOSI";
485 case ARMISD::FTOUI: return "ARMISD::FTOUI";
486 case ARMISD::SITOF: return "ARMISD::SITOF";
487 case ARMISD::UITOF: return "ARMISD::UITOF";
489 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
490 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
491 case ARMISD::RRX: return "ARMISD::RRX";
493 case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
494 case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
496 case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
497 case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
499 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
501 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
503 case ARMISD::VCEQ: return "ARMISD::VCEQ";
504 case ARMISD::VCGE: return "ARMISD::VCGE";
505 case ARMISD::VCGEU: return "ARMISD::VCGEU";
506 case ARMISD::VCGT: return "ARMISD::VCGT";
507 case ARMISD::VCGTU: return "ARMISD::VCGTU";
508 case ARMISD::VTST: return "ARMISD::VTST";
510 case ARMISD::VSHL: return "ARMISD::VSHL";
511 case ARMISD::VSHRs: return "ARMISD::VSHRs";
512 case ARMISD::VSHRu: return "ARMISD::VSHRu";
513 case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
514 case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
515 case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
516 case ARMISD::VSHRN: return "ARMISD::VSHRN";
517 case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
518 case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
519 case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
520 case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
521 case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
522 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
523 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
524 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
525 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
526 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
527 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
528 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
529 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
530 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
531 case ARMISD::VDUP: return "ARMISD::VDUP";
532 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
533 case ARMISD::VEXT: return "ARMISD::VEXT";
534 case ARMISD::VREV64: return "ARMISD::VREV64";
535 case ARMISD::VREV32: return "ARMISD::VREV32";
536 case ARMISD::VREV16: return "ARMISD::VREV16";
537 case ARMISD::VZIP: return "ARMISD::VZIP";
538 case ARMISD::VUZP: return "ARMISD::VUZP";
539 case ARMISD::VTRN: return "ARMISD::VTRN";
543 /// getFunctionAlignment - Return the Log2 alignment of this function.
544 unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
545 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
548 //===----------------------------------------------------------------------===//
550 //===----------------------------------------------------------------------===//
552 /// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
553 static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
555 default: llvm_unreachable("Unknown condition code!");
556 case ISD::SETNE: return ARMCC::NE;
557 case ISD::SETEQ: return ARMCC::EQ;
558 case ISD::SETGT: return ARMCC::GT;
559 case ISD::SETGE: return ARMCC::GE;
560 case ISD::SETLT: return ARMCC::LT;
561 case ISD::SETLE: return ARMCC::LE;
562 case ISD::SETUGT: return ARMCC::HI;
563 case ISD::SETUGE: return ARMCC::HS;
564 case ISD::SETULT: return ARMCC::LO;
565 case ISD::SETULE: return ARMCC::LS;
569 /// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
570 static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
571 ARMCC::CondCodes &CondCode2) {
572 CondCode2 = ARMCC::AL;
574 default: llvm_unreachable("Unknown FP condition!");
576 case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
578 case ISD::SETOGT: CondCode = ARMCC::GT; break;
580 case ISD::SETOGE: CondCode = ARMCC::GE; break;
581 case ISD::SETOLT: CondCode = ARMCC::MI; break;
582 case ISD::SETOLE: CondCode = ARMCC::LS; break;
583 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
584 case ISD::SETO: CondCode = ARMCC::VC; break;
585 case ISD::SETUO: CondCode = ARMCC::VS; break;
586 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
587 case ISD::SETUGT: CondCode = ARMCC::HI; break;
588 case ISD::SETUGE: CondCode = ARMCC::PL; break;
590 case ISD::SETULT: CondCode = ARMCC::LT; break;
592 case ISD::SETULE: CondCode = ARMCC::LE; break;
594 case ISD::SETUNE: CondCode = ARMCC::NE; break;
598 //===----------------------------------------------------------------------===//
599 // Calling Convention Implementation
600 //===----------------------------------------------------------------------===//
602 #include "ARMGenCallingConv.inc"
604 // APCS f64 is in register pairs, possibly split to stack
605 static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
606 CCValAssign::LocInfo &LocInfo,
607 CCState &State, bool CanFail) {
608 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
610 // Try to get the first register.
611 if (unsigned Reg = State.AllocateReg(RegList, 4))
612 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
614 // For the 2nd half of a v2f64, do not fail.
618 // Put the whole thing on the stack.
619 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
620 State.AllocateStack(8, 4),
625 // Try to get the second register.
626 if (unsigned Reg = State.AllocateReg(RegList, 4))
627 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
629 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
630 State.AllocateStack(4, 4),
635 static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
636 CCValAssign::LocInfo &LocInfo,
637 ISD::ArgFlagsTy &ArgFlags,
639 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
641 if (LocVT == MVT::v2f64 &&
642 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
644 return true; // we handled it
647 // AAPCS f64 is in aligned register pairs
648 static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
649 CCValAssign::LocInfo &LocInfo,
650 CCState &State, bool CanFail) {
651 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
652 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
654 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
656 // For the 2nd half of a v2f64, do not just fail.
660 // Put the whole thing on the stack.
661 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
662 State.AllocateStack(8, 8),
668 for (i = 0; i < 2; ++i)
669 if (HiRegList[i] == Reg)
672 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
673 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
678 static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
679 CCValAssign::LocInfo &LocInfo,
680 ISD::ArgFlagsTy &ArgFlags,
682 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
684 if (LocVT == MVT::v2f64 &&
685 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
687 return true; // we handled it
690 static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
691 CCValAssign::LocInfo &LocInfo, CCState &State) {
692 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 };
693 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 };
695 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2);
697 return false; // we didn't handle it
700 for (i = 0; i < 2; ++i)
701 if (HiRegList[i] == Reg)
704 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
705 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
710 static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
711 CCValAssign::LocInfo &LocInfo,
712 ISD::ArgFlagsTy &ArgFlags,
714 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
716 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
718 return true; // we handled it
721 static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
722 CCValAssign::LocInfo &LocInfo,
723 ISD::ArgFlagsTy &ArgFlags,
725 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
729 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
730 /// given CallingConvention value.
731 CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
733 bool isVarArg) const {
736 llvm_unreachable("Unsupported calling convention");
738 case CallingConv::Fast:
739 // Use target triple & subtarget features to do actual dispatch.
740 if (Subtarget->isAAPCS_ABI()) {
741 if (Subtarget->hasVFP2() &&
742 FloatABIType == FloatABI::Hard && !isVarArg)
743 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
745 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
747 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
748 case CallingConv::ARM_AAPCS_VFP:
749 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
750 case CallingConv::ARM_AAPCS:
751 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
752 case CallingConv::ARM_APCS:
753 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
757 /// LowerCallResult - Lower the result values of a call into the
758 /// appropriate copies out of appropriate physical registers.
760 ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
761 CallingConv::ID CallConv, bool isVarArg,
762 const SmallVectorImpl<ISD::InputArg> &Ins,
763 DebugLoc dl, SelectionDAG &DAG,
764 SmallVectorImpl<SDValue> &InVals) {
766 // Assign locations to each value returned by this call.
767 SmallVector<CCValAssign, 16> RVLocs;
768 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
769 RVLocs, *DAG.getContext());
770 CCInfo.AnalyzeCallResult(Ins,
771 CCAssignFnForNode(CallConv, /* Return*/ true,
774 // Copy all of the result registers out of their specified physreg.
775 for (unsigned i = 0; i != RVLocs.size(); ++i) {
776 CCValAssign VA = RVLocs[i];
779 if (VA.needsCustom()) {
780 // Handle f64 or half of a v2f64.
781 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
783 Chain = Lo.getValue(1);
784 InFlag = Lo.getValue(2);
785 VA = RVLocs[++i]; // skip ahead to next loc
786 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
788 Chain = Hi.getValue(1);
789 InFlag = Hi.getValue(2);
790 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
792 if (VA.getLocVT() == MVT::v2f64) {
793 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
794 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
795 DAG.getConstant(0, MVT::i32));
797 VA = RVLocs[++i]; // skip ahead to next loc
798 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
799 Chain = Lo.getValue(1);
800 InFlag = Lo.getValue(2);
801 VA = RVLocs[++i]; // skip ahead to next loc
802 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
803 Chain = Hi.getValue(1);
804 InFlag = Hi.getValue(2);
805 Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
806 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
807 DAG.getConstant(1, MVT::i32));
810 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
812 Chain = Val.getValue(1);
813 InFlag = Val.getValue(2);
816 switch (VA.getLocInfo()) {
817 default: llvm_unreachable("Unknown loc info!");
818 case CCValAssign::Full: break;
819 case CCValAssign::BCvt:
820 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val);
824 InVals.push_back(Val);
830 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
831 /// by "Src" to address "Dst" of size "Size". Alignment information is
832 /// specified by the specific parameter attribute. The copy will be passed as
833 /// a byval function parameter.
834 /// Sometimes what we are copying is the end of a larger object, the part that
835 /// does not fit in registers.
837 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
838 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
840 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
841 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
842 /*AlwaysInline=*/false, NULL, 0, NULL, 0);
845 /// LowerMemOpCallTo - Store the argument to the stack.
847 ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
848 SDValue StackPtr, SDValue Arg,
849 DebugLoc dl, SelectionDAG &DAG,
850 const CCValAssign &VA,
851 ISD::ArgFlagsTy Flags) {
852 unsigned LocMemOffset = VA.getLocMemOffset();
853 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
854 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
855 if (Flags.isByVal()) {
856 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
858 return DAG.getStore(Chain, dl, Arg, PtrOff,
859 PseudoSourceValue::getStack(), LocMemOffset);
862 void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
863 SDValue Chain, SDValue &Arg,
864 RegsToPassVector &RegsToPass,
865 CCValAssign &VA, CCValAssign &NextVA,
867 SmallVector<SDValue, 8> &MemOpChains,
868 ISD::ArgFlagsTy Flags) {
870 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
871 DAG.getVTList(MVT::i32, MVT::i32), Arg);
872 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
874 if (NextVA.isRegLoc())
875 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
877 assert(NextVA.isMemLoc());
878 if (StackPtr.getNode() == 0)
879 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
881 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
887 /// LowerCall - Lowering a call into a callseq_start <-
888 /// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
891 ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
892 CallingConv::ID CallConv, bool isVarArg,
894 const SmallVectorImpl<ISD::OutputArg> &Outs,
895 const SmallVectorImpl<ISD::InputArg> &Ins,
896 DebugLoc dl, SelectionDAG &DAG,
897 SmallVectorImpl<SDValue> &InVals) {
899 // Analyze operands of the call, assigning locations to each operand.
900 SmallVector<CCValAssign, 16> ArgLocs;
901 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
903 CCInfo.AnalyzeCallOperands(Outs,
904 CCAssignFnForNode(CallConv, /* Return*/ false,
907 // Get a count of how many bytes are to be pushed on the stack.
908 unsigned NumBytes = CCInfo.getNextStackOffset();
910 // Adjust the stack pointer for the new arguments...
911 // These operations are automatically eliminated by the prolog/epilog pass
912 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
914 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32);
916 RegsToPassVector RegsToPass;
917 SmallVector<SDValue, 8> MemOpChains;
919 // Walk the register/memloc assignments, inserting copies/loads. In the case
920 // of tail call optimization, arguments are handled later.
921 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
924 CCValAssign &VA = ArgLocs[i];
925 SDValue Arg = Outs[realArgIdx].Val;
926 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
928 // Promote the value if needed.
929 switch (VA.getLocInfo()) {
930 default: llvm_unreachable("Unknown loc info!");
931 case CCValAssign::Full: break;
932 case CCValAssign::SExt:
933 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
935 case CCValAssign::ZExt:
936 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
938 case CCValAssign::AExt:
939 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
941 case CCValAssign::BCvt:
942 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
946 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
947 if (VA.needsCustom()) {
948 if (VA.getLocVT() == MVT::v2f64) {
949 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
950 DAG.getConstant(0, MVT::i32));
951 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
952 DAG.getConstant(1, MVT::i32));
954 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
955 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
957 VA = ArgLocs[++i]; // skip ahead to next loc
959 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
960 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
962 assert(VA.isMemLoc());
963 if (StackPtr.getNode() == 0)
964 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
966 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
967 dl, DAG, VA, Flags));
970 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
971 StackPtr, MemOpChains, Flags);
973 } else if (VA.isRegLoc()) {
974 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
976 assert(VA.isMemLoc());
977 if (StackPtr.getNode() == 0)
978 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
980 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
981 dl, DAG, VA, Flags));
985 if (!MemOpChains.empty())
986 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
987 &MemOpChains[0], MemOpChains.size());
989 // Build a sequence of copy-to-reg nodes chained together with token chain
990 // and flag operands which copy the outgoing args into the appropriate regs.
992 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
993 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
994 RegsToPass[i].second, InFlag);
995 InFlag = Chain.getValue(1);
998 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
999 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1000 // node so that legalize doesn't hack it.
1001 bool isDirect = false;
1002 bool isARMFunc = false;
1003 bool isLocalARMFunc = false;
1004 MachineFunction &MF = DAG.getMachineFunction();
1005 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1006 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1007 GlobalValue *GV = G->getGlobal();
1009 bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1010 bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1011 getTargetMachine().getRelocationModel() != Reloc::Static;
1012 isARMFunc = !Subtarget->isThumb() || isStub;
1013 // ARM call to a local ARM function is predicable.
1014 isLocalARMFunc = !Subtarget->isThumb() && !isExt;
1015 // tBX takes a register source operand.
1016 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1017 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1018 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
1021 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1022 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1023 Callee = DAG.getLoad(getPointerTy(), dl,
1024 DAG.getEntryNode(), CPAddr,
1025 PseudoSourceValue::getConstantPool(), 0);
1026 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1027 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1028 getPointerTy(), Callee, PICLabel);
1030 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy());
1031 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1033 bool isStub = Subtarget->isTargetDarwin() &&
1034 getTargetMachine().getRelocationModel() != Reloc::Static;
1035 isARMFunc = !Subtarget->isThumb() || isStub;
1036 // tBX takes a register source operand.
1037 const char *Sym = S->getSymbol();
1038 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1039 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1040 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1041 Sym, ARMPCLabelIndex, 4);
1042 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1043 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1044 Callee = DAG.getLoad(getPointerTy(), dl,
1045 DAG.getEntryNode(), CPAddr,
1046 PseudoSourceValue::getConstantPool(), 0);
1047 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1048 Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1049 getPointerTy(), Callee, PICLabel);
1051 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1054 // FIXME: handle tail calls differently.
1056 if (Subtarget->isThumb()) {
1057 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1058 CallOpc = ARMISD::CALL_NOLINK;
1060 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1062 CallOpc = (isDirect || Subtarget->hasV5TOps())
1063 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1064 : ARMISD::CALL_NOLINK;
1066 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) {
1067 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK
1068 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag);
1069 InFlag = Chain.getValue(1);
1072 std::vector<SDValue> Ops;
1073 Ops.push_back(Chain);
1074 Ops.push_back(Callee);
1076 // Add argument registers to the end of the list so that they are known live
1078 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1079 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1080 RegsToPass[i].second.getValueType()));
1082 if (InFlag.getNode())
1083 Ops.push_back(InFlag);
1084 // Returns a chain and a flag for retval copy to use.
1085 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1086 &Ops[0], Ops.size());
1087 InFlag = Chain.getValue(1);
1089 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1090 DAG.getIntPtrConstant(0, true), InFlag);
1092 InFlag = Chain.getValue(1);
1094 // Handle result values, copying them out of physregs into vregs that we
1096 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1101 ARMTargetLowering::LowerReturn(SDValue Chain,
1102 CallingConv::ID CallConv, bool isVarArg,
1103 const SmallVectorImpl<ISD::OutputArg> &Outs,
1104 DebugLoc dl, SelectionDAG &DAG) {
1106 // CCValAssign - represent the assignment of the return value to a location.
1107 SmallVector<CCValAssign, 16> RVLocs;
1109 // CCState - Info about the registers and stack slots.
1110 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1113 // Analyze outgoing return values.
1114 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1117 // If this is the first return lowered for this function, add
1118 // the regs to the liveout set for the function.
1119 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1120 for (unsigned i = 0; i != RVLocs.size(); ++i)
1121 if (RVLocs[i].isRegLoc())
1122 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1127 // Copy the result values into the output registers.
1128 for (unsigned i = 0, realRVLocIdx = 0;
1130 ++i, ++realRVLocIdx) {
1131 CCValAssign &VA = RVLocs[i];
1132 assert(VA.isRegLoc() && "Can only return in registers!");
1134 SDValue Arg = Outs[realRVLocIdx].Val;
1136 switch (VA.getLocInfo()) {
1137 default: llvm_unreachable("Unknown loc info!");
1138 case CCValAssign::Full: break;
1139 case CCValAssign::BCvt:
1140 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg);
1144 if (VA.needsCustom()) {
1145 if (VA.getLocVT() == MVT::v2f64) {
1146 // Extract the first half and return it in two registers.
1147 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1148 DAG.getConstant(0, MVT::i32));
1149 SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1150 DAG.getVTList(MVT::i32, MVT::i32), Half);
1152 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1153 Flag = Chain.getValue(1);
1154 VA = RVLocs[++i]; // skip ahead to next loc
1155 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1156 HalfGPRs.getValue(1), Flag);
1157 Flag = Chain.getValue(1);
1158 VA = RVLocs[++i]; // skip ahead to next loc
1160 // Extract the 2nd half and fall through to handle it as an f64 value.
1161 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1162 DAG.getConstant(1, MVT::i32));
1164 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1166 SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1167 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1168 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1169 Flag = Chain.getValue(1);
1170 VA = RVLocs[++i]; // skip ahead to next loc
1171 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1174 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1176 // Guarantee that all emitted copies are
1177 // stuck together, avoiding something bad.
1178 Flag = Chain.getValue(1);
1183 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1185 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1190 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1191 // their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1192 // one of the above mentioned nodes. It has to be wrapped because otherwise
1193 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1194 // be used to form addressing mode. These wrapped nodes will be selected
1196 static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1197 EVT PtrVT = Op.getValueType();
1198 // FIXME there is no actual debug info here
1199 DebugLoc dl = Op.getDebugLoc();
1200 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1202 if (CP->isMachineConstantPoolEntry())
1203 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1204 CP->getAlignment());
1206 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1207 CP->getAlignment());
1208 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1211 SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1212 MachineFunction &MF = DAG.getMachineFunction();
1213 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1214 unsigned ARMPCLabelIndex = 0;
1215 DebugLoc DL = Op.getDebugLoc();
1216 EVT PtrVT = getPointerTy();
1217 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1218 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1220 if (RelocM == Reloc::Static) {
1221 CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1223 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1224 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1225 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
1226 ARMCP::CPBlockAddress,
1228 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1230 CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1231 SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1232 PseudoSourceValue::getConstantPool(), 0);
1233 if (RelocM == Reloc::Static)
1235 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1236 return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1239 // Lower ISD::GlobalTLSAddress using the "general dynamic" model
1241 ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1242 SelectionDAG &DAG) {
1243 DebugLoc dl = GA->getDebugLoc();
1244 EVT PtrVT = getPointerTy();
1245 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1246 MachineFunction &MF = DAG.getMachineFunction();
1247 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1248 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1249 ARMConstantPoolValue *CPV =
1250 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1251 ARMCP::CPValue, PCAdj, "tlsgd", true);
1252 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1253 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1254 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1255 PseudoSourceValue::getConstantPool(), 0);
1256 SDValue Chain = Argument.getValue(1);
1258 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1259 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1261 // call __tls_get_addr.
1264 Entry.Node = Argument;
1265 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
1266 Args.push_back(Entry);
1267 // FIXME: is there useful debug info available here?
1268 std::pair<SDValue, SDValue> CallResult =
1269 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
1270 false, false, false, false,
1271 0, CallingConv::C, false, /*isReturnValueUsed=*/true,
1272 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
1273 return CallResult.first;
1276 // Lower ISD::GlobalTLSAddress using the "initial exec" or
1277 // "local exec" model.
1279 ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
1280 SelectionDAG &DAG) {
1281 GlobalValue *GV = GA->getGlobal();
1282 DebugLoc dl = GA->getDebugLoc();
1284 SDValue Chain = DAG.getEntryNode();
1285 EVT PtrVT = getPointerTy();
1286 // Get the Thread Pointer
1287 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1289 if (GV->isDeclaration()) {
1290 MachineFunction &MF = DAG.getMachineFunction();
1291 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1292 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1293 // Initial exec model.
1294 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1295 ARMConstantPoolValue *CPV =
1296 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
1297 ARMCP::CPValue, PCAdj, "gottpoff", true);
1298 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1299 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1300 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1301 PseudoSourceValue::getConstantPool(), 0);
1302 Chain = Offset.getValue(1);
1304 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1305 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
1307 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1308 PseudoSourceValue::getConstantPool(), 0);
1311 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
1312 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1313 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
1314 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
1315 PseudoSourceValue::getConstantPool(), 0);
1318 // The address of the thread local variable is the add of the thread
1319 // pointer with the offset of the variable.
1320 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
1324 ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
1325 // TODO: implement the "local dynamic" model
1326 assert(Subtarget->isTargetELF() &&
1327 "TLS not implemented for non-ELF targets");
1328 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1329 // If the relocation model is PIC, use the "General Dynamic" TLS Model,
1330 // otherwise use the "Local Exec" TLS Model
1331 if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
1332 return LowerToTLSGeneralDynamicModel(GA, DAG);
1334 return LowerToTLSExecModels(GA, DAG);
1337 SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
1338 SelectionDAG &DAG) {
1339 EVT PtrVT = getPointerTy();
1340 DebugLoc dl = Op.getDebugLoc();
1341 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1342 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1343 if (RelocM == Reloc::PIC_) {
1344 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
1345 ARMConstantPoolValue *CPV =
1346 new ARMConstantPoolValue(GV, UseGOTOFF ? "GOTOFF" : "GOT");
1347 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1348 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1349 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
1351 PseudoSourceValue::getConstantPool(), 0);
1352 SDValue Chain = Result.getValue(1);
1353 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
1354 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
1356 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1357 PseudoSourceValue::getGOT(), 0);
1360 // If we have T2 ops, we can materialize the address directly via movt/movw
1361 // pair. This is always cheaper.
1362 if (Subtarget->useMovt()) {
1363 return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
1364 DAG.getTargetGlobalAddress(GV, PtrVT));
1366 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1367 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1368 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1369 PseudoSourceValue::getConstantPool(), 0);
1374 SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
1375 SelectionDAG &DAG) {
1376 MachineFunction &MF = DAG.getMachineFunction();
1377 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1378 unsigned ARMPCLabelIndex = 0;
1379 EVT PtrVT = getPointerTy();
1380 DebugLoc dl = Op.getDebugLoc();
1381 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1382 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1384 if (RelocM == Reloc::Static)
1385 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
1387 ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1388 unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
1389 ARMConstantPoolValue *CPV =
1390 new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
1391 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1393 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1395 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1396 PseudoSourceValue::getConstantPool(), 0);
1397 SDValue Chain = Result.getValue(1);
1399 if (RelocM == Reloc::PIC_) {
1400 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1401 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1404 if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
1405 Result = DAG.getLoad(PtrVT, dl, Chain, Result,
1406 PseudoSourceValue::getGOT(), 0);
1411 SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
1413 assert(Subtarget->isTargetELF() &&
1414 "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
1415 MachineFunction &MF = DAG.getMachineFunction();
1416 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1417 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1418 EVT PtrVT = getPointerTy();
1419 DebugLoc dl = Op.getDebugLoc();
1420 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1421 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
1422 "_GLOBAL_OFFSET_TABLE_",
1423 ARMPCLabelIndex, PCAdj);
1424 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1425 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1426 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1427 PseudoSourceValue::getConstantPool(), 0);
1428 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1429 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1433 ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
1434 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1435 DebugLoc dl = Op.getDebugLoc();
1437 default: return SDValue(); // Don't custom lower most intrinsics.
1438 case Intrinsic::arm_thread_pointer: {
1439 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1440 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
1442 case Intrinsic::eh_sjlj_lsda: {
1443 MachineFunction &MF = DAG.getMachineFunction();
1444 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1445 unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
1446 EVT PtrVT = getPointerTy();
1447 DebugLoc dl = Op.getDebugLoc();
1448 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1450 unsigned PCAdj = (RelocM != Reloc::PIC_)
1451 ? 0 : (Subtarget->isThumb() ? 4 : 8);
1452 ARMConstantPoolValue *CPV =
1453 new ARMConstantPoolValue(MF.getFunction(), ARMPCLabelIndex,
1454 ARMCP::CPLSDA, PCAdj);
1455 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1456 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1458 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
1459 PseudoSourceValue::getConstantPool(), 0);
1460 SDValue Chain = Result.getValue(1);
1462 if (RelocM == Reloc::PIC_) {
1463 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1464 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
1468 case Intrinsic::eh_sjlj_setjmp:
1469 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1));
1473 static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
1474 unsigned VarArgsFrameIndex) {
1475 // vastart just stores the address of the VarArgsFrameIndex slot into the
1476 // memory location argument.
1477 DebugLoc dl = Op.getDebugLoc();
1478 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1479 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1480 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1481 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1485 ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) {
1486 SDNode *Node = Op.getNode();
1487 DebugLoc dl = Node->getDebugLoc();
1488 EVT VT = Node->getValueType(0);
1489 SDValue Chain = Op.getOperand(0);
1490 SDValue Size = Op.getOperand(1);
1491 SDValue Align = Op.getOperand(2);
1493 // Chain the dynamic stack allocation so that it doesn't modify the stack
1494 // pointer when other instructions are using the stack.
1495 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
1497 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue();
1498 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment();
1499 if (AlignVal > StackAlign)
1500 // Do this now since selection pass cannot introduce new target
1501 // independent node.
1502 Align = DAG.getConstant(-(uint64_t)AlignVal, VT);
1504 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up
1505 // using a "add r, sp, r" instead. Negate the size now so we don't have to
1506 // do even more horrible hack later.
1507 MachineFunction &MF = DAG.getMachineFunction();
1508 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1509 if (AFI->isThumb1OnlyFunction()) {
1511 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size);
1513 uint32_t Val = C->getZExtValue();
1514 if (Val <= 508 && ((Val & 3) == 0))
1518 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size);
1521 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
1522 SDValue Ops1[] = { Chain, Size, Align };
1523 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3);
1524 Chain = Res.getValue(1);
1525 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1526 DAG.getIntPtrConstant(0, true), SDValue());
1527 SDValue Ops2[] = { Res, Chain };
1528 return DAG.getMergeValues(Ops2, 2, dl);
1532 ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
1533 SDValue &Root, SelectionDAG &DAG,
1535 MachineFunction &MF = DAG.getMachineFunction();
1536 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1538 TargetRegisterClass *RC;
1539 if (AFI->isThumb1OnlyFunction())
1540 RC = ARM::tGPRRegisterClass;
1542 RC = ARM::GPRRegisterClass;
1544 // Transform the arguments stored in physical registers into virtual ones.
1545 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1546 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1549 if (NextVA.isMemLoc()) {
1550 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
1551 MachineFrameInfo *MFI = MF.getFrameInfo();
1552 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
1555 // Create load node to retrieve arguments from the stack.
1556 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1557 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
1558 PseudoSourceValue::getFixedStack(FI), 0);
1560 Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
1561 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
1564 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
1568 ARMTargetLowering::LowerFormalArguments(SDValue Chain,
1569 CallingConv::ID CallConv, bool isVarArg,
1570 const SmallVectorImpl<ISD::InputArg>
1572 DebugLoc dl, SelectionDAG &DAG,
1573 SmallVectorImpl<SDValue> &InVals) {
1575 MachineFunction &MF = DAG.getMachineFunction();
1576 MachineFrameInfo *MFI = MF.getFrameInfo();
1578 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1580 // Assign locations to all of the incoming arguments.
1581 SmallVector<CCValAssign, 16> ArgLocs;
1582 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1584 CCInfo.AnalyzeFormalArguments(Ins,
1585 CCAssignFnForNode(CallConv, /* Return*/ false,
1588 SmallVector<SDValue, 16> ArgValues;
1590 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1591 CCValAssign &VA = ArgLocs[i];
1593 // Arguments stored in registers.
1594 if (VA.isRegLoc()) {
1595 EVT RegVT = VA.getLocVT();
1598 if (VA.needsCustom()) {
1599 // f64 and vector types are split up into multiple registers or
1600 // combinations of registers and stack slots.
1603 if (VA.getLocVT() == MVT::v2f64) {
1604 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
1606 VA = ArgLocs[++i]; // skip ahead to next loc
1607 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
1609 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1610 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1611 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
1612 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
1613 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
1615 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
1618 TargetRegisterClass *RC;
1620 if (RegVT == MVT::f32)
1621 RC = ARM::SPRRegisterClass;
1622 else if (RegVT == MVT::f64)
1623 RC = ARM::DPRRegisterClass;
1624 else if (RegVT == MVT::v2f64)
1625 RC = ARM::QPRRegisterClass;
1626 else if (RegVT == MVT::i32)
1627 RC = (AFI->isThumb1OnlyFunction() ?
1628 ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
1630 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
1632 // Transform the arguments in physical registers into virtual ones.
1633 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1634 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1637 // If this is an 8 or 16-bit value, it is really passed promoted
1638 // to 32 bits. Insert an assert[sz]ext to capture this, then
1639 // truncate to the right size.
1640 switch (VA.getLocInfo()) {
1641 default: llvm_unreachable("Unknown loc info!");
1642 case CCValAssign::Full: break;
1643 case CCValAssign::BCvt:
1644 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1646 case CCValAssign::SExt:
1647 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1648 DAG.getValueType(VA.getValVT()));
1649 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1651 case CCValAssign::ZExt:
1652 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1653 DAG.getValueType(VA.getValVT()));
1654 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1658 InVals.push_back(ArgValue);
1660 } else { // VA.isRegLoc()
1663 assert(VA.isMemLoc());
1664 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
1666 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
1667 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1670 // Create load nodes to retrieve arguments from the stack.
1671 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1672 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1673 PseudoSourceValue::getFixedStack(FI), 0));
1679 static const unsigned GPRArgRegs[] = {
1680 ARM::R0, ARM::R1, ARM::R2, ARM::R3
1683 unsigned NumGPRs = CCInfo.getFirstUnallocated
1684 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
1686 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
1687 unsigned VARegSize = (4 - NumGPRs) * 4;
1688 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
1689 unsigned ArgOffset = CCInfo.getNextStackOffset();
1690 if (VARegSaveSize) {
1691 // If this function is vararg, store any remaining integer argument regs
1692 // to their spots on the stack so that they may be loaded by deferencing
1693 // the result of va_next.
1694 AFI->setVarArgsRegSaveSize(VARegSaveSize);
1695 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
1696 VARegSaveSize - VARegSize,
1698 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
1700 SmallVector<SDValue, 4> MemOps;
1701 for (; NumGPRs < 4; ++NumGPRs) {
1702 TargetRegisterClass *RC;
1703 if (AFI->isThumb1OnlyFunction())
1704 RC = ARM::tGPRRegisterClass;
1706 RC = ARM::GPRRegisterClass;
1708 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
1709 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1710 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1711 PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0);
1712 MemOps.push_back(Store);
1713 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1714 DAG.getConstant(4, getPointerTy()));
1716 if (!MemOps.empty())
1717 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1718 &MemOps[0], MemOps.size());
1720 // This will point to the next argument passed via stack.
1721 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
1727 /// isFloatingPointZero - Return true if this is +0.0.
1728 static bool isFloatingPointZero(SDValue Op) {
1729 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
1730 return CFP->getValueAPF().isPosZero();
1731 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
1732 // Maybe this has already been legalized into the constant pool?
1733 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
1734 SDValue WrapperOp = Op.getOperand(1).getOperand(0);
1735 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
1736 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
1737 return CFP->getValueAPF().isPosZero();
1743 /// Returns appropriate ARM CMP (cmp) and corresponding condition code for
1744 /// the given operands.
1746 ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
1747 SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
1748 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1749 unsigned C = RHSC->getZExtValue();
1750 if (!isLegalICmpImmediate(C)) {
1751 // Constant does not fit, try adjusting it by one?
1756 if (isLegalICmpImmediate(C-1)) {
1757 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1758 RHS = DAG.getConstant(C-1, MVT::i32);
1763 if (C > 0 && isLegalICmpImmediate(C-1)) {
1764 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1765 RHS = DAG.getConstant(C-1, MVT::i32);
1770 if (isLegalICmpImmediate(C+1)) {
1771 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1772 RHS = DAG.getConstant(C+1, MVT::i32);
1777 if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
1778 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1779 RHS = DAG.getConstant(C+1, MVT::i32);
1786 ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
1787 ARMISD::NodeType CompareType;
1790 CompareType = ARMISD::CMP;
1795 CompareType = ARMISD::CMPZ;
1798 ARMCC = DAG.getConstant(CondCode, MVT::i32);
1799 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS);
1802 /// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
1803 static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
1806 if (!isFloatingPointZero(RHS))
1807 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS);
1809 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS);
1810 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
1813 SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
1814 EVT VT = Op.getValueType();
1815 SDValue LHS = Op.getOperand(0);
1816 SDValue RHS = Op.getOperand(1);
1817 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1818 SDValue TrueVal = Op.getOperand(2);
1819 SDValue FalseVal = Op.getOperand(3);
1820 DebugLoc dl = Op.getDebugLoc();
1822 if (LHS.getValueType() == MVT::i32) {
1824 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1825 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1826 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
1829 ARMCC::CondCodes CondCode, CondCode2;
1830 FPCCToARMCC(CC, CondCode, CondCode2);
1832 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1833 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1834 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1835 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
1837 if (CondCode2 != ARMCC::AL) {
1838 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32);
1839 // FIXME: Needs another CMP because flag can have but one use.
1840 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
1841 Result = DAG.getNode(ARMISD::CMOV, dl, VT,
1842 Result, TrueVal, ARMCC2, CCR, Cmp2);
1847 SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
1848 SDValue Chain = Op.getOperand(0);
1849 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1850 SDValue LHS = Op.getOperand(2);
1851 SDValue RHS = Op.getOperand(3);
1852 SDValue Dest = Op.getOperand(4);
1853 DebugLoc dl = Op.getDebugLoc();
1855 if (LHS.getValueType() == MVT::i32) {
1857 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1858 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
1859 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
1860 Chain, Dest, ARMCC, CCR,Cmp);
1863 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
1864 ARMCC::CondCodes CondCode, CondCode2;
1865 FPCCToARMCC(CC, CondCode, CondCode2);
1867 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
1868 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
1869 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1870 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag);
1871 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp };
1872 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1873 if (CondCode2 != ARMCC::AL) {
1874 ARMCC = DAG.getConstant(CondCode2, MVT::i32);
1875 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) };
1876 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
1881 SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) {
1882 SDValue Chain = Op.getOperand(0);
1883 SDValue Table = Op.getOperand(1);
1884 SDValue Index = Op.getOperand(2);
1885 DebugLoc dl = Op.getDebugLoc();
1887 EVT PTy = getPointerTy();
1888 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1889 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
1890 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
1891 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
1892 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
1893 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
1894 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
1895 if (Subtarget->isThumb2()) {
1896 // Thumb2 uses a two-level jump. That is, it jumps into the jump table
1897 // which does another jump to the destination. This also makes it easier
1898 // to translate it to TBB / TBH later.
1899 // FIXME: This might not work if the function is extremely large.
1900 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
1901 Addr, Op.getOperand(2), JTI, UId);
1903 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1904 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
1905 PseudoSourceValue::getJumpTable(), 0);
1906 Chain = Addr.getValue(1);
1907 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
1908 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1910 Addr = DAG.getLoad(PTy, dl, Chain, Addr,
1911 PseudoSourceValue::getJumpTable(), 0);
1912 Chain = Addr.getValue(1);
1913 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
1917 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
1918 DebugLoc dl = Op.getDebugLoc();
1920 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI;
1921 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
1922 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
1925 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
1926 EVT VT = Op.getValueType();
1927 DebugLoc dl = Op.getDebugLoc();
1929 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF;
1931 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0));
1932 return DAG.getNode(Opc, dl, VT, Op);
1935 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
1936 // Implement fcopysign with a fabs and a conditional fneg.
1937 SDValue Tmp0 = Op.getOperand(0);
1938 SDValue Tmp1 = Op.getOperand(1);
1939 DebugLoc dl = Op.getDebugLoc();
1940 EVT VT = Op.getValueType();
1941 EVT SrcVT = Tmp1.getValueType();
1942 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0);
1943 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl);
1944 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32);
1945 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
1946 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp);
1949 SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
1950 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
1951 MFI->setFrameAddressIsTaken(true);
1952 EVT VT = Op.getValueType();
1953 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
1954 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1955 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
1956 ? ARM::R7 : ARM::R11;
1957 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
1959 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
1964 ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
1966 SDValue Dst, SDValue Src,
1967 SDValue Size, unsigned Align,
1969 const Value *DstSV, uint64_t DstSVOff,
1970 const Value *SrcSV, uint64_t SrcSVOff){
1971 // Do repeated 4-byte loads and stores. To be improved.
1972 // This requires 4-byte alignment.
1973 if ((Align & 3) != 0)
1975 // This requires the copy size to be a constant, preferrably
1976 // within a subtarget-specific limit.
1977 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
1980 uint64_t SizeVal = ConstantSize->getZExtValue();
1981 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
1984 unsigned BytesLeft = SizeVal & 3;
1985 unsigned NumMemOps = SizeVal >> 2;
1986 unsigned EmittedNumMemOps = 0;
1988 unsigned VTSize = 4;
1990 const unsigned MAX_LOADS_IN_LDM = 6;
1991 SDValue TFOps[MAX_LOADS_IN_LDM];
1992 SDValue Loads[MAX_LOADS_IN_LDM];
1993 uint64_t SrcOff = 0, DstOff = 0;
1995 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
1996 // same number of stores. The loads and stores will get combined into
1997 // ldm/stm later on.
1998 while (EmittedNumMemOps < NumMemOps) {
2000 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2001 Loads[i] = DAG.getLoad(VT, dl, Chain,
2002 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2003 DAG.getConstant(SrcOff, MVT::i32)),
2004 SrcSV, SrcSVOff + SrcOff);
2005 TFOps[i] = Loads[i].getValue(1);
2008 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2011 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
2012 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2013 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2014 DAG.getConstant(DstOff, MVT::i32)),
2015 DstSV, DstSVOff + DstOff);
2018 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2020 EmittedNumMemOps += i;
2026 // Issue loads / stores for the trailing (1 - 3) bytes.
2027 unsigned BytesLeftSave = BytesLeft;
2030 if (BytesLeft >= 2) {
2038 Loads[i] = DAG.getLoad(VT, dl, Chain,
2039 DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
2040 DAG.getConstant(SrcOff, MVT::i32)),
2041 SrcSV, SrcSVOff + SrcOff);
2042 TFOps[i] = Loads[i].getValue(1);
2045 BytesLeft -= VTSize;
2047 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2050 BytesLeft = BytesLeftSave;
2052 if (BytesLeft >= 2) {
2060 TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
2061 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
2062 DAG.getConstant(DstOff, MVT::i32)),
2063 DstSV, DstSVOff + DstOff);
2066 BytesLeft -= VTSize;
2068 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i);
2071 static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) {
2072 SDValue Op = N->getOperand(0);
2073 DebugLoc dl = N->getDebugLoc();
2074 if (N->getValueType(0) == MVT::f64) {
2075 // Turn i64->f64 into VMOVDRR.
2076 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2077 DAG.getConstant(0, MVT::i32));
2078 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
2079 DAG.getConstant(1, MVT::i32));
2080 return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
2083 // Turn f64->i64 into VMOVRRD.
2084 SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
2085 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
2087 // Merge the pieces into a single i64 value.
2088 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
2091 /// getZeroVector - Returns a vector of specified type with all zero elements.
2093 static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2094 assert(VT.isVector() && "Expected a vector type");
2096 // Zero vectors are used to represent vector negation and in those cases
2097 // will be implemented with the NEON VNEG instruction. However, VNEG does
2098 // not support i64 elements, so sometimes the zero vectors will need to be
2099 // explicitly constructed. For those cases, and potentially other uses in
2100 // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
2101 // to their dest type. This ensures they get CSE'd.
2103 SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
2104 SmallVector<SDValue, 8> Ops;
2107 if (VT.getSizeInBits() == 64) {
2108 Ops.assign(8, Cst); TVT = MVT::v8i8;
2110 Ops.assign(16, Cst); TVT = MVT::v16i8;
2112 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2114 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2117 /// getOnesVector - Returns a vector of specified type with all bits set.
2119 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2120 assert(VT.isVector() && "Expected a vector type");
2122 // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
2123 // dest type. This ensures they get CSE'd.
2125 SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
2126 SmallVector<SDValue, 8> Ops;
2129 if (VT.getSizeInBits() == 64) {
2130 Ops.assign(8, Cst); TVT = MVT::v8i8;
2132 Ops.assign(16, Cst); TVT = MVT::v16i8;
2134 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
2136 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2139 /// LowerShiftRightParts - Lower SRA_PARTS, which returns two
2140 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2141 SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
2142 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2143 EVT VT = Op.getValueType();
2144 unsigned VTBits = VT.getSizeInBits();
2145 DebugLoc dl = Op.getDebugLoc();
2146 SDValue ShOpLo = Op.getOperand(0);
2147 SDValue ShOpHi = Op.getOperand(1);
2148 SDValue ShAmt = Op.getOperand(2);
2150 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
2152 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
2154 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2155 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2156 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
2157 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2158 DAG.getConstant(VTBits, MVT::i32));
2159 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
2160 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2161 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
2163 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2164 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2166 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
2167 SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
2170 SDValue Ops[2] = { Lo, Hi };
2171 return DAG.getMergeValues(Ops, 2, dl);
2174 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
2175 /// i32 values and take a 2 x i32 value to shift plus a shift amount.
2176 SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
2177 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
2178 EVT VT = Op.getValueType();
2179 unsigned VTBits = VT.getSizeInBits();
2180 DebugLoc dl = Op.getDebugLoc();
2181 SDValue ShOpLo = Op.getOperand(0);
2182 SDValue ShOpHi = Op.getOperand(1);
2183 SDValue ShAmt = Op.getOperand(2);
2186 assert(Op.getOpcode() == ISD::SHL_PARTS);
2187 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
2188 DAG.getConstant(VTBits, MVT::i32), ShAmt);
2189 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
2190 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
2191 DAG.getConstant(VTBits, MVT::i32));
2192 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
2193 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
2195 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2196 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2197 SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
2199 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
2200 SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
2203 SDValue Ops[2] = { Lo, Hi };
2204 return DAG.getMergeValues(Ops, 2, dl);
2207 static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
2208 const ARMSubtarget *ST) {
2209 EVT VT = N->getValueType(0);
2210 DebugLoc dl = N->getDebugLoc();
2212 // Lower vector shifts on NEON to use VSHL.
2213 if (VT.isVector()) {
2214 assert(ST->hasNEON() && "unexpected vector shift");
2216 // Left shifts translate directly to the vshiftu intrinsic.
2217 if (N->getOpcode() == ISD::SHL)
2218 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2219 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
2220 N->getOperand(0), N->getOperand(1));
2222 assert((N->getOpcode() == ISD::SRA ||
2223 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
2225 // NEON uses the same intrinsics for both left and right shifts. For
2226 // right shifts, the shift amounts are negative, so negate the vector of
2228 EVT ShiftVT = N->getOperand(1).getValueType();
2229 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
2230 getZeroVector(ShiftVT, DAG, dl),
2232 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
2233 Intrinsic::arm_neon_vshifts :
2234 Intrinsic::arm_neon_vshiftu);
2235 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
2236 DAG.getConstant(vshiftInt, MVT::i32),
2237 N->getOperand(0), NegatedCount);
2240 // We can get here for a node like i32 = ISD::SHL i32, i64
2244 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
2245 "Unknown shift to lower!");
2247 // We only lower SRA, SRL of 1 here, all others use generic lowering.
2248 if (!isa<ConstantSDNode>(N->getOperand(1)) ||
2249 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
2252 // If we are in thumb mode, we don't have RRX.
2253 if (ST->isThumb1Only()) return SDValue();
2255 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
2256 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2257 DAG.getConstant(0, MVT::i32));
2258 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
2259 DAG.getConstant(1, MVT::i32));
2261 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
2262 // captures the result into a carry flag.
2263 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
2264 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1);
2266 // The low part is an ARMISD::RRX operand, which shifts the carry in.
2267 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
2269 // Merge the pieces into a single i64 value.
2270 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
2273 static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
2274 SDValue TmpOp0, TmpOp1;
2275 bool Invert = false;
2279 SDValue Op0 = Op.getOperand(0);
2280 SDValue Op1 = Op.getOperand(1);
2281 SDValue CC = Op.getOperand(2);
2282 EVT VT = Op.getValueType();
2283 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
2284 DebugLoc dl = Op.getDebugLoc();
2286 if (Op.getOperand(1).getValueType().isFloatingPoint()) {
2287 switch (SetCCOpcode) {
2288 default: llvm_unreachable("Illegal FP comparison"); break;
2290 case ISD::SETNE: Invert = true; // Fallthrough
2292 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2294 case ISD::SETLT: Swap = true; // Fallthrough
2296 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2298 case ISD::SETLE: Swap = true; // Fallthrough
2300 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2301 case ISD::SETUGE: Swap = true; // Fallthrough
2302 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
2303 case ISD::SETUGT: Swap = true; // Fallthrough
2304 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
2305 case ISD::SETUEQ: Invert = true; // Fallthrough
2307 // Expand this to (OLT | OGT).
2311 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2312 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
2314 case ISD::SETUO: Invert = true; // Fallthrough
2316 // Expand this to (OLT | OGE).
2320 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
2321 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
2325 // Integer comparisons.
2326 switch (SetCCOpcode) {
2327 default: llvm_unreachable("Illegal integer comparison"); break;
2328 case ISD::SETNE: Invert = true;
2329 case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
2330 case ISD::SETLT: Swap = true;
2331 case ISD::SETGT: Opc = ARMISD::VCGT; break;
2332 case ISD::SETLE: Swap = true;
2333 case ISD::SETGE: Opc = ARMISD::VCGE; break;
2334 case ISD::SETULT: Swap = true;
2335 case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
2336 case ISD::SETULE: Swap = true;
2337 case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
2340 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
2341 if (Opc == ARMISD::VCEQ) {
2344 if (ISD::isBuildVectorAllZeros(Op1.getNode()))
2346 else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
2349 // Ignore bitconvert.
2350 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT)
2351 AndOp = AndOp.getOperand(0);
2353 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
2355 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0));
2356 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1));
2363 std::swap(Op0, Op1);
2365 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
2368 Result = DAG.getNOT(dl, Result, VT);
2373 /// isVMOVSplat - Check if the specified splat value corresponds to an immediate
2374 /// VMOV instruction, and if so, return the constant being splatted.
2375 static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef,
2376 unsigned SplatBitSize, SelectionDAG &DAG) {
2377 switch (SplatBitSize) {
2379 // Any 1-byte value is OK.
2380 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
2381 return DAG.getTargetConstant(SplatBits, MVT::i8);
2384 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
2385 if ((SplatBits & ~0xff) == 0 ||
2386 (SplatBits & ~0xff00) == 0)
2387 return DAG.getTargetConstant(SplatBits, MVT::i16);
2391 // NEON's 32-bit VMOV supports splat values where:
2392 // * only one byte is nonzero, or
2393 // * the least significant byte is 0xff and the second byte is nonzero, or
2394 // * the least significant 2 bytes are 0xff and the third is nonzero.
2395 if ((SplatBits & ~0xff) == 0 ||
2396 (SplatBits & ~0xff00) == 0 ||
2397 (SplatBits & ~0xff0000) == 0 ||
2398 (SplatBits & ~0xff000000) == 0)
2399 return DAG.getTargetConstant(SplatBits, MVT::i32);
2401 if ((SplatBits & ~0xffff) == 0 &&
2402 ((SplatBits | SplatUndef) & 0xff) == 0xff)
2403 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32);
2405 if ((SplatBits & ~0xffffff) == 0 &&
2406 ((SplatBits | SplatUndef) & 0xffff) == 0xffff)
2407 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32);
2409 // Note: there are a few 32-bit splat values (specifically: 00ffff00,
2410 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
2411 // VMOV.I32. A (very) minor optimization would be to replicate the value
2412 // and fall through here to test for a valid 64-bit splat. But, then the
2413 // caller would also need to check and handle the change in size.
2417 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
2418 uint64_t BitMask = 0xff;
2420 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
2421 if (((SplatBits | SplatUndef) & BitMask) == BitMask)
2423 else if ((SplatBits & BitMask) != 0)
2427 return DAG.getTargetConstant(Val, MVT::i64);
2431 llvm_unreachable("unexpected size for isVMOVSplat");
2438 /// getVMOVImm - If this is a build_vector of constants which can be
2439 /// formed by using a VMOV instruction of the specified element size,
2440 /// return the constant being splatted. The ByteSize field indicates the
2441 /// number of bytes of each element [1248].
2442 SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
2443 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N);
2444 APInt SplatBits, SplatUndef;
2445 unsigned SplatBitSize;
2447 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
2448 HasAnyUndefs, ByteSize * 8))
2451 if (SplatBitSize > ByteSize * 8)
2454 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
2458 static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
2459 bool &ReverseVEXT, unsigned &Imm) {
2460 unsigned NumElts = VT.getVectorNumElements();
2461 ReverseVEXT = false;
2464 // If this is a VEXT shuffle, the immediate value is the index of the first
2465 // element. The other shuffle indices must be the successive elements after
2467 unsigned ExpectedElt = Imm;
2468 for (unsigned i = 1; i < NumElts; ++i) {
2469 // Increment the expected index. If it wraps around, it may still be
2470 // a VEXT but the source vectors must be swapped.
2472 if (ExpectedElt == NumElts * 2) {
2477 if (ExpectedElt != static_cast<unsigned>(M[i]))
2481 // Adjust the index value if the source operands will be swapped.
2488 /// isVREVMask - Check if a vector shuffle corresponds to a VREV
2489 /// instruction with the specified blocksize. (The order of the elements
2490 /// within each block of the vector is reversed.)
2491 static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
2492 unsigned BlockSize) {
2493 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
2494 "Only possible block sizes for VREV are: 16, 32, 64");
2496 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2500 unsigned NumElts = VT.getVectorNumElements();
2501 unsigned BlockElts = M[0] + 1;
2503 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
2506 for (unsigned i = 0; i < NumElts; ++i) {
2507 if ((unsigned) M[i] !=
2508 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
2515 static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
2516 unsigned &WhichResult) {
2517 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2521 unsigned NumElts = VT.getVectorNumElements();
2522 WhichResult = (M[0] == 0 ? 0 : 1);
2523 for (unsigned i = 0; i < NumElts; i += 2) {
2524 if ((unsigned) M[i] != i + WhichResult ||
2525 (unsigned) M[i+1] != i + NumElts + WhichResult)
2531 static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
2532 unsigned &WhichResult) {
2533 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2537 unsigned NumElts = VT.getVectorNumElements();
2538 WhichResult = (M[0] == 0 ? 0 : 1);
2539 for (unsigned i = 0; i != NumElts; ++i) {
2540 if ((unsigned) M[i] != 2 * i + WhichResult)
2544 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2545 if (VT.is64BitVector() && EltSz == 32)
2551 static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
2552 unsigned &WhichResult) {
2553 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
2557 unsigned NumElts = VT.getVectorNumElements();
2558 WhichResult = (M[0] == 0 ? 0 : 1);
2559 unsigned Idx = WhichResult * NumElts / 2;
2560 for (unsigned i = 0; i != NumElts; i += 2) {
2561 if ((unsigned) M[i] != Idx ||
2562 (unsigned) M[i+1] != Idx + NumElts)
2567 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
2568 if (VT.is64BitVector() && EltSz == 32)
2574 static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2575 // Canonicalize all-zeros and all-ones vectors.
2576 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
2577 if (ConstVal->isNullValue())
2578 return getZeroVector(VT, DAG, dl);
2579 if (ConstVal->isAllOnesValue())
2580 return getOnesVector(VT, DAG, dl);
2583 if (VT.is64BitVector()) {
2584 switch (Val.getValueType().getSizeInBits()) {
2585 case 8: CanonicalVT = MVT::v8i8; break;
2586 case 16: CanonicalVT = MVT::v4i16; break;
2587 case 32: CanonicalVT = MVT::v2i32; break;
2588 case 64: CanonicalVT = MVT::v1i64; break;
2589 default: llvm_unreachable("unexpected splat element type"); break;
2592 assert(VT.is128BitVector() && "unknown splat vector size");
2593 switch (Val.getValueType().getSizeInBits()) {
2594 case 8: CanonicalVT = MVT::v16i8; break;
2595 case 16: CanonicalVT = MVT::v8i16; break;
2596 case 32: CanonicalVT = MVT::v4i32; break;
2597 case 64: CanonicalVT = MVT::v2i64; break;
2598 default: llvm_unreachable("unexpected splat element type"); break;
2602 // Build a canonical splat for this value.
2603 SmallVector<SDValue, 8> Ops;
2604 Ops.assign(CanonicalVT.getVectorNumElements(), Val);
2605 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0],
2607 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res);
2610 // If this is a case we can't handle, return null and let the default
2611 // expansion code take care of it.
2612 static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
2613 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2614 DebugLoc dl = Op.getDebugLoc();
2615 EVT VT = Op.getValueType();
2617 APInt SplatBits, SplatUndef;
2618 unsigned SplatBitSize;
2620 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
2621 if (SplatBitSize <= 64) {
2622 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(),
2623 SplatUndef.getZExtValue(), SplatBitSize, DAG);
2625 return BuildSplat(Val, VT, DAG, dl);
2629 // If there are only 2 elements in a 128-bit vector, insert them into an
2630 // undef vector. This handles the common case for 128-bit vector argument
2631 // passing, where the insertions should be translated to subreg accesses
2632 // with no real instructions.
2633 if (VT.is128BitVector() && Op.getNumOperands() == 2) {
2634 SDValue Val = DAG.getUNDEF(VT);
2635 SDValue Op0 = Op.getOperand(0);
2636 SDValue Op1 = Op.getOperand(1);
2637 if (Op0.getOpcode() != ISD::UNDEF)
2638 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0,
2639 DAG.getIntPtrConstant(0));
2640 if (Op1.getOpcode() != ISD::UNDEF)
2641 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1,
2642 DAG.getIntPtrConstant(1));
2649 /// isShuffleMaskLegal - Targets can use this to indicate that they only
2650 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
2651 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
2652 /// are assumed to be legal.
2654 ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
2656 if (VT.getVectorNumElements() == 4 &&
2657 (VT.is128BitVector() || VT.is64BitVector())) {
2658 unsigned PFIndexes[4];
2659 for (unsigned i = 0; i != 4; ++i) {
2663 PFIndexes[i] = M[i];
2666 // Compute the index in the perfect shuffle table.
2667 unsigned PFTableIndex =
2668 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2669 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2670 unsigned Cost = (PFEntry >> 30);
2677 unsigned Imm, WhichResult;
2679 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
2680 isVREVMask(M, VT, 64) ||
2681 isVREVMask(M, VT, 32) ||
2682 isVREVMask(M, VT, 16) ||
2683 isVEXTMask(M, VT, ReverseVEXT, Imm) ||
2684 isVTRNMask(M, VT, WhichResult) ||
2685 isVUZPMask(M, VT, WhichResult) ||
2686 isVZIPMask(M, VT, WhichResult));
2689 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2690 /// the specified operations to build the shuffle.
2691 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
2692 SDValue RHS, SelectionDAG &DAG,
2694 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2695 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2696 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2699 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2708 OP_VUZPL, // VUZP, left result
2709 OP_VUZPR, // VUZP, right result
2710 OP_VZIPL, // VZIP, left result
2711 OP_VZIPR, // VZIP, right result
2712 OP_VTRNL, // VTRN, left result
2713 OP_VTRNR // VTRN, right result
2716 if (OpNum == OP_COPY) {
2717 if (LHSID == (1*9+2)*9+3) return LHS;
2718 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2722 SDValue OpLHS, OpRHS;
2723 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
2724 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
2725 EVT VT = OpLHS.getValueType();
2728 default: llvm_unreachable("Unknown shuffle opcode!");
2730 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
2735 return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
2736 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
2740 return DAG.getNode(ARMISD::VEXT, dl, VT,
2742 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
2745 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2746 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
2749 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2750 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
2753 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2754 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
2758 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2759 SDValue V1 = Op.getOperand(0);
2760 SDValue V2 = Op.getOperand(1);
2761 DebugLoc dl = Op.getDebugLoc();
2762 EVT VT = Op.getValueType();
2763 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
2764 SmallVector<int, 8> ShuffleMask;
2766 // Convert shuffles that are directly supported on NEON to target-specific
2767 // DAG nodes, instead of keeping them as shuffles and matching them again
2768 // during code selection. This is more efficient and avoids the possibility
2769 // of inconsistencies between legalization and selection.
2770 // FIXME: floating-point vectors should be canonicalized to integer vectors
2771 // of the same time so that they get CSEd properly.
2772 SVN->getMask(ShuffleMask);
2774 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
2775 int Lane = SVN->getSplatIndex();
2776 // If this is undef splat, generate it via "just" vdup, if possible.
2777 if (Lane == -1) Lane = 0;
2779 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
2780 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
2782 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
2783 DAG.getConstant(Lane, MVT::i32));
2788 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
2791 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
2792 DAG.getConstant(Imm, MVT::i32));
2795 if (isVREVMask(ShuffleMask, VT, 64))
2796 return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
2797 if (isVREVMask(ShuffleMask, VT, 32))
2798 return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
2799 if (isVREVMask(ShuffleMask, VT, 16))
2800 return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
2802 // Check for Neon shuffles that modify both input vectors in place.
2803 // If both results are used, i.e., if there are two shuffles with the same
2804 // source operands and with masks corresponding to both results of one of
2805 // these operations, DAG memoization will ensure that a single node is
2806 // used for both shuffles.
2807 unsigned WhichResult;
2808 if (isVTRNMask(ShuffleMask, VT, WhichResult))
2809 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
2810 V1, V2).getValue(WhichResult);
2811 if (isVUZPMask(ShuffleMask, VT, WhichResult))
2812 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
2813 V1, V2).getValue(WhichResult);
2814 if (isVZIPMask(ShuffleMask, VT, WhichResult))
2815 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
2816 V1, V2).getValue(WhichResult);
2818 // If the shuffle is not directly supported and it has 4 elements, use
2819 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2820 if (VT.getVectorNumElements() == 4 &&
2821 (VT.is128BitVector() || VT.is64BitVector())) {
2822 unsigned PFIndexes[4];
2823 for (unsigned i = 0; i != 4; ++i) {
2824 if (ShuffleMask[i] < 0)
2827 PFIndexes[i] = ShuffleMask[i];
2830 // Compute the index in the perfect shuffle table.
2831 unsigned PFTableIndex =
2832 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2834 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2835 unsigned Cost = (PFEntry >> 30);
2838 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
2844 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2845 EVT VT = Op.getValueType();
2846 DebugLoc dl = Op.getDebugLoc();
2847 SDValue Vec = Op.getOperand(0);
2848 SDValue Lane = Op.getOperand(1);
2849 assert(VT == MVT::i32 &&
2850 Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
2851 "unexpected type for custom-lowering vector extract");
2852 return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
2855 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2856 // The only time a CONCAT_VECTORS operation can have legal types is when
2857 // two 64-bit vectors are concatenated to a 128-bit vector.
2858 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
2859 "unexpected CONCAT_VECTORS");
2860 DebugLoc dl = Op.getDebugLoc();
2861 SDValue Val = DAG.getUNDEF(MVT::v2f64);
2862 SDValue Op0 = Op.getOperand(0);
2863 SDValue Op1 = Op.getOperand(1);
2864 if (Op0.getOpcode() != ISD::UNDEF)
2865 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2866 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0),
2867 DAG.getIntPtrConstant(0));
2868 if (Op1.getOpcode() != ISD::UNDEF)
2869 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
2870 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1),
2871 DAG.getIntPtrConstant(1));
2872 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val);
2875 SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
2876 switch (Op.getOpcode()) {
2877 default: llvm_unreachable("Don't know how to custom lower this!");
2878 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2879 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
2880 case ISD::GlobalAddress:
2881 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
2882 LowerGlobalAddressELF(Op, DAG);
2883 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2884 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
2885 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
2886 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2887 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2888 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
2889 case ISD::SINT_TO_FP:
2890 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
2891 case ISD::FP_TO_SINT:
2892 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
2893 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
2894 case ISD::RETURNADDR: break;
2895 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2896 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
2897 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2898 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
2901 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
2902 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
2903 case ISD::SRL_PARTS:
2904 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
2905 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
2906 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2907 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2908 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2909 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2914 /// ReplaceNodeResults - Replace the results of node with an illegal result
2915 /// type with new values built out of custom code.
2916 void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
2917 SmallVectorImpl<SDValue>&Results,
2918 SelectionDAG &DAG) {
2919 switch (N->getOpcode()) {
2921 llvm_unreachable("Don't know how to custom expand this!");
2923 case ISD::BIT_CONVERT:
2924 Results.push_back(ExpandBIT_CONVERT(N, DAG));
2928 SDValue Res = LowerShift(N, DAG, Subtarget);
2930 Results.push_back(Res);
2936 //===----------------------------------------------------------------------===//
2937 // ARM Scheduler Hooks
2938 //===----------------------------------------------------------------------===//
2941 ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2942 MachineBasicBlock *BB,
2943 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
2944 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2945 DebugLoc dl = MI->getDebugLoc();
2946 switch (MI->getOpcode()) {
2948 llvm_unreachable("Unexpected instr type to insert");
2949 case ARM::tMOVCCr_pseudo: {
2950 // To "insert" a SELECT_CC instruction, we actually have to insert the
2951 // diamond control-flow pattern. The incoming instruction knows the
2952 // destination vreg to set, the condition code register to branch on, the
2953 // true/false values to select between, and a branch opcode to use.
2954 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2955 MachineFunction::iterator It = BB;
2961 // cmpTY ccX, r1, r2
2963 // fallthrough --> copy0MBB
2964 MachineBasicBlock *thisMBB = BB;
2965 MachineFunction *F = BB->getParent();
2966 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
2967 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
2968 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
2969 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
2970 F->insert(It, copy0MBB);
2971 F->insert(It, sinkMBB);
2972 // Update machine-CFG edges by first adding all successors of the current
2973 // block to the new block which will contain the Phi node for the select.
2974 // Also inform sdisel of the edge changes.
2975 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
2976 E = BB->succ_end(); I != E; ++I) {
2977 EM->insert(std::make_pair(*I, sinkMBB));
2978 sinkMBB->addSuccessor(*I);
2980 // Next, remove all successors of the current block, and add the true
2981 // and fallthrough blocks as its successors.
2982 while (!BB->succ_empty())
2983 BB->removeSuccessor(BB->succ_begin());
2984 BB->addSuccessor(copy0MBB);
2985 BB->addSuccessor(sinkMBB);
2988 // %FalseValue = ...
2989 // # fallthrough to sinkMBB
2992 // Update machine-CFG edges
2993 BB->addSuccessor(sinkMBB);
2996 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2999 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg())
3000 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
3001 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3003 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3010 case ARM::t2SUBrSPi_:
3011 case ARM::t2SUBrSPi12_:
3012 case ARM::t2SUBrSPs_: {
3013 MachineFunction *MF = BB->getParent();
3014 unsigned DstReg = MI->getOperand(0).getReg();
3015 unsigned SrcReg = MI->getOperand(1).getReg();
3016 bool DstIsDead = MI->getOperand(0).isDead();
3017 bool SrcIsKill = MI->getOperand(1).isKill();
3019 if (SrcReg != ARM::SP) {
3020 // Copy the source to SP from virtual register.
3021 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg);
3022 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3023 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr;
3024 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP)
3025 .addReg(SrcReg, getKillRegState(SrcIsKill));
3029 bool NeedPred = false, NeedCC = false, NeedOp3 = false;
3030 switch (MI->getOpcode()) {
3032 llvm_unreachable("Unexpected pseudo instruction!");
3038 OpOpc = ARM::tADDspr;
3041 OpOpc = ARM::tSUBspi;
3043 case ARM::t2SUBrSPi_:
3044 OpOpc = ARM::t2SUBrSPi;
3045 NeedPred = true; NeedCC = true;
3047 case ARM::t2SUBrSPi12_:
3048 OpOpc = ARM::t2SUBrSPi12;
3051 case ARM::t2SUBrSPs_:
3052 OpOpc = ARM::t2SUBrSPs;
3053 NeedPred = true; NeedCC = true; NeedOp3 = true;
3056 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP);
3057 if (OpOpc == ARM::tAND)
3058 AddDefaultT1CC(MIB);
3059 MIB.addReg(ARM::SP);
3060 MIB.addOperand(MI->getOperand(2));
3062 MIB.addOperand(MI->getOperand(3));
3064 AddDefaultPred(MIB);
3068 // Copy the result from SP to virtual register.
3069 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg);
3070 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass)
3071 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr;
3072 BuildMI(BB, dl, TII->get(CopyOpc))
3073 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead))
3075 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
3081 //===----------------------------------------------------------------------===//
3082 // ARM Optimization Hooks
3083 //===----------------------------------------------------------------------===//
3086 SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
3087 TargetLowering::DAGCombinerInfo &DCI) {
3088 SelectionDAG &DAG = DCI.DAG;
3089 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3090 EVT VT = N->getValueType(0);
3091 unsigned Opc = N->getOpcode();
3092 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
3093 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
3094 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
3095 ISD::CondCode CC = ISD::SETCC_INVALID;
3098 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
3100 SDValue CCOp = Slct.getOperand(0);
3101 if (CCOp.getOpcode() == ISD::SETCC)
3102 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
3105 bool DoXform = false;
3107 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
3110 if (LHS.getOpcode() == ISD::Constant &&
3111 cast<ConstantSDNode>(LHS)->isNullValue()) {
3113 } else if (CC != ISD::SETCC_INVALID &&
3114 RHS.getOpcode() == ISD::Constant &&
3115 cast<ConstantSDNode>(RHS)->isNullValue()) {
3116 std::swap(LHS, RHS);
3117 SDValue Op0 = Slct.getOperand(0);
3118 EVT OpVT = isSlctCC ? Op0.getValueType() :
3119 Op0.getOperand(0).getValueType();
3120 bool isInt = OpVT.isInteger();
3121 CC = ISD::getSetCCInverse(CC, isInt);
3123 if (!TLI.isCondCodeLegal(CC, OpVT))
3124 return SDValue(); // Inverse operator isn't legal.
3131 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
3133 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
3134 Slct.getOperand(0), Slct.getOperand(1), CC);
3135 SDValue CCOp = Slct.getOperand(0);
3137 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
3138 CCOp.getOperand(0), CCOp.getOperand(1), CC);
3139 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
3140 CCOp, OtherOp, Result);
3145 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3146 static SDValue PerformADDCombine(SDNode *N,
3147 TargetLowering::DAGCombinerInfo &DCI) {
3148 // added by evan in r37685 with no testcase.
3149 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3151 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
3152 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
3153 SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
3154 if (Result.getNode()) return Result;
3156 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3157 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3158 if (Result.getNode()) return Result;
3164 /// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
3165 static SDValue PerformSUBCombine(SDNode *N,
3166 TargetLowering::DAGCombinerInfo &DCI) {
3167 // added by evan in r37685 with no testcase.
3168 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
3170 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
3171 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
3172 SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
3173 if (Result.getNode()) return Result;
3179 /// PerformVMOVRRDCombine - Target-specific dag combine xforms for ARMISD::VMOVRRD.
3180 static SDValue PerformVMOVRRDCombine(SDNode *N,
3181 TargetLowering::DAGCombinerInfo &DCI) {
3182 // fmrrd(fmdrr x, y) -> x,y
3183 SDValue InDouble = N->getOperand(0);
3184 if (InDouble.getOpcode() == ARMISD::VMOVDRR)
3185 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
3189 /// getVShiftImm - Check if this is a valid build_vector for the immediate
3190 /// operand of a vector shift operation, where all the elements of the
3191 /// build_vector must have the same constant integer value.
3192 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3193 // Ignore bit_converts.
3194 while (Op.getOpcode() == ISD::BIT_CONVERT)
3195 Op = Op.getOperand(0);
3196 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3197 APInt SplatBits, SplatUndef;
3198 unsigned SplatBitSize;
3200 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3201 HasAnyUndefs, ElementBits) ||
3202 SplatBitSize > ElementBits)
3204 Cnt = SplatBits.getSExtValue();
3208 /// isVShiftLImm - Check if this is a valid build_vector for the immediate
3209 /// operand of a vector shift left operation. That value must be in the range:
3210 /// 0 <= Value < ElementBits for a left shift; or
3211 /// 0 <= Value <= ElementBits for a long left shift.
3212 static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
3213 assert(VT.isVector() && "vector shift count is not a vector type");
3214 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3215 if (! getVShiftImm(Op, ElementBits, Cnt))
3217 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
3220 /// isVShiftRImm - Check if this is a valid build_vector for the immediate
3221 /// operand of a vector shift right operation. For a shift opcode, the value
3222 /// is positive, but for an intrinsic the value count must be negative. The
3223 /// absolute value must be in the range:
3224 /// 1 <= |Value| <= ElementBits for a right shift; or
3225 /// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
3226 static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
3228 assert(VT.isVector() && "vector shift count is not a vector type");
3229 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3230 if (! getVShiftImm(Op, ElementBits, Cnt))
3234 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
3237 /// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
3238 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3239 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3242 // Don't do anything for most intrinsics.
3245 // Vector shifts: check for immediate versions and lower them.
3246 // Note: This is done during DAG combining instead of DAG legalizing because
3247 // the build_vectors for 64-bit vector element shift counts are generally
3248 // not legal, and it is hard to see their values after they get legalized to
3249 // loads from a constant pool.
3250 case Intrinsic::arm_neon_vshifts:
3251 case Intrinsic::arm_neon_vshiftu:
3252 case Intrinsic::arm_neon_vshiftls:
3253 case Intrinsic::arm_neon_vshiftlu:
3254 case Intrinsic::arm_neon_vshiftn:
3255 case Intrinsic::arm_neon_vrshifts:
3256 case Intrinsic::arm_neon_vrshiftu:
3257 case Intrinsic::arm_neon_vrshiftn:
3258 case Intrinsic::arm_neon_vqshifts:
3259 case Intrinsic::arm_neon_vqshiftu:
3260 case Intrinsic::arm_neon_vqshiftsu:
3261 case Intrinsic::arm_neon_vqshiftns:
3262 case Intrinsic::arm_neon_vqshiftnu:
3263 case Intrinsic::arm_neon_vqshiftnsu:
3264 case Intrinsic::arm_neon_vqrshiftns:
3265 case Intrinsic::arm_neon_vqrshiftnu:
3266 case Intrinsic::arm_neon_vqrshiftnsu: {
3267 EVT VT = N->getOperand(1).getValueType();
3269 unsigned VShiftOpc = 0;
3272 case Intrinsic::arm_neon_vshifts:
3273 case Intrinsic::arm_neon_vshiftu:
3274 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
3275 VShiftOpc = ARMISD::VSHL;
3278 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
3279 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
3280 ARMISD::VSHRs : ARMISD::VSHRu);
3285 case Intrinsic::arm_neon_vshiftls:
3286 case Intrinsic::arm_neon_vshiftlu:
3287 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
3289 llvm_unreachable("invalid shift count for vshll intrinsic");
3291 case Intrinsic::arm_neon_vrshifts:
3292 case Intrinsic::arm_neon_vrshiftu:
3293 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
3297 case Intrinsic::arm_neon_vqshifts:
3298 case Intrinsic::arm_neon_vqshiftu:
3299 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3303 case Intrinsic::arm_neon_vqshiftsu:
3304 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
3306 llvm_unreachable("invalid shift count for vqshlu intrinsic");
3308 case Intrinsic::arm_neon_vshiftn:
3309 case Intrinsic::arm_neon_vrshiftn:
3310 case Intrinsic::arm_neon_vqshiftns:
3311 case Intrinsic::arm_neon_vqshiftnu:
3312 case Intrinsic::arm_neon_vqshiftnsu:
3313 case Intrinsic::arm_neon_vqrshiftns:
3314 case Intrinsic::arm_neon_vqrshiftnu:
3315 case Intrinsic::arm_neon_vqrshiftnsu:
3316 // Narrowing shifts require an immediate right shift.
3317 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
3319 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic");
3322 llvm_unreachable("unhandled vector shift");
3326 case Intrinsic::arm_neon_vshifts:
3327 case Intrinsic::arm_neon_vshiftu:
3328 // Opcode already set above.
3330 case Intrinsic::arm_neon_vshiftls:
3331 case Intrinsic::arm_neon_vshiftlu:
3332 if (Cnt == VT.getVectorElementType().getSizeInBits())
3333 VShiftOpc = ARMISD::VSHLLi;
3335 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
3336 ARMISD::VSHLLs : ARMISD::VSHLLu);
3338 case Intrinsic::arm_neon_vshiftn:
3339 VShiftOpc = ARMISD::VSHRN; break;
3340 case Intrinsic::arm_neon_vrshifts:
3341 VShiftOpc = ARMISD::VRSHRs; break;
3342 case Intrinsic::arm_neon_vrshiftu:
3343 VShiftOpc = ARMISD::VRSHRu; break;
3344 case Intrinsic::arm_neon_vrshiftn:
3345 VShiftOpc = ARMISD::VRSHRN; break;
3346 case Intrinsic::arm_neon_vqshifts:
3347 VShiftOpc = ARMISD::VQSHLs; break;
3348 case Intrinsic::arm_neon_vqshiftu:
3349 VShiftOpc = ARMISD::VQSHLu; break;
3350 case Intrinsic::arm_neon_vqshiftsu:
3351 VShiftOpc = ARMISD::VQSHLsu; break;
3352 case Intrinsic::arm_neon_vqshiftns:
3353 VShiftOpc = ARMISD::VQSHRNs; break;
3354 case Intrinsic::arm_neon_vqshiftnu:
3355 VShiftOpc = ARMISD::VQSHRNu; break;
3356 case Intrinsic::arm_neon_vqshiftnsu:
3357 VShiftOpc = ARMISD::VQSHRNsu; break;
3358 case Intrinsic::arm_neon_vqrshiftns:
3359 VShiftOpc = ARMISD::VQRSHRNs; break;
3360 case Intrinsic::arm_neon_vqrshiftnu:
3361 VShiftOpc = ARMISD::VQRSHRNu; break;
3362 case Intrinsic::arm_neon_vqrshiftnsu:
3363 VShiftOpc = ARMISD::VQRSHRNsu; break;
3366 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3367 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3370 case Intrinsic::arm_neon_vshiftins: {
3371 EVT VT = N->getOperand(1).getValueType();
3373 unsigned VShiftOpc = 0;
3375 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
3376 VShiftOpc = ARMISD::VSLI;
3377 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
3378 VShiftOpc = ARMISD::VSRI;
3380 llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
3383 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
3384 N->getOperand(1), N->getOperand(2),
3385 DAG.getConstant(Cnt, MVT::i32));
3388 case Intrinsic::arm_neon_vqrshifts:
3389 case Intrinsic::arm_neon_vqrshiftu:
3390 // No immediate versions of these to check for.
3397 /// PerformShiftCombine - Checks for immediate versions of vector shifts and
3398 /// lowers them. As with the vector shift intrinsics, this is done during DAG
3399 /// combining instead of DAG legalizing because the build_vectors for 64-bit
3400 /// vector element shift counts are generally not legal, and it is hard to see
3401 /// their values after they get legalized to loads from a constant pool.
3402 static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
3403 const ARMSubtarget *ST) {
3404 EVT VT = N->getValueType(0);
3406 // Nothing to be done for scalar shifts.
3407 if (! VT.isVector())
3410 assert(ST->hasNEON() && "unexpected vector shift");
3413 switch (N->getOpcode()) {
3414 default: llvm_unreachable("unexpected shift opcode");
3417 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
3418 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
3419 DAG.getConstant(Cnt, MVT::i32));
3424 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
3425 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
3426 ARMISD::VSHRs : ARMISD::VSHRu);
3427 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
3428 DAG.getConstant(Cnt, MVT::i32));
3434 /// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
3435 /// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
3436 static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
3437 const ARMSubtarget *ST) {
3438 SDValue N0 = N->getOperand(0);
3440 // Check for sign- and zero-extensions of vector extract operations of 8-
3441 // and 16-bit vector elements. NEON supports these directly. They are
3442 // handled during DAG combining because type legalization will promote them
3443 // to 32-bit types and it is messy to recognize the operations after that.
3444 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
3445 SDValue Vec = N0.getOperand(0);
3446 SDValue Lane = N0.getOperand(1);
3447 EVT VT = N->getValueType(0);
3448 EVT EltVT = N0.getValueType();
3449 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3451 if (VT == MVT::i32 &&
3452 (EltVT == MVT::i8 || EltVT == MVT::i16) &&
3453 TLI.isTypeLegal(Vec.getValueType())) {
3456 switch (N->getOpcode()) {
3457 default: llvm_unreachable("unexpected opcode");
3458 case ISD::SIGN_EXTEND:
3459 Opc = ARMISD::VGETLANEs;
3461 case ISD::ZERO_EXTEND:
3462 case ISD::ANY_EXTEND:
3463 Opc = ARMISD::VGETLANEu;
3466 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
3473 SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
3474 DAGCombinerInfo &DCI) const {
3475 switch (N->getOpcode()) {
3477 case ISD::ADD: return PerformADDCombine(N, DCI);
3478 case ISD::SUB: return PerformSUBCombine(N, DCI);
3479 case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
3480 case ISD::INTRINSIC_WO_CHAIN:
3481 return PerformIntrinsicCombine(N, DCI.DAG);
3485 return PerformShiftCombine(N, DCI.DAG, Subtarget);
3486 case ISD::SIGN_EXTEND:
3487 case ISD::ZERO_EXTEND:
3488 case ISD::ANY_EXTEND:
3489 return PerformExtendCombine(N, DCI.DAG, Subtarget);
3494 bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
3495 if (!Subtarget->hasV6Ops())
3496 // Pre-v6 does not support unaligned mem access.
3498 else if (!Subtarget->hasV6Ops()) {
3499 // v6 may or may not support unaligned mem access.
3500 if (!Subtarget->isTargetDarwin())
3504 switch (VT.getSimpleVT().SimpleTy) {
3511 // FIXME: VLD1 etc with standard alignment is legal.
3515 static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
3520 switch (VT.getSimpleVT().SimpleTy) {
3521 default: return false;
3536 if ((V & (Scale - 1)) != 0)
3539 return V == (V & ((1LL << 5) - 1));
3542 static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
3543 const ARMSubtarget *Subtarget) {
3550 switch (VT.getSimpleVT().SimpleTy) {
3551 default: return false;
3556 // + imm12 or - imm8
3558 return V == (V & ((1LL << 8) - 1));
3559 return V == (V & ((1LL << 12) - 1));
3562 // Same as ARM mode. FIXME: NEON?
3563 if (!Subtarget->hasVFP2())
3568 return V == (V & ((1LL << 8) - 1));
3572 /// isLegalAddressImmediate - Return true if the integer value can be used
3573 /// as the offset of the target addressing mode for load / store of the
3575 static bool isLegalAddressImmediate(int64_t V, EVT VT,
3576 const ARMSubtarget *Subtarget) {
3583 if (Subtarget->isThumb1Only())
3584 return isLegalT1AddressImmediate(V, VT);
3585 else if (Subtarget->isThumb2())
3586 return isLegalT2AddressImmediate(V, VT, Subtarget);
3591 switch (VT.getSimpleVT().SimpleTy) {
3592 default: return false;
3597 return V == (V & ((1LL << 12) - 1));
3600 return V == (V & ((1LL << 8) - 1));
3603 if (!Subtarget->hasVFP2()) // FIXME: NEON?
3608 return V == (V & ((1LL << 8) - 1));
3612 bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
3614 int Scale = AM.Scale;
3618 switch (VT.getSimpleVT().SimpleTy) {
3619 default: return false;
3628 return Scale == 2 || Scale == 4 || Scale == 8;
3631 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3635 // Note, we allow "void" uses (basically, uses that aren't loads or
3636 // stores), because arm allows folding a scale into many arithmetic
3637 // operations. This should be made more precise and revisited later.
3639 // Allow r << imm, but the imm has to be a multiple of two.
3640 if (Scale & 1) return false;
3641 return isPowerOf2_32(Scale);
3645 /// isLegalAddressingMode - Return true if the addressing mode represented
3646 /// by AM is legal for this target, for a load/store of the specified type.
3647 bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3648 const Type *Ty) const {
3649 EVT VT = getValueType(Ty, true);
3650 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
3653 // Can never fold addr of global into load/store.
3658 case 0: // no scale reg, must be "r+i" or "r", or "i".
3661 if (Subtarget->isThumb1Only())
3665 // ARM doesn't support any R+R*scale+imm addr modes.
3672 if (Subtarget->isThumb2())
3673 return isLegalT2ScaledAddressingMode(AM, VT);
3675 int Scale = AM.Scale;
3676 switch (VT.getSimpleVT().SimpleTy) {
3677 default: return false;
3681 if (Scale < 0) Scale = -Scale;
3685 return isPowerOf2_32(Scale & ~1);
3689 if (((unsigned)AM.HasBaseReg + Scale) <= 2)
3694 // Note, we allow "void" uses (basically, uses that aren't loads or
3695 // stores), because arm allows folding a scale into many arithmetic
3696 // operations. This should be made more precise and revisited later.
3698 // Allow r << imm, but the imm has to be a multiple of two.
3699 if (Scale & 1) return false;
3700 return isPowerOf2_32(Scale);
3707 /// isLegalICmpImmediate - Return true if the specified immediate is legal
3708 /// icmp immediate, that is the target has icmp instructions which can compare
3709 /// a register against the immediate without having to materialize the
3710 /// immediate into a register.
3711 bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
3712 if (!Subtarget->isThumb())
3713 return ARM_AM::getSOImmVal(Imm) != -1;
3714 if (Subtarget->isThumb2())
3715 return ARM_AM::getT2SOImmVal(Imm) != -1;
3716 return Imm >= 0 && Imm <= 255;
3719 static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
3720 bool isSEXTLoad, SDValue &Base,
3721 SDValue &Offset, bool &isInc,
3722 SelectionDAG &DAG) {
3723 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3726 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
3728 Base = Ptr->getOperand(0);
3729 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3730 int RHSC = (int)RHS->getZExtValue();
3731 if (RHSC < 0 && RHSC > -256) {
3732 assert(Ptr->getOpcode() == ISD::ADD);
3734 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3738 isInc = (Ptr->getOpcode() == ISD::ADD);
3739 Offset = Ptr->getOperand(1);
3741 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
3743 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3744 int RHSC = (int)RHS->getZExtValue();
3745 if (RHSC < 0 && RHSC > -0x1000) {
3746 assert(Ptr->getOpcode() == ISD::ADD);
3748 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3749 Base = Ptr->getOperand(0);
3754 if (Ptr->getOpcode() == ISD::ADD) {
3756 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0));
3757 if (ShOpcVal != ARM_AM::no_shift) {
3758 Base = Ptr->getOperand(1);
3759 Offset = Ptr->getOperand(0);
3761 Base = Ptr->getOperand(0);
3762 Offset = Ptr->getOperand(1);
3767 isInc = (Ptr->getOpcode() == ISD::ADD);
3768 Base = Ptr->getOperand(0);
3769 Offset = Ptr->getOperand(1);
3773 // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
3777 static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
3778 bool isSEXTLoad, SDValue &Base,
3779 SDValue &Offset, bool &isInc,
3780 SelectionDAG &DAG) {
3781 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
3784 Base = Ptr->getOperand(0);
3785 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
3786 int RHSC = (int)RHS->getZExtValue();
3787 if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
3788 assert(Ptr->getOpcode() == ISD::ADD);
3790 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
3792 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
3793 isInc = Ptr->getOpcode() == ISD::ADD;
3794 Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
3802 /// getPreIndexedAddressParts - returns true by value, base pointer and
3803 /// offset pointer and addressing mode by reference if the node's address
3804 /// can be legally represented as pre-indexed load / store address.
3806 ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
3808 ISD::MemIndexedMode &AM,
3809 SelectionDAG &DAG) const {
3810 if (Subtarget->isThumb1Only())
3815 bool isSEXTLoad = false;
3816 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3817 Ptr = LD->getBasePtr();
3818 VT = LD->getMemoryVT();
3819 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3820 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3821 Ptr = ST->getBasePtr();
3822 VT = ST->getMemoryVT();
3827 bool isLegal = false;
3828 if (Subtarget->isThumb2())
3829 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3830 Offset, isInc, DAG);
3832 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
3833 Offset, isInc, DAG);
3837 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
3841 /// getPostIndexedAddressParts - returns true by value, base pointer and
3842 /// offset pointer and addressing mode by reference if this node can be
3843 /// combined with a load / store to form a post-indexed load / store.
3844 bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
3847 ISD::MemIndexedMode &AM,
3848 SelectionDAG &DAG) const {
3849 if (Subtarget->isThumb1Only())
3854 bool isSEXTLoad = false;
3855 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
3856 VT = LD->getMemoryVT();
3857 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
3858 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
3859 VT = ST->getMemoryVT();
3864 bool isLegal = false;
3865 if (Subtarget->isThumb2())
3866 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3869 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
3874 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
3878 void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3882 const SelectionDAG &DAG,
3883 unsigned Depth) const {
3884 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3885 switch (Op.getOpcode()) {
3887 case ARMISD::CMOV: {
3888 // Bits are known zero/one if known on the LHS and RHS.
3889 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
3890 if (KnownZero == 0 && KnownOne == 0) return;
3892 APInt KnownZeroRHS, KnownOneRHS;
3893 DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
3894 KnownZeroRHS, KnownOneRHS, Depth+1);
3895 KnownZero &= KnownZeroRHS;
3896 KnownOne &= KnownOneRHS;
3902 //===----------------------------------------------------------------------===//
3903 // ARM Inline Assembly Support
3904 //===----------------------------------------------------------------------===//
3906 /// getConstraintType - Given a constraint letter, return the type of
3907 /// constraint it is for this target.
3908 ARMTargetLowering::ConstraintType
3909 ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
3910 if (Constraint.size() == 1) {
3911 switch (Constraint[0]) {
3913 case 'l': return C_RegisterClass;
3914 case 'w': return C_RegisterClass;
3917 return TargetLowering::getConstraintType(Constraint);
3920 std::pair<unsigned, const TargetRegisterClass*>
3921 ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3923 if (Constraint.size() == 1) {
3924 // GCC RS6000 Constraint Letters
3925 switch (Constraint[0]) {
3927 if (Subtarget->isThumb1Only())
3928 return std::make_pair(0U, ARM::tGPRRegisterClass);
3930 return std::make_pair(0U, ARM::GPRRegisterClass);
3932 return std::make_pair(0U, ARM::GPRRegisterClass);
3935 return std::make_pair(0U, ARM::SPRRegisterClass);
3937 return std::make_pair(0U, ARM::DPRRegisterClass);
3941 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3944 std::vector<unsigned> ARMTargetLowering::
3945 getRegClassForInlineAsmConstraint(const std::string &Constraint,
3947 if (Constraint.size() != 1)
3948 return std::vector<unsigned>();
3950 switch (Constraint[0]) { // GCC ARM Constraint Letters
3953 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3954 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3957 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3,
3958 ARM::R4, ARM::R5, ARM::R6, ARM::R7,
3959 ARM::R8, ARM::R9, ARM::R10, ARM::R11,
3960 ARM::R12, ARM::LR, 0);
3963 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3,
3964 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
3965 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
3966 ARM::S12,ARM::S13,ARM::S14,ARM::S15,
3967 ARM::S16,ARM::S17,ARM::S18,ARM::S19,
3968 ARM::S20,ARM::S21,ARM::S22,ARM::S23,
3969 ARM::S24,ARM::S25,ARM::S26,ARM::S27,
3970 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
3972 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
3973 ARM::D4, ARM::D5, ARM::D6, ARM::D7,
3974 ARM::D8, ARM::D9, ARM::D10,ARM::D11,
3975 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
3979 return std::vector<unsigned>();
3982 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3983 /// vector. If it is invalid, don't add anything to Ops.
3984 void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3987 std::vector<SDValue>&Ops,
3988 SelectionDAG &DAG) const {
3989 SDValue Result(0, 0);
3991 switch (Constraint) {
3993 case 'I': case 'J': case 'K': case 'L':
3994 case 'M': case 'N': case 'O':
3995 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
3999 int64_t CVal64 = C->getSExtValue();
4000 int CVal = (int) CVal64;
4001 // None of these constraints allow values larger than 32 bits. Check
4002 // that the value fits in an int.
4006 switch (Constraint) {
4008 if (Subtarget->isThumb1Only()) {
4009 // This must be a constant between 0 and 255, for ADD
4011 if (CVal >= 0 && CVal <= 255)
4013 } else if (Subtarget->isThumb2()) {
4014 // A constant that can be used as an immediate value in a
4015 // data-processing instruction.
4016 if (ARM_AM::getT2SOImmVal(CVal) != -1)
4019 // A constant that can be used as an immediate value in a
4020 // data-processing instruction.
4021 if (ARM_AM::getSOImmVal(CVal) != -1)
4027 if (Subtarget->isThumb()) { // FIXME thumb2
4028 // This must be a constant between -255 and -1, for negated ADD
4029 // immediates. This can be used in GCC with an "n" modifier that
4030 // prints the negated value, for use with SUB instructions. It is
4031 // not useful otherwise but is implemented for compatibility.
4032 if (CVal >= -255 && CVal <= -1)
4035 // This must be a constant between -4095 and 4095. It is not clear
4036 // what this constraint is intended for. Implemented for
4037 // compatibility with GCC.
4038 if (CVal >= -4095 && CVal <= 4095)
4044 if (Subtarget->isThumb1Only()) {
4045 // A 32-bit value where only one byte has a nonzero value. Exclude
4046 // zero to match GCC. This constraint is used by GCC internally for
4047 // constants that can be loaded with a move/shift combination.
4048 // It is not useful otherwise but is implemented for compatibility.
4049 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
4051 } else if (Subtarget->isThumb2()) {
4052 // A constant whose bitwise inverse can be used as an immediate
4053 // value in a data-processing instruction. This can be used in GCC
4054 // with a "B" modifier that prints the inverted value, for use with
4055 // BIC and MVN instructions. It is not useful otherwise but is
4056 // implemented for compatibility.
4057 if (ARM_AM::getT2SOImmVal(~CVal) != -1)
4060 // A constant whose bitwise inverse can be used as an immediate
4061 // value in a data-processing instruction. This can be used in GCC
4062 // with a "B" modifier that prints the inverted value, for use with
4063 // BIC and MVN instructions. It is not useful otherwise but is
4064 // implemented for compatibility.
4065 if (ARM_AM::getSOImmVal(~CVal) != -1)
4071 if (Subtarget->isThumb1Only()) {
4072 // This must be a constant between -7 and 7,
4073 // for 3-operand ADD/SUB immediate instructions.
4074 if (CVal >= -7 && CVal < 7)
4076 } else if (Subtarget->isThumb2()) {
4077 // A constant whose negation can be used as an immediate value in a
4078 // data-processing instruction. This can be used in GCC with an "n"
4079 // modifier that prints the negated value, for use with SUB
4080 // instructions. It is not useful otherwise but is implemented for
4082 if (ARM_AM::getT2SOImmVal(-CVal) != -1)
4085 // A constant whose negation can be used as an immediate value in a
4086 // data-processing instruction. This can be used in GCC with an "n"
4087 // modifier that prints the negated value, for use with SUB
4088 // instructions. It is not useful otherwise but is implemented for
4090 if (ARM_AM::getSOImmVal(-CVal) != -1)
4096 if (Subtarget->isThumb()) { // FIXME thumb2
4097 // This must be a multiple of 4 between 0 and 1020, for
4098 // ADD sp + immediate.
4099 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
4102 // A power of two or a constant between 0 and 32. This is used in
4103 // GCC for the shift amount on shifted register operands, but it is
4104 // useful in general for any shift amounts.
4105 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
4111 if (Subtarget->isThumb()) { // FIXME thumb2
4112 // This must be a constant between 0 and 31, for shift amounts.
4113 if (CVal >= 0 && CVal <= 31)
4119 if (Subtarget->isThumb()) { // FIXME thumb2
4120 // This must be a multiple of 4 between -508 and 508, for
4121 // ADD/SUB sp = sp + immediate.
4122 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
4127 Result = DAG.getTargetConstant(CVal, Op.getValueType());
4131 if (Result.getNode()) {
4132 Ops.push_back(Result);
4135 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
4140 ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
4141 // The ARM target isn't yet aware of offsets.
4145 int ARM::getVFPf32Imm(const APFloat &FPImm) {
4146 APInt Imm = FPImm.bitcastToAPInt();
4147 uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
4148 int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
4149 int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
4151 // We can handle 4 bits of mantissa.
4152 // mantissa = (16+UInt(e:f:g:h))/16.
4153 if (Mantissa & 0x7ffff)
4156 if ((Mantissa & 0xf) != Mantissa)
4159 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4160 if (Exp < -3 || Exp > 4)
4162 Exp = ((Exp+3) & 0x7) ^ 4;
4164 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4167 int ARM::getVFPf64Imm(const APFloat &FPImm) {
4168 APInt Imm = FPImm.bitcastToAPInt();
4169 uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
4170 int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
4171 uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
4173 // We can handle 4 bits of mantissa.
4174 // mantissa = (16+UInt(e:f:g:h))/16.
4175 if (Mantissa & 0xffffffffffffLL)
4178 if ((Mantissa & 0xf) != Mantissa)
4181 // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
4182 if (Exp < -3 || Exp > 4)
4184 Exp = ((Exp+3) & 0x7) ^ 4;
4186 return ((int)Sign << 7) | (Exp << 4) | Mantissa;
4189 /// isFPImmLegal - Returns true if the target can instruction select the
4190 /// specified FP immediate natively. If false, the legalizer will
4191 /// materialize the FP immediate as a load from a constant pool.
4192 bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4193 if (!Subtarget->hasVFP3())
4196 return ARM::getVFPf32Imm(Imm) != -1;
4198 return ARM::getVFPf64Imm(Imm) != -1;