1 //===-- PPCISelLowering.cpp - PPC 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 implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCMachineFunctionInfo.h"
16 #include "PPCPerfectShuffle.h"
17 #include "PPCTargetMachine.h"
18 #include "MCTargetDesc/PPCPredicates.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/VectorExtras.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
28 #include "llvm/CallingConv.h"
29 #include "llvm/Constants.h"
30 #include "llvm/Function.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/DerivedTypes.h"
40 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
41 CCValAssign::LocInfo &LocInfo,
42 ISD::ArgFlagsTy &ArgFlags,
44 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
46 CCValAssign::LocInfo &LocInfo,
47 ISD::ArgFlagsTy &ArgFlags,
49 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
51 CCValAssign::LocInfo &LocInfo,
52 ISD::ArgFlagsTy &ArgFlags,
55 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
56 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
59 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
60 if (TM.getSubtargetImpl()->isDarwin())
61 return new TargetLoweringObjectFileMachO();
63 return new TargetLoweringObjectFileELF();
66 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
67 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
71 // Use _setjmp/_longjmp instead of setjmp/longjmp.
72 setUseUnderscoreSetJmp(true);
73 setUseUnderscoreLongJmp(true);
75 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
76 // arguments are at least 4/8 bytes aligned.
77 setMinStackArgumentAlignment(TM.getSubtarget<PPCSubtarget>().isPPC64() ? 8:4);
79 // Set up the register classes.
80 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
81 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
82 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
84 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
85 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
86 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
88 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
90 // PowerPC has pre-inc load and store's.
91 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
93 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
96 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
98 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
102 // This is used in the ppcf128->int sequence. Note it has different semantics
103 // from FP_ROUND: that rounds to nearest, this rounds to zero.
104 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
106 // PowerPC has no SREM/UREM instructions
107 setOperationAction(ISD::SREM, MVT::i32, Expand);
108 setOperationAction(ISD::UREM, MVT::i32, Expand);
109 setOperationAction(ISD::SREM, MVT::i64, Expand);
110 setOperationAction(ISD::UREM, MVT::i64, Expand);
112 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
113 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
114 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
115 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
116 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
117 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
118 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
119 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
120 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
122 // We don't support sin/cos/sqrt/fmod/pow
123 setOperationAction(ISD::FSIN , MVT::f64, Expand);
124 setOperationAction(ISD::FCOS , MVT::f64, Expand);
125 setOperationAction(ISD::FREM , MVT::f64, Expand);
126 setOperationAction(ISD::FPOW , MVT::f64, Expand);
127 setOperationAction(ISD::FMA , MVT::f64, Expand);
128 setOperationAction(ISD::FSIN , MVT::f32, Expand);
129 setOperationAction(ISD::FCOS , MVT::f32, Expand);
130 setOperationAction(ISD::FREM , MVT::f32, Expand);
131 setOperationAction(ISD::FPOW , MVT::f32, Expand);
132 setOperationAction(ISD::FMA , MVT::f32, Expand);
134 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
136 // If we're enabling GP optimizations, use hardware square root
137 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
138 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
139 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
142 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
143 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
145 // PowerPC does not have BSWAP, CTPOP or CTTZ
146 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
147 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
148 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
149 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
150 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
151 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
153 // PowerPC does not have ROTR
154 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
155 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
157 // PowerPC does not have Select
158 setOperationAction(ISD::SELECT, MVT::i32, Expand);
159 setOperationAction(ISD::SELECT, MVT::i64, Expand);
160 setOperationAction(ISD::SELECT, MVT::f32, Expand);
161 setOperationAction(ISD::SELECT, MVT::f64, Expand);
163 // PowerPC wants to turn select_cc of FP into fsel when possible.
164 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
165 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
167 // PowerPC wants to optimize integer setcc a bit
168 setOperationAction(ISD::SETCC, MVT::i32, Custom);
170 // PowerPC does not have BRCOND which requires SetCC
171 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
173 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
175 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
176 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
178 // PowerPC does not have [U|S]INT_TO_FP
179 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
180 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
182 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
183 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
184 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
185 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
187 // We cannot sextinreg(i1). Expand to shifts.
188 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
190 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
191 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
192 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
193 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
196 // We want to legalize GlobalAddress and ConstantPool nodes into the
197 // appropriate instructions to materialize the address.
198 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
199 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
200 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
201 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
202 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
203 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
204 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
205 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
206 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
207 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
210 setOperationAction(ISD::TRAP, MVT::Other, Legal);
212 // TRAMPOLINE is custom lowered.
213 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
214 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
216 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
217 setOperationAction(ISD::VASTART , MVT::Other, Custom);
219 // VAARG is custom lowered with the 32-bit SVR4 ABI.
220 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
221 && !TM.getSubtarget<PPCSubtarget>().isPPC64()) {
222 setOperationAction(ISD::VAARG, MVT::Other, Custom);
223 setOperationAction(ISD::VAARG, MVT::i64, Custom);
225 setOperationAction(ISD::VAARG, MVT::Other, Expand);
227 // Use the default implementation.
228 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
229 setOperationAction(ISD::VAEND , MVT::Other, Expand);
230 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
231 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
232 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
233 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
235 // We want to custom lower some of our intrinsics.
236 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
238 // Comparisons that require checking two conditions.
239 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
240 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
241 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
242 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
243 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
244 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
245 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
246 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
247 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
248 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
249 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
250 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
252 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
253 // They also have instructions for converting between i64 and fp.
254 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
255 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
256 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
257 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
258 // This is just the low 32 bits of a (signed) fp->i64 conversion.
259 // We cannot do this with Promote because i64 is not a legal type.
260 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
262 // FIXME: disable this lowered code. This generates 64-bit register values,
263 // and we don't model the fact that the top part is clobbered by calls. We
264 // need to flag these together so that the value isn't live across a call.
265 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
267 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
268 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
271 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
272 // 64-bit PowerPC implementations can support i64 types directly
273 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
274 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
275 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
276 // 64-bit PowerPC wants to expand i128 shifts itself.
277 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
278 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
279 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
281 // 32-bit PowerPC wants to expand i64 shifts itself.
282 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
283 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
284 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
287 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
288 // First set operation action for all vector types to expand. Then we
289 // will selectively turn on ones that can be effectively codegen'd.
290 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
291 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
292 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
294 // add/sub are legal for all supported vector VT's.
295 setOperationAction(ISD::ADD , VT, Legal);
296 setOperationAction(ISD::SUB , VT, Legal);
298 // We promote all shuffles to v16i8.
299 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
300 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
302 // We promote all non-typed operations to v4i32.
303 setOperationAction(ISD::AND , VT, Promote);
304 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
305 setOperationAction(ISD::OR , VT, Promote);
306 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
307 setOperationAction(ISD::XOR , VT, Promote);
308 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
309 setOperationAction(ISD::LOAD , VT, Promote);
310 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
311 setOperationAction(ISD::SELECT, VT, Promote);
312 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
313 setOperationAction(ISD::STORE, VT, Promote);
314 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
316 // No other operations are legal.
317 setOperationAction(ISD::MUL , VT, Expand);
318 setOperationAction(ISD::SDIV, VT, Expand);
319 setOperationAction(ISD::SREM, VT, Expand);
320 setOperationAction(ISD::UDIV, VT, Expand);
321 setOperationAction(ISD::UREM, VT, Expand);
322 setOperationAction(ISD::FDIV, VT, Expand);
323 setOperationAction(ISD::FNEG, VT, Expand);
324 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
325 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
326 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
327 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
328 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
329 setOperationAction(ISD::UDIVREM, VT, Expand);
330 setOperationAction(ISD::SDIVREM, VT, Expand);
331 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
332 setOperationAction(ISD::FPOW, VT, Expand);
333 setOperationAction(ISD::CTPOP, VT, Expand);
334 setOperationAction(ISD::CTLZ, VT, Expand);
335 setOperationAction(ISD::CTTZ, VT, Expand);
338 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
339 // with merges, splats, etc.
340 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
342 setOperationAction(ISD::AND , MVT::v4i32, Legal);
343 setOperationAction(ISD::OR , MVT::v4i32, Legal);
344 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
345 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
346 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
347 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
349 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
350 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
351 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
352 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
354 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
355 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
356 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
357 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
359 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
360 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
362 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
363 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
364 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
365 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
369 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
371 setBooleanContents(ZeroOrOneBooleanContent);
372 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
374 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
375 setStackPointerRegisterToSaveRestore(PPC::X1);
376 setExceptionPointerRegister(PPC::X3);
377 setExceptionSelectorRegister(PPC::X4);
379 setStackPointerRegisterToSaveRestore(PPC::R1);
380 setExceptionPointerRegister(PPC::R3);
381 setExceptionSelectorRegister(PPC::R4);
384 // We have target-specific dag combine patterns for the following nodes:
385 setTargetDAGCombine(ISD::SINT_TO_FP);
386 setTargetDAGCombine(ISD::STORE);
387 setTargetDAGCombine(ISD::BR_CC);
388 setTargetDAGCombine(ISD::BSWAP);
390 // Darwin long double math library functions have $LDBL128 appended.
391 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
392 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
393 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
394 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
395 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
396 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
397 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
398 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
399 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
400 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
401 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
404 setMinFunctionAlignment(2);
405 if (PPCSubTarget.isDarwin())
406 setPrefFunctionAlignment(4);
408 setInsertFencesForAtomic(true);
410 setSchedulingPreference(Sched::Hybrid);
412 computeRegisterProperties();
415 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
416 /// function arguments in the caller parameter area.
417 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
418 const TargetMachine &TM = getTargetMachine();
419 // Darwin passes everything on 4 byte boundary.
420 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
426 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
429 case PPCISD::FSEL: return "PPCISD::FSEL";
430 case PPCISD::FCFID: return "PPCISD::FCFID";
431 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
432 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
433 case PPCISD::STFIWX: return "PPCISD::STFIWX";
434 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
435 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
436 case PPCISD::VPERM: return "PPCISD::VPERM";
437 case PPCISD::Hi: return "PPCISD::Hi";
438 case PPCISD::Lo: return "PPCISD::Lo";
439 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
440 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
441 case PPCISD::LOAD: return "PPCISD::LOAD";
442 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
443 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
444 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
445 case PPCISD::SRL: return "PPCISD::SRL";
446 case PPCISD::SRA: return "PPCISD::SRA";
447 case PPCISD::SHL: return "PPCISD::SHL";
448 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
449 case PPCISD::STD_32: return "PPCISD::STD_32";
450 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
451 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
452 case PPCISD::NOP: return "PPCISD::NOP";
453 case PPCISD::MTCTR: return "PPCISD::MTCTR";
454 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
455 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
456 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
457 case PPCISD::MFCR: return "PPCISD::MFCR";
458 case PPCISD::VCMP: return "PPCISD::VCMP";
459 case PPCISD::VCMPo: return "PPCISD::VCMPo";
460 case PPCISD::LBRX: return "PPCISD::LBRX";
461 case PPCISD::STBRX: return "PPCISD::STBRX";
462 case PPCISD::LARX: return "PPCISD::LARX";
463 case PPCISD::STCX: return "PPCISD::STCX";
464 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
465 case PPCISD::MFFS: return "PPCISD::MFFS";
466 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
467 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
468 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
469 case PPCISD::MTFSF: return "PPCISD::MTFSF";
470 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
474 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
478 //===----------------------------------------------------------------------===//
479 // Node matching predicates, for use by the tblgen matching code.
480 //===----------------------------------------------------------------------===//
482 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
483 static bool isFloatingPointZero(SDValue Op) {
484 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
485 return CFP->getValueAPF().isZero();
486 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
487 // Maybe this has already been legalized into the constant pool?
488 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
489 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
490 return CFP->getValueAPF().isZero();
495 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
496 /// true if Op is undef or if it matches the specified value.
497 static bool isConstantOrUndef(int Op, int Val) {
498 return Op < 0 || Op == Val;
501 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
502 /// VPKUHUM instruction.
503 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
505 for (unsigned i = 0; i != 16; ++i)
506 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
509 for (unsigned i = 0; i != 8; ++i)
510 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
511 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
517 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
518 /// VPKUWUM instruction.
519 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
521 for (unsigned i = 0; i != 16; i += 2)
522 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
523 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
526 for (unsigned i = 0; i != 8; i += 2)
527 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
528 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
529 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
530 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
536 /// isVMerge - Common function, used to match vmrg* shuffles.
538 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
539 unsigned LHSStart, unsigned RHSStart) {
540 assert(N->getValueType(0) == MVT::v16i8 &&
541 "PPC only supports shuffles by bytes!");
542 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
543 "Unsupported merge size!");
545 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
546 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
547 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
548 LHSStart+j+i*UnitSize) ||
549 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
550 RHSStart+j+i*UnitSize))
556 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
557 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
558 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
561 return isVMerge(N, UnitSize, 8, 24);
562 return isVMerge(N, UnitSize, 8, 8);
565 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
566 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
567 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
570 return isVMerge(N, UnitSize, 0, 16);
571 return isVMerge(N, UnitSize, 0, 0);
575 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
576 /// amount, otherwise return -1.
577 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
578 assert(N->getValueType(0) == MVT::v16i8 &&
579 "PPC only supports shuffles by bytes!");
581 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
583 // Find the first non-undef value in the shuffle mask.
585 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
588 if (i == 16) return -1; // all undef.
590 // Otherwise, check to see if the rest of the elements are consecutively
591 // numbered from this value.
592 unsigned ShiftAmt = SVOp->getMaskElt(i);
593 if (ShiftAmt < i) return -1;
597 // Check the rest of the elements to see if they are consecutive.
598 for (++i; i != 16; ++i)
599 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
602 // Check the rest of the elements to see if they are consecutive.
603 for (++i; i != 16; ++i)
604 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
610 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
611 /// specifies a splat of a single element that is suitable for input to
612 /// VSPLTB/VSPLTH/VSPLTW.
613 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
614 assert(N->getValueType(0) == MVT::v16i8 &&
615 (EltSize == 1 || EltSize == 2 || EltSize == 4));
617 // This is a splat operation if each element of the permute is the same, and
618 // if the value doesn't reference the second vector.
619 unsigned ElementBase = N->getMaskElt(0);
621 // FIXME: Handle UNDEF elements too!
622 if (ElementBase >= 16)
625 // Check that the indices are consecutive, in the case of a multi-byte element
626 // splatted with a v16i8 mask.
627 for (unsigned i = 1; i != EltSize; ++i)
628 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
631 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
632 if (N->getMaskElt(i) < 0) continue;
633 for (unsigned j = 0; j != EltSize; ++j)
634 if (N->getMaskElt(i+j) != N->getMaskElt(j))
640 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
642 bool PPC::isAllNegativeZeroVector(SDNode *N) {
643 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
645 APInt APVal, APUndef;
649 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
650 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
651 return CFP->getValueAPF().isNegZero();
656 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
657 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
658 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
659 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
660 assert(isSplatShuffleMask(SVOp, EltSize));
661 return SVOp->getMaskElt(0) / EltSize;
664 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
665 /// by using a vspltis[bhw] instruction of the specified element size, return
666 /// the constant being splatted. The ByteSize field indicates the number of
667 /// bytes of each element [124] -> [bhw].
668 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
671 // If ByteSize of the splat is bigger than the element size of the
672 // build_vector, then we have a case where we are checking for a splat where
673 // multiple elements of the buildvector are folded together into a single
674 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
675 unsigned EltSize = 16/N->getNumOperands();
676 if (EltSize < ByteSize) {
677 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
678 SDValue UniquedVals[4];
679 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
681 // See if all of the elements in the buildvector agree across.
682 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
683 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
684 // If the element isn't a constant, bail fully out.
685 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
688 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
689 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
690 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
691 return SDValue(); // no match.
694 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
695 // either constant or undef values that are identical for each chunk. See
696 // if these chunks can form into a larger vspltis*.
698 // Check to see if all of the leading entries are either 0 or -1. If
699 // neither, then this won't fit into the immediate field.
700 bool LeadingZero = true;
701 bool LeadingOnes = true;
702 for (unsigned i = 0; i != Multiple-1; ++i) {
703 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
705 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
706 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
708 // Finally, check the least significant entry.
710 if (UniquedVals[Multiple-1].getNode() == 0)
711 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
712 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
714 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
717 if (UniquedVals[Multiple-1].getNode() == 0)
718 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
719 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
720 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
721 return DAG.getTargetConstant(Val, MVT::i32);
727 // Check to see if this buildvec has a single non-undef value in its elements.
728 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
729 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
730 if (OpVal.getNode() == 0)
731 OpVal = N->getOperand(i);
732 else if (OpVal != N->getOperand(i))
736 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
738 unsigned ValSizeInBytes = EltSize;
740 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
741 Value = CN->getZExtValue();
742 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
743 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
744 Value = FloatToBits(CN->getValueAPF().convertToFloat());
747 // If the splat value is larger than the element value, then we can never do
748 // this splat. The only case that we could fit the replicated bits into our
749 // immediate field for would be zero, and we prefer to use vxor for it.
750 if (ValSizeInBytes < ByteSize) return SDValue();
752 // If the element value is larger than the splat value, cut it in half and
753 // check to see if the two halves are equal. Continue doing this until we
754 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
755 while (ValSizeInBytes > ByteSize) {
756 ValSizeInBytes >>= 1;
758 // If the top half equals the bottom half, we're still ok.
759 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
760 (Value & ((1 << (8*ValSizeInBytes))-1)))
764 // Properly sign extend the value.
765 int ShAmt = (4-ByteSize)*8;
766 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
768 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
769 if (MaskVal == 0) return SDValue();
771 // Finally, if this value fits in a 5 bit sext field, return it
772 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
773 return DAG.getTargetConstant(MaskVal, MVT::i32);
777 //===----------------------------------------------------------------------===//
778 // Addressing Mode Selection
779 //===----------------------------------------------------------------------===//
781 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
782 /// or 64-bit immediate, and if the value can be accurately represented as a
783 /// sign extension from a 16-bit value. If so, this returns true and the
785 static bool isIntS16Immediate(SDNode *N, short &Imm) {
786 if (N->getOpcode() != ISD::Constant)
789 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
790 if (N->getValueType(0) == MVT::i32)
791 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
793 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
795 static bool isIntS16Immediate(SDValue Op, short &Imm) {
796 return isIntS16Immediate(Op.getNode(), Imm);
800 /// SelectAddressRegReg - Given the specified addressed, check to see if it
801 /// can be represented as an indexed [r+r] operation. Returns false if it
802 /// can be more efficiently represented with [r+imm].
803 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
805 SelectionDAG &DAG) const {
807 if (N.getOpcode() == ISD::ADD) {
808 if (isIntS16Immediate(N.getOperand(1), imm))
810 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
813 Base = N.getOperand(0);
814 Index = N.getOperand(1);
816 } else if (N.getOpcode() == ISD::OR) {
817 if (isIntS16Immediate(N.getOperand(1), imm))
818 return false; // r+i can fold it if we can.
820 // If this is an or of disjoint bitfields, we can codegen this as an add
821 // (for better address arithmetic) if the LHS and RHS of the OR are provably
823 APInt LHSKnownZero, LHSKnownOne;
824 APInt RHSKnownZero, RHSKnownOne;
825 DAG.ComputeMaskedBits(N.getOperand(0),
826 APInt::getAllOnesValue(N.getOperand(0)
827 .getValueSizeInBits()),
828 LHSKnownZero, LHSKnownOne);
830 if (LHSKnownZero.getBoolValue()) {
831 DAG.ComputeMaskedBits(N.getOperand(1),
832 APInt::getAllOnesValue(N.getOperand(1)
833 .getValueSizeInBits()),
834 RHSKnownZero, RHSKnownOne);
835 // If all of the bits are known zero on the LHS or RHS, the add won't
837 if (~(LHSKnownZero | RHSKnownZero) == 0) {
838 Base = N.getOperand(0);
839 Index = N.getOperand(1);
848 /// Returns true if the address N can be represented by a base register plus
849 /// a signed 16-bit displacement [r+imm], and if it is not better
850 /// represented as reg+reg.
851 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
853 SelectionDAG &DAG) const {
854 // FIXME dl should come from parent load or store, not from address
855 DebugLoc dl = N.getDebugLoc();
856 // If this can be more profitably realized as r+r, fail.
857 if (SelectAddressRegReg(N, Disp, Base, DAG))
860 if (N.getOpcode() == ISD::ADD) {
862 if (isIntS16Immediate(N.getOperand(1), imm)) {
863 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
864 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
865 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
867 Base = N.getOperand(0);
869 return true; // [r+i]
870 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
871 // Match LOAD (ADD (X, Lo(G))).
872 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
873 && "Cannot handle constant offsets yet!");
874 Disp = N.getOperand(1).getOperand(0); // The global address.
875 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
876 Disp.getOpcode() == ISD::TargetConstantPool ||
877 Disp.getOpcode() == ISD::TargetJumpTable);
878 Base = N.getOperand(0);
879 return true; // [&g+r]
881 } else if (N.getOpcode() == ISD::OR) {
883 if (isIntS16Immediate(N.getOperand(1), imm)) {
884 // If this is an or of disjoint bitfields, we can codegen this as an add
885 // (for better address arithmetic) if the LHS and RHS of the OR are
886 // provably disjoint.
887 APInt LHSKnownZero, LHSKnownOne;
888 DAG.ComputeMaskedBits(N.getOperand(0),
889 APInt::getAllOnesValue(N.getOperand(0)
890 .getValueSizeInBits()),
891 LHSKnownZero, LHSKnownOne);
893 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
894 // If all of the bits are known zero on the LHS or RHS, the add won't
896 Base = N.getOperand(0);
897 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
901 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
902 // Loading from a constant address.
904 // If this address fits entirely in a 16-bit sext immediate field, codegen
907 if (isIntS16Immediate(CN, Imm)) {
908 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
909 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
910 CN->getValueType(0));
914 // Handle 32-bit sext immediates with LIS + addr mode.
915 if (CN->getValueType(0) == MVT::i32 ||
916 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
917 int Addr = (int)CN->getZExtValue();
919 // Otherwise, break this down into an LIS + disp.
920 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
922 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
923 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
924 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
929 Disp = DAG.getTargetConstant(0, getPointerTy());
930 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
931 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
934 return true; // [r+0]
937 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
938 /// represented as an indexed [r+r] operation.
939 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
941 SelectionDAG &DAG) const {
942 // Check to see if we can easily represent this as an [r+r] address. This
943 // will fail if it thinks that the address is more profitably represented as
944 // reg+imm, e.g. where imm = 0.
945 if (SelectAddressRegReg(N, Base, Index, DAG))
948 // If the operand is an addition, always emit this as [r+r], since this is
949 // better (for code size, and execution, as the memop does the add for free)
950 // than emitting an explicit add.
951 if (N.getOpcode() == ISD::ADD) {
952 Base = N.getOperand(0);
953 Index = N.getOperand(1);
957 // Otherwise, do it the hard way, using R0 as the base register.
958 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
964 /// SelectAddressRegImmShift - Returns true if the address N can be
965 /// represented by a base register plus a signed 14-bit displacement
966 /// [r+imm*4]. Suitable for use by STD and friends.
967 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
969 SelectionDAG &DAG) const {
970 // FIXME dl should come from the parent load or store, not the address
971 DebugLoc dl = N.getDebugLoc();
972 // If this can be more profitably realized as r+r, fail.
973 if (SelectAddressRegReg(N, Disp, Base, DAG))
976 if (N.getOpcode() == ISD::ADD) {
978 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
979 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
980 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
981 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
983 Base = N.getOperand(0);
985 return true; // [r+i]
986 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
987 // Match LOAD (ADD (X, Lo(G))).
988 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
989 && "Cannot handle constant offsets yet!");
990 Disp = N.getOperand(1).getOperand(0); // The global address.
991 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
992 Disp.getOpcode() == ISD::TargetConstantPool ||
993 Disp.getOpcode() == ISD::TargetJumpTable);
994 Base = N.getOperand(0);
995 return true; // [&g+r]
997 } else if (N.getOpcode() == ISD::OR) {
999 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1000 // If this is an or of disjoint bitfields, we can codegen this as an add
1001 // (for better address arithmetic) if the LHS and RHS of the OR are
1002 // provably disjoint.
1003 APInt LHSKnownZero, LHSKnownOne;
1004 DAG.ComputeMaskedBits(N.getOperand(0),
1005 APInt::getAllOnesValue(N.getOperand(0)
1006 .getValueSizeInBits()),
1007 LHSKnownZero, LHSKnownOne);
1008 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1009 // If all of the bits are known zero on the LHS or RHS, the add won't
1011 Base = N.getOperand(0);
1012 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1016 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1017 // Loading from a constant address. Verify low two bits are clear.
1018 if ((CN->getZExtValue() & 3) == 0) {
1019 // If this address fits entirely in a 14-bit sext immediate field, codegen
1022 if (isIntS16Immediate(CN, Imm)) {
1023 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1024 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1025 CN->getValueType(0));
1029 // Fold the low-part of 32-bit absolute addresses into addr mode.
1030 if (CN->getValueType(0) == MVT::i32 ||
1031 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1032 int Addr = (int)CN->getZExtValue();
1034 // Otherwise, break this down into an LIS + disp.
1035 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1036 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1037 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1038 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1044 Disp = DAG.getTargetConstant(0, getPointerTy());
1045 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1046 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1049 return true; // [r+0]
1053 /// getPreIndexedAddressParts - returns true by value, base pointer and
1054 /// offset pointer and addressing mode by reference if the node's address
1055 /// can be legally represented as pre-indexed load / store address.
1056 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1058 ISD::MemIndexedMode &AM,
1059 SelectionDAG &DAG) const {
1060 // Disabled by default for now.
1061 if (!EnablePPCPreinc) return false;
1065 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1066 Ptr = LD->getBasePtr();
1067 VT = LD->getMemoryVT();
1069 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1070 Ptr = ST->getBasePtr();
1071 VT = ST->getMemoryVT();
1075 // PowerPC doesn't have preinc load/store instructions for vectors.
1079 // TODO: Check reg+reg first.
1081 // LDU/STU use reg+imm*4, others use reg+imm.
1082 if (VT != MVT::i64) {
1084 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1088 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1092 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1093 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1094 // sext i32 to i64 when addr mode is r+i.
1095 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1096 LD->getExtensionType() == ISD::SEXTLOAD &&
1097 isa<ConstantSDNode>(Offset))
1105 //===----------------------------------------------------------------------===//
1106 // LowerOperation implementation
1107 //===----------------------------------------------------------------------===//
1109 /// GetLabelAccessInfo - Return true if we should reference labels using a
1110 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1111 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1112 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1113 HiOpFlags = PPCII::MO_HA16;
1114 LoOpFlags = PPCII::MO_LO16;
1116 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1117 // non-darwin platform. We don't support PIC on other platforms yet.
1118 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1119 TM.getSubtarget<PPCSubtarget>().isDarwin();
1121 HiOpFlags |= PPCII::MO_PIC_FLAG;
1122 LoOpFlags |= PPCII::MO_PIC_FLAG;
1125 // If this is a reference to a global value that requires a non-lazy-ptr, make
1126 // sure that instruction lowering adds it.
1127 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1128 HiOpFlags |= PPCII::MO_NLP_FLAG;
1129 LoOpFlags |= PPCII::MO_NLP_FLAG;
1131 if (GV->hasHiddenVisibility()) {
1132 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1133 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1140 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1141 SelectionDAG &DAG) {
1142 EVT PtrVT = HiPart.getValueType();
1143 SDValue Zero = DAG.getConstant(0, PtrVT);
1144 DebugLoc DL = HiPart.getDebugLoc();
1146 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1147 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1149 // With PIC, the first instruction is actually "GR+hi(&G)".
1151 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1152 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1154 // Generate non-pic code that has direct accesses to the constant pool.
1155 // The address of the global is just (hi(&g)+lo(&g)).
1156 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1159 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1160 SelectionDAG &DAG) const {
1161 EVT PtrVT = Op.getValueType();
1162 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1163 const Constant *C = CP->getConstVal();
1165 unsigned MOHiFlag, MOLoFlag;
1166 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1168 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1170 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1171 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1174 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1175 EVT PtrVT = Op.getValueType();
1176 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1178 unsigned MOHiFlag, MOLoFlag;
1179 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1180 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1181 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1182 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1185 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1186 SelectionDAG &DAG) const {
1187 EVT PtrVT = Op.getValueType();
1189 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1191 unsigned MOHiFlag, MOLoFlag;
1192 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1193 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1194 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1195 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1198 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1199 SelectionDAG &DAG) const {
1200 EVT PtrVT = Op.getValueType();
1201 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1202 DebugLoc DL = GSDN->getDebugLoc();
1203 const GlobalValue *GV = GSDN->getGlobal();
1205 // 64-bit SVR4 ABI code is always position-independent.
1206 // The actual address of the GlobalValue is stored in the TOC.
1207 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1208 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1209 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1210 DAG.getRegister(PPC::X2, MVT::i64));
1213 unsigned MOHiFlag, MOLoFlag;
1214 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1217 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1219 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1221 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1223 // If the global reference is actually to a non-lazy-pointer, we have to do an
1224 // extra load to get the address of the global.
1225 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1226 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1227 false, false, false, 0);
1231 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1232 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1233 DebugLoc dl = Op.getDebugLoc();
1235 // If we're comparing for equality to zero, expose the fact that this is
1236 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1237 // fold the new nodes.
1238 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1239 if (C->isNullValue() && CC == ISD::SETEQ) {
1240 EVT VT = Op.getOperand(0).getValueType();
1241 SDValue Zext = Op.getOperand(0);
1242 if (VT.bitsLT(MVT::i32)) {
1244 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1246 unsigned Log2b = Log2_32(VT.getSizeInBits());
1247 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1248 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1249 DAG.getConstant(Log2b, MVT::i32));
1250 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1252 // Leave comparisons against 0 and -1 alone for now, since they're usually
1253 // optimized. FIXME: revisit this when we can custom lower all setcc
1255 if (C->isAllOnesValue() || C->isNullValue())
1259 // If we have an integer seteq/setne, turn it into a compare against zero
1260 // by xor'ing the rhs with the lhs, which is faster than setting a
1261 // condition register, reading it back out, and masking the correct bit. The
1262 // normal approach here uses sub to do this instead of xor. Using xor exposes
1263 // the result to other bit-twiddling opportunities.
1264 EVT LHSVT = Op.getOperand(0).getValueType();
1265 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1266 EVT VT = Op.getValueType();
1267 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1269 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1274 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1275 const PPCSubtarget &Subtarget) const {
1276 SDNode *Node = Op.getNode();
1277 EVT VT = Node->getValueType(0);
1278 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1279 SDValue InChain = Node->getOperand(0);
1280 SDValue VAListPtr = Node->getOperand(1);
1281 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1282 DebugLoc dl = Node->getDebugLoc();
1284 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1287 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1288 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1290 InChain = GprIndex.getValue(1);
1292 if (VT == MVT::i64) {
1293 // Check if GprIndex is even
1294 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1295 DAG.getConstant(1, MVT::i32));
1296 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1297 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1298 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1299 DAG.getConstant(1, MVT::i32));
1300 // Align GprIndex to be even if it isn't
1301 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1305 // fpr index is 1 byte after gpr
1306 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1307 DAG.getConstant(1, MVT::i32));
1310 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1311 FprPtr, MachinePointerInfo(SV), MVT::i8,
1313 InChain = FprIndex.getValue(1);
1315 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1316 DAG.getConstant(8, MVT::i32));
1318 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1319 DAG.getConstant(4, MVT::i32));
1322 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1323 MachinePointerInfo(), false, false,
1325 InChain = OverflowArea.getValue(1);
1327 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1328 MachinePointerInfo(), false, false,
1330 InChain = RegSaveArea.getValue(1);
1332 // select overflow_area if index > 8
1333 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1334 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1336 // adjustment constant gpr_index * 4/8
1337 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1338 VT.isInteger() ? GprIndex : FprIndex,
1339 DAG.getConstant(VT.isInteger() ? 4 : 8,
1342 // OurReg = RegSaveArea + RegConstant
1343 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1346 // Floating types are 32 bytes into RegSaveArea
1347 if (VT.isFloatingPoint())
1348 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1349 DAG.getConstant(32, MVT::i32));
1351 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1352 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1353 VT.isInteger() ? GprIndex : FprIndex,
1354 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1357 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1358 VT.isInteger() ? VAListPtr : FprPtr,
1359 MachinePointerInfo(SV),
1360 MVT::i8, false, false, 0);
1362 // determine if we should load from reg_save_area or overflow_area
1363 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1365 // increase overflow_area by 4/8 if gpr/fpr > 8
1366 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1367 DAG.getConstant(VT.isInteger() ? 4 : 8,
1370 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1373 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1375 MachinePointerInfo(),
1376 MVT::i32, false, false, 0);
1378 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1379 false, false, false, 0);
1382 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1383 SelectionDAG &DAG) const {
1384 return Op.getOperand(0);
1387 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1388 SelectionDAG &DAG) const {
1389 SDValue Chain = Op.getOperand(0);
1390 SDValue Trmp = Op.getOperand(1); // trampoline
1391 SDValue FPtr = Op.getOperand(2); // nested function
1392 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1393 DebugLoc dl = Op.getDebugLoc();
1395 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1396 bool isPPC64 = (PtrVT == MVT::i64);
1398 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1401 TargetLowering::ArgListTy Args;
1402 TargetLowering::ArgListEntry Entry;
1404 Entry.Ty = IntPtrTy;
1405 Entry.Node = Trmp; Args.push_back(Entry);
1407 // TrampSize == (isPPC64 ? 48 : 40);
1408 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1409 isPPC64 ? MVT::i64 : MVT::i32);
1410 Args.push_back(Entry);
1412 Entry.Node = FPtr; Args.push_back(Entry);
1413 Entry.Node = Nest; Args.push_back(Entry);
1415 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1416 std::pair<SDValue, SDValue> CallResult =
1417 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
1418 false, false, false, false, 0, CallingConv::C, false,
1419 /*isReturnValueUsed=*/true,
1420 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1423 return CallResult.second;
1426 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1427 const PPCSubtarget &Subtarget) const {
1428 MachineFunction &MF = DAG.getMachineFunction();
1429 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1431 DebugLoc dl = Op.getDebugLoc();
1433 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1434 // vastart just stores the address of the VarArgsFrameIndex slot into the
1435 // memory location argument.
1436 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1437 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1438 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1439 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1440 MachinePointerInfo(SV),
1444 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1445 // We suppose the given va_list is already allocated.
1448 // char gpr; /* index into the array of 8 GPRs
1449 // * stored in the register save area
1450 // * gpr=0 corresponds to r3,
1451 // * gpr=1 to r4, etc.
1453 // char fpr; /* index into the array of 8 FPRs
1454 // * stored in the register save area
1455 // * fpr=0 corresponds to f1,
1456 // * fpr=1 to f2, etc.
1458 // char *overflow_arg_area;
1459 // /* location on stack that holds
1460 // * the next overflow argument
1462 // char *reg_save_area;
1463 // /* where r3:r10 and f1:f8 (if saved)
1469 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1470 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1473 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1475 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1477 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1480 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1481 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1483 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1484 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1486 uint64_t FPROffset = 1;
1487 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1489 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1491 // Store first byte : number of int regs
1492 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1494 MachinePointerInfo(SV),
1495 MVT::i8, false, false, 0);
1496 uint64_t nextOffset = FPROffset;
1497 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1500 // Store second byte : number of float regs
1501 SDValue secondStore =
1502 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1503 MachinePointerInfo(SV, nextOffset), MVT::i8,
1505 nextOffset += StackOffset;
1506 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1508 // Store second word : arguments given on stack
1509 SDValue thirdStore =
1510 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1511 MachinePointerInfo(SV, nextOffset),
1513 nextOffset += FrameOffset;
1514 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1516 // Store third word : arguments given in registers
1517 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1518 MachinePointerInfo(SV, nextOffset),
1523 #include "PPCGenCallingConv.inc"
1525 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1526 CCValAssign::LocInfo &LocInfo,
1527 ISD::ArgFlagsTy &ArgFlags,
1532 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1534 CCValAssign::LocInfo &LocInfo,
1535 ISD::ArgFlagsTy &ArgFlags,
1537 static const unsigned ArgRegs[] = {
1538 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1539 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1541 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1543 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1545 // Skip one register if the first unallocated register has an even register
1546 // number and there are still argument registers available which have not been
1547 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1548 // need to skip a register if RegNum is odd.
1549 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1550 State.AllocateReg(ArgRegs[RegNum]);
1553 // Always return false here, as this function only makes sure that the first
1554 // unallocated register has an odd register number and does not actually
1555 // allocate a register for the current argument.
1559 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1561 CCValAssign::LocInfo &LocInfo,
1562 ISD::ArgFlagsTy &ArgFlags,
1564 static const unsigned ArgRegs[] = {
1565 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1569 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1571 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1573 // If there is only one Floating-point register left we need to put both f64
1574 // values of a split ppc_fp128 value on the stack.
1575 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1576 State.AllocateReg(ArgRegs[RegNum]);
1579 // Always return false here, as this function only makes sure that the two f64
1580 // values a ppc_fp128 value is split into are both passed in registers or both
1581 // passed on the stack and does not actually allocate a register for the
1582 // current argument.
1586 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1588 static const unsigned *GetFPR() {
1589 static const unsigned FPR[] = {
1590 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1591 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1597 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1599 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1600 unsigned PtrByteSize) {
1601 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1602 if (Flags.isByVal())
1603 ArgSize = Flags.getByValSize();
1604 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1610 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1611 CallingConv::ID CallConv, bool isVarArg,
1612 const SmallVectorImpl<ISD::InputArg>
1614 DebugLoc dl, SelectionDAG &DAG,
1615 SmallVectorImpl<SDValue> &InVals)
1617 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1618 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1621 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1627 PPCTargetLowering::LowerFormalArguments_SVR4(
1629 CallingConv::ID CallConv, bool isVarArg,
1630 const SmallVectorImpl<ISD::InputArg>
1632 DebugLoc dl, SelectionDAG &DAG,
1633 SmallVectorImpl<SDValue> &InVals) const {
1635 // 32-bit SVR4 ABI Stack Frame Layout:
1636 // +-----------------------------------+
1637 // +--> | Back chain |
1638 // | +-----------------------------------+
1639 // | | Floating-point register save area |
1640 // | +-----------------------------------+
1641 // | | General register save area |
1642 // | +-----------------------------------+
1643 // | | CR save word |
1644 // | +-----------------------------------+
1645 // | | VRSAVE save word |
1646 // | +-----------------------------------+
1647 // | | Alignment padding |
1648 // | +-----------------------------------+
1649 // | | Vector register save area |
1650 // | +-----------------------------------+
1651 // | | Local variable space |
1652 // | +-----------------------------------+
1653 // | | Parameter list area |
1654 // | +-----------------------------------+
1655 // | | LR save word |
1656 // | +-----------------------------------+
1657 // SP--> +--- | Back chain |
1658 // +-----------------------------------+
1661 // System V Application Binary Interface PowerPC Processor Supplement
1662 // AltiVec Technology Programming Interface Manual
1664 MachineFunction &MF = DAG.getMachineFunction();
1665 MachineFrameInfo *MFI = MF.getFrameInfo();
1666 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1668 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1669 // Potential tail calls could cause overwriting of argument stack slots.
1670 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1671 unsigned PtrByteSize = 4;
1673 // Assign locations to all of the incoming arguments.
1674 SmallVector<CCValAssign, 16> ArgLocs;
1675 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1676 getTargetMachine(), ArgLocs, *DAG.getContext());
1678 // Reserve space for the linkage area on the stack.
1679 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1681 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1683 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1684 CCValAssign &VA = ArgLocs[i];
1686 // Arguments stored in registers.
1687 if (VA.isRegLoc()) {
1688 TargetRegisterClass *RC;
1689 EVT ValVT = VA.getValVT();
1691 switch (ValVT.getSimpleVT().SimpleTy) {
1693 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1695 RC = PPC::GPRCRegisterClass;
1698 RC = PPC::F4RCRegisterClass;
1701 RC = PPC::F8RCRegisterClass;
1707 RC = PPC::VRRCRegisterClass;
1711 // Transform the arguments stored in physical registers into virtual ones.
1712 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1713 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1715 InVals.push_back(ArgValue);
1717 // Argument stored in memory.
1718 assert(VA.isMemLoc());
1720 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1721 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1724 // Create load nodes to retrieve arguments from the stack.
1725 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1726 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1727 MachinePointerInfo(),
1728 false, false, false, 0));
1732 // Assign locations to all of the incoming aggregate by value arguments.
1733 // Aggregates passed by value are stored in the local variable space of the
1734 // caller's stack frame, right above the parameter list area.
1735 SmallVector<CCValAssign, 16> ByValArgLocs;
1736 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1737 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1739 // Reserve stack space for the allocations in CCInfo.
1740 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1742 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1744 // Area that is at least reserved in the caller of this function.
1745 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1747 // Set the size that is at least reserved in caller of this function. Tail
1748 // call optimized function's reserved stack space needs to be aligned so that
1749 // taking the difference between two stack areas will result in an aligned
1751 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1754 std::max(MinReservedArea,
1755 PPCFrameLowering::getMinCallFrameSize(false, false));
1757 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1758 getStackAlignment();
1759 unsigned AlignMask = TargetAlign-1;
1760 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1762 FI->setMinReservedArea(MinReservedArea);
1764 SmallVector<SDValue, 8> MemOps;
1766 // If the function takes variable number of arguments, make a frame index for
1767 // the start of the first vararg value... for expansion of llvm.va_start.
1769 static const unsigned GPArgRegs[] = {
1770 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1771 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1773 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1775 static const unsigned FPArgRegs[] = {
1776 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1779 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1781 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1783 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1786 // Make room for NumGPArgRegs and NumFPArgRegs.
1787 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1788 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1790 FuncInfo->setVarArgsStackOffset(
1791 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1792 CCInfo.getNextStackOffset(), true));
1794 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1795 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1797 // The fixed integer arguments of a variadic function are stored to the
1798 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1799 // the result of va_next.
1800 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1801 // Get an existing live-in vreg, or add a new one.
1802 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1804 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1806 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1807 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1808 MachinePointerInfo(), false, false, 0);
1809 MemOps.push_back(Store);
1810 // Increment the address by four for the next argument to store
1811 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1812 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1815 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1817 // The double arguments are stored to the VarArgsFrameIndex
1819 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1820 // Get an existing live-in vreg, or add a new one.
1821 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1823 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1825 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1826 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1827 MachinePointerInfo(), false, false, 0);
1828 MemOps.push_back(Store);
1829 // Increment the address by eight for the next argument to store
1830 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1832 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1836 if (!MemOps.empty())
1837 Chain = DAG.getNode(ISD::TokenFactor, dl,
1838 MVT::Other, &MemOps[0], MemOps.size());
1844 PPCTargetLowering::LowerFormalArguments_Darwin(
1846 CallingConv::ID CallConv, bool isVarArg,
1847 const SmallVectorImpl<ISD::InputArg>
1849 DebugLoc dl, SelectionDAG &DAG,
1850 SmallVectorImpl<SDValue> &InVals) const {
1851 // TODO: add description of PPC stack frame format, or at least some docs.
1853 MachineFunction &MF = DAG.getMachineFunction();
1854 MachineFrameInfo *MFI = MF.getFrameInfo();
1855 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1857 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1858 bool isPPC64 = PtrVT == MVT::i64;
1859 // Potential tail calls could cause overwriting of argument stack slots.
1860 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1861 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1863 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1864 // Area that is at least reserved in caller of this function.
1865 unsigned MinReservedArea = ArgOffset;
1867 static const unsigned GPR_32[] = { // 32-bit registers.
1868 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1869 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1871 static const unsigned GPR_64[] = { // 64-bit registers.
1872 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1873 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1876 static const unsigned *FPR = GetFPR();
1878 static const unsigned VR[] = {
1879 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1880 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1883 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1884 const unsigned Num_FPR_Regs = 13;
1885 const unsigned Num_VR_Regs = array_lengthof( VR);
1887 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1889 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1891 // In 32-bit non-varargs functions, the stack space for vectors is after the
1892 // stack space for non-vectors. We do not use this space unless we have
1893 // too many vectors to fit in registers, something that only occurs in
1894 // constructed examples:), but we have to walk the arglist to figure
1895 // that out...for the pathological case, compute VecArgOffset as the
1896 // start of the vector parameter area. Computing VecArgOffset is the
1897 // entire point of the following loop.
1898 unsigned VecArgOffset = ArgOffset;
1899 if (!isVarArg && !isPPC64) {
1900 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1902 EVT ObjectVT = Ins[ArgNo].VT;
1903 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1904 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1906 if (Flags.isByVal()) {
1907 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1908 ObjSize = Flags.getByValSize();
1910 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1911 VecArgOffset += ArgSize;
1915 switch(ObjectVT.getSimpleVT().SimpleTy) {
1916 default: llvm_unreachable("Unhandled argument type!");
1919 VecArgOffset += isPPC64 ? 8 : 4;
1921 case MVT::i64: // PPC64
1929 // Nothing to do, we're only looking at Nonvector args here.
1934 // We've found where the vector parameter area in memory is. Skip the
1935 // first 12 parameters; these don't use that memory.
1936 VecArgOffset = ((VecArgOffset+15)/16)*16;
1937 VecArgOffset += 12*16;
1939 // Add DAG nodes to load the arguments or copy them out of registers. On
1940 // entry to a function on PPC, the arguments start after the linkage area,
1941 // although the first ones are often in registers.
1943 SmallVector<SDValue, 8> MemOps;
1944 unsigned nAltivecParamsAtEnd = 0;
1945 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1947 bool needsLoad = false;
1948 EVT ObjectVT = Ins[ArgNo].VT;
1949 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1950 unsigned ArgSize = ObjSize;
1951 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1953 unsigned CurArgOffset = ArgOffset;
1955 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1956 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1957 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1958 if (isVarArg || isPPC64) {
1959 MinReservedArea = ((MinReservedArea+15)/16)*16;
1960 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1963 } else nAltivecParamsAtEnd++;
1965 // Calculate min reserved area.
1966 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1970 // FIXME the codegen can be much improved in some cases.
1971 // We do not have to keep everything in memory.
1972 if (Flags.isByVal()) {
1973 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1974 ObjSize = Flags.getByValSize();
1975 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1976 // Objects of size 1 and 2 are right justified, everything else is
1977 // left justified. This means the memory address is adjusted forwards.
1978 if (ObjSize==1 || ObjSize==2) {
1979 CurArgOffset = CurArgOffset + (4 - ObjSize);
1981 // The value of the object is its address.
1982 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
1983 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1984 InVals.push_back(FIN);
1985 if (ObjSize==1 || ObjSize==2) {
1986 if (GPR_idx != Num_GPR_Regs) {
1989 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1991 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1992 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1993 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1994 MachinePointerInfo(),
1995 ObjSize==1 ? MVT::i8 : MVT::i16,
1997 MemOps.push_back(Store);
2001 ArgOffset += PtrByteSize;
2005 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2006 // Store whatever pieces of the object are in registers
2007 // to memory. ArgVal will be address of the beginning of
2009 if (GPR_idx != Num_GPR_Regs) {
2012 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2014 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2015 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2016 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2017 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2018 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2019 MachinePointerInfo(),
2021 MemOps.push_back(Store);
2023 ArgOffset += PtrByteSize;
2025 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2032 switch (ObjectVT.getSimpleVT().SimpleTy) {
2033 default: llvm_unreachable("Unhandled argument type!");
2036 if (GPR_idx != Num_GPR_Regs) {
2037 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2038 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2042 ArgSize = PtrByteSize;
2044 // All int arguments reserve stack space in the Darwin ABI.
2045 ArgOffset += PtrByteSize;
2049 case MVT::i64: // PPC64
2050 if (GPR_idx != Num_GPR_Regs) {
2051 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2052 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2054 if (ObjectVT == MVT::i32) {
2055 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2056 // value to MVT::i64 and then truncate to the correct register size.
2058 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2059 DAG.getValueType(ObjectVT));
2060 else if (Flags.isZExt())
2061 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2062 DAG.getValueType(ObjectVT));
2064 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2070 ArgSize = PtrByteSize;
2072 // All int arguments reserve stack space in the Darwin ABI.
2078 // Every 4 bytes of argument space consumes one of the GPRs available for
2079 // argument passing.
2080 if (GPR_idx != Num_GPR_Regs) {
2082 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2085 if (FPR_idx != Num_FPR_Regs) {
2088 if (ObjectVT == MVT::f32)
2089 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2091 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2093 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2099 // All FP arguments reserve stack space in the Darwin ABI.
2100 ArgOffset += isPPC64 ? 8 : ObjSize;
2106 // Note that vector arguments in registers don't reserve stack space,
2107 // except in varargs functions.
2108 if (VR_idx != Num_VR_Regs) {
2109 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2110 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2112 while ((ArgOffset % 16) != 0) {
2113 ArgOffset += PtrByteSize;
2114 if (GPR_idx != Num_GPR_Regs)
2118 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2122 if (!isVarArg && !isPPC64) {
2123 // Vectors go after all the nonvectors.
2124 CurArgOffset = VecArgOffset;
2127 // Vectors are aligned.
2128 ArgOffset = ((ArgOffset+15)/16)*16;
2129 CurArgOffset = ArgOffset;
2137 // We need to load the argument to a virtual register if we determined above
2138 // that we ran out of physical registers of the appropriate type.
2140 int FI = MFI->CreateFixedObject(ObjSize,
2141 CurArgOffset + (ArgSize - ObjSize),
2143 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2144 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2145 false, false, false, 0);
2148 InVals.push_back(ArgVal);
2151 // Set the size that is at least reserved in caller of this function. Tail
2152 // call optimized function's reserved stack space needs to be aligned so that
2153 // taking the difference between two stack areas will result in an aligned
2155 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2156 // Add the Altivec parameters at the end, if needed.
2157 if (nAltivecParamsAtEnd) {
2158 MinReservedArea = ((MinReservedArea+15)/16)*16;
2159 MinReservedArea += 16*nAltivecParamsAtEnd;
2162 std::max(MinReservedArea,
2163 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2164 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2165 getStackAlignment();
2166 unsigned AlignMask = TargetAlign-1;
2167 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2168 FI->setMinReservedArea(MinReservedArea);
2170 // If the function takes variable number of arguments, make a frame index for
2171 // the start of the first vararg value... for expansion of llvm.va_start.
2173 int Depth = ArgOffset;
2175 FuncInfo->setVarArgsFrameIndex(
2176 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2178 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2180 // If this function is vararg, store any remaining integer argument regs
2181 // to their spots on the stack so that they may be loaded by deferencing the
2182 // result of va_next.
2183 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2187 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2189 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2191 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2192 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2193 MachinePointerInfo(), false, false, 0);
2194 MemOps.push_back(Store);
2195 // Increment the address by four for the next argument to store
2196 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2197 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2201 if (!MemOps.empty())
2202 Chain = DAG.getNode(ISD::TokenFactor, dl,
2203 MVT::Other, &MemOps[0], MemOps.size());
2208 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2209 /// linkage area for the Darwin ABI.
2211 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2215 const SmallVectorImpl<ISD::OutputArg>
2217 const SmallVectorImpl<SDValue> &OutVals,
2218 unsigned &nAltivecParamsAtEnd) {
2219 // Count how many bytes are to be pushed on the stack, including the linkage
2220 // area, and parameter passing area. We start with 24/48 bytes, which is
2221 // prereserved space for [SP][CR][LR][3 x unused].
2222 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2223 unsigned NumOps = Outs.size();
2224 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2226 // Add up all the space actually used.
2227 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2228 // they all go in registers, but we must reserve stack space for them for
2229 // possible use by the caller. In varargs or 64-bit calls, parameters are
2230 // assigned stack space in order, with padding so Altivec parameters are
2232 nAltivecParamsAtEnd = 0;
2233 for (unsigned i = 0; i != NumOps; ++i) {
2234 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2235 EVT ArgVT = Outs[i].VT;
2236 // Varargs Altivec parameters are padded to a 16 byte boundary.
2237 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2238 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2239 if (!isVarArg && !isPPC64) {
2240 // Non-varargs Altivec parameters go after all the non-Altivec
2241 // parameters; handle those later so we know how much padding we need.
2242 nAltivecParamsAtEnd++;
2245 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2246 NumBytes = ((NumBytes+15)/16)*16;
2248 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2251 // Allow for Altivec parameters at the end, if needed.
2252 if (nAltivecParamsAtEnd) {
2253 NumBytes = ((NumBytes+15)/16)*16;
2254 NumBytes += 16*nAltivecParamsAtEnd;
2257 // The prolog code of the callee may store up to 8 GPR argument registers to
2258 // the stack, allowing va_start to index over them in memory if its varargs.
2259 // Because we cannot tell if this is needed on the caller side, we have to
2260 // conservatively assume that it is needed. As such, make sure we have at
2261 // least enough stack space for the caller to store the 8 GPRs.
2262 NumBytes = std::max(NumBytes,
2263 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2265 // Tail call needs the stack to be aligned.
2266 if (CC==CallingConv::Fast && GuaranteedTailCallOpt) {
2267 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2268 getStackAlignment();
2269 unsigned AlignMask = TargetAlign-1;
2270 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2276 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2277 /// adjusted to accommodate the arguments for the tailcall.
2278 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2279 unsigned ParamSize) {
2281 if (!isTailCall) return 0;
2283 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2284 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2285 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2286 // Remember only if the new adjustement is bigger.
2287 if (SPDiff < FI->getTailCallSPDelta())
2288 FI->setTailCallSPDelta(SPDiff);
2293 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2294 /// for tail call optimization. Targets which want to do tail call
2295 /// optimization should implement this function.
2297 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2298 CallingConv::ID CalleeCC,
2300 const SmallVectorImpl<ISD::InputArg> &Ins,
2301 SelectionDAG& DAG) const {
2302 if (!GuaranteedTailCallOpt)
2305 // Variable argument functions are not supported.
2309 MachineFunction &MF = DAG.getMachineFunction();
2310 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2311 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2312 // Functions containing by val parameters are not supported.
2313 for (unsigned i = 0; i != Ins.size(); i++) {
2314 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2315 if (Flags.isByVal()) return false;
2318 // Non PIC/GOT tail calls are supported.
2319 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2322 // At the moment we can only do local tail calls (in same module, hidden
2323 // or protected) if we are generating PIC.
2324 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2325 return G->getGlobal()->hasHiddenVisibility()
2326 || G->getGlobal()->hasProtectedVisibility();
2332 /// isCallCompatibleAddress - Return the immediate to use if the specified
2333 /// 32-bit value is representable in the immediate field of a BxA instruction.
2334 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2335 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2338 int Addr = C->getZExtValue();
2339 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2340 (Addr << 6 >> 6) != Addr)
2341 return 0; // Top 6 bits have to be sext of immediate.
2343 return DAG.getConstant((int)C->getZExtValue() >> 2,
2344 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2349 struct TailCallArgumentInfo {
2354 TailCallArgumentInfo() : FrameIdx(0) {}
2359 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2361 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2363 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2364 SmallVector<SDValue, 8> &MemOpChains,
2366 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2367 SDValue Arg = TailCallArgs[i].Arg;
2368 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2369 int FI = TailCallArgs[i].FrameIdx;
2370 // Store relative to framepointer.
2371 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2372 MachinePointerInfo::getFixedStack(FI),
2377 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2378 /// the appropriate stack slot for the tail call optimized function call.
2379 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2380 MachineFunction &MF,
2389 // Calculate the new stack slot for the return address.
2390 int SlotSize = isPPC64 ? 8 : 4;
2391 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2393 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2394 NewRetAddrLoc, true);
2395 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2396 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2397 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2398 MachinePointerInfo::getFixedStack(NewRetAddr),
2401 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2402 // slot as the FP is never overwritten.
2405 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2406 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2408 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2409 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2410 MachinePointerInfo::getFixedStack(NewFPIdx),
2417 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2418 /// the position of the argument.
2420 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2421 SDValue Arg, int SPDiff, unsigned ArgOffset,
2422 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2423 int Offset = ArgOffset + SPDiff;
2424 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2425 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2426 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2427 SDValue FIN = DAG.getFrameIndex(FI, VT);
2428 TailCallArgumentInfo Info;
2430 Info.FrameIdxOp = FIN;
2432 TailCallArguments.push_back(Info);
2435 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2436 /// stack slot. Returns the chain as result and the loaded frame pointers in
2437 /// LROpOut/FPOpout. Used when tail calling.
2438 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2444 DebugLoc dl) const {
2446 // Load the LR and FP stack slot for later adjusting.
2447 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2448 LROpOut = getReturnAddrFrameIndex(DAG);
2449 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2450 false, false, false, 0);
2451 Chain = SDValue(LROpOut.getNode(), 1);
2453 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2454 // slot as the FP is never overwritten.
2456 FPOpOut = getFramePointerFrameIndex(DAG);
2457 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2458 false, false, false, 0);
2459 Chain = SDValue(FPOpOut.getNode(), 1);
2465 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2466 /// by "Src" to address "Dst" of size "Size". Alignment information is
2467 /// specified by the specific parameter attribute. The copy will be passed as
2468 /// a byval function parameter.
2469 /// Sometimes what we are copying is the end of a larger object, the part that
2470 /// does not fit in registers.
2472 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2473 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2475 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2476 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2477 false, false, MachinePointerInfo(0),
2478 MachinePointerInfo(0));
2481 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2484 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2485 SDValue Arg, SDValue PtrOff, int SPDiff,
2486 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2487 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2488 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2490 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2495 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2497 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2498 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2499 DAG.getConstant(ArgOffset, PtrVT));
2501 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2502 MachinePointerInfo(), false, false, 0));
2503 // Calculate and remember argument location.
2504 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2509 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2510 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2511 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2512 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2513 MachineFunction &MF = DAG.getMachineFunction();
2515 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2516 // might overwrite each other in case of tail call optimization.
2517 SmallVector<SDValue, 8> MemOpChains2;
2518 // Do not flag preceding copytoreg stuff together with the following stuff.
2520 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2522 if (!MemOpChains2.empty())
2523 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2524 &MemOpChains2[0], MemOpChains2.size());
2526 // Store the return address to the appropriate stack slot.
2527 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2528 isPPC64, isDarwinABI, dl);
2530 // Emit callseq_end just before tailcall node.
2531 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2532 DAG.getIntPtrConstant(0, true), InFlag);
2533 InFlag = Chain.getValue(1);
2537 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2538 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2539 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2540 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2541 const PPCSubtarget &PPCSubTarget) {
2543 bool isPPC64 = PPCSubTarget.isPPC64();
2544 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2546 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2547 NodeTys.push_back(MVT::Other); // Returns a chain
2548 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2550 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2552 bool needIndirectCall = true;
2553 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2554 // If this is an absolute destination address, use the munged value.
2555 Callee = SDValue(Dest, 0);
2556 needIndirectCall = false;
2559 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2560 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2561 // Use indirect calls for ALL functions calls in JIT mode, since the
2562 // far-call stubs may be outside relocation limits for a BL instruction.
2563 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2564 unsigned OpFlags = 0;
2565 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2566 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2567 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2568 (G->getGlobal()->isDeclaration() ||
2569 G->getGlobal()->isWeakForLinker())) {
2570 // PC-relative references to external symbols should go through $stub,
2571 // unless we're building with the leopard linker or later, which
2572 // automatically synthesizes these stubs.
2573 OpFlags = PPCII::MO_DARWIN_STUB;
2576 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2577 // every direct call is) turn it into a TargetGlobalAddress /
2578 // TargetExternalSymbol node so that legalize doesn't hack it.
2579 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2580 Callee.getValueType(),
2582 needIndirectCall = false;
2586 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2587 unsigned char OpFlags = 0;
2589 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2590 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2591 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2592 // PC-relative references to external symbols should go through $stub,
2593 // unless we're building with the leopard linker or later, which
2594 // automatically synthesizes these stubs.
2595 OpFlags = PPCII::MO_DARWIN_STUB;
2598 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2600 needIndirectCall = false;
2603 if (needIndirectCall) {
2604 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2605 // to do the call, we can't use PPCISD::CALL.
2606 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2608 if (isSVR4ABI && isPPC64) {
2609 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2610 // entry point, but to the function descriptor (the function entry point
2611 // address is part of the function descriptor though).
2612 // The function descriptor is a three doubleword structure with the
2613 // following fields: function entry point, TOC base address and
2614 // environment pointer.
2615 // Thus for a call through a function pointer, the following actions need
2617 // 1. Save the TOC of the caller in the TOC save area of its stack
2618 // frame (this is done in LowerCall_Darwin()).
2619 // 2. Load the address of the function entry point from the function
2621 // 3. Load the TOC of the callee from the function descriptor into r2.
2622 // 4. Load the environment pointer from the function descriptor into
2624 // 5. Branch to the function entry point address.
2625 // 6. On return of the callee, the TOC of the caller needs to be
2626 // restored (this is done in FinishCall()).
2628 // All those operations are flagged together to ensure that no other
2629 // operations can be scheduled in between. E.g. without flagging the
2630 // operations together, a TOC access in the caller could be scheduled
2631 // between the load of the callee TOC and the branch to the callee, which
2632 // results in the TOC access going through the TOC of the callee instead
2633 // of going through the TOC of the caller, which leads to incorrect code.
2635 // Load the address of the function entry point from the function
2637 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2638 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2639 InFlag.getNode() ? 3 : 2);
2640 Chain = LoadFuncPtr.getValue(1);
2641 InFlag = LoadFuncPtr.getValue(2);
2643 // Load environment pointer into r11.
2644 // Offset of the environment pointer within the function descriptor.
2645 SDValue PtrOff = DAG.getIntPtrConstant(16);
2647 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2648 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2650 Chain = LoadEnvPtr.getValue(1);
2651 InFlag = LoadEnvPtr.getValue(2);
2653 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2655 Chain = EnvVal.getValue(0);
2656 InFlag = EnvVal.getValue(1);
2658 // Load TOC of the callee into r2. We are using a target-specific load
2659 // with r2 hard coded, because the result of a target-independent load
2660 // would never go directly into r2, since r2 is a reserved register (which
2661 // prevents the register allocator from allocating it), resulting in an
2662 // additional register being allocated and an unnecessary move instruction
2664 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2665 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2667 Chain = LoadTOCPtr.getValue(0);
2668 InFlag = LoadTOCPtr.getValue(1);
2670 MTCTROps[0] = Chain;
2671 MTCTROps[1] = LoadFuncPtr;
2672 MTCTROps[2] = InFlag;
2675 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2676 2 + (InFlag.getNode() != 0));
2677 InFlag = Chain.getValue(1);
2680 NodeTys.push_back(MVT::Other);
2681 NodeTys.push_back(MVT::Glue);
2682 Ops.push_back(Chain);
2683 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2685 // Add CTR register as callee so a bctr can be emitted later.
2687 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2690 // If this is a direct call, pass the chain and the callee.
2691 if (Callee.getNode()) {
2692 Ops.push_back(Chain);
2693 Ops.push_back(Callee);
2695 // If this is a tail call add stack pointer delta.
2697 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2699 // Add argument registers to the end of the list so that they are known live
2701 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2702 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2703 RegsToPass[i].second.getValueType()));
2709 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2710 CallingConv::ID CallConv, bool isVarArg,
2711 const SmallVectorImpl<ISD::InputArg> &Ins,
2712 DebugLoc dl, SelectionDAG &DAG,
2713 SmallVectorImpl<SDValue> &InVals) const {
2715 SmallVector<CCValAssign, 16> RVLocs;
2716 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2717 getTargetMachine(), RVLocs, *DAG.getContext());
2718 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2720 // Copy all of the result registers out of their specified physreg.
2721 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2722 CCValAssign &VA = RVLocs[i];
2723 EVT VT = VA.getValVT();
2724 assert(VA.isRegLoc() && "Can only return in registers!");
2725 Chain = DAG.getCopyFromReg(Chain, dl,
2726 VA.getLocReg(), VT, InFlag).getValue(1);
2727 InVals.push_back(Chain.getValue(0));
2728 InFlag = Chain.getValue(2);
2735 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2736 bool isTailCall, bool isVarArg,
2738 SmallVector<std::pair<unsigned, SDValue>, 8>
2740 SDValue InFlag, SDValue Chain,
2742 int SPDiff, unsigned NumBytes,
2743 const SmallVectorImpl<ISD::InputArg> &Ins,
2744 SmallVectorImpl<SDValue> &InVals) const {
2745 std::vector<EVT> NodeTys;
2746 SmallVector<SDValue, 8> Ops;
2747 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2748 isTailCall, RegsToPass, Ops, NodeTys,
2751 // When performing tail call optimization the callee pops its arguments off
2752 // the stack. Account for this here so these bytes can be pushed back on in
2753 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2754 int BytesCalleePops =
2755 (CallConv==CallingConv::Fast && GuaranteedTailCallOpt) ? NumBytes : 0;
2757 if (InFlag.getNode())
2758 Ops.push_back(InFlag);
2762 // If this is the first return lowered for this function, add the regs
2763 // to the liveout set for the function.
2764 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2765 SmallVector<CCValAssign, 16> RVLocs;
2766 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2767 getTargetMachine(), RVLocs, *DAG.getContext());
2768 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2769 for (unsigned i = 0; i != RVLocs.size(); ++i)
2770 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2773 assert(((Callee.getOpcode() == ISD::Register &&
2774 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2775 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2776 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2777 isa<ConstantSDNode>(Callee)) &&
2778 "Expecting an global address, external symbol, absolute value or register");
2780 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2783 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2784 InFlag = Chain.getValue(1);
2786 // Add a NOP immediately after the branch instruction when using the 64-bit
2787 // SVR4 ABI. At link time, if caller and callee are in a different module and
2788 // thus have a different TOC, the call will be replaced with a call to a stub
2789 // function which saves the current TOC, loads the TOC of the callee and
2790 // branches to the callee. The NOP will be replaced with a load instruction
2791 // which restores the TOC of the caller from the TOC save slot of the current
2792 // stack frame. If caller and callee belong to the same module (and have the
2793 // same TOC), the NOP will remain unchanged.
2794 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2795 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2796 if (CallOpc == PPCISD::BCTRL_SVR4) {
2797 // This is a call through a function pointer.
2798 // Restore the caller TOC from the save area into R2.
2799 // See PrepareCall() for more information about calls through function
2800 // pointers in the 64-bit SVR4 ABI.
2801 // We are using a target-specific load with r2 hard coded, because the
2802 // result of a target-independent load would never go directly into r2,
2803 // since r2 is a reserved register (which prevents the register allocator
2804 // from allocating it), resulting in an additional register being
2805 // allocated and an unnecessary move instruction being generated.
2806 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2807 InFlag = Chain.getValue(1);
2809 // Otherwise insert NOP.
2810 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Glue, InFlag);
2814 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2815 DAG.getIntPtrConstant(BytesCalleePops, true),
2818 InFlag = Chain.getValue(1);
2820 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2821 Ins, dl, DAG, InVals);
2825 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2826 CallingConv::ID CallConv, bool isVarArg,
2828 const SmallVectorImpl<ISD::OutputArg> &Outs,
2829 const SmallVectorImpl<SDValue> &OutVals,
2830 const SmallVectorImpl<ISD::InputArg> &Ins,
2831 DebugLoc dl, SelectionDAG &DAG,
2832 SmallVectorImpl<SDValue> &InVals) const {
2834 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2837 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2838 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2839 isTailCall, Outs, OutVals, Ins,
2842 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2843 isTailCall, Outs, OutVals, Ins,
2848 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2849 CallingConv::ID CallConv, bool isVarArg,
2851 const SmallVectorImpl<ISD::OutputArg> &Outs,
2852 const SmallVectorImpl<SDValue> &OutVals,
2853 const SmallVectorImpl<ISD::InputArg> &Ins,
2854 DebugLoc dl, SelectionDAG &DAG,
2855 SmallVectorImpl<SDValue> &InVals) const {
2856 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2857 // of the 32-bit SVR4 ABI stack frame layout.
2859 assert((CallConv == CallingConv::C ||
2860 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2862 unsigned PtrByteSize = 4;
2864 MachineFunction &MF = DAG.getMachineFunction();
2866 // Mark this function as potentially containing a function that contains a
2867 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2868 // and restoring the callers stack pointer in this functions epilog. This is
2869 // done because by tail calling the called function might overwrite the value
2870 // in this function's (MF) stack pointer stack slot 0(SP).
2871 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2872 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2874 // Count how many bytes are to be pushed on the stack, including the linkage
2875 // area, parameter list area and the part of the local variable space which
2876 // contains copies of aggregates which are passed by value.
2878 // Assign locations to all of the outgoing arguments.
2879 SmallVector<CCValAssign, 16> ArgLocs;
2880 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2881 getTargetMachine(), ArgLocs, *DAG.getContext());
2883 // Reserve space for the linkage area on the stack.
2884 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2887 // Handle fixed and variable vector arguments differently.
2888 // Fixed vector arguments go into registers as long as registers are
2889 // available. Variable vector arguments always go into memory.
2890 unsigned NumArgs = Outs.size();
2892 for (unsigned i = 0; i != NumArgs; ++i) {
2893 MVT ArgVT = Outs[i].VT;
2894 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2897 if (Outs[i].IsFixed) {
2898 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2901 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2907 errs() << "Call operand #" << i << " has unhandled type "
2908 << EVT(ArgVT).getEVTString() << "\n";
2910 llvm_unreachable(0);
2914 // All arguments are treated the same.
2915 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2918 // Assign locations to all of the outgoing aggregate by value arguments.
2919 SmallVector<CCValAssign, 16> ByValArgLocs;
2920 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2921 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2923 // Reserve stack space for the allocations in CCInfo.
2924 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2926 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2928 // Size of the linkage area, parameter list area and the part of the local
2929 // space variable where copies of aggregates which are passed by value are
2931 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2933 // Calculate by how many bytes the stack has to be adjusted in case of tail
2934 // call optimization.
2935 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2937 // Adjust the stack pointer for the new arguments...
2938 // These operations are automatically eliminated by the prolog/epilog pass
2939 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2940 SDValue CallSeqStart = Chain;
2942 // Load the return address and frame pointer so it can be moved somewhere else
2945 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2948 // Set up a copy of the stack pointer for use loading and storing any
2949 // arguments that may not fit in the registers available for argument
2951 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2953 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2954 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2955 SmallVector<SDValue, 8> MemOpChains;
2957 bool seenFloatArg = false;
2958 // Walk the register/memloc assignments, inserting copies/loads.
2959 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2962 CCValAssign &VA = ArgLocs[i];
2963 SDValue Arg = OutVals[i];
2964 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2966 if (Flags.isByVal()) {
2967 // Argument is an aggregate which is passed by value, thus we need to
2968 // create a copy of it in the local variable space of the current stack
2969 // frame (which is the stack frame of the caller) and pass the address of
2970 // this copy to the callee.
2971 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2972 CCValAssign &ByValVA = ByValArgLocs[j++];
2973 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2975 // Memory reserved in the local variable space of the callers stack frame.
2976 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2978 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2979 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2981 // Create a copy of the argument in the local area of the current
2983 SDValue MemcpyCall =
2984 CreateCopyOfByValArgument(Arg, PtrOff,
2985 CallSeqStart.getNode()->getOperand(0),
2988 // This must go outside the CALLSEQ_START..END.
2989 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2990 CallSeqStart.getNode()->getOperand(1));
2991 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2992 NewCallSeqStart.getNode());
2993 Chain = CallSeqStart = NewCallSeqStart;
2995 // Pass the address of the aggregate copy on the stack either in a
2996 // physical register or in the parameter list area of the current stack
2997 // frame to the callee.
3001 if (VA.isRegLoc()) {
3002 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3003 // Put argument in a physical register.
3004 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3006 // Put argument in the parameter list area of the current stack frame.
3007 assert(VA.isMemLoc());
3008 unsigned LocMemOffset = VA.getLocMemOffset();
3011 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3012 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3014 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3015 MachinePointerInfo(),
3018 // Calculate and remember argument location.
3019 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3025 if (!MemOpChains.empty())
3026 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3027 &MemOpChains[0], MemOpChains.size());
3029 // Set CR6 to true if this is a vararg call with floating args passed in
3032 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3034 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3037 // Build a sequence of copy-to-reg nodes chained together with token chain
3038 // and flag operands which copy the outgoing args into the appropriate regs.
3040 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3041 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3042 RegsToPass[i].second, InFlag);
3043 InFlag = Chain.getValue(1);
3047 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3048 false, TailCallArguments);
3050 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3051 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3056 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3057 CallingConv::ID CallConv, bool isVarArg,
3059 const SmallVectorImpl<ISD::OutputArg> &Outs,
3060 const SmallVectorImpl<SDValue> &OutVals,
3061 const SmallVectorImpl<ISD::InputArg> &Ins,
3062 DebugLoc dl, SelectionDAG &DAG,
3063 SmallVectorImpl<SDValue> &InVals) const {
3065 unsigned NumOps = Outs.size();
3067 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3068 bool isPPC64 = PtrVT == MVT::i64;
3069 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3071 MachineFunction &MF = DAG.getMachineFunction();
3073 // Mark this function as potentially containing a function that contains a
3074 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3075 // and restoring the callers stack pointer in this functions epilog. This is
3076 // done because by tail calling the called function might overwrite the value
3077 // in this function's (MF) stack pointer stack slot 0(SP).
3078 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
3079 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3081 unsigned nAltivecParamsAtEnd = 0;
3083 // Count how many bytes are to be pushed on the stack, including the linkage
3084 // area, and parameter passing area. We start with 24/48 bytes, which is
3085 // prereserved space for [SP][CR][LR][3 x unused].
3087 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3089 nAltivecParamsAtEnd);
3091 // Calculate by how many bytes the stack has to be adjusted in case of tail
3092 // call optimization.
3093 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3095 // To protect arguments on the stack from being clobbered in a tail call,
3096 // force all the loads to happen before doing any other lowering.
3098 Chain = DAG.getStackArgumentTokenFactor(Chain);
3100 // Adjust the stack pointer for the new arguments...
3101 // These operations are automatically eliminated by the prolog/epilog pass
3102 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3103 SDValue CallSeqStart = Chain;
3105 // Load the return address and frame pointer so it can be move somewhere else
3108 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3111 // Set up a copy of the stack pointer for use loading and storing any
3112 // arguments that may not fit in the registers available for argument
3116 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3118 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3120 // Figure out which arguments are going to go in registers, and which in
3121 // memory. Also, if this is a vararg function, floating point operations
3122 // must be stored to our stack, and loaded into integer regs as well, if
3123 // any integer regs are available for argument passing.
3124 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3125 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3127 static const unsigned GPR_32[] = { // 32-bit registers.
3128 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3129 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3131 static const unsigned GPR_64[] = { // 64-bit registers.
3132 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3133 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3135 static const unsigned *FPR = GetFPR();
3137 static const unsigned VR[] = {
3138 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3139 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3141 const unsigned NumGPRs = array_lengthof(GPR_32);
3142 const unsigned NumFPRs = 13;
3143 const unsigned NumVRs = array_lengthof(VR);
3145 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
3147 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3148 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3150 SmallVector<SDValue, 8> MemOpChains;
3151 for (unsigned i = 0; i != NumOps; ++i) {
3152 SDValue Arg = OutVals[i];
3153 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3155 // PtrOff will be used to store the current argument to the stack if a
3156 // register cannot be found for it.
3159 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3161 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3163 // On PPC64, promote integers to 64-bit values.
3164 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3165 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3166 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3167 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3170 // FIXME memcpy is used way more than necessary. Correctness first.
3171 if (Flags.isByVal()) {
3172 unsigned Size = Flags.getByValSize();
3173 if (Size==1 || Size==2) {
3174 // Very small objects are passed right-justified.
3175 // Everything else is passed left-justified.
3176 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3177 if (GPR_idx != NumGPRs) {
3178 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3179 MachinePointerInfo(), VT,
3181 MemOpChains.push_back(Load.getValue(1));
3182 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3184 ArgOffset += PtrByteSize;
3186 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3187 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3188 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3189 CallSeqStart.getNode()->getOperand(0),
3191 // This must go outside the CALLSEQ_START..END.
3192 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3193 CallSeqStart.getNode()->getOperand(1));
3194 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3195 NewCallSeqStart.getNode());
3196 Chain = CallSeqStart = NewCallSeqStart;
3197 ArgOffset += PtrByteSize;
3201 // Copy entire object into memory. There are cases where gcc-generated
3202 // code assumes it is there, even if it could be put entirely into
3203 // registers. (This is not what the doc says.)
3204 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3205 CallSeqStart.getNode()->getOperand(0),
3207 // This must go outside the CALLSEQ_START..END.
3208 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3209 CallSeqStart.getNode()->getOperand(1));
3210 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3211 Chain = CallSeqStart = NewCallSeqStart;
3212 // And copy the pieces of it that fit into registers.
3213 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3214 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3215 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3216 if (GPR_idx != NumGPRs) {
3217 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3218 MachinePointerInfo(),
3219 false, false, false, 0);
3220 MemOpChains.push_back(Load.getValue(1));
3221 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3222 ArgOffset += PtrByteSize;
3224 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3231 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3232 default: llvm_unreachable("Unexpected ValueType for argument!");
3235 if (GPR_idx != NumGPRs) {
3236 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3238 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3239 isPPC64, isTailCall, false, MemOpChains,
3240 TailCallArguments, dl);
3242 ArgOffset += PtrByteSize;
3246 if (FPR_idx != NumFPRs) {
3247 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3250 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3251 MachinePointerInfo(), false, false, 0);
3252 MemOpChains.push_back(Store);
3254 // Float varargs are always shadowed in available integer registers
3255 if (GPR_idx != NumGPRs) {
3256 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3257 MachinePointerInfo(), false, false,
3259 MemOpChains.push_back(Load.getValue(1));
3260 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3262 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3263 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3264 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3265 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3266 MachinePointerInfo(),
3267 false, false, false, 0);
3268 MemOpChains.push_back(Load.getValue(1));
3269 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3272 // If we have any FPRs remaining, we may also have GPRs remaining.
3273 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3275 if (GPR_idx != NumGPRs)
3277 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3278 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3282 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3283 isPPC64, isTailCall, false, MemOpChains,
3284 TailCallArguments, dl);
3289 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3296 // These go aligned on the stack, or in the corresponding R registers
3297 // when within range. The Darwin PPC ABI doc claims they also go in
3298 // V registers; in fact gcc does this only for arguments that are
3299 // prototyped, not for those that match the ... We do it for all
3300 // arguments, seems to work.
3301 while (ArgOffset % 16 !=0) {
3302 ArgOffset += PtrByteSize;
3303 if (GPR_idx != NumGPRs)
3306 // We could elide this store in the case where the object fits
3307 // entirely in R registers. Maybe later.
3308 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3309 DAG.getConstant(ArgOffset, PtrVT));
3310 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3311 MachinePointerInfo(), false, false, 0);
3312 MemOpChains.push_back(Store);
3313 if (VR_idx != NumVRs) {
3314 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3315 MachinePointerInfo(),
3316 false, false, false, 0);
3317 MemOpChains.push_back(Load.getValue(1));
3318 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3321 for (unsigned i=0; i<16; i+=PtrByteSize) {
3322 if (GPR_idx == NumGPRs)
3324 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3325 DAG.getConstant(i, PtrVT));
3326 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3327 false, false, false, 0);
3328 MemOpChains.push_back(Load.getValue(1));
3329 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3334 // Non-varargs Altivec params generally go in registers, but have
3335 // stack space allocated at the end.
3336 if (VR_idx != NumVRs) {
3337 // Doesn't have GPR space allocated.
3338 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3339 } else if (nAltivecParamsAtEnd==0) {
3340 // We are emitting Altivec params in order.
3341 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3342 isPPC64, isTailCall, true, MemOpChains,
3343 TailCallArguments, dl);
3349 // If all Altivec parameters fit in registers, as they usually do,
3350 // they get stack space following the non-Altivec parameters. We
3351 // don't track this here because nobody below needs it.
3352 // If there are more Altivec parameters than fit in registers emit
3354 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3356 // Offset is aligned; skip 1st 12 params which go in V registers.
3357 ArgOffset = ((ArgOffset+15)/16)*16;
3359 for (unsigned i = 0; i != NumOps; ++i) {
3360 SDValue Arg = OutVals[i];
3361 EVT ArgType = Outs[i].VT;
3362 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3363 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3366 // We are emitting Altivec params in order.
3367 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3368 isPPC64, isTailCall, true, MemOpChains,
3369 TailCallArguments, dl);
3376 if (!MemOpChains.empty())
3377 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3378 &MemOpChains[0], MemOpChains.size());
3380 // Check if this is an indirect call (MTCTR/BCTRL).
3381 // See PrepareCall() for more information about calls through function
3382 // pointers in the 64-bit SVR4 ABI.
3383 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3384 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3385 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3386 !isBLACompatibleAddress(Callee, DAG)) {
3387 // Load r2 into a virtual register and store it to the TOC save area.
3388 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3389 // TOC save area offset.
3390 SDValue PtrOff = DAG.getIntPtrConstant(40);
3391 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3392 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3396 // On Darwin, R12 must contain the address of an indirect callee. This does
3397 // not mean the MTCTR instruction must use R12; it's easier to model this as
3398 // an extra parameter, so do that.
3400 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3401 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3402 !isBLACompatibleAddress(Callee, DAG))
3403 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3404 PPC::R12), Callee));
3406 // Build a sequence of copy-to-reg nodes chained together with token chain
3407 // and flag operands which copy the outgoing args into the appropriate regs.
3409 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3410 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3411 RegsToPass[i].second, InFlag);
3412 InFlag = Chain.getValue(1);
3416 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3417 FPOp, true, TailCallArguments);
3419 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3420 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3425 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3426 MachineFunction &MF, bool isVarArg,
3427 const SmallVectorImpl<ISD::OutputArg> &Outs,
3428 LLVMContext &Context) const {
3429 SmallVector<CCValAssign, 16> RVLocs;
3430 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3432 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3436 PPCTargetLowering::LowerReturn(SDValue Chain,
3437 CallingConv::ID CallConv, bool isVarArg,
3438 const SmallVectorImpl<ISD::OutputArg> &Outs,
3439 const SmallVectorImpl<SDValue> &OutVals,
3440 DebugLoc dl, SelectionDAG &DAG) const {
3442 SmallVector<CCValAssign, 16> RVLocs;
3443 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3444 getTargetMachine(), RVLocs, *DAG.getContext());
3445 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3447 // If this is the first return lowered for this function, add the regs to the
3448 // liveout set for the function.
3449 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3450 for (unsigned i = 0; i != RVLocs.size(); ++i)
3451 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3456 // Copy the result values into the output registers.
3457 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3458 CCValAssign &VA = RVLocs[i];
3459 assert(VA.isRegLoc() && "Can only return in registers!");
3460 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3462 Flag = Chain.getValue(1);
3466 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3468 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3471 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3472 const PPCSubtarget &Subtarget) const {
3473 // When we pop the dynamic allocation we need to restore the SP link.
3474 DebugLoc dl = Op.getDebugLoc();
3476 // Get the corect type for pointers.
3477 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3479 // Construct the stack pointer operand.
3480 bool isPPC64 = Subtarget.isPPC64();
3481 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3482 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3484 // Get the operands for the STACKRESTORE.
3485 SDValue Chain = Op.getOperand(0);
3486 SDValue SaveSP = Op.getOperand(1);
3488 // Load the old link SP.
3489 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3490 MachinePointerInfo(),
3491 false, false, false, 0);
3493 // Restore the stack pointer.
3494 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3496 // Store the old link SP.
3497 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3504 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3505 MachineFunction &MF = DAG.getMachineFunction();
3506 bool isPPC64 = PPCSubTarget.isPPC64();
3507 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3508 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3510 // Get current frame pointer save index. The users of this index will be
3511 // primarily DYNALLOC instructions.
3512 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3513 int RASI = FI->getReturnAddrSaveIndex();
3515 // If the frame pointer save index hasn't been defined yet.
3517 // Find out what the fix offset of the frame pointer save area.
3518 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3519 // Allocate the frame index for frame pointer save area.
3520 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3522 FI->setReturnAddrSaveIndex(RASI);
3524 return DAG.getFrameIndex(RASI, PtrVT);
3528 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3529 MachineFunction &MF = DAG.getMachineFunction();
3530 bool isPPC64 = PPCSubTarget.isPPC64();
3531 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3532 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3534 // Get current frame pointer save index. The users of this index will be
3535 // primarily DYNALLOC instructions.
3536 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3537 int FPSI = FI->getFramePointerSaveIndex();
3539 // If the frame pointer save index hasn't been defined yet.
3541 // Find out what the fix offset of the frame pointer save area.
3542 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3545 // Allocate the frame index for frame pointer save area.
3546 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3548 FI->setFramePointerSaveIndex(FPSI);
3550 return DAG.getFrameIndex(FPSI, PtrVT);
3553 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3555 const PPCSubtarget &Subtarget) const {
3557 SDValue Chain = Op.getOperand(0);
3558 SDValue Size = Op.getOperand(1);
3559 DebugLoc dl = Op.getDebugLoc();
3561 // Get the corect type for pointers.
3562 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3564 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3565 DAG.getConstant(0, PtrVT), Size);
3566 // Construct a node for the frame pointer save index.
3567 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3568 // Build a DYNALLOC node.
3569 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3570 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3571 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3574 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3576 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3577 // Not FP? Not a fsel.
3578 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3579 !Op.getOperand(2).getValueType().isFloatingPoint())
3582 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3584 // Cannot handle SETEQ/SETNE.
3585 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3587 EVT ResVT = Op.getValueType();
3588 EVT CmpVT = Op.getOperand(0).getValueType();
3589 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3590 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3591 DebugLoc dl = Op.getDebugLoc();
3593 // If the RHS of the comparison is a 0.0, we don't need to do the
3594 // subtraction at all.
3595 if (isFloatingPointZero(RHS))
3597 default: break; // SETUO etc aren't handled by fsel.
3600 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3603 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3604 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3605 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3608 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3611 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3612 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3613 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3614 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3619 default: break; // SETUO etc aren't handled by fsel.
3622 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3623 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3624 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3625 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3628 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3629 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3630 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3631 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3634 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3635 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3636 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3637 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3640 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3641 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3642 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3643 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3648 // FIXME: Split this code up when LegalizeDAGTypes lands.
3649 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3650 DebugLoc dl) const {
3651 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3652 SDValue Src = Op.getOperand(0);
3653 if (Src.getValueType() == MVT::f32)
3654 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3657 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3658 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3660 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3665 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3669 // Convert the FP value to an int value through memory.
3670 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3672 // Emit a store to the stack slot.
3673 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3674 MachinePointerInfo(), false, false, 0);
3676 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3678 if (Op.getValueType() == MVT::i32)
3679 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3680 DAG.getConstant(4, FIPtr.getValueType()));
3681 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3682 false, false, false, 0);
3685 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3686 SelectionDAG &DAG) const {
3687 DebugLoc dl = Op.getDebugLoc();
3688 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3689 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3692 if (Op.getOperand(0).getValueType() == MVT::i64) {
3693 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3694 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3695 if (Op.getValueType() == MVT::f32)
3696 FP = DAG.getNode(ISD::FP_ROUND, dl,
3697 MVT::f32, FP, DAG.getIntPtrConstant(0));
3701 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3702 "Unhandled SINT_TO_FP type in custom expander!");
3703 // Since we only generate this in 64-bit mode, we can take advantage of
3704 // 64-bit registers. In particular, sign extend the input value into the
3705 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3706 // then lfd it and fcfid it.
3707 MachineFunction &MF = DAG.getMachineFunction();
3708 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3709 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3710 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3711 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3713 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3716 // STD the extended value into the stack slot.
3717 MachineMemOperand *MMO =
3718 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3719 MachineMemOperand::MOStore, 8, 8);
3720 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3722 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3723 Ops, 4, MVT::i64, MMO);
3724 // Load the value as a double.
3725 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3726 false, false, false, 0);
3728 // FCFID it and return it.
3729 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3730 if (Op.getValueType() == MVT::f32)
3731 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3735 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3736 SelectionDAG &DAG) const {
3737 DebugLoc dl = Op.getDebugLoc();
3739 The rounding mode is in bits 30:31 of FPSR, and has the following
3746 FLT_ROUNDS, on the other hand, expects the following:
3753 To perform the conversion, we do:
3754 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3757 MachineFunction &MF = DAG.getMachineFunction();
3758 EVT VT = Op.getValueType();
3759 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3760 std::vector<EVT> NodeTys;
3761 SDValue MFFSreg, InFlag;
3763 // Save FP Control Word to register
3764 NodeTys.push_back(MVT::f64); // return register
3765 NodeTys.push_back(MVT::Glue); // unused in this context
3766 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3768 // Save FP register to stack slot
3769 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3770 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3771 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3772 StackSlot, MachinePointerInfo(), false, false,0);
3774 // Load FP Control Word from low 32 bits of stack slot.
3775 SDValue Four = DAG.getConstant(4, PtrVT);
3776 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3777 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3778 false, false, false, 0);
3780 // Transform as necessary
3782 DAG.getNode(ISD::AND, dl, MVT::i32,
3783 CWD, DAG.getConstant(3, MVT::i32));
3785 DAG.getNode(ISD::SRL, dl, MVT::i32,
3786 DAG.getNode(ISD::AND, dl, MVT::i32,
3787 DAG.getNode(ISD::XOR, dl, MVT::i32,
3788 CWD, DAG.getConstant(3, MVT::i32)),
3789 DAG.getConstant(3, MVT::i32)),
3790 DAG.getConstant(1, MVT::i32));
3793 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3795 return DAG.getNode((VT.getSizeInBits() < 16 ?
3796 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3799 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3800 EVT VT = Op.getValueType();
3801 unsigned BitWidth = VT.getSizeInBits();
3802 DebugLoc dl = Op.getDebugLoc();
3803 assert(Op.getNumOperands() == 3 &&
3804 VT == Op.getOperand(1).getValueType() &&
3807 // Expand into a bunch of logical ops. Note that these ops
3808 // depend on the PPC behavior for oversized shift amounts.
3809 SDValue Lo = Op.getOperand(0);
3810 SDValue Hi = Op.getOperand(1);
3811 SDValue Amt = Op.getOperand(2);
3812 EVT AmtVT = Amt.getValueType();
3814 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3815 DAG.getConstant(BitWidth, AmtVT), Amt);
3816 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3817 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3818 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3819 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3820 DAG.getConstant(-BitWidth, AmtVT));
3821 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3822 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3823 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3824 SDValue OutOps[] = { OutLo, OutHi };
3825 return DAG.getMergeValues(OutOps, 2, dl);
3828 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3829 EVT VT = Op.getValueType();
3830 DebugLoc dl = Op.getDebugLoc();
3831 unsigned BitWidth = VT.getSizeInBits();
3832 assert(Op.getNumOperands() == 3 &&
3833 VT == Op.getOperand(1).getValueType() &&
3836 // Expand into a bunch of logical ops. Note that these ops
3837 // depend on the PPC behavior for oversized shift amounts.
3838 SDValue Lo = Op.getOperand(0);
3839 SDValue Hi = Op.getOperand(1);
3840 SDValue Amt = Op.getOperand(2);
3841 EVT AmtVT = Amt.getValueType();
3843 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3844 DAG.getConstant(BitWidth, AmtVT), Amt);
3845 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3846 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3847 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3848 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3849 DAG.getConstant(-BitWidth, AmtVT));
3850 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3851 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3852 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3853 SDValue OutOps[] = { OutLo, OutHi };
3854 return DAG.getMergeValues(OutOps, 2, dl);
3857 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3858 DebugLoc dl = Op.getDebugLoc();
3859 EVT VT = Op.getValueType();
3860 unsigned BitWidth = VT.getSizeInBits();
3861 assert(Op.getNumOperands() == 3 &&
3862 VT == Op.getOperand(1).getValueType() &&
3865 // Expand into a bunch of logical ops, followed by a select_cc.
3866 SDValue Lo = Op.getOperand(0);
3867 SDValue Hi = Op.getOperand(1);
3868 SDValue Amt = Op.getOperand(2);
3869 EVT AmtVT = Amt.getValueType();
3871 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3872 DAG.getConstant(BitWidth, AmtVT), Amt);
3873 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3874 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3875 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3876 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3877 DAG.getConstant(-BitWidth, AmtVT));
3878 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3879 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3880 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3881 Tmp4, Tmp6, ISD::SETLE);
3882 SDValue OutOps[] = { OutLo, OutHi };
3883 return DAG.getMergeValues(OutOps, 2, dl);
3886 //===----------------------------------------------------------------------===//
3887 // Vector related lowering.
3890 /// BuildSplatI - Build a canonical splati of Val with an element size of
3891 /// SplatSize. Cast the result to VT.
3892 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3893 SelectionDAG &DAG, DebugLoc dl) {
3894 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3896 static const EVT VTys[] = { // canonical VT to use for each size.
3897 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3900 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3902 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3906 EVT CanonicalVT = VTys[SplatSize-1];
3908 // Build a canonical splat for this value.
3909 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3910 SmallVector<SDValue, 8> Ops;
3911 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3912 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3913 &Ops[0], Ops.size());
3914 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3917 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3918 /// specified intrinsic ID.
3919 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3920 SelectionDAG &DAG, DebugLoc dl,
3921 EVT DestVT = MVT::Other) {
3922 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3923 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3924 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3927 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3928 /// specified intrinsic ID.
3929 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3930 SDValue Op2, SelectionDAG &DAG,
3931 DebugLoc dl, EVT DestVT = MVT::Other) {
3932 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3933 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3934 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3938 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3939 /// amount. The result has the specified value type.
3940 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3941 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3942 // Force LHS/RHS to be the right type.
3943 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3944 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3947 for (unsigned i = 0; i != 16; ++i)
3949 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3950 return DAG.getNode(ISD::BITCAST, dl, VT, T);
3953 // If this is a case we can't handle, return null and let the default
3954 // expansion code take care of it. If we CAN select this case, and if it
3955 // selects to a single instruction, return Op. Otherwise, if we can codegen
3956 // this case more efficiently than a constant pool load, lower it to the
3957 // sequence of ops that should be used.
3958 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3959 SelectionDAG &DAG) const {
3960 DebugLoc dl = Op.getDebugLoc();
3961 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3962 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3964 // Check if this is a splat of a constant value.
3965 APInt APSplatBits, APSplatUndef;
3966 unsigned SplatBitSize;
3968 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3969 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3972 unsigned SplatBits = APSplatBits.getZExtValue();
3973 unsigned SplatUndef = APSplatUndef.getZExtValue();
3974 unsigned SplatSize = SplatBitSize / 8;
3976 // First, handle single instruction cases.
3979 if (SplatBits == 0) {
3980 // Canonicalize all zero vectors to be v4i32.
3981 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3982 SDValue Z = DAG.getConstant(0, MVT::i32);
3983 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3984 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
3989 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3990 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3992 if (SextVal >= -16 && SextVal <= 15)
3993 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3996 // Two instruction sequences.
3998 // If this value is in the range [-32,30] and is even, use:
3999 // tmp = VSPLTI[bhw], result = add tmp, tmp
4000 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4001 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4002 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4003 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4006 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4007 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4009 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4010 // Make -1 and vspltisw -1:
4011 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4013 // Make the VSLW intrinsic, computing 0x8000_0000.
4014 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4017 // xor by OnesV to invert it.
4018 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4019 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4022 // Check to see if this is a wide variety of vsplti*, binop self cases.
4023 static const signed char SplatCsts[] = {
4024 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4025 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4028 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4029 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4030 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4031 int i = SplatCsts[idx];
4033 // Figure out what shift amount will be used by altivec if shifted by i in
4035 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4037 // vsplti + shl self.
4038 if (SextVal == (i << (int)TypeShiftAmt)) {
4039 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4040 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4041 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4042 Intrinsic::ppc_altivec_vslw
4044 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4045 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4048 // vsplti + srl self.
4049 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4050 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4051 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4052 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4053 Intrinsic::ppc_altivec_vsrw
4055 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4056 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4059 // vsplti + sra self.
4060 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4061 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4062 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4063 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4064 Intrinsic::ppc_altivec_vsraw
4066 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4067 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4070 // vsplti + rol self.
4071 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4072 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4073 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4074 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4075 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4076 Intrinsic::ppc_altivec_vrlw
4078 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4079 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4082 // t = vsplti c, result = vsldoi t, t, 1
4083 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4084 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4085 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4087 // t = vsplti c, result = vsldoi t, t, 2
4088 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4089 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4090 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4092 // t = vsplti c, result = vsldoi t, t, 3
4093 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4094 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4095 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4099 // Three instruction sequences.
4101 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4102 if (SextVal >= 0 && SextVal <= 31) {
4103 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4104 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4105 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4106 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4108 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4109 if (SextVal >= -31 && SextVal <= 0) {
4110 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4111 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4112 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4113 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4119 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4120 /// the specified operations to build the shuffle.
4121 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4122 SDValue RHS, SelectionDAG &DAG,
4124 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4125 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4126 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4129 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4141 if (OpNum == OP_COPY) {
4142 if (LHSID == (1*9+2)*9+3) return LHS;
4143 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4147 SDValue OpLHS, OpRHS;
4148 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4149 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4153 default: llvm_unreachable("Unknown i32 permute!");
4155 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4156 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4157 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4158 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4161 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4162 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4163 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4164 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4167 for (unsigned i = 0; i != 16; ++i)
4168 ShufIdxs[i] = (i&3)+0;
4171 for (unsigned i = 0; i != 16; ++i)
4172 ShufIdxs[i] = (i&3)+4;
4175 for (unsigned i = 0; i != 16; ++i)
4176 ShufIdxs[i] = (i&3)+8;
4179 for (unsigned i = 0; i != 16; ++i)
4180 ShufIdxs[i] = (i&3)+12;
4183 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4185 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4187 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4189 EVT VT = OpLHS.getValueType();
4190 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4191 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4192 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4193 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4196 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4197 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4198 /// return the code it can be lowered into. Worst case, it can always be
4199 /// lowered into a vperm.
4200 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4201 SelectionDAG &DAG) const {
4202 DebugLoc dl = Op.getDebugLoc();
4203 SDValue V1 = Op.getOperand(0);
4204 SDValue V2 = Op.getOperand(1);
4205 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4206 EVT VT = Op.getValueType();
4208 // Cases that are handled by instructions that take permute immediates
4209 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4210 // selected by the instruction selector.
4211 if (V2.getOpcode() == ISD::UNDEF) {
4212 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4213 PPC::isSplatShuffleMask(SVOp, 2) ||
4214 PPC::isSplatShuffleMask(SVOp, 4) ||
4215 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4216 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4217 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4218 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4219 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4220 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4221 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4222 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4223 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4228 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4229 // and produce a fixed permutation. If any of these match, do not lower to
4231 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4232 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4233 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4234 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4235 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4236 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4237 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4238 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4239 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4242 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4243 // perfect shuffle table to emit an optimal matching sequence.
4244 SmallVector<int, 16> PermMask;
4245 SVOp->getMask(PermMask);
4247 unsigned PFIndexes[4];
4248 bool isFourElementShuffle = true;
4249 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4250 unsigned EltNo = 8; // Start out undef.
4251 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4252 if (PermMask[i*4+j] < 0)
4253 continue; // Undef, ignore it.
4255 unsigned ByteSource = PermMask[i*4+j];
4256 if ((ByteSource & 3) != j) {
4257 isFourElementShuffle = false;
4262 EltNo = ByteSource/4;
4263 } else if (EltNo != ByteSource/4) {
4264 isFourElementShuffle = false;
4268 PFIndexes[i] = EltNo;
4271 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4272 // perfect shuffle vector to determine if it is cost effective to do this as
4273 // discrete instructions, or whether we should use a vperm.
4274 if (isFourElementShuffle) {
4275 // Compute the index in the perfect shuffle table.
4276 unsigned PFTableIndex =
4277 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4279 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4280 unsigned Cost = (PFEntry >> 30);
4282 // Determining when to avoid vperm is tricky. Many things affect the cost
4283 // of vperm, particularly how many times the perm mask needs to be computed.
4284 // For example, if the perm mask can be hoisted out of a loop or is already
4285 // used (perhaps because there are multiple permutes with the same shuffle
4286 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4287 // the loop requires an extra register.
4289 // As a compromise, we only emit discrete instructions if the shuffle can be
4290 // generated in 3 or fewer operations. When we have loop information
4291 // available, if this block is within a loop, we should avoid using vperm
4292 // for 3-operation perms and use a constant pool load instead.
4294 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4297 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4298 // vector that will get spilled to the constant pool.
4299 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4301 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4302 // that it is in input element units, not in bytes. Convert now.
4303 EVT EltVT = V1.getValueType().getVectorElementType();
4304 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4306 SmallVector<SDValue, 16> ResultMask;
4307 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4308 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4310 for (unsigned j = 0; j != BytesPerElement; ++j)
4311 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4315 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4316 &ResultMask[0], ResultMask.size());
4317 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4320 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4321 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4322 /// information about the intrinsic.
4323 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4325 unsigned IntrinsicID =
4326 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4329 switch (IntrinsicID) {
4330 default: return false;
4331 // Comparison predicates.
4332 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4333 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4334 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4335 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4336 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4337 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4338 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4339 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4340 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4341 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4342 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4343 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4344 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4346 // Normal Comparisons.
4347 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4348 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4349 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4350 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4351 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4352 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4353 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4354 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4355 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4356 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4357 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4358 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4359 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4364 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4365 /// lower, do it, otherwise return null.
4366 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4367 SelectionDAG &DAG) const {
4368 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4369 // opcode number of the comparison.
4370 DebugLoc dl = Op.getDebugLoc();
4373 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4374 return SDValue(); // Don't custom lower most intrinsics.
4376 // If this is a non-dot comparison, make the VCMP node and we are done.
4378 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4379 Op.getOperand(1), Op.getOperand(2),
4380 DAG.getConstant(CompareOpc, MVT::i32));
4381 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4384 // Create the PPCISD altivec 'dot' comparison node.
4386 Op.getOperand(2), // LHS
4387 Op.getOperand(3), // RHS
4388 DAG.getConstant(CompareOpc, MVT::i32)
4390 std::vector<EVT> VTs;
4391 VTs.push_back(Op.getOperand(2).getValueType());
4392 VTs.push_back(MVT::Glue);
4393 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4395 // Now that we have the comparison, emit a copy from the CR to a GPR.
4396 // This is flagged to the above dot comparison.
4397 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4398 DAG.getRegister(PPC::CR6, MVT::i32),
4399 CompNode.getValue(1));
4401 // Unpack the result based on how the target uses it.
4402 unsigned BitNo; // Bit # of CR6.
4403 bool InvertBit; // Invert result?
4404 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4405 default: // Can't happen, don't crash on invalid number though.
4406 case 0: // Return the value of the EQ bit of CR6.
4407 BitNo = 0; InvertBit = false;
4409 case 1: // Return the inverted value of the EQ bit of CR6.
4410 BitNo = 0; InvertBit = true;
4412 case 2: // Return the value of the LT bit of CR6.
4413 BitNo = 2; InvertBit = false;
4415 case 3: // Return the inverted value of the LT bit of CR6.
4416 BitNo = 2; InvertBit = true;
4420 // Shift the bit into the low position.
4421 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4422 DAG.getConstant(8-(3-BitNo), MVT::i32));
4424 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4425 DAG.getConstant(1, MVT::i32));
4427 // If we are supposed to, toggle the bit.
4429 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4430 DAG.getConstant(1, MVT::i32));
4434 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4435 SelectionDAG &DAG) const {
4436 DebugLoc dl = Op.getDebugLoc();
4437 // Create a stack slot that is 16-byte aligned.
4438 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4439 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4440 EVT PtrVT = getPointerTy();
4441 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4443 // Store the input value into Value#0 of the stack slot.
4444 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4445 Op.getOperand(0), FIdx, MachinePointerInfo(),
4448 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4449 false, false, false, 0);
4452 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4453 DebugLoc dl = Op.getDebugLoc();
4454 if (Op.getValueType() == MVT::v4i32) {
4455 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4457 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4458 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4460 SDValue RHSSwap = // = vrlw RHS, 16
4461 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4463 // Shrinkify inputs to v8i16.
4464 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4465 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4466 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4468 // Low parts multiplied together, generating 32-bit results (we ignore the
4470 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4471 LHS, RHS, DAG, dl, MVT::v4i32);
4473 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4474 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4475 // Shift the high parts up 16 bits.
4476 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4478 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4479 } else if (Op.getValueType() == MVT::v8i16) {
4480 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4482 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4484 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4485 LHS, RHS, Zero, DAG, dl);
4486 } else if (Op.getValueType() == MVT::v16i8) {
4487 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4489 // Multiply the even 8-bit parts, producing 16-bit sums.
4490 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4491 LHS, RHS, DAG, dl, MVT::v8i16);
4492 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4494 // Multiply the odd 8-bit parts, producing 16-bit sums.
4495 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4496 LHS, RHS, DAG, dl, MVT::v8i16);
4497 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4499 // Merge the results together.
4501 for (unsigned i = 0; i != 8; ++i) {
4503 Ops[i*2+1] = 2*i+1+16;
4505 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4507 llvm_unreachable("Unknown mul to lower!");
4511 /// LowerOperation - Provide custom lowering hooks for some operations.
4513 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4514 switch (Op.getOpcode()) {
4515 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4516 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4517 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4518 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4519 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4520 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4521 case ISD::SETCC: return LowerSETCC(Op, DAG);
4522 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4523 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4525 return LowerVASTART(Op, DAG, PPCSubTarget);
4528 return LowerVAARG(Op, DAG, PPCSubTarget);
4530 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4531 case ISD::DYNAMIC_STACKALLOC:
4532 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4534 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4535 case ISD::FP_TO_UINT:
4536 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4538 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4539 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4541 // Lower 64-bit shifts.
4542 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4543 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4544 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4546 // Vector-related lowering.
4547 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4548 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4549 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4550 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4551 case ISD::MUL: return LowerMUL(Op, DAG);
4553 // Frame & Return address.
4554 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4555 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4560 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4561 SmallVectorImpl<SDValue>&Results,
4562 SelectionDAG &DAG) const {
4563 const TargetMachine &TM = getTargetMachine();
4564 DebugLoc dl = N->getDebugLoc();
4565 switch (N->getOpcode()) {
4567 assert(false && "Do not know how to custom type legalize this operation!");
4570 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4571 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4574 EVT VT = N->getValueType(0);
4576 if (VT == MVT::i64) {
4577 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4579 Results.push_back(NewNode);
4580 Results.push_back(NewNode.getValue(1));
4584 case ISD::FP_ROUND_INREG: {
4585 assert(N->getValueType(0) == MVT::ppcf128);
4586 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4587 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4588 MVT::f64, N->getOperand(0),
4589 DAG.getIntPtrConstant(0));
4590 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4591 MVT::f64, N->getOperand(0),
4592 DAG.getIntPtrConstant(1));
4594 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4595 // of the long double, and puts FPSCR back the way it was. We do not
4596 // actually model FPSCR.
4597 std::vector<EVT> NodeTys;
4598 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4600 NodeTys.push_back(MVT::f64); // Return register
4601 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4602 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4603 MFFSreg = Result.getValue(0);
4604 InFlag = Result.getValue(1);
4607 NodeTys.push_back(MVT::Glue); // Returns a flag
4608 Ops[0] = DAG.getConstant(31, MVT::i32);
4610 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4611 InFlag = Result.getValue(0);
4614 NodeTys.push_back(MVT::Glue); // Returns a flag
4615 Ops[0] = DAG.getConstant(30, MVT::i32);
4617 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4618 InFlag = Result.getValue(0);
4621 NodeTys.push_back(MVT::f64); // result of add
4622 NodeTys.push_back(MVT::Glue); // Returns a flag
4626 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4627 FPreg = Result.getValue(0);
4628 InFlag = Result.getValue(1);
4631 NodeTys.push_back(MVT::f64);
4632 Ops[0] = DAG.getConstant(1, MVT::i32);
4636 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4637 FPreg = Result.getValue(0);
4639 // We know the low half is about to be thrown away, so just use something
4641 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4645 case ISD::FP_TO_SINT:
4646 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4652 //===----------------------------------------------------------------------===//
4653 // Other Lowering Code
4654 //===----------------------------------------------------------------------===//
4657 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4658 bool is64bit, unsigned BinOpcode) const {
4659 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4660 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4662 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4663 MachineFunction *F = BB->getParent();
4664 MachineFunction::iterator It = BB;
4667 unsigned dest = MI->getOperand(0).getReg();
4668 unsigned ptrA = MI->getOperand(1).getReg();
4669 unsigned ptrB = MI->getOperand(2).getReg();
4670 unsigned incr = MI->getOperand(3).getReg();
4671 DebugLoc dl = MI->getDebugLoc();
4673 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4674 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4675 F->insert(It, loopMBB);
4676 F->insert(It, exitMBB);
4677 exitMBB->splice(exitMBB->begin(), BB,
4678 llvm::next(MachineBasicBlock::iterator(MI)),
4680 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4682 MachineRegisterInfo &RegInfo = F->getRegInfo();
4683 unsigned TmpReg = (!BinOpcode) ? incr :
4684 RegInfo.createVirtualRegister(
4685 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4686 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4690 // fallthrough --> loopMBB
4691 BB->addSuccessor(loopMBB);
4694 // l[wd]arx dest, ptr
4695 // add r0, dest, incr
4696 // st[wd]cx. r0, ptr
4698 // fallthrough --> exitMBB
4700 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4701 .addReg(ptrA).addReg(ptrB);
4703 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4704 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4705 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4706 BuildMI(BB, dl, TII->get(PPC::BCC))
4707 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4708 BB->addSuccessor(loopMBB);
4709 BB->addSuccessor(exitMBB);
4718 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4719 MachineBasicBlock *BB,
4720 bool is8bit, // operation
4721 unsigned BinOpcode) const {
4722 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4723 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4724 // In 64 bit mode we have to use 64 bits for addresses, even though the
4725 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4726 // registers without caring whether they're 32 or 64, but here we're
4727 // doing actual arithmetic on the addresses.
4728 bool is64bit = PPCSubTarget.isPPC64();
4729 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4731 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4732 MachineFunction *F = BB->getParent();
4733 MachineFunction::iterator It = BB;
4736 unsigned dest = MI->getOperand(0).getReg();
4737 unsigned ptrA = MI->getOperand(1).getReg();
4738 unsigned ptrB = MI->getOperand(2).getReg();
4739 unsigned incr = MI->getOperand(3).getReg();
4740 DebugLoc dl = MI->getDebugLoc();
4742 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4743 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4744 F->insert(It, loopMBB);
4745 F->insert(It, exitMBB);
4746 exitMBB->splice(exitMBB->begin(), BB,
4747 llvm::next(MachineBasicBlock::iterator(MI)),
4749 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4751 MachineRegisterInfo &RegInfo = F->getRegInfo();
4752 const TargetRegisterClass *RC =
4753 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4754 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4755 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4756 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4757 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4758 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4759 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4760 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4761 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4762 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4763 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4764 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4765 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4767 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4771 // fallthrough --> loopMBB
4772 BB->addSuccessor(loopMBB);
4774 // The 4-byte load must be aligned, while a char or short may be
4775 // anywhere in the word. Hence all this nasty bookkeeping code.
4776 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4777 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4778 // xori shift, shift1, 24 [16]
4779 // rlwinm ptr, ptr1, 0, 0, 29
4780 // slw incr2, incr, shift
4781 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4782 // slw mask, mask2, shift
4784 // lwarx tmpDest, ptr
4785 // add tmp, tmpDest, incr2
4786 // andc tmp2, tmpDest, mask
4787 // and tmp3, tmp, mask
4788 // or tmp4, tmp3, tmp2
4791 // fallthrough --> exitMBB
4792 // srw dest, tmpDest, shift
4793 if (ptrA != ZeroReg) {
4794 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4795 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4796 .addReg(ptrA).addReg(ptrB);
4800 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4801 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4802 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4803 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4805 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4806 .addReg(Ptr1Reg).addImm(0).addImm(61);
4808 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4809 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4810 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4811 .addReg(incr).addReg(ShiftReg);
4813 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4815 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4816 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4818 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4819 .addReg(Mask2Reg).addReg(ShiftReg);
4822 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4823 .addReg(ZeroReg).addReg(PtrReg);
4825 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4826 .addReg(Incr2Reg).addReg(TmpDestReg);
4827 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4828 .addReg(TmpDestReg).addReg(MaskReg);
4829 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4830 .addReg(TmpReg).addReg(MaskReg);
4831 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4832 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4833 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4834 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4835 BuildMI(BB, dl, TII->get(PPC::BCC))
4836 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4837 BB->addSuccessor(loopMBB);
4838 BB->addSuccessor(exitMBB);
4843 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4849 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4850 MachineBasicBlock *BB) const {
4851 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4853 // To "insert" these instructions we actually have to insert their
4854 // control-flow patterns.
4855 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4856 MachineFunction::iterator It = BB;
4859 MachineFunction *F = BB->getParent();
4861 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4862 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4863 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4864 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4865 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4867 // The incoming instruction knows the destination vreg to set, the
4868 // condition code register to branch on, the true/false values to
4869 // select between, and a branch opcode to use.
4874 // cmpTY ccX, r1, r2
4876 // fallthrough --> copy0MBB
4877 MachineBasicBlock *thisMBB = BB;
4878 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4879 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4880 unsigned SelectPred = MI->getOperand(4).getImm();
4881 DebugLoc dl = MI->getDebugLoc();
4882 F->insert(It, copy0MBB);
4883 F->insert(It, sinkMBB);
4885 // Transfer the remainder of BB and its successor edges to sinkMBB.
4886 sinkMBB->splice(sinkMBB->begin(), BB,
4887 llvm::next(MachineBasicBlock::iterator(MI)),
4889 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4891 // Next, add the true and fallthrough blocks as its successors.
4892 BB->addSuccessor(copy0MBB);
4893 BB->addSuccessor(sinkMBB);
4895 BuildMI(BB, dl, TII->get(PPC::BCC))
4896 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4899 // %FalseValue = ...
4900 // # fallthrough to sinkMBB
4903 // Update machine-CFG edges
4904 BB->addSuccessor(sinkMBB);
4907 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4910 BuildMI(*BB, BB->begin(), dl,
4911 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4912 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4913 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4915 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4916 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4917 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4918 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4919 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4920 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4921 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4922 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4924 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4925 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4926 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4927 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4928 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4929 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4930 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4931 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4933 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4934 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4935 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4936 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4937 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4938 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4939 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4940 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4942 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4943 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4944 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4945 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4946 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4947 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4948 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4949 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4951 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4952 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4953 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4954 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4955 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4956 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4957 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4958 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4960 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4961 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4962 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4963 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4964 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4965 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4966 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4967 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4969 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4970 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4971 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4972 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4973 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4974 BB = EmitAtomicBinary(MI, BB, false, 0);
4975 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4976 BB = EmitAtomicBinary(MI, BB, true, 0);
4978 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4979 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4980 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4982 unsigned dest = MI->getOperand(0).getReg();
4983 unsigned ptrA = MI->getOperand(1).getReg();
4984 unsigned ptrB = MI->getOperand(2).getReg();
4985 unsigned oldval = MI->getOperand(3).getReg();
4986 unsigned newval = MI->getOperand(4).getReg();
4987 DebugLoc dl = MI->getDebugLoc();
4989 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4990 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4991 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4992 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4993 F->insert(It, loop1MBB);
4994 F->insert(It, loop2MBB);
4995 F->insert(It, midMBB);
4996 F->insert(It, exitMBB);
4997 exitMBB->splice(exitMBB->begin(), BB,
4998 llvm::next(MachineBasicBlock::iterator(MI)),
5000 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5004 // fallthrough --> loopMBB
5005 BB->addSuccessor(loop1MBB);
5008 // l[wd]arx dest, ptr
5009 // cmp[wd] dest, oldval
5012 // st[wd]cx. newval, ptr
5016 // st[wd]cx. dest, ptr
5019 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5020 .addReg(ptrA).addReg(ptrB);
5021 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5022 .addReg(oldval).addReg(dest);
5023 BuildMI(BB, dl, TII->get(PPC::BCC))
5024 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5025 BB->addSuccessor(loop2MBB);
5026 BB->addSuccessor(midMBB);
5029 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5030 .addReg(newval).addReg(ptrA).addReg(ptrB);
5031 BuildMI(BB, dl, TII->get(PPC::BCC))
5032 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5033 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5034 BB->addSuccessor(loop1MBB);
5035 BB->addSuccessor(exitMBB);
5038 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5039 .addReg(dest).addReg(ptrA).addReg(ptrB);
5040 BB->addSuccessor(exitMBB);
5045 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5046 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5047 // We must use 64-bit registers for addresses when targeting 64-bit,
5048 // since we're actually doing arithmetic on them. Other registers
5050 bool is64bit = PPCSubTarget.isPPC64();
5051 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5053 unsigned dest = MI->getOperand(0).getReg();
5054 unsigned ptrA = MI->getOperand(1).getReg();
5055 unsigned ptrB = MI->getOperand(2).getReg();
5056 unsigned oldval = MI->getOperand(3).getReg();
5057 unsigned newval = MI->getOperand(4).getReg();
5058 DebugLoc dl = MI->getDebugLoc();
5060 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5061 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5062 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5063 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5064 F->insert(It, loop1MBB);
5065 F->insert(It, loop2MBB);
5066 F->insert(It, midMBB);
5067 F->insert(It, exitMBB);
5068 exitMBB->splice(exitMBB->begin(), BB,
5069 llvm::next(MachineBasicBlock::iterator(MI)),
5071 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5073 MachineRegisterInfo &RegInfo = F->getRegInfo();
5074 const TargetRegisterClass *RC =
5075 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5076 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5077 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5078 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5079 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5080 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5081 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5082 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5083 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5084 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5085 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5086 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5087 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5088 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5089 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5091 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5092 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5095 // fallthrough --> loopMBB
5096 BB->addSuccessor(loop1MBB);
5098 // The 4-byte load must be aligned, while a char or short may be
5099 // anywhere in the word. Hence all this nasty bookkeeping code.
5100 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5101 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5102 // xori shift, shift1, 24 [16]
5103 // rlwinm ptr, ptr1, 0, 0, 29
5104 // slw newval2, newval, shift
5105 // slw oldval2, oldval,shift
5106 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5107 // slw mask, mask2, shift
5108 // and newval3, newval2, mask
5109 // and oldval3, oldval2, mask
5111 // lwarx tmpDest, ptr
5112 // and tmp, tmpDest, mask
5113 // cmpw tmp, oldval3
5116 // andc tmp2, tmpDest, mask
5117 // or tmp4, tmp2, newval3
5122 // stwcx. tmpDest, ptr
5124 // srw dest, tmpDest, shift
5125 if (ptrA != ZeroReg) {
5126 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5127 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5128 .addReg(ptrA).addReg(ptrB);
5132 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5133 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5134 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5135 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5137 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5138 .addReg(Ptr1Reg).addImm(0).addImm(61);
5140 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5141 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5142 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5143 .addReg(newval).addReg(ShiftReg);
5144 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5145 .addReg(oldval).addReg(ShiftReg);
5147 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5149 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5150 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5151 .addReg(Mask3Reg).addImm(65535);
5153 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5154 .addReg(Mask2Reg).addReg(ShiftReg);
5155 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5156 .addReg(NewVal2Reg).addReg(MaskReg);
5157 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5158 .addReg(OldVal2Reg).addReg(MaskReg);
5161 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5162 .addReg(ZeroReg).addReg(PtrReg);
5163 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5164 .addReg(TmpDestReg).addReg(MaskReg);
5165 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5166 .addReg(TmpReg).addReg(OldVal3Reg);
5167 BuildMI(BB, dl, TII->get(PPC::BCC))
5168 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5169 BB->addSuccessor(loop2MBB);
5170 BB->addSuccessor(midMBB);
5173 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5174 .addReg(TmpDestReg).addReg(MaskReg);
5175 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5176 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5177 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5178 .addReg(ZeroReg).addReg(PtrReg);
5179 BuildMI(BB, dl, TII->get(PPC::BCC))
5180 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5181 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5182 BB->addSuccessor(loop1MBB);
5183 BB->addSuccessor(exitMBB);
5186 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5187 .addReg(ZeroReg).addReg(PtrReg);
5188 BB->addSuccessor(exitMBB);
5193 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5196 llvm_unreachable("Unexpected instr type to insert");
5199 MI->eraseFromParent(); // The pseudo instruction is gone now.
5203 //===----------------------------------------------------------------------===//
5204 // Target Optimization Hooks
5205 //===----------------------------------------------------------------------===//
5207 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5208 DAGCombinerInfo &DCI) const {
5209 const TargetMachine &TM = getTargetMachine();
5210 SelectionDAG &DAG = DCI.DAG;
5211 DebugLoc dl = N->getDebugLoc();
5212 switch (N->getOpcode()) {
5215 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5216 if (C->isNullValue()) // 0 << V -> 0.
5217 return N->getOperand(0);
5221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5222 if (C->isNullValue()) // 0 >>u V -> 0.
5223 return N->getOperand(0);
5227 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5228 if (C->isNullValue() || // 0 >>s V -> 0.
5229 C->isAllOnesValue()) // -1 >>s V -> -1.
5230 return N->getOperand(0);
5234 case ISD::SINT_TO_FP:
5235 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5236 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5237 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5238 // We allow the src/dst to be either f32/f64, but the intermediate
5239 // type must be i64.
5240 if (N->getOperand(0).getValueType() == MVT::i64 &&
5241 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5242 SDValue Val = N->getOperand(0).getOperand(0);
5243 if (Val.getValueType() == MVT::f32) {
5244 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5245 DCI.AddToWorklist(Val.getNode());
5248 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5249 DCI.AddToWorklist(Val.getNode());
5250 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5251 DCI.AddToWorklist(Val.getNode());
5252 if (N->getValueType(0) == MVT::f32) {
5253 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5254 DAG.getIntPtrConstant(0));
5255 DCI.AddToWorklist(Val.getNode());
5258 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5259 // If the intermediate type is i32, we can avoid the load/store here
5266 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5267 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5268 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5269 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5270 N->getOperand(1).getValueType() == MVT::i32 &&
5271 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5272 SDValue Val = N->getOperand(1).getOperand(0);
5273 if (Val.getValueType() == MVT::f32) {
5274 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5275 DCI.AddToWorklist(Val.getNode());
5277 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5278 DCI.AddToWorklist(Val.getNode());
5280 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5281 N->getOperand(2), N->getOperand(3));
5282 DCI.AddToWorklist(Val.getNode());
5286 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5287 if (cast<StoreSDNode>(N)->isUnindexed() &&
5288 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5289 N->getOperand(1).getNode()->hasOneUse() &&
5290 (N->getOperand(1).getValueType() == MVT::i32 ||
5291 N->getOperand(1).getValueType() == MVT::i16)) {
5292 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5293 // Do an any-extend to 32-bits if this is a half-word input.
5294 if (BSwapOp.getValueType() == MVT::i16)
5295 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5298 N->getOperand(0), BSwapOp, N->getOperand(2),
5299 DAG.getValueType(N->getOperand(1).getValueType())
5302 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5303 Ops, array_lengthof(Ops),
5304 cast<StoreSDNode>(N)->getMemoryVT(),
5305 cast<StoreSDNode>(N)->getMemOperand());
5309 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5310 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5311 N->getOperand(0).hasOneUse() &&
5312 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5313 SDValue Load = N->getOperand(0);
5314 LoadSDNode *LD = cast<LoadSDNode>(Load);
5315 // Create the byte-swapping load.
5317 LD->getChain(), // Chain
5318 LD->getBasePtr(), // Ptr
5319 DAG.getValueType(N->getValueType(0)) // VT
5322 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5323 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5324 LD->getMemoryVT(), LD->getMemOperand());
5326 // If this is an i16 load, insert the truncate.
5327 SDValue ResVal = BSLoad;
5328 if (N->getValueType(0) == MVT::i16)
5329 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5331 // First, combine the bswap away. This makes the value produced by the
5333 DCI.CombineTo(N, ResVal);
5335 // Next, combine the load away, we give it a bogus result value but a real
5336 // chain result. The result value is dead because the bswap is dead.
5337 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5339 // Return N so it doesn't get rechecked!
5340 return SDValue(N, 0);
5344 case PPCISD::VCMP: {
5345 // If a VCMPo node already exists with exactly the same operands as this
5346 // node, use its result instead of this node (VCMPo computes both a CR6 and
5347 // a normal output).
5349 if (!N->getOperand(0).hasOneUse() &&
5350 !N->getOperand(1).hasOneUse() &&
5351 !N->getOperand(2).hasOneUse()) {
5353 // Scan all of the users of the LHS, looking for VCMPo's that match.
5354 SDNode *VCMPoNode = 0;
5356 SDNode *LHSN = N->getOperand(0).getNode();
5357 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5359 if (UI->getOpcode() == PPCISD::VCMPo &&
5360 UI->getOperand(1) == N->getOperand(1) &&
5361 UI->getOperand(2) == N->getOperand(2) &&
5362 UI->getOperand(0) == N->getOperand(0)) {
5367 // If there is no VCMPo node, or if the flag value has a single use, don't
5369 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5372 // Look at the (necessarily single) use of the flag value. If it has a
5373 // chain, this transformation is more complex. Note that multiple things
5374 // could use the value result, which we should ignore.
5375 SDNode *FlagUser = 0;
5376 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5377 FlagUser == 0; ++UI) {
5378 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5380 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5381 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5388 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5389 // give up for right now.
5390 if (FlagUser->getOpcode() == PPCISD::MFCR)
5391 return SDValue(VCMPoNode, 0);
5396 // If this is a branch on an altivec predicate comparison, lower this so
5397 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5398 // lowering is done pre-legalize, because the legalizer lowers the predicate
5399 // compare down to code that is difficult to reassemble.
5400 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5401 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5405 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5406 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5407 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5408 assert(isDot && "Can't compare against a vector result!");
5410 // If this is a comparison against something other than 0/1, then we know
5411 // that the condition is never/always true.
5412 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5413 if (Val != 0 && Val != 1) {
5414 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5415 return N->getOperand(0);
5416 // Always !=, turn it into an unconditional branch.
5417 return DAG.getNode(ISD::BR, dl, MVT::Other,
5418 N->getOperand(0), N->getOperand(4));
5421 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5423 // Create the PPCISD altivec 'dot' comparison node.
5424 std::vector<EVT> VTs;
5426 LHS.getOperand(2), // LHS of compare
5427 LHS.getOperand(3), // RHS of compare
5428 DAG.getConstant(CompareOpc, MVT::i32)
5430 VTs.push_back(LHS.getOperand(2).getValueType());
5431 VTs.push_back(MVT::Glue);
5432 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5434 // Unpack the result based on how the target uses it.
5435 PPC::Predicate CompOpc;
5436 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5437 default: // Can't happen, don't crash on invalid number though.
5438 case 0: // Branch on the value of the EQ bit of CR6.
5439 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5441 case 1: // Branch on the inverted value of the EQ bit of CR6.
5442 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5444 case 2: // Branch on the value of the LT bit of CR6.
5445 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5447 case 3: // Branch on the inverted value of the LT bit of CR6.
5448 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5452 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5453 DAG.getConstant(CompOpc, MVT::i32),
5454 DAG.getRegister(PPC::CR6, MVT::i32),
5455 N->getOperand(4), CompNode.getValue(1));
5464 //===----------------------------------------------------------------------===//
5465 // Inline Assembly Support
5466 //===----------------------------------------------------------------------===//
5468 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5472 const SelectionDAG &DAG,
5473 unsigned Depth) const {
5474 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5475 switch (Op.getOpcode()) {
5477 case PPCISD::LBRX: {
5478 // lhbrx is known to have the top bits cleared out.
5479 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5480 KnownZero = 0xFFFF0000;
5483 case ISD::INTRINSIC_WO_CHAIN: {
5484 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5486 case Intrinsic::ppc_altivec_vcmpbfp_p:
5487 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5488 case Intrinsic::ppc_altivec_vcmpequb_p:
5489 case Intrinsic::ppc_altivec_vcmpequh_p:
5490 case Intrinsic::ppc_altivec_vcmpequw_p:
5491 case Intrinsic::ppc_altivec_vcmpgefp_p:
5492 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5493 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5494 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5495 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5496 case Intrinsic::ppc_altivec_vcmpgtub_p:
5497 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5498 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5499 KnownZero = ~1U; // All bits but the low one are known to be zero.
5507 /// getConstraintType - Given a constraint, return the type of
5508 /// constraint it is for this target.
5509 PPCTargetLowering::ConstraintType
5510 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5511 if (Constraint.size() == 1) {
5512 switch (Constraint[0]) {
5519 return C_RegisterClass;
5522 return TargetLowering::getConstraintType(Constraint);
5525 /// Examine constraint type and operand type and determine a weight value.
5526 /// This object must already have been set up with the operand type
5527 /// and the current alternative constraint selected.
5528 TargetLowering::ConstraintWeight
5529 PPCTargetLowering::getSingleConstraintMatchWeight(
5530 AsmOperandInfo &info, const char *constraint) const {
5531 ConstraintWeight weight = CW_Invalid;
5532 Value *CallOperandVal = info.CallOperandVal;
5533 // If we don't have a value, we can't do a match,
5534 // but allow it at the lowest weight.
5535 if (CallOperandVal == NULL)
5537 Type *type = CallOperandVal->getType();
5538 // Look at the constraint type.
5539 switch (*constraint) {
5541 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5544 if (type->isIntegerTy())
5545 weight = CW_Register;
5548 if (type->isFloatTy())
5549 weight = CW_Register;
5552 if (type->isDoubleTy())
5553 weight = CW_Register;
5556 if (type->isVectorTy())
5557 weight = CW_Register;
5560 weight = CW_Register;
5566 std::pair<unsigned, const TargetRegisterClass*>
5567 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5569 if (Constraint.size() == 1) {
5570 // GCC RS6000 Constraint Letters
5571 switch (Constraint[0]) {
5574 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5575 return std::make_pair(0U, PPC::G8RCRegisterClass);
5576 return std::make_pair(0U, PPC::GPRCRegisterClass);
5579 return std::make_pair(0U, PPC::F4RCRegisterClass);
5580 else if (VT == MVT::f64)
5581 return std::make_pair(0U, PPC::F8RCRegisterClass);
5584 return std::make_pair(0U, PPC::VRRCRegisterClass);
5586 return std::make_pair(0U, PPC::CRRCRegisterClass);
5590 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5594 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5595 /// vector. If it is invalid, don't add anything to Ops.
5596 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5597 std::string &Constraint,
5598 std::vector<SDValue>&Ops,
5599 SelectionDAG &DAG) const {
5600 SDValue Result(0,0);
5602 // Only support length 1 constraints.
5603 if (Constraint.length() > 1) return;
5605 char Letter = Constraint[0];
5616 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5617 if (!CST) return; // Must be an immediate to match.
5618 unsigned Value = CST->getZExtValue();
5620 default: llvm_unreachable("Unknown constraint letter!");
5621 case 'I': // "I" is a signed 16-bit constant.
5622 if ((short)Value == (int)Value)
5623 Result = DAG.getTargetConstant(Value, Op.getValueType());
5625 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5626 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5627 if ((short)Value == 0)
5628 Result = DAG.getTargetConstant(Value, Op.getValueType());
5630 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5631 if ((Value >> 16) == 0)
5632 Result = DAG.getTargetConstant(Value, Op.getValueType());
5634 case 'M': // "M" is a constant that is greater than 31.
5636 Result = DAG.getTargetConstant(Value, Op.getValueType());
5638 case 'N': // "N" is a positive constant that is an exact power of two.
5639 if ((int)Value > 0 && isPowerOf2_32(Value))
5640 Result = DAG.getTargetConstant(Value, Op.getValueType());
5642 case 'O': // "O" is the constant zero.
5644 Result = DAG.getTargetConstant(Value, Op.getValueType());
5646 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5647 if ((short)-Value == (int)-Value)
5648 Result = DAG.getTargetConstant(Value, Op.getValueType());
5655 if (Result.getNode()) {
5656 Ops.push_back(Result);
5660 // Handle standard constraint letters.
5661 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5664 // isLegalAddressingMode - Return true if the addressing mode represented
5665 // by AM is legal for this target, for a load/store of the specified type.
5666 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5668 // FIXME: PPC does not allow r+i addressing modes for vectors!
5670 // PPC allows a sign-extended 16-bit immediate field.
5671 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5674 // No global is ever allowed as a base.
5678 // PPC only support r+r,
5680 case 0: // "r+i" or just "i", depending on HasBaseReg.
5683 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5685 // Otherwise we have r+r or r+i.
5688 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5690 // Allow 2*r as r+r.
5693 // No other scales are supported.
5700 /// isLegalAddressImmediate - Return true if the integer value can be used
5701 /// as the offset of the target addressing mode for load / store of the
5703 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5704 // PPC allows a sign-extended 16-bit immediate field.
5705 return (V > -(1 << 16) && V < (1 << 16)-1);
5708 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5712 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5713 SelectionDAG &DAG) const {
5714 MachineFunction &MF = DAG.getMachineFunction();
5715 MachineFrameInfo *MFI = MF.getFrameInfo();
5716 MFI->setReturnAddressIsTaken(true);
5718 DebugLoc dl = Op.getDebugLoc();
5719 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5721 // Make sure the function does not optimize away the store of the RA to
5723 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5724 FuncInfo->setLRStoreRequired();
5725 bool isPPC64 = PPCSubTarget.isPPC64();
5726 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5729 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5732 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5733 isPPC64? MVT::i64 : MVT::i32);
5734 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5735 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5737 MachinePointerInfo(), false, false, false, 0);
5740 // Just load the return address off the stack.
5741 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5742 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5743 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5746 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5747 SelectionDAG &DAG) const {
5748 DebugLoc dl = Op.getDebugLoc();
5749 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5751 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5752 bool isPPC64 = PtrVT == MVT::i64;
5754 MachineFunction &MF = DAG.getMachineFunction();
5755 MachineFrameInfo *MFI = MF.getFrameInfo();
5756 MFI->setFrameAddressIsTaken(true);
5757 bool is31 = (DisableFramePointerElim(MF) || MFI->hasVarSizedObjects()) &&
5758 MFI->getStackSize() &&
5759 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5760 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5761 (is31 ? PPC::R31 : PPC::R1);
5762 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5765 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5766 FrameAddr, MachinePointerInfo(), false, false,
5772 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5773 // The PowerPC target isn't yet aware of offsets.
5777 /// getOptimalMemOpType - Returns the target specific optimal type for load
5778 /// and store operations as a result of memset, memcpy, and memmove
5779 /// lowering. If DstAlign is zero that means it's safe to destination
5780 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5781 /// means there isn't a need to check it against alignment requirement,
5782 /// probably because the source does not need to be loaded. If
5783 /// 'IsZeroVal' is true, that means it's safe to return a
5784 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5785 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5786 /// constant so it does not need to be loaded.
5787 /// It returns EVT::Other if the type should be determined using generic
5788 /// target-independent logic.
5789 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5790 unsigned DstAlign, unsigned SrcAlign,
5793 MachineFunction &MF) const {
5794 if (this->PPCSubTarget.isPPC64()) {