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/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
40 CCValAssign::LocInfo &LocInfo,
41 ISD::ArgFlagsTy &ArgFlags,
43 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
54 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
55 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
57 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
58 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
60 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
61 if (TM.getSubtargetImpl()->isDarwin())
62 return new TargetLoweringObjectFileMachO();
64 return new TargetLoweringObjectFileELF();
67 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
68 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
72 // Use _setjmp/_longjmp instead of setjmp/longjmp.
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(true);
76 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
77 // arguments are at least 4/8 bytes aligned.
78 setMinStackArgumentAlignment(TM.getSubtarget<PPCSubtarget>().isPPC64() ? 8:4);
80 // Set up the register classes.
81 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
82 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
83 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
85 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
86 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
87 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
89 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
91 // PowerPC has pre-inc load and store's.
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
93 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
96 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
98 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
101 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
103 // This is used in the ppcf128->int sequence. Note it has different semantics
104 // from FP_ROUND: that rounds to nearest, this rounds to zero.
105 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
107 // We do not currently implment this libm ops for PowerPC.
108 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
109 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
110 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
111 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
112 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
114 // PowerPC has no SREM/UREM instructions
115 setOperationAction(ISD::SREM, MVT::i32, Expand);
116 setOperationAction(ISD::UREM, MVT::i32, Expand);
117 setOperationAction(ISD::SREM, MVT::i64, Expand);
118 setOperationAction(ISD::UREM, MVT::i64, Expand);
120 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
121 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
122 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
123 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
124 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
125 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
126 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
127 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
128 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
130 // We don't support sin/cos/sqrt/fmod/pow
131 setOperationAction(ISD::FSIN , MVT::f64, Expand);
132 setOperationAction(ISD::FCOS , MVT::f64, Expand);
133 setOperationAction(ISD::FREM , MVT::f64, Expand);
134 setOperationAction(ISD::FPOW , MVT::f64, Expand);
135 setOperationAction(ISD::FMA , MVT::f64, Expand);
136 setOperationAction(ISD::FSIN , MVT::f32, Expand);
137 setOperationAction(ISD::FCOS , MVT::f32, Expand);
138 setOperationAction(ISD::FREM , MVT::f32, Expand);
139 setOperationAction(ISD::FPOW , MVT::f32, Expand);
140 setOperationAction(ISD::FMA , MVT::f32, Expand);
142 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
144 // If we're enabling GP optimizations, use hardware square root
145 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
146 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
147 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
150 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
151 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
153 // PowerPC does not have BSWAP, CTPOP or CTTZ
154 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
155 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
156 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
157 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
158 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
159 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
160 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
161 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
162 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
163 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
165 // PowerPC does not have ROTR
166 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
167 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
169 // PowerPC does not have Select
170 setOperationAction(ISD::SELECT, MVT::i32, Expand);
171 setOperationAction(ISD::SELECT, MVT::i64, Expand);
172 setOperationAction(ISD::SELECT, MVT::f32, Expand);
173 setOperationAction(ISD::SELECT, MVT::f64, Expand);
175 // PowerPC wants to turn select_cc of FP into fsel when possible.
176 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
177 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
179 // PowerPC wants to optimize integer setcc a bit
180 setOperationAction(ISD::SETCC, MVT::i32, Custom);
182 // PowerPC does not have BRCOND which requires SetCC
183 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
185 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
187 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
188 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
190 // PowerPC does not have [U|S]INT_TO_FP
191 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
192 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
194 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
195 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
196 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
197 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
199 // We cannot sextinreg(i1). Expand to shifts.
200 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
202 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
203 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
204 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
205 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
208 // We want to legalize GlobalAddress and ConstantPool nodes into the
209 // appropriate instructions to materialize the address.
210 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
211 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
212 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
213 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
214 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
215 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
216 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
217 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
218 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
219 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
222 setOperationAction(ISD::TRAP, MVT::Other, Legal);
224 // TRAMPOLINE is custom lowered.
225 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
226 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
228 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
229 setOperationAction(ISD::VASTART , MVT::Other, Custom);
231 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()) {
232 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
233 // VAARG always uses double-word chunks, so promote anything smaller.
234 setOperationAction(ISD::VAARG, MVT::i1, Promote);
235 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
236 setOperationAction(ISD::VAARG, MVT::i8, Promote);
237 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
238 setOperationAction(ISD::VAARG, MVT::i16, Promote);
239 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
240 setOperationAction(ISD::VAARG, MVT::i32, Promote);
241 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
242 setOperationAction(ISD::VAARG, MVT::Other, Expand);
244 // VAARG is custom lowered with the 32-bit SVR4 ABI.
245 setOperationAction(ISD::VAARG, MVT::Other, Custom);
246 setOperationAction(ISD::VAARG, MVT::i64, Custom);
249 setOperationAction(ISD::VAARG, MVT::Other, Expand);
251 // Use the default implementation.
252 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
253 setOperationAction(ISD::VAEND , MVT::Other, Expand);
254 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
255 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
256 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
257 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
259 // We want to custom lower some of our intrinsics.
260 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
262 // Comparisons that require checking two conditions.
263 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
264 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
265 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
266 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
267 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
270 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
271 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
272 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
273 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
274 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
276 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
277 // They also have instructions for converting between i64 and fp.
278 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
279 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
280 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
281 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
282 // This is just the low 32 bits of a (signed) fp->i64 conversion.
283 // We cannot do this with Promote because i64 is not a legal type.
284 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
286 // FIXME: disable this lowered code. This generates 64-bit register values,
287 // and we don't model the fact that the top part is clobbered by calls. We
288 // need to flag these together so that the value isn't live across a call.
289 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
291 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
292 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
295 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
296 // 64-bit PowerPC implementations can support i64 types directly
297 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
298 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
299 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
300 // 64-bit PowerPC wants to expand i128 shifts itself.
301 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
302 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
303 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
305 // 32-bit PowerPC wants to expand i64 shifts itself.
306 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
307 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
308 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
311 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
312 // First set operation action for all vector types to expand. Then we
313 // will selectively turn on ones that can be effectively codegen'd.
314 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
315 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
316 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
318 // add/sub are legal for all supported vector VT's.
319 setOperationAction(ISD::ADD , VT, Legal);
320 setOperationAction(ISD::SUB , VT, Legal);
322 // We promote all shuffles to v16i8.
323 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
324 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
326 // We promote all non-typed operations to v4i32.
327 setOperationAction(ISD::AND , VT, Promote);
328 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
329 setOperationAction(ISD::OR , VT, Promote);
330 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
331 setOperationAction(ISD::XOR , VT, Promote);
332 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
333 setOperationAction(ISD::LOAD , VT, Promote);
334 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
335 setOperationAction(ISD::SELECT, VT, Promote);
336 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
337 setOperationAction(ISD::STORE, VT, Promote);
338 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
340 // No other operations are legal.
341 setOperationAction(ISD::MUL , VT, Expand);
342 setOperationAction(ISD::SDIV, VT, Expand);
343 setOperationAction(ISD::SREM, VT, Expand);
344 setOperationAction(ISD::UDIV, VT, Expand);
345 setOperationAction(ISD::UREM, VT, Expand);
346 setOperationAction(ISD::FDIV, VT, Expand);
347 setOperationAction(ISD::FNEG, VT, Expand);
348 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
349 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
350 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
351 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
352 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
353 setOperationAction(ISD::UDIVREM, VT, Expand);
354 setOperationAction(ISD::SDIVREM, VT, Expand);
355 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
356 setOperationAction(ISD::FPOW, VT, Expand);
357 setOperationAction(ISD::CTPOP, VT, Expand);
358 setOperationAction(ISD::CTLZ, VT, Expand);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
360 setOperationAction(ISD::CTTZ, VT, Expand);
361 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
364 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
365 // with merges, splats, etc.
366 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
368 setOperationAction(ISD::AND , MVT::v4i32, Legal);
369 setOperationAction(ISD::OR , MVT::v4i32, Legal);
370 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
371 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
372 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
373 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
375 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
376 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
377 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
378 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
380 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
381 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
382 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
383 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
385 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
386 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
388 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
389 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
390 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
391 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
394 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport())
395 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
397 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
398 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
400 setBooleanContents(ZeroOrOneBooleanContent);
401 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
403 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
404 setStackPointerRegisterToSaveRestore(PPC::X1);
405 setExceptionPointerRegister(PPC::X3);
406 setExceptionSelectorRegister(PPC::X4);
408 setStackPointerRegisterToSaveRestore(PPC::R1);
409 setExceptionPointerRegister(PPC::R3);
410 setExceptionSelectorRegister(PPC::R4);
413 // We have target-specific dag combine patterns for the following nodes:
414 setTargetDAGCombine(ISD::SINT_TO_FP);
415 setTargetDAGCombine(ISD::STORE);
416 setTargetDAGCombine(ISD::BR_CC);
417 setTargetDAGCombine(ISD::BSWAP);
419 // Darwin long double math library functions have $LDBL128 appended.
420 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
421 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
422 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
423 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
424 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
425 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
426 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
427 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
428 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
429 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
430 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
433 setMinFunctionAlignment(2);
434 if (PPCSubTarget.isDarwin())
435 setPrefFunctionAlignment(4);
437 setInsertFencesForAtomic(true);
439 setSchedulingPreference(Sched::Hybrid);
441 computeRegisterProperties();
444 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
445 /// function arguments in the caller parameter area.
446 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
447 const TargetMachine &TM = getTargetMachine();
448 // Darwin passes everything on 4 byte boundary.
449 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
452 // 16byte and wider vectors are passed on 16byte boundary.
453 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
454 if (VTy->getBitWidth() >= 128)
457 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
458 if (PPCSubTarget.isPPC64())
464 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
467 case PPCISD::FSEL: return "PPCISD::FSEL";
468 case PPCISD::FCFID: return "PPCISD::FCFID";
469 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
470 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
471 case PPCISD::STFIWX: return "PPCISD::STFIWX";
472 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
473 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
474 case PPCISD::VPERM: return "PPCISD::VPERM";
475 case PPCISD::Hi: return "PPCISD::Hi";
476 case PPCISD::Lo: return "PPCISD::Lo";
477 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
478 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
479 case PPCISD::LOAD: return "PPCISD::LOAD";
480 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
481 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
482 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
483 case PPCISD::SRL: return "PPCISD::SRL";
484 case PPCISD::SRA: return "PPCISD::SRA";
485 case PPCISD::SHL: return "PPCISD::SHL";
486 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
487 case PPCISD::STD_32: return "PPCISD::STD_32";
488 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
489 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
490 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
491 case PPCISD::NOP: return "PPCISD::NOP";
492 case PPCISD::MTCTR: return "PPCISD::MTCTR";
493 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
494 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
495 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
496 case PPCISD::MFCR: return "PPCISD::MFCR";
497 case PPCISD::VCMP: return "PPCISD::VCMP";
498 case PPCISD::VCMPo: return "PPCISD::VCMPo";
499 case PPCISD::LBRX: return "PPCISD::LBRX";
500 case PPCISD::STBRX: return "PPCISD::STBRX";
501 case PPCISD::LARX: return "PPCISD::LARX";
502 case PPCISD::STCX: return "PPCISD::STCX";
503 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
504 case PPCISD::MFFS: return "PPCISD::MFFS";
505 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
506 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
507 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
508 case PPCISD::MTFSF: return "PPCISD::MTFSF";
509 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
513 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
517 //===----------------------------------------------------------------------===//
518 // Node matching predicates, for use by the tblgen matching code.
519 //===----------------------------------------------------------------------===//
521 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
522 static bool isFloatingPointZero(SDValue Op) {
523 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
524 return CFP->getValueAPF().isZero();
525 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
526 // Maybe this has already been legalized into the constant pool?
527 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
528 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
529 return CFP->getValueAPF().isZero();
534 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
535 /// true if Op is undef or if it matches the specified value.
536 static bool isConstantOrUndef(int Op, int Val) {
537 return Op < 0 || Op == Val;
540 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
541 /// VPKUHUM instruction.
542 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
544 for (unsigned i = 0; i != 16; ++i)
545 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
548 for (unsigned i = 0; i != 8; ++i)
549 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
550 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
556 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
557 /// VPKUWUM instruction.
558 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
560 for (unsigned i = 0; i != 16; i += 2)
561 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
562 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
565 for (unsigned i = 0; i != 8; i += 2)
566 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
567 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
568 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
569 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
575 /// isVMerge - Common function, used to match vmrg* shuffles.
577 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
578 unsigned LHSStart, unsigned RHSStart) {
579 assert(N->getValueType(0) == MVT::v16i8 &&
580 "PPC only supports shuffles by bytes!");
581 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
582 "Unsupported merge size!");
584 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
585 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
586 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
587 LHSStart+j+i*UnitSize) ||
588 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
589 RHSStart+j+i*UnitSize))
595 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
596 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
597 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
600 return isVMerge(N, UnitSize, 8, 24);
601 return isVMerge(N, UnitSize, 8, 8);
604 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
605 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
606 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
609 return isVMerge(N, UnitSize, 0, 16);
610 return isVMerge(N, UnitSize, 0, 0);
614 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
615 /// amount, otherwise return -1.
616 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
617 assert(N->getValueType(0) == MVT::v16i8 &&
618 "PPC only supports shuffles by bytes!");
620 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
622 // Find the first non-undef value in the shuffle mask.
624 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
627 if (i == 16) return -1; // all undef.
629 // Otherwise, check to see if the rest of the elements are consecutively
630 // numbered from this value.
631 unsigned ShiftAmt = SVOp->getMaskElt(i);
632 if (ShiftAmt < i) return -1;
636 // Check the rest of the elements to see if they are consecutive.
637 for (++i; i != 16; ++i)
638 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
641 // Check the rest of the elements to see if they are consecutive.
642 for (++i; i != 16; ++i)
643 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
649 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
650 /// specifies a splat of a single element that is suitable for input to
651 /// VSPLTB/VSPLTH/VSPLTW.
652 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
653 assert(N->getValueType(0) == MVT::v16i8 &&
654 (EltSize == 1 || EltSize == 2 || EltSize == 4));
656 // This is a splat operation if each element of the permute is the same, and
657 // if the value doesn't reference the second vector.
658 unsigned ElementBase = N->getMaskElt(0);
660 // FIXME: Handle UNDEF elements too!
661 if (ElementBase >= 16)
664 // Check that the indices are consecutive, in the case of a multi-byte element
665 // splatted with a v16i8 mask.
666 for (unsigned i = 1; i != EltSize; ++i)
667 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
670 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
671 if (N->getMaskElt(i) < 0) continue;
672 for (unsigned j = 0; j != EltSize; ++j)
673 if (N->getMaskElt(i+j) != N->getMaskElt(j))
679 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
681 bool PPC::isAllNegativeZeroVector(SDNode *N) {
682 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
684 APInt APVal, APUndef;
688 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
689 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
690 return CFP->getValueAPF().isNegZero();
695 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
696 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
697 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
698 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
699 assert(isSplatShuffleMask(SVOp, EltSize));
700 return SVOp->getMaskElt(0) / EltSize;
703 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
704 /// by using a vspltis[bhw] instruction of the specified element size, return
705 /// the constant being splatted. The ByteSize field indicates the number of
706 /// bytes of each element [124] -> [bhw].
707 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
710 // If ByteSize of the splat is bigger than the element size of the
711 // build_vector, then we have a case where we are checking for a splat where
712 // multiple elements of the buildvector are folded together into a single
713 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
714 unsigned EltSize = 16/N->getNumOperands();
715 if (EltSize < ByteSize) {
716 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
717 SDValue UniquedVals[4];
718 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
720 // See if all of the elements in the buildvector agree across.
721 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
722 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
723 // If the element isn't a constant, bail fully out.
724 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
727 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
728 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
729 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
730 return SDValue(); // no match.
733 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
734 // either constant or undef values that are identical for each chunk. See
735 // if these chunks can form into a larger vspltis*.
737 // Check to see if all of the leading entries are either 0 or -1. If
738 // neither, then this won't fit into the immediate field.
739 bool LeadingZero = true;
740 bool LeadingOnes = true;
741 for (unsigned i = 0; i != Multiple-1; ++i) {
742 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
744 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
745 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
747 // Finally, check the least significant entry.
749 if (UniquedVals[Multiple-1].getNode() == 0)
750 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
751 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
753 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
756 if (UniquedVals[Multiple-1].getNode() == 0)
757 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
758 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
759 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
760 return DAG.getTargetConstant(Val, MVT::i32);
766 // Check to see if this buildvec has a single non-undef value in its elements.
767 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
768 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
769 if (OpVal.getNode() == 0)
770 OpVal = N->getOperand(i);
771 else if (OpVal != N->getOperand(i))
775 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
777 unsigned ValSizeInBytes = EltSize;
779 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
780 Value = CN->getZExtValue();
781 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
782 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
783 Value = FloatToBits(CN->getValueAPF().convertToFloat());
786 // If the splat value is larger than the element value, then we can never do
787 // this splat. The only case that we could fit the replicated bits into our
788 // immediate field for would be zero, and we prefer to use vxor for it.
789 if (ValSizeInBytes < ByteSize) return SDValue();
791 // If the element value is larger than the splat value, cut it in half and
792 // check to see if the two halves are equal. Continue doing this until we
793 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
794 while (ValSizeInBytes > ByteSize) {
795 ValSizeInBytes >>= 1;
797 // If the top half equals the bottom half, we're still ok.
798 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
799 (Value & ((1 << (8*ValSizeInBytes))-1)))
803 // Properly sign extend the value.
804 int ShAmt = (4-ByteSize)*8;
805 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
807 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
808 if (MaskVal == 0) return SDValue();
810 // Finally, if this value fits in a 5 bit sext field, return it
811 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
812 return DAG.getTargetConstant(MaskVal, MVT::i32);
816 //===----------------------------------------------------------------------===//
817 // Addressing Mode Selection
818 //===----------------------------------------------------------------------===//
820 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
821 /// or 64-bit immediate, and if the value can be accurately represented as a
822 /// sign extension from a 16-bit value. If so, this returns true and the
824 static bool isIntS16Immediate(SDNode *N, short &Imm) {
825 if (N->getOpcode() != ISD::Constant)
828 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
829 if (N->getValueType(0) == MVT::i32)
830 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
832 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
834 static bool isIntS16Immediate(SDValue Op, short &Imm) {
835 return isIntS16Immediate(Op.getNode(), Imm);
839 /// SelectAddressRegReg - Given the specified addressed, check to see if it
840 /// can be represented as an indexed [r+r] operation. Returns false if it
841 /// can be more efficiently represented with [r+imm].
842 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
844 SelectionDAG &DAG) const {
846 if (N.getOpcode() == ISD::ADD) {
847 if (isIntS16Immediate(N.getOperand(1), imm))
849 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
852 Base = N.getOperand(0);
853 Index = N.getOperand(1);
855 } else if (N.getOpcode() == ISD::OR) {
856 if (isIntS16Immediate(N.getOperand(1), imm))
857 return false; // r+i can fold it if we can.
859 // If this is an or of disjoint bitfields, we can codegen this as an add
860 // (for better address arithmetic) if the LHS and RHS of the OR are provably
862 APInt LHSKnownZero, LHSKnownOne;
863 APInt RHSKnownZero, RHSKnownOne;
864 DAG.ComputeMaskedBits(N.getOperand(0),
865 LHSKnownZero, LHSKnownOne);
867 if (LHSKnownZero.getBoolValue()) {
868 DAG.ComputeMaskedBits(N.getOperand(1),
869 RHSKnownZero, RHSKnownOne);
870 // If all of the bits are known zero on the LHS or RHS, the add won't
872 if (~(LHSKnownZero | RHSKnownZero) == 0) {
873 Base = N.getOperand(0);
874 Index = N.getOperand(1);
883 /// Returns true if the address N can be represented by a base register plus
884 /// a signed 16-bit displacement [r+imm], and if it is not better
885 /// represented as reg+reg.
886 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
888 SelectionDAG &DAG) const {
889 // FIXME dl should come from parent load or store, not from address
890 DebugLoc dl = N.getDebugLoc();
891 // If this can be more profitably realized as r+r, fail.
892 if (SelectAddressRegReg(N, Disp, Base, DAG))
895 if (N.getOpcode() == ISD::ADD) {
897 if (isIntS16Immediate(N.getOperand(1), imm)) {
898 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
899 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
900 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
902 Base = N.getOperand(0);
904 return true; // [r+i]
905 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
906 // Match LOAD (ADD (X, Lo(G))).
907 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
908 && "Cannot handle constant offsets yet!");
909 Disp = N.getOperand(1).getOperand(0); // The global address.
910 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
911 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
912 Disp.getOpcode() == ISD::TargetConstantPool ||
913 Disp.getOpcode() == ISD::TargetJumpTable);
914 Base = N.getOperand(0);
915 return true; // [&g+r]
917 } else if (N.getOpcode() == ISD::OR) {
919 if (isIntS16Immediate(N.getOperand(1), imm)) {
920 // If this is an or of disjoint bitfields, we can codegen this as an add
921 // (for better address arithmetic) if the LHS and RHS of the OR are
922 // provably disjoint.
923 APInt LHSKnownZero, LHSKnownOne;
924 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
926 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
927 // If all of the bits are known zero on the LHS or RHS, the add won't
929 Base = N.getOperand(0);
930 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
934 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
935 // Loading from a constant address.
937 // If this address fits entirely in a 16-bit sext immediate field, codegen
940 if (isIntS16Immediate(CN, Imm)) {
941 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
942 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
943 CN->getValueType(0));
947 // Handle 32-bit sext immediates with LIS + addr mode.
948 if (CN->getValueType(0) == MVT::i32 ||
949 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
950 int Addr = (int)CN->getZExtValue();
952 // Otherwise, break this down into an LIS + disp.
953 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
955 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
956 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
957 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
962 Disp = DAG.getTargetConstant(0, getPointerTy());
963 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
964 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
967 return true; // [r+0]
970 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
971 /// represented as an indexed [r+r] operation.
972 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
974 SelectionDAG &DAG) const {
975 // Check to see if we can easily represent this as an [r+r] address. This
976 // will fail if it thinks that the address is more profitably represented as
977 // reg+imm, e.g. where imm = 0.
978 if (SelectAddressRegReg(N, Base, Index, DAG))
981 // If the operand is an addition, always emit this as [r+r], since this is
982 // better (for code size, and execution, as the memop does the add for free)
983 // than emitting an explicit add.
984 if (N.getOpcode() == ISD::ADD) {
985 Base = N.getOperand(0);
986 Index = N.getOperand(1);
990 // Otherwise, do it the hard way, using R0 as the base register.
991 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
997 /// SelectAddressRegImmShift - Returns true if the address N can be
998 /// represented by a base register plus a signed 14-bit displacement
999 /// [r+imm*4]. Suitable for use by STD and friends.
1000 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1002 SelectionDAG &DAG) const {
1003 // FIXME dl should come from the parent load or store, not the address
1004 DebugLoc dl = N.getDebugLoc();
1005 // If this can be more profitably realized as r+r, fail.
1006 if (SelectAddressRegReg(N, Disp, Base, DAG))
1009 if (N.getOpcode() == ISD::ADD) {
1011 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1012 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1013 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1014 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1016 Base = N.getOperand(0);
1018 return true; // [r+i]
1019 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1020 // Match LOAD (ADD (X, Lo(G))).
1021 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1022 && "Cannot handle constant offsets yet!");
1023 Disp = N.getOperand(1).getOperand(0); // The global address.
1024 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1025 Disp.getOpcode() == ISD::TargetConstantPool ||
1026 Disp.getOpcode() == ISD::TargetJumpTable);
1027 Base = N.getOperand(0);
1028 return true; // [&g+r]
1030 } else if (N.getOpcode() == ISD::OR) {
1032 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1033 // If this is an or of disjoint bitfields, we can codegen this as an add
1034 // (for better address arithmetic) if the LHS and RHS of the OR are
1035 // provably disjoint.
1036 APInt LHSKnownZero, LHSKnownOne;
1037 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1038 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1039 // If all of the bits are known zero on the LHS or RHS, the add won't
1041 Base = N.getOperand(0);
1042 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1046 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1047 // Loading from a constant address. Verify low two bits are clear.
1048 if ((CN->getZExtValue() & 3) == 0) {
1049 // If this address fits entirely in a 14-bit sext immediate field, codegen
1052 if (isIntS16Immediate(CN, Imm)) {
1053 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1054 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1055 CN->getValueType(0));
1059 // Fold the low-part of 32-bit absolute addresses into addr mode.
1060 if (CN->getValueType(0) == MVT::i32 ||
1061 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1062 int Addr = (int)CN->getZExtValue();
1064 // Otherwise, break this down into an LIS + disp.
1065 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1066 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1067 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1068 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1074 Disp = DAG.getTargetConstant(0, getPointerTy());
1075 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1076 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1079 return true; // [r+0]
1083 /// getPreIndexedAddressParts - returns true by value, base pointer and
1084 /// offset pointer and addressing mode by reference if the node's address
1085 /// can be legally represented as pre-indexed load / store address.
1086 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1088 ISD::MemIndexedMode &AM,
1089 SelectionDAG &DAG) const {
1090 if (DisablePPCPreinc) return false;
1094 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1095 Ptr = LD->getBasePtr();
1096 VT = LD->getMemoryVT();
1098 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1099 Ptr = ST->getBasePtr();
1100 VT = ST->getMemoryVT();
1104 // PowerPC doesn't have preinc load/store instructions for vectors.
1108 // TODO: Check reg+reg first.
1110 // LDU/STU use reg+imm*4, others use reg+imm.
1111 if (VT != MVT::i64) {
1113 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1117 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1121 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1122 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1123 // sext i32 to i64 when addr mode is r+i.
1124 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1125 LD->getExtensionType() == ISD::SEXTLOAD &&
1126 isa<ConstantSDNode>(Offset))
1134 //===----------------------------------------------------------------------===//
1135 // LowerOperation implementation
1136 //===----------------------------------------------------------------------===//
1138 /// GetLabelAccessInfo - Return true if we should reference labels using a
1139 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1140 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1141 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1142 HiOpFlags = PPCII::MO_HA16;
1143 LoOpFlags = PPCII::MO_LO16;
1145 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1146 // non-darwin platform. We don't support PIC on other platforms yet.
1147 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1148 TM.getSubtarget<PPCSubtarget>().isDarwin();
1150 HiOpFlags |= PPCII::MO_PIC_FLAG;
1151 LoOpFlags |= PPCII::MO_PIC_FLAG;
1154 // If this is a reference to a global value that requires a non-lazy-ptr, make
1155 // sure that instruction lowering adds it.
1156 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1157 HiOpFlags |= PPCII::MO_NLP_FLAG;
1158 LoOpFlags |= PPCII::MO_NLP_FLAG;
1160 if (GV->hasHiddenVisibility()) {
1161 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1162 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1169 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1170 SelectionDAG &DAG) {
1171 EVT PtrVT = HiPart.getValueType();
1172 SDValue Zero = DAG.getConstant(0, PtrVT);
1173 DebugLoc DL = HiPart.getDebugLoc();
1175 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1176 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1178 // With PIC, the first instruction is actually "GR+hi(&G)".
1180 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1181 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1183 // Generate non-pic code that has direct accesses to the constant pool.
1184 // The address of the global is just (hi(&g)+lo(&g)).
1185 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1188 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1189 SelectionDAG &DAG) const {
1190 EVT PtrVT = Op.getValueType();
1191 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1192 const Constant *C = CP->getConstVal();
1194 unsigned MOHiFlag, MOLoFlag;
1195 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1197 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1199 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1200 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1203 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1204 EVT PtrVT = Op.getValueType();
1205 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1207 unsigned MOHiFlag, MOLoFlag;
1208 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1209 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1210 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1211 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1214 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1215 SelectionDAG &DAG) const {
1216 EVT PtrVT = Op.getValueType();
1218 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1220 unsigned MOHiFlag, MOLoFlag;
1221 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1222 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1223 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1224 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1227 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1228 SelectionDAG &DAG) const {
1230 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1231 DebugLoc dl = GA->getDebugLoc();
1232 const GlobalValue *GV = GA->getGlobal();
1233 EVT PtrVT = getPointerTy();
1234 bool is64bit = PPCSubTarget.isPPC64();
1236 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1238 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1239 PPCII::MO_TPREL16_HA);
1240 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1241 PPCII::MO_TPREL16_LO);
1243 if (model != TLSModel::LocalExec)
1244 llvm_unreachable("only local-exec TLS mode supported");
1245 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1246 is64bit ? MVT::i64 : MVT::i32);
1247 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1248 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1251 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1252 SelectionDAG &DAG) const {
1253 EVT PtrVT = Op.getValueType();
1254 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1255 DebugLoc DL = GSDN->getDebugLoc();
1256 const GlobalValue *GV = GSDN->getGlobal();
1258 // 64-bit SVR4 ABI code is always position-independent.
1259 // The actual address of the GlobalValue is stored in the TOC.
1260 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1261 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1262 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1263 DAG.getRegister(PPC::X2, MVT::i64));
1266 unsigned MOHiFlag, MOLoFlag;
1267 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1270 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1272 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1274 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1276 // If the global reference is actually to a non-lazy-pointer, we have to do an
1277 // extra load to get the address of the global.
1278 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1279 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1280 false, false, false, 0);
1284 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1285 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1286 DebugLoc dl = Op.getDebugLoc();
1288 // If we're comparing for equality to zero, expose the fact that this is
1289 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1290 // fold the new nodes.
1291 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1292 if (C->isNullValue() && CC == ISD::SETEQ) {
1293 EVT VT = Op.getOperand(0).getValueType();
1294 SDValue Zext = Op.getOperand(0);
1295 if (VT.bitsLT(MVT::i32)) {
1297 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1299 unsigned Log2b = Log2_32(VT.getSizeInBits());
1300 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1301 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1302 DAG.getConstant(Log2b, MVT::i32));
1303 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1305 // Leave comparisons against 0 and -1 alone for now, since they're usually
1306 // optimized. FIXME: revisit this when we can custom lower all setcc
1308 if (C->isAllOnesValue() || C->isNullValue())
1312 // If we have an integer seteq/setne, turn it into a compare against zero
1313 // by xor'ing the rhs with the lhs, which is faster than setting a
1314 // condition register, reading it back out, and masking the correct bit. The
1315 // normal approach here uses sub to do this instead of xor. Using xor exposes
1316 // the result to other bit-twiddling opportunities.
1317 EVT LHSVT = Op.getOperand(0).getValueType();
1318 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1319 EVT VT = Op.getValueType();
1320 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1322 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1327 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1328 const PPCSubtarget &Subtarget) const {
1329 SDNode *Node = Op.getNode();
1330 EVT VT = Node->getValueType(0);
1331 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1332 SDValue InChain = Node->getOperand(0);
1333 SDValue VAListPtr = Node->getOperand(1);
1334 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1335 DebugLoc dl = Node->getDebugLoc();
1337 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1340 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1341 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1343 InChain = GprIndex.getValue(1);
1345 if (VT == MVT::i64) {
1346 // Check if GprIndex is even
1347 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1348 DAG.getConstant(1, MVT::i32));
1349 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1350 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1351 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1352 DAG.getConstant(1, MVT::i32));
1353 // Align GprIndex to be even if it isn't
1354 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1358 // fpr index is 1 byte after gpr
1359 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1360 DAG.getConstant(1, MVT::i32));
1363 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1364 FprPtr, MachinePointerInfo(SV), MVT::i8,
1366 InChain = FprIndex.getValue(1);
1368 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1369 DAG.getConstant(8, MVT::i32));
1371 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1372 DAG.getConstant(4, MVT::i32));
1375 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1376 MachinePointerInfo(), false, false,
1378 InChain = OverflowArea.getValue(1);
1380 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1381 MachinePointerInfo(), false, false,
1383 InChain = RegSaveArea.getValue(1);
1385 // select overflow_area if index > 8
1386 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1387 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1389 // adjustment constant gpr_index * 4/8
1390 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1391 VT.isInteger() ? GprIndex : FprIndex,
1392 DAG.getConstant(VT.isInteger() ? 4 : 8,
1395 // OurReg = RegSaveArea + RegConstant
1396 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1399 // Floating types are 32 bytes into RegSaveArea
1400 if (VT.isFloatingPoint())
1401 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1402 DAG.getConstant(32, MVT::i32));
1404 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1405 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1406 VT.isInteger() ? GprIndex : FprIndex,
1407 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1410 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1411 VT.isInteger() ? VAListPtr : FprPtr,
1412 MachinePointerInfo(SV),
1413 MVT::i8, false, false, 0);
1415 // determine if we should load from reg_save_area or overflow_area
1416 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1418 // increase overflow_area by 4/8 if gpr/fpr > 8
1419 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1420 DAG.getConstant(VT.isInteger() ? 4 : 8,
1423 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1426 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1428 MachinePointerInfo(),
1429 MVT::i32, false, false, 0);
1431 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1432 false, false, false, 0);
1435 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1436 SelectionDAG &DAG) const {
1437 return Op.getOperand(0);
1440 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1441 SelectionDAG &DAG) const {
1442 SDValue Chain = Op.getOperand(0);
1443 SDValue Trmp = Op.getOperand(1); // trampoline
1444 SDValue FPtr = Op.getOperand(2); // nested function
1445 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1446 DebugLoc dl = Op.getDebugLoc();
1448 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1449 bool isPPC64 = (PtrVT == MVT::i64);
1451 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1454 TargetLowering::ArgListTy Args;
1455 TargetLowering::ArgListEntry Entry;
1457 Entry.Ty = IntPtrTy;
1458 Entry.Node = Trmp; Args.push_back(Entry);
1460 // TrampSize == (isPPC64 ? 48 : 40);
1461 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1462 isPPC64 ? MVT::i64 : MVT::i32);
1463 Args.push_back(Entry);
1465 Entry.Node = FPtr; Args.push_back(Entry);
1466 Entry.Node = Nest; Args.push_back(Entry);
1468 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1469 TargetLowering::CallLoweringInfo CLI(Chain,
1470 Type::getVoidTy(*DAG.getContext()),
1471 false, false, false, false, 0,
1473 /*isTailCall=*/false,
1474 /*doesNotRet=*/false,
1475 /*isReturnValueUsed=*/true,
1476 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1478 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1480 return CallResult.second;
1483 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1484 const PPCSubtarget &Subtarget) const {
1485 MachineFunction &MF = DAG.getMachineFunction();
1486 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1488 DebugLoc dl = Op.getDebugLoc();
1490 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1491 // vastart just stores the address of the VarArgsFrameIndex slot into the
1492 // memory location argument.
1493 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1494 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1495 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1496 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1497 MachinePointerInfo(SV),
1501 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1502 // We suppose the given va_list is already allocated.
1505 // char gpr; /* index into the array of 8 GPRs
1506 // * stored in the register save area
1507 // * gpr=0 corresponds to r3,
1508 // * gpr=1 to r4, etc.
1510 // char fpr; /* index into the array of 8 FPRs
1511 // * stored in the register save area
1512 // * fpr=0 corresponds to f1,
1513 // * fpr=1 to f2, etc.
1515 // char *overflow_arg_area;
1516 // /* location on stack that holds
1517 // * the next overflow argument
1519 // char *reg_save_area;
1520 // /* where r3:r10 and f1:f8 (if saved)
1526 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1527 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1530 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1532 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1534 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1537 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1538 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1540 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1541 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1543 uint64_t FPROffset = 1;
1544 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1546 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1548 // Store first byte : number of int regs
1549 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1551 MachinePointerInfo(SV),
1552 MVT::i8, false, false, 0);
1553 uint64_t nextOffset = FPROffset;
1554 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1557 // Store second byte : number of float regs
1558 SDValue secondStore =
1559 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1560 MachinePointerInfo(SV, nextOffset), MVT::i8,
1562 nextOffset += StackOffset;
1563 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1565 // Store second word : arguments given on stack
1566 SDValue thirdStore =
1567 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1568 MachinePointerInfo(SV, nextOffset),
1570 nextOffset += FrameOffset;
1571 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1573 // Store third word : arguments given in registers
1574 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1575 MachinePointerInfo(SV, nextOffset),
1580 #include "PPCGenCallingConv.inc"
1582 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1583 CCValAssign::LocInfo &LocInfo,
1584 ISD::ArgFlagsTy &ArgFlags,
1589 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1591 CCValAssign::LocInfo &LocInfo,
1592 ISD::ArgFlagsTy &ArgFlags,
1594 static const uint16_t ArgRegs[] = {
1595 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1596 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1598 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1600 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1602 // Skip one register if the first unallocated register has an even register
1603 // number and there are still argument registers available which have not been
1604 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1605 // need to skip a register if RegNum is odd.
1606 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1607 State.AllocateReg(ArgRegs[RegNum]);
1610 // Always return false here, as this function only makes sure that the first
1611 // unallocated register has an odd register number and does not actually
1612 // allocate a register for the current argument.
1616 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1618 CCValAssign::LocInfo &LocInfo,
1619 ISD::ArgFlagsTy &ArgFlags,
1621 static const uint16_t ArgRegs[] = {
1622 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1626 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1628 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1630 // If there is only one Floating-point register left we need to put both f64
1631 // values of a split ppc_fp128 value on the stack.
1632 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1633 State.AllocateReg(ArgRegs[RegNum]);
1636 // Always return false here, as this function only makes sure that the two f64
1637 // values a ppc_fp128 value is split into are both passed in registers or both
1638 // passed on the stack and does not actually allocate a register for the
1639 // current argument.
1643 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1645 static const uint16_t *GetFPR() {
1646 static const uint16_t FPR[] = {
1647 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1648 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1654 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1656 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1657 unsigned PtrByteSize) {
1658 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1659 if (Flags.isByVal())
1660 ArgSize = Flags.getByValSize();
1661 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1667 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1668 CallingConv::ID CallConv, bool isVarArg,
1669 const SmallVectorImpl<ISD::InputArg>
1671 DebugLoc dl, SelectionDAG &DAG,
1672 SmallVectorImpl<SDValue> &InVals)
1674 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1675 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1678 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1684 PPCTargetLowering::LowerFormalArguments_SVR4(
1686 CallingConv::ID CallConv, bool isVarArg,
1687 const SmallVectorImpl<ISD::InputArg>
1689 DebugLoc dl, SelectionDAG &DAG,
1690 SmallVectorImpl<SDValue> &InVals) const {
1692 // 32-bit SVR4 ABI Stack Frame Layout:
1693 // +-----------------------------------+
1694 // +--> | Back chain |
1695 // | +-----------------------------------+
1696 // | | Floating-point register save area |
1697 // | +-----------------------------------+
1698 // | | General register save area |
1699 // | +-----------------------------------+
1700 // | | CR save word |
1701 // | +-----------------------------------+
1702 // | | VRSAVE save word |
1703 // | +-----------------------------------+
1704 // | | Alignment padding |
1705 // | +-----------------------------------+
1706 // | | Vector register save area |
1707 // | +-----------------------------------+
1708 // | | Local variable space |
1709 // | +-----------------------------------+
1710 // | | Parameter list area |
1711 // | +-----------------------------------+
1712 // | | LR save word |
1713 // | +-----------------------------------+
1714 // SP--> +--- | Back chain |
1715 // +-----------------------------------+
1718 // System V Application Binary Interface PowerPC Processor Supplement
1719 // AltiVec Technology Programming Interface Manual
1721 MachineFunction &MF = DAG.getMachineFunction();
1722 MachineFrameInfo *MFI = MF.getFrameInfo();
1723 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1725 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1726 // Potential tail calls could cause overwriting of argument stack slots.
1727 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1728 (CallConv == CallingConv::Fast));
1729 unsigned PtrByteSize = 4;
1731 // Assign locations to all of the incoming arguments.
1732 SmallVector<CCValAssign, 16> ArgLocs;
1733 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1734 getTargetMachine(), ArgLocs, *DAG.getContext());
1736 // Reserve space for the linkage area on the stack.
1737 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1739 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1741 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1742 CCValAssign &VA = ArgLocs[i];
1744 // Arguments stored in registers.
1745 if (VA.isRegLoc()) {
1746 const TargetRegisterClass *RC;
1747 EVT ValVT = VA.getValVT();
1749 switch (ValVT.getSimpleVT().SimpleTy) {
1751 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1753 RC = &PPC::GPRCRegClass;
1756 RC = &PPC::F4RCRegClass;
1759 RC = &PPC::F8RCRegClass;
1765 RC = &PPC::VRRCRegClass;
1769 // Transform the arguments stored in physical registers into virtual ones.
1770 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1771 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1773 InVals.push_back(ArgValue);
1775 // Argument stored in memory.
1776 assert(VA.isMemLoc());
1778 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1779 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1782 // Create load nodes to retrieve arguments from the stack.
1783 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1784 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1785 MachinePointerInfo(),
1786 false, false, false, 0));
1790 // Assign locations to all of the incoming aggregate by value arguments.
1791 // Aggregates passed by value are stored in the local variable space of the
1792 // caller's stack frame, right above the parameter list area.
1793 SmallVector<CCValAssign, 16> ByValArgLocs;
1794 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1795 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1797 // Reserve stack space for the allocations in CCInfo.
1798 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1800 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1802 // Area that is at least reserved in the caller of this function.
1803 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1805 // Set the size that is at least reserved in caller of this function. Tail
1806 // call optimized function's reserved stack space needs to be aligned so that
1807 // taking the difference between two stack areas will result in an aligned
1809 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1812 std::max(MinReservedArea,
1813 PPCFrameLowering::getMinCallFrameSize(false, false));
1815 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1816 getStackAlignment();
1817 unsigned AlignMask = TargetAlign-1;
1818 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1820 FI->setMinReservedArea(MinReservedArea);
1822 SmallVector<SDValue, 8> MemOps;
1824 // If the function takes variable number of arguments, make a frame index for
1825 // the start of the first vararg value... for expansion of llvm.va_start.
1827 static const uint16_t GPArgRegs[] = {
1828 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1829 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1831 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1833 static const uint16_t FPArgRegs[] = {
1834 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1837 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1839 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1841 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1844 // Make room for NumGPArgRegs and NumFPArgRegs.
1845 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1846 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1848 FuncInfo->setVarArgsStackOffset(
1849 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1850 CCInfo.getNextStackOffset(), true));
1852 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1853 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1855 // The fixed integer arguments of a variadic function are stored to the
1856 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1857 // the result of va_next.
1858 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1859 // Get an existing live-in vreg, or add a new one.
1860 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1862 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1864 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1865 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1866 MachinePointerInfo(), false, false, 0);
1867 MemOps.push_back(Store);
1868 // Increment the address by four for the next argument to store
1869 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1870 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1873 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1875 // The double arguments are stored to the VarArgsFrameIndex
1877 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1878 // Get an existing live-in vreg, or add a new one.
1879 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1881 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1883 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1884 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1885 MachinePointerInfo(), false, false, 0);
1886 MemOps.push_back(Store);
1887 // Increment the address by eight for the next argument to store
1888 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1890 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1894 if (!MemOps.empty())
1895 Chain = DAG.getNode(ISD::TokenFactor, dl,
1896 MVT::Other, &MemOps[0], MemOps.size());
1902 PPCTargetLowering::LowerFormalArguments_Darwin(
1904 CallingConv::ID CallConv, bool isVarArg,
1905 const SmallVectorImpl<ISD::InputArg>
1907 DebugLoc dl, SelectionDAG &DAG,
1908 SmallVectorImpl<SDValue> &InVals) const {
1909 // TODO: add description of PPC stack frame format, or at least some docs.
1911 MachineFunction &MF = DAG.getMachineFunction();
1912 MachineFrameInfo *MFI = MF.getFrameInfo();
1913 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1915 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1916 bool isPPC64 = PtrVT == MVT::i64;
1917 // Potential tail calls could cause overwriting of argument stack slots.
1918 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1919 (CallConv == CallingConv::Fast));
1920 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1922 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1923 // Area that is at least reserved in caller of this function.
1924 unsigned MinReservedArea = ArgOffset;
1926 static const uint16_t GPR_32[] = { // 32-bit registers.
1927 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1928 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1930 static const uint16_t GPR_64[] = { // 64-bit registers.
1931 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1932 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1935 static const uint16_t *FPR = GetFPR();
1937 static const uint16_t VR[] = {
1938 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1939 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1942 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1943 const unsigned Num_FPR_Regs = 13;
1944 const unsigned Num_VR_Regs = array_lengthof( VR);
1946 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1948 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
1950 // In 32-bit non-varargs functions, the stack space for vectors is after the
1951 // stack space for non-vectors. We do not use this space unless we have
1952 // too many vectors to fit in registers, something that only occurs in
1953 // constructed examples:), but we have to walk the arglist to figure
1954 // that out...for the pathological case, compute VecArgOffset as the
1955 // start of the vector parameter area. Computing VecArgOffset is the
1956 // entire point of the following loop.
1957 unsigned VecArgOffset = ArgOffset;
1958 if (!isVarArg && !isPPC64) {
1959 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1961 EVT ObjectVT = Ins[ArgNo].VT;
1962 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1964 if (Flags.isByVal()) {
1965 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1966 unsigned ObjSize = Flags.getByValSize();
1968 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1969 VecArgOffset += ArgSize;
1973 switch(ObjectVT.getSimpleVT().SimpleTy) {
1974 default: llvm_unreachable("Unhandled argument type!");
1977 VecArgOffset += isPPC64 ? 8 : 4;
1979 case MVT::i64: // PPC64
1987 // Nothing to do, we're only looking at Nonvector args here.
1992 // We've found where the vector parameter area in memory is. Skip the
1993 // first 12 parameters; these don't use that memory.
1994 VecArgOffset = ((VecArgOffset+15)/16)*16;
1995 VecArgOffset += 12*16;
1997 // Add DAG nodes to load the arguments or copy them out of registers. On
1998 // entry to a function on PPC, the arguments start after the linkage area,
1999 // although the first ones are often in registers.
2001 SmallVector<SDValue, 8> MemOps;
2002 unsigned nAltivecParamsAtEnd = 0;
2003 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2005 bool needsLoad = false;
2006 EVT ObjectVT = Ins[ArgNo].VT;
2007 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2008 unsigned ArgSize = ObjSize;
2009 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2011 unsigned CurArgOffset = ArgOffset;
2013 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2014 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2015 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2016 if (isVarArg || isPPC64) {
2017 MinReservedArea = ((MinReservedArea+15)/16)*16;
2018 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2021 } else nAltivecParamsAtEnd++;
2023 // Calculate min reserved area.
2024 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2028 // FIXME the codegen can be much improved in some cases.
2029 // We do not have to keep everything in memory.
2030 if (Flags.isByVal()) {
2031 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2032 ObjSize = Flags.getByValSize();
2033 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2034 // Objects of size 1 and 2 are right justified, everything else is
2035 // left justified. This means the memory address is adjusted forwards.
2036 if (ObjSize==1 || ObjSize==2) {
2037 CurArgOffset = CurArgOffset + (4 - ObjSize);
2039 // The value of the object is its address.
2040 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2041 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2042 InVals.push_back(FIN);
2043 if (ObjSize==1 || ObjSize==2) {
2044 if (GPR_idx != Num_GPR_Regs) {
2047 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2049 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2050 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2051 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2052 MachinePointerInfo(),
2053 ObjSize==1 ? MVT::i8 : MVT::i16,
2055 MemOps.push_back(Store);
2059 ArgOffset += PtrByteSize;
2063 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2064 // Store whatever pieces of the object are in registers
2065 // to memory. ArgVal will be address of the beginning of
2067 if (GPR_idx != Num_GPR_Regs) {
2070 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2072 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2073 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2074 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2075 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2076 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2077 MachinePointerInfo(),
2079 MemOps.push_back(Store);
2081 ArgOffset += PtrByteSize;
2083 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2090 switch (ObjectVT.getSimpleVT().SimpleTy) {
2091 default: llvm_unreachable("Unhandled argument type!");
2094 if (GPR_idx != Num_GPR_Regs) {
2095 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2096 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2100 ArgSize = PtrByteSize;
2102 // All int arguments reserve stack space in the Darwin ABI.
2103 ArgOffset += PtrByteSize;
2107 case MVT::i64: // PPC64
2108 if (GPR_idx != Num_GPR_Regs) {
2109 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2110 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2112 if (ObjectVT == MVT::i32) {
2113 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2114 // value to MVT::i64 and then truncate to the correct register size.
2116 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2117 DAG.getValueType(ObjectVT));
2118 else if (Flags.isZExt())
2119 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2120 DAG.getValueType(ObjectVT));
2122 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2128 ArgSize = PtrByteSize;
2130 // All int arguments reserve stack space in the Darwin ABI.
2136 // Every 4 bytes of argument space consumes one of the GPRs available for
2137 // argument passing.
2138 if (GPR_idx != Num_GPR_Regs) {
2140 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2143 if (FPR_idx != Num_FPR_Regs) {
2146 if (ObjectVT == MVT::f32)
2147 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2149 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2151 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2157 // All FP arguments reserve stack space in the Darwin ABI.
2158 ArgOffset += isPPC64 ? 8 : ObjSize;
2164 // Note that vector arguments in registers don't reserve stack space,
2165 // except in varargs functions.
2166 if (VR_idx != Num_VR_Regs) {
2167 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2168 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2170 while ((ArgOffset % 16) != 0) {
2171 ArgOffset += PtrByteSize;
2172 if (GPR_idx != Num_GPR_Regs)
2176 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2180 if (!isVarArg && !isPPC64) {
2181 // Vectors go after all the nonvectors.
2182 CurArgOffset = VecArgOffset;
2185 // Vectors are aligned.
2186 ArgOffset = ((ArgOffset+15)/16)*16;
2187 CurArgOffset = ArgOffset;
2195 // We need to load the argument to a virtual register if we determined above
2196 // that we ran out of physical registers of the appropriate type.
2198 int FI = MFI->CreateFixedObject(ObjSize,
2199 CurArgOffset + (ArgSize - ObjSize),
2201 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2202 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2203 false, false, false, 0);
2206 InVals.push_back(ArgVal);
2209 // Set the size that is at least reserved in caller of this function. Tail
2210 // call optimized function's reserved stack space needs to be aligned so that
2211 // taking the difference between two stack areas will result in an aligned
2213 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2214 // Add the Altivec parameters at the end, if needed.
2215 if (nAltivecParamsAtEnd) {
2216 MinReservedArea = ((MinReservedArea+15)/16)*16;
2217 MinReservedArea += 16*nAltivecParamsAtEnd;
2220 std::max(MinReservedArea,
2221 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2222 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2223 getStackAlignment();
2224 unsigned AlignMask = TargetAlign-1;
2225 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2226 FI->setMinReservedArea(MinReservedArea);
2228 // If the function takes variable number of arguments, make a frame index for
2229 // the start of the first vararg value... for expansion of llvm.va_start.
2231 int Depth = ArgOffset;
2233 FuncInfo->setVarArgsFrameIndex(
2234 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2236 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2238 // If this function is vararg, store any remaining integer argument regs
2239 // to their spots on the stack so that they may be loaded by deferencing the
2240 // result of va_next.
2241 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2245 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2247 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2249 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2250 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2251 MachinePointerInfo(), false, false, 0);
2252 MemOps.push_back(Store);
2253 // Increment the address by four for the next argument to store
2254 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2255 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2259 if (!MemOps.empty())
2260 Chain = DAG.getNode(ISD::TokenFactor, dl,
2261 MVT::Other, &MemOps[0], MemOps.size());
2266 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2267 /// linkage area for the Darwin ABI.
2269 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2273 const SmallVectorImpl<ISD::OutputArg>
2275 const SmallVectorImpl<SDValue> &OutVals,
2276 unsigned &nAltivecParamsAtEnd) {
2277 // Count how many bytes are to be pushed on the stack, including the linkage
2278 // area, and parameter passing area. We start with 24/48 bytes, which is
2279 // prereserved space for [SP][CR][LR][3 x unused].
2280 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2281 unsigned NumOps = Outs.size();
2282 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2284 // Add up all the space actually used.
2285 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2286 // they all go in registers, but we must reserve stack space for them for
2287 // possible use by the caller. In varargs or 64-bit calls, parameters are
2288 // assigned stack space in order, with padding so Altivec parameters are
2290 nAltivecParamsAtEnd = 0;
2291 for (unsigned i = 0; i != NumOps; ++i) {
2292 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2293 EVT ArgVT = Outs[i].VT;
2294 // Varargs Altivec parameters are padded to a 16 byte boundary.
2295 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2296 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2297 if (!isVarArg && !isPPC64) {
2298 // Non-varargs Altivec parameters go after all the non-Altivec
2299 // parameters; handle those later so we know how much padding we need.
2300 nAltivecParamsAtEnd++;
2303 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2304 NumBytes = ((NumBytes+15)/16)*16;
2306 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2309 // Allow for Altivec parameters at the end, if needed.
2310 if (nAltivecParamsAtEnd) {
2311 NumBytes = ((NumBytes+15)/16)*16;
2312 NumBytes += 16*nAltivecParamsAtEnd;
2315 // The prolog code of the callee may store up to 8 GPR argument registers to
2316 // the stack, allowing va_start to index over them in memory if its varargs.
2317 // Because we cannot tell if this is needed on the caller side, we have to
2318 // conservatively assume that it is needed. As such, make sure we have at
2319 // least enough stack space for the caller to store the 8 GPRs.
2320 NumBytes = std::max(NumBytes,
2321 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2323 // Tail call needs the stack to be aligned.
2324 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2325 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2326 getFrameLowering()->getStackAlignment();
2327 unsigned AlignMask = TargetAlign-1;
2328 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2334 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2335 /// adjusted to accommodate the arguments for the tailcall.
2336 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2337 unsigned ParamSize) {
2339 if (!isTailCall) return 0;
2341 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2342 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2343 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2344 // Remember only if the new adjustement is bigger.
2345 if (SPDiff < FI->getTailCallSPDelta())
2346 FI->setTailCallSPDelta(SPDiff);
2351 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2352 /// for tail call optimization. Targets which want to do tail call
2353 /// optimization should implement this function.
2355 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2356 CallingConv::ID CalleeCC,
2358 const SmallVectorImpl<ISD::InputArg> &Ins,
2359 SelectionDAG& DAG) const {
2360 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2363 // Variable argument functions are not supported.
2367 MachineFunction &MF = DAG.getMachineFunction();
2368 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2369 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2370 // Functions containing by val parameters are not supported.
2371 for (unsigned i = 0; i != Ins.size(); i++) {
2372 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2373 if (Flags.isByVal()) return false;
2376 // Non PIC/GOT tail calls are supported.
2377 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2380 // At the moment we can only do local tail calls (in same module, hidden
2381 // or protected) if we are generating PIC.
2382 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2383 return G->getGlobal()->hasHiddenVisibility()
2384 || G->getGlobal()->hasProtectedVisibility();
2390 /// isCallCompatibleAddress - Return the immediate to use if the specified
2391 /// 32-bit value is representable in the immediate field of a BxA instruction.
2392 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2393 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2396 int Addr = C->getZExtValue();
2397 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2398 (Addr << 6 >> 6) != Addr)
2399 return 0; // Top 6 bits have to be sext of immediate.
2401 return DAG.getConstant((int)C->getZExtValue() >> 2,
2402 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2407 struct TailCallArgumentInfo {
2412 TailCallArgumentInfo() : FrameIdx(0) {}
2417 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2419 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2421 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2422 SmallVector<SDValue, 8> &MemOpChains,
2424 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2425 SDValue Arg = TailCallArgs[i].Arg;
2426 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2427 int FI = TailCallArgs[i].FrameIdx;
2428 // Store relative to framepointer.
2429 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2430 MachinePointerInfo::getFixedStack(FI),
2435 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2436 /// the appropriate stack slot for the tail call optimized function call.
2437 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2438 MachineFunction &MF,
2447 // Calculate the new stack slot for the return address.
2448 int SlotSize = isPPC64 ? 8 : 4;
2449 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2451 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2452 NewRetAddrLoc, true);
2453 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2454 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2455 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2456 MachinePointerInfo::getFixedStack(NewRetAddr),
2459 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2460 // slot as the FP is never overwritten.
2463 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2464 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2466 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2467 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2468 MachinePointerInfo::getFixedStack(NewFPIdx),
2475 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2476 /// the position of the argument.
2478 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2479 SDValue Arg, int SPDiff, unsigned ArgOffset,
2480 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2481 int Offset = ArgOffset + SPDiff;
2482 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2483 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2484 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2485 SDValue FIN = DAG.getFrameIndex(FI, VT);
2486 TailCallArgumentInfo Info;
2488 Info.FrameIdxOp = FIN;
2490 TailCallArguments.push_back(Info);
2493 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2494 /// stack slot. Returns the chain as result and the loaded frame pointers in
2495 /// LROpOut/FPOpout. Used when tail calling.
2496 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2502 DebugLoc dl) const {
2504 // Load the LR and FP stack slot for later adjusting.
2505 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2506 LROpOut = getReturnAddrFrameIndex(DAG);
2507 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2508 false, false, false, 0);
2509 Chain = SDValue(LROpOut.getNode(), 1);
2511 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2512 // slot as the FP is never overwritten.
2514 FPOpOut = getFramePointerFrameIndex(DAG);
2515 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2516 false, false, false, 0);
2517 Chain = SDValue(FPOpOut.getNode(), 1);
2523 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2524 /// by "Src" to address "Dst" of size "Size". Alignment information is
2525 /// specified by the specific parameter attribute. The copy will be passed as
2526 /// a byval function parameter.
2527 /// Sometimes what we are copying is the end of a larger object, the part that
2528 /// does not fit in registers.
2530 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2531 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2533 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2534 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2535 false, false, MachinePointerInfo(0),
2536 MachinePointerInfo(0));
2539 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2542 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2543 SDValue Arg, SDValue PtrOff, int SPDiff,
2544 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2545 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2546 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2548 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2553 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2555 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2556 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2557 DAG.getConstant(ArgOffset, PtrVT));
2559 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2560 MachinePointerInfo(), false, false, 0));
2561 // Calculate and remember argument location.
2562 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2567 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2568 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2569 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2570 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2571 MachineFunction &MF = DAG.getMachineFunction();
2573 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2574 // might overwrite each other in case of tail call optimization.
2575 SmallVector<SDValue, 8> MemOpChains2;
2576 // Do not flag preceding copytoreg stuff together with the following stuff.
2578 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2580 if (!MemOpChains2.empty())
2581 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2582 &MemOpChains2[0], MemOpChains2.size());
2584 // Store the return address to the appropriate stack slot.
2585 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2586 isPPC64, isDarwinABI, dl);
2588 // Emit callseq_end just before tailcall node.
2589 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2590 DAG.getIntPtrConstant(0, true), InFlag);
2591 InFlag = Chain.getValue(1);
2595 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2596 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2597 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2598 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2599 const PPCSubtarget &PPCSubTarget) {
2601 bool isPPC64 = PPCSubTarget.isPPC64();
2602 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2604 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2605 NodeTys.push_back(MVT::Other); // Returns a chain
2606 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2608 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2610 bool needIndirectCall = true;
2611 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2612 // If this is an absolute destination address, use the munged value.
2613 Callee = SDValue(Dest, 0);
2614 needIndirectCall = false;
2617 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2618 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2619 // Use indirect calls for ALL functions calls in JIT mode, since the
2620 // far-call stubs may be outside relocation limits for a BL instruction.
2621 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2622 unsigned OpFlags = 0;
2623 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2624 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2625 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2626 (G->getGlobal()->isDeclaration() ||
2627 G->getGlobal()->isWeakForLinker())) {
2628 // PC-relative references to external symbols should go through $stub,
2629 // unless we're building with the leopard linker or later, which
2630 // automatically synthesizes these stubs.
2631 OpFlags = PPCII::MO_DARWIN_STUB;
2634 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2635 // every direct call is) turn it into a TargetGlobalAddress /
2636 // TargetExternalSymbol node so that legalize doesn't hack it.
2637 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2638 Callee.getValueType(),
2640 needIndirectCall = false;
2644 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2645 unsigned char OpFlags = 0;
2647 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2648 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2649 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2650 // PC-relative references to external symbols should go through $stub,
2651 // unless we're building with the leopard linker or later, which
2652 // automatically synthesizes these stubs.
2653 OpFlags = PPCII::MO_DARWIN_STUB;
2656 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2658 needIndirectCall = false;
2661 if (needIndirectCall) {
2662 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2663 // to do the call, we can't use PPCISD::CALL.
2664 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2666 if (isSVR4ABI && isPPC64) {
2667 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2668 // entry point, but to the function descriptor (the function entry point
2669 // address is part of the function descriptor though).
2670 // The function descriptor is a three doubleword structure with the
2671 // following fields: function entry point, TOC base address and
2672 // environment pointer.
2673 // Thus for a call through a function pointer, the following actions need
2675 // 1. Save the TOC of the caller in the TOC save area of its stack
2676 // frame (this is done in LowerCall_Darwin()).
2677 // 2. Load the address of the function entry point from the function
2679 // 3. Load the TOC of the callee from the function descriptor into r2.
2680 // 4. Load the environment pointer from the function descriptor into
2682 // 5. Branch to the function entry point address.
2683 // 6. On return of the callee, the TOC of the caller needs to be
2684 // restored (this is done in FinishCall()).
2686 // All those operations are flagged together to ensure that no other
2687 // operations can be scheduled in between. E.g. without flagging the
2688 // operations together, a TOC access in the caller could be scheduled
2689 // between the load of the callee TOC and the branch to the callee, which
2690 // results in the TOC access going through the TOC of the callee instead
2691 // of going through the TOC of the caller, which leads to incorrect code.
2693 // Load the address of the function entry point from the function
2695 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2696 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2697 InFlag.getNode() ? 3 : 2);
2698 Chain = LoadFuncPtr.getValue(1);
2699 InFlag = LoadFuncPtr.getValue(2);
2701 // Load environment pointer into r11.
2702 // Offset of the environment pointer within the function descriptor.
2703 SDValue PtrOff = DAG.getIntPtrConstant(16);
2705 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2706 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2708 Chain = LoadEnvPtr.getValue(1);
2709 InFlag = LoadEnvPtr.getValue(2);
2711 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2713 Chain = EnvVal.getValue(0);
2714 InFlag = EnvVal.getValue(1);
2716 // Load TOC of the callee into r2. We are using a target-specific load
2717 // with r2 hard coded, because the result of a target-independent load
2718 // would never go directly into r2, since r2 is a reserved register (which
2719 // prevents the register allocator from allocating it), resulting in an
2720 // additional register being allocated and an unnecessary move instruction
2722 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2723 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2725 Chain = LoadTOCPtr.getValue(0);
2726 InFlag = LoadTOCPtr.getValue(1);
2728 MTCTROps[0] = Chain;
2729 MTCTROps[1] = LoadFuncPtr;
2730 MTCTROps[2] = InFlag;
2733 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2734 2 + (InFlag.getNode() != 0));
2735 InFlag = Chain.getValue(1);
2738 NodeTys.push_back(MVT::Other);
2739 NodeTys.push_back(MVT::Glue);
2740 Ops.push_back(Chain);
2741 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2743 // Add CTR register as callee so a bctr can be emitted later.
2745 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2748 // If this is a direct call, pass the chain and the callee.
2749 if (Callee.getNode()) {
2750 Ops.push_back(Chain);
2751 Ops.push_back(Callee);
2753 // If this is a tail call add stack pointer delta.
2755 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2757 // Add argument registers to the end of the list so that they are known live
2759 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2760 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2761 RegsToPass[i].second.getValueType()));
2767 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2768 CallingConv::ID CallConv, bool isVarArg,
2769 const SmallVectorImpl<ISD::InputArg> &Ins,
2770 DebugLoc dl, SelectionDAG &DAG,
2771 SmallVectorImpl<SDValue> &InVals) const {
2773 SmallVector<CCValAssign, 16> RVLocs;
2774 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2775 getTargetMachine(), RVLocs, *DAG.getContext());
2776 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2778 // Copy all of the result registers out of their specified physreg.
2779 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2780 CCValAssign &VA = RVLocs[i];
2781 EVT VT = VA.getValVT();
2782 assert(VA.isRegLoc() && "Can only return in registers!");
2783 Chain = DAG.getCopyFromReg(Chain, dl,
2784 VA.getLocReg(), VT, InFlag).getValue(1);
2785 InVals.push_back(Chain.getValue(0));
2786 InFlag = Chain.getValue(2);
2793 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2794 bool isTailCall, bool isVarArg,
2796 SmallVector<std::pair<unsigned, SDValue>, 8>
2798 SDValue InFlag, SDValue Chain,
2800 int SPDiff, unsigned NumBytes,
2801 const SmallVectorImpl<ISD::InputArg> &Ins,
2802 SmallVectorImpl<SDValue> &InVals) const {
2803 std::vector<EVT> NodeTys;
2804 SmallVector<SDValue, 8> Ops;
2805 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2806 isTailCall, RegsToPass, Ops, NodeTys,
2809 // When performing tail call optimization the callee pops its arguments off
2810 // the stack. Account for this here so these bytes can be pushed back on in
2811 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2812 int BytesCalleePops =
2813 (CallConv == CallingConv::Fast &&
2814 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
2816 // Add a register mask operand representing the call-preserved registers.
2817 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2818 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2819 assert(Mask && "Missing call preserved mask for calling convention");
2820 Ops.push_back(DAG.getRegisterMask(Mask));
2822 if (InFlag.getNode())
2823 Ops.push_back(InFlag);
2827 // If this is the first return lowered for this function, add the regs
2828 // to the liveout set for the function.
2829 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2830 SmallVector<CCValAssign, 16> RVLocs;
2831 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2832 getTargetMachine(), RVLocs, *DAG.getContext());
2833 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2834 for (unsigned i = 0; i != RVLocs.size(); ++i)
2835 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2838 assert(((Callee.getOpcode() == ISD::Register &&
2839 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2840 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2841 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2842 isa<ConstantSDNode>(Callee)) &&
2843 "Expecting an global address, external symbol, absolute value or register");
2845 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2848 // Add a NOP immediately after the branch instruction when using the 64-bit
2849 // SVR4 ABI. At link time, if caller and callee are in a different module and
2850 // thus have a different TOC, the call will be replaced with a call to a stub
2851 // function which saves the current TOC, loads the TOC of the callee and
2852 // branches to the callee. The NOP will be replaced with a load instruction
2853 // which restores the TOC of the caller from the TOC save slot of the current
2854 // stack frame. If caller and callee belong to the same module (and have the
2855 // same TOC), the NOP will remain unchanged.
2857 bool needsTOCRestore = false;
2858 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2859 if (CallOpc == PPCISD::BCTRL_SVR4) {
2860 // This is a call through a function pointer.
2861 // Restore the caller TOC from the save area into R2.
2862 // See PrepareCall() for more information about calls through function
2863 // pointers in the 64-bit SVR4 ABI.
2864 // We are using a target-specific load with r2 hard coded, because the
2865 // result of a target-independent load would never go directly into r2,
2866 // since r2 is a reserved register (which prevents the register allocator
2867 // from allocating it), resulting in an additional register being
2868 // allocated and an unnecessary move instruction being generated.
2869 needsTOCRestore = true;
2870 } else if (CallOpc == PPCISD::CALL_SVR4) {
2871 // Otherwise insert NOP.
2872 CallOpc = PPCISD::CALL_NOP_SVR4;
2876 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2877 InFlag = Chain.getValue(1);
2879 if (needsTOCRestore) {
2880 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2881 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2882 InFlag = Chain.getValue(1);
2885 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2886 DAG.getIntPtrConstant(BytesCalleePops, true),
2889 InFlag = Chain.getValue(1);
2891 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2892 Ins, dl, DAG, InVals);
2896 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2897 SmallVectorImpl<SDValue> &InVals) const {
2898 SelectionDAG &DAG = CLI.DAG;
2899 DebugLoc &dl = CLI.DL;
2900 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2901 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2902 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2903 SDValue Chain = CLI.Chain;
2904 SDValue Callee = CLI.Callee;
2905 bool &isTailCall = CLI.IsTailCall;
2906 CallingConv::ID CallConv = CLI.CallConv;
2907 bool isVarArg = CLI.IsVarArg;
2910 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2913 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2914 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2915 isTailCall, Outs, OutVals, Ins,
2918 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2919 isTailCall, Outs, OutVals, Ins,
2924 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2925 CallingConv::ID CallConv, bool isVarArg,
2927 const SmallVectorImpl<ISD::OutputArg> &Outs,
2928 const SmallVectorImpl<SDValue> &OutVals,
2929 const SmallVectorImpl<ISD::InputArg> &Ins,
2930 DebugLoc dl, SelectionDAG &DAG,
2931 SmallVectorImpl<SDValue> &InVals) const {
2932 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2933 // of the 32-bit SVR4 ABI stack frame layout.
2935 assert((CallConv == CallingConv::C ||
2936 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2938 unsigned PtrByteSize = 4;
2940 MachineFunction &MF = DAG.getMachineFunction();
2942 // Mark this function as potentially containing a function that contains a
2943 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2944 // and restoring the callers stack pointer in this functions epilog. This is
2945 // done because by tail calling the called function might overwrite the value
2946 // in this function's (MF) stack pointer stack slot 0(SP).
2947 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2948 CallConv == CallingConv::Fast)
2949 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2951 // Count how many bytes are to be pushed on the stack, including the linkage
2952 // area, parameter list area and the part of the local variable space which
2953 // contains copies of aggregates which are passed by value.
2955 // Assign locations to all of the outgoing arguments.
2956 SmallVector<CCValAssign, 16> ArgLocs;
2957 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2958 getTargetMachine(), ArgLocs, *DAG.getContext());
2960 // Reserve space for the linkage area on the stack.
2961 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2964 // Handle fixed and variable vector arguments differently.
2965 // Fixed vector arguments go into registers as long as registers are
2966 // available. Variable vector arguments always go into memory.
2967 unsigned NumArgs = Outs.size();
2969 for (unsigned i = 0; i != NumArgs; ++i) {
2970 MVT ArgVT = Outs[i].VT;
2971 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2974 if (Outs[i].IsFixed) {
2975 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2978 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2984 errs() << "Call operand #" << i << " has unhandled type "
2985 << EVT(ArgVT).getEVTString() << "\n";
2987 llvm_unreachable(0);
2991 // All arguments are treated the same.
2992 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2995 // Assign locations to all of the outgoing aggregate by value arguments.
2996 SmallVector<CCValAssign, 16> ByValArgLocs;
2997 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2998 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3000 // Reserve stack space for the allocations in CCInfo.
3001 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3003 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3005 // Size of the linkage area, parameter list area and the part of the local
3006 // space variable where copies of aggregates which are passed by value are
3008 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3010 // Calculate by how many bytes the stack has to be adjusted in case of tail
3011 // call optimization.
3012 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3014 // Adjust the stack pointer for the new arguments...
3015 // These operations are automatically eliminated by the prolog/epilog pass
3016 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3017 SDValue CallSeqStart = Chain;
3019 // Load the return address and frame pointer so it can be moved somewhere else
3022 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3025 // Set up a copy of the stack pointer for use loading and storing any
3026 // arguments that may not fit in the registers available for argument
3028 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3030 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3031 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3032 SmallVector<SDValue, 8> MemOpChains;
3034 bool seenFloatArg = false;
3035 // Walk the register/memloc assignments, inserting copies/loads.
3036 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3039 CCValAssign &VA = ArgLocs[i];
3040 SDValue Arg = OutVals[i];
3041 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3043 if (Flags.isByVal()) {
3044 // Argument is an aggregate which is passed by value, thus we need to
3045 // create a copy of it in the local variable space of the current stack
3046 // frame (which is the stack frame of the caller) and pass the address of
3047 // this copy to the callee.
3048 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3049 CCValAssign &ByValVA = ByValArgLocs[j++];
3050 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3052 // Memory reserved in the local variable space of the callers stack frame.
3053 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3055 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3056 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3058 // Create a copy of the argument in the local area of the current
3060 SDValue MemcpyCall =
3061 CreateCopyOfByValArgument(Arg, PtrOff,
3062 CallSeqStart.getNode()->getOperand(0),
3065 // This must go outside the CALLSEQ_START..END.
3066 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3067 CallSeqStart.getNode()->getOperand(1));
3068 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3069 NewCallSeqStart.getNode());
3070 Chain = CallSeqStart = NewCallSeqStart;
3072 // Pass the address of the aggregate copy on the stack either in a
3073 // physical register or in the parameter list area of the current stack
3074 // frame to the callee.
3078 if (VA.isRegLoc()) {
3079 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3080 // Put argument in a physical register.
3081 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3083 // Put argument in the parameter list area of the current stack frame.
3084 assert(VA.isMemLoc());
3085 unsigned LocMemOffset = VA.getLocMemOffset();
3088 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3089 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3091 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3092 MachinePointerInfo(),
3095 // Calculate and remember argument location.
3096 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3102 if (!MemOpChains.empty())
3103 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3104 &MemOpChains[0], MemOpChains.size());
3106 // Set CR6 to true if this is a vararg call with floating args passed in
3109 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3111 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3114 // Build a sequence of copy-to-reg nodes chained together with token chain
3115 // and flag operands which copy the outgoing args into the appropriate regs.
3117 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3118 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3119 RegsToPass[i].second, InFlag);
3120 InFlag = Chain.getValue(1);
3124 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3125 false, TailCallArguments);
3127 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3128 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3133 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3134 CallingConv::ID CallConv, bool isVarArg,
3136 const SmallVectorImpl<ISD::OutputArg> &Outs,
3137 const SmallVectorImpl<SDValue> &OutVals,
3138 const SmallVectorImpl<ISD::InputArg> &Ins,
3139 DebugLoc dl, SelectionDAG &DAG,
3140 SmallVectorImpl<SDValue> &InVals) const {
3142 unsigned NumOps = Outs.size();
3144 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3145 bool isPPC64 = PtrVT == MVT::i64;
3146 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3148 MachineFunction &MF = DAG.getMachineFunction();
3150 // Mark this function as potentially containing a function that contains a
3151 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3152 // and restoring the callers stack pointer in this functions epilog. This is
3153 // done because by tail calling the called function might overwrite the value
3154 // in this function's (MF) stack pointer stack slot 0(SP).
3155 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3156 CallConv == CallingConv::Fast)
3157 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3159 unsigned nAltivecParamsAtEnd = 0;
3161 // Count how many bytes are to be pushed on the stack, including the linkage
3162 // area, and parameter passing area. We start with 24/48 bytes, which is
3163 // prereserved space for [SP][CR][LR][3 x unused].
3165 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3167 nAltivecParamsAtEnd);
3169 // Calculate by how many bytes the stack has to be adjusted in case of tail
3170 // call optimization.
3171 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3173 // To protect arguments on the stack from being clobbered in a tail call,
3174 // force all the loads to happen before doing any other lowering.
3176 Chain = DAG.getStackArgumentTokenFactor(Chain);
3178 // Adjust the stack pointer for the new arguments...
3179 // These operations are automatically eliminated by the prolog/epilog pass
3180 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3181 SDValue CallSeqStart = Chain;
3183 // Load the return address and frame pointer so it can be move somewhere else
3186 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3189 // Set up a copy of the stack pointer for use loading and storing any
3190 // arguments that may not fit in the registers available for argument
3194 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3196 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3198 // Figure out which arguments are going to go in registers, and which in
3199 // memory. Also, if this is a vararg function, floating point operations
3200 // must be stored to our stack, and loaded into integer regs as well, if
3201 // any integer regs are available for argument passing.
3202 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3203 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3205 static const uint16_t GPR_32[] = { // 32-bit registers.
3206 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3207 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3209 static const uint16_t GPR_64[] = { // 64-bit registers.
3210 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3211 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3213 static const uint16_t *FPR = GetFPR();
3215 static const uint16_t VR[] = {
3216 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3217 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3219 const unsigned NumGPRs = array_lengthof(GPR_32);
3220 const unsigned NumFPRs = 13;
3221 const unsigned NumVRs = array_lengthof(VR);
3223 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3225 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3226 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3228 SmallVector<SDValue, 8> MemOpChains;
3229 for (unsigned i = 0; i != NumOps; ++i) {
3230 SDValue Arg = OutVals[i];
3231 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3233 // PtrOff will be used to store the current argument to the stack if a
3234 // register cannot be found for it.
3237 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3239 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3241 // On PPC64, promote integers to 64-bit values.
3242 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3243 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3244 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3245 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3248 // FIXME memcpy is used way more than necessary. Correctness first.
3249 if (Flags.isByVal()) {
3250 unsigned Size = Flags.getByValSize();
3251 if (Size==1 || Size==2) {
3252 // Very small objects are passed right-justified.
3253 // Everything else is passed left-justified.
3254 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3255 if (GPR_idx != NumGPRs) {
3256 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3257 MachinePointerInfo(), VT,
3259 MemOpChains.push_back(Load.getValue(1));
3260 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3262 ArgOffset += PtrByteSize;
3264 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3265 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3266 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3267 CallSeqStart.getNode()->getOperand(0),
3269 // This must go outside the CALLSEQ_START..END.
3270 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3271 CallSeqStart.getNode()->getOperand(1));
3272 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3273 NewCallSeqStart.getNode());
3274 Chain = CallSeqStart = NewCallSeqStart;
3275 ArgOffset += PtrByteSize;
3279 // Copy entire object into memory. There are cases where gcc-generated
3280 // code assumes it is there, even if it could be put entirely into
3281 // registers. (This is not what the doc says.)
3282 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3283 CallSeqStart.getNode()->getOperand(0),
3285 // This must go outside the CALLSEQ_START..END.
3286 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3287 CallSeqStart.getNode()->getOperand(1));
3288 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3289 Chain = CallSeqStart = NewCallSeqStart;
3290 // And copy the pieces of it that fit into registers.
3291 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3292 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3293 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3294 if (GPR_idx != NumGPRs) {
3295 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3296 MachinePointerInfo(),
3297 false, false, false, 0);
3298 MemOpChains.push_back(Load.getValue(1));
3299 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3300 ArgOffset += PtrByteSize;
3302 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3309 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3310 default: llvm_unreachable("Unexpected ValueType for argument!");
3313 if (GPR_idx != NumGPRs) {
3314 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3316 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3317 isPPC64, isTailCall, false, MemOpChains,
3318 TailCallArguments, dl);
3320 ArgOffset += PtrByteSize;
3324 if (FPR_idx != NumFPRs) {
3325 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3328 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3329 MachinePointerInfo(), false, false, 0);
3330 MemOpChains.push_back(Store);
3332 // Float varargs are always shadowed in available integer registers
3333 if (GPR_idx != NumGPRs) {
3334 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3335 MachinePointerInfo(), false, false,
3337 MemOpChains.push_back(Load.getValue(1));
3338 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3340 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3341 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3342 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3343 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3344 MachinePointerInfo(),
3345 false, false, false, 0);
3346 MemOpChains.push_back(Load.getValue(1));
3347 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3350 // If we have any FPRs remaining, we may also have GPRs remaining.
3351 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3353 if (GPR_idx != NumGPRs)
3355 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3356 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3360 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3361 isPPC64, isTailCall, false, MemOpChains,
3362 TailCallArguments, dl);
3367 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3374 // These go aligned on the stack, or in the corresponding R registers
3375 // when within range. The Darwin PPC ABI doc claims they also go in
3376 // V registers; in fact gcc does this only for arguments that are
3377 // prototyped, not for those that match the ... We do it for all
3378 // arguments, seems to work.
3379 while (ArgOffset % 16 !=0) {
3380 ArgOffset += PtrByteSize;
3381 if (GPR_idx != NumGPRs)
3384 // We could elide this store in the case where the object fits
3385 // entirely in R registers. Maybe later.
3386 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3387 DAG.getConstant(ArgOffset, PtrVT));
3388 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3389 MachinePointerInfo(), false, false, 0);
3390 MemOpChains.push_back(Store);
3391 if (VR_idx != NumVRs) {
3392 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3393 MachinePointerInfo(),
3394 false, false, false, 0);
3395 MemOpChains.push_back(Load.getValue(1));
3396 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3399 for (unsigned i=0; i<16; i+=PtrByteSize) {
3400 if (GPR_idx == NumGPRs)
3402 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3403 DAG.getConstant(i, PtrVT));
3404 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3405 false, false, false, 0);
3406 MemOpChains.push_back(Load.getValue(1));
3407 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3412 // Non-varargs Altivec params generally go in registers, but have
3413 // stack space allocated at the end.
3414 if (VR_idx != NumVRs) {
3415 // Doesn't have GPR space allocated.
3416 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3417 } else if (nAltivecParamsAtEnd==0) {
3418 // We are emitting Altivec params in order.
3419 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3420 isPPC64, isTailCall, true, MemOpChains,
3421 TailCallArguments, dl);
3427 // If all Altivec parameters fit in registers, as they usually do,
3428 // they get stack space following the non-Altivec parameters. We
3429 // don't track this here because nobody below needs it.
3430 // If there are more Altivec parameters than fit in registers emit
3432 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3434 // Offset is aligned; skip 1st 12 params which go in V registers.
3435 ArgOffset = ((ArgOffset+15)/16)*16;
3437 for (unsigned i = 0; i != NumOps; ++i) {
3438 SDValue Arg = OutVals[i];
3439 EVT ArgType = Outs[i].VT;
3440 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3441 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3444 // We are emitting Altivec params in order.
3445 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3446 isPPC64, isTailCall, true, MemOpChains,
3447 TailCallArguments, dl);
3454 if (!MemOpChains.empty())
3455 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3456 &MemOpChains[0], MemOpChains.size());
3458 // Check if this is an indirect call (MTCTR/BCTRL).
3459 // See PrepareCall() for more information about calls through function
3460 // pointers in the 64-bit SVR4 ABI.
3461 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3462 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3463 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3464 !isBLACompatibleAddress(Callee, DAG)) {
3465 // Load r2 into a virtual register and store it to the TOC save area.
3466 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3467 // TOC save area offset.
3468 SDValue PtrOff = DAG.getIntPtrConstant(40);
3469 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3470 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3474 // On Darwin, R12 must contain the address of an indirect callee. This does
3475 // not mean the MTCTR instruction must use R12; it's easier to model this as
3476 // an extra parameter, so do that.
3478 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3479 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3480 !isBLACompatibleAddress(Callee, DAG))
3481 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3482 PPC::R12), Callee));
3484 // Build a sequence of copy-to-reg nodes chained together with token chain
3485 // and flag operands which copy the outgoing args into the appropriate regs.
3487 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3488 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3489 RegsToPass[i].second, InFlag);
3490 InFlag = Chain.getValue(1);
3494 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3495 FPOp, true, TailCallArguments);
3497 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3498 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3503 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3504 MachineFunction &MF, bool isVarArg,
3505 const SmallVectorImpl<ISD::OutputArg> &Outs,
3506 LLVMContext &Context) const {
3507 SmallVector<CCValAssign, 16> RVLocs;
3508 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3510 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3514 PPCTargetLowering::LowerReturn(SDValue Chain,
3515 CallingConv::ID CallConv, bool isVarArg,
3516 const SmallVectorImpl<ISD::OutputArg> &Outs,
3517 const SmallVectorImpl<SDValue> &OutVals,
3518 DebugLoc dl, SelectionDAG &DAG) const {
3520 SmallVector<CCValAssign, 16> RVLocs;
3521 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3522 getTargetMachine(), RVLocs, *DAG.getContext());
3523 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3525 // If this is the first return lowered for this function, add the regs to the
3526 // liveout set for the function.
3527 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3528 for (unsigned i = 0; i != RVLocs.size(); ++i)
3529 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3534 // Copy the result values into the output registers.
3535 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3536 CCValAssign &VA = RVLocs[i];
3537 assert(VA.isRegLoc() && "Can only return in registers!");
3538 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3540 Flag = Chain.getValue(1);
3544 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3546 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3549 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3550 const PPCSubtarget &Subtarget) const {
3551 // When we pop the dynamic allocation we need to restore the SP link.
3552 DebugLoc dl = Op.getDebugLoc();
3554 // Get the corect type for pointers.
3555 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3557 // Construct the stack pointer operand.
3558 bool isPPC64 = Subtarget.isPPC64();
3559 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3560 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3562 // Get the operands for the STACKRESTORE.
3563 SDValue Chain = Op.getOperand(0);
3564 SDValue SaveSP = Op.getOperand(1);
3566 // Load the old link SP.
3567 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3568 MachinePointerInfo(),
3569 false, false, false, 0);
3571 // Restore the stack pointer.
3572 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3574 // Store the old link SP.
3575 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3582 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3583 MachineFunction &MF = DAG.getMachineFunction();
3584 bool isPPC64 = PPCSubTarget.isPPC64();
3585 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3586 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3588 // Get current frame pointer save index. The users of this index will be
3589 // primarily DYNALLOC instructions.
3590 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3591 int RASI = FI->getReturnAddrSaveIndex();
3593 // If the frame pointer save index hasn't been defined yet.
3595 // Find out what the fix offset of the frame pointer save area.
3596 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3597 // Allocate the frame index for frame pointer save area.
3598 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3600 FI->setReturnAddrSaveIndex(RASI);
3602 return DAG.getFrameIndex(RASI, PtrVT);
3606 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3607 MachineFunction &MF = DAG.getMachineFunction();
3608 bool isPPC64 = PPCSubTarget.isPPC64();
3609 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3610 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3612 // Get current frame pointer save index. The users of this index will be
3613 // primarily DYNALLOC instructions.
3614 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3615 int FPSI = FI->getFramePointerSaveIndex();
3617 // If the frame pointer save index hasn't been defined yet.
3619 // Find out what the fix offset of the frame pointer save area.
3620 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3623 // Allocate the frame index for frame pointer save area.
3624 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3626 FI->setFramePointerSaveIndex(FPSI);
3628 return DAG.getFrameIndex(FPSI, PtrVT);
3631 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3633 const PPCSubtarget &Subtarget) const {
3635 SDValue Chain = Op.getOperand(0);
3636 SDValue Size = Op.getOperand(1);
3637 DebugLoc dl = Op.getDebugLoc();
3639 // Get the corect type for pointers.
3640 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3642 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3643 DAG.getConstant(0, PtrVT), Size);
3644 // Construct a node for the frame pointer save index.
3645 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3646 // Build a DYNALLOC node.
3647 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3648 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3649 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3652 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3654 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3655 // Not FP? Not a fsel.
3656 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3657 !Op.getOperand(2).getValueType().isFloatingPoint())
3660 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3662 // Cannot handle SETEQ/SETNE.
3663 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3665 EVT ResVT = Op.getValueType();
3666 EVT CmpVT = Op.getOperand(0).getValueType();
3667 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3668 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3669 DebugLoc dl = Op.getDebugLoc();
3671 // If the RHS of the comparison is a 0.0, we don't need to do the
3672 // subtraction at all.
3673 if (isFloatingPointZero(RHS))
3675 default: break; // SETUO etc aren't handled by fsel.
3678 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3681 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3682 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3683 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3686 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3689 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3690 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3691 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3692 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3697 default: break; // SETUO etc aren't handled by fsel.
3700 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3701 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3702 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3703 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3706 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3707 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3708 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3709 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3712 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3713 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3714 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3715 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3718 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3719 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3720 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3721 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3726 // FIXME: Split this code up when LegalizeDAGTypes lands.
3727 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3728 DebugLoc dl) const {
3729 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3730 SDValue Src = Op.getOperand(0);
3731 if (Src.getValueType() == MVT::f32)
3732 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3735 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3736 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3738 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3743 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3747 // Convert the FP value to an int value through memory.
3748 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3750 // Emit a store to the stack slot.
3751 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3752 MachinePointerInfo(), false, false, 0);
3754 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3756 if (Op.getValueType() == MVT::i32)
3757 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3758 DAG.getConstant(4, FIPtr.getValueType()));
3759 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3760 false, false, false, 0);
3763 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3764 SelectionDAG &DAG) const {
3765 DebugLoc dl = Op.getDebugLoc();
3766 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3767 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3770 if (Op.getOperand(0).getValueType() == MVT::i64) {
3771 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3772 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3773 if (Op.getValueType() == MVT::f32)
3774 FP = DAG.getNode(ISD::FP_ROUND, dl,
3775 MVT::f32, FP, DAG.getIntPtrConstant(0));
3779 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3780 "Unhandled SINT_TO_FP type in custom expander!");
3781 // Since we only generate this in 64-bit mode, we can take advantage of
3782 // 64-bit registers. In particular, sign extend the input value into the
3783 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3784 // then lfd it and fcfid it.
3785 MachineFunction &MF = DAG.getMachineFunction();
3786 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3787 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3788 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3789 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3791 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3794 // STD the extended value into the stack slot.
3795 MachineMemOperand *MMO =
3796 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3797 MachineMemOperand::MOStore, 8, 8);
3798 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3800 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3801 Ops, 4, MVT::i64, MMO);
3802 // Load the value as a double.
3803 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3804 false, false, false, 0);
3806 // FCFID it and return it.
3807 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3808 if (Op.getValueType() == MVT::f32)
3809 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3813 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3814 SelectionDAG &DAG) const {
3815 DebugLoc dl = Op.getDebugLoc();
3817 The rounding mode is in bits 30:31 of FPSR, and has the following
3824 FLT_ROUNDS, on the other hand, expects the following:
3831 To perform the conversion, we do:
3832 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3835 MachineFunction &MF = DAG.getMachineFunction();
3836 EVT VT = Op.getValueType();
3837 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3838 std::vector<EVT> NodeTys;
3839 SDValue MFFSreg, InFlag;
3841 // Save FP Control Word to register
3842 NodeTys.push_back(MVT::f64); // return register
3843 NodeTys.push_back(MVT::Glue); // unused in this context
3844 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3846 // Save FP register to stack slot
3847 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3848 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3849 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3850 StackSlot, MachinePointerInfo(), false, false,0);
3852 // Load FP Control Word from low 32 bits of stack slot.
3853 SDValue Four = DAG.getConstant(4, PtrVT);
3854 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3855 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3856 false, false, false, 0);
3858 // Transform as necessary
3860 DAG.getNode(ISD::AND, dl, MVT::i32,
3861 CWD, DAG.getConstant(3, MVT::i32));
3863 DAG.getNode(ISD::SRL, dl, MVT::i32,
3864 DAG.getNode(ISD::AND, dl, MVT::i32,
3865 DAG.getNode(ISD::XOR, dl, MVT::i32,
3866 CWD, DAG.getConstant(3, MVT::i32)),
3867 DAG.getConstant(3, MVT::i32)),
3868 DAG.getConstant(1, MVT::i32));
3871 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3873 return DAG.getNode((VT.getSizeInBits() < 16 ?
3874 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3877 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3878 EVT VT = Op.getValueType();
3879 unsigned BitWidth = VT.getSizeInBits();
3880 DebugLoc dl = Op.getDebugLoc();
3881 assert(Op.getNumOperands() == 3 &&
3882 VT == Op.getOperand(1).getValueType() &&
3885 // Expand into a bunch of logical ops. Note that these ops
3886 // depend on the PPC behavior for oversized shift amounts.
3887 SDValue Lo = Op.getOperand(0);
3888 SDValue Hi = Op.getOperand(1);
3889 SDValue Amt = Op.getOperand(2);
3890 EVT AmtVT = Amt.getValueType();
3892 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3893 DAG.getConstant(BitWidth, AmtVT), Amt);
3894 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3895 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3896 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3897 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3898 DAG.getConstant(-BitWidth, AmtVT));
3899 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3900 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3901 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3902 SDValue OutOps[] = { OutLo, OutHi };
3903 return DAG.getMergeValues(OutOps, 2, dl);
3906 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3907 EVT VT = Op.getValueType();
3908 DebugLoc dl = Op.getDebugLoc();
3909 unsigned BitWidth = VT.getSizeInBits();
3910 assert(Op.getNumOperands() == 3 &&
3911 VT == Op.getOperand(1).getValueType() &&
3914 // Expand into a bunch of logical ops. Note that these ops
3915 // depend on the PPC behavior for oversized shift amounts.
3916 SDValue Lo = Op.getOperand(0);
3917 SDValue Hi = Op.getOperand(1);
3918 SDValue Amt = Op.getOperand(2);
3919 EVT AmtVT = Amt.getValueType();
3921 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3922 DAG.getConstant(BitWidth, AmtVT), Amt);
3923 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3924 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3925 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3926 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3927 DAG.getConstant(-BitWidth, AmtVT));
3928 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3929 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3930 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3931 SDValue OutOps[] = { OutLo, OutHi };
3932 return DAG.getMergeValues(OutOps, 2, dl);
3935 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3936 DebugLoc dl = Op.getDebugLoc();
3937 EVT VT = Op.getValueType();
3938 unsigned BitWidth = VT.getSizeInBits();
3939 assert(Op.getNumOperands() == 3 &&
3940 VT == Op.getOperand(1).getValueType() &&
3943 // Expand into a bunch of logical ops, followed by a select_cc.
3944 SDValue Lo = Op.getOperand(0);
3945 SDValue Hi = Op.getOperand(1);
3946 SDValue Amt = Op.getOperand(2);
3947 EVT AmtVT = Amt.getValueType();
3949 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3950 DAG.getConstant(BitWidth, AmtVT), Amt);
3951 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3952 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3953 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3954 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3955 DAG.getConstant(-BitWidth, AmtVT));
3956 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3957 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3958 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3959 Tmp4, Tmp6, ISD::SETLE);
3960 SDValue OutOps[] = { OutLo, OutHi };
3961 return DAG.getMergeValues(OutOps, 2, dl);
3964 //===----------------------------------------------------------------------===//
3965 // Vector related lowering.
3968 /// BuildSplatI - Build a canonical splati of Val with an element size of
3969 /// SplatSize. Cast the result to VT.
3970 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3971 SelectionDAG &DAG, DebugLoc dl) {
3972 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3974 static const EVT VTys[] = { // canonical VT to use for each size.
3975 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3978 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3980 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3984 EVT CanonicalVT = VTys[SplatSize-1];
3986 // Build a canonical splat for this value.
3987 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3988 SmallVector<SDValue, 8> Ops;
3989 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3990 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3991 &Ops[0], Ops.size());
3992 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3995 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3996 /// specified intrinsic ID.
3997 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3998 SelectionDAG &DAG, DebugLoc dl,
3999 EVT DestVT = MVT::Other) {
4000 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4001 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4002 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4005 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4006 /// specified intrinsic ID.
4007 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4008 SDValue Op2, SelectionDAG &DAG,
4009 DebugLoc dl, EVT DestVT = MVT::Other) {
4010 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4011 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4012 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4016 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4017 /// amount. The result has the specified value type.
4018 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4019 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4020 // Force LHS/RHS to be the right type.
4021 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4022 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4025 for (unsigned i = 0; i != 16; ++i)
4027 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4028 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4031 // If this is a case we can't handle, return null and let the default
4032 // expansion code take care of it. If we CAN select this case, and if it
4033 // selects to a single instruction, return Op. Otherwise, if we can codegen
4034 // this case more efficiently than a constant pool load, lower it to the
4035 // sequence of ops that should be used.
4036 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4037 SelectionDAG &DAG) const {
4038 DebugLoc dl = Op.getDebugLoc();
4039 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4040 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4042 // Check if this is a splat of a constant value.
4043 APInt APSplatBits, APSplatUndef;
4044 unsigned SplatBitSize;
4046 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4047 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4050 unsigned SplatBits = APSplatBits.getZExtValue();
4051 unsigned SplatUndef = APSplatUndef.getZExtValue();
4052 unsigned SplatSize = SplatBitSize / 8;
4054 // First, handle single instruction cases.
4057 if (SplatBits == 0) {
4058 // Canonicalize all zero vectors to be v4i32.
4059 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4060 SDValue Z = DAG.getConstant(0, MVT::i32);
4061 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4062 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4067 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4068 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4070 if (SextVal >= -16 && SextVal <= 15)
4071 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4074 // Two instruction sequences.
4076 // If this value is in the range [-32,30] and is even, use:
4077 // tmp = VSPLTI[bhw], result = add tmp, tmp
4078 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4079 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4080 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4081 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4084 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4085 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4087 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4088 // Make -1 and vspltisw -1:
4089 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4091 // Make the VSLW intrinsic, computing 0x8000_0000.
4092 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4095 // xor by OnesV to invert it.
4096 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4097 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4100 // Check to see if this is a wide variety of vsplti*, binop self cases.
4101 static const signed char SplatCsts[] = {
4102 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4103 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4106 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4107 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4108 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4109 int i = SplatCsts[idx];
4111 // Figure out what shift amount will be used by altivec if shifted by i in
4113 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4115 // vsplti + shl self.
4116 if (SextVal == (i << (int)TypeShiftAmt)) {
4117 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4118 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4119 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4120 Intrinsic::ppc_altivec_vslw
4122 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4123 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4126 // vsplti + srl self.
4127 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4128 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4129 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4130 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4131 Intrinsic::ppc_altivec_vsrw
4133 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4134 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4137 // vsplti + sra self.
4138 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4139 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4140 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4141 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4142 Intrinsic::ppc_altivec_vsraw
4144 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4145 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4148 // vsplti + rol self.
4149 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4150 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4151 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4152 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4153 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4154 Intrinsic::ppc_altivec_vrlw
4156 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4157 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4160 // t = vsplti c, result = vsldoi t, t, 1
4161 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4162 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4163 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4165 // t = vsplti c, result = vsldoi t, t, 2
4166 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4167 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4168 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4170 // t = vsplti c, result = vsldoi t, t, 3
4171 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4172 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4173 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4177 // Three instruction sequences.
4179 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4180 if (SextVal >= 0 && SextVal <= 31) {
4181 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4182 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4183 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4184 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4186 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4187 if (SextVal >= -31 && SextVal <= 0) {
4188 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4189 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4190 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4191 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4197 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4198 /// the specified operations to build the shuffle.
4199 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4200 SDValue RHS, SelectionDAG &DAG,
4202 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4203 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4204 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4207 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4219 if (OpNum == OP_COPY) {
4220 if (LHSID == (1*9+2)*9+3) return LHS;
4221 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4225 SDValue OpLHS, OpRHS;
4226 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4227 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4231 default: llvm_unreachable("Unknown i32 permute!");
4233 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4234 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4235 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4236 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4239 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4240 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4241 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4242 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4245 for (unsigned i = 0; i != 16; ++i)
4246 ShufIdxs[i] = (i&3)+0;
4249 for (unsigned i = 0; i != 16; ++i)
4250 ShufIdxs[i] = (i&3)+4;
4253 for (unsigned i = 0; i != 16; ++i)
4254 ShufIdxs[i] = (i&3)+8;
4257 for (unsigned i = 0; i != 16; ++i)
4258 ShufIdxs[i] = (i&3)+12;
4261 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4263 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4265 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4267 EVT VT = OpLHS.getValueType();
4268 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4269 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4270 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4271 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4274 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4275 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4276 /// return the code it can be lowered into. Worst case, it can always be
4277 /// lowered into a vperm.
4278 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4279 SelectionDAG &DAG) const {
4280 DebugLoc dl = Op.getDebugLoc();
4281 SDValue V1 = Op.getOperand(0);
4282 SDValue V2 = Op.getOperand(1);
4283 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4284 EVT VT = Op.getValueType();
4286 // Cases that are handled by instructions that take permute immediates
4287 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4288 // selected by the instruction selector.
4289 if (V2.getOpcode() == ISD::UNDEF) {
4290 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4291 PPC::isSplatShuffleMask(SVOp, 2) ||
4292 PPC::isSplatShuffleMask(SVOp, 4) ||
4293 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4294 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4295 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4296 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4297 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4298 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4299 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4300 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4301 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4306 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4307 // and produce a fixed permutation. If any of these match, do not lower to
4309 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4310 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4311 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4312 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4313 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4314 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4315 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4316 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4317 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4320 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4321 // perfect shuffle table to emit an optimal matching sequence.
4322 ArrayRef<int> PermMask = SVOp->getMask();
4324 unsigned PFIndexes[4];
4325 bool isFourElementShuffle = true;
4326 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4327 unsigned EltNo = 8; // Start out undef.
4328 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4329 if (PermMask[i*4+j] < 0)
4330 continue; // Undef, ignore it.
4332 unsigned ByteSource = PermMask[i*4+j];
4333 if ((ByteSource & 3) != j) {
4334 isFourElementShuffle = false;
4339 EltNo = ByteSource/4;
4340 } else if (EltNo != ByteSource/4) {
4341 isFourElementShuffle = false;
4345 PFIndexes[i] = EltNo;
4348 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4349 // perfect shuffle vector to determine if it is cost effective to do this as
4350 // discrete instructions, or whether we should use a vperm.
4351 if (isFourElementShuffle) {
4352 // Compute the index in the perfect shuffle table.
4353 unsigned PFTableIndex =
4354 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4356 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4357 unsigned Cost = (PFEntry >> 30);
4359 // Determining when to avoid vperm is tricky. Many things affect the cost
4360 // of vperm, particularly how many times the perm mask needs to be computed.
4361 // For example, if the perm mask can be hoisted out of a loop or is already
4362 // used (perhaps because there are multiple permutes with the same shuffle
4363 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4364 // the loop requires an extra register.
4366 // As a compromise, we only emit discrete instructions if the shuffle can be
4367 // generated in 3 or fewer operations. When we have loop information
4368 // available, if this block is within a loop, we should avoid using vperm
4369 // for 3-operation perms and use a constant pool load instead.
4371 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4374 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4375 // vector that will get spilled to the constant pool.
4376 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4378 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4379 // that it is in input element units, not in bytes. Convert now.
4380 EVT EltVT = V1.getValueType().getVectorElementType();
4381 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4383 SmallVector<SDValue, 16> ResultMask;
4384 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4385 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4387 for (unsigned j = 0; j != BytesPerElement; ++j)
4388 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4392 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4393 &ResultMask[0], ResultMask.size());
4394 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4397 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4398 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4399 /// information about the intrinsic.
4400 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4402 unsigned IntrinsicID =
4403 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4406 switch (IntrinsicID) {
4407 default: return false;
4408 // Comparison predicates.
4409 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4410 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4411 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4412 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4413 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4414 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4415 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4416 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4417 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4418 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4419 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4420 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4421 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4423 // Normal Comparisons.
4424 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4425 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4426 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4427 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4428 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4429 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4430 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4431 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4432 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4433 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4434 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4435 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4436 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4441 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4442 /// lower, do it, otherwise return null.
4443 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4444 SelectionDAG &DAG) const {
4445 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4446 // opcode number of the comparison.
4447 DebugLoc dl = Op.getDebugLoc();
4450 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4451 return SDValue(); // Don't custom lower most intrinsics.
4453 // If this is a non-dot comparison, make the VCMP node and we are done.
4455 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4456 Op.getOperand(1), Op.getOperand(2),
4457 DAG.getConstant(CompareOpc, MVT::i32));
4458 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4461 // Create the PPCISD altivec 'dot' comparison node.
4463 Op.getOperand(2), // LHS
4464 Op.getOperand(3), // RHS
4465 DAG.getConstant(CompareOpc, MVT::i32)
4467 std::vector<EVT> VTs;
4468 VTs.push_back(Op.getOperand(2).getValueType());
4469 VTs.push_back(MVT::Glue);
4470 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4472 // Now that we have the comparison, emit a copy from the CR to a GPR.
4473 // This is flagged to the above dot comparison.
4474 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4475 DAG.getRegister(PPC::CR6, MVT::i32),
4476 CompNode.getValue(1));
4478 // Unpack the result based on how the target uses it.
4479 unsigned BitNo; // Bit # of CR6.
4480 bool InvertBit; // Invert result?
4481 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4482 default: // Can't happen, don't crash on invalid number though.
4483 case 0: // Return the value of the EQ bit of CR6.
4484 BitNo = 0; InvertBit = false;
4486 case 1: // Return the inverted value of the EQ bit of CR6.
4487 BitNo = 0; InvertBit = true;
4489 case 2: // Return the value of the LT bit of CR6.
4490 BitNo = 2; InvertBit = false;
4492 case 3: // Return the inverted value of the LT bit of CR6.
4493 BitNo = 2; InvertBit = true;
4497 // Shift the bit into the low position.
4498 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4499 DAG.getConstant(8-(3-BitNo), MVT::i32));
4501 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4502 DAG.getConstant(1, MVT::i32));
4504 // If we are supposed to, toggle the bit.
4506 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4507 DAG.getConstant(1, MVT::i32));
4511 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4512 SelectionDAG &DAG) const {
4513 DebugLoc dl = Op.getDebugLoc();
4514 // Create a stack slot that is 16-byte aligned.
4515 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4516 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4517 EVT PtrVT = getPointerTy();
4518 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4520 // Store the input value into Value#0 of the stack slot.
4521 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4522 Op.getOperand(0), FIdx, MachinePointerInfo(),
4525 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4526 false, false, false, 0);
4529 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4530 DebugLoc dl = Op.getDebugLoc();
4531 if (Op.getValueType() == MVT::v4i32) {
4532 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4534 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4535 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4537 SDValue RHSSwap = // = vrlw RHS, 16
4538 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4540 // Shrinkify inputs to v8i16.
4541 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4542 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4543 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4545 // Low parts multiplied together, generating 32-bit results (we ignore the
4547 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4548 LHS, RHS, DAG, dl, MVT::v4i32);
4550 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4551 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4552 // Shift the high parts up 16 bits.
4553 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4555 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4556 } else if (Op.getValueType() == MVT::v8i16) {
4557 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4559 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4561 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4562 LHS, RHS, Zero, DAG, dl);
4563 } else if (Op.getValueType() == MVT::v16i8) {
4564 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4566 // Multiply the even 8-bit parts, producing 16-bit sums.
4567 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4568 LHS, RHS, DAG, dl, MVT::v8i16);
4569 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4571 // Multiply the odd 8-bit parts, producing 16-bit sums.
4572 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4573 LHS, RHS, DAG, dl, MVT::v8i16);
4574 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4576 // Merge the results together.
4578 for (unsigned i = 0; i != 8; ++i) {
4580 Ops[i*2+1] = 2*i+1+16;
4582 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4584 llvm_unreachable("Unknown mul to lower!");
4588 /// LowerOperation - Provide custom lowering hooks for some operations.
4590 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4591 switch (Op.getOpcode()) {
4592 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4593 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4594 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4595 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4596 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4597 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4598 case ISD::SETCC: return LowerSETCC(Op, DAG);
4599 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4600 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4602 return LowerVASTART(Op, DAG, PPCSubTarget);
4605 return LowerVAARG(Op, DAG, PPCSubTarget);
4607 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4608 case ISD::DYNAMIC_STACKALLOC:
4609 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4611 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4612 case ISD::FP_TO_UINT:
4613 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4615 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4616 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4618 // Lower 64-bit shifts.
4619 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4620 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4621 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4623 // Vector-related lowering.
4624 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4625 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4626 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4627 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4628 case ISD::MUL: return LowerMUL(Op, DAG);
4630 // Frame & Return address.
4631 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4632 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4636 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4637 SmallVectorImpl<SDValue>&Results,
4638 SelectionDAG &DAG) const {
4639 const TargetMachine &TM = getTargetMachine();
4640 DebugLoc dl = N->getDebugLoc();
4641 switch (N->getOpcode()) {
4643 llvm_unreachable("Do not know how to custom type legalize this operation!");
4645 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4646 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4649 EVT VT = N->getValueType(0);
4651 if (VT == MVT::i64) {
4652 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4654 Results.push_back(NewNode);
4655 Results.push_back(NewNode.getValue(1));
4659 case ISD::FP_ROUND_INREG: {
4660 assert(N->getValueType(0) == MVT::ppcf128);
4661 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4662 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4663 MVT::f64, N->getOperand(0),
4664 DAG.getIntPtrConstant(0));
4665 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4666 MVT::f64, N->getOperand(0),
4667 DAG.getIntPtrConstant(1));
4669 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4670 // of the long double, and puts FPSCR back the way it was. We do not
4671 // actually model FPSCR.
4672 std::vector<EVT> NodeTys;
4673 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4675 NodeTys.push_back(MVT::f64); // Return register
4676 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4677 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4678 MFFSreg = Result.getValue(0);
4679 InFlag = Result.getValue(1);
4682 NodeTys.push_back(MVT::Glue); // Returns a flag
4683 Ops[0] = DAG.getConstant(31, MVT::i32);
4685 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4686 InFlag = Result.getValue(0);
4689 NodeTys.push_back(MVT::Glue); // Returns a flag
4690 Ops[0] = DAG.getConstant(30, MVT::i32);
4692 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4693 InFlag = Result.getValue(0);
4696 NodeTys.push_back(MVT::f64); // result of add
4697 NodeTys.push_back(MVT::Glue); // Returns a flag
4701 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4702 FPreg = Result.getValue(0);
4703 InFlag = Result.getValue(1);
4706 NodeTys.push_back(MVT::f64);
4707 Ops[0] = DAG.getConstant(1, MVT::i32);
4711 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4712 FPreg = Result.getValue(0);
4714 // We know the low half is about to be thrown away, so just use something
4716 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4720 case ISD::FP_TO_SINT:
4721 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4727 //===----------------------------------------------------------------------===//
4728 // Other Lowering Code
4729 //===----------------------------------------------------------------------===//
4732 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4733 bool is64bit, unsigned BinOpcode) const {
4734 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4735 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4737 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4738 MachineFunction *F = BB->getParent();
4739 MachineFunction::iterator It = BB;
4742 unsigned dest = MI->getOperand(0).getReg();
4743 unsigned ptrA = MI->getOperand(1).getReg();
4744 unsigned ptrB = MI->getOperand(2).getReg();
4745 unsigned incr = MI->getOperand(3).getReg();
4746 DebugLoc dl = MI->getDebugLoc();
4748 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4749 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4750 F->insert(It, loopMBB);
4751 F->insert(It, exitMBB);
4752 exitMBB->splice(exitMBB->begin(), BB,
4753 llvm::next(MachineBasicBlock::iterator(MI)),
4755 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4757 MachineRegisterInfo &RegInfo = F->getRegInfo();
4758 unsigned TmpReg = (!BinOpcode) ? incr :
4759 RegInfo.createVirtualRegister(
4760 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4761 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4765 // fallthrough --> loopMBB
4766 BB->addSuccessor(loopMBB);
4769 // l[wd]arx dest, ptr
4770 // add r0, dest, incr
4771 // st[wd]cx. r0, ptr
4773 // fallthrough --> exitMBB
4775 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4776 .addReg(ptrA).addReg(ptrB);
4778 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4779 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4780 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4781 BuildMI(BB, dl, TII->get(PPC::BCC))
4782 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4783 BB->addSuccessor(loopMBB);
4784 BB->addSuccessor(exitMBB);
4793 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4794 MachineBasicBlock *BB,
4795 bool is8bit, // operation
4796 unsigned BinOpcode) const {
4797 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4798 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4799 // In 64 bit mode we have to use 64 bits for addresses, even though the
4800 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4801 // registers without caring whether they're 32 or 64, but here we're
4802 // doing actual arithmetic on the addresses.
4803 bool is64bit = PPCSubTarget.isPPC64();
4804 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4806 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4807 MachineFunction *F = BB->getParent();
4808 MachineFunction::iterator It = BB;
4811 unsigned dest = MI->getOperand(0).getReg();
4812 unsigned ptrA = MI->getOperand(1).getReg();
4813 unsigned ptrB = MI->getOperand(2).getReg();
4814 unsigned incr = MI->getOperand(3).getReg();
4815 DebugLoc dl = MI->getDebugLoc();
4817 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4818 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4819 F->insert(It, loopMBB);
4820 F->insert(It, exitMBB);
4821 exitMBB->splice(exitMBB->begin(), BB,
4822 llvm::next(MachineBasicBlock::iterator(MI)),
4824 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4826 MachineRegisterInfo &RegInfo = F->getRegInfo();
4827 const TargetRegisterClass *RC =
4828 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4829 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4830 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4831 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4832 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4833 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4834 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4835 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4836 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4837 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4838 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4839 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4840 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4842 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4846 // fallthrough --> loopMBB
4847 BB->addSuccessor(loopMBB);
4849 // The 4-byte load must be aligned, while a char or short may be
4850 // anywhere in the word. Hence all this nasty bookkeeping code.
4851 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4852 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4853 // xori shift, shift1, 24 [16]
4854 // rlwinm ptr, ptr1, 0, 0, 29
4855 // slw incr2, incr, shift
4856 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4857 // slw mask, mask2, shift
4859 // lwarx tmpDest, ptr
4860 // add tmp, tmpDest, incr2
4861 // andc tmp2, tmpDest, mask
4862 // and tmp3, tmp, mask
4863 // or tmp4, tmp3, tmp2
4866 // fallthrough --> exitMBB
4867 // srw dest, tmpDest, shift
4868 if (ptrA != ZeroReg) {
4869 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4870 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4871 .addReg(ptrA).addReg(ptrB);
4875 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4876 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4877 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4878 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4880 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4881 .addReg(Ptr1Reg).addImm(0).addImm(61);
4883 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4884 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4885 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4886 .addReg(incr).addReg(ShiftReg);
4888 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4890 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4891 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4893 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4894 .addReg(Mask2Reg).addReg(ShiftReg);
4897 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4898 .addReg(ZeroReg).addReg(PtrReg);
4900 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4901 .addReg(Incr2Reg).addReg(TmpDestReg);
4902 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4903 .addReg(TmpDestReg).addReg(MaskReg);
4904 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4905 .addReg(TmpReg).addReg(MaskReg);
4906 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4907 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4908 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4909 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4910 BuildMI(BB, dl, TII->get(PPC::BCC))
4911 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4912 BB->addSuccessor(loopMBB);
4913 BB->addSuccessor(exitMBB);
4918 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4924 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4925 MachineBasicBlock *BB) const {
4926 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4928 // To "insert" these instructions we actually have to insert their
4929 // control-flow patterns.
4930 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4931 MachineFunction::iterator It = BB;
4934 MachineFunction *F = BB->getParent();
4936 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4937 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4938 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4939 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4940 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4942 // The incoming instruction knows the destination vreg to set, the
4943 // condition code register to branch on, the true/false values to
4944 // select between, and a branch opcode to use.
4949 // cmpTY ccX, r1, r2
4951 // fallthrough --> copy0MBB
4952 MachineBasicBlock *thisMBB = BB;
4953 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4954 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4955 unsigned SelectPred = MI->getOperand(4).getImm();
4956 DebugLoc dl = MI->getDebugLoc();
4957 F->insert(It, copy0MBB);
4958 F->insert(It, sinkMBB);
4960 // Transfer the remainder of BB and its successor edges to sinkMBB.
4961 sinkMBB->splice(sinkMBB->begin(), BB,
4962 llvm::next(MachineBasicBlock::iterator(MI)),
4964 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4966 // Next, add the true and fallthrough blocks as its successors.
4967 BB->addSuccessor(copy0MBB);
4968 BB->addSuccessor(sinkMBB);
4970 BuildMI(BB, dl, TII->get(PPC::BCC))
4971 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4974 // %FalseValue = ...
4975 // # fallthrough to sinkMBB
4978 // Update machine-CFG edges
4979 BB->addSuccessor(sinkMBB);
4982 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4985 BuildMI(*BB, BB->begin(), dl,
4986 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4987 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4988 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4990 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4991 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4992 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4993 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4994 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4995 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4996 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4997 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4999 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5000 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5001 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5002 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5003 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5004 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5005 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5006 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5008 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5009 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5010 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5011 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5012 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5013 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5014 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5015 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5017 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5018 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5019 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5020 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5021 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5022 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5023 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5024 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5026 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5027 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5028 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5029 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5030 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5031 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5032 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5033 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5035 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5036 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5037 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5038 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5039 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5040 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5041 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5042 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5044 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5045 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5046 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5047 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5048 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5049 BB = EmitAtomicBinary(MI, BB, false, 0);
5050 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5051 BB = EmitAtomicBinary(MI, BB, true, 0);
5053 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5054 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5055 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5057 unsigned dest = MI->getOperand(0).getReg();
5058 unsigned ptrA = MI->getOperand(1).getReg();
5059 unsigned ptrB = MI->getOperand(2).getReg();
5060 unsigned oldval = MI->getOperand(3).getReg();
5061 unsigned newval = MI->getOperand(4).getReg();
5062 DebugLoc dl = MI->getDebugLoc();
5064 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5065 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5066 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5067 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5068 F->insert(It, loop1MBB);
5069 F->insert(It, loop2MBB);
5070 F->insert(It, midMBB);
5071 F->insert(It, exitMBB);
5072 exitMBB->splice(exitMBB->begin(), BB,
5073 llvm::next(MachineBasicBlock::iterator(MI)),
5075 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5079 // fallthrough --> loopMBB
5080 BB->addSuccessor(loop1MBB);
5083 // l[wd]arx dest, ptr
5084 // cmp[wd] dest, oldval
5087 // st[wd]cx. newval, ptr
5091 // st[wd]cx. dest, ptr
5094 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5095 .addReg(ptrA).addReg(ptrB);
5096 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5097 .addReg(oldval).addReg(dest);
5098 BuildMI(BB, dl, TII->get(PPC::BCC))
5099 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5100 BB->addSuccessor(loop2MBB);
5101 BB->addSuccessor(midMBB);
5104 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5105 .addReg(newval).addReg(ptrA).addReg(ptrB);
5106 BuildMI(BB, dl, TII->get(PPC::BCC))
5107 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5108 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5109 BB->addSuccessor(loop1MBB);
5110 BB->addSuccessor(exitMBB);
5113 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5114 .addReg(dest).addReg(ptrA).addReg(ptrB);
5115 BB->addSuccessor(exitMBB);
5120 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5121 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5122 // We must use 64-bit registers for addresses when targeting 64-bit,
5123 // since we're actually doing arithmetic on them. Other registers
5125 bool is64bit = PPCSubTarget.isPPC64();
5126 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5128 unsigned dest = MI->getOperand(0).getReg();
5129 unsigned ptrA = MI->getOperand(1).getReg();
5130 unsigned ptrB = MI->getOperand(2).getReg();
5131 unsigned oldval = MI->getOperand(3).getReg();
5132 unsigned newval = MI->getOperand(4).getReg();
5133 DebugLoc dl = MI->getDebugLoc();
5135 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5136 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5137 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5138 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5139 F->insert(It, loop1MBB);
5140 F->insert(It, loop2MBB);
5141 F->insert(It, midMBB);
5142 F->insert(It, exitMBB);
5143 exitMBB->splice(exitMBB->begin(), BB,
5144 llvm::next(MachineBasicBlock::iterator(MI)),
5146 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5148 MachineRegisterInfo &RegInfo = F->getRegInfo();
5149 const TargetRegisterClass *RC =
5150 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5151 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5152 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5153 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5154 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5155 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5156 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5157 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5158 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5159 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5160 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5161 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5162 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5163 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5164 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5166 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5167 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5170 // fallthrough --> loopMBB
5171 BB->addSuccessor(loop1MBB);
5173 // The 4-byte load must be aligned, while a char or short may be
5174 // anywhere in the word. Hence all this nasty bookkeeping code.
5175 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5176 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5177 // xori shift, shift1, 24 [16]
5178 // rlwinm ptr, ptr1, 0, 0, 29
5179 // slw newval2, newval, shift
5180 // slw oldval2, oldval,shift
5181 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5182 // slw mask, mask2, shift
5183 // and newval3, newval2, mask
5184 // and oldval3, oldval2, mask
5186 // lwarx tmpDest, ptr
5187 // and tmp, tmpDest, mask
5188 // cmpw tmp, oldval3
5191 // andc tmp2, tmpDest, mask
5192 // or tmp4, tmp2, newval3
5197 // stwcx. tmpDest, ptr
5199 // srw dest, tmpDest, shift
5200 if (ptrA != ZeroReg) {
5201 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5202 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5203 .addReg(ptrA).addReg(ptrB);
5207 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5208 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5209 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5210 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5212 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5213 .addReg(Ptr1Reg).addImm(0).addImm(61);
5215 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5216 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5217 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5218 .addReg(newval).addReg(ShiftReg);
5219 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5220 .addReg(oldval).addReg(ShiftReg);
5222 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5224 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5225 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5226 .addReg(Mask3Reg).addImm(65535);
5228 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5229 .addReg(Mask2Reg).addReg(ShiftReg);
5230 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5231 .addReg(NewVal2Reg).addReg(MaskReg);
5232 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5233 .addReg(OldVal2Reg).addReg(MaskReg);
5236 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5237 .addReg(ZeroReg).addReg(PtrReg);
5238 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5239 .addReg(TmpDestReg).addReg(MaskReg);
5240 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5241 .addReg(TmpReg).addReg(OldVal3Reg);
5242 BuildMI(BB, dl, TII->get(PPC::BCC))
5243 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5244 BB->addSuccessor(loop2MBB);
5245 BB->addSuccessor(midMBB);
5248 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5249 .addReg(TmpDestReg).addReg(MaskReg);
5250 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5251 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5252 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5253 .addReg(ZeroReg).addReg(PtrReg);
5254 BuildMI(BB, dl, TII->get(PPC::BCC))
5255 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5256 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5257 BB->addSuccessor(loop1MBB);
5258 BB->addSuccessor(exitMBB);
5261 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5262 .addReg(ZeroReg).addReg(PtrReg);
5263 BB->addSuccessor(exitMBB);
5268 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5271 llvm_unreachable("Unexpected instr type to insert");
5274 MI->eraseFromParent(); // The pseudo instruction is gone now.
5278 //===----------------------------------------------------------------------===//
5279 // Target Optimization Hooks
5280 //===----------------------------------------------------------------------===//
5282 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5283 DAGCombinerInfo &DCI) const {
5284 const TargetMachine &TM = getTargetMachine();
5285 SelectionDAG &DAG = DCI.DAG;
5286 DebugLoc dl = N->getDebugLoc();
5287 switch (N->getOpcode()) {
5290 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5291 if (C->isNullValue()) // 0 << V -> 0.
5292 return N->getOperand(0);
5296 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5297 if (C->isNullValue()) // 0 >>u V -> 0.
5298 return N->getOperand(0);
5302 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5303 if (C->isNullValue() || // 0 >>s V -> 0.
5304 C->isAllOnesValue()) // -1 >>s V -> -1.
5305 return N->getOperand(0);
5309 case ISD::SINT_TO_FP:
5310 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5311 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5312 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5313 // We allow the src/dst to be either f32/f64, but the intermediate
5314 // type must be i64.
5315 if (N->getOperand(0).getValueType() == MVT::i64 &&
5316 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5317 SDValue Val = N->getOperand(0).getOperand(0);
5318 if (Val.getValueType() == MVT::f32) {
5319 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5320 DCI.AddToWorklist(Val.getNode());
5323 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5324 DCI.AddToWorklist(Val.getNode());
5325 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5326 DCI.AddToWorklist(Val.getNode());
5327 if (N->getValueType(0) == MVT::f32) {
5328 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5329 DAG.getIntPtrConstant(0));
5330 DCI.AddToWorklist(Val.getNode());
5333 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5334 // If the intermediate type is i32, we can avoid the load/store here
5341 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5342 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5343 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5344 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5345 N->getOperand(1).getValueType() == MVT::i32 &&
5346 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5347 SDValue Val = N->getOperand(1).getOperand(0);
5348 if (Val.getValueType() == MVT::f32) {
5349 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5350 DCI.AddToWorklist(Val.getNode());
5352 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5353 DCI.AddToWorklist(Val.getNode());
5355 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5356 N->getOperand(2), N->getOperand(3));
5357 DCI.AddToWorklist(Val.getNode());
5361 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5362 if (cast<StoreSDNode>(N)->isUnindexed() &&
5363 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5364 N->getOperand(1).getNode()->hasOneUse() &&
5365 (N->getOperand(1).getValueType() == MVT::i32 ||
5366 N->getOperand(1).getValueType() == MVT::i16)) {
5367 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5368 // Do an any-extend to 32-bits if this is a half-word input.
5369 if (BSwapOp.getValueType() == MVT::i16)
5370 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5373 N->getOperand(0), BSwapOp, N->getOperand(2),
5374 DAG.getValueType(N->getOperand(1).getValueType())
5377 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5378 Ops, array_lengthof(Ops),
5379 cast<StoreSDNode>(N)->getMemoryVT(),
5380 cast<StoreSDNode>(N)->getMemOperand());
5384 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5385 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5386 N->getOperand(0).hasOneUse() &&
5387 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5388 SDValue Load = N->getOperand(0);
5389 LoadSDNode *LD = cast<LoadSDNode>(Load);
5390 // Create the byte-swapping load.
5392 LD->getChain(), // Chain
5393 LD->getBasePtr(), // Ptr
5394 DAG.getValueType(N->getValueType(0)) // VT
5397 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5398 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5399 LD->getMemoryVT(), LD->getMemOperand());
5401 // If this is an i16 load, insert the truncate.
5402 SDValue ResVal = BSLoad;
5403 if (N->getValueType(0) == MVT::i16)
5404 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5406 // First, combine the bswap away. This makes the value produced by the
5408 DCI.CombineTo(N, ResVal);
5410 // Next, combine the load away, we give it a bogus result value but a real
5411 // chain result. The result value is dead because the bswap is dead.
5412 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5414 // Return N so it doesn't get rechecked!
5415 return SDValue(N, 0);
5419 case PPCISD::VCMP: {
5420 // If a VCMPo node already exists with exactly the same operands as this
5421 // node, use its result instead of this node (VCMPo computes both a CR6 and
5422 // a normal output).
5424 if (!N->getOperand(0).hasOneUse() &&
5425 !N->getOperand(1).hasOneUse() &&
5426 !N->getOperand(2).hasOneUse()) {
5428 // Scan all of the users of the LHS, looking for VCMPo's that match.
5429 SDNode *VCMPoNode = 0;
5431 SDNode *LHSN = N->getOperand(0).getNode();
5432 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5434 if (UI->getOpcode() == PPCISD::VCMPo &&
5435 UI->getOperand(1) == N->getOperand(1) &&
5436 UI->getOperand(2) == N->getOperand(2) &&
5437 UI->getOperand(0) == N->getOperand(0)) {
5442 // If there is no VCMPo node, or if the flag value has a single use, don't
5444 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5447 // Look at the (necessarily single) use of the flag value. If it has a
5448 // chain, this transformation is more complex. Note that multiple things
5449 // could use the value result, which we should ignore.
5450 SDNode *FlagUser = 0;
5451 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5452 FlagUser == 0; ++UI) {
5453 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5455 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5456 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5463 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5464 // give up for right now.
5465 if (FlagUser->getOpcode() == PPCISD::MFCR)
5466 return SDValue(VCMPoNode, 0);
5471 // If this is a branch on an altivec predicate comparison, lower this so
5472 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5473 // lowering is done pre-legalize, because the legalizer lowers the predicate
5474 // compare down to code that is difficult to reassemble.
5475 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5476 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5480 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5481 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5482 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5483 assert(isDot && "Can't compare against a vector result!");
5485 // If this is a comparison against something other than 0/1, then we know
5486 // that the condition is never/always true.
5487 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5488 if (Val != 0 && Val != 1) {
5489 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5490 return N->getOperand(0);
5491 // Always !=, turn it into an unconditional branch.
5492 return DAG.getNode(ISD::BR, dl, MVT::Other,
5493 N->getOperand(0), N->getOperand(4));
5496 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5498 // Create the PPCISD altivec 'dot' comparison node.
5499 std::vector<EVT> VTs;
5501 LHS.getOperand(2), // LHS of compare
5502 LHS.getOperand(3), // RHS of compare
5503 DAG.getConstant(CompareOpc, MVT::i32)
5505 VTs.push_back(LHS.getOperand(2).getValueType());
5506 VTs.push_back(MVT::Glue);
5507 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5509 // Unpack the result based on how the target uses it.
5510 PPC::Predicate CompOpc;
5511 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5512 default: // Can't happen, don't crash on invalid number though.
5513 case 0: // Branch on the value of the EQ bit of CR6.
5514 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5516 case 1: // Branch on the inverted value of the EQ bit of CR6.
5517 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5519 case 2: // Branch on the value of the LT bit of CR6.
5520 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5522 case 3: // Branch on the inverted value of the LT bit of CR6.
5523 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5527 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5528 DAG.getConstant(CompOpc, MVT::i32),
5529 DAG.getRegister(PPC::CR6, MVT::i32),
5530 N->getOperand(4), CompNode.getValue(1));
5539 //===----------------------------------------------------------------------===//
5540 // Inline Assembly Support
5541 //===----------------------------------------------------------------------===//
5543 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5546 const SelectionDAG &DAG,
5547 unsigned Depth) const {
5548 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
5549 switch (Op.getOpcode()) {
5551 case PPCISD::LBRX: {
5552 // lhbrx is known to have the top bits cleared out.
5553 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5554 KnownZero = 0xFFFF0000;
5557 case ISD::INTRINSIC_WO_CHAIN: {
5558 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5560 case Intrinsic::ppc_altivec_vcmpbfp_p:
5561 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5562 case Intrinsic::ppc_altivec_vcmpequb_p:
5563 case Intrinsic::ppc_altivec_vcmpequh_p:
5564 case Intrinsic::ppc_altivec_vcmpequw_p:
5565 case Intrinsic::ppc_altivec_vcmpgefp_p:
5566 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5567 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5568 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5569 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5570 case Intrinsic::ppc_altivec_vcmpgtub_p:
5571 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5572 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5573 KnownZero = ~1U; // All bits but the low one are known to be zero.
5581 /// getConstraintType - Given a constraint, return the type of
5582 /// constraint it is for this target.
5583 PPCTargetLowering::ConstraintType
5584 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5585 if (Constraint.size() == 1) {
5586 switch (Constraint[0]) {
5593 return C_RegisterClass;
5596 return TargetLowering::getConstraintType(Constraint);
5599 /// Examine constraint type and operand type and determine a weight value.
5600 /// This object must already have been set up with the operand type
5601 /// and the current alternative constraint selected.
5602 TargetLowering::ConstraintWeight
5603 PPCTargetLowering::getSingleConstraintMatchWeight(
5604 AsmOperandInfo &info, const char *constraint) const {
5605 ConstraintWeight weight = CW_Invalid;
5606 Value *CallOperandVal = info.CallOperandVal;
5607 // If we don't have a value, we can't do a match,
5608 // but allow it at the lowest weight.
5609 if (CallOperandVal == NULL)
5611 Type *type = CallOperandVal->getType();
5612 // Look at the constraint type.
5613 switch (*constraint) {
5615 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5618 if (type->isIntegerTy())
5619 weight = CW_Register;
5622 if (type->isFloatTy())
5623 weight = CW_Register;
5626 if (type->isDoubleTy())
5627 weight = CW_Register;
5630 if (type->isVectorTy())
5631 weight = CW_Register;
5634 weight = CW_Register;
5640 std::pair<unsigned, const TargetRegisterClass*>
5641 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5643 if (Constraint.size() == 1) {
5644 // GCC RS6000 Constraint Letters
5645 switch (Constraint[0]) {
5648 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5649 return std::make_pair(0U, &PPC::G8RCRegClass);
5650 return std::make_pair(0U, &PPC::GPRCRegClass);
5653 return std::make_pair(0U, &PPC::F4RCRegClass);
5655 return std::make_pair(0U, &PPC::F8RCRegClass);
5658 return std::make_pair(0U, &PPC::VRRCRegClass);
5660 return std::make_pair(0U, &PPC::CRRCRegClass);
5664 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5668 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5669 /// vector. If it is invalid, don't add anything to Ops.
5670 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5671 std::string &Constraint,
5672 std::vector<SDValue>&Ops,
5673 SelectionDAG &DAG) const {
5674 SDValue Result(0,0);
5676 // Only support length 1 constraints.
5677 if (Constraint.length() > 1) return;
5679 char Letter = Constraint[0];
5690 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5691 if (!CST) return; // Must be an immediate to match.
5692 unsigned Value = CST->getZExtValue();
5694 default: llvm_unreachable("Unknown constraint letter!");
5695 case 'I': // "I" is a signed 16-bit constant.
5696 if ((short)Value == (int)Value)
5697 Result = DAG.getTargetConstant(Value, Op.getValueType());
5699 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5700 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5701 if ((short)Value == 0)
5702 Result = DAG.getTargetConstant(Value, Op.getValueType());
5704 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5705 if ((Value >> 16) == 0)
5706 Result = DAG.getTargetConstant(Value, Op.getValueType());
5708 case 'M': // "M" is a constant that is greater than 31.
5710 Result = DAG.getTargetConstant(Value, Op.getValueType());
5712 case 'N': // "N" is a positive constant that is an exact power of two.
5713 if ((int)Value > 0 && isPowerOf2_32(Value))
5714 Result = DAG.getTargetConstant(Value, Op.getValueType());
5716 case 'O': // "O" is the constant zero.
5718 Result = DAG.getTargetConstant(Value, Op.getValueType());
5720 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5721 if ((short)-Value == (int)-Value)
5722 Result = DAG.getTargetConstant(Value, Op.getValueType());
5729 if (Result.getNode()) {
5730 Ops.push_back(Result);
5734 // Handle standard constraint letters.
5735 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5738 // isLegalAddressingMode - Return true if the addressing mode represented
5739 // by AM is legal for this target, for a load/store of the specified type.
5740 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5742 // FIXME: PPC does not allow r+i addressing modes for vectors!
5744 // PPC allows a sign-extended 16-bit immediate field.
5745 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5748 // No global is ever allowed as a base.
5752 // PPC only support r+r,
5754 case 0: // "r+i" or just "i", depending on HasBaseReg.
5757 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5759 // Otherwise we have r+r or r+i.
5762 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5764 // Allow 2*r as r+r.
5767 // No other scales are supported.
5774 /// isLegalAddressImmediate - Return true if the integer value can be used
5775 /// as the offset of the target addressing mode for load / store of the
5777 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5778 // PPC allows a sign-extended 16-bit immediate field.
5779 return (V > -(1 << 16) && V < (1 << 16)-1);
5782 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
5786 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5787 SelectionDAG &DAG) const {
5788 MachineFunction &MF = DAG.getMachineFunction();
5789 MachineFrameInfo *MFI = MF.getFrameInfo();
5790 MFI->setReturnAddressIsTaken(true);
5792 DebugLoc dl = Op.getDebugLoc();
5793 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5795 // Make sure the function does not optimize away the store of the RA to
5797 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5798 FuncInfo->setLRStoreRequired();
5799 bool isPPC64 = PPCSubTarget.isPPC64();
5800 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5803 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5806 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5807 isPPC64? MVT::i64 : MVT::i32);
5808 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5809 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5811 MachinePointerInfo(), false, false, false, 0);
5814 // Just load the return address off the stack.
5815 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5816 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5817 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5820 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5821 SelectionDAG &DAG) const {
5822 DebugLoc dl = Op.getDebugLoc();
5823 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5825 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5826 bool isPPC64 = PtrVT == MVT::i64;
5828 MachineFunction &MF = DAG.getMachineFunction();
5829 MachineFrameInfo *MFI = MF.getFrameInfo();
5830 MFI->setFrameAddressIsTaken(true);
5831 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
5832 MFI->hasVarSizedObjects()) &&
5833 MFI->getStackSize() &&
5834 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5835 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5836 (is31 ? PPC::R31 : PPC::R1);
5837 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5840 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5841 FrameAddr, MachinePointerInfo(), false, false,
5847 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5848 // The PowerPC target isn't yet aware of offsets.
5852 /// getOptimalMemOpType - Returns the target specific optimal type for load
5853 /// and store operations as a result of memset, memcpy, and memmove
5854 /// lowering. If DstAlign is zero that means it's safe to destination
5855 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5856 /// means there isn't a need to check it against alignment requirement,
5857 /// probably because the source does not need to be loaded. If
5858 /// 'IsZeroVal' is true, that means it's safe to return a
5859 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5860 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5861 /// constant so it does not need to be loaded.
5862 /// It returns EVT::Other if the type should be determined using generic
5863 /// target-independent logic.
5864 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5865 unsigned DstAlign, unsigned SrcAlign,
5868 MachineFunction &MF) const {
5869 if (this->PPCSubTarget.isPPC64()) {
5876 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
5878 return TargetLowering::getSchedulingPreference(N);