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()) {
69 const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
73 // Use _setjmp/_longjmp instead of setjmp/longjmp.
74 setUseUnderscoreSetJmp(true);
75 setUseUnderscoreLongJmp(true);
77 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
78 // arguments are at least 4/8 bytes aligned.
79 bool isPPC64 = Subtarget->isPPC64();
80 setMinStackArgumentAlignment(isPPC64 ? 8:4);
82 // Set up the register classes.
83 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
84 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
85 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
87 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
88 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
91 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
93 // PowerPC has pre-inc load and store's.
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
96 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
97 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
98 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
101 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
102 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
103 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
105 // This is used in the ppcf128->int sequence. Note it has different semantics
106 // from FP_ROUND: that rounds to nearest, this rounds to zero.
107 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
109 // We do not currently implement these libm ops for PowerPC.
110 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
111 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
112 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
113 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
114 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
116 // PowerPC has no SREM/UREM instructions
117 setOperationAction(ISD::SREM, MVT::i32, Expand);
118 setOperationAction(ISD::UREM, MVT::i32, Expand);
119 setOperationAction(ISD::SREM, MVT::i64, Expand);
120 setOperationAction(ISD::UREM, MVT::i64, Expand);
122 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
123 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
124 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
125 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
126 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
127 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
128 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
129 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
130 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
132 // We don't support sin/cos/sqrt/fmod/pow
133 setOperationAction(ISD::FSIN , MVT::f64, Expand);
134 setOperationAction(ISD::FCOS , MVT::f64, Expand);
135 setOperationAction(ISD::FREM , MVT::f64, Expand);
136 setOperationAction(ISD::FPOW , MVT::f64, Expand);
137 setOperationAction(ISD::FMA , MVT::f64, Legal);
138 setOperationAction(ISD::FSIN , MVT::f32, Expand);
139 setOperationAction(ISD::FCOS , MVT::f32, Expand);
140 setOperationAction(ISD::FREM , MVT::f32, Expand);
141 setOperationAction(ISD::FPOW , MVT::f32, Expand);
142 setOperationAction(ISD::FMA , MVT::f32, Legal);
144 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
146 // If we're enabling GP optimizations, use hardware square root
147 if (!Subtarget->hasFSQRT()) {
148 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
149 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
152 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
153 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
155 // PowerPC does not have BSWAP, CTPOP or CTTZ
156 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
157 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
158 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
160 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
161 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
162 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
163 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
164 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
165 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
167 // PowerPC does not have ROTR
168 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
169 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
171 // PowerPC does not have Select
172 setOperationAction(ISD::SELECT, MVT::i32, Expand);
173 setOperationAction(ISD::SELECT, MVT::i64, Expand);
174 setOperationAction(ISD::SELECT, MVT::f32, Expand);
175 setOperationAction(ISD::SELECT, MVT::f64, Expand);
177 // PowerPC wants to turn select_cc of FP into fsel when possible.
178 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
179 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
181 // PowerPC wants to optimize integer setcc a bit
182 setOperationAction(ISD::SETCC, MVT::i32, Custom);
184 // PowerPC does not have BRCOND which requires SetCC
185 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
187 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
189 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
190 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
192 // PowerPC does not have [U|S]INT_TO_FP
193 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
194 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
196 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
197 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
198 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
199 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
201 // We cannot sextinreg(i1). Expand to shifts.
202 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
204 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
205 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
206 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
207 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
210 // We want to legalize GlobalAddress and ConstantPool nodes into the
211 // appropriate instructions to materialize the address.
212 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
213 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
214 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
215 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
216 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
217 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
218 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
219 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
220 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
221 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
224 setOperationAction(ISD::TRAP, MVT::Other, Legal);
226 // TRAMPOLINE is custom lowered.
227 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
228 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
230 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
231 setOperationAction(ISD::VASTART , MVT::Other, Custom);
233 if (Subtarget->isSVR4ABI()) {
235 // VAARG always uses double-word chunks, so promote anything smaller.
236 setOperationAction(ISD::VAARG, MVT::i1, Promote);
237 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
238 setOperationAction(ISD::VAARG, MVT::i8, Promote);
239 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
240 setOperationAction(ISD::VAARG, MVT::i16, Promote);
241 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
242 setOperationAction(ISD::VAARG, MVT::i32, Promote);
243 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
244 setOperationAction(ISD::VAARG, MVT::Other, Expand);
246 // VAARG is custom lowered with the 32-bit SVR4 ABI.
247 setOperationAction(ISD::VAARG, MVT::Other, Custom);
248 setOperationAction(ISD::VAARG, MVT::i64, Custom);
251 setOperationAction(ISD::VAARG, MVT::Other, Expand);
253 // Use the default implementation.
254 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
255 setOperationAction(ISD::VAEND , MVT::Other, Expand);
256 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
257 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
258 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
259 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
261 // We want to custom lower some of our intrinsics.
262 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
264 // Comparisons that require checking two conditions.
265 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
266 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
267 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
270 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
271 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
272 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
273 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
274 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
275 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
276 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
278 if (Subtarget->has64BitSupport()) {
279 // They also have instructions for converting between i64 and fp.
280 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
281 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
282 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
283 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
284 // This is just the low 32 bits of a (signed) fp->i64 conversion.
285 // We cannot do this with Promote because i64 is not a legal type.
286 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
288 // FIXME: disable this lowered code. This generates 64-bit register values,
289 // and we don't model the fact that the top part is clobbered by calls. We
290 // need to flag these together so that the value isn't live across a call.
291 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
293 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
294 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
297 if (Subtarget->use64BitRegs()) {
298 // 64-bit PowerPC implementations can support i64 types directly
299 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
300 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
301 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
302 // 64-bit PowerPC wants to expand i128 shifts itself.
303 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
304 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
305 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
307 // 32-bit PowerPC wants to expand i64 shifts itself.
308 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
309 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
310 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
313 if (Subtarget->hasAltivec()) {
314 // First set operation action for all vector types to expand. Then we
315 // will selectively turn on ones that can be effectively codegen'd.
316 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
317 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
318 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
320 // add/sub are legal for all supported vector VT's.
321 setOperationAction(ISD::ADD , VT, Legal);
322 setOperationAction(ISD::SUB , VT, Legal);
324 // We promote all shuffles to v16i8.
325 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
326 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
328 // We promote all non-typed operations to v4i32.
329 setOperationAction(ISD::AND , VT, Promote);
330 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
331 setOperationAction(ISD::OR , VT, Promote);
332 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
333 setOperationAction(ISD::XOR , VT, Promote);
334 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
335 setOperationAction(ISD::LOAD , VT, Promote);
336 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
337 setOperationAction(ISD::SELECT, VT, Promote);
338 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
339 setOperationAction(ISD::STORE, VT, Promote);
340 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
342 // No other operations are legal.
343 setOperationAction(ISD::MUL , VT, Expand);
344 setOperationAction(ISD::SDIV, VT, Expand);
345 setOperationAction(ISD::SREM, VT, Expand);
346 setOperationAction(ISD::UDIV, VT, Expand);
347 setOperationAction(ISD::UREM, VT, Expand);
348 setOperationAction(ISD::FDIV, VT, Expand);
349 setOperationAction(ISD::FNEG, VT, Expand);
350 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
351 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
352 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
353 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
354 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
355 setOperationAction(ISD::UDIVREM, VT, Expand);
356 setOperationAction(ISD::SDIVREM, VT, Expand);
357 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
358 setOperationAction(ISD::FPOW, VT, Expand);
359 setOperationAction(ISD::CTPOP, VT, Expand);
360 setOperationAction(ISD::CTLZ, VT, Expand);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
362 setOperationAction(ISD::CTTZ, VT, Expand);
363 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
366 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
367 // with merges, splats, etc.
368 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
370 setOperationAction(ISD::AND , MVT::v4i32, Legal);
371 setOperationAction(ISD::OR , MVT::v4i32, Legal);
372 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
373 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
374 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
375 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
377 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
378 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
379 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
380 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
382 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
383 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
384 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
385 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
386 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
388 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
389 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
391 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
392 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
393 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
394 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
397 if (Subtarget->has64BitSupport()) {
398 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
399 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
402 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
403 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
405 setBooleanContents(ZeroOrOneBooleanContent);
406 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
409 setStackPointerRegisterToSaveRestore(PPC::X1);
410 setExceptionPointerRegister(PPC::X3);
411 setExceptionSelectorRegister(PPC::X4);
413 setStackPointerRegisterToSaveRestore(PPC::R1);
414 setExceptionPointerRegister(PPC::R3);
415 setExceptionSelectorRegister(PPC::R4);
418 // We have target-specific dag combine patterns for the following nodes:
419 setTargetDAGCombine(ISD::SINT_TO_FP);
420 setTargetDAGCombine(ISD::STORE);
421 setTargetDAGCombine(ISD::BR_CC);
422 setTargetDAGCombine(ISD::BSWAP);
424 // Darwin long double math library functions have $LDBL128 appended.
425 if (Subtarget->isDarwin()) {
426 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
427 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
428 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
429 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
430 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
431 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
432 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
433 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
434 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
435 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
438 setMinFunctionAlignment(2);
439 if (PPCSubTarget.isDarwin())
440 setPrefFunctionAlignment(4);
442 if (isPPC64 && Subtarget->isJITCodeModel())
443 // Temporary workaround for the inability of PPC64 JIT to handle jump
445 setSupportJumpTables(false);
447 setInsertFencesForAtomic(true);
449 setSchedulingPreference(Sched::Hybrid);
451 computeRegisterProperties();
454 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
455 /// function arguments in the caller parameter area.
456 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
457 const TargetMachine &TM = getTargetMachine();
458 // Darwin passes everything on 4 byte boundary.
459 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
462 // 16byte and wider vectors are passed on 16byte boundary.
463 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
464 if (VTy->getBitWidth() >= 128)
467 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
468 if (PPCSubTarget.isPPC64())
474 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
477 case PPCISD::FSEL: return "PPCISD::FSEL";
478 case PPCISD::FCFID: return "PPCISD::FCFID";
479 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
480 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
481 case PPCISD::STFIWX: return "PPCISD::STFIWX";
482 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
483 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
484 case PPCISD::VPERM: return "PPCISD::VPERM";
485 case PPCISD::Hi: return "PPCISD::Hi";
486 case PPCISD::Lo: return "PPCISD::Lo";
487 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
488 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
489 case PPCISD::LOAD: return "PPCISD::LOAD";
490 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
491 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
492 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
493 case PPCISD::SRL: return "PPCISD::SRL";
494 case PPCISD::SRA: return "PPCISD::SRA";
495 case PPCISD::SHL: return "PPCISD::SHL";
496 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
497 case PPCISD::STD_32: return "PPCISD::STD_32";
498 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
499 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
500 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
501 case PPCISD::NOP: return "PPCISD::NOP";
502 case PPCISD::MTCTR: return "PPCISD::MTCTR";
503 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
504 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
505 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
506 case PPCISD::MFCR: return "PPCISD::MFCR";
507 case PPCISD::VCMP: return "PPCISD::VCMP";
508 case PPCISD::VCMPo: return "PPCISD::VCMPo";
509 case PPCISD::LBRX: return "PPCISD::LBRX";
510 case PPCISD::STBRX: return "PPCISD::STBRX";
511 case PPCISD::LARX: return "PPCISD::LARX";
512 case PPCISD::STCX: return "PPCISD::STCX";
513 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
514 case PPCISD::MFFS: return "PPCISD::MFFS";
515 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
516 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
517 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
518 case PPCISD::MTFSF: return "PPCISD::MTFSF";
519 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
523 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
527 //===----------------------------------------------------------------------===//
528 // Node matching predicates, for use by the tblgen matching code.
529 //===----------------------------------------------------------------------===//
531 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
532 static bool isFloatingPointZero(SDValue Op) {
533 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
534 return CFP->getValueAPF().isZero();
535 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
536 // Maybe this has already been legalized into the constant pool?
537 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
538 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
539 return CFP->getValueAPF().isZero();
544 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
545 /// true if Op is undef or if it matches the specified value.
546 static bool isConstantOrUndef(int Op, int Val) {
547 return Op < 0 || Op == Val;
550 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
551 /// VPKUHUM instruction.
552 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
554 for (unsigned i = 0; i != 16; ++i)
555 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
558 for (unsigned i = 0; i != 8; ++i)
559 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
560 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
566 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
567 /// VPKUWUM instruction.
568 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
570 for (unsigned i = 0; i != 16; i += 2)
571 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
572 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
575 for (unsigned i = 0; i != 8; i += 2)
576 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
577 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
578 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
579 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
585 /// isVMerge - Common function, used to match vmrg* shuffles.
587 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
588 unsigned LHSStart, unsigned RHSStart) {
589 assert(N->getValueType(0) == MVT::v16i8 &&
590 "PPC only supports shuffles by bytes!");
591 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
592 "Unsupported merge size!");
594 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
595 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
596 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
597 LHSStart+j+i*UnitSize) ||
598 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
599 RHSStart+j+i*UnitSize))
605 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
606 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
607 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
610 return isVMerge(N, UnitSize, 8, 24);
611 return isVMerge(N, UnitSize, 8, 8);
614 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
615 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
616 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
619 return isVMerge(N, UnitSize, 0, 16);
620 return isVMerge(N, UnitSize, 0, 0);
624 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
625 /// amount, otherwise return -1.
626 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
627 assert(N->getValueType(0) == MVT::v16i8 &&
628 "PPC only supports shuffles by bytes!");
630 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
632 // Find the first non-undef value in the shuffle mask.
634 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
637 if (i == 16) return -1; // all undef.
639 // Otherwise, check to see if the rest of the elements are consecutively
640 // numbered from this value.
641 unsigned ShiftAmt = SVOp->getMaskElt(i);
642 if (ShiftAmt < i) return -1;
646 // Check the rest of the elements to see if they are consecutive.
647 for (++i; i != 16; ++i)
648 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
651 // Check the rest of the elements to see if they are consecutive.
652 for (++i; i != 16; ++i)
653 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
659 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
660 /// specifies a splat of a single element that is suitable for input to
661 /// VSPLTB/VSPLTH/VSPLTW.
662 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
663 assert(N->getValueType(0) == MVT::v16i8 &&
664 (EltSize == 1 || EltSize == 2 || EltSize == 4));
666 // This is a splat operation if each element of the permute is the same, and
667 // if the value doesn't reference the second vector.
668 unsigned ElementBase = N->getMaskElt(0);
670 // FIXME: Handle UNDEF elements too!
671 if (ElementBase >= 16)
674 // Check that the indices are consecutive, in the case of a multi-byte element
675 // splatted with a v16i8 mask.
676 for (unsigned i = 1; i != EltSize; ++i)
677 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
680 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
681 if (N->getMaskElt(i) < 0) continue;
682 for (unsigned j = 0; j != EltSize; ++j)
683 if (N->getMaskElt(i+j) != N->getMaskElt(j))
689 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
691 bool PPC::isAllNegativeZeroVector(SDNode *N) {
692 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
694 APInt APVal, APUndef;
698 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
699 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
700 return CFP->getValueAPF().isNegZero();
705 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
706 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
707 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
708 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
709 assert(isSplatShuffleMask(SVOp, EltSize));
710 return SVOp->getMaskElt(0) / EltSize;
713 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
714 /// by using a vspltis[bhw] instruction of the specified element size, return
715 /// the constant being splatted. The ByteSize field indicates the number of
716 /// bytes of each element [124] -> [bhw].
717 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
720 // If ByteSize of the splat is bigger than the element size of the
721 // build_vector, then we have a case where we are checking for a splat where
722 // multiple elements of the buildvector are folded together into a single
723 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
724 unsigned EltSize = 16/N->getNumOperands();
725 if (EltSize < ByteSize) {
726 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
727 SDValue UniquedVals[4];
728 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
730 // See if all of the elements in the buildvector agree across.
731 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
732 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
733 // If the element isn't a constant, bail fully out.
734 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
737 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
738 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
739 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
740 return SDValue(); // no match.
743 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
744 // either constant or undef values that are identical for each chunk. See
745 // if these chunks can form into a larger vspltis*.
747 // Check to see if all of the leading entries are either 0 or -1. If
748 // neither, then this won't fit into the immediate field.
749 bool LeadingZero = true;
750 bool LeadingOnes = true;
751 for (unsigned i = 0; i != Multiple-1; ++i) {
752 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
754 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
755 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
757 // Finally, check the least significant entry.
759 if (UniquedVals[Multiple-1].getNode() == 0)
760 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
761 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
763 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
766 if (UniquedVals[Multiple-1].getNode() == 0)
767 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
768 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
769 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
770 return DAG.getTargetConstant(Val, MVT::i32);
776 // Check to see if this buildvec has a single non-undef value in its elements.
777 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
778 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
779 if (OpVal.getNode() == 0)
780 OpVal = N->getOperand(i);
781 else if (OpVal != N->getOperand(i))
785 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
787 unsigned ValSizeInBytes = EltSize;
789 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
790 Value = CN->getZExtValue();
791 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
792 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
793 Value = FloatToBits(CN->getValueAPF().convertToFloat());
796 // If the splat value is larger than the element value, then we can never do
797 // this splat. The only case that we could fit the replicated bits into our
798 // immediate field for would be zero, and we prefer to use vxor for it.
799 if (ValSizeInBytes < ByteSize) return SDValue();
801 // If the element value is larger than the splat value, cut it in half and
802 // check to see if the two halves are equal. Continue doing this until we
803 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
804 while (ValSizeInBytes > ByteSize) {
805 ValSizeInBytes >>= 1;
807 // If the top half equals the bottom half, we're still ok.
808 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
809 (Value & ((1 << (8*ValSizeInBytes))-1)))
813 // Properly sign extend the value.
814 int MaskVal = SignExtend32(Value, ByteSize * 8);
816 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
817 if (MaskVal == 0) return SDValue();
819 // Finally, if this value fits in a 5 bit sext field, return it
820 if (SignExtend32<5>(MaskVal) == MaskVal)
821 return DAG.getTargetConstant(MaskVal, MVT::i32);
825 //===----------------------------------------------------------------------===//
826 // Addressing Mode Selection
827 //===----------------------------------------------------------------------===//
829 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
830 /// or 64-bit immediate, and if the value can be accurately represented as a
831 /// sign extension from a 16-bit value. If so, this returns true and the
833 static bool isIntS16Immediate(SDNode *N, short &Imm) {
834 if (N->getOpcode() != ISD::Constant)
837 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
838 if (N->getValueType(0) == MVT::i32)
839 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
841 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
843 static bool isIntS16Immediate(SDValue Op, short &Imm) {
844 return isIntS16Immediate(Op.getNode(), Imm);
848 /// SelectAddressRegReg - Given the specified addressed, check to see if it
849 /// can be represented as an indexed [r+r] operation. Returns false if it
850 /// can be more efficiently represented with [r+imm].
851 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
853 SelectionDAG &DAG) const {
855 if (N.getOpcode() == ISD::ADD) {
856 if (isIntS16Immediate(N.getOperand(1), imm))
858 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
861 Base = N.getOperand(0);
862 Index = N.getOperand(1);
864 } else if (N.getOpcode() == ISD::OR) {
865 if (isIntS16Immediate(N.getOperand(1), imm))
866 return false; // r+i can fold it if we can.
868 // If this is an or of disjoint bitfields, we can codegen this as an add
869 // (for better address arithmetic) if the LHS and RHS of the OR are provably
871 APInt LHSKnownZero, LHSKnownOne;
872 APInt RHSKnownZero, RHSKnownOne;
873 DAG.ComputeMaskedBits(N.getOperand(0),
874 LHSKnownZero, LHSKnownOne);
876 if (LHSKnownZero.getBoolValue()) {
877 DAG.ComputeMaskedBits(N.getOperand(1),
878 RHSKnownZero, RHSKnownOne);
879 // If all of the bits are known zero on the LHS or RHS, the add won't
881 if (~(LHSKnownZero | RHSKnownZero) == 0) {
882 Base = N.getOperand(0);
883 Index = N.getOperand(1);
892 /// Returns true if the address N can be represented by a base register plus
893 /// a signed 16-bit displacement [r+imm], and if it is not better
894 /// represented as reg+reg.
895 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
897 SelectionDAG &DAG) const {
898 // FIXME dl should come from parent load or store, not from address
899 DebugLoc dl = N.getDebugLoc();
900 // If this can be more profitably realized as r+r, fail.
901 if (SelectAddressRegReg(N, Disp, Base, DAG))
904 if (N.getOpcode() == ISD::ADD) {
906 if (isIntS16Immediate(N.getOperand(1), imm)) {
907 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
908 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
909 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
911 Base = N.getOperand(0);
913 return true; // [r+i]
914 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
915 // Match LOAD (ADD (X, Lo(G))).
916 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
917 && "Cannot handle constant offsets yet!");
918 Disp = N.getOperand(1).getOperand(0); // The global address.
919 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
920 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
921 Disp.getOpcode() == ISD::TargetConstantPool ||
922 Disp.getOpcode() == ISD::TargetJumpTable);
923 Base = N.getOperand(0);
924 return true; // [&g+r]
926 } else if (N.getOpcode() == ISD::OR) {
928 if (isIntS16Immediate(N.getOperand(1), imm)) {
929 // If this is an or of disjoint bitfields, we can codegen this as an add
930 // (for better address arithmetic) if the LHS and RHS of the OR are
931 // provably disjoint.
932 APInt LHSKnownZero, LHSKnownOne;
933 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
935 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
936 // If all of the bits are known zero on the LHS or RHS, the add won't
938 Base = N.getOperand(0);
939 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
943 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
944 // Loading from a constant address.
946 // If this address fits entirely in a 16-bit sext immediate field, codegen
949 if (isIntS16Immediate(CN, Imm)) {
950 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
951 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
952 CN->getValueType(0));
956 // Handle 32-bit sext immediates with LIS + addr mode.
957 if (CN->getValueType(0) == MVT::i32 ||
958 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
959 int Addr = (int)CN->getZExtValue();
961 // Otherwise, break this down into an LIS + disp.
962 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
964 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
965 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
966 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
971 Disp = DAG.getTargetConstant(0, getPointerTy());
972 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
973 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
976 return true; // [r+0]
979 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
980 /// represented as an indexed [r+r] operation.
981 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
983 SelectionDAG &DAG) const {
984 // Check to see if we can easily represent this as an [r+r] address. This
985 // will fail if it thinks that the address is more profitably represented as
986 // reg+imm, e.g. where imm = 0.
987 if (SelectAddressRegReg(N, Base, Index, DAG))
990 // If the operand is an addition, always emit this as [r+r], since this is
991 // better (for code size, and execution, as the memop does the add for free)
992 // than emitting an explicit add.
993 if (N.getOpcode() == ISD::ADD) {
994 Base = N.getOperand(0);
995 Index = N.getOperand(1);
999 // Otherwise, do it the hard way, using R0 as the base register.
1000 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1006 /// SelectAddressRegImmShift - Returns true if the address N can be
1007 /// represented by a base register plus a signed 14-bit displacement
1008 /// [r+imm*4]. Suitable for use by STD and friends.
1009 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1011 SelectionDAG &DAG) const {
1012 // FIXME dl should come from the parent load or store, not the address
1013 DebugLoc dl = N.getDebugLoc();
1014 // If this can be more profitably realized as r+r, fail.
1015 if (SelectAddressRegReg(N, Disp, Base, DAG))
1018 if (N.getOpcode() == ISD::ADD) {
1020 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1021 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1022 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1023 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1025 Base = N.getOperand(0);
1027 return true; // [r+i]
1028 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1029 // Match LOAD (ADD (X, Lo(G))).
1030 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1031 && "Cannot handle constant offsets yet!");
1032 Disp = N.getOperand(1).getOperand(0); // The global address.
1033 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1034 Disp.getOpcode() == ISD::TargetConstantPool ||
1035 Disp.getOpcode() == ISD::TargetJumpTable);
1036 Base = N.getOperand(0);
1037 return true; // [&g+r]
1039 } else if (N.getOpcode() == ISD::OR) {
1041 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1042 // If this is an or of disjoint bitfields, we can codegen this as an add
1043 // (for better address arithmetic) if the LHS and RHS of the OR are
1044 // provably disjoint.
1045 APInt LHSKnownZero, LHSKnownOne;
1046 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1047 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1048 // If all of the bits are known zero on the LHS or RHS, the add won't
1050 Base = N.getOperand(0);
1051 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1055 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1056 // Loading from a constant address. Verify low two bits are clear.
1057 if ((CN->getZExtValue() & 3) == 0) {
1058 // If this address fits entirely in a 14-bit sext immediate field, codegen
1061 if (isIntS16Immediate(CN, Imm)) {
1062 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1063 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1064 CN->getValueType(0));
1068 // Fold the low-part of 32-bit absolute addresses into addr mode.
1069 if (CN->getValueType(0) == MVT::i32 ||
1070 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1071 int Addr = (int)CN->getZExtValue();
1073 // Otherwise, break this down into an LIS + disp.
1074 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1075 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1076 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1077 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1083 Disp = DAG.getTargetConstant(0, getPointerTy());
1084 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1085 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1088 return true; // [r+0]
1092 /// getPreIndexedAddressParts - returns true by value, base pointer and
1093 /// offset pointer and addressing mode by reference if the node's address
1094 /// can be legally represented as pre-indexed load / store address.
1095 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1097 ISD::MemIndexedMode &AM,
1098 SelectionDAG &DAG) const {
1099 if (DisablePPCPreinc) return false;
1103 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1104 Ptr = LD->getBasePtr();
1105 VT = LD->getMemoryVT();
1107 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1108 Ptr = ST->getBasePtr();
1109 VT = ST->getMemoryVT();
1113 // PowerPC doesn't have preinc load/store instructions for vectors.
1117 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1122 // LDU/STU use reg+imm*4, others use reg+imm.
1123 if (VT != MVT::i64) {
1125 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1129 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1133 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1134 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1135 // sext i32 to i64 when addr mode is r+i.
1136 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1137 LD->getExtensionType() == ISD::SEXTLOAD &&
1138 isa<ConstantSDNode>(Offset))
1146 //===----------------------------------------------------------------------===//
1147 // LowerOperation implementation
1148 //===----------------------------------------------------------------------===//
1150 /// GetLabelAccessInfo - Return true if we should reference labels using a
1151 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1152 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1153 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1154 HiOpFlags = PPCII::MO_HA16;
1155 LoOpFlags = PPCII::MO_LO16;
1157 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1158 // non-darwin platform. We don't support PIC on other platforms yet.
1159 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1160 TM.getSubtarget<PPCSubtarget>().isDarwin();
1162 HiOpFlags |= PPCII::MO_PIC_FLAG;
1163 LoOpFlags |= PPCII::MO_PIC_FLAG;
1166 // If this is a reference to a global value that requires a non-lazy-ptr, make
1167 // sure that instruction lowering adds it.
1168 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1169 HiOpFlags |= PPCII::MO_NLP_FLAG;
1170 LoOpFlags |= PPCII::MO_NLP_FLAG;
1172 if (GV->hasHiddenVisibility()) {
1173 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1174 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1181 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1182 SelectionDAG &DAG) {
1183 EVT PtrVT = HiPart.getValueType();
1184 SDValue Zero = DAG.getConstant(0, PtrVT);
1185 DebugLoc DL = HiPart.getDebugLoc();
1187 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1188 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1190 // With PIC, the first instruction is actually "GR+hi(&G)".
1192 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1193 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1195 // Generate non-pic code that has direct accesses to the constant pool.
1196 // The address of the global is just (hi(&g)+lo(&g)).
1197 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1200 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1201 SelectionDAG &DAG) const {
1202 EVT PtrVT = Op.getValueType();
1203 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1204 const Constant *C = CP->getConstVal();
1206 // 64-bit SVR4 ABI code is always position-independent.
1207 // The actual address of the GlobalValue is stored in the TOC.
1208 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1209 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1210 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1211 DAG.getRegister(PPC::X2, MVT::i64));
1214 unsigned MOHiFlag, MOLoFlag;
1215 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1217 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1219 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1220 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1223 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1224 EVT PtrVT = Op.getValueType();
1225 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1227 // 64-bit SVR4 ABI code is always position-independent.
1228 // The actual address of the GlobalValue is stored in the TOC.
1229 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1230 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1231 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1232 DAG.getRegister(PPC::X2, MVT::i64));
1235 unsigned MOHiFlag, MOLoFlag;
1236 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1237 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1238 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1239 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1242 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1243 SelectionDAG &DAG) const {
1244 EVT PtrVT = Op.getValueType();
1246 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1248 unsigned MOHiFlag, MOLoFlag;
1249 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1250 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1251 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1252 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1255 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1256 SelectionDAG &DAG) const {
1258 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1259 DebugLoc dl = GA->getDebugLoc();
1260 const GlobalValue *GV = GA->getGlobal();
1261 EVT PtrVT = getPointerTy();
1262 bool is64bit = PPCSubTarget.isPPC64();
1264 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1266 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1267 PPCII::MO_TPREL16_HA);
1268 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1269 PPCII::MO_TPREL16_LO);
1271 if (model != TLSModel::LocalExec)
1272 llvm_unreachable("only local-exec TLS mode supported");
1273 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1274 is64bit ? MVT::i64 : MVT::i32);
1275 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1276 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1279 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1280 SelectionDAG &DAG) const {
1281 EVT PtrVT = Op.getValueType();
1282 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1283 DebugLoc DL = GSDN->getDebugLoc();
1284 const GlobalValue *GV = GSDN->getGlobal();
1286 // 64-bit SVR4 ABI code is always position-independent.
1287 // The actual address of the GlobalValue is stored in the TOC.
1288 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1289 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1290 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1291 DAG.getRegister(PPC::X2, MVT::i64));
1294 unsigned MOHiFlag, MOLoFlag;
1295 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1298 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1300 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1302 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1304 // If the global reference is actually to a non-lazy-pointer, we have to do an
1305 // extra load to get the address of the global.
1306 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1307 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1308 false, false, false, 0);
1312 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1313 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1314 DebugLoc dl = Op.getDebugLoc();
1316 // If we're comparing for equality to zero, expose the fact that this is
1317 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1318 // fold the new nodes.
1319 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1320 if (C->isNullValue() && CC == ISD::SETEQ) {
1321 EVT VT = Op.getOperand(0).getValueType();
1322 SDValue Zext = Op.getOperand(0);
1323 if (VT.bitsLT(MVT::i32)) {
1325 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1327 unsigned Log2b = Log2_32(VT.getSizeInBits());
1328 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1329 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1330 DAG.getConstant(Log2b, MVT::i32));
1331 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1333 // Leave comparisons against 0 and -1 alone for now, since they're usually
1334 // optimized. FIXME: revisit this when we can custom lower all setcc
1336 if (C->isAllOnesValue() || C->isNullValue())
1340 // If we have an integer seteq/setne, turn it into a compare against zero
1341 // by xor'ing the rhs with the lhs, which is faster than setting a
1342 // condition register, reading it back out, and masking the correct bit. The
1343 // normal approach here uses sub to do this instead of xor. Using xor exposes
1344 // the result to other bit-twiddling opportunities.
1345 EVT LHSVT = Op.getOperand(0).getValueType();
1346 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1347 EVT VT = Op.getValueType();
1348 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1350 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1355 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1356 const PPCSubtarget &Subtarget) const {
1357 SDNode *Node = Op.getNode();
1358 EVT VT = Node->getValueType(0);
1359 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1360 SDValue InChain = Node->getOperand(0);
1361 SDValue VAListPtr = Node->getOperand(1);
1362 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1363 DebugLoc dl = Node->getDebugLoc();
1365 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1368 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1369 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1371 InChain = GprIndex.getValue(1);
1373 if (VT == MVT::i64) {
1374 // Check if GprIndex is even
1375 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1376 DAG.getConstant(1, MVT::i32));
1377 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1378 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1379 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1380 DAG.getConstant(1, MVT::i32));
1381 // Align GprIndex to be even if it isn't
1382 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1386 // fpr index is 1 byte after gpr
1387 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1388 DAG.getConstant(1, MVT::i32));
1391 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1392 FprPtr, MachinePointerInfo(SV), MVT::i8,
1394 InChain = FprIndex.getValue(1);
1396 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1397 DAG.getConstant(8, MVT::i32));
1399 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1400 DAG.getConstant(4, MVT::i32));
1403 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1404 MachinePointerInfo(), false, false,
1406 InChain = OverflowArea.getValue(1);
1408 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1409 MachinePointerInfo(), false, false,
1411 InChain = RegSaveArea.getValue(1);
1413 // select overflow_area if index > 8
1414 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1415 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1417 // adjustment constant gpr_index * 4/8
1418 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1419 VT.isInteger() ? GprIndex : FprIndex,
1420 DAG.getConstant(VT.isInteger() ? 4 : 8,
1423 // OurReg = RegSaveArea + RegConstant
1424 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1427 // Floating types are 32 bytes into RegSaveArea
1428 if (VT.isFloatingPoint())
1429 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1430 DAG.getConstant(32, MVT::i32));
1432 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1433 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1434 VT.isInteger() ? GprIndex : FprIndex,
1435 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1438 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1439 VT.isInteger() ? VAListPtr : FprPtr,
1440 MachinePointerInfo(SV),
1441 MVT::i8, false, false, 0);
1443 // determine if we should load from reg_save_area or overflow_area
1444 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1446 // increase overflow_area by 4/8 if gpr/fpr > 8
1447 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1448 DAG.getConstant(VT.isInteger() ? 4 : 8,
1451 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1454 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1456 MachinePointerInfo(),
1457 MVT::i32, false, false, 0);
1459 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1460 false, false, false, 0);
1463 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1464 SelectionDAG &DAG) const {
1465 return Op.getOperand(0);
1468 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1469 SelectionDAG &DAG) const {
1470 SDValue Chain = Op.getOperand(0);
1471 SDValue Trmp = Op.getOperand(1); // trampoline
1472 SDValue FPtr = Op.getOperand(2); // nested function
1473 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1474 DebugLoc dl = Op.getDebugLoc();
1476 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1477 bool isPPC64 = (PtrVT == MVT::i64);
1479 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1482 TargetLowering::ArgListTy Args;
1483 TargetLowering::ArgListEntry Entry;
1485 Entry.Ty = IntPtrTy;
1486 Entry.Node = Trmp; Args.push_back(Entry);
1488 // TrampSize == (isPPC64 ? 48 : 40);
1489 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1490 isPPC64 ? MVT::i64 : MVT::i32);
1491 Args.push_back(Entry);
1493 Entry.Node = FPtr; Args.push_back(Entry);
1494 Entry.Node = Nest; Args.push_back(Entry);
1496 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1497 TargetLowering::CallLoweringInfo CLI(Chain,
1498 Type::getVoidTy(*DAG.getContext()),
1499 false, false, false, false, 0,
1501 /*isTailCall=*/false,
1502 /*doesNotRet=*/false,
1503 /*isReturnValueUsed=*/true,
1504 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1506 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1508 return CallResult.second;
1511 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1512 const PPCSubtarget &Subtarget) const {
1513 MachineFunction &MF = DAG.getMachineFunction();
1514 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1516 DebugLoc dl = Op.getDebugLoc();
1518 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1519 // vastart just stores the address of the VarArgsFrameIndex slot into the
1520 // memory location argument.
1521 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1522 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1523 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1524 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1525 MachinePointerInfo(SV),
1529 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1530 // We suppose the given va_list is already allocated.
1533 // char gpr; /* index into the array of 8 GPRs
1534 // * stored in the register save area
1535 // * gpr=0 corresponds to r3,
1536 // * gpr=1 to r4, etc.
1538 // char fpr; /* index into the array of 8 FPRs
1539 // * stored in the register save area
1540 // * fpr=0 corresponds to f1,
1541 // * fpr=1 to f2, etc.
1543 // char *overflow_arg_area;
1544 // /* location on stack that holds
1545 // * the next overflow argument
1547 // char *reg_save_area;
1548 // /* where r3:r10 and f1:f8 (if saved)
1554 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1555 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1558 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1560 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1562 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1565 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1566 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1568 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1569 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1571 uint64_t FPROffset = 1;
1572 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1574 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1576 // Store first byte : number of int regs
1577 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1579 MachinePointerInfo(SV),
1580 MVT::i8, false, false, 0);
1581 uint64_t nextOffset = FPROffset;
1582 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1585 // Store second byte : number of float regs
1586 SDValue secondStore =
1587 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1588 MachinePointerInfo(SV, nextOffset), MVT::i8,
1590 nextOffset += StackOffset;
1591 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1593 // Store second word : arguments given on stack
1594 SDValue thirdStore =
1595 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1596 MachinePointerInfo(SV, nextOffset),
1598 nextOffset += FrameOffset;
1599 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1601 // Store third word : arguments given in registers
1602 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1603 MachinePointerInfo(SV, nextOffset),
1608 #include "PPCGenCallingConv.inc"
1610 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1611 CCValAssign::LocInfo &LocInfo,
1612 ISD::ArgFlagsTy &ArgFlags,
1617 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1619 CCValAssign::LocInfo &LocInfo,
1620 ISD::ArgFlagsTy &ArgFlags,
1622 static const uint16_t ArgRegs[] = {
1623 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1624 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1626 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1628 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1630 // Skip one register if the first unallocated register has an even register
1631 // number and there are still argument registers available which have not been
1632 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1633 // need to skip a register if RegNum is odd.
1634 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1635 State.AllocateReg(ArgRegs[RegNum]);
1638 // Always return false here, as this function only makes sure that the first
1639 // unallocated register has an odd register number and does not actually
1640 // allocate a register for the current argument.
1644 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1646 CCValAssign::LocInfo &LocInfo,
1647 ISD::ArgFlagsTy &ArgFlags,
1649 static const uint16_t ArgRegs[] = {
1650 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1654 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1656 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1658 // If there is only one Floating-point register left we need to put both f64
1659 // values of a split ppc_fp128 value on the stack.
1660 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1661 State.AllocateReg(ArgRegs[RegNum]);
1664 // Always return false here, as this function only makes sure that the two f64
1665 // values a ppc_fp128 value is split into are both passed in registers or both
1666 // passed on the stack and does not actually allocate a register for the
1667 // current argument.
1671 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1673 static const uint16_t *GetFPR() {
1674 static const uint16_t FPR[] = {
1675 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1676 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1682 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1684 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1685 unsigned PtrByteSize) {
1686 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1687 if (Flags.isByVal())
1688 ArgSize = Flags.getByValSize();
1689 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1695 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1696 CallingConv::ID CallConv, bool isVarArg,
1697 const SmallVectorImpl<ISD::InputArg>
1699 DebugLoc dl, SelectionDAG &DAG,
1700 SmallVectorImpl<SDValue> &InVals)
1702 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1703 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1706 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1712 PPCTargetLowering::LowerFormalArguments_SVR4(
1714 CallingConv::ID CallConv, bool isVarArg,
1715 const SmallVectorImpl<ISD::InputArg>
1717 DebugLoc dl, SelectionDAG &DAG,
1718 SmallVectorImpl<SDValue> &InVals) const {
1720 // 32-bit SVR4 ABI Stack Frame Layout:
1721 // +-----------------------------------+
1722 // +--> | Back chain |
1723 // | +-----------------------------------+
1724 // | | Floating-point register save area |
1725 // | +-----------------------------------+
1726 // | | General register save area |
1727 // | +-----------------------------------+
1728 // | | CR save word |
1729 // | +-----------------------------------+
1730 // | | VRSAVE save word |
1731 // | +-----------------------------------+
1732 // | | Alignment padding |
1733 // | +-----------------------------------+
1734 // | | Vector register save area |
1735 // | +-----------------------------------+
1736 // | | Local variable space |
1737 // | +-----------------------------------+
1738 // | | Parameter list area |
1739 // | +-----------------------------------+
1740 // | | LR save word |
1741 // | +-----------------------------------+
1742 // SP--> +--- | Back chain |
1743 // +-----------------------------------+
1746 // System V Application Binary Interface PowerPC Processor Supplement
1747 // AltiVec Technology Programming Interface Manual
1749 MachineFunction &MF = DAG.getMachineFunction();
1750 MachineFrameInfo *MFI = MF.getFrameInfo();
1751 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1753 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1754 // Potential tail calls could cause overwriting of argument stack slots.
1755 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1756 (CallConv == CallingConv::Fast));
1757 unsigned PtrByteSize = 4;
1759 // Assign locations to all of the incoming arguments.
1760 SmallVector<CCValAssign, 16> ArgLocs;
1761 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1762 getTargetMachine(), ArgLocs, *DAG.getContext());
1764 // Reserve space for the linkage area on the stack.
1765 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1767 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1769 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1770 CCValAssign &VA = ArgLocs[i];
1772 // Arguments stored in registers.
1773 if (VA.isRegLoc()) {
1774 const TargetRegisterClass *RC;
1775 EVT ValVT = VA.getValVT();
1777 switch (ValVT.getSimpleVT().SimpleTy) {
1779 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1781 RC = &PPC::GPRCRegClass;
1784 RC = &PPC::F4RCRegClass;
1787 RC = &PPC::F8RCRegClass;
1793 RC = &PPC::VRRCRegClass;
1797 // Transform the arguments stored in physical registers into virtual ones.
1798 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1799 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1801 InVals.push_back(ArgValue);
1803 // Argument stored in memory.
1804 assert(VA.isMemLoc());
1806 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1807 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1810 // Create load nodes to retrieve arguments from the stack.
1811 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1812 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1813 MachinePointerInfo(),
1814 false, false, false, 0));
1818 // Assign locations to all of the incoming aggregate by value arguments.
1819 // Aggregates passed by value are stored in the local variable space of the
1820 // caller's stack frame, right above the parameter list area.
1821 SmallVector<CCValAssign, 16> ByValArgLocs;
1822 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1823 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1825 // Reserve stack space for the allocations in CCInfo.
1826 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1828 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1830 // Area that is at least reserved in the caller of this function.
1831 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1833 // Set the size that is at least reserved in caller of this function. Tail
1834 // call optimized function's reserved stack space needs to be aligned so that
1835 // taking the difference between two stack areas will result in an aligned
1837 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1840 std::max(MinReservedArea,
1841 PPCFrameLowering::getMinCallFrameSize(false, false));
1843 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1844 getStackAlignment();
1845 unsigned AlignMask = TargetAlign-1;
1846 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1848 FI->setMinReservedArea(MinReservedArea);
1850 SmallVector<SDValue, 8> MemOps;
1852 // If the function takes variable number of arguments, make a frame index for
1853 // the start of the first vararg value... for expansion of llvm.va_start.
1855 static const uint16_t GPArgRegs[] = {
1856 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1857 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1859 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1861 static const uint16_t FPArgRegs[] = {
1862 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1865 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1867 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1869 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1872 // Make room for NumGPArgRegs and NumFPArgRegs.
1873 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1874 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1876 FuncInfo->setVarArgsStackOffset(
1877 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1878 CCInfo.getNextStackOffset(), true));
1880 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1881 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1883 // The fixed integer arguments of a variadic function are stored to the
1884 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1885 // the result of va_next.
1886 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1887 // Get an existing live-in vreg, or add a new one.
1888 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1890 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1892 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1893 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1894 MachinePointerInfo(), false, false, 0);
1895 MemOps.push_back(Store);
1896 // Increment the address by four for the next argument to store
1897 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1898 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1901 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1903 // The double arguments are stored to the VarArgsFrameIndex
1905 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1906 // Get an existing live-in vreg, or add a new one.
1907 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1909 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1911 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1912 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1913 MachinePointerInfo(), false, false, 0);
1914 MemOps.push_back(Store);
1915 // Increment the address by eight for the next argument to store
1916 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1918 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1922 if (!MemOps.empty())
1923 Chain = DAG.getNode(ISD::TokenFactor, dl,
1924 MVT::Other, &MemOps[0], MemOps.size());
1930 PPCTargetLowering::LowerFormalArguments_Darwin(
1932 CallingConv::ID CallConv, bool isVarArg,
1933 const SmallVectorImpl<ISD::InputArg>
1935 DebugLoc dl, SelectionDAG &DAG,
1936 SmallVectorImpl<SDValue> &InVals) const {
1937 // TODO: add description of PPC stack frame format, or at least some docs.
1939 MachineFunction &MF = DAG.getMachineFunction();
1940 MachineFrameInfo *MFI = MF.getFrameInfo();
1941 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1943 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1944 bool isPPC64 = PtrVT == MVT::i64;
1945 // Potential tail calls could cause overwriting of argument stack slots.
1946 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1947 (CallConv == CallingConv::Fast));
1948 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1950 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1951 // Area that is at least reserved in caller of this function.
1952 unsigned MinReservedArea = ArgOffset;
1954 static const uint16_t GPR_32[] = { // 32-bit registers.
1955 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1956 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1958 static const uint16_t GPR_64[] = { // 64-bit registers.
1959 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1960 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1963 static const uint16_t *FPR = GetFPR();
1965 static const uint16_t VR[] = {
1966 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1967 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1970 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1971 const unsigned Num_FPR_Regs = 13;
1972 const unsigned Num_VR_Regs = array_lengthof( VR);
1974 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1976 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
1978 // In 32-bit non-varargs functions, the stack space for vectors is after the
1979 // stack space for non-vectors. We do not use this space unless we have
1980 // too many vectors to fit in registers, something that only occurs in
1981 // constructed examples:), but we have to walk the arglist to figure
1982 // that out...for the pathological case, compute VecArgOffset as the
1983 // start of the vector parameter area. Computing VecArgOffset is the
1984 // entire point of the following loop.
1985 unsigned VecArgOffset = ArgOffset;
1986 if (!isVarArg && !isPPC64) {
1987 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1989 EVT ObjectVT = Ins[ArgNo].VT;
1990 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1992 if (Flags.isByVal()) {
1993 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1994 unsigned ObjSize = Flags.getByValSize();
1996 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1997 VecArgOffset += ArgSize;
2001 switch(ObjectVT.getSimpleVT().SimpleTy) {
2002 default: llvm_unreachable("Unhandled argument type!");
2005 VecArgOffset += isPPC64 ? 8 : 4;
2007 case MVT::i64: // PPC64
2015 // Nothing to do, we're only looking at Nonvector args here.
2020 // We've found where the vector parameter area in memory is. Skip the
2021 // first 12 parameters; these don't use that memory.
2022 VecArgOffset = ((VecArgOffset+15)/16)*16;
2023 VecArgOffset += 12*16;
2025 // Add DAG nodes to load the arguments or copy them out of registers. On
2026 // entry to a function on PPC, the arguments start after the linkage area,
2027 // although the first ones are often in registers.
2029 SmallVector<SDValue, 8> MemOps;
2030 unsigned nAltivecParamsAtEnd = 0;
2031 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2033 bool needsLoad = false;
2034 EVT ObjectVT = Ins[ArgNo].VT;
2035 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2036 unsigned ArgSize = ObjSize;
2037 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2039 unsigned CurArgOffset = ArgOffset;
2041 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2042 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2043 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2044 if (isVarArg || isPPC64) {
2045 MinReservedArea = ((MinReservedArea+15)/16)*16;
2046 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2049 } else nAltivecParamsAtEnd++;
2051 // Calculate min reserved area.
2052 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2056 // FIXME the codegen can be much improved in some cases.
2057 // We do not have to keep everything in memory.
2058 if (Flags.isByVal()) {
2059 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2060 ObjSize = Flags.getByValSize();
2061 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2062 // Objects of size 1 and 2 are right justified, everything else is
2063 // left justified. This means the memory address is adjusted forwards.
2064 if (ObjSize==1 || ObjSize==2) {
2065 CurArgOffset = CurArgOffset + (4 - ObjSize);
2067 // The value of the object is its address.
2068 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2069 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2070 InVals.push_back(FIN);
2071 if (ObjSize==1 || ObjSize==2) {
2072 if (GPR_idx != Num_GPR_Regs) {
2075 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2077 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2078 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2079 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2080 MachinePointerInfo(),
2081 ObjSize==1 ? MVT::i8 : MVT::i16,
2083 MemOps.push_back(Store);
2087 ArgOffset += PtrByteSize;
2091 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2092 // Store whatever pieces of the object are in registers
2093 // to memory. ArgVal will be address of the beginning of
2095 if (GPR_idx != Num_GPR_Regs) {
2098 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2100 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2101 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2102 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2103 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2104 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2105 MachinePointerInfo(),
2107 MemOps.push_back(Store);
2109 ArgOffset += PtrByteSize;
2111 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2118 switch (ObjectVT.getSimpleVT().SimpleTy) {
2119 default: llvm_unreachable("Unhandled argument type!");
2122 if (GPR_idx != Num_GPR_Regs) {
2123 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2124 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2128 ArgSize = PtrByteSize;
2130 // All int arguments reserve stack space in the Darwin ABI.
2131 ArgOffset += PtrByteSize;
2135 case MVT::i64: // PPC64
2136 if (GPR_idx != Num_GPR_Regs) {
2137 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2138 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2140 if (ObjectVT == MVT::i32) {
2141 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2142 // value to MVT::i64 and then truncate to the correct register size.
2144 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2145 DAG.getValueType(ObjectVT));
2146 else if (Flags.isZExt())
2147 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2148 DAG.getValueType(ObjectVT));
2150 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2156 ArgSize = PtrByteSize;
2158 // All int arguments reserve stack space in the Darwin ABI.
2164 // Every 4 bytes of argument space consumes one of the GPRs available for
2165 // argument passing.
2166 if (GPR_idx != Num_GPR_Regs) {
2168 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2171 if (FPR_idx != Num_FPR_Regs) {
2174 if (ObjectVT == MVT::f32)
2175 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2177 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2179 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2185 // All FP arguments reserve stack space in the Darwin ABI.
2186 ArgOffset += isPPC64 ? 8 : ObjSize;
2192 // Note that vector arguments in registers don't reserve stack space,
2193 // except in varargs functions.
2194 if (VR_idx != Num_VR_Regs) {
2195 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2196 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2198 while ((ArgOffset % 16) != 0) {
2199 ArgOffset += PtrByteSize;
2200 if (GPR_idx != Num_GPR_Regs)
2204 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2208 if (!isVarArg && !isPPC64) {
2209 // Vectors go after all the nonvectors.
2210 CurArgOffset = VecArgOffset;
2213 // Vectors are aligned.
2214 ArgOffset = ((ArgOffset+15)/16)*16;
2215 CurArgOffset = ArgOffset;
2223 // We need to load the argument to a virtual register if we determined above
2224 // that we ran out of physical registers of the appropriate type.
2226 int FI = MFI->CreateFixedObject(ObjSize,
2227 CurArgOffset + (ArgSize - ObjSize),
2229 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2230 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2231 false, false, false, 0);
2234 InVals.push_back(ArgVal);
2237 // Set the size that is at least reserved in caller of this function. Tail
2238 // call optimized function's reserved stack space needs to be aligned so that
2239 // taking the difference between two stack areas will result in an aligned
2241 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2242 // Add the Altivec parameters at the end, if needed.
2243 if (nAltivecParamsAtEnd) {
2244 MinReservedArea = ((MinReservedArea+15)/16)*16;
2245 MinReservedArea += 16*nAltivecParamsAtEnd;
2248 std::max(MinReservedArea,
2249 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2250 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2251 getStackAlignment();
2252 unsigned AlignMask = TargetAlign-1;
2253 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2254 FI->setMinReservedArea(MinReservedArea);
2256 // If the function takes variable number of arguments, make a frame index for
2257 // the start of the first vararg value... for expansion of llvm.va_start.
2259 int Depth = ArgOffset;
2261 FuncInfo->setVarArgsFrameIndex(
2262 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2264 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2266 // If this function is vararg, store any remaining integer argument regs
2267 // to their spots on the stack so that they may be loaded by deferencing the
2268 // result of va_next.
2269 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2273 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2275 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2277 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2278 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2279 MachinePointerInfo(), false, false, 0);
2280 MemOps.push_back(Store);
2281 // Increment the address by four for the next argument to store
2282 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2283 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2287 if (!MemOps.empty())
2288 Chain = DAG.getNode(ISD::TokenFactor, dl,
2289 MVT::Other, &MemOps[0], MemOps.size());
2294 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2295 /// linkage area for the Darwin ABI.
2297 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2301 const SmallVectorImpl<ISD::OutputArg>
2303 const SmallVectorImpl<SDValue> &OutVals,
2304 unsigned &nAltivecParamsAtEnd) {
2305 // Count how many bytes are to be pushed on the stack, including the linkage
2306 // area, and parameter passing area. We start with 24/48 bytes, which is
2307 // prereserved space for [SP][CR][LR][3 x unused].
2308 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2309 unsigned NumOps = Outs.size();
2310 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2312 // Add up all the space actually used.
2313 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2314 // they all go in registers, but we must reserve stack space for them for
2315 // possible use by the caller. In varargs or 64-bit calls, parameters are
2316 // assigned stack space in order, with padding so Altivec parameters are
2318 nAltivecParamsAtEnd = 0;
2319 for (unsigned i = 0; i != NumOps; ++i) {
2320 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2321 EVT ArgVT = Outs[i].VT;
2322 // Varargs Altivec parameters are padded to a 16 byte boundary.
2323 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2324 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2325 if (!isVarArg && !isPPC64) {
2326 // Non-varargs Altivec parameters go after all the non-Altivec
2327 // parameters; handle those later so we know how much padding we need.
2328 nAltivecParamsAtEnd++;
2331 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2332 NumBytes = ((NumBytes+15)/16)*16;
2334 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2337 // Allow for Altivec parameters at the end, if needed.
2338 if (nAltivecParamsAtEnd) {
2339 NumBytes = ((NumBytes+15)/16)*16;
2340 NumBytes += 16*nAltivecParamsAtEnd;
2343 // The prolog code of the callee may store up to 8 GPR argument registers to
2344 // the stack, allowing va_start to index over them in memory if its varargs.
2345 // Because we cannot tell if this is needed on the caller side, we have to
2346 // conservatively assume that it is needed. As such, make sure we have at
2347 // least enough stack space for the caller to store the 8 GPRs.
2348 NumBytes = std::max(NumBytes,
2349 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2351 // Tail call needs the stack to be aligned.
2352 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2353 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2354 getFrameLowering()->getStackAlignment();
2355 unsigned AlignMask = TargetAlign-1;
2356 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2362 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2363 /// adjusted to accommodate the arguments for the tailcall.
2364 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2365 unsigned ParamSize) {
2367 if (!isTailCall) return 0;
2369 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2370 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2371 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2372 // Remember only if the new adjustement is bigger.
2373 if (SPDiff < FI->getTailCallSPDelta())
2374 FI->setTailCallSPDelta(SPDiff);
2379 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2380 /// for tail call optimization. Targets which want to do tail call
2381 /// optimization should implement this function.
2383 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2384 CallingConv::ID CalleeCC,
2386 const SmallVectorImpl<ISD::InputArg> &Ins,
2387 SelectionDAG& DAG) const {
2388 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2391 // Variable argument functions are not supported.
2395 MachineFunction &MF = DAG.getMachineFunction();
2396 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2397 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2398 // Functions containing by val parameters are not supported.
2399 for (unsigned i = 0; i != Ins.size(); i++) {
2400 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2401 if (Flags.isByVal()) return false;
2404 // Non PIC/GOT tail calls are supported.
2405 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2408 // At the moment we can only do local tail calls (in same module, hidden
2409 // or protected) if we are generating PIC.
2410 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2411 return G->getGlobal()->hasHiddenVisibility()
2412 || G->getGlobal()->hasProtectedVisibility();
2418 /// isCallCompatibleAddress - Return the immediate to use if the specified
2419 /// 32-bit value is representable in the immediate field of a BxA instruction.
2420 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2421 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2424 int Addr = C->getZExtValue();
2425 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2426 SignExtend32<26>(Addr) != Addr)
2427 return 0; // Top 6 bits have to be sext of immediate.
2429 return DAG.getConstant((int)C->getZExtValue() >> 2,
2430 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2435 struct TailCallArgumentInfo {
2440 TailCallArgumentInfo() : FrameIdx(0) {}
2445 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2447 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2449 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2450 SmallVector<SDValue, 8> &MemOpChains,
2452 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2453 SDValue Arg = TailCallArgs[i].Arg;
2454 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2455 int FI = TailCallArgs[i].FrameIdx;
2456 // Store relative to framepointer.
2457 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2458 MachinePointerInfo::getFixedStack(FI),
2463 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2464 /// the appropriate stack slot for the tail call optimized function call.
2465 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2466 MachineFunction &MF,
2475 // Calculate the new stack slot for the return address.
2476 int SlotSize = isPPC64 ? 8 : 4;
2477 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2479 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2480 NewRetAddrLoc, true);
2481 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2482 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2483 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2484 MachinePointerInfo::getFixedStack(NewRetAddr),
2487 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2488 // slot as the FP is never overwritten.
2491 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2492 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2494 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2495 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2496 MachinePointerInfo::getFixedStack(NewFPIdx),
2503 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2504 /// the position of the argument.
2506 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2507 SDValue Arg, int SPDiff, unsigned ArgOffset,
2508 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2509 int Offset = ArgOffset + SPDiff;
2510 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2511 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2512 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2513 SDValue FIN = DAG.getFrameIndex(FI, VT);
2514 TailCallArgumentInfo Info;
2516 Info.FrameIdxOp = FIN;
2518 TailCallArguments.push_back(Info);
2521 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2522 /// stack slot. Returns the chain as result and the loaded frame pointers in
2523 /// LROpOut/FPOpout. Used when tail calling.
2524 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2530 DebugLoc dl) const {
2532 // Load the LR and FP stack slot for later adjusting.
2533 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2534 LROpOut = getReturnAddrFrameIndex(DAG);
2535 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2536 false, false, false, 0);
2537 Chain = SDValue(LROpOut.getNode(), 1);
2539 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2540 // slot as the FP is never overwritten.
2542 FPOpOut = getFramePointerFrameIndex(DAG);
2543 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2544 false, false, false, 0);
2545 Chain = SDValue(FPOpOut.getNode(), 1);
2551 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2552 /// by "Src" to address "Dst" of size "Size". Alignment information is
2553 /// specified by the specific parameter attribute. The copy will be passed as
2554 /// a byval function parameter.
2555 /// Sometimes what we are copying is the end of a larger object, the part that
2556 /// does not fit in registers.
2558 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2559 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2561 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2562 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2563 false, false, MachinePointerInfo(0),
2564 MachinePointerInfo(0));
2567 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2570 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2571 SDValue Arg, SDValue PtrOff, int SPDiff,
2572 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2573 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2574 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2576 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2581 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2583 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2584 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2585 DAG.getConstant(ArgOffset, PtrVT));
2587 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2588 MachinePointerInfo(), false, false, 0));
2589 // Calculate and remember argument location.
2590 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2595 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2596 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2597 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2598 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2599 MachineFunction &MF = DAG.getMachineFunction();
2601 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2602 // might overwrite each other in case of tail call optimization.
2603 SmallVector<SDValue, 8> MemOpChains2;
2604 // Do not flag preceding copytoreg stuff together with the following stuff.
2606 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2608 if (!MemOpChains2.empty())
2609 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2610 &MemOpChains2[0], MemOpChains2.size());
2612 // Store the return address to the appropriate stack slot.
2613 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2614 isPPC64, isDarwinABI, dl);
2616 // Emit callseq_end just before tailcall node.
2617 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2618 DAG.getIntPtrConstant(0, true), InFlag);
2619 InFlag = Chain.getValue(1);
2623 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2624 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2625 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2626 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2627 const PPCSubtarget &PPCSubTarget) {
2629 bool isPPC64 = PPCSubTarget.isPPC64();
2630 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2632 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2633 NodeTys.push_back(MVT::Other); // Returns a chain
2634 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2636 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2638 bool needIndirectCall = true;
2639 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2640 // If this is an absolute destination address, use the munged value.
2641 Callee = SDValue(Dest, 0);
2642 needIndirectCall = false;
2645 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2646 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2647 // Use indirect calls for ALL functions calls in JIT mode, since the
2648 // far-call stubs may be outside relocation limits for a BL instruction.
2649 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2650 unsigned OpFlags = 0;
2651 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2652 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2653 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2654 (G->getGlobal()->isDeclaration() ||
2655 G->getGlobal()->isWeakForLinker())) {
2656 // PC-relative references to external symbols should go through $stub,
2657 // unless we're building with the leopard linker or later, which
2658 // automatically synthesizes these stubs.
2659 OpFlags = PPCII::MO_DARWIN_STUB;
2662 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2663 // every direct call is) turn it into a TargetGlobalAddress /
2664 // TargetExternalSymbol node so that legalize doesn't hack it.
2665 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2666 Callee.getValueType(),
2668 needIndirectCall = false;
2672 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2673 unsigned char OpFlags = 0;
2675 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2676 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2677 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2678 // PC-relative references to external symbols should go through $stub,
2679 // unless we're building with the leopard linker or later, which
2680 // automatically synthesizes these stubs.
2681 OpFlags = PPCII::MO_DARWIN_STUB;
2684 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2686 needIndirectCall = false;
2689 if (needIndirectCall) {
2690 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2691 // to do the call, we can't use PPCISD::CALL.
2692 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2694 if (isSVR4ABI && isPPC64) {
2695 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2696 // entry point, but to the function descriptor (the function entry point
2697 // address is part of the function descriptor though).
2698 // The function descriptor is a three doubleword structure with the
2699 // following fields: function entry point, TOC base address and
2700 // environment pointer.
2701 // Thus for a call through a function pointer, the following actions need
2703 // 1. Save the TOC of the caller in the TOC save area of its stack
2704 // frame (this is done in LowerCall_Darwin()).
2705 // 2. Load the address of the function entry point from the function
2707 // 3. Load the TOC of the callee from the function descriptor into r2.
2708 // 4. Load the environment pointer from the function descriptor into
2710 // 5. Branch to the function entry point address.
2711 // 6. On return of the callee, the TOC of the caller needs to be
2712 // restored (this is done in FinishCall()).
2714 // All those operations are flagged together to ensure that no other
2715 // operations can be scheduled in between. E.g. without flagging the
2716 // operations together, a TOC access in the caller could be scheduled
2717 // between the load of the callee TOC and the branch to the callee, which
2718 // results in the TOC access going through the TOC of the callee instead
2719 // of going through the TOC of the caller, which leads to incorrect code.
2721 // Load the address of the function entry point from the function
2723 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2724 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2725 InFlag.getNode() ? 3 : 2);
2726 Chain = LoadFuncPtr.getValue(1);
2727 InFlag = LoadFuncPtr.getValue(2);
2729 // Load environment pointer into r11.
2730 // Offset of the environment pointer within the function descriptor.
2731 SDValue PtrOff = DAG.getIntPtrConstant(16);
2733 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2734 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2736 Chain = LoadEnvPtr.getValue(1);
2737 InFlag = LoadEnvPtr.getValue(2);
2739 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2741 Chain = EnvVal.getValue(0);
2742 InFlag = EnvVal.getValue(1);
2744 // Load TOC of the callee into r2. We are using a target-specific load
2745 // with r2 hard coded, because the result of a target-independent load
2746 // would never go directly into r2, since r2 is a reserved register (which
2747 // prevents the register allocator from allocating it), resulting in an
2748 // additional register being allocated and an unnecessary move instruction
2750 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2751 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2753 Chain = LoadTOCPtr.getValue(0);
2754 InFlag = LoadTOCPtr.getValue(1);
2756 MTCTROps[0] = Chain;
2757 MTCTROps[1] = LoadFuncPtr;
2758 MTCTROps[2] = InFlag;
2761 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2762 2 + (InFlag.getNode() != 0));
2763 InFlag = Chain.getValue(1);
2766 NodeTys.push_back(MVT::Other);
2767 NodeTys.push_back(MVT::Glue);
2768 Ops.push_back(Chain);
2769 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2771 // Add CTR register as callee so a bctr can be emitted later.
2773 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2776 // If this is a direct call, pass the chain and the callee.
2777 if (Callee.getNode()) {
2778 Ops.push_back(Chain);
2779 Ops.push_back(Callee);
2781 // If this is a tail call add stack pointer delta.
2783 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2785 // Add argument registers to the end of the list so that they are known live
2787 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2788 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2789 RegsToPass[i].second.getValueType()));
2795 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2796 CallingConv::ID CallConv, bool isVarArg,
2797 const SmallVectorImpl<ISD::InputArg> &Ins,
2798 DebugLoc dl, SelectionDAG &DAG,
2799 SmallVectorImpl<SDValue> &InVals) const {
2801 SmallVector<CCValAssign, 16> RVLocs;
2802 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2803 getTargetMachine(), RVLocs, *DAG.getContext());
2804 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2806 // Copy all of the result registers out of their specified physreg.
2807 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2808 CCValAssign &VA = RVLocs[i];
2809 EVT VT = VA.getValVT();
2810 assert(VA.isRegLoc() && "Can only return in registers!");
2811 Chain = DAG.getCopyFromReg(Chain, dl,
2812 VA.getLocReg(), VT, InFlag).getValue(1);
2813 InVals.push_back(Chain.getValue(0));
2814 InFlag = Chain.getValue(2);
2821 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2822 bool isTailCall, bool isVarArg,
2824 SmallVector<std::pair<unsigned, SDValue>, 8>
2826 SDValue InFlag, SDValue Chain,
2828 int SPDiff, unsigned NumBytes,
2829 const SmallVectorImpl<ISD::InputArg> &Ins,
2830 SmallVectorImpl<SDValue> &InVals) const {
2831 std::vector<EVT> NodeTys;
2832 SmallVector<SDValue, 8> Ops;
2833 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2834 isTailCall, RegsToPass, Ops, NodeTys,
2837 // When performing tail call optimization the callee pops its arguments off
2838 // the stack. Account for this here so these bytes can be pushed back on in
2839 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2840 int BytesCalleePops =
2841 (CallConv == CallingConv::Fast &&
2842 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
2844 // Add a register mask operand representing the call-preserved registers.
2845 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2846 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2847 assert(Mask && "Missing call preserved mask for calling convention");
2848 Ops.push_back(DAG.getRegisterMask(Mask));
2850 if (InFlag.getNode())
2851 Ops.push_back(InFlag);
2855 // If this is the first return lowered for this function, add the regs
2856 // to the liveout set for the function.
2857 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2858 SmallVector<CCValAssign, 16> RVLocs;
2859 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2860 getTargetMachine(), RVLocs, *DAG.getContext());
2861 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2862 for (unsigned i = 0; i != RVLocs.size(); ++i)
2863 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2866 assert(((Callee.getOpcode() == ISD::Register &&
2867 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2868 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2869 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2870 isa<ConstantSDNode>(Callee)) &&
2871 "Expecting an global address, external symbol, absolute value or register");
2873 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2876 // Add a NOP immediately after the branch instruction when using the 64-bit
2877 // SVR4 ABI. At link time, if caller and callee are in a different module and
2878 // thus have a different TOC, the call will be replaced with a call to a stub
2879 // function which saves the current TOC, loads the TOC of the callee and
2880 // branches to the callee. The NOP will be replaced with a load instruction
2881 // which restores the TOC of the caller from the TOC save slot of the current
2882 // stack frame. If caller and callee belong to the same module (and have the
2883 // same TOC), the NOP will remain unchanged.
2885 bool needsTOCRestore = false;
2886 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2887 if (CallOpc == PPCISD::BCTRL_SVR4) {
2888 // This is a call through a function pointer.
2889 // Restore the caller TOC from the save area into R2.
2890 // See PrepareCall() for more information about calls through function
2891 // pointers in the 64-bit SVR4 ABI.
2892 // We are using a target-specific load with r2 hard coded, because the
2893 // result of a target-independent load would never go directly into r2,
2894 // since r2 is a reserved register (which prevents the register allocator
2895 // from allocating it), resulting in an additional register being
2896 // allocated and an unnecessary move instruction being generated.
2897 needsTOCRestore = true;
2898 } else if (CallOpc == PPCISD::CALL_SVR4) {
2899 // Otherwise insert NOP.
2900 CallOpc = PPCISD::CALL_NOP_SVR4;
2904 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2905 InFlag = Chain.getValue(1);
2907 if (needsTOCRestore) {
2908 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2909 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2910 InFlag = Chain.getValue(1);
2913 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2914 DAG.getIntPtrConstant(BytesCalleePops, true),
2917 InFlag = Chain.getValue(1);
2919 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2920 Ins, dl, DAG, InVals);
2924 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2925 SmallVectorImpl<SDValue> &InVals) const {
2926 SelectionDAG &DAG = CLI.DAG;
2927 DebugLoc &dl = CLI.DL;
2928 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2929 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2930 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2931 SDValue Chain = CLI.Chain;
2932 SDValue Callee = CLI.Callee;
2933 bool &isTailCall = CLI.IsTailCall;
2934 CallingConv::ID CallConv = CLI.CallConv;
2935 bool isVarArg = CLI.IsVarArg;
2938 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2941 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2942 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2943 isTailCall, Outs, OutVals, Ins,
2946 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2947 isTailCall, Outs, OutVals, Ins,
2952 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2953 CallingConv::ID CallConv, bool isVarArg,
2955 const SmallVectorImpl<ISD::OutputArg> &Outs,
2956 const SmallVectorImpl<SDValue> &OutVals,
2957 const SmallVectorImpl<ISD::InputArg> &Ins,
2958 DebugLoc dl, SelectionDAG &DAG,
2959 SmallVectorImpl<SDValue> &InVals) const {
2960 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2961 // of the 32-bit SVR4 ABI stack frame layout.
2963 assert((CallConv == CallingConv::C ||
2964 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2966 unsigned PtrByteSize = 4;
2968 MachineFunction &MF = DAG.getMachineFunction();
2970 // Mark this function as potentially containing a function that contains a
2971 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2972 // and restoring the callers stack pointer in this functions epilog. This is
2973 // done because by tail calling the called function might overwrite the value
2974 // in this function's (MF) stack pointer stack slot 0(SP).
2975 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2976 CallConv == CallingConv::Fast)
2977 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2979 // Count how many bytes are to be pushed on the stack, including the linkage
2980 // area, parameter list area and the part of the local variable space which
2981 // contains copies of aggregates which are passed by value.
2983 // Assign locations to all of the outgoing arguments.
2984 SmallVector<CCValAssign, 16> ArgLocs;
2985 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2986 getTargetMachine(), ArgLocs, *DAG.getContext());
2988 // Reserve space for the linkage area on the stack.
2989 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2992 // Handle fixed and variable vector arguments differently.
2993 // Fixed vector arguments go into registers as long as registers are
2994 // available. Variable vector arguments always go into memory.
2995 unsigned NumArgs = Outs.size();
2997 for (unsigned i = 0; i != NumArgs; ++i) {
2998 MVT ArgVT = Outs[i].VT;
2999 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3002 if (Outs[i].IsFixed) {
3003 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3006 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3012 errs() << "Call operand #" << i << " has unhandled type "
3013 << EVT(ArgVT).getEVTString() << "\n";
3015 llvm_unreachable(0);
3019 // All arguments are treated the same.
3020 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3023 // Assign locations to all of the outgoing aggregate by value arguments.
3024 SmallVector<CCValAssign, 16> ByValArgLocs;
3025 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3026 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3028 // Reserve stack space for the allocations in CCInfo.
3029 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3031 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3033 // Size of the linkage area, parameter list area and the part of the local
3034 // space variable where copies of aggregates which are passed by value are
3036 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3038 // Calculate by how many bytes the stack has to be adjusted in case of tail
3039 // call optimization.
3040 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3042 // Adjust the stack pointer for the new arguments...
3043 // These operations are automatically eliminated by the prolog/epilog pass
3044 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3045 SDValue CallSeqStart = Chain;
3047 // Load the return address and frame pointer so it can be moved somewhere else
3050 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3053 // Set up a copy of the stack pointer for use loading and storing any
3054 // arguments that may not fit in the registers available for argument
3056 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3058 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3059 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3060 SmallVector<SDValue, 8> MemOpChains;
3062 bool seenFloatArg = false;
3063 // Walk the register/memloc assignments, inserting copies/loads.
3064 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3067 CCValAssign &VA = ArgLocs[i];
3068 SDValue Arg = OutVals[i];
3069 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3071 if (Flags.isByVal()) {
3072 // Argument is an aggregate which is passed by value, thus we need to
3073 // create a copy of it in the local variable space of the current stack
3074 // frame (which is the stack frame of the caller) and pass the address of
3075 // this copy to the callee.
3076 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3077 CCValAssign &ByValVA = ByValArgLocs[j++];
3078 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3080 // Memory reserved in the local variable space of the callers stack frame.
3081 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3083 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3084 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3086 // Create a copy of the argument in the local area of the current
3088 SDValue MemcpyCall =
3089 CreateCopyOfByValArgument(Arg, PtrOff,
3090 CallSeqStart.getNode()->getOperand(0),
3093 // This must go outside the CALLSEQ_START..END.
3094 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3095 CallSeqStart.getNode()->getOperand(1));
3096 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3097 NewCallSeqStart.getNode());
3098 Chain = CallSeqStart = NewCallSeqStart;
3100 // Pass the address of the aggregate copy on the stack either in a
3101 // physical register or in the parameter list area of the current stack
3102 // frame to the callee.
3106 if (VA.isRegLoc()) {
3107 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3108 // Put argument in a physical register.
3109 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3111 // Put argument in the parameter list area of the current stack frame.
3112 assert(VA.isMemLoc());
3113 unsigned LocMemOffset = VA.getLocMemOffset();
3116 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3117 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3119 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3120 MachinePointerInfo(),
3123 // Calculate and remember argument location.
3124 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3130 if (!MemOpChains.empty())
3131 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3132 &MemOpChains[0], MemOpChains.size());
3134 // Set CR6 to true if this is a vararg call with floating args passed in
3137 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3139 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3142 // Build a sequence of copy-to-reg nodes chained together with token chain
3143 // and flag operands which copy the outgoing args into the appropriate regs.
3145 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3146 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3147 RegsToPass[i].second, InFlag);
3148 InFlag = Chain.getValue(1);
3152 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3153 false, TailCallArguments);
3155 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3156 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3161 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3162 CallingConv::ID CallConv, bool isVarArg,
3164 const SmallVectorImpl<ISD::OutputArg> &Outs,
3165 const SmallVectorImpl<SDValue> &OutVals,
3166 const SmallVectorImpl<ISD::InputArg> &Ins,
3167 DebugLoc dl, SelectionDAG &DAG,
3168 SmallVectorImpl<SDValue> &InVals) const {
3170 unsigned NumOps = Outs.size();
3172 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3173 bool isPPC64 = PtrVT == MVT::i64;
3174 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3176 MachineFunction &MF = DAG.getMachineFunction();
3178 // Mark this function as potentially containing a function that contains a
3179 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3180 // and restoring the callers stack pointer in this functions epilog. This is
3181 // done because by tail calling the called function might overwrite the value
3182 // in this function's (MF) stack pointer stack slot 0(SP).
3183 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3184 CallConv == CallingConv::Fast)
3185 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3187 unsigned nAltivecParamsAtEnd = 0;
3189 // Count how many bytes are to be pushed on the stack, including the linkage
3190 // area, and parameter passing area. We start with 24/48 bytes, which is
3191 // prereserved space for [SP][CR][LR][3 x unused].
3193 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3195 nAltivecParamsAtEnd);
3197 // Calculate by how many bytes the stack has to be adjusted in case of tail
3198 // call optimization.
3199 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3201 // To protect arguments on the stack from being clobbered in a tail call,
3202 // force all the loads to happen before doing any other lowering.
3204 Chain = DAG.getStackArgumentTokenFactor(Chain);
3206 // Adjust the stack pointer for the new arguments...
3207 // These operations are automatically eliminated by the prolog/epilog pass
3208 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3209 SDValue CallSeqStart = Chain;
3211 // Load the return address and frame pointer so it can be move somewhere else
3214 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3217 // Set up a copy of the stack pointer for use loading and storing any
3218 // arguments that may not fit in the registers available for argument
3222 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3224 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3226 // Figure out which arguments are going to go in registers, and which in
3227 // memory. Also, if this is a vararg function, floating point operations
3228 // must be stored to our stack, and loaded into integer regs as well, if
3229 // any integer regs are available for argument passing.
3230 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3231 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3233 static const uint16_t GPR_32[] = { // 32-bit registers.
3234 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3235 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3237 static const uint16_t GPR_64[] = { // 64-bit registers.
3238 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3239 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3241 static const uint16_t *FPR = GetFPR();
3243 static const uint16_t VR[] = {
3244 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3245 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3247 const unsigned NumGPRs = array_lengthof(GPR_32);
3248 const unsigned NumFPRs = 13;
3249 const unsigned NumVRs = array_lengthof(VR);
3251 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3253 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3254 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3256 SmallVector<SDValue, 8> MemOpChains;
3257 for (unsigned i = 0; i != NumOps; ++i) {
3258 SDValue Arg = OutVals[i];
3259 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3261 // PtrOff will be used to store the current argument to the stack if a
3262 // register cannot be found for it.
3265 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3267 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3269 // On PPC64, promote integers to 64-bit values.
3270 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3271 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3272 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3273 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3276 // FIXME memcpy is used way more than necessary. Correctness first.
3277 if (Flags.isByVal()) {
3278 unsigned Size = Flags.getByValSize();
3279 if (Size==1 || Size==2) {
3280 // Very small objects are passed right-justified.
3281 // Everything else is passed left-justified.
3282 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3283 if (GPR_idx != NumGPRs) {
3284 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3285 MachinePointerInfo(), VT,
3287 MemOpChains.push_back(Load.getValue(1));
3288 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3290 ArgOffset += PtrByteSize;
3292 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3293 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3294 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3295 CallSeqStart.getNode()->getOperand(0),
3297 // This must go outside the CALLSEQ_START..END.
3298 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3299 CallSeqStart.getNode()->getOperand(1));
3300 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3301 NewCallSeqStart.getNode());
3302 Chain = CallSeqStart = NewCallSeqStart;
3303 ArgOffset += PtrByteSize;
3307 // Copy entire object into memory. There are cases where gcc-generated
3308 // code assumes it is there, even if it could be put entirely into
3309 // registers. (This is not what the doc says.)
3310 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3311 CallSeqStart.getNode()->getOperand(0),
3313 // This must go outside the CALLSEQ_START..END.
3314 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3315 CallSeqStart.getNode()->getOperand(1));
3316 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3317 Chain = CallSeqStart = NewCallSeqStart;
3318 // And copy the pieces of it that fit into registers.
3319 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3320 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3321 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3322 if (GPR_idx != NumGPRs) {
3323 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3324 MachinePointerInfo(),
3325 false, false, false, 0);
3326 MemOpChains.push_back(Load.getValue(1));
3327 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3328 ArgOffset += PtrByteSize;
3330 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3337 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3338 default: llvm_unreachable("Unexpected ValueType for argument!");
3341 if (GPR_idx != NumGPRs) {
3342 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3344 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3345 isPPC64, isTailCall, false, MemOpChains,
3346 TailCallArguments, dl);
3348 ArgOffset += PtrByteSize;
3352 if (FPR_idx != NumFPRs) {
3353 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3356 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3357 MachinePointerInfo(), false, false, 0);
3358 MemOpChains.push_back(Store);
3360 // Float varargs are always shadowed in available integer registers
3361 if (GPR_idx != NumGPRs) {
3362 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3363 MachinePointerInfo(), false, false,
3365 MemOpChains.push_back(Load.getValue(1));
3366 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3368 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3369 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3370 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3371 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3372 MachinePointerInfo(),
3373 false, false, false, 0);
3374 MemOpChains.push_back(Load.getValue(1));
3375 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3378 // If we have any FPRs remaining, we may also have GPRs remaining.
3379 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3381 if (GPR_idx != NumGPRs)
3383 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3384 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3388 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3389 isPPC64, isTailCall, false, MemOpChains,
3390 TailCallArguments, dl);
3395 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3402 // These go aligned on the stack, or in the corresponding R registers
3403 // when within range. The Darwin PPC ABI doc claims they also go in
3404 // V registers; in fact gcc does this only for arguments that are
3405 // prototyped, not for those that match the ... We do it for all
3406 // arguments, seems to work.
3407 while (ArgOffset % 16 !=0) {
3408 ArgOffset += PtrByteSize;
3409 if (GPR_idx != NumGPRs)
3412 // We could elide this store in the case where the object fits
3413 // entirely in R registers. Maybe later.
3414 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3415 DAG.getConstant(ArgOffset, PtrVT));
3416 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3417 MachinePointerInfo(), false, false, 0);
3418 MemOpChains.push_back(Store);
3419 if (VR_idx != NumVRs) {
3420 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3421 MachinePointerInfo(),
3422 false, false, false, 0);
3423 MemOpChains.push_back(Load.getValue(1));
3424 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3427 for (unsigned i=0; i<16; i+=PtrByteSize) {
3428 if (GPR_idx == NumGPRs)
3430 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3431 DAG.getConstant(i, PtrVT));
3432 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3433 false, false, false, 0);
3434 MemOpChains.push_back(Load.getValue(1));
3435 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3440 // Non-varargs Altivec params generally go in registers, but have
3441 // stack space allocated at the end.
3442 if (VR_idx != NumVRs) {
3443 // Doesn't have GPR space allocated.
3444 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3445 } else if (nAltivecParamsAtEnd==0) {
3446 // We are emitting Altivec params in order.
3447 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3448 isPPC64, isTailCall, true, MemOpChains,
3449 TailCallArguments, dl);
3455 // If all Altivec parameters fit in registers, as they usually do,
3456 // they get stack space following the non-Altivec parameters. We
3457 // don't track this here because nobody below needs it.
3458 // If there are more Altivec parameters than fit in registers emit
3460 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3462 // Offset is aligned; skip 1st 12 params which go in V registers.
3463 ArgOffset = ((ArgOffset+15)/16)*16;
3465 for (unsigned i = 0; i != NumOps; ++i) {
3466 SDValue Arg = OutVals[i];
3467 EVT ArgType = Outs[i].VT;
3468 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3469 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3472 // We are emitting Altivec params in order.
3473 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3474 isPPC64, isTailCall, true, MemOpChains,
3475 TailCallArguments, dl);
3482 if (!MemOpChains.empty())
3483 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3484 &MemOpChains[0], MemOpChains.size());
3486 // Check if this is an indirect call (MTCTR/BCTRL).
3487 // See PrepareCall() for more information about calls through function
3488 // pointers in the 64-bit SVR4 ABI.
3489 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3490 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3491 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3492 !isBLACompatibleAddress(Callee, DAG)) {
3493 // Load r2 into a virtual register and store it to the TOC save area.
3494 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3495 // TOC save area offset.
3496 SDValue PtrOff = DAG.getIntPtrConstant(40);
3497 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3498 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3502 // On Darwin, R12 must contain the address of an indirect callee. This does
3503 // not mean the MTCTR instruction must use R12; it's easier to model this as
3504 // an extra parameter, so do that.
3506 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3507 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3508 !isBLACompatibleAddress(Callee, DAG))
3509 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3510 PPC::R12), Callee));
3512 // Build a sequence of copy-to-reg nodes chained together with token chain
3513 // and flag operands which copy the outgoing args into the appropriate regs.
3515 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3516 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3517 RegsToPass[i].second, InFlag);
3518 InFlag = Chain.getValue(1);
3522 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3523 FPOp, true, TailCallArguments);
3525 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3526 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3531 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3532 MachineFunction &MF, bool isVarArg,
3533 const SmallVectorImpl<ISD::OutputArg> &Outs,
3534 LLVMContext &Context) const {
3535 SmallVector<CCValAssign, 16> RVLocs;
3536 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3538 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3542 PPCTargetLowering::LowerReturn(SDValue Chain,
3543 CallingConv::ID CallConv, bool isVarArg,
3544 const SmallVectorImpl<ISD::OutputArg> &Outs,
3545 const SmallVectorImpl<SDValue> &OutVals,
3546 DebugLoc dl, SelectionDAG &DAG) const {
3548 SmallVector<CCValAssign, 16> RVLocs;
3549 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3550 getTargetMachine(), RVLocs, *DAG.getContext());
3551 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3553 // If this is the first return lowered for this function, add the regs to the
3554 // liveout set for the function.
3555 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3556 for (unsigned i = 0; i != RVLocs.size(); ++i)
3557 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3562 // Copy the result values into the output registers.
3563 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3564 CCValAssign &VA = RVLocs[i];
3565 assert(VA.isRegLoc() && "Can only return in registers!");
3566 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3568 Flag = Chain.getValue(1);
3572 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3574 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3577 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3578 const PPCSubtarget &Subtarget) const {
3579 // When we pop the dynamic allocation we need to restore the SP link.
3580 DebugLoc dl = Op.getDebugLoc();
3582 // Get the corect type for pointers.
3583 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3585 // Construct the stack pointer operand.
3586 bool isPPC64 = Subtarget.isPPC64();
3587 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3588 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3590 // Get the operands for the STACKRESTORE.
3591 SDValue Chain = Op.getOperand(0);
3592 SDValue SaveSP = Op.getOperand(1);
3594 // Load the old link SP.
3595 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3596 MachinePointerInfo(),
3597 false, false, false, 0);
3599 // Restore the stack pointer.
3600 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3602 // Store the old link SP.
3603 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3610 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3611 MachineFunction &MF = DAG.getMachineFunction();
3612 bool isPPC64 = PPCSubTarget.isPPC64();
3613 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3614 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3616 // Get current frame pointer save index. The users of this index will be
3617 // primarily DYNALLOC instructions.
3618 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3619 int RASI = FI->getReturnAddrSaveIndex();
3621 // If the frame pointer save index hasn't been defined yet.
3623 // Find out what the fix offset of the frame pointer save area.
3624 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3625 // Allocate the frame index for frame pointer save area.
3626 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3628 FI->setReturnAddrSaveIndex(RASI);
3630 return DAG.getFrameIndex(RASI, PtrVT);
3634 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3635 MachineFunction &MF = DAG.getMachineFunction();
3636 bool isPPC64 = PPCSubTarget.isPPC64();
3637 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3638 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3640 // Get current frame pointer save index. The users of this index will be
3641 // primarily DYNALLOC instructions.
3642 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3643 int FPSI = FI->getFramePointerSaveIndex();
3645 // If the frame pointer save index hasn't been defined yet.
3647 // Find out what the fix offset of the frame pointer save area.
3648 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3651 // Allocate the frame index for frame pointer save area.
3652 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3654 FI->setFramePointerSaveIndex(FPSI);
3656 return DAG.getFrameIndex(FPSI, PtrVT);
3659 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3661 const PPCSubtarget &Subtarget) const {
3663 SDValue Chain = Op.getOperand(0);
3664 SDValue Size = Op.getOperand(1);
3665 DebugLoc dl = Op.getDebugLoc();
3667 // Get the corect type for pointers.
3668 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3670 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3671 DAG.getConstant(0, PtrVT), Size);
3672 // Construct a node for the frame pointer save index.
3673 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3674 // Build a DYNALLOC node.
3675 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3676 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3677 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3680 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3682 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3683 // Not FP? Not a fsel.
3684 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3685 !Op.getOperand(2).getValueType().isFloatingPoint())
3688 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3690 // Cannot handle SETEQ/SETNE.
3691 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3693 EVT ResVT = Op.getValueType();
3694 EVT CmpVT = Op.getOperand(0).getValueType();
3695 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3696 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3697 DebugLoc dl = Op.getDebugLoc();
3699 // If the RHS of the comparison is a 0.0, we don't need to do the
3700 // subtraction at all.
3701 if (isFloatingPointZero(RHS))
3703 default: break; // SETUO etc aren't handled by fsel.
3706 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3709 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3710 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3711 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3714 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3717 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3718 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3719 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3720 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3725 default: break; // SETUO etc aren't handled by fsel.
3728 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3729 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3730 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3731 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3734 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3735 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3736 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3737 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3740 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3741 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3742 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3743 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3746 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3747 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3748 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3749 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3754 // FIXME: Split this code up when LegalizeDAGTypes lands.
3755 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3756 DebugLoc dl) const {
3757 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3758 SDValue Src = Op.getOperand(0);
3759 if (Src.getValueType() == MVT::f32)
3760 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3763 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3764 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3766 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3771 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3775 // Convert the FP value to an int value through memory.
3776 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3778 // Emit a store to the stack slot.
3779 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3780 MachinePointerInfo(), false, false, 0);
3782 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3784 if (Op.getValueType() == MVT::i32)
3785 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3786 DAG.getConstant(4, FIPtr.getValueType()));
3787 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3788 false, false, false, 0);
3791 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3792 SelectionDAG &DAG) const {
3793 DebugLoc dl = Op.getDebugLoc();
3794 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3795 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3798 if (Op.getOperand(0).getValueType() == MVT::i64) {
3799 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3800 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3801 if (Op.getValueType() == MVT::f32)
3802 FP = DAG.getNode(ISD::FP_ROUND, dl,
3803 MVT::f32, FP, DAG.getIntPtrConstant(0));
3807 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3808 "Unhandled SINT_TO_FP type in custom expander!");
3809 // Since we only generate this in 64-bit mode, we can take advantage of
3810 // 64-bit registers. In particular, sign extend the input value into the
3811 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3812 // then lfd it and fcfid it.
3813 MachineFunction &MF = DAG.getMachineFunction();
3814 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3815 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3816 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3817 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3819 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3822 // STD the extended value into the stack slot.
3823 MachineMemOperand *MMO =
3824 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3825 MachineMemOperand::MOStore, 8, 8);
3826 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3828 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3829 Ops, 4, MVT::i64, MMO);
3830 // Load the value as a double.
3831 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3832 false, false, false, 0);
3834 // FCFID it and return it.
3835 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3836 if (Op.getValueType() == MVT::f32)
3837 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3841 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3842 SelectionDAG &DAG) const {
3843 DebugLoc dl = Op.getDebugLoc();
3845 The rounding mode is in bits 30:31 of FPSR, and has the following
3852 FLT_ROUNDS, on the other hand, expects the following:
3859 To perform the conversion, we do:
3860 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3863 MachineFunction &MF = DAG.getMachineFunction();
3864 EVT VT = Op.getValueType();
3865 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3866 std::vector<EVT> NodeTys;
3867 SDValue MFFSreg, InFlag;
3869 // Save FP Control Word to register
3870 NodeTys.push_back(MVT::f64); // return register
3871 NodeTys.push_back(MVT::Glue); // unused in this context
3872 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3874 // Save FP register to stack slot
3875 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3876 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3877 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3878 StackSlot, MachinePointerInfo(), false, false,0);
3880 // Load FP Control Word from low 32 bits of stack slot.
3881 SDValue Four = DAG.getConstant(4, PtrVT);
3882 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3883 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3884 false, false, false, 0);
3886 // Transform as necessary
3888 DAG.getNode(ISD::AND, dl, MVT::i32,
3889 CWD, DAG.getConstant(3, MVT::i32));
3891 DAG.getNode(ISD::SRL, dl, MVT::i32,
3892 DAG.getNode(ISD::AND, dl, MVT::i32,
3893 DAG.getNode(ISD::XOR, dl, MVT::i32,
3894 CWD, DAG.getConstant(3, MVT::i32)),
3895 DAG.getConstant(3, MVT::i32)),
3896 DAG.getConstant(1, MVT::i32));
3899 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3901 return DAG.getNode((VT.getSizeInBits() < 16 ?
3902 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3905 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3906 EVT VT = Op.getValueType();
3907 unsigned BitWidth = VT.getSizeInBits();
3908 DebugLoc dl = Op.getDebugLoc();
3909 assert(Op.getNumOperands() == 3 &&
3910 VT == Op.getOperand(1).getValueType() &&
3913 // Expand into a bunch of logical ops. Note that these ops
3914 // depend on the PPC behavior for oversized shift amounts.
3915 SDValue Lo = Op.getOperand(0);
3916 SDValue Hi = Op.getOperand(1);
3917 SDValue Amt = Op.getOperand(2);
3918 EVT AmtVT = Amt.getValueType();
3920 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3921 DAG.getConstant(BitWidth, AmtVT), Amt);
3922 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3923 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3924 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3925 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3926 DAG.getConstant(-BitWidth, AmtVT));
3927 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3928 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3929 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3930 SDValue OutOps[] = { OutLo, OutHi };
3931 return DAG.getMergeValues(OutOps, 2, dl);
3934 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3935 EVT VT = Op.getValueType();
3936 DebugLoc dl = Op.getDebugLoc();
3937 unsigned BitWidth = VT.getSizeInBits();
3938 assert(Op.getNumOperands() == 3 &&
3939 VT == Op.getOperand(1).getValueType() &&
3942 // Expand into a bunch of logical ops. Note that these ops
3943 // depend on the PPC behavior for oversized shift amounts.
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::SRL, dl, VT, Hi, Tmp5);
3957 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3958 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3959 SDValue OutOps[] = { OutLo, OutHi };
3960 return DAG.getMergeValues(OutOps, 2, dl);
3963 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3964 DebugLoc dl = Op.getDebugLoc();
3965 EVT VT = Op.getValueType();
3966 unsigned BitWidth = VT.getSizeInBits();
3967 assert(Op.getNumOperands() == 3 &&
3968 VT == Op.getOperand(1).getValueType() &&
3971 // Expand into a bunch of logical ops, followed by a select_cc.
3972 SDValue Lo = Op.getOperand(0);
3973 SDValue Hi = Op.getOperand(1);
3974 SDValue Amt = Op.getOperand(2);
3975 EVT AmtVT = Amt.getValueType();
3977 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3978 DAG.getConstant(BitWidth, AmtVT), Amt);
3979 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3980 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3981 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3982 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3983 DAG.getConstant(-BitWidth, AmtVT));
3984 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3985 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3986 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3987 Tmp4, Tmp6, ISD::SETLE);
3988 SDValue OutOps[] = { OutLo, OutHi };
3989 return DAG.getMergeValues(OutOps, 2, dl);
3992 //===----------------------------------------------------------------------===//
3993 // Vector related lowering.
3996 /// BuildSplatI - Build a canonical splati of Val with an element size of
3997 /// SplatSize. Cast the result to VT.
3998 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3999 SelectionDAG &DAG, DebugLoc dl) {
4000 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4002 static const EVT VTys[] = { // canonical VT to use for each size.
4003 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4006 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4008 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4012 EVT CanonicalVT = VTys[SplatSize-1];
4014 // Build a canonical splat for this value.
4015 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4016 SmallVector<SDValue, 8> Ops;
4017 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4018 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4019 &Ops[0], Ops.size());
4020 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4023 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4024 /// specified intrinsic ID.
4025 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4026 SelectionDAG &DAG, DebugLoc dl,
4027 EVT DestVT = MVT::Other) {
4028 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4029 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4030 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4033 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4034 /// specified intrinsic ID.
4035 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4036 SDValue Op2, SelectionDAG &DAG,
4037 DebugLoc dl, EVT DestVT = MVT::Other) {
4038 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4039 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4040 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4044 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4045 /// amount. The result has the specified value type.
4046 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4047 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4048 // Force LHS/RHS to be the right type.
4049 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4050 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4053 for (unsigned i = 0; i != 16; ++i)
4055 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4056 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4059 // If this is a case we can't handle, return null and let the default
4060 // expansion code take care of it. If we CAN select this case, and if it
4061 // selects to a single instruction, return Op. Otherwise, if we can codegen
4062 // this case more efficiently than a constant pool load, lower it to the
4063 // sequence of ops that should be used.
4064 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4065 SelectionDAG &DAG) const {
4066 DebugLoc dl = Op.getDebugLoc();
4067 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4068 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4070 // Check if this is a splat of a constant value.
4071 APInt APSplatBits, APSplatUndef;
4072 unsigned SplatBitSize;
4074 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4075 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4078 unsigned SplatBits = APSplatBits.getZExtValue();
4079 unsigned SplatUndef = APSplatUndef.getZExtValue();
4080 unsigned SplatSize = SplatBitSize / 8;
4082 // First, handle single instruction cases.
4085 if (SplatBits == 0) {
4086 // Canonicalize all zero vectors to be v4i32.
4087 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4088 SDValue Z = DAG.getConstant(0, MVT::i32);
4089 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4090 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4095 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4096 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4098 if (SextVal >= -16 && SextVal <= 15)
4099 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4102 // Two instruction sequences.
4104 // If this value is in the range [-32,30] and is even, use:
4105 // tmp = VSPLTI[bhw], result = add tmp, tmp
4106 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4107 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4108 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4109 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4112 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4113 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4115 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4116 // Make -1 and vspltisw -1:
4117 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4119 // Make the VSLW intrinsic, computing 0x8000_0000.
4120 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4123 // xor by OnesV to invert it.
4124 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4125 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4128 // Check to see if this is a wide variety of vsplti*, binop self cases.
4129 static const signed char SplatCsts[] = {
4130 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4131 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4134 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4135 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4136 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4137 int i = SplatCsts[idx];
4139 // Figure out what shift amount will be used by altivec if shifted by i in
4141 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4143 // vsplti + shl self.
4144 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4145 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4146 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4147 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4148 Intrinsic::ppc_altivec_vslw
4150 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4151 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4154 // vsplti + srl self.
4155 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4156 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4157 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4158 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4159 Intrinsic::ppc_altivec_vsrw
4161 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4162 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4165 // vsplti + sra self.
4166 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4167 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4168 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4169 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4170 Intrinsic::ppc_altivec_vsraw
4172 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4173 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4176 // vsplti + rol self.
4177 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4178 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4179 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4180 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4181 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4182 Intrinsic::ppc_altivec_vrlw
4184 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4185 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4188 // t = vsplti c, result = vsldoi t, t, 1
4189 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
4190 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4191 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4193 // t = vsplti c, result = vsldoi t, t, 2
4194 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
4195 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4196 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4198 // t = vsplti c, result = vsldoi t, t, 3
4199 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4200 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4201 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4205 // Three instruction sequences.
4207 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4208 if (SextVal >= 0 && SextVal <= 31) {
4209 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4210 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4211 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4212 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4214 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4215 if (SextVal >= -31 && SextVal <= 0) {
4216 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4217 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4218 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4219 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4225 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4226 /// the specified operations to build the shuffle.
4227 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4228 SDValue RHS, SelectionDAG &DAG,
4230 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4231 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4232 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4235 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4247 if (OpNum == OP_COPY) {
4248 if (LHSID == (1*9+2)*9+3) return LHS;
4249 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4253 SDValue OpLHS, OpRHS;
4254 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4255 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4259 default: llvm_unreachable("Unknown i32 permute!");
4261 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4262 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4263 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4264 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4267 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4268 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4269 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4270 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4273 for (unsigned i = 0; i != 16; ++i)
4274 ShufIdxs[i] = (i&3)+0;
4277 for (unsigned i = 0; i != 16; ++i)
4278 ShufIdxs[i] = (i&3)+4;
4281 for (unsigned i = 0; i != 16; ++i)
4282 ShufIdxs[i] = (i&3)+8;
4285 for (unsigned i = 0; i != 16; ++i)
4286 ShufIdxs[i] = (i&3)+12;
4289 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4291 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4293 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4295 EVT VT = OpLHS.getValueType();
4296 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4297 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4298 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4299 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4302 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4303 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4304 /// return the code it can be lowered into. Worst case, it can always be
4305 /// lowered into a vperm.
4306 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4307 SelectionDAG &DAG) const {
4308 DebugLoc dl = Op.getDebugLoc();
4309 SDValue V1 = Op.getOperand(0);
4310 SDValue V2 = Op.getOperand(1);
4311 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4312 EVT VT = Op.getValueType();
4314 // Cases that are handled by instructions that take permute immediates
4315 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4316 // selected by the instruction selector.
4317 if (V2.getOpcode() == ISD::UNDEF) {
4318 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4319 PPC::isSplatShuffleMask(SVOp, 2) ||
4320 PPC::isSplatShuffleMask(SVOp, 4) ||
4321 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4322 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4323 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4324 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4325 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4326 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4327 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4328 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4329 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4334 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4335 // and produce a fixed permutation. If any of these match, do not lower to
4337 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4338 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4339 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4340 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4341 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4342 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4343 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4344 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4345 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4348 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4349 // perfect shuffle table to emit an optimal matching sequence.
4350 ArrayRef<int> PermMask = SVOp->getMask();
4352 unsigned PFIndexes[4];
4353 bool isFourElementShuffle = true;
4354 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4355 unsigned EltNo = 8; // Start out undef.
4356 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4357 if (PermMask[i*4+j] < 0)
4358 continue; // Undef, ignore it.
4360 unsigned ByteSource = PermMask[i*4+j];
4361 if ((ByteSource & 3) != j) {
4362 isFourElementShuffle = false;
4367 EltNo = ByteSource/4;
4368 } else if (EltNo != ByteSource/4) {
4369 isFourElementShuffle = false;
4373 PFIndexes[i] = EltNo;
4376 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4377 // perfect shuffle vector to determine if it is cost effective to do this as
4378 // discrete instructions, or whether we should use a vperm.
4379 if (isFourElementShuffle) {
4380 // Compute the index in the perfect shuffle table.
4381 unsigned PFTableIndex =
4382 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4384 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4385 unsigned Cost = (PFEntry >> 30);
4387 // Determining when to avoid vperm is tricky. Many things affect the cost
4388 // of vperm, particularly how many times the perm mask needs to be computed.
4389 // For example, if the perm mask can be hoisted out of a loop or is already
4390 // used (perhaps because there are multiple permutes with the same shuffle
4391 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4392 // the loop requires an extra register.
4394 // As a compromise, we only emit discrete instructions if the shuffle can be
4395 // generated in 3 or fewer operations. When we have loop information
4396 // available, if this block is within a loop, we should avoid using vperm
4397 // for 3-operation perms and use a constant pool load instead.
4399 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4402 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4403 // vector that will get spilled to the constant pool.
4404 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4406 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4407 // that it is in input element units, not in bytes. Convert now.
4408 EVT EltVT = V1.getValueType().getVectorElementType();
4409 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4411 SmallVector<SDValue, 16> ResultMask;
4412 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4413 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4415 for (unsigned j = 0; j != BytesPerElement; ++j)
4416 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4420 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4421 &ResultMask[0], ResultMask.size());
4422 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4425 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4426 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4427 /// information about the intrinsic.
4428 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4430 unsigned IntrinsicID =
4431 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4434 switch (IntrinsicID) {
4435 default: return false;
4436 // Comparison predicates.
4437 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4438 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4439 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4440 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4441 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4442 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4443 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4444 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4445 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4446 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4447 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4448 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4449 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4451 // Normal Comparisons.
4452 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4453 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4454 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4455 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4456 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4457 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4458 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4459 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4460 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4461 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4462 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4463 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4464 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4469 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4470 /// lower, do it, otherwise return null.
4471 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4472 SelectionDAG &DAG) const {
4473 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4474 // opcode number of the comparison.
4475 DebugLoc dl = Op.getDebugLoc();
4478 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4479 return SDValue(); // Don't custom lower most intrinsics.
4481 // If this is a non-dot comparison, make the VCMP node and we are done.
4483 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4484 Op.getOperand(1), Op.getOperand(2),
4485 DAG.getConstant(CompareOpc, MVT::i32));
4486 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4489 // Create the PPCISD altivec 'dot' comparison node.
4491 Op.getOperand(2), // LHS
4492 Op.getOperand(3), // RHS
4493 DAG.getConstant(CompareOpc, MVT::i32)
4495 std::vector<EVT> VTs;
4496 VTs.push_back(Op.getOperand(2).getValueType());
4497 VTs.push_back(MVT::Glue);
4498 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4500 // Now that we have the comparison, emit a copy from the CR to a GPR.
4501 // This is flagged to the above dot comparison.
4502 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4503 DAG.getRegister(PPC::CR6, MVT::i32),
4504 CompNode.getValue(1));
4506 // Unpack the result based on how the target uses it.
4507 unsigned BitNo; // Bit # of CR6.
4508 bool InvertBit; // Invert result?
4509 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4510 default: // Can't happen, don't crash on invalid number though.
4511 case 0: // Return the value of the EQ bit of CR6.
4512 BitNo = 0; InvertBit = false;
4514 case 1: // Return the inverted value of the EQ bit of CR6.
4515 BitNo = 0; InvertBit = true;
4517 case 2: // Return the value of the LT bit of CR6.
4518 BitNo = 2; InvertBit = false;
4520 case 3: // Return the inverted value of the LT bit of CR6.
4521 BitNo = 2; InvertBit = true;
4525 // Shift the bit into the low position.
4526 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4527 DAG.getConstant(8-(3-BitNo), MVT::i32));
4529 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4530 DAG.getConstant(1, MVT::i32));
4532 // If we are supposed to, toggle the bit.
4534 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4535 DAG.getConstant(1, MVT::i32));
4539 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4540 SelectionDAG &DAG) const {
4541 DebugLoc dl = Op.getDebugLoc();
4542 // Create a stack slot that is 16-byte aligned.
4543 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4544 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4545 EVT PtrVT = getPointerTy();
4546 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4548 // Store the input value into Value#0 of the stack slot.
4549 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4550 Op.getOperand(0), FIdx, MachinePointerInfo(),
4553 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4554 false, false, false, 0);
4557 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4558 DebugLoc dl = Op.getDebugLoc();
4559 if (Op.getValueType() == MVT::v4i32) {
4560 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4562 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4563 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4565 SDValue RHSSwap = // = vrlw RHS, 16
4566 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4568 // Shrinkify inputs to v8i16.
4569 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4570 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4571 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4573 // Low parts multiplied together, generating 32-bit results (we ignore the
4575 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4576 LHS, RHS, DAG, dl, MVT::v4i32);
4578 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4579 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4580 // Shift the high parts up 16 bits.
4581 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4583 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4584 } else if (Op.getValueType() == MVT::v8i16) {
4585 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4587 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4589 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4590 LHS, RHS, Zero, DAG, dl);
4591 } else if (Op.getValueType() == MVT::v16i8) {
4592 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4594 // Multiply the even 8-bit parts, producing 16-bit sums.
4595 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4596 LHS, RHS, DAG, dl, MVT::v8i16);
4597 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4599 // Multiply the odd 8-bit parts, producing 16-bit sums.
4600 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4601 LHS, RHS, DAG, dl, MVT::v8i16);
4602 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4604 // Merge the results together.
4606 for (unsigned i = 0; i != 8; ++i) {
4608 Ops[i*2+1] = 2*i+1+16;
4610 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4612 llvm_unreachable("Unknown mul to lower!");
4616 /// LowerOperation - Provide custom lowering hooks for some operations.
4618 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4619 switch (Op.getOpcode()) {
4620 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4621 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4622 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4623 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4624 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4625 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4626 case ISD::SETCC: return LowerSETCC(Op, DAG);
4627 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4628 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4630 return LowerVASTART(Op, DAG, PPCSubTarget);
4633 return LowerVAARG(Op, DAG, PPCSubTarget);
4635 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4636 case ISD::DYNAMIC_STACKALLOC:
4637 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4639 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4640 case ISD::FP_TO_UINT:
4641 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4643 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4644 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4646 // Lower 64-bit shifts.
4647 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4648 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4649 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4651 // Vector-related lowering.
4652 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4653 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4654 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4655 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4656 case ISD::MUL: return LowerMUL(Op, DAG);
4658 // Frame & Return address.
4659 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4660 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4664 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4665 SmallVectorImpl<SDValue>&Results,
4666 SelectionDAG &DAG) const {
4667 const TargetMachine &TM = getTargetMachine();
4668 DebugLoc dl = N->getDebugLoc();
4669 switch (N->getOpcode()) {
4671 llvm_unreachable("Do not know how to custom type legalize this operation!");
4673 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4674 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4677 EVT VT = N->getValueType(0);
4679 if (VT == MVT::i64) {
4680 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4682 Results.push_back(NewNode);
4683 Results.push_back(NewNode.getValue(1));
4687 case ISD::FP_ROUND_INREG: {
4688 assert(N->getValueType(0) == MVT::ppcf128);
4689 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4690 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4691 MVT::f64, N->getOperand(0),
4692 DAG.getIntPtrConstant(0));
4693 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4694 MVT::f64, N->getOperand(0),
4695 DAG.getIntPtrConstant(1));
4697 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4698 // of the long double, and puts FPSCR back the way it was. We do not
4699 // actually model FPSCR.
4700 std::vector<EVT> NodeTys;
4701 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4703 NodeTys.push_back(MVT::f64); // Return register
4704 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4705 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4706 MFFSreg = Result.getValue(0);
4707 InFlag = Result.getValue(1);
4710 NodeTys.push_back(MVT::Glue); // Returns a flag
4711 Ops[0] = DAG.getConstant(31, MVT::i32);
4713 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4714 InFlag = Result.getValue(0);
4717 NodeTys.push_back(MVT::Glue); // Returns a flag
4718 Ops[0] = DAG.getConstant(30, MVT::i32);
4720 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4721 InFlag = Result.getValue(0);
4724 NodeTys.push_back(MVT::f64); // result of add
4725 NodeTys.push_back(MVT::Glue); // Returns a flag
4729 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4730 FPreg = Result.getValue(0);
4731 InFlag = Result.getValue(1);
4734 NodeTys.push_back(MVT::f64);
4735 Ops[0] = DAG.getConstant(1, MVT::i32);
4739 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4740 FPreg = Result.getValue(0);
4742 // We know the low half is about to be thrown away, so just use something
4744 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4748 case ISD::FP_TO_SINT:
4749 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4755 //===----------------------------------------------------------------------===//
4756 // Other Lowering Code
4757 //===----------------------------------------------------------------------===//
4760 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4761 bool is64bit, unsigned BinOpcode) const {
4762 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4763 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4765 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4766 MachineFunction *F = BB->getParent();
4767 MachineFunction::iterator It = BB;
4770 unsigned dest = MI->getOperand(0).getReg();
4771 unsigned ptrA = MI->getOperand(1).getReg();
4772 unsigned ptrB = MI->getOperand(2).getReg();
4773 unsigned incr = MI->getOperand(3).getReg();
4774 DebugLoc dl = MI->getDebugLoc();
4776 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4777 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4778 F->insert(It, loopMBB);
4779 F->insert(It, exitMBB);
4780 exitMBB->splice(exitMBB->begin(), BB,
4781 llvm::next(MachineBasicBlock::iterator(MI)),
4783 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4785 MachineRegisterInfo &RegInfo = F->getRegInfo();
4786 unsigned TmpReg = (!BinOpcode) ? incr :
4787 RegInfo.createVirtualRegister(
4788 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4789 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4793 // fallthrough --> loopMBB
4794 BB->addSuccessor(loopMBB);
4797 // l[wd]arx dest, ptr
4798 // add r0, dest, incr
4799 // st[wd]cx. r0, ptr
4801 // fallthrough --> exitMBB
4803 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4804 .addReg(ptrA).addReg(ptrB);
4806 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4807 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4808 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4809 BuildMI(BB, dl, TII->get(PPC::BCC))
4810 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4811 BB->addSuccessor(loopMBB);
4812 BB->addSuccessor(exitMBB);
4821 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4822 MachineBasicBlock *BB,
4823 bool is8bit, // operation
4824 unsigned BinOpcode) const {
4825 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4826 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4827 // In 64 bit mode we have to use 64 bits for addresses, even though the
4828 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4829 // registers without caring whether they're 32 or 64, but here we're
4830 // doing actual arithmetic on the addresses.
4831 bool is64bit = PPCSubTarget.isPPC64();
4832 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4834 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4835 MachineFunction *F = BB->getParent();
4836 MachineFunction::iterator It = BB;
4839 unsigned dest = MI->getOperand(0).getReg();
4840 unsigned ptrA = MI->getOperand(1).getReg();
4841 unsigned ptrB = MI->getOperand(2).getReg();
4842 unsigned incr = MI->getOperand(3).getReg();
4843 DebugLoc dl = MI->getDebugLoc();
4845 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4846 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4847 F->insert(It, loopMBB);
4848 F->insert(It, exitMBB);
4849 exitMBB->splice(exitMBB->begin(), BB,
4850 llvm::next(MachineBasicBlock::iterator(MI)),
4852 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4854 MachineRegisterInfo &RegInfo = F->getRegInfo();
4855 const TargetRegisterClass *RC =
4856 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4857 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4858 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4859 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4860 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4861 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4862 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4863 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4864 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4865 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4866 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4867 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4868 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4870 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4874 // fallthrough --> loopMBB
4875 BB->addSuccessor(loopMBB);
4877 // The 4-byte load must be aligned, while a char or short may be
4878 // anywhere in the word. Hence all this nasty bookkeeping code.
4879 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4880 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4881 // xori shift, shift1, 24 [16]
4882 // rlwinm ptr, ptr1, 0, 0, 29
4883 // slw incr2, incr, shift
4884 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4885 // slw mask, mask2, shift
4887 // lwarx tmpDest, ptr
4888 // add tmp, tmpDest, incr2
4889 // andc tmp2, tmpDest, mask
4890 // and tmp3, tmp, mask
4891 // or tmp4, tmp3, tmp2
4894 // fallthrough --> exitMBB
4895 // srw dest, tmpDest, shift
4896 if (ptrA != ZeroReg) {
4897 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4898 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4899 .addReg(ptrA).addReg(ptrB);
4903 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4904 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4905 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4906 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4908 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4909 .addReg(Ptr1Reg).addImm(0).addImm(61);
4911 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4912 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4913 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4914 .addReg(incr).addReg(ShiftReg);
4916 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4918 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4919 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4921 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4922 .addReg(Mask2Reg).addReg(ShiftReg);
4925 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4926 .addReg(ZeroReg).addReg(PtrReg);
4928 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4929 .addReg(Incr2Reg).addReg(TmpDestReg);
4930 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4931 .addReg(TmpDestReg).addReg(MaskReg);
4932 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4933 .addReg(TmpReg).addReg(MaskReg);
4934 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4935 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4936 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4937 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4938 BuildMI(BB, dl, TII->get(PPC::BCC))
4939 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4940 BB->addSuccessor(loopMBB);
4941 BB->addSuccessor(exitMBB);
4946 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4952 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4953 MachineBasicBlock *BB) const {
4954 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4956 // To "insert" these instructions we actually have to insert their
4957 // control-flow patterns.
4958 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4959 MachineFunction::iterator It = BB;
4962 MachineFunction *F = BB->getParent();
4964 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4965 MI->getOpcode() == PPC::SELECT_CC_I8)) {
4966 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
4967 PPC::ISEL8 : PPC::ISEL;
4968 unsigned SelectPred = MI->getOperand(4).getImm();
4969 DebugLoc dl = MI->getDebugLoc();
4971 // The SelectPred is ((BI << 5) | BO) for a BCC
4972 unsigned BO = SelectPred & 0xF;
4973 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
4975 unsigned TrueOpNo, FalseOpNo;
4982 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
4985 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
4986 .addReg(MI->getOperand(TrueOpNo).getReg())
4987 .addReg(MI->getOperand(FalseOpNo).getReg())
4988 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
4989 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4990 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4991 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4992 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4993 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4996 // The incoming instruction knows the destination vreg to set, the
4997 // condition code register to branch on, the true/false values to
4998 // select between, and a branch opcode to use.
5003 // cmpTY ccX, r1, r2
5005 // fallthrough --> copy0MBB
5006 MachineBasicBlock *thisMBB = BB;
5007 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5008 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5009 unsigned SelectPred = MI->getOperand(4).getImm();
5010 DebugLoc dl = MI->getDebugLoc();
5011 F->insert(It, copy0MBB);
5012 F->insert(It, sinkMBB);
5014 // Transfer the remainder of BB and its successor edges to sinkMBB.
5015 sinkMBB->splice(sinkMBB->begin(), BB,
5016 llvm::next(MachineBasicBlock::iterator(MI)),
5018 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5020 // Next, add the true and fallthrough blocks as its successors.
5021 BB->addSuccessor(copy0MBB);
5022 BB->addSuccessor(sinkMBB);
5024 BuildMI(BB, dl, TII->get(PPC::BCC))
5025 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5028 // %FalseValue = ...
5029 // # fallthrough to sinkMBB
5032 // Update machine-CFG edges
5033 BB->addSuccessor(sinkMBB);
5036 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5039 BuildMI(*BB, BB->begin(), dl,
5040 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5041 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5042 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5044 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5045 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5046 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5047 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5048 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5049 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5050 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5051 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5053 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5054 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5055 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5056 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5057 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5058 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5059 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5060 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5062 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5063 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5064 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5065 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5066 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5067 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5068 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5069 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5071 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5072 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5073 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5074 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5075 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5076 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5077 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5078 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5080 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5081 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5082 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5083 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5084 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5085 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5086 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5087 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5089 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5090 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5091 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5092 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5093 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5094 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5095 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5096 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5098 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5099 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5100 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5101 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5102 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5103 BB = EmitAtomicBinary(MI, BB, false, 0);
5104 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5105 BB = EmitAtomicBinary(MI, BB, true, 0);
5107 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5108 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5109 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5111 unsigned dest = MI->getOperand(0).getReg();
5112 unsigned ptrA = MI->getOperand(1).getReg();
5113 unsigned ptrB = MI->getOperand(2).getReg();
5114 unsigned oldval = MI->getOperand(3).getReg();
5115 unsigned newval = MI->getOperand(4).getReg();
5116 DebugLoc dl = MI->getDebugLoc();
5118 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5119 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5120 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5121 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5122 F->insert(It, loop1MBB);
5123 F->insert(It, loop2MBB);
5124 F->insert(It, midMBB);
5125 F->insert(It, exitMBB);
5126 exitMBB->splice(exitMBB->begin(), BB,
5127 llvm::next(MachineBasicBlock::iterator(MI)),
5129 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5133 // fallthrough --> loopMBB
5134 BB->addSuccessor(loop1MBB);
5137 // l[wd]arx dest, ptr
5138 // cmp[wd] dest, oldval
5141 // st[wd]cx. newval, ptr
5145 // st[wd]cx. dest, ptr
5148 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5149 .addReg(ptrA).addReg(ptrB);
5150 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5151 .addReg(oldval).addReg(dest);
5152 BuildMI(BB, dl, TII->get(PPC::BCC))
5153 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5154 BB->addSuccessor(loop2MBB);
5155 BB->addSuccessor(midMBB);
5158 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5159 .addReg(newval).addReg(ptrA).addReg(ptrB);
5160 BuildMI(BB, dl, TII->get(PPC::BCC))
5161 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5162 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5163 BB->addSuccessor(loop1MBB);
5164 BB->addSuccessor(exitMBB);
5167 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5168 .addReg(dest).addReg(ptrA).addReg(ptrB);
5169 BB->addSuccessor(exitMBB);
5174 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5175 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5176 // We must use 64-bit registers for addresses when targeting 64-bit,
5177 // since we're actually doing arithmetic on them. Other registers
5179 bool is64bit = PPCSubTarget.isPPC64();
5180 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5182 unsigned dest = MI->getOperand(0).getReg();
5183 unsigned ptrA = MI->getOperand(1).getReg();
5184 unsigned ptrB = MI->getOperand(2).getReg();
5185 unsigned oldval = MI->getOperand(3).getReg();
5186 unsigned newval = MI->getOperand(4).getReg();
5187 DebugLoc dl = MI->getDebugLoc();
5189 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5190 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5191 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5192 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5193 F->insert(It, loop1MBB);
5194 F->insert(It, loop2MBB);
5195 F->insert(It, midMBB);
5196 F->insert(It, exitMBB);
5197 exitMBB->splice(exitMBB->begin(), BB,
5198 llvm::next(MachineBasicBlock::iterator(MI)),
5200 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5202 MachineRegisterInfo &RegInfo = F->getRegInfo();
5203 const TargetRegisterClass *RC =
5204 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5205 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5206 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5207 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5208 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5209 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5210 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5211 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5212 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5213 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5214 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5215 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5216 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5217 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5218 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5220 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5221 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5224 // fallthrough --> loopMBB
5225 BB->addSuccessor(loop1MBB);
5227 // The 4-byte load must be aligned, while a char or short may be
5228 // anywhere in the word. Hence all this nasty bookkeeping code.
5229 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5230 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5231 // xori shift, shift1, 24 [16]
5232 // rlwinm ptr, ptr1, 0, 0, 29
5233 // slw newval2, newval, shift
5234 // slw oldval2, oldval,shift
5235 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5236 // slw mask, mask2, shift
5237 // and newval3, newval2, mask
5238 // and oldval3, oldval2, mask
5240 // lwarx tmpDest, ptr
5241 // and tmp, tmpDest, mask
5242 // cmpw tmp, oldval3
5245 // andc tmp2, tmpDest, mask
5246 // or tmp4, tmp2, newval3
5251 // stwcx. tmpDest, ptr
5253 // srw dest, tmpDest, shift
5254 if (ptrA != ZeroReg) {
5255 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5256 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5257 .addReg(ptrA).addReg(ptrB);
5261 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5262 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5263 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5264 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5266 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5267 .addReg(Ptr1Reg).addImm(0).addImm(61);
5269 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5270 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5271 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5272 .addReg(newval).addReg(ShiftReg);
5273 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5274 .addReg(oldval).addReg(ShiftReg);
5276 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5278 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5279 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5280 .addReg(Mask3Reg).addImm(65535);
5282 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5283 .addReg(Mask2Reg).addReg(ShiftReg);
5284 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5285 .addReg(NewVal2Reg).addReg(MaskReg);
5286 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5287 .addReg(OldVal2Reg).addReg(MaskReg);
5290 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5291 .addReg(ZeroReg).addReg(PtrReg);
5292 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5293 .addReg(TmpDestReg).addReg(MaskReg);
5294 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5295 .addReg(TmpReg).addReg(OldVal3Reg);
5296 BuildMI(BB, dl, TII->get(PPC::BCC))
5297 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5298 BB->addSuccessor(loop2MBB);
5299 BB->addSuccessor(midMBB);
5302 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5303 .addReg(TmpDestReg).addReg(MaskReg);
5304 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5305 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5306 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5307 .addReg(ZeroReg).addReg(PtrReg);
5308 BuildMI(BB, dl, TII->get(PPC::BCC))
5309 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5310 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5311 BB->addSuccessor(loop1MBB);
5312 BB->addSuccessor(exitMBB);
5315 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5316 .addReg(ZeroReg).addReg(PtrReg);
5317 BB->addSuccessor(exitMBB);
5322 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5325 llvm_unreachable("Unexpected instr type to insert");
5328 MI->eraseFromParent(); // The pseudo instruction is gone now.
5332 //===----------------------------------------------------------------------===//
5333 // Target Optimization Hooks
5334 //===----------------------------------------------------------------------===//
5336 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5337 DAGCombinerInfo &DCI) const {
5338 const TargetMachine &TM = getTargetMachine();
5339 SelectionDAG &DAG = DCI.DAG;
5340 DebugLoc dl = N->getDebugLoc();
5341 switch (N->getOpcode()) {
5344 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5345 if (C->isNullValue()) // 0 << V -> 0.
5346 return N->getOperand(0);
5350 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5351 if (C->isNullValue()) // 0 >>u V -> 0.
5352 return N->getOperand(0);
5356 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5357 if (C->isNullValue() || // 0 >>s V -> 0.
5358 C->isAllOnesValue()) // -1 >>s V -> -1.
5359 return N->getOperand(0);
5363 case ISD::SINT_TO_FP:
5364 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5365 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5366 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5367 // We allow the src/dst to be either f32/f64, but the intermediate
5368 // type must be i64.
5369 if (N->getOperand(0).getValueType() == MVT::i64 &&
5370 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5371 SDValue Val = N->getOperand(0).getOperand(0);
5372 if (Val.getValueType() == MVT::f32) {
5373 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5374 DCI.AddToWorklist(Val.getNode());
5377 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5378 DCI.AddToWorklist(Val.getNode());
5379 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5380 DCI.AddToWorklist(Val.getNode());
5381 if (N->getValueType(0) == MVT::f32) {
5382 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5383 DAG.getIntPtrConstant(0));
5384 DCI.AddToWorklist(Val.getNode());
5387 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5388 // If the intermediate type is i32, we can avoid the load/store here
5395 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5396 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5397 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5398 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5399 N->getOperand(1).getValueType() == MVT::i32 &&
5400 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5401 SDValue Val = N->getOperand(1).getOperand(0);
5402 if (Val.getValueType() == MVT::f32) {
5403 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5404 DCI.AddToWorklist(Val.getNode());
5406 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5407 DCI.AddToWorklist(Val.getNode());
5409 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5410 N->getOperand(2), N->getOperand(3));
5411 DCI.AddToWorklist(Val.getNode());
5415 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5416 if (cast<StoreSDNode>(N)->isUnindexed() &&
5417 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5418 N->getOperand(1).getNode()->hasOneUse() &&
5419 (N->getOperand(1).getValueType() == MVT::i32 ||
5420 N->getOperand(1).getValueType() == MVT::i16)) {
5421 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5422 // Do an any-extend to 32-bits if this is a half-word input.
5423 if (BSwapOp.getValueType() == MVT::i16)
5424 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5427 N->getOperand(0), BSwapOp, N->getOperand(2),
5428 DAG.getValueType(N->getOperand(1).getValueType())
5431 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5432 Ops, array_lengthof(Ops),
5433 cast<StoreSDNode>(N)->getMemoryVT(),
5434 cast<StoreSDNode>(N)->getMemOperand());
5438 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5439 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5440 N->getOperand(0).hasOneUse() &&
5441 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5442 SDValue Load = N->getOperand(0);
5443 LoadSDNode *LD = cast<LoadSDNode>(Load);
5444 // Create the byte-swapping load.
5446 LD->getChain(), // Chain
5447 LD->getBasePtr(), // Ptr
5448 DAG.getValueType(N->getValueType(0)) // VT
5451 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5452 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5453 LD->getMemoryVT(), LD->getMemOperand());
5455 // If this is an i16 load, insert the truncate.
5456 SDValue ResVal = BSLoad;
5457 if (N->getValueType(0) == MVT::i16)
5458 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5460 // First, combine the bswap away. This makes the value produced by the
5462 DCI.CombineTo(N, ResVal);
5464 // Next, combine the load away, we give it a bogus result value but a real
5465 // chain result. The result value is dead because the bswap is dead.
5466 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5468 // Return N so it doesn't get rechecked!
5469 return SDValue(N, 0);
5473 case PPCISD::VCMP: {
5474 // If a VCMPo node already exists with exactly the same operands as this
5475 // node, use its result instead of this node (VCMPo computes both a CR6 and
5476 // a normal output).
5478 if (!N->getOperand(0).hasOneUse() &&
5479 !N->getOperand(1).hasOneUse() &&
5480 !N->getOperand(2).hasOneUse()) {
5482 // Scan all of the users of the LHS, looking for VCMPo's that match.
5483 SDNode *VCMPoNode = 0;
5485 SDNode *LHSN = N->getOperand(0).getNode();
5486 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5488 if (UI->getOpcode() == PPCISD::VCMPo &&
5489 UI->getOperand(1) == N->getOperand(1) &&
5490 UI->getOperand(2) == N->getOperand(2) &&
5491 UI->getOperand(0) == N->getOperand(0)) {
5496 // If there is no VCMPo node, or if the flag value has a single use, don't
5498 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5501 // Look at the (necessarily single) use of the flag value. If it has a
5502 // chain, this transformation is more complex. Note that multiple things
5503 // could use the value result, which we should ignore.
5504 SDNode *FlagUser = 0;
5505 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5506 FlagUser == 0; ++UI) {
5507 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5509 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5510 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5517 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5518 // give up for right now.
5519 if (FlagUser->getOpcode() == PPCISD::MFCR)
5520 return SDValue(VCMPoNode, 0);
5525 // If this is a branch on an altivec predicate comparison, lower this so
5526 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5527 // lowering is done pre-legalize, because the legalizer lowers the predicate
5528 // compare down to code that is difficult to reassemble.
5529 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5530 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5534 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5535 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5536 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5537 assert(isDot && "Can't compare against a vector result!");
5539 // If this is a comparison against something other than 0/1, then we know
5540 // that the condition is never/always true.
5541 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5542 if (Val != 0 && Val != 1) {
5543 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5544 return N->getOperand(0);
5545 // Always !=, turn it into an unconditional branch.
5546 return DAG.getNode(ISD::BR, dl, MVT::Other,
5547 N->getOperand(0), N->getOperand(4));
5550 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5552 // Create the PPCISD altivec 'dot' comparison node.
5553 std::vector<EVT> VTs;
5555 LHS.getOperand(2), // LHS of compare
5556 LHS.getOperand(3), // RHS of compare
5557 DAG.getConstant(CompareOpc, MVT::i32)
5559 VTs.push_back(LHS.getOperand(2).getValueType());
5560 VTs.push_back(MVT::Glue);
5561 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5563 // Unpack the result based on how the target uses it.
5564 PPC::Predicate CompOpc;
5565 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5566 default: // Can't happen, don't crash on invalid number though.
5567 case 0: // Branch on the value of the EQ bit of CR6.
5568 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5570 case 1: // Branch on the inverted value of the EQ bit of CR6.
5571 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5573 case 2: // Branch on the value of the LT bit of CR6.
5574 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5576 case 3: // Branch on the inverted value of the LT bit of CR6.
5577 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5581 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5582 DAG.getConstant(CompOpc, MVT::i32),
5583 DAG.getRegister(PPC::CR6, MVT::i32),
5584 N->getOperand(4), CompNode.getValue(1));
5593 //===----------------------------------------------------------------------===//
5594 // Inline Assembly Support
5595 //===----------------------------------------------------------------------===//
5597 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5600 const SelectionDAG &DAG,
5601 unsigned Depth) const {
5602 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
5603 switch (Op.getOpcode()) {
5605 case PPCISD::LBRX: {
5606 // lhbrx is known to have the top bits cleared out.
5607 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5608 KnownZero = 0xFFFF0000;
5611 case ISD::INTRINSIC_WO_CHAIN: {
5612 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5614 case Intrinsic::ppc_altivec_vcmpbfp_p:
5615 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5616 case Intrinsic::ppc_altivec_vcmpequb_p:
5617 case Intrinsic::ppc_altivec_vcmpequh_p:
5618 case Intrinsic::ppc_altivec_vcmpequw_p:
5619 case Intrinsic::ppc_altivec_vcmpgefp_p:
5620 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5621 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5622 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5623 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5624 case Intrinsic::ppc_altivec_vcmpgtub_p:
5625 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5626 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5627 KnownZero = ~1U; // All bits but the low one are known to be zero.
5635 /// getConstraintType - Given a constraint, return the type of
5636 /// constraint it is for this target.
5637 PPCTargetLowering::ConstraintType
5638 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5639 if (Constraint.size() == 1) {
5640 switch (Constraint[0]) {
5647 return C_RegisterClass;
5650 return TargetLowering::getConstraintType(Constraint);
5653 /// Examine constraint type and operand type and determine a weight value.
5654 /// This object must already have been set up with the operand type
5655 /// and the current alternative constraint selected.
5656 TargetLowering::ConstraintWeight
5657 PPCTargetLowering::getSingleConstraintMatchWeight(
5658 AsmOperandInfo &info, const char *constraint) const {
5659 ConstraintWeight weight = CW_Invalid;
5660 Value *CallOperandVal = info.CallOperandVal;
5661 // If we don't have a value, we can't do a match,
5662 // but allow it at the lowest weight.
5663 if (CallOperandVal == NULL)
5665 Type *type = CallOperandVal->getType();
5666 // Look at the constraint type.
5667 switch (*constraint) {
5669 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5672 if (type->isIntegerTy())
5673 weight = CW_Register;
5676 if (type->isFloatTy())
5677 weight = CW_Register;
5680 if (type->isDoubleTy())
5681 weight = CW_Register;
5684 if (type->isVectorTy())
5685 weight = CW_Register;
5688 weight = CW_Register;
5694 std::pair<unsigned, const TargetRegisterClass*>
5695 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5697 if (Constraint.size() == 1) {
5698 // GCC RS6000 Constraint Letters
5699 switch (Constraint[0]) {
5702 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5703 return std::make_pair(0U, &PPC::G8RCRegClass);
5704 return std::make_pair(0U, &PPC::GPRCRegClass);
5707 return std::make_pair(0U, &PPC::F4RCRegClass);
5709 return std::make_pair(0U, &PPC::F8RCRegClass);
5712 return std::make_pair(0U, &PPC::VRRCRegClass);
5714 return std::make_pair(0U, &PPC::CRRCRegClass);
5718 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5722 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5723 /// vector. If it is invalid, don't add anything to Ops.
5724 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5725 std::string &Constraint,
5726 std::vector<SDValue>&Ops,
5727 SelectionDAG &DAG) const {
5728 SDValue Result(0,0);
5730 // Only support length 1 constraints.
5731 if (Constraint.length() > 1) return;
5733 char Letter = Constraint[0];
5744 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5745 if (!CST) return; // Must be an immediate to match.
5746 unsigned Value = CST->getZExtValue();
5748 default: llvm_unreachable("Unknown constraint letter!");
5749 case 'I': // "I" is a signed 16-bit constant.
5750 if ((short)Value == (int)Value)
5751 Result = DAG.getTargetConstant(Value, Op.getValueType());
5753 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5754 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5755 if ((short)Value == 0)
5756 Result = DAG.getTargetConstant(Value, Op.getValueType());
5758 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5759 if ((Value >> 16) == 0)
5760 Result = DAG.getTargetConstant(Value, Op.getValueType());
5762 case 'M': // "M" is a constant that is greater than 31.
5764 Result = DAG.getTargetConstant(Value, Op.getValueType());
5766 case 'N': // "N" is a positive constant that is an exact power of two.
5767 if ((int)Value > 0 && isPowerOf2_32(Value))
5768 Result = DAG.getTargetConstant(Value, Op.getValueType());
5770 case 'O': // "O" is the constant zero.
5772 Result = DAG.getTargetConstant(Value, Op.getValueType());
5774 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5775 if ((short)-Value == (int)-Value)
5776 Result = DAG.getTargetConstant(Value, Op.getValueType());
5783 if (Result.getNode()) {
5784 Ops.push_back(Result);
5788 // Handle standard constraint letters.
5789 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5792 // isLegalAddressingMode - Return true if the addressing mode represented
5793 // by AM is legal for this target, for a load/store of the specified type.
5794 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5796 // FIXME: PPC does not allow r+i addressing modes for vectors!
5798 // PPC allows a sign-extended 16-bit immediate field.
5799 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5802 // No global is ever allowed as a base.
5806 // PPC only support r+r,
5808 case 0: // "r+i" or just "i", depending on HasBaseReg.
5811 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5813 // Otherwise we have r+r or r+i.
5816 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5818 // Allow 2*r as r+r.
5821 // No other scales are supported.
5828 /// isLegalAddressImmediate - Return true if the integer value can be used
5829 /// as the offset of the target addressing mode for load / store of the
5831 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5832 // PPC allows a sign-extended 16-bit immediate field.
5833 return (V > -(1 << 16) && V < (1 << 16)-1);
5836 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
5840 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5841 SelectionDAG &DAG) const {
5842 MachineFunction &MF = DAG.getMachineFunction();
5843 MachineFrameInfo *MFI = MF.getFrameInfo();
5844 MFI->setReturnAddressIsTaken(true);
5846 DebugLoc dl = Op.getDebugLoc();
5847 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5849 // Make sure the function does not optimize away the store of the RA to
5851 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5852 FuncInfo->setLRStoreRequired();
5853 bool isPPC64 = PPCSubTarget.isPPC64();
5854 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5857 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5860 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5861 isPPC64? MVT::i64 : MVT::i32);
5862 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5863 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5865 MachinePointerInfo(), false, false, false, 0);
5868 // Just load the return address off the stack.
5869 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5870 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5871 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5874 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5875 SelectionDAG &DAG) const {
5876 DebugLoc dl = Op.getDebugLoc();
5877 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5879 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5880 bool isPPC64 = PtrVT == MVT::i64;
5882 MachineFunction &MF = DAG.getMachineFunction();
5883 MachineFrameInfo *MFI = MF.getFrameInfo();
5884 MFI->setFrameAddressIsTaken(true);
5885 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
5886 MFI->hasVarSizedObjects()) &&
5887 MFI->getStackSize() &&
5888 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5889 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5890 (is31 ? PPC::R31 : PPC::R1);
5891 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5894 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5895 FrameAddr, MachinePointerInfo(), false, false,
5901 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5902 // The PowerPC target isn't yet aware of offsets.
5906 /// getOptimalMemOpType - Returns the target specific optimal type for load
5907 /// and store operations as a result of memset, memcpy, and memmove
5908 /// lowering. If DstAlign is zero that means it's safe to destination
5909 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5910 /// means there isn't a need to check it against alignment requirement,
5911 /// probably because the source does not need to be loaded. If
5912 /// 'IsZeroVal' is true, that means it's safe to return a
5913 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5914 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5915 /// constant so it does not need to be loaded.
5916 /// It returns EVT::Other if the type should be determined using generic
5917 /// target-independent logic.
5918 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5919 unsigned DstAlign, unsigned SrcAlign,
5922 MachineFunction &MF) const {
5923 if (this->PPCSubTarget.isPPC64()) {
5930 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
5931 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
5932 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
5933 /// is expanded to mul + add.
5934 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
5938 switch (VT.getSimpleVT().SimpleTy) {
5950 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
5952 return TargetLowering::getSchedulingPreference(N);