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 TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
58 if (TM.getSubtargetImpl()->isDarwin())
59 return new TargetLoweringObjectFileMachO();
61 return new TargetLoweringObjectFileELF();
64 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
65 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
69 // Use _setjmp/_longjmp instead of setjmp/longjmp.
70 setUseUnderscoreSetJmp(true);
71 setUseUnderscoreLongJmp(true);
73 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
74 // arguments are at least 4/8 bytes aligned.
75 setMinStackArgumentAlignment(TM.getSubtarget<PPCSubtarget>().isPPC64() ? 8:4);
77 // Set up the register classes.
78 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
79 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
80 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
82 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
83 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
84 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
86 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
88 // PowerPC has pre-inc load and store's.
89 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
90 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
91 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
93 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
94 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
95 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
96 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
98 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
100 // This is used in the ppcf128->int sequence. Note it has different semantics
101 // from FP_ROUND: that rounds to nearest, this rounds to zero.
102 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
104 // We do not currently implment this libm ops for PowerPC.
105 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
106 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
107 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
108 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
109 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
111 // PowerPC has no SREM/UREM instructions
112 setOperationAction(ISD::SREM, MVT::i32, Expand);
113 setOperationAction(ISD::UREM, MVT::i32, Expand);
114 setOperationAction(ISD::SREM, MVT::i64, Expand);
115 setOperationAction(ISD::UREM, MVT::i64, Expand);
117 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
118 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
119 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
120 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
121 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
122 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
123 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
124 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
125 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
127 // We don't support sin/cos/sqrt/fmod/pow
128 setOperationAction(ISD::FSIN , MVT::f64, Expand);
129 setOperationAction(ISD::FCOS , MVT::f64, Expand);
130 setOperationAction(ISD::FREM , MVT::f64, Expand);
131 setOperationAction(ISD::FPOW , MVT::f64, Expand);
132 setOperationAction(ISD::FMA , MVT::f64, Expand);
133 setOperationAction(ISD::FSIN , MVT::f32, Expand);
134 setOperationAction(ISD::FCOS , MVT::f32, Expand);
135 setOperationAction(ISD::FREM , MVT::f32, Expand);
136 setOperationAction(ISD::FPOW , MVT::f32, Expand);
137 setOperationAction(ISD::FMA , MVT::f32, Expand);
139 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
141 // If we're enabling GP optimizations, use hardware square root
142 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
143 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
144 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
147 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
148 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
150 // PowerPC does not have BSWAP, CTPOP or CTTZ
151 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
152 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
153 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
154 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
155 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
156 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
157 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
158 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
160 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
162 // PowerPC does not have ROTR
163 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
164 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
166 // PowerPC does not have Select
167 setOperationAction(ISD::SELECT, MVT::i32, Expand);
168 setOperationAction(ISD::SELECT, MVT::i64, Expand);
169 setOperationAction(ISD::SELECT, MVT::f32, Expand);
170 setOperationAction(ISD::SELECT, MVT::f64, Expand);
172 // PowerPC wants to turn select_cc of FP into fsel when possible.
173 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
174 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
176 // PowerPC wants to optimize integer setcc a bit
177 setOperationAction(ISD::SETCC, MVT::i32, Custom);
179 // PowerPC does not have BRCOND which requires SetCC
180 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
182 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
184 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
185 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
187 // PowerPC does not have [U|S]INT_TO_FP
188 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
189 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
191 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
192 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
193 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
194 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
196 // We cannot sextinreg(i1). Expand to shifts.
197 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
199 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
200 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
201 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
202 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
205 // We want to legalize GlobalAddress and ConstantPool nodes into the
206 // appropriate instructions to materialize the address.
207 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
208 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
209 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
210 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
211 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
212 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
213 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
214 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
215 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
216 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
219 setOperationAction(ISD::TRAP, MVT::Other, Legal);
221 // TRAMPOLINE is custom lowered.
222 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
223 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
225 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
226 setOperationAction(ISD::VASTART , MVT::Other, Custom);
228 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()) {
229 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
230 // VAARG always uses double-word chunks, so promote anything smaller.
231 setOperationAction(ISD::VAARG, MVT::i1, Promote);
232 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
233 setOperationAction(ISD::VAARG, MVT::i8, Promote);
234 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
235 setOperationAction(ISD::VAARG, MVT::i16, Promote);
236 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
237 setOperationAction(ISD::VAARG, MVT::i32, Promote);
238 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
239 setOperationAction(ISD::VAARG, MVT::Other, Expand);
241 // VAARG is custom lowered with the 32-bit SVR4 ABI.
242 setOperationAction(ISD::VAARG, MVT::Other, Custom);
243 setOperationAction(ISD::VAARG, MVT::i64, Custom);
246 setOperationAction(ISD::VAARG, MVT::Other, Expand);
248 // Use the default implementation.
249 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
250 setOperationAction(ISD::VAEND , MVT::Other, Expand);
251 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
252 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
253 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
254 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
256 // We want to custom lower some of our intrinsics.
257 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
259 // Comparisons that require checking two conditions.
260 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
261 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
262 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
263 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
264 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
265 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
266 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
267 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
268 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
269 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
270 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
271 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
273 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
274 // They also have instructions for converting between i64 and fp.
275 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
276 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
277 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
278 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
279 // This is just the low 32 bits of a (signed) fp->i64 conversion.
280 // We cannot do this with Promote because i64 is not a legal type.
281 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
283 // FIXME: disable this lowered code. This generates 64-bit register values,
284 // and we don't model the fact that the top part is clobbered by calls. We
285 // need to flag these together so that the value isn't live across a call.
286 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
288 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
289 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
292 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
293 // 64-bit PowerPC implementations can support i64 types directly
294 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
295 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
296 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
297 // 64-bit PowerPC wants to expand i128 shifts itself.
298 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
299 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
300 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
302 // 32-bit PowerPC wants to expand i64 shifts itself.
303 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
304 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
305 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
308 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
309 // First set operation action for all vector types to expand. Then we
310 // will selectively turn on ones that can be effectively codegen'd.
311 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
312 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
313 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
315 // add/sub are legal for all supported vector VT's.
316 setOperationAction(ISD::ADD , VT, Legal);
317 setOperationAction(ISD::SUB , VT, Legal);
319 // We promote all shuffles to v16i8.
320 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
321 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
323 // We promote all non-typed operations to v4i32.
324 setOperationAction(ISD::AND , VT, Promote);
325 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
326 setOperationAction(ISD::OR , VT, Promote);
327 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
328 setOperationAction(ISD::XOR , VT, Promote);
329 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
330 setOperationAction(ISD::LOAD , VT, Promote);
331 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
332 setOperationAction(ISD::SELECT, VT, Promote);
333 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
334 setOperationAction(ISD::STORE, VT, Promote);
335 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
337 // No other operations are legal.
338 setOperationAction(ISD::MUL , VT, Expand);
339 setOperationAction(ISD::SDIV, VT, Expand);
340 setOperationAction(ISD::SREM, VT, Expand);
341 setOperationAction(ISD::UDIV, VT, Expand);
342 setOperationAction(ISD::UREM, VT, Expand);
343 setOperationAction(ISD::FDIV, VT, Expand);
344 setOperationAction(ISD::FNEG, VT, Expand);
345 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
346 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
347 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
348 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
349 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
350 setOperationAction(ISD::UDIVREM, VT, Expand);
351 setOperationAction(ISD::SDIVREM, VT, Expand);
352 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
353 setOperationAction(ISD::FPOW, VT, Expand);
354 setOperationAction(ISD::CTPOP, VT, Expand);
355 setOperationAction(ISD::CTLZ, VT, Expand);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
357 setOperationAction(ISD::CTTZ, VT, Expand);
358 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
361 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
362 // with merges, splats, etc.
363 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
365 setOperationAction(ISD::AND , MVT::v4i32, Legal);
366 setOperationAction(ISD::OR , MVT::v4i32, Legal);
367 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
368 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
369 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
370 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
372 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
373 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
374 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
375 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
377 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
378 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
379 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
380 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
382 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
383 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
385 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
386 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
387 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
388 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
391 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport())
392 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
394 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
395 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
397 setBooleanContents(ZeroOrOneBooleanContent);
398 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
400 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
401 setStackPointerRegisterToSaveRestore(PPC::X1);
402 setExceptionPointerRegister(PPC::X3);
403 setExceptionSelectorRegister(PPC::X4);
405 setStackPointerRegisterToSaveRestore(PPC::R1);
406 setExceptionPointerRegister(PPC::R3);
407 setExceptionSelectorRegister(PPC::R4);
410 // We have target-specific dag combine patterns for the following nodes:
411 setTargetDAGCombine(ISD::SINT_TO_FP);
412 setTargetDAGCombine(ISD::STORE);
413 setTargetDAGCombine(ISD::BR_CC);
414 setTargetDAGCombine(ISD::BSWAP);
416 // Darwin long double math library functions have $LDBL128 appended.
417 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
418 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
419 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
420 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
421 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
422 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
423 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
424 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
425 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
426 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
427 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
430 setMinFunctionAlignment(2);
431 if (PPCSubTarget.isDarwin())
432 setPrefFunctionAlignment(4);
434 setInsertFencesForAtomic(true);
436 setSchedulingPreference(Sched::Hybrid);
438 computeRegisterProperties();
441 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
442 /// function arguments in the caller parameter area.
443 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
444 const TargetMachine &TM = getTargetMachine();
445 // Darwin passes everything on 4 byte boundary.
446 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
449 // 16byte and wider vectors are passed on 16byte boundary.
450 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
451 if (VTy->getBitWidth() >= 128)
454 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
455 if (PPCSubTarget.isPPC64())
461 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
464 case PPCISD::FSEL: return "PPCISD::FSEL";
465 case PPCISD::FCFID: return "PPCISD::FCFID";
466 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
467 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
468 case PPCISD::STFIWX: return "PPCISD::STFIWX";
469 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
470 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
471 case PPCISD::VPERM: return "PPCISD::VPERM";
472 case PPCISD::Hi: return "PPCISD::Hi";
473 case PPCISD::Lo: return "PPCISD::Lo";
474 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
475 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
476 case PPCISD::LOAD: return "PPCISD::LOAD";
477 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
478 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
479 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
480 case PPCISD::SRL: return "PPCISD::SRL";
481 case PPCISD::SRA: return "PPCISD::SRA";
482 case PPCISD::SHL: return "PPCISD::SHL";
483 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
484 case PPCISD::STD_32: return "PPCISD::STD_32";
485 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
486 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
487 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
488 case PPCISD::NOP: return "PPCISD::NOP";
489 case PPCISD::MTCTR: return "PPCISD::MTCTR";
490 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
491 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
492 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
493 case PPCISD::MFCR: return "PPCISD::MFCR";
494 case PPCISD::VCMP: return "PPCISD::VCMP";
495 case PPCISD::VCMPo: return "PPCISD::VCMPo";
496 case PPCISD::LBRX: return "PPCISD::LBRX";
497 case PPCISD::STBRX: return "PPCISD::STBRX";
498 case PPCISD::LARX: return "PPCISD::LARX";
499 case PPCISD::STCX: return "PPCISD::STCX";
500 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
501 case PPCISD::MFFS: return "PPCISD::MFFS";
502 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
503 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
504 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
505 case PPCISD::MTFSF: return "PPCISD::MTFSF";
506 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
510 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
514 //===----------------------------------------------------------------------===//
515 // Node matching predicates, for use by the tblgen matching code.
516 //===----------------------------------------------------------------------===//
518 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
519 static bool isFloatingPointZero(SDValue Op) {
520 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
521 return CFP->getValueAPF().isZero();
522 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
523 // Maybe this has already been legalized into the constant pool?
524 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
525 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
526 return CFP->getValueAPF().isZero();
531 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
532 /// true if Op is undef or if it matches the specified value.
533 static bool isConstantOrUndef(int Op, int Val) {
534 return Op < 0 || Op == Val;
537 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
538 /// VPKUHUM instruction.
539 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
541 for (unsigned i = 0; i != 16; ++i)
542 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
545 for (unsigned i = 0; i != 8; ++i)
546 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
547 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
553 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
554 /// VPKUWUM instruction.
555 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
557 for (unsigned i = 0; i != 16; i += 2)
558 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
559 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
562 for (unsigned i = 0; i != 8; i += 2)
563 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
564 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
565 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
566 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
572 /// isVMerge - Common function, used to match vmrg* shuffles.
574 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
575 unsigned LHSStart, unsigned RHSStart) {
576 assert(N->getValueType(0) == MVT::v16i8 &&
577 "PPC only supports shuffles by bytes!");
578 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
579 "Unsupported merge size!");
581 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
582 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
583 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
584 LHSStart+j+i*UnitSize) ||
585 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
586 RHSStart+j+i*UnitSize))
592 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
593 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
594 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
597 return isVMerge(N, UnitSize, 8, 24);
598 return isVMerge(N, UnitSize, 8, 8);
601 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
602 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
603 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
606 return isVMerge(N, UnitSize, 0, 16);
607 return isVMerge(N, UnitSize, 0, 0);
611 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
612 /// amount, otherwise return -1.
613 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
614 assert(N->getValueType(0) == MVT::v16i8 &&
615 "PPC only supports shuffles by bytes!");
617 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
619 // Find the first non-undef value in the shuffle mask.
621 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
624 if (i == 16) return -1; // all undef.
626 // Otherwise, check to see if the rest of the elements are consecutively
627 // numbered from this value.
628 unsigned ShiftAmt = SVOp->getMaskElt(i);
629 if (ShiftAmt < i) return -1;
633 // Check the rest of the elements to see if they are consecutive.
634 for (++i; i != 16; ++i)
635 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
638 // Check the rest of the elements to see if they are consecutive.
639 for (++i; i != 16; ++i)
640 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
646 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
647 /// specifies a splat of a single element that is suitable for input to
648 /// VSPLTB/VSPLTH/VSPLTW.
649 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
650 assert(N->getValueType(0) == MVT::v16i8 &&
651 (EltSize == 1 || EltSize == 2 || EltSize == 4));
653 // This is a splat operation if each element of the permute is the same, and
654 // if the value doesn't reference the second vector.
655 unsigned ElementBase = N->getMaskElt(0);
657 // FIXME: Handle UNDEF elements too!
658 if (ElementBase >= 16)
661 // Check that the indices are consecutive, in the case of a multi-byte element
662 // splatted with a v16i8 mask.
663 for (unsigned i = 1; i != EltSize; ++i)
664 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
667 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
668 if (N->getMaskElt(i) < 0) continue;
669 for (unsigned j = 0; j != EltSize; ++j)
670 if (N->getMaskElt(i+j) != N->getMaskElt(j))
676 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
678 bool PPC::isAllNegativeZeroVector(SDNode *N) {
679 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
681 APInt APVal, APUndef;
685 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
686 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
687 return CFP->getValueAPF().isNegZero();
692 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
693 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
694 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
695 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
696 assert(isSplatShuffleMask(SVOp, EltSize));
697 return SVOp->getMaskElt(0) / EltSize;
700 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
701 /// by using a vspltis[bhw] instruction of the specified element size, return
702 /// the constant being splatted. The ByteSize field indicates the number of
703 /// bytes of each element [124] -> [bhw].
704 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
707 // If ByteSize of the splat is bigger than the element size of the
708 // build_vector, then we have a case where we are checking for a splat where
709 // multiple elements of the buildvector are folded together into a single
710 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
711 unsigned EltSize = 16/N->getNumOperands();
712 if (EltSize < ByteSize) {
713 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
714 SDValue UniquedVals[4];
715 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
717 // See if all of the elements in the buildvector agree across.
718 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
719 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
720 // If the element isn't a constant, bail fully out.
721 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
724 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
725 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
726 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
727 return SDValue(); // no match.
730 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
731 // either constant or undef values that are identical for each chunk. See
732 // if these chunks can form into a larger vspltis*.
734 // Check to see if all of the leading entries are either 0 or -1. If
735 // neither, then this won't fit into the immediate field.
736 bool LeadingZero = true;
737 bool LeadingOnes = true;
738 for (unsigned i = 0; i != Multiple-1; ++i) {
739 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
741 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
742 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
744 // Finally, check the least significant entry.
746 if (UniquedVals[Multiple-1].getNode() == 0)
747 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
748 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
750 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
753 if (UniquedVals[Multiple-1].getNode() == 0)
754 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
755 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
756 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
757 return DAG.getTargetConstant(Val, MVT::i32);
763 // Check to see if this buildvec has a single non-undef value in its elements.
764 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
765 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
766 if (OpVal.getNode() == 0)
767 OpVal = N->getOperand(i);
768 else if (OpVal != N->getOperand(i))
772 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
774 unsigned ValSizeInBytes = EltSize;
776 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
777 Value = CN->getZExtValue();
778 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
779 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
780 Value = FloatToBits(CN->getValueAPF().convertToFloat());
783 // If the splat value is larger than the element value, then we can never do
784 // this splat. The only case that we could fit the replicated bits into our
785 // immediate field for would be zero, and we prefer to use vxor for it.
786 if (ValSizeInBytes < ByteSize) return SDValue();
788 // If the element value is larger than the splat value, cut it in half and
789 // check to see if the two halves are equal. Continue doing this until we
790 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
791 while (ValSizeInBytes > ByteSize) {
792 ValSizeInBytes >>= 1;
794 // If the top half equals the bottom half, we're still ok.
795 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
796 (Value & ((1 << (8*ValSizeInBytes))-1)))
800 // Properly sign extend the value.
801 int ShAmt = (4-ByteSize)*8;
802 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
804 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
805 if (MaskVal == 0) return SDValue();
807 // Finally, if this value fits in a 5 bit sext field, return it
808 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
809 return DAG.getTargetConstant(MaskVal, MVT::i32);
813 //===----------------------------------------------------------------------===//
814 // Addressing Mode Selection
815 //===----------------------------------------------------------------------===//
817 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
818 /// or 64-bit immediate, and if the value can be accurately represented as a
819 /// sign extension from a 16-bit value. If so, this returns true and the
821 static bool isIntS16Immediate(SDNode *N, short &Imm) {
822 if (N->getOpcode() != ISD::Constant)
825 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
826 if (N->getValueType(0) == MVT::i32)
827 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
829 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
831 static bool isIntS16Immediate(SDValue Op, short &Imm) {
832 return isIntS16Immediate(Op.getNode(), Imm);
836 /// SelectAddressRegReg - Given the specified addressed, check to see if it
837 /// can be represented as an indexed [r+r] operation. Returns false if it
838 /// can be more efficiently represented with [r+imm].
839 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
841 SelectionDAG &DAG) const {
843 if (N.getOpcode() == ISD::ADD) {
844 if (isIntS16Immediate(N.getOperand(1), imm))
846 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
849 Base = N.getOperand(0);
850 Index = N.getOperand(1);
852 } else if (N.getOpcode() == ISD::OR) {
853 if (isIntS16Immediate(N.getOperand(1), imm))
854 return false; // r+i can fold it if we can.
856 // If this is an or of disjoint bitfields, we can codegen this as an add
857 // (for better address arithmetic) if the LHS and RHS of the OR are provably
859 APInt LHSKnownZero, LHSKnownOne;
860 APInt RHSKnownZero, RHSKnownOne;
861 DAG.ComputeMaskedBits(N.getOperand(0),
862 LHSKnownZero, LHSKnownOne);
864 if (LHSKnownZero.getBoolValue()) {
865 DAG.ComputeMaskedBits(N.getOperand(1),
866 RHSKnownZero, RHSKnownOne);
867 // If all of the bits are known zero on the LHS or RHS, the add won't
869 if (~(LHSKnownZero | RHSKnownZero) == 0) {
870 Base = N.getOperand(0);
871 Index = N.getOperand(1);
880 /// Returns true if the address N can be represented by a base register plus
881 /// a signed 16-bit displacement [r+imm], and if it is not better
882 /// represented as reg+reg.
883 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
885 SelectionDAG &DAG) const {
886 // FIXME dl should come from parent load or store, not from address
887 DebugLoc dl = N.getDebugLoc();
888 // If this can be more profitably realized as r+r, fail.
889 if (SelectAddressRegReg(N, Disp, Base, DAG))
892 if (N.getOpcode() == ISD::ADD) {
894 if (isIntS16Immediate(N.getOperand(1), imm)) {
895 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
896 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
897 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
899 Base = N.getOperand(0);
901 return true; // [r+i]
902 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
903 // Match LOAD (ADD (X, Lo(G))).
904 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
905 && "Cannot handle constant offsets yet!");
906 Disp = N.getOperand(1).getOperand(0); // The global address.
907 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
908 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
909 Disp.getOpcode() == ISD::TargetConstantPool ||
910 Disp.getOpcode() == ISD::TargetJumpTable);
911 Base = N.getOperand(0);
912 return true; // [&g+r]
914 } else if (N.getOpcode() == ISD::OR) {
916 if (isIntS16Immediate(N.getOperand(1), imm)) {
917 // If this is an or of disjoint bitfields, we can codegen this as an add
918 // (for better address arithmetic) if the LHS and RHS of the OR are
919 // provably disjoint.
920 APInt LHSKnownZero, LHSKnownOne;
921 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
923 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
924 // If all of the bits are known zero on the LHS or RHS, the add won't
926 Base = N.getOperand(0);
927 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
931 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
932 // Loading from a constant address.
934 // If this address fits entirely in a 16-bit sext immediate field, codegen
937 if (isIntS16Immediate(CN, Imm)) {
938 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
939 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
940 CN->getValueType(0));
944 // Handle 32-bit sext immediates with LIS + addr mode.
945 if (CN->getValueType(0) == MVT::i32 ||
946 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
947 int Addr = (int)CN->getZExtValue();
949 // Otherwise, break this down into an LIS + disp.
950 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
952 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
953 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
954 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
959 Disp = DAG.getTargetConstant(0, getPointerTy());
960 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
961 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
964 return true; // [r+0]
967 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
968 /// represented as an indexed [r+r] operation.
969 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
971 SelectionDAG &DAG) const {
972 // Check to see if we can easily represent this as an [r+r] address. This
973 // will fail if it thinks that the address is more profitably represented as
974 // reg+imm, e.g. where imm = 0.
975 if (SelectAddressRegReg(N, Base, Index, DAG))
978 // If the operand is an addition, always emit this as [r+r], since this is
979 // better (for code size, and execution, as the memop does the add for free)
980 // than emitting an explicit add.
981 if (N.getOpcode() == ISD::ADD) {
982 Base = N.getOperand(0);
983 Index = N.getOperand(1);
987 // Otherwise, do it the hard way, using R0 as the base register.
988 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
994 /// SelectAddressRegImmShift - Returns true if the address N can be
995 /// represented by a base register plus a signed 14-bit displacement
996 /// [r+imm*4]. Suitable for use by STD and friends.
997 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
999 SelectionDAG &DAG) const {
1000 // FIXME dl should come from the parent load or store, not the address
1001 DebugLoc dl = N.getDebugLoc();
1002 // If this can be more profitably realized as r+r, fail.
1003 if (SelectAddressRegReg(N, Disp, Base, DAG))
1006 if (N.getOpcode() == ISD::ADD) {
1008 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1009 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1010 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1011 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1013 Base = N.getOperand(0);
1015 return true; // [r+i]
1016 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1017 // Match LOAD (ADD (X, Lo(G))).
1018 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1019 && "Cannot handle constant offsets yet!");
1020 Disp = N.getOperand(1).getOperand(0); // The global address.
1021 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1022 Disp.getOpcode() == ISD::TargetConstantPool ||
1023 Disp.getOpcode() == ISD::TargetJumpTable);
1024 Base = N.getOperand(0);
1025 return true; // [&g+r]
1027 } else if (N.getOpcode() == ISD::OR) {
1029 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1030 // If this is an or of disjoint bitfields, we can codegen this as an add
1031 // (for better address arithmetic) if the LHS and RHS of the OR are
1032 // provably disjoint.
1033 APInt LHSKnownZero, LHSKnownOne;
1034 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1035 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1036 // If all of the bits are known zero on the LHS or RHS, the add won't
1038 Base = N.getOperand(0);
1039 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1043 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1044 // Loading from a constant address. Verify low two bits are clear.
1045 if ((CN->getZExtValue() & 3) == 0) {
1046 // If this address fits entirely in a 14-bit sext immediate field, codegen
1049 if (isIntS16Immediate(CN, Imm)) {
1050 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1051 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1052 CN->getValueType(0));
1056 // Fold the low-part of 32-bit absolute addresses into addr mode.
1057 if (CN->getValueType(0) == MVT::i32 ||
1058 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1059 int Addr = (int)CN->getZExtValue();
1061 // Otherwise, break this down into an LIS + disp.
1062 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1063 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1064 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1065 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1071 Disp = DAG.getTargetConstant(0, getPointerTy());
1072 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1073 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1076 return true; // [r+0]
1080 /// getPreIndexedAddressParts - returns true by value, base pointer and
1081 /// offset pointer and addressing mode by reference if the node's address
1082 /// can be legally represented as pre-indexed load / store address.
1083 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1085 ISD::MemIndexedMode &AM,
1086 SelectionDAG &DAG) const {
1087 if (DisablePPCPreinc) return false;
1091 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1092 Ptr = LD->getBasePtr();
1093 VT = LD->getMemoryVT();
1095 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1096 Ptr = ST->getBasePtr();
1097 VT = ST->getMemoryVT();
1101 // PowerPC doesn't have preinc load/store instructions for vectors.
1105 // TODO: Check reg+reg first.
1107 // LDU/STU use reg+imm*4, others use reg+imm.
1108 if (VT != MVT::i64) {
1110 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1114 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1118 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1119 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1120 // sext i32 to i64 when addr mode is r+i.
1121 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1122 LD->getExtensionType() == ISD::SEXTLOAD &&
1123 isa<ConstantSDNode>(Offset))
1131 //===----------------------------------------------------------------------===//
1132 // LowerOperation implementation
1133 //===----------------------------------------------------------------------===//
1135 /// GetLabelAccessInfo - Return true if we should reference labels using a
1136 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1137 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1138 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1139 HiOpFlags = PPCII::MO_HA16;
1140 LoOpFlags = PPCII::MO_LO16;
1142 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1143 // non-darwin platform. We don't support PIC on other platforms yet.
1144 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1145 TM.getSubtarget<PPCSubtarget>().isDarwin();
1147 HiOpFlags |= PPCII::MO_PIC_FLAG;
1148 LoOpFlags |= PPCII::MO_PIC_FLAG;
1151 // If this is a reference to a global value that requires a non-lazy-ptr, make
1152 // sure that instruction lowering adds it.
1153 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1154 HiOpFlags |= PPCII::MO_NLP_FLAG;
1155 LoOpFlags |= PPCII::MO_NLP_FLAG;
1157 if (GV->hasHiddenVisibility()) {
1158 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1159 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1166 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1167 SelectionDAG &DAG) {
1168 EVT PtrVT = HiPart.getValueType();
1169 SDValue Zero = DAG.getConstant(0, PtrVT);
1170 DebugLoc DL = HiPart.getDebugLoc();
1172 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1173 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1175 // With PIC, the first instruction is actually "GR+hi(&G)".
1177 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1178 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1180 // Generate non-pic code that has direct accesses to the constant pool.
1181 // The address of the global is just (hi(&g)+lo(&g)).
1182 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1185 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1186 SelectionDAG &DAG) const {
1187 EVT PtrVT = Op.getValueType();
1188 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1189 const Constant *C = CP->getConstVal();
1191 unsigned MOHiFlag, MOLoFlag;
1192 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1194 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1196 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1197 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1200 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1201 EVT PtrVT = Op.getValueType();
1202 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1204 unsigned MOHiFlag, MOLoFlag;
1205 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1206 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1207 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1208 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1211 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1212 SelectionDAG &DAG) const {
1213 EVT PtrVT = Op.getValueType();
1215 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1217 unsigned MOHiFlag, MOLoFlag;
1218 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1219 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1220 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1221 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1224 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1225 SelectionDAG &DAG) const {
1227 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1228 DebugLoc dl = GA->getDebugLoc();
1229 const GlobalValue *GV = GA->getGlobal();
1230 EVT PtrVT = getPointerTy();
1231 bool is64bit = PPCSubTarget.isPPC64();
1233 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1235 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1236 PPCII::MO_TPREL16_HA);
1237 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1238 PPCII::MO_TPREL16_LO);
1240 if (model != TLSModel::LocalExec)
1241 llvm_unreachable("only local-exec TLS mode supported");
1242 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1243 is64bit ? MVT::i64 : MVT::i32);
1244 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1245 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1248 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1249 SelectionDAG &DAG) const {
1250 EVT PtrVT = Op.getValueType();
1251 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1252 DebugLoc DL = GSDN->getDebugLoc();
1253 const GlobalValue *GV = GSDN->getGlobal();
1255 // 64-bit SVR4 ABI code is always position-independent.
1256 // The actual address of the GlobalValue is stored in the TOC.
1257 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1258 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1259 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1260 DAG.getRegister(PPC::X2, MVT::i64));
1263 unsigned MOHiFlag, MOLoFlag;
1264 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1267 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1269 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1271 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1273 // If the global reference is actually to a non-lazy-pointer, we have to do an
1274 // extra load to get the address of the global.
1275 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1276 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1277 false, false, false, 0);
1281 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1282 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1283 DebugLoc dl = Op.getDebugLoc();
1285 // If we're comparing for equality to zero, expose the fact that this is
1286 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1287 // fold the new nodes.
1288 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1289 if (C->isNullValue() && CC == ISD::SETEQ) {
1290 EVT VT = Op.getOperand(0).getValueType();
1291 SDValue Zext = Op.getOperand(0);
1292 if (VT.bitsLT(MVT::i32)) {
1294 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1296 unsigned Log2b = Log2_32(VT.getSizeInBits());
1297 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1298 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1299 DAG.getConstant(Log2b, MVT::i32));
1300 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1302 // Leave comparisons against 0 and -1 alone for now, since they're usually
1303 // optimized. FIXME: revisit this when we can custom lower all setcc
1305 if (C->isAllOnesValue() || C->isNullValue())
1309 // If we have an integer seteq/setne, turn it into a compare against zero
1310 // by xor'ing the rhs with the lhs, which is faster than setting a
1311 // condition register, reading it back out, and masking the correct bit. The
1312 // normal approach here uses sub to do this instead of xor. Using xor exposes
1313 // the result to other bit-twiddling opportunities.
1314 EVT LHSVT = Op.getOperand(0).getValueType();
1315 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1316 EVT VT = Op.getValueType();
1317 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1319 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1324 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1325 const PPCSubtarget &Subtarget) const {
1326 SDNode *Node = Op.getNode();
1327 EVT VT = Node->getValueType(0);
1328 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1329 SDValue InChain = Node->getOperand(0);
1330 SDValue VAListPtr = Node->getOperand(1);
1331 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1332 DebugLoc dl = Node->getDebugLoc();
1334 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1337 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1338 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1340 InChain = GprIndex.getValue(1);
1342 if (VT == MVT::i64) {
1343 // Check if GprIndex is even
1344 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1345 DAG.getConstant(1, MVT::i32));
1346 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1347 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1348 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1349 DAG.getConstant(1, MVT::i32));
1350 // Align GprIndex to be even if it isn't
1351 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1355 // fpr index is 1 byte after gpr
1356 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1357 DAG.getConstant(1, MVT::i32));
1360 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1361 FprPtr, MachinePointerInfo(SV), MVT::i8,
1363 InChain = FprIndex.getValue(1);
1365 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1366 DAG.getConstant(8, MVT::i32));
1368 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1369 DAG.getConstant(4, MVT::i32));
1372 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1373 MachinePointerInfo(), false, false,
1375 InChain = OverflowArea.getValue(1);
1377 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1378 MachinePointerInfo(), false, false,
1380 InChain = RegSaveArea.getValue(1);
1382 // select overflow_area if index > 8
1383 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1384 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1386 // adjustment constant gpr_index * 4/8
1387 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1388 VT.isInteger() ? GprIndex : FprIndex,
1389 DAG.getConstant(VT.isInteger() ? 4 : 8,
1392 // OurReg = RegSaveArea + RegConstant
1393 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1396 // Floating types are 32 bytes into RegSaveArea
1397 if (VT.isFloatingPoint())
1398 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1399 DAG.getConstant(32, MVT::i32));
1401 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1402 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1403 VT.isInteger() ? GprIndex : FprIndex,
1404 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1407 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1408 VT.isInteger() ? VAListPtr : FprPtr,
1409 MachinePointerInfo(SV),
1410 MVT::i8, false, false, 0);
1412 // determine if we should load from reg_save_area or overflow_area
1413 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1415 // increase overflow_area by 4/8 if gpr/fpr > 8
1416 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1417 DAG.getConstant(VT.isInteger() ? 4 : 8,
1420 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1423 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1425 MachinePointerInfo(),
1426 MVT::i32, false, false, 0);
1428 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1429 false, false, false, 0);
1432 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1433 SelectionDAG &DAG) const {
1434 return Op.getOperand(0);
1437 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1438 SelectionDAG &DAG) const {
1439 SDValue Chain = Op.getOperand(0);
1440 SDValue Trmp = Op.getOperand(1); // trampoline
1441 SDValue FPtr = Op.getOperand(2); // nested function
1442 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1443 DebugLoc dl = Op.getDebugLoc();
1445 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1446 bool isPPC64 = (PtrVT == MVT::i64);
1448 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1451 TargetLowering::ArgListTy Args;
1452 TargetLowering::ArgListEntry Entry;
1454 Entry.Ty = IntPtrTy;
1455 Entry.Node = Trmp; Args.push_back(Entry);
1457 // TrampSize == (isPPC64 ? 48 : 40);
1458 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1459 isPPC64 ? MVT::i64 : MVT::i32);
1460 Args.push_back(Entry);
1462 Entry.Node = FPtr; Args.push_back(Entry);
1463 Entry.Node = Nest; Args.push_back(Entry);
1465 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1466 TargetLowering::CallLoweringInfo CLI(Chain,
1467 Type::getVoidTy(*DAG.getContext()),
1468 false, false, false, false, 0,
1470 /*isTailCall=*/false,
1471 /*doesNotRet=*/false,
1472 /*isReturnValueUsed=*/true,
1473 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1475 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1477 return CallResult.second;
1480 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1481 const PPCSubtarget &Subtarget) const {
1482 MachineFunction &MF = DAG.getMachineFunction();
1483 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1485 DebugLoc dl = Op.getDebugLoc();
1487 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1488 // vastart just stores the address of the VarArgsFrameIndex slot into the
1489 // memory location argument.
1490 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1491 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1492 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1493 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1494 MachinePointerInfo(SV),
1498 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1499 // We suppose the given va_list is already allocated.
1502 // char gpr; /* index into the array of 8 GPRs
1503 // * stored in the register save area
1504 // * gpr=0 corresponds to r3,
1505 // * gpr=1 to r4, etc.
1507 // char fpr; /* index into the array of 8 FPRs
1508 // * stored in the register save area
1509 // * fpr=0 corresponds to f1,
1510 // * fpr=1 to f2, etc.
1512 // char *overflow_arg_area;
1513 // /* location on stack that holds
1514 // * the next overflow argument
1516 // char *reg_save_area;
1517 // /* where r3:r10 and f1:f8 (if saved)
1523 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1524 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1527 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1529 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1531 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1534 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1535 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1537 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1538 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1540 uint64_t FPROffset = 1;
1541 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1543 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1545 // Store first byte : number of int regs
1546 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1548 MachinePointerInfo(SV),
1549 MVT::i8, false, false, 0);
1550 uint64_t nextOffset = FPROffset;
1551 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1554 // Store second byte : number of float regs
1555 SDValue secondStore =
1556 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1557 MachinePointerInfo(SV, nextOffset), MVT::i8,
1559 nextOffset += StackOffset;
1560 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1562 // Store second word : arguments given on stack
1563 SDValue thirdStore =
1564 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1565 MachinePointerInfo(SV, nextOffset),
1567 nextOffset += FrameOffset;
1568 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1570 // Store third word : arguments given in registers
1571 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1572 MachinePointerInfo(SV, nextOffset),
1577 #include "PPCGenCallingConv.inc"
1579 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1580 CCValAssign::LocInfo &LocInfo,
1581 ISD::ArgFlagsTy &ArgFlags,
1586 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1588 CCValAssign::LocInfo &LocInfo,
1589 ISD::ArgFlagsTy &ArgFlags,
1591 static const uint16_t ArgRegs[] = {
1592 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1593 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1595 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1597 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1599 // Skip one register if the first unallocated register has an even register
1600 // number and there are still argument registers available which have not been
1601 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1602 // need to skip a register if RegNum is odd.
1603 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1604 State.AllocateReg(ArgRegs[RegNum]);
1607 // Always return false here, as this function only makes sure that the first
1608 // unallocated register has an odd register number and does not actually
1609 // allocate a register for the current argument.
1613 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1615 CCValAssign::LocInfo &LocInfo,
1616 ISD::ArgFlagsTy &ArgFlags,
1618 static const uint16_t ArgRegs[] = {
1619 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1623 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1625 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1627 // If there is only one Floating-point register left we need to put both f64
1628 // values of a split ppc_fp128 value on the stack.
1629 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1630 State.AllocateReg(ArgRegs[RegNum]);
1633 // Always return false here, as this function only makes sure that the two f64
1634 // values a ppc_fp128 value is split into are both passed in registers or both
1635 // passed on the stack and does not actually allocate a register for the
1636 // current argument.
1640 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1642 static const uint16_t *GetFPR() {
1643 static const uint16_t FPR[] = {
1644 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1645 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1651 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1653 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1654 unsigned PtrByteSize) {
1655 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1656 if (Flags.isByVal())
1657 ArgSize = Flags.getByValSize();
1658 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1664 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1665 CallingConv::ID CallConv, bool isVarArg,
1666 const SmallVectorImpl<ISD::InputArg>
1668 DebugLoc dl, SelectionDAG &DAG,
1669 SmallVectorImpl<SDValue> &InVals)
1671 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1672 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1675 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1681 PPCTargetLowering::LowerFormalArguments_SVR4(
1683 CallingConv::ID CallConv, bool isVarArg,
1684 const SmallVectorImpl<ISD::InputArg>
1686 DebugLoc dl, SelectionDAG &DAG,
1687 SmallVectorImpl<SDValue> &InVals) const {
1689 // 32-bit SVR4 ABI Stack Frame Layout:
1690 // +-----------------------------------+
1691 // +--> | Back chain |
1692 // | +-----------------------------------+
1693 // | | Floating-point register save area |
1694 // | +-----------------------------------+
1695 // | | General register save area |
1696 // | +-----------------------------------+
1697 // | | CR save word |
1698 // | +-----------------------------------+
1699 // | | VRSAVE save word |
1700 // | +-----------------------------------+
1701 // | | Alignment padding |
1702 // | +-----------------------------------+
1703 // | | Vector register save area |
1704 // | +-----------------------------------+
1705 // | | Local variable space |
1706 // | +-----------------------------------+
1707 // | | Parameter list area |
1708 // | +-----------------------------------+
1709 // | | LR save word |
1710 // | +-----------------------------------+
1711 // SP--> +--- | Back chain |
1712 // +-----------------------------------+
1715 // System V Application Binary Interface PowerPC Processor Supplement
1716 // AltiVec Technology Programming Interface Manual
1718 MachineFunction &MF = DAG.getMachineFunction();
1719 MachineFrameInfo *MFI = MF.getFrameInfo();
1720 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1722 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1723 // Potential tail calls could cause overwriting of argument stack slots.
1724 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1725 (CallConv == CallingConv::Fast));
1726 unsigned PtrByteSize = 4;
1728 // Assign locations to all of the incoming arguments.
1729 SmallVector<CCValAssign, 16> ArgLocs;
1730 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1731 getTargetMachine(), ArgLocs, *DAG.getContext());
1733 // Reserve space for the linkage area on the stack.
1734 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1736 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1738 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1739 CCValAssign &VA = ArgLocs[i];
1741 // Arguments stored in registers.
1742 if (VA.isRegLoc()) {
1743 const TargetRegisterClass *RC;
1744 EVT ValVT = VA.getValVT();
1746 switch (ValVT.getSimpleVT().SimpleTy) {
1748 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1750 RC = &PPC::GPRCRegClass;
1753 RC = &PPC::F4RCRegClass;
1756 RC = &PPC::F8RCRegClass;
1762 RC = &PPC::VRRCRegClass;
1766 // Transform the arguments stored in physical registers into virtual ones.
1767 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1768 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1770 InVals.push_back(ArgValue);
1772 // Argument stored in memory.
1773 assert(VA.isMemLoc());
1775 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1776 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1779 // Create load nodes to retrieve arguments from the stack.
1780 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1781 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1782 MachinePointerInfo(),
1783 false, false, false, 0));
1787 // Assign locations to all of the incoming aggregate by value arguments.
1788 // Aggregates passed by value are stored in the local variable space of the
1789 // caller's stack frame, right above the parameter list area.
1790 SmallVector<CCValAssign, 16> ByValArgLocs;
1791 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1792 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1794 // Reserve stack space for the allocations in CCInfo.
1795 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1797 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1799 // Area that is at least reserved in the caller of this function.
1800 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1802 // Set the size that is at least reserved in caller of this function. Tail
1803 // call optimized function's reserved stack space needs to be aligned so that
1804 // taking the difference between two stack areas will result in an aligned
1806 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1809 std::max(MinReservedArea,
1810 PPCFrameLowering::getMinCallFrameSize(false, false));
1812 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1813 getStackAlignment();
1814 unsigned AlignMask = TargetAlign-1;
1815 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1817 FI->setMinReservedArea(MinReservedArea);
1819 SmallVector<SDValue, 8> MemOps;
1821 // If the function takes variable number of arguments, make a frame index for
1822 // the start of the first vararg value... for expansion of llvm.va_start.
1824 static const uint16_t GPArgRegs[] = {
1825 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1826 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1828 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1830 static const uint16_t FPArgRegs[] = {
1831 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1834 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1836 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1838 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1841 // Make room for NumGPArgRegs and NumFPArgRegs.
1842 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1843 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1845 FuncInfo->setVarArgsStackOffset(
1846 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1847 CCInfo.getNextStackOffset(), true));
1849 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1850 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1852 // The fixed integer arguments of a variadic function are stored to the
1853 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1854 // the result of va_next.
1855 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1856 // Get an existing live-in vreg, or add a new one.
1857 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1859 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1861 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1862 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1863 MachinePointerInfo(), false, false, 0);
1864 MemOps.push_back(Store);
1865 // Increment the address by four for the next argument to store
1866 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1867 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1870 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1872 // The double arguments are stored to the VarArgsFrameIndex
1874 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1875 // Get an existing live-in vreg, or add a new one.
1876 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1878 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1880 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1881 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1882 MachinePointerInfo(), false, false, 0);
1883 MemOps.push_back(Store);
1884 // Increment the address by eight for the next argument to store
1885 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1887 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1891 if (!MemOps.empty())
1892 Chain = DAG.getNode(ISD::TokenFactor, dl,
1893 MVT::Other, &MemOps[0], MemOps.size());
1899 PPCTargetLowering::LowerFormalArguments_Darwin(
1901 CallingConv::ID CallConv, bool isVarArg,
1902 const SmallVectorImpl<ISD::InputArg>
1904 DebugLoc dl, SelectionDAG &DAG,
1905 SmallVectorImpl<SDValue> &InVals) const {
1906 // TODO: add description of PPC stack frame format, or at least some docs.
1908 MachineFunction &MF = DAG.getMachineFunction();
1909 MachineFrameInfo *MFI = MF.getFrameInfo();
1910 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1912 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1913 bool isPPC64 = PtrVT == MVT::i64;
1914 // Potential tail calls could cause overwriting of argument stack slots.
1915 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1916 (CallConv == CallingConv::Fast));
1917 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1919 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1920 // Area that is at least reserved in caller of this function.
1921 unsigned MinReservedArea = ArgOffset;
1923 static const uint16_t GPR_32[] = { // 32-bit registers.
1924 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1925 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1927 static const uint16_t GPR_64[] = { // 64-bit registers.
1928 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1929 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1932 static const uint16_t *FPR = GetFPR();
1934 static const uint16_t VR[] = {
1935 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1936 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1939 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1940 const unsigned Num_FPR_Regs = 13;
1941 const unsigned Num_VR_Regs = array_lengthof( VR);
1943 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1945 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
1947 // In 32-bit non-varargs functions, the stack space for vectors is after the
1948 // stack space for non-vectors. We do not use this space unless we have
1949 // too many vectors to fit in registers, something that only occurs in
1950 // constructed examples:), but we have to walk the arglist to figure
1951 // that out...for the pathological case, compute VecArgOffset as the
1952 // start of the vector parameter area. Computing VecArgOffset is the
1953 // entire point of the following loop.
1954 unsigned VecArgOffset = ArgOffset;
1955 if (!isVarArg && !isPPC64) {
1956 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1958 EVT ObjectVT = Ins[ArgNo].VT;
1959 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1961 if (Flags.isByVal()) {
1962 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1963 unsigned ObjSize = Flags.getByValSize();
1965 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1966 VecArgOffset += ArgSize;
1970 switch(ObjectVT.getSimpleVT().SimpleTy) {
1971 default: llvm_unreachable("Unhandled argument type!");
1974 VecArgOffset += isPPC64 ? 8 : 4;
1976 case MVT::i64: // PPC64
1984 // Nothing to do, we're only looking at Nonvector args here.
1989 // We've found where the vector parameter area in memory is. Skip the
1990 // first 12 parameters; these don't use that memory.
1991 VecArgOffset = ((VecArgOffset+15)/16)*16;
1992 VecArgOffset += 12*16;
1994 // Add DAG nodes to load the arguments or copy them out of registers. On
1995 // entry to a function on PPC, the arguments start after the linkage area,
1996 // although the first ones are often in registers.
1998 SmallVector<SDValue, 8> MemOps;
1999 unsigned nAltivecParamsAtEnd = 0;
2000 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2002 bool needsLoad = false;
2003 EVT ObjectVT = Ins[ArgNo].VT;
2004 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2005 unsigned ArgSize = ObjSize;
2006 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2008 unsigned CurArgOffset = ArgOffset;
2010 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2011 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2012 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2013 if (isVarArg || isPPC64) {
2014 MinReservedArea = ((MinReservedArea+15)/16)*16;
2015 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2018 } else nAltivecParamsAtEnd++;
2020 // Calculate min reserved area.
2021 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2025 // FIXME the codegen can be much improved in some cases.
2026 // We do not have to keep everything in memory.
2027 if (Flags.isByVal()) {
2028 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2029 ObjSize = Flags.getByValSize();
2030 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2031 // Objects of size 1 and 2 are right justified, everything else is
2032 // left justified. This means the memory address is adjusted forwards.
2033 if (ObjSize==1 || ObjSize==2) {
2034 CurArgOffset = CurArgOffset + (4 - ObjSize);
2036 // The value of the object is its address.
2037 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2038 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2039 InVals.push_back(FIN);
2040 if (ObjSize==1 || ObjSize==2) {
2041 if (GPR_idx != Num_GPR_Regs) {
2044 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2046 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2047 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2048 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2049 MachinePointerInfo(),
2050 ObjSize==1 ? MVT::i8 : MVT::i16,
2052 MemOps.push_back(Store);
2056 ArgOffset += PtrByteSize;
2060 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2061 // Store whatever pieces of the object are in registers
2062 // to memory. ArgVal will be address of the beginning of
2064 if (GPR_idx != Num_GPR_Regs) {
2067 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2069 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2070 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2071 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2072 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2073 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2074 MachinePointerInfo(),
2076 MemOps.push_back(Store);
2078 ArgOffset += PtrByteSize;
2080 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2087 switch (ObjectVT.getSimpleVT().SimpleTy) {
2088 default: llvm_unreachable("Unhandled argument type!");
2091 if (GPR_idx != Num_GPR_Regs) {
2092 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2093 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2097 ArgSize = PtrByteSize;
2099 // All int arguments reserve stack space in the Darwin ABI.
2100 ArgOffset += PtrByteSize;
2104 case MVT::i64: // PPC64
2105 if (GPR_idx != Num_GPR_Regs) {
2106 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2107 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2109 if (ObjectVT == MVT::i32) {
2110 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2111 // value to MVT::i64 and then truncate to the correct register size.
2113 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2114 DAG.getValueType(ObjectVT));
2115 else if (Flags.isZExt())
2116 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2117 DAG.getValueType(ObjectVT));
2119 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2125 ArgSize = PtrByteSize;
2127 // All int arguments reserve stack space in the Darwin ABI.
2133 // Every 4 bytes of argument space consumes one of the GPRs available for
2134 // argument passing.
2135 if (GPR_idx != Num_GPR_Regs) {
2137 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2140 if (FPR_idx != Num_FPR_Regs) {
2143 if (ObjectVT == MVT::f32)
2144 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2146 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2148 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2154 // All FP arguments reserve stack space in the Darwin ABI.
2155 ArgOffset += isPPC64 ? 8 : ObjSize;
2161 // Note that vector arguments in registers don't reserve stack space,
2162 // except in varargs functions.
2163 if (VR_idx != Num_VR_Regs) {
2164 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2165 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2167 while ((ArgOffset % 16) != 0) {
2168 ArgOffset += PtrByteSize;
2169 if (GPR_idx != Num_GPR_Regs)
2173 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2177 if (!isVarArg && !isPPC64) {
2178 // Vectors go after all the nonvectors.
2179 CurArgOffset = VecArgOffset;
2182 // Vectors are aligned.
2183 ArgOffset = ((ArgOffset+15)/16)*16;
2184 CurArgOffset = ArgOffset;
2192 // We need to load the argument to a virtual register if we determined above
2193 // that we ran out of physical registers of the appropriate type.
2195 int FI = MFI->CreateFixedObject(ObjSize,
2196 CurArgOffset + (ArgSize - ObjSize),
2198 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2199 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2200 false, false, false, 0);
2203 InVals.push_back(ArgVal);
2206 // Set the size that is at least reserved in caller of this function. Tail
2207 // call optimized function's reserved stack space needs to be aligned so that
2208 // taking the difference between two stack areas will result in an aligned
2210 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2211 // Add the Altivec parameters at the end, if needed.
2212 if (nAltivecParamsAtEnd) {
2213 MinReservedArea = ((MinReservedArea+15)/16)*16;
2214 MinReservedArea += 16*nAltivecParamsAtEnd;
2217 std::max(MinReservedArea,
2218 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2219 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2220 getStackAlignment();
2221 unsigned AlignMask = TargetAlign-1;
2222 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2223 FI->setMinReservedArea(MinReservedArea);
2225 // If the function takes variable number of arguments, make a frame index for
2226 // the start of the first vararg value... for expansion of llvm.va_start.
2228 int Depth = ArgOffset;
2230 FuncInfo->setVarArgsFrameIndex(
2231 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2233 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2235 // If this function is vararg, store any remaining integer argument regs
2236 // to their spots on the stack so that they may be loaded by deferencing the
2237 // result of va_next.
2238 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2242 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2244 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2246 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2247 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2248 MachinePointerInfo(), false, false, 0);
2249 MemOps.push_back(Store);
2250 // Increment the address by four for the next argument to store
2251 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2252 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2256 if (!MemOps.empty())
2257 Chain = DAG.getNode(ISD::TokenFactor, dl,
2258 MVT::Other, &MemOps[0], MemOps.size());
2263 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2264 /// linkage area for the Darwin ABI.
2266 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2270 const SmallVectorImpl<ISD::OutputArg>
2272 const SmallVectorImpl<SDValue> &OutVals,
2273 unsigned &nAltivecParamsAtEnd) {
2274 // Count how many bytes are to be pushed on the stack, including the linkage
2275 // area, and parameter passing area. We start with 24/48 bytes, which is
2276 // prereserved space for [SP][CR][LR][3 x unused].
2277 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2278 unsigned NumOps = Outs.size();
2279 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2281 // Add up all the space actually used.
2282 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2283 // they all go in registers, but we must reserve stack space for them for
2284 // possible use by the caller. In varargs or 64-bit calls, parameters are
2285 // assigned stack space in order, with padding so Altivec parameters are
2287 nAltivecParamsAtEnd = 0;
2288 for (unsigned i = 0; i != NumOps; ++i) {
2289 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2290 EVT ArgVT = Outs[i].VT;
2291 // Varargs Altivec parameters are padded to a 16 byte boundary.
2292 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2293 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2294 if (!isVarArg && !isPPC64) {
2295 // Non-varargs Altivec parameters go after all the non-Altivec
2296 // parameters; handle those later so we know how much padding we need.
2297 nAltivecParamsAtEnd++;
2300 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2301 NumBytes = ((NumBytes+15)/16)*16;
2303 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2306 // Allow for Altivec parameters at the end, if needed.
2307 if (nAltivecParamsAtEnd) {
2308 NumBytes = ((NumBytes+15)/16)*16;
2309 NumBytes += 16*nAltivecParamsAtEnd;
2312 // The prolog code of the callee may store up to 8 GPR argument registers to
2313 // the stack, allowing va_start to index over them in memory if its varargs.
2314 // Because we cannot tell if this is needed on the caller side, we have to
2315 // conservatively assume that it is needed. As such, make sure we have at
2316 // least enough stack space for the caller to store the 8 GPRs.
2317 NumBytes = std::max(NumBytes,
2318 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2320 // Tail call needs the stack to be aligned.
2321 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2322 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2323 getFrameLowering()->getStackAlignment();
2324 unsigned AlignMask = TargetAlign-1;
2325 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2331 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2332 /// adjusted to accommodate the arguments for the tailcall.
2333 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2334 unsigned ParamSize) {
2336 if (!isTailCall) return 0;
2338 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2339 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2340 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2341 // Remember only if the new adjustement is bigger.
2342 if (SPDiff < FI->getTailCallSPDelta())
2343 FI->setTailCallSPDelta(SPDiff);
2348 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2349 /// for tail call optimization. Targets which want to do tail call
2350 /// optimization should implement this function.
2352 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2353 CallingConv::ID CalleeCC,
2355 const SmallVectorImpl<ISD::InputArg> &Ins,
2356 SelectionDAG& DAG) const {
2357 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2360 // Variable argument functions are not supported.
2364 MachineFunction &MF = DAG.getMachineFunction();
2365 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2366 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2367 // Functions containing by val parameters are not supported.
2368 for (unsigned i = 0; i != Ins.size(); i++) {
2369 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2370 if (Flags.isByVal()) return false;
2373 // Non PIC/GOT tail calls are supported.
2374 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2377 // At the moment we can only do local tail calls (in same module, hidden
2378 // or protected) if we are generating PIC.
2379 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2380 return G->getGlobal()->hasHiddenVisibility()
2381 || G->getGlobal()->hasProtectedVisibility();
2387 /// isCallCompatibleAddress - Return the immediate to use if the specified
2388 /// 32-bit value is representable in the immediate field of a BxA instruction.
2389 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2390 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2393 int Addr = C->getZExtValue();
2394 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2395 (Addr << 6 >> 6) != Addr)
2396 return 0; // Top 6 bits have to be sext of immediate.
2398 return DAG.getConstant((int)C->getZExtValue() >> 2,
2399 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2404 struct TailCallArgumentInfo {
2409 TailCallArgumentInfo() : FrameIdx(0) {}
2414 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2416 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2418 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2419 SmallVector<SDValue, 8> &MemOpChains,
2421 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2422 SDValue Arg = TailCallArgs[i].Arg;
2423 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2424 int FI = TailCallArgs[i].FrameIdx;
2425 // Store relative to framepointer.
2426 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2427 MachinePointerInfo::getFixedStack(FI),
2432 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2433 /// the appropriate stack slot for the tail call optimized function call.
2434 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2435 MachineFunction &MF,
2444 // Calculate the new stack slot for the return address.
2445 int SlotSize = isPPC64 ? 8 : 4;
2446 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2448 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2449 NewRetAddrLoc, true);
2450 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2451 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2452 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2453 MachinePointerInfo::getFixedStack(NewRetAddr),
2456 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2457 // slot as the FP is never overwritten.
2460 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2461 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2463 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2464 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2465 MachinePointerInfo::getFixedStack(NewFPIdx),
2472 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2473 /// the position of the argument.
2475 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2476 SDValue Arg, int SPDiff, unsigned ArgOffset,
2477 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2478 int Offset = ArgOffset + SPDiff;
2479 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2480 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2481 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2482 SDValue FIN = DAG.getFrameIndex(FI, VT);
2483 TailCallArgumentInfo Info;
2485 Info.FrameIdxOp = FIN;
2487 TailCallArguments.push_back(Info);
2490 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2491 /// stack slot. Returns the chain as result and the loaded frame pointers in
2492 /// LROpOut/FPOpout. Used when tail calling.
2493 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2499 DebugLoc dl) const {
2501 // Load the LR and FP stack slot for later adjusting.
2502 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2503 LROpOut = getReturnAddrFrameIndex(DAG);
2504 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2505 false, false, false, 0);
2506 Chain = SDValue(LROpOut.getNode(), 1);
2508 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2509 // slot as the FP is never overwritten.
2511 FPOpOut = getFramePointerFrameIndex(DAG);
2512 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2513 false, false, false, 0);
2514 Chain = SDValue(FPOpOut.getNode(), 1);
2520 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2521 /// by "Src" to address "Dst" of size "Size". Alignment information is
2522 /// specified by the specific parameter attribute. The copy will be passed as
2523 /// a byval function parameter.
2524 /// Sometimes what we are copying is the end of a larger object, the part that
2525 /// does not fit in registers.
2527 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2528 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2530 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2531 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2532 false, false, MachinePointerInfo(0),
2533 MachinePointerInfo(0));
2536 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2539 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2540 SDValue Arg, SDValue PtrOff, int SPDiff,
2541 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2542 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2543 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2545 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2550 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2552 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2553 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2554 DAG.getConstant(ArgOffset, PtrVT));
2556 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2557 MachinePointerInfo(), false, false, 0));
2558 // Calculate and remember argument location.
2559 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2564 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2565 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2566 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2567 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2568 MachineFunction &MF = DAG.getMachineFunction();
2570 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2571 // might overwrite each other in case of tail call optimization.
2572 SmallVector<SDValue, 8> MemOpChains2;
2573 // Do not flag preceding copytoreg stuff together with the following stuff.
2575 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2577 if (!MemOpChains2.empty())
2578 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2579 &MemOpChains2[0], MemOpChains2.size());
2581 // Store the return address to the appropriate stack slot.
2582 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2583 isPPC64, isDarwinABI, dl);
2585 // Emit callseq_end just before tailcall node.
2586 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2587 DAG.getIntPtrConstant(0, true), InFlag);
2588 InFlag = Chain.getValue(1);
2592 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2593 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2594 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2595 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2596 const PPCSubtarget &PPCSubTarget) {
2598 bool isPPC64 = PPCSubTarget.isPPC64();
2599 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2601 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2602 NodeTys.push_back(MVT::Other); // Returns a chain
2603 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2605 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2607 bool needIndirectCall = true;
2608 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2609 // If this is an absolute destination address, use the munged value.
2610 Callee = SDValue(Dest, 0);
2611 needIndirectCall = false;
2614 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2615 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2616 // Use indirect calls for ALL functions calls in JIT mode, since the
2617 // far-call stubs may be outside relocation limits for a BL instruction.
2618 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2619 unsigned OpFlags = 0;
2620 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2621 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2622 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2623 (G->getGlobal()->isDeclaration() ||
2624 G->getGlobal()->isWeakForLinker())) {
2625 // PC-relative references to external symbols should go through $stub,
2626 // unless we're building with the leopard linker or later, which
2627 // automatically synthesizes these stubs.
2628 OpFlags = PPCII::MO_DARWIN_STUB;
2631 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2632 // every direct call is) turn it into a TargetGlobalAddress /
2633 // TargetExternalSymbol node so that legalize doesn't hack it.
2634 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2635 Callee.getValueType(),
2637 needIndirectCall = false;
2641 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2642 unsigned char OpFlags = 0;
2644 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2645 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2646 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2647 // PC-relative references to external symbols should go through $stub,
2648 // unless we're building with the leopard linker or later, which
2649 // automatically synthesizes these stubs.
2650 OpFlags = PPCII::MO_DARWIN_STUB;
2653 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2655 needIndirectCall = false;
2658 if (needIndirectCall) {
2659 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2660 // to do the call, we can't use PPCISD::CALL.
2661 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2663 if (isSVR4ABI && isPPC64) {
2664 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2665 // entry point, but to the function descriptor (the function entry point
2666 // address is part of the function descriptor though).
2667 // The function descriptor is a three doubleword structure with the
2668 // following fields: function entry point, TOC base address and
2669 // environment pointer.
2670 // Thus for a call through a function pointer, the following actions need
2672 // 1. Save the TOC of the caller in the TOC save area of its stack
2673 // frame (this is done in LowerCall_Darwin()).
2674 // 2. Load the address of the function entry point from the function
2676 // 3. Load the TOC of the callee from the function descriptor into r2.
2677 // 4. Load the environment pointer from the function descriptor into
2679 // 5. Branch to the function entry point address.
2680 // 6. On return of the callee, the TOC of the caller needs to be
2681 // restored (this is done in FinishCall()).
2683 // All those operations are flagged together to ensure that no other
2684 // operations can be scheduled in between. E.g. without flagging the
2685 // operations together, a TOC access in the caller could be scheduled
2686 // between the load of the callee TOC and the branch to the callee, which
2687 // results in the TOC access going through the TOC of the callee instead
2688 // of going through the TOC of the caller, which leads to incorrect code.
2690 // Load the address of the function entry point from the function
2692 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2693 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2694 InFlag.getNode() ? 3 : 2);
2695 Chain = LoadFuncPtr.getValue(1);
2696 InFlag = LoadFuncPtr.getValue(2);
2698 // Load environment pointer into r11.
2699 // Offset of the environment pointer within the function descriptor.
2700 SDValue PtrOff = DAG.getIntPtrConstant(16);
2702 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2703 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2705 Chain = LoadEnvPtr.getValue(1);
2706 InFlag = LoadEnvPtr.getValue(2);
2708 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2710 Chain = EnvVal.getValue(0);
2711 InFlag = EnvVal.getValue(1);
2713 // Load TOC of the callee into r2. We are using a target-specific load
2714 // with r2 hard coded, because the result of a target-independent load
2715 // would never go directly into r2, since r2 is a reserved register (which
2716 // prevents the register allocator from allocating it), resulting in an
2717 // additional register being allocated and an unnecessary move instruction
2719 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2720 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2722 Chain = LoadTOCPtr.getValue(0);
2723 InFlag = LoadTOCPtr.getValue(1);
2725 MTCTROps[0] = Chain;
2726 MTCTROps[1] = LoadFuncPtr;
2727 MTCTROps[2] = InFlag;
2730 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2731 2 + (InFlag.getNode() != 0));
2732 InFlag = Chain.getValue(1);
2735 NodeTys.push_back(MVT::Other);
2736 NodeTys.push_back(MVT::Glue);
2737 Ops.push_back(Chain);
2738 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2740 // Add CTR register as callee so a bctr can be emitted later.
2742 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2745 // If this is a direct call, pass the chain and the callee.
2746 if (Callee.getNode()) {
2747 Ops.push_back(Chain);
2748 Ops.push_back(Callee);
2750 // If this is a tail call add stack pointer delta.
2752 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2754 // Add argument registers to the end of the list so that they are known live
2756 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2757 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2758 RegsToPass[i].second.getValueType()));
2764 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2765 CallingConv::ID CallConv, bool isVarArg,
2766 const SmallVectorImpl<ISD::InputArg> &Ins,
2767 DebugLoc dl, SelectionDAG &DAG,
2768 SmallVectorImpl<SDValue> &InVals) const {
2770 SmallVector<CCValAssign, 16> RVLocs;
2771 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2772 getTargetMachine(), RVLocs, *DAG.getContext());
2773 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2775 // Copy all of the result registers out of their specified physreg.
2776 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2777 CCValAssign &VA = RVLocs[i];
2778 EVT VT = VA.getValVT();
2779 assert(VA.isRegLoc() && "Can only return in registers!");
2780 Chain = DAG.getCopyFromReg(Chain, dl,
2781 VA.getLocReg(), VT, InFlag).getValue(1);
2782 InVals.push_back(Chain.getValue(0));
2783 InFlag = Chain.getValue(2);
2790 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2791 bool isTailCall, bool isVarArg,
2793 SmallVector<std::pair<unsigned, SDValue>, 8>
2795 SDValue InFlag, SDValue Chain,
2797 int SPDiff, unsigned NumBytes,
2798 const SmallVectorImpl<ISD::InputArg> &Ins,
2799 SmallVectorImpl<SDValue> &InVals) const {
2800 std::vector<EVT> NodeTys;
2801 SmallVector<SDValue, 8> Ops;
2802 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2803 isTailCall, RegsToPass, Ops, NodeTys,
2806 // When performing tail call optimization the callee pops its arguments off
2807 // the stack. Account for this here so these bytes can be pushed back on in
2808 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2809 int BytesCalleePops =
2810 (CallConv == CallingConv::Fast &&
2811 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
2813 // Add a register mask operand representing the call-preserved registers.
2814 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2815 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2816 assert(Mask && "Missing call preserved mask for calling convention");
2817 Ops.push_back(DAG.getRegisterMask(Mask));
2819 if (InFlag.getNode())
2820 Ops.push_back(InFlag);
2824 // If this is the first return lowered for this function, add the regs
2825 // to the liveout set for the function.
2826 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2827 SmallVector<CCValAssign, 16> RVLocs;
2828 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2829 getTargetMachine(), RVLocs, *DAG.getContext());
2830 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2831 for (unsigned i = 0; i != RVLocs.size(); ++i)
2832 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2835 assert(((Callee.getOpcode() == ISD::Register &&
2836 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2837 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2838 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2839 isa<ConstantSDNode>(Callee)) &&
2840 "Expecting an global address, external symbol, absolute value or register");
2842 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2845 // Add a NOP immediately after the branch instruction when using the 64-bit
2846 // SVR4 ABI. At link time, if caller and callee are in a different module and
2847 // thus have a different TOC, the call will be replaced with a call to a stub
2848 // function which saves the current TOC, loads the TOC of the callee and
2849 // branches to the callee. The NOP will be replaced with a load instruction
2850 // which restores the TOC of the caller from the TOC save slot of the current
2851 // stack frame. If caller and callee belong to the same module (and have the
2852 // same TOC), the NOP will remain unchanged.
2854 bool needsTOCRestore = false;
2855 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2856 if (CallOpc == PPCISD::BCTRL_SVR4) {
2857 // This is a call through a function pointer.
2858 // Restore the caller TOC from the save area into R2.
2859 // See PrepareCall() for more information about calls through function
2860 // pointers in the 64-bit SVR4 ABI.
2861 // We are using a target-specific load with r2 hard coded, because the
2862 // result of a target-independent load would never go directly into r2,
2863 // since r2 is a reserved register (which prevents the register allocator
2864 // from allocating it), resulting in an additional register being
2865 // allocated and an unnecessary move instruction being generated.
2866 needsTOCRestore = true;
2867 } else if (CallOpc == PPCISD::CALL_SVR4) {
2868 // Otherwise insert NOP.
2869 CallOpc = PPCISD::CALL_NOP_SVR4;
2873 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2874 InFlag = Chain.getValue(1);
2876 if (needsTOCRestore) {
2877 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2878 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2879 InFlag = Chain.getValue(1);
2882 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2883 DAG.getIntPtrConstant(BytesCalleePops, true),
2886 InFlag = Chain.getValue(1);
2888 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2889 Ins, dl, DAG, InVals);
2893 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2894 SmallVectorImpl<SDValue> &InVals) const {
2895 SelectionDAG &DAG = CLI.DAG;
2896 DebugLoc &dl = CLI.DL;
2897 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2898 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2899 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2900 SDValue Chain = CLI.Chain;
2901 SDValue Callee = CLI.Callee;
2902 bool &isTailCall = CLI.IsTailCall;
2903 CallingConv::ID CallConv = CLI.CallConv;
2904 bool isVarArg = CLI.IsVarArg;
2907 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2910 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2911 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2912 isTailCall, Outs, OutVals, Ins,
2915 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2916 isTailCall, Outs, OutVals, Ins,
2921 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2922 CallingConv::ID CallConv, bool isVarArg,
2924 const SmallVectorImpl<ISD::OutputArg> &Outs,
2925 const SmallVectorImpl<SDValue> &OutVals,
2926 const SmallVectorImpl<ISD::InputArg> &Ins,
2927 DebugLoc dl, SelectionDAG &DAG,
2928 SmallVectorImpl<SDValue> &InVals) const {
2929 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2930 // of the 32-bit SVR4 ABI stack frame layout.
2932 assert((CallConv == CallingConv::C ||
2933 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2935 unsigned PtrByteSize = 4;
2937 MachineFunction &MF = DAG.getMachineFunction();
2939 // Mark this function as potentially containing a function that contains a
2940 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2941 // and restoring the callers stack pointer in this functions epilog. This is
2942 // done because by tail calling the called function might overwrite the value
2943 // in this function's (MF) stack pointer stack slot 0(SP).
2944 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2945 CallConv == CallingConv::Fast)
2946 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2948 // Count how many bytes are to be pushed on the stack, including the linkage
2949 // area, parameter list area and the part of the local variable space which
2950 // contains copies of aggregates which are passed by value.
2952 // Assign locations to all of the outgoing arguments.
2953 SmallVector<CCValAssign, 16> ArgLocs;
2954 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2955 getTargetMachine(), ArgLocs, *DAG.getContext());
2957 // Reserve space for the linkage area on the stack.
2958 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2961 // Handle fixed and variable vector arguments differently.
2962 // Fixed vector arguments go into registers as long as registers are
2963 // available. Variable vector arguments always go into memory.
2964 unsigned NumArgs = Outs.size();
2966 for (unsigned i = 0; i != NumArgs; ++i) {
2967 MVT ArgVT = Outs[i].VT;
2968 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2971 if (Outs[i].IsFixed) {
2972 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2975 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2981 errs() << "Call operand #" << i << " has unhandled type "
2982 << EVT(ArgVT).getEVTString() << "\n";
2984 llvm_unreachable(0);
2988 // All arguments are treated the same.
2989 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2992 // Assign locations to all of the outgoing aggregate by value arguments.
2993 SmallVector<CCValAssign, 16> ByValArgLocs;
2994 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2995 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2997 // Reserve stack space for the allocations in CCInfo.
2998 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3000 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3002 // Size of the linkage area, parameter list area and the part of the local
3003 // space variable where copies of aggregates which are passed by value are
3005 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3007 // Calculate by how many bytes the stack has to be adjusted in case of tail
3008 // call optimization.
3009 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3011 // Adjust the stack pointer for the new arguments...
3012 // These operations are automatically eliminated by the prolog/epilog pass
3013 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3014 SDValue CallSeqStart = Chain;
3016 // Load the return address and frame pointer so it can be moved somewhere else
3019 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3022 // Set up a copy of the stack pointer for use loading and storing any
3023 // arguments that may not fit in the registers available for argument
3025 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3027 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3028 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3029 SmallVector<SDValue, 8> MemOpChains;
3031 bool seenFloatArg = false;
3032 // Walk the register/memloc assignments, inserting copies/loads.
3033 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3036 CCValAssign &VA = ArgLocs[i];
3037 SDValue Arg = OutVals[i];
3038 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3040 if (Flags.isByVal()) {
3041 // Argument is an aggregate which is passed by value, thus we need to
3042 // create a copy of it in the local variable space of the current stack
3043 // frame (which is the stack frame of the caller) and pass the address of
3044 // this copy to the callee.
3045 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3046 CCValAssign &ByValVA = ByValArgLocs[j++];
3047 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3049 // Memory reserved in the local variable space of the callers stack frame.
3050 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3052 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3053 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3055 // Create a copy of the argument in the local area of the current
3057 SDValue MemcpyCall =
3058 CreateCopyOfByValArgument(Arg, PtrOff,
3059 CallSeqStart.getNode()->getOperand(0),
3062 // This must go outside the CALLSEQ_START..END.
3063 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3064 CallSeqStart.getNode()->getOperand(1));
3065 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3066 NewCallSeqStart.getNode());
3067 Chain = CallSeqStart = NewCallSeqStart;
3069 // Pass the address of the aggregate copy on the stack either in a
3070 // physical register or in the parameter list area of the current stack
3071 // frame to the callee.
3075 if (VA.isRegLoc()) {
3076 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3077 // Put argument in a physical register.
3078 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3080 // Put argument in the parameter list area of the current stack frame.
3081 assert(VA.isMemLoc());
3082 unsigned LocMemOffset = VA.getLocMemOffset();
3085 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3086 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3088 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3089 MachinePointerInfo(),
3092 // Calculate and remember argument location.
3093 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3099 if (!MemOpChains.empty())
3100 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3101 &MemOpChains[0], MemOpChains.size());
3103 // Set CR6 to true if this is a vararg call with floating args passed in
3106 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3108 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3111 // Build a sequence of copy-to-reg nodes chained together with token chain
3112 // and flag operands which copy the outgoing args into the appropriate regs.
3114 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3115 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3116 RegsToPass[i].second, InFlag);
3117 InFlag = Chain.getValue(1);
3121 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3122 false, TailCallArguments);
3124 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3125 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3130 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3131 CallingConv::ID CallConv, bool isVarArg,
3133 const SmallVectorImpl<ISD::OutputArg> &Outs,
3134 const SmallVectorImpl<SDValue> &OutVals,
3135 const SmallVectorImpl<ISD::InputArg> &Ins,
3136 DebugLoc dl, SelectionDAG &DAG,
3137 SmallVectorImpl<SDValue> &InVals) const {
3139 unsigned NumOps = Outs.size();
3141 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3142 bool isPPC64 = PtrVT == MVT::i64;
3143 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3145 MachineFunction &MF = DAG.getMachineFunction();
3147 // Mark this function as potentially containing a function that contains a
3148 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3149 // and restoring the callers stack pointer in this functions epilog. This is
3150 // done because by tail calling the called function might overwrite the value
3151 // in this function's (MF) stack pointer stack slot 0(SP).
3152 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3153 CallConv == CallingConv::Fast)
3154 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3156 unsigned nAltivecParamsAtEnd = 0;
3158 // Count how many bytes are to be pushed on the stack, including the linkage
3159 // area, and parameter passing area. We start with 24/48 bytes, which is
3160 // prereserved space for [SP][CR][LR][3 x unused].
3162 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3164 nAltivecParamsAtEnd);
3166 // Calculate by how many bytes the stack has to be adjusted in case of tail
3167 // call optimization.
3168 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3170 // To protect arguments on the stack from being clobbered in a tail call,
3171 // force all the loads to happen before doing any other lowering.
3173 Chain = DAG.getStackArgumentTokenFactor(Chain);
3175 // Adjust the stack pointer for the new arguments...
3176 // These operations are automatically eliminated by the prolog/epilog pass
3177 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3178 SDValue CallSeqStart = Chain;
3180 // Load the return address and frame pointer so it can be move somewhere else
3183 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3186 // Set up a copy of the stack pointer for use loading and storing any
3187 // arguments that may not fit in the registers available for argument
3191 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3193 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3195 // Figure out which arguments are going to go in registers, and which in
3196 // memory. Also, if this is a vararg function, floating point operations
3197 // must be stored to our stack, and loaded into integer regs as well, if
3198 // any integer regs are available for argument passing.
3199 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3200 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3202 static const uint16_t GPR_32[] = { // 32-bit registers.
3203 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3204 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3206 static const uint16_t GPR_64[] = { // 64-bit registers.
3207 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3208 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3210 static const uint16_t *FPR = GetFPR();
3212 static const uint16_t VR[] = {
3213 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3214 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3216 const unsigned NumGPRs = array_lengthof(GPR_32);
3217 const unsigned NumFPRs = 13;
3218 const unsigned NumVRs = array_lengthof(VR);
3220 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3222 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3223 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3225 SmallVector<SDValue, 8> MemOpChains;
3226 for (unsigned i = 0; i != NumOps; ++i) {
3227 SDValue Arg = OutVals[i];
3228 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3230 // PtrOff will be used to store the current argument to the stack if a
3231 // register cannot be found for it.
3234 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3236 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3238 // On PPC64, promote integers to 64-bit values.
3239 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3240 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3241 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3242 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3245 // FIXME memcpy is used way more than necessary. Correctness first.
3246 if (Flags.isByVal()) {
3247 unsigned Size = Flags.getByValSize();
3248 if (Size==1 || Size==2) {
3249 // Very small objects are passed right-justified.
3250 // Everything else is passed left-justified.
3251 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3252 if (GPR_idx != NumGPRs) {
3253 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3254 MachinePointerInfo(), VT,
3256 MemOpChains.push_back(Load.getValue(1));
3257 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3259 ArgOffset += PtrByteSize;
3261 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3262 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3263 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3264 CallSeqStart.getNode()->getOperand(0),
3266 // This must go outside the CALLSEQ_START..END.
3267 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3268 CallSeqStart.getNode()->getOperand(1));
3269 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3270 NewCallSeqStart.getNode());
3271 Chain = CallSeqStart = NewCallSeqStart;
3272 ArgOffset += PtrByteSize;
3276 // Copy entire object into memory. There are cases where gcc-generated
3277 // code assumes it is there, even if it could be put entirely into
3278 // registers. (This is not what the doc says.)
3279 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3280 CallSeqStart.getNode()->getOperand(0),
3282 // This must go outside the CALLSEQ_START..END.
3283 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3284 CallSeqStart.getNode()->getOperand(1));
3285 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3286 Chain = CallSeqStart = NewCallSeqStart;
3287 // And copy the pieces of it that fit into registers.
3288 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3289 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3290 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3291 if (GPR_idx != NumGPRs) {
3292 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3293 MachinePointerInfo(),
3294 false, false, false, 0);
3295 MemOpChains.push_back(Load.getValue(1));
3296 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3297 ArgOffset += PtrByteSize;
3299 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3306 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3307 default: llvm_unreachable("Unexpected ValueType for argument!");
3310 if (GPR_idx != NumGPRs) {
3311 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3313 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3314 isPPC64, isTailCall, false, MemOpChains,
3315 TailCallArguments, dl);
3317 ArgOffset += PtrByteSize;
3321 if (FPR_idx != NumFPRs) {
3322 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3325 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3326 MachinePointerInfo(), false, false, 0);
3327 MemOpChains.push_back(Store);
3329 // Float varargs are always shadowed in available integer registers
3330 if (GPR_idx != NumGPRs) {
3331 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3332 MachinePointerInfo(), false, false,
3334 MemOpChains.push_back(Load.getValue(1));
3335 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3337 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3338 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3339 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3340 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3341 MachinePointerInfo(),
3342 false, false, false, 0);
3343 MemOpChains.push_back(Load.getValue(1));
3344 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3347 // If we have any FPRs remaining, we may also have GPRs remaining.
3348 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3350 if (GPR_idx != NumGPRs)
3352 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3353 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3357 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3358 isPPC64, isTailCall, false, MemOpChains,
3359 TailCallArguments, dl);
3364 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3371 // These go aligned on the stack, or in the corresponding R registers
3372 // when within range. The Darwin PPC ABI doc claims they also go in
3373 // V registers; in fact gcc does this only for arguments that are
3374 // prototyped, not for those that match the ... We do it for all
3375 // arguments, seems to work.
3376 while (ArgOffset % 16 !=0) {
3377 ArgOffset += PtrByteSize;
3378 if (GPR_idx != NumGPRs)
3381 // We could elide this store in the case where the object fits
3382 // entirely in R registers. Maybe later.
3383 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3384 DAG.getConstant(ArgOffset, PtrVT));
3385 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3386 MachinePointerInfo(), false, false, 0);
3387 MemOpChains.push_back(Store);
3388 if (VR_idx != NumVRs) {
3389 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3390 MachinePointerInfo(),
3391 false, false, false, 0);
3392 MemOpChains.push_back(Load.getValue(1));
3393 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3396 for (unsigned i=0; i<16; i+=PtrByteSize) {
3397 if (GPR_idx == NumGPRs)
3399 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3400 DAG.getConstant(i, PtrVT));
3401 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3402 false, false, false, 0);
3403 MemOpChains.push_back(Load.getValue(1));
3404 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3409 // Non-varargs Altivec params generally go in registers, but have
3410 // stack space allocated at the end.
3411 if (VR_idx != NumVRs) {
3412 // Doesn't have GPR space allocated.
3413 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3414 } else if (nAltivecParamsAtEnd==0) {
3415 // We are emitting Altivec params in order.
3416 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3417 isPPC64, isTailCall, true, MemOpChains,
3418 TailCallArguments, dl);
3424 // If all Altivec parameters fit in registers, as they usually do,
3425 // they get stack space following the non-Altivec parameters. We
3426 // don't track this here because nobody below needs it.
3427 // If there are more Altivec parameters than fit in registers emit
3429 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3431 // Offset is aligned; skip 1st 12 params which go in V registers.
3432 ArgOffset = ((ArgOffset+15)/16)*16;
3434 for (unsigned i = 0; i != NumOps; ++i) {
3435 SDValue Arg = OutVals[i];
3436 EVT ArgType = Outs[i].VT;
3437 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3438 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3441 // We are emitting Altivec params in order.
3442 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3443 isPPC64, isTailCall, true, MemOpChains,
3444 TailCallArguments, dl);
3451 if (!MemOpChains.empty())
3452 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3453 &MemOpChains[0], MemOpChains.size());
3455 // Check if this is an indirect call (MTCTR/BCTRL).
3456 // See PrepareCall() for more information about calls through function
3457 // pointers in the 64-bit SVR4 ABI.
3458 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3459 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3460 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3461 !isBLACompatibleAddress(Callee, DAG)) {
3462 // Load r2 into a virtual register and store it to the TOC save area.
3463 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3464 // TOC save area offset.
3465 SDValue PtrOff = DAG.getIntPtrConstant(40);
3466 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3467 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3471 // On Darwin, R12 must contain the address of an indirect callee. This does
3472 // not mean the MTCTR instruction must use R12; it's easier to model this as
3473 // an extra parameter, so do that.
3475 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3476 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3477 !isBLACompatibleAddress(Callee, DAG))
3478 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3479 PPC::R12), Callee));
3481 // Build a sequence of copy-to-reg nodes chained together with token chain
3482 // and flag operands which copy the outgoing args into the appropriate regs.
3484 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3485 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3486 RegsToPass[i].second, InFlag);
3487 InFlag = Chain.getValue(1);
3491 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3492 FPOp, true, TailCallArguments);
3494 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3495 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3500 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3501 MachineFunction &MF, bool isVarArg,
3502 const SmallVectorImpl<ISD::OutputArg> &Outs,
3503 LLVMContext &Context) const {
3504 SmallVector<CCValAssign, 16> RVLocs;
3505 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3507 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3511 PPCTargetLowering::LowerReturn(SDValue Chain,
3512 CallingConv::ID CallConv, bool isVarArg,
3513 const SmallVectorImpl<ISD::OutputArg> &Outs,
3514 const SmallVectorImpl<SDValue> &OutVals,
3515 DebugLoc dl, SelectionDAG &DAG) const {
3517 SmallVector<CCValAssign, 16> RVLocs;
3518 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3519 getTargetMachine(), RVLocs, *DAG.getContext());
3520 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3522 // If this is the first return lowered for this function, add the regs to the
3523 // liveout set for the function.
3524 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3525 for (unsigned i = 0; i != RVLocs.size(); ++i)
3526 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3531 // Copy the result values into the output registers.
3532 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3533 CCValAssign &VA = RVLocs[i];
3534 assert(VA.isRegLoc() && "Can only return in registers!");
3535 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3537 Flag = Chain.getValue(1);
3541 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3543 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3546 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3547 const PPCSubtarget &Subtarget) const {
3548 // When we pop the dynamic allocation we need to restore the SP link.
3549 DebugLoc dl = Op.getDebugLoc();
3551 // Get the corect type for pointers.
3552 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3554 // Construct the stack pointer operand.
3555 bool isPPC64 = Subtarget.isPPC64();
3556 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3557 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3559 // Get the operands for the STACKRESTORE.
3560 SDValue Chain = Op.getOperand(0);
3561 SDValue SaveSP = Op.getOperand(1);
3563 // Load the old link SP.
3564 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3565 MachinePointerInfo(),
3566 false, false, false, 0);
3568 // Restore the stack pointer.
3569 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3571 // Store the old link SP.
3572 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3579 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3580 MachineFunction &MF = DAG.getMachineFunction();
3581 bool isPPC64 = PPCSubTarget.isPPC64();
3582 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3583 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3585 // Get current frame pointer save index. The users of this index will be
3586 // primarily DYNALLOC instructions.
3587 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3588 int RASI = FI->getReturnAddrSaveIndex();
3590 // If the frame pointer save index hasn't been defined yet.
3592 // Find out what the fix offset of the frame pointer save area.
3593 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3594 // Allocate the frame index for frame pointer save area.
3595 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3597 FI->setReturnAddrSaveIndex(RASI);
3599 return DAG.getFrameIndex(RASI, PtrVT);
3603 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3604 MachineFunction &MF = DAG.getMachineFunction();
3605 bool isPPC64 = PPCSubTarget.isPPC64();
3606 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3607 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3609 // Get current frame pointer save index. The users of this index will be
3610 // primarily DYNALLOC instructions.
3611 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3612 int FPSI = FI->getFramePointerSaveIndex();
3614 // If the frame pointer save index hasn't been defined yet.
3616 // Find out what the fix offset of the frame pointer save area.
3617 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3620 // Allocate the frame index for frame pointer save area.
3621 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3623 FI->setFramePointerSaveIndex(FPSI);
3625 return DAG.getFrameIndex(FPSI, PtrVT);
3628 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3630 const PPCSubtarget &Subtarget) const {
3632 SDValue Chain = Op.getOperand(0);
3633 SDValue Size = Op.getOperand(1);
3634 DebugLoc dl = Op.getDebugLoc();
3636 // Get the corect type for pointers.
3637 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3639 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3640 DAG.getConstant(0, PtrVT), Size);
3641 // Construct a node for the frame pointer save index.
3642 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3643 // Build a DYNALLOC node.
3644 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3645 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3646 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3649 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3651 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3652 // Not FP? Not a fsel.
3653 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3654 !Op.getOperand(2).getValueType().isFloatingPoint())
3657 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3659 // Cannot handle SETEQ/SETNE.
3660 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3662 EVT ResVT = Op.getValueType();
3663 EVT CmpVT = Op.getOperand(0).getValueType();
3664 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3665 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3666 DebugLoc dl = Op.getDebugLoc();
3668 // If the RHS of the comparison is a 0.0, we don't need to do the
3669 // subtraction at all.
3670 if (isFloatingPointZero(RHS))
3672 default: break; // SETUO etc aren't handled by fsel.
3675 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3678 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3679 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3680 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3683 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3686 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3687 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3688 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3689 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3694 default: break; // SETUO etc aren't handled by fsel.
3697 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3698 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3699 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3700 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3703 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3704 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3705 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3706 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3709 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3710 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3711 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3712 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3715 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3716 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3717 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3718 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3723 // FIXME: Split this code up when LegalizeDAGTypes lands.
3724 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3725 DebugLoc dl) const {
3726 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3727 SDValue Src = Op.getOperand(0);
3728 if (Src.getValueType() == MVT::f32)
3729 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3732 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3733 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3735 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3740 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3744 // Convert the FP value to an int value through memory.
3745 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3747 // Emit a store to the stack slot.
3748 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3749 MachinePointerInfo(), false, false, 0);
3751 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3753 if (Op.getValueType() == MVT::i32)
3754 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3755 DAG.getConstant(4, FIPtr.getValueType()));
3756 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3757 false, false, false, 0);
3760 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3761 SelectionDAG &DAG) const {
3762 DebugLoc dl = Op.getDebugLoc();
3763 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3764 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3767 if (Op.getOperand(0).getValueType() == MVT::i64) {
3768 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3769 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3770 if (Op.getValueType() == MVT::f32)
3771 FP = DAG.getNode(ISD::FP_ROUND, dl,
3772 MVT::f32, FP, DAG.getIntPtrConstant(0));
3776 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3777 "Unhandled SINT_TO_FP type in custom expander!");
3778 // Since we only generate this in 64-bit mode, we can take advantage of
3779 // 64-bit registers. In particular, sign extend the input value into the
3780 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3781 // then lfd it and fcfid it.
3782 MachineFunction &MF = DAG.getMachineFunction();
3783 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3784 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3785 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3786 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3788 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3791 // STD the extended value into the stack slot.
3792 MachineMemOperand *MMO =
3793 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3794 MachineMemOperand::MOStore, 8, 8);
3795 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3797 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3798 Ops, 4, MVT::i64, MMO);
3799 // Load the value as a double.
3800 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3801 false, false, false, 0);
3803 // FCFID it and return it.
3804 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3805 if (Op.getValueType() == MVT::f32)
3806 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3810 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3811 SelectionDAG &DAG) const {
3812 DebugLoc dl = Op.getDebugLoc();
3814 The rounding mode is in bits 30:31 of FPSR, and has the following
3821 FLT_ROUNDS, on the other hand, expects the following:
3828 To perform the conversion, we do:
3829 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3832 MachineFunction &MF = DAG.getMachineFunction();
3833 EVT VT = Op.getValueType();
3834 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3835 std::vector<EVT> NodeTys;
3836 SDValue MFFSreg, InFlag;
3838 // Save FP Control Word to register
3839 NodeTys.push_back(MVT::f64); // return register
3840 NodeTys.push_back(MVT::Glue); // unused in this context
3841 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3843 // Save FP register to stack slot
3844 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3845 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3846 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3847 StackSlot, MachinePointerInfo(), false, false,0);
3849 // Load FP Control Word from low 32 bits of stack slot.
3850 SDValue Four = DAG.getConstant(4, PtrVT);
3851 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3852 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3853 false, false, false, 0);
3855 // Transform as necessary
3857 DAG.getNode(ISD::AND, dl, MVT::i32,
3858 CWD, DAG.getConstant(3, MVT::i32));
3860 DAG.getNode(ISD::SRL, dl, MVT::i32,
3861 DAG.getNode(ISD::AND, dl, MVT::i32,
3862 DAG.getNode(ISD::XOR, dl, MVT::i32,
3863 CWD, DAG.getConstant(3, MVT::i32)),
3864 DAG.getConstant(3, MVT::i32)),
3865 DAG.getConstant(1, MVT::i32));
3868 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3870 return DAG.getNode((VT.getSizeInBits() < 16 ?
3871 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3874 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3875 EVT VT = Op.getValueType();
3876 unsigned BitWidth = VT.getSizeInBits();
3877 DebugLoc dl = Op.getDebugLoc();
3878 assert(Op.getNumOperands() == 3 &&
3879 VT == Op.getOperand(1).getValueType() &&
3882 // Expand into a bunch of logical ops. Note that these ops
3883 // depend on the PPC behavior for oversized shift amounts.
3884 SDValue Lo = Op.getOperand(0);
3885 SDValue Hi = Op.getOperand(1);
3886 SDValue Amt = Op.getOperand(2);
3887 EVT AmtVT = Amt.getValueType();
3889 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3890 DAG.getConstant(BitWidth, AmtVT), Amt);
3891 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3892 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3893 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3894 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3895 DAG.getConstant(-BitWidth, AmtVT));
3896 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3897 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3898 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3899 SDValue OutOps[] = { OutLo, OutHi };
3900 return DAG.getMergeValues(OutOps, 2, dl);
3903 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3904 EVT VT = Op.getValueType();
3905 DebugLoc dl = Op.getDebugLoc();
3906 unsigned BitWidth = VT.getSizeInBits();
3907 assert(Op.getNumOperands() == 3 &&
3908 VT == Op.getOperand(1).getValueType() &&
3911 // Expand into a bunch of logical ops. Note that these ops
3912 // depend on the PPC behavior for oversized shift amounts.
3913 SDValue Lo = Op.getOperand(0);
3914 SDValue Hi = Op.getOperand(1);
3915 SDValue Amt = Op.getOperand(2);
3916 EVT AmtVT = Amt.getValueType();
3918 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3919 DAG.getConstant(BitWidth, AmtVT), Amt);
3920 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3921 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3922 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3923 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3924 DAG.getConstant(-BitWidth, AmtVT));
3925 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3926 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3927 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3928 SDValue OutOps[] = { OutLo, OutHi };
3929 return DAG.getMergeValues(OutOps, 2, dl);
3932 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3933 DebugLoc dl = Op.getDebugLoc();
3934 EVT VT = Op.getValueType();
3935 unsigned BitWidth = VT.getSizeInBits();
3936 assert(Op.getNumOperands() == 3 &&
3937 VT == Op.getOperand(1).getValueType() &&
3940 // Expand into a bunch of logical ops, followed by a select_cc.
3941 SDValue Lo = Op.getOperand(0);
3942 SDValue Hi = Op.getOperand(1);
3943 SDValue Amt = Op.getOperand(2);
3944 EVT AmtVT = Amt.getValueType();
3946 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3947 DAG.getConstant(BitWidth, AmtVT), Amt);
3948 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3949 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3950 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3951 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3952 DAG.getConstant(-BitWidth, AmtVT));
3953 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3954 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3955 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3956 Tmp4, Tmp6, ISD::SETLE);
3957 SDValue OutOps[] = { OutLo, OutHi };
3958 return DAG.getMergeValues(OutOps, 2, dl);
3961 //===----------------------------------------------------------------------===//
3962 // Vector related lowering.
3965 /// BuildSplatI - Build a canonical splati of Val with an element size of
3966 /// SplatSize. Cast the result to VT.
3967 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3968 SelectionDAG &DAG, DebugLoc dl) {
3969 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3971 static const EVT VTys[] = { // canonical VT to use for each size.
3972 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3975 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3977 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3981 EVT CanonicalVT = VTys[SplatSize-1];
3983 // Build a canonical splat for this value.
3984 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3985 SmallVector<SDValue, 8> Ops;
3986 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3987 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3988 &Ops[0], Ops.size());
3989 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3992 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3993 /// specified intrinsic ID.
3994 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3995 SelectionDAG &DAG, DebugLoc dl,
3996 EVT DestVT = MVT::Other) {
3997 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3998 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3999 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4002 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4003 /// specified intrinsic ID.
4004 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4005 SDValue Op2, SelectionDAG &DAG,
4006 DebugLoc dl, EVT DestVT = MVT::Other) {
4007 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4008 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4009 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4013 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4014 /// amount. The result has the specified value type.
4015 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4016 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4017 // Force LHS/RHS to be the right type.
4018 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4019 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4022 for (unsigned i = 0; i != 16; ++i)
4024 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4025 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4028 // If this is a case we can't handle, return null and let the default
4029 // expansion code take care of it. If we CAN select this case, and if it
4030 // selects to a single instruction, return Op. Otherwise, if we can codegen
4031 // this case more efficiently than a constant pool load, lower it to the
4032 // sequence of ops that should be used.
4033 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4034 SelectionDAG &DAG) const {
4035 DebugLoc dl = Op.getDebugLoc();
4036 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4037 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4039 // Check if this is a splat of a constant value.
4040 APInt APSplatBits, APSplatUndef;
4041 unsigned SplatBitSize;
4043 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4044 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4047 unsigned SplatBits = APSplatBits.getZExtValue();
4048 unsigned SplatUndef = APSplatUndef.getZExtValue();
4049 unsigned SplatSize = SplatBitSize / 8;
4051 // First, handle single instruction cases.
4054 if (SplatBits == 0) {
4055 // Canonicalize all zero vectors to be v4i32.
4056 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4057 SDValue Z = DAG.getConstant(0, MVT::i32);
4058 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4059 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4064 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4065 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4067 if (SextVal >= -16 && SextVal <= 15)
4068 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4071 // Two instruction sequences.
4073 // If this value is in the range [-32,30] and is even, use:
4074 // tmp = VSPLTI[bhw], result = add tmp, tmp
4075 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4076 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4077 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4078 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4081 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4082 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4084 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4085 // Make -1 and vspltisw -1:
4086 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4088 // Make the VSLW intrinsic, computing 0x8000_0000.
4089 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4092 // xor by OnesV to invert it.
4093 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4094 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4097 // Check to see if this is a wide variety of vsplti*, binop self cases.
4098 static const signed char SplatCsts[] = {
4099 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4100 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4103 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4104 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4105 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4106 int i = SplatCsts[idx];
4108 // Figure out what shift amount will be used by altivec if shifted by i in
4110 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4112 // vsplti + shl self.
4113 if (SextVal == (i << (int)TypeShiftAmt)) {
4114 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4115 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4116 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4117 Intrinsic::ppc_altivec_vslw
4119 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4120 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4123 // vsplti + srl self.
4124 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4125 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4126 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4127 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4128 Intrinsic::ppc_altivec_vsrw
4130 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4131 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4134 // vsplti + sra self.
4135 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4136 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4137 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4138 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4139 Intrinsic::ppc_altivec_vsraw
4141 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4142 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4145 // vsplti + rol self.
4146 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4147 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4148 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4149 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4150 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4151 Intrinsic::ppc_altivec_vrlw
4153 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4154 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4157 // t = vsplti c, result = vsldoi t, t, 1
4158 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4159 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4160 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4162 // t = vsplti c, result = vsldoi t, t, 2
4163 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4164 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4165 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4167 // t = vsplti c, result = vsldoi t, t, 3
4168 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4169 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4170 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4174 // Three instruction sequences.
4176 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4177 if (SextVal >= 0 && SextVal <= 31) {
4178 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4179 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4180 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4181 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4183 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4184 if (SextVal >= -31 && SextVal <= 0) {
4185 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4186 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4187 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4188 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4194 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4195 /// the specified operations to build the shuffle.
4196 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4197 SDValue RHS, SelectionDAG &DAG,
4199 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4200 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4201 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4204 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4216 if (OpNum == OP_COPY) {
4217 if (LHSID == (1*9+2)*9+3) return LHS;
4218 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4222 SDValue OpLHS, OpRHS;
4223 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4224 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4228 default: llvm_unreachable("Unknown i32 permute!");
4230 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4231 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4232 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4233 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4236 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4237 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4238 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4239 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4242 for (unsigned i = 0; i != 16; ++i)
4243 ShufIdxs[i] = (i&3)+0;
4246 for (unsigned i = 0; i != 16; ++i)
4247 ShufIdxs[i] = (i&3)+4;
4250 for (unsigned i = 0; i != 16; ++i)
4251 ShufIdxs[i] = (i&3)+8;
4254 for (unsigned i = 0; i != 16; ++i)
4255 ShufIdxs[i] = (i&3)+12;
4258 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4260 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4262 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4264 EVT VT = OpLHS.getValueType();
4265 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4266 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4267 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4268 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4271 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4272 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4273 /// return the code it can be lowered into. Worst case, it can always be
4274 /// lowered into a vperm.
4275 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4276 SelectionDAG &DAG) const {
4277 DebugLoc dl = Op.getDebugLoc();
4278 SDValue V1 = Op.getOperand(0);
4279 SDValue V2 = Op.getOperand(1);
4280 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4281 EVT VT = Op.getValueType();
4283 // Cases that are handled by instructions that take permute immediates
4284 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4285 // selected by the instruction selector.
4286 if (V2.getOpcode() == ISD::UNDEF) {
4287 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4288 PPC::isSplatShuffleMask(SVOp, 2) ||
4289 PPC::isSplatShuffleMask(SVOp, 4) ||
4290 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4291 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4292 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4293 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4294 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4295 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4296 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4297 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4298 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4303 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4304 // and produce a fixed permutation. If any of these match, do not lower to
4306 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4307 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4308 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4309 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4310 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4311 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4312 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4313 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4314 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4317 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4318 // perfect shuffle table to emit an optimal matching sequence.
4319 ArrayRef<int> PermMask = SVOp->getMask();
4321 unsigned PFIndexes[4];
4322 bool isFourElementShuffle = true;
4323 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4324 unsigned EltNo = 8; // Start out undef.
4325 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4326 if (PermMask[i*4+j] < 0)
4327 continue; // Undef, ignore it.
4329 unsigned ByteSource = PermMask[i*4+j];
4330 if ((ByteSource & 3) != j) {
4331 isFourElementShuffle = false;
4336 EltNo = ByteSource/4;
4337 } else if (EltNo != ByteSource/4) {
4338 isFourElementShuffle = false;
4342 PFIndexes[i] = EltNo;
4345 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4346 // perfect shuffle vector to determine if it is cost effective to do this as
4347 // discrete instructions, or whether we should use a vperm.
4348 if (isFourElementShuffle) {
4349 // Compute the index in the perfect shuffle table.
4350 unsigned PFTableIndex =
4351 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4353 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4354 unsigned Cost = (PFEntry >> 30);
4356 // Determining when to avoid vperm is tricky. Many things affect the cost
4357 // of vperm, particularly how many times the perm mask needs to be computed.
4358 // For example, if the perm mask can be hoisted out of a loop or is already
4359 // used (perhaps because there are multiple permutes with the same shuffle
4360 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4361 // the loop requires an extra register.
4363 // As a compromise, we only emit discrete instructions if the shuffle can be
4364 // generated in 3 or fewer operations. When we have loop information
4365 // available, if this block is within a loop, we should avoid using vperm
4366 // for 3-operation perms and use a constant pool load instead.
4368 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4371 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4372 // vector that will get spilled to the constant pool.
4373 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4375 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4376 // that it is in input element units, not in bytes. Convert now.
4377 EVT EltVT = V1.getValueType().getVectorElementType();
4378 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4380 SmallVector<SDValue, 16> ResultMask;
4381 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4382 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4384 for (unsigned j = 0; j != BytesPerElement; ++j)
4385 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4389 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4390 &ResultMask[0], ResultMask.size());
4391 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4394 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4395 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4396 /// information about the intrinsic.
4397 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4399 unsigned IntrinsicID =
4400 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4403 switch (IntrinsicID) {
4404 default: return false;
4405 // Comparison predicates.
4406 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4407 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4408 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4409 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4410 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4411 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4412 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4413 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4414 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4415 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4416 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4417 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4418 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4420 // Normal Comparisons.
4421 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4422 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4423 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4424 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4425 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4426 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4427 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4428 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4429 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4430 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4431 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4432 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4433 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4438 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4439 /// lower, do it, otherwise return null.
4440 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4441 SelectionDAG &DAG) const {
4442 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4443 // opcode number of the comparison.
4444 DebugLoc dl = Op.getDebugLoc();
4447 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4448 return SDValue(); // Don't custom lower most intrinsics.
4450 // If this is a non-dot comparison, make the VCMP node and we are done.
4452 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4453 Op.getOperand(1), Op.getOperand(2),
4454 DAG.getConstant(CompareOpc, MVT::i32));
4455 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4458 // Create the PPCISD altivec 'dot' comparison node.
4460 Op.getOperand(2), // LHS
4461 Op.getOperand(3), // RHS
4462 DAG.getConstant(CompareOpc, MVT::i32)
4464 std::vector<EVT> VTs;
4465 VTs.push_back(Op.getOperand(2).getValueType());
4466 VTs.push_back(MVT::Glue);
4467 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4469 // Now that we have the comparison, emit a copy from the CR to a GPR.
4470 // This is flagged to the above dot comparison.
4471 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4472 DAG.getRegister(PPC::CR6, MVT::i32),
4473 CompNode.getValue(1));
4475 // Unpack the result based on how the target uses it.
4476 unsigned BitNo; // Bit # of CR6.
4477 bool InvertBit; // Invert result?
4478 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4479 default: // Can't happen, don't crash on invalid number though.
4480 case 0: // Return the value of the EQ bit of CR6.
4481 BitNo = 0; InvertBit = false;
4483 case 1: // Return the inverted value of the EQ bit of CR6.
4484 BitNo = 0; InvertBit = true;
4486 case 2: // Return the value of the LT bit of CR6.
4487 BitNo = 2; InvertBit = false;
4489 case 3: // Return the inverted value of the LT bit of CR6.
4490 BitNo = 2; InvertBit = true;
4494 // Shift the bit into the low position.
4495 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4496 DAG.getConstant(8-(3-BitNo), MVT::i32));
4498 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4499 DAG.getConstant(1, MVT::i32));
4501 // If we are supposed to, toggle the bit.
4503 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4504 DAG.getConstant(1, MVT::i32));
4508 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4509 SelectionDAG &DAG) const {
4510 DebugLoc dl = Op.getDebugLoc();
4511 // Create a stack slot that is 16-byte aligned.
4512 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4513 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4514 EVT PtrVT = getPointerTy();
4515 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4517 // Store the input value into Value#0 of the stack slot.
4518 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4519 Op.getOperand(0), FIdx, MachinePointerInfo(),
4522 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4523 false, false, false, 0);
4526 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4527 DebugLoc dl = Op.getDebugLoc();
4528 if (Op.getValueType() == MVT::v4i32) {
4529 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4531 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4532 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4534 SDValue RHSSwap = // = vrlw RHS, 16
4535 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4537 // Shrinkify inputs to v8i16.
4538 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4539 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4540 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4542 // Low parts multiplied together, generating 32-bit results (we ignore the
4544 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4545 LHS, RHS, DAG, dl, MVT::v4i32);
4547 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4548 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4549 // Shift the high parts up 16 bits.
4550 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4552 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4553 } else if (Op.getValueType() == MVT::v8i16) {
4554 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4556 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4558 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4559 LHS, RHS, Zero, DAG, dl);
4560 } else if (Op.getValueType() == MVT::v16i8) {
4561 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4563 // Multiply the even 8-bit parts, producing 16-bit sums.
4564 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4565 LHS, RHS, DAG, dl, MVT::v8i16);
4566 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4568 // Multiply the odd 8-bit parts, producing 16-bit sums.
4569 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4570 LHS, RHS, DAG, dl, MVT::v8i16);
4571 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4573 // Merge the results together.
4575 for (unsigned i = 0; i != 8; ++i) {
4577 Ops[i*2+1] = 2*i+1+16;
4579 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4581 llvm_unreachable("Unknown mul to lower!");
4585 /// LowerOperation - Provide custom lowering hooks for some operations.
4587 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4588 switch (Op.getOpcode()) {
4589 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4590 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4591 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4592 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4593 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4594 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4595 case ISD::SETCC: return LowerSETCC(Op, DAG);
4596 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4597 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4599 return LowerVASTART(Op, DAG, PPCSubTarget);
4602 return LowerVAARG(Op, DAG, PPCSubTarget);
4604 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4605 case ISD::DYNAMIC_STACKALLOC:
4606 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4608 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4609 case ISD::FP_TO_UINT:
4610 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4612 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4613 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4615 // Lower 64-bit shifts.
4616 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4617 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4618 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4620 // Vector-related lowering.
4621 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4622 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4623 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4624 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4625 case ISD::MUL: return LowerMUL(Op, DAG);
4627 // Frame & Return address.
4628 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4629 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4633 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4634 SmallVectorImpl<SDValue>&Results,
4635 SelectionDAG &DAG) const {
4636 const TargetMachine &TM = getTargetMachine();
4637 DebugLoc dl = N->getDebugLoc();
4638 switch (N->getOpcode()) {
4640 llvm_unreachable("Do not know how to custom type legalize this operation!");
4642 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4643 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4646 EVT VT = N->getValueType(0);
4648 if (VT == MVT::i64) {
4649 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4651 Results.push_back(NewNode);
4652 Results.push_back(NewNode.getValue(1));
4656 case ISD::FP_ROUND_INREG: {
4657 assert(N->getValueType(0) == MVT::ppcf128);
4658 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4659 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4660 MVT::f64, N->getOperand(0),
4661 DAG.getIntPtrConstant(0));
4662 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4663 MVT::f64, N->getOperand(0),
4664 DAG.getIntPtrConstant(1));
4666 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4667 // of the long double, and puts FPSCR back the way it was. We do not
4668 // actually model FPSCR.
4669 std::vector<EVT> NodeTys;
4670 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4672 NodeTys.push_back(MVT::f64); // Return register
4673 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4674 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4675 MFFSreg = Result.getValue(0);
4676 InFlag = Result.getValue(1);
4679 NodeTys.push_back(MVT::Glue); // Returns a flag
4680 Ops[0] = DAG.getConstant(31, MVT::i32);
4682 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4683 InFlag = Result.getValue(0);
4686 NodeTys.push_back(MVT::Glue); // Returns a flag
4687 Ops[0] = DAG.getConstant(30, MVT::i32);
4689 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4690 InFlag = Result.getValue(0);
4693 NodeTys.push_back(MVT::f64); // result of add
4694 NodeTys.push_back(MVT::Glue); // Returns a flag
4698 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4699 FPreg = Result.getValue(0);
4700 InFlag = Result.getValue(1);
4703 NodeTys.push_back(MVT::f64);
4704 Ops[0] = DAG.getConstant(1, MVT::i32);
4708 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4709 FPreg = Result.getValue(0);
4711 // We know the low half is about to be thrown away, so just use something
4713 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4717 case ISD::FP_TO_SINT:
4718 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4724 //===----------------------------------------------------------------------===//
4725 // Other Lowering Code
4726 //===----------------------------------------------------------------------===//
4729 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4730 bool is64bit, unsigned BinOpcode) const {
4731 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4732 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4734 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4735 MachineFunction *F = BB->getParent();
4736 MachineFunction::iterator It = BB;
4739 unsigned dest = MI->getOperand(0).getReg();
4740 unsigned ptrA = MI->getOperand(1).getReg();
4741 unsigned ptrB = MI->getOperand(2).getReg();
4742 unsigned incr = MI->getOperand(3).getReg();
4743 DebugLoc dl = MI->getDebugLoc();
4745 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4746 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4747 F->insert(It, loopMBB);
4748 F->insert(It, exitMBB);
4749 exitMBB->splice(exitMBB->begin(), BB,
4750 llvm::next(MachineBasicBlock::iterator(MI)),
4752 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4754 MachineRegisterInfo &RegInfo = F->getRegInfo();
4755 unsigned TmpReg = (!BinOpcode) ? incr :
4756 RegInfo.createVirtualRegister(
4757 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4758 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4762 // fallthrough --> loopMBB
4763 BB->addSuccessor(loopMBB);
4766 // l[wd]arx dest, ptr
4767 // add r0, dest, incr
4768 // st[wd]cx. r0, ptr
4770 // fallthrough --> exitMBB
4772 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4773 .addReg(ptrA).addReg(ptrB);
4775 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4776 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4777 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4778 BuildMI(BB, dl, TII->get(PPC::BCC))
4779 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4780 BB->addSuccessor(loopMBB);
4781 BB->addSuccessor(exitMBB);
4790 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4791 MachineBasicBlock *BB,
4792 bool is8bit, // operation
4793 unsigned BinOpcode) const {
4794 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4795 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4796 // In 64 bit mode we have to use 64 bits for addresses, even though the
4797 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4798 // registers without caring whether they're 32 or 64, but here we're
4799 // doing actual arithmetic on the addresses.
4800 bool is64bit = PPCSubTarget.isPPC64();
4801 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4803 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4804 MachineFunction *F = BB->getParent();
4805 MachineFunction::iterator It = BB;
4808 unsigned dest = MI->getOperand(0).getReg();
4809 unsigned ptrA = MI->getOperand(1).getReg();
4810 unsigned ptrB = MI->getOperand(2).getReg();
4811 unsigned incr = MI->getOperand(3).getReg();
4812 DebugLoc dl = MI->getDebugLoc();
4814 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4815 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4816 F->insert(It, loopMBB);
4817 F->insert(It, exitMBB);
4818 exitMBB->splice(exitMBB->begin(), BB,
4819 llvm::next(MachineBasicBlock::iterator(MI)),
4821 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4823 MachineRegisterInfo &RegInfo = F->getRegInfo();
4824 const TargetRegisterClass *RC =
4825 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4826 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4827 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4828 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4829 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4830 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4831 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4832 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4833 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4834 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4835 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4836 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4837 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4839 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4843 // fallthrough --> loopMBB
4844 BB->addSuccessor(loopMBB);
4846 // The 4-byte load must be aligned, while a char or short may be
4847 // anywhere in the word. Hence all this nasty bookkeeping code.
4848 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4849 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4850 // xori shift, shift1, 24 [16]
4851 // rlwinm ptr, ptr1, 0, 0, 29
4852 // slw incr2, incr, shift
4853 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4854 // slw mask, mask2, shift
4856 // lwarx tmpDest, ptr
4857 // add tmp, tmpDest, incr2
4858 // andc tmp2, tmpDest, mask
4859 // and tmp3, tmp, mask
4860 // or tmp4, tmp3, tmp2
4863 // fallthrough --> exitMBB
4864 // srw dest, tmpDest, shift
4865 if (ptrA != ZeroReg) {
4866 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4867 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4868 .addReg(ptrA).addReg(ptrB);
4872 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4873 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4874 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4875 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4877 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4878 .addReg(Ptr1Reg).addImm(0).addImm(61);
4880 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4881 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4882 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4883 .addReg(incr).addReg(ShiftReg);
4885 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4887 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4888 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4890 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4891 .addReg(Mask2Reg).addReg(ShiftReg);
4894 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4895 .addReg(ZeroReg).addReg(PtrReg);
4897 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4898 .addReg(Incr2Reg).addReg(TmpDestReg);
4899 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4900 .addReg(TmpDestReg).addReg(MaskReg);
4901 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4902 .addReg(TmpReg).addReg(MaskReg);
4903 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4904 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4905 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4906 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4907 BuildMI(BB, dl, TII->get(PPC::BCC))
4908 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4909 BB->addSuccessor(loopMBB);
4910 BB->addSuccessor(exitMBB);
4915 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4921 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4922 MachineBasicBlock *BB) const {
4923 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4925 // To "insert" these instructions we actually have to insert their
4926 // control-flow patterns.
4927 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4928 MachineFunction::iterator It = BB;
4931 MachineFunction *F = BB->getParent();
4933 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4934 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4935 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4936 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4937 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4939 // The incoming instruction knows the destination vreg to set, the
4940 // condition code register to branch on, the true/false values to
4941 // select between, and a branch opcode to use.
4946 // cmpTY ccX, r1, r2
4948 // fallthrough --> copy0MBB
4949 MachineBasicBlock *thisMBB = BB;
4950 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4951 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4952 unsigned SelectPred = MI->getOperand(4).getImm();
4953 DebugLoc dl = MI->getDebugLoc();
4954 F->insert(It, copy0MBB);
4955 F->insert(It, sinkMBB);
4957 // Transfer the remainder of BB and its successor edges to sinkMBB.
4958 sinkMBB->splice(sinkMBB->begin(), BB,
4959 llvm::next(MachineBasicBlock::iterator(MI)),
4961 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4963 // Next, add the true and fallthrough blocks as its successors.
4964 BB->addSuccessor(copy0MBB);
4965 BB->addSuccessor(sinkMBB);
4967 BuildMI(BB, dl, TII->get(PPC::BCC))
4968 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4971 // %FalseValue = ...
4972 // # fallthrough to sinkMBB
4975 // Update machine-CFG edges
4976 BB->addSuccessor(sinkMBB);
4979 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4982 BuildMI(*BB, BB->begin(), dl,
4983 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4984 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4985 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4987 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4988 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4989 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4990 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4991 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4992 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4993 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4994 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4996 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4997 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4998 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4999 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5000 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5001 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5002 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5003 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5005 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5006 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5007 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5008 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5009 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5010 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5011 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5012 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5014 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5015 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5016 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5017 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5018 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5019 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5020 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5021 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5023 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5024 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5025 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5026 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5027 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5028 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5029 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5030 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5032 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5033 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5034 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5035 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5036 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5037 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5038 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5039 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5041 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5042 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5043 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5044 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5045 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5046 BB = EmitAtomicBinary(MI, BB, false, 0);
5047 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5048 BB = EmitAtomicBinary(MI, BB, true, 0);
5050 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5051 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5052 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5054 unsigned dest = MI->getOperand(0).getReg();
5055 unsigned ptrA = MI->getOperand(1).getReg();
5056 unsigned ptrB = MI->getOperand(2).getReg();
5057 unsigned oldval = MI->getOperand(3).getReg();
5058 unsigned newval = MI->getOperand(4).getReg();
5059 DebugLoc dl = MI->getDebugLoc();
5061 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5062 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5063 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5064 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5065 F->insert(It, loop1MBB);
5066 F->insert(It, loop2MBB);
5067 F->insert(It, midMBB);
5068 F->insert(It, exitMBB);
5069 exitMBB->splice(exitMBB->begin(), BB,
5070 llvm::next(MachineBasicBlock::iterator(MI)),
5072 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5076 // fallthrough --> loopMBB
5077 BB->addSuccessor(loop1MBB);
5080 // l[wd]arx dest, ptr
5081 // cmp[wd] dest, oldval
5084 // st[wd]cx. newval, ptr
5088 // st[wd]cx. dest, ptr
5091 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5092 .addReg(ptrA).addReg(ptrB);
5093 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5094 .addReg(oldval).addReg(dest);
5095 BuildMI(BB, dl, TII->get(PPC::BCC))
5096 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5097 BB->addSuccessor(loop2MBB);
5098 BB->addSuccessor(midMBB);
5101 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5102 .addReg(newval).addReg(ptrA).addReg(ptrB);
5103 BuildMI(BB, dl, TII->get(PPC::BCC))
5104 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5105 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5106 BB->addSuccessor(loop1MBB);
5107 BB->addSuccessor(exitMBB);
5110 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5111 .addReg(dest).addReg(ptrA).addReg(ptrB);
5112 BB->addSuccessor(exitMBB);
5117 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5118 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5119 // We must use 64-bit registers for addresses when targeting 64-bit,
5120 // since we're actually doing arithmetic on them. Other registers
5122 bool is64bit = PPCSubTarget.isPPC64();
5123 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5125 unsigned dest = MI->getOperand(0).getReg();
5126 unsigned ptrA = MI->getOperand(1).getReg();
5127 unsigned ptrB = MI->getOperand(2).getReg();
5128 unsigned oldval = MI->getOperand(3).getReg();
5129 unsigned newval = MI->getOperand(4).getReg();
5130 DebugLoc dl = MI->getDebugLoc();
5132 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5133 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5134 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5135 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5136 F->insert(It, loop1MBB);
5137 F->insert(It, loop2MBB);
5138 F->insert(It, midMBB);
5139 F->insert(It, exitMBB);
5140 exitMBB->splice(exitMBB->begin(), BB,
5141 llvm::next(MachineBasicBlock::iterator(MI)),
5143 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5145 MachineRegisterInfo &RegInfo = F->getRegInfo();
5146 const TargetRegisterClass *RC =
5147 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5148 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5149 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5150 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5151 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5152 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5153 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5154 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5155 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5156 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5157 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5158 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5159 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5160 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5161 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5163 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5164 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5167 // fallthrough --> loopMBB
5168 BB->addSuccessor(loop1MBB);
5170 // The 4-byte load must be aligned, while a char or short may be
5171 // anywhere in the word. Hence all this nasty bookkeeping code.
5172 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5173 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5174 // xori shift, shift1, 24 [16]
5175 // rlwinm ptr, ptr1, 0, 0, 29
5176 // slw newval2, newval, shift
5177 // slw oldval2, oldval,shift
5178 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5179 // slw mask, mask2, shift
5180 // and newval3, newval2, mask
5181 // and oldval3, oldval2, mask
5183 // lwarx tmpDest, ptr
5184 // and tmp, tmpDest, mask
5185 // cmpw tmp, oldval3
5188 // andc tmp2, tmpDest, mask
5189 // or tmp4, tmp2, newval3
5194 // stwcx. tmpDest, ptr
5196 // srw dest, tmpDest, shift
5197 if (ptrA != ZeroReg) {
5198 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5199 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5200 .addReg(ptrA).addReg(ptrB);
5204 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5205 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5206 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5207 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5209 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5210 .addReg(Ptr1Reg).addImm(0).addImm(61);
5212 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5213 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5214 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5215 .addReg(newval).addReg(ShiftReg);
5216 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5217 .addReg(oldval).addReg(ShiftReg);
5219 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5221 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5222 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5223 .addReg(Mask3Reg).addImm(65535);
5225 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5226 .addReg(Mask2Reg).addReg(ShiftReg);
5227 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5228 .addReg(NewVal2Reg).addReg(MaskReg);
5229 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5230 .addReg(OldVal2Reg).addReg(MaskReg);
5233 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5234 .addReg(ZeroReg).addReg(PtrReg);
5235 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5236 .addReg(TmpDestReg).addReg(MaskReg);
5237 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5238 .addReg(TmpReg).addReg(OldVal3Reg);
5239 BuildMI(BB, dl, TII->get(PPC::BCC))
5240 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5241 BB->addSuccessor(loop2MBB);
5242 BB->addSuccessor(midMBB);
5245 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5246 .addReg(TmpDestReg).addReg(MaskReg);
5247 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5248 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5249 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5250 .addReg(ZeroReg).addReg(PtrReg);
5251 BuildMI(BB, dl, TII->get(PPC::BCC))
5252 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5253 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5254 BB->addSuccessor(loop1MBB);
5255 BB->addSuccessor(exitMBB);
5258 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5259 .addReg(ZeroReg).addReg(PtrReg);
5260 BB->addSuccessor(exitMBB);
5265 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5268 llvm_unreachable("Unexpected instr type to insert");
5271 MI->eraseFromParent(); // The pseudo instruction is gone now.
5275 //===----------------------------------------------------------------------===//
5276 // Target Optimization Hooks
5277 //===----------------------------------------------------------------------===//
5279 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5280 DAGCombinerInfo &DCI) const {
5281 const TargetMachine &TM = getTargetMachine();
5282 SelectionDAG &DAG = DCI.DAG;
5283 DebugLoc dl = N->getDebugLoc();
5284 switch (N->getOpcode()) {
5287 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5288 if (C->isNullValue()) // 0 << V -> 0.
5289 return N->getOperand(0);
5293 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5294 if (C->isNullValue()) // 0 >>u V -> 0.
5295 return N->getOperand(0);
5299 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5300 if (C->isNullValue() || // 0 >>s V -> 0.
5301 C->isAllOnesValue()) // -1 >>s V -> -1.
5302 return N->getOperand(0);
5306 case ISD::SINT_TO_FP:
5307 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5308 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5309 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5310 // We allow the src/dst to be either f32/f64, but the intermediate
5311 // type must be i64.
5312 if (N->getOperand(0).getValueType() == MVT::i64 &&
5313 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5314 SDValue Val = N->getOperand(0).getOperand(0);
5315 if (Val.getValueType() == MVT::f32) {
5316 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5317 DCI.AddToWorklist(Val.getNode());
5320 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5321 DCI.AddToWorklist(Val.getNode());
5322 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5323 DCI.AddToWorklist(Val.getNode());
5324 if (N->getValueType(0) == MVT::f32) {
5325 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5326 DAG.getIntPtrConstant(0));
5327 DCI.AddToWorklist(Val.getNode());
5330 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5331 // If the intermediate type is i32, we can avoid the load/store here
5338 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5339 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5340 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5341 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5342 N->getOperand(1).getValueType() == MVT::i32 &&
5343 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5344 SDValue Val = N->getOperand(1).getOperand(0);
5345 if (Val.getValueType() == MVT::f32) {
5346 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5347 DCI.AddToWorklist(Val.getNode());
5349 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5350 DCI.AddToWorklist(Val.getNode());
5352 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5353 N->getOperand(2), N->getOperand(3));
5354 DCI.AddToWorklist(Val.getNode());
5358 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5359 if (cast<StoreSDNode>(N)->isUnindexed() &&
5360 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5361 N->getOperand(1).getNode()->hasOneUse() &&
5362 (N->getOperand(1).getValueType() == MVT::i32 ||
5363 N->getOperand(1).getValueType() == MVT::i16)) {
5364 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5365 // Do an any-extend to 32-bits if this is a half-word input.
5366 if (BSwapOp.getValueType() == MVT::i16)
5367 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5370 N->getOperand(0), BSwapOp, N->getOperand(2),
5371 DAG.getValueType(N->getOperand(1).getValueType())
5374 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5375 Ops, array_lengthof(Ops),
5376 cast<StoreSDNode>(N)->getMemoryVT(),
5377 cast<StoreSDNode>(N)->getMemOperand());
5381 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5382 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5383 N->getOperand(0).hasOneUse() &&
5384 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5385 SDValue Load = N->getOperand(0);
5386 LoadSDNode *LD = cast<LoadSDNode>(Load);
5387 // Create the byte-swapping load.
5389 LD->getChain(), // Chain
5390 LD->getBasePtr(), // Ptr
5391 DAG.getValueType(N->getValueType(0)) // VT
5394 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5395 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5396 LD->getMemoryVT(), LD->getMemOperand());
5398 // If this is an i16 load, insert the truncate.
5399 SDValue ResVal = BSLoad;
5400 if (N->getValueType(0) == MVT::i16)
5401 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5403 // First, combine the bswap away. This makes the value produced by the
5405 DCI.CombineTo(N, ResVal);
5407 // Next, combine the load away, we give it a bogus result value but a real
5408 // chain result. The result value is dead because the bswap is dead.
5409 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5411 // Return N so it doesn't get rechecked!
5412 return SDValue(N, 0);
5416 case PPCISD::VCMP: {
5417 // If a VCMPo node already exists with exactly the same operands as this
5418 // node, use its result instead of this node (VCMPo computes both a CR6 and
5419 // a normal output).
5421 if (!N->getOperand(0).hasOneUse() &&
5422 !N->getOperand(1).hasOneUse() &&
5423 !N->getOperand(2).hasOneUse()) {
5425 // Scan all of the users of the LHS, looking for VCMPo's that match.
5426 SDNode *VCMPoNode = 0;
5428 SDNode *LHSN = N->getOperand(0).getNode();
5429 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5431 if (UI->getOpcode() == PPCISD::VCMPo &&
5432 UI->getOperand(1) == N->getOperand(1) &&
5433 UI->getOperand(2) == N->getOperand(2) &&
5434 UI->getOperand(0) == N->getOperand(0)) {
5439 // If there is no VCMPo node, or if the flag value has a single use, don't
5441 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5444 // Look at the (necessarily single) use of the flag value. If it has a
5445 // chain, this transformation is more complex. Note that multiple things
5446 // could use the value result, which we should ignore.
5447 SDNode *FlagUser = 0;
5448 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5449 FlagUser == 0; ++UI) {
5450 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5452 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5453 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5460 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5461 // give up for right now.
5462 if (FlagUser->getOpcode() == PPCISD::MFCR)
5463 return SDValue(VCMPoNode, 0);
5468 // If this is a branch on an altivec predicate comparison, lower this so
5469 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5470 // lowering is done pre-legalize, because the legalizer lowers the predicate
5471 // compare down to code that is difficult to reassemble.
5472 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5473 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5477 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5478 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5479 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5480 assert(isDot && "Can't compare against a vector result!");
5482 // If this is a comparison against something other than 0/1, then we know
5483 // that the condition is never/always true.
5484 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5485 if (Val != 0 && Val != 1) {
5486 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5487 return N->getOperand(0);
5488 // Always !=, turn it into an unconditional branch.
5489 return DAG.getNode(ISD::BR, dl, MVT::Other,
5490 N->getOperand(0), N->getOperand(4));
5493 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5495 // Create the PPCISD altivec 'dot' comparison node.
5496 std::vector<EVT> VTs;
5498 LHS.getOperand(2), // LHS of compare
5499 LHS.getOperand(3), // RHS of compare
5500 DAG.getConstant(CompareOpc, MVT::i32)
5502 VTs.push_back(LHS.getOperand(2).getValueType());
5503 VTs.push_back(MVT::Glue);
5504 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5506 // Unpack the result based on how the target uses it.
5507 PPC::Predicate CompOpc;
5508 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5509 default: // Can't happen, don't crash on invalid number though.
5510 case 0: // Branch on the value of the EQ bit of CR6.
5511 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5513 case 1: // Branch on the inverted value of the EQ bit of CR6.
5514 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5516 case 2: // Branch on the value of the LT bit of CR6.
5517 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5519 case 3: // Branch on the inverted value of the LT bit of CR6.
5520 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5524 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5525 DAG.getConstant(CompOpc, MVT::i32),
5526 DAG.getRegister(PPC::CR6, MVT::i32),
5527 N->getOperand(4), CompNode.getValue(1));
5536 //===----------------------------------------------------------------------===//
5537 // Inline Assembly Support
5538 //===----------------------------------------------------------------------===//
5540 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5543 const SelectionDAG &DAG,
5544 unsigned Depth) const {
5545 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
5546 switch (Op.getOpcode()) {
5548 case PPCISD::LBRX: {
5549 // lhbrx is known to have the top bits cleared out.
5550 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5551 KnownZero = 0xFFFF0000;
5554 case ISD::INTRINSIC_WO_CHAIN: {
5555 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5557 case Intrinsic::ppc_altivec_vcmpbfp_p:
5558 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5559 case Intrinsic::ppc_altivec_vcmpequb_p:
5560 case Intrinsic::ppc_altivec_vcmpequh_p:
5561 case Intrinsic::ppc_altivec_vcmpequw_p:
5562 case Intrinsic::ppc_altivec_vcmpgefp_p:
5563 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5564 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5565 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5566 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5567 case Intrinsic::ppc_altivec_vcmpgtub_p:
5568 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5569 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5570 KnownZero = ~1U; // All bits but the low one are known to be zero.
5578 /// getConstraintType - Given a constraint, return the type of
5579 /// constraint it is for this target.
5580 PPCTargetLowering::ConstraintType
5581 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5582 if (Constraint.size() == 1) {
5583 switch (Constraint[0]) {
5590 return C_RegisterClass;
5593 return TargetLowering::getConstraintType(Constraint);
5596 /// Examine constraint type and operand type and determine a weight value.
5597 /// This object must already have been set up with the operand type
5598 /// and the current alternative constraint selected.
5599 TargetLowering::ConstraintWeight
5600 PPCTargetLowering::getSingleConstraintMatchWeight(
5601 AsmOperandInfo &info, const char *constraint) const {
5602 ConstraintWeight weight = CW_Invalid;
5603 Value *CallOperandVal = info.CallOperandVal;
5604 // If we don't have a value, we can't do a match,
5605 // but allow it at the lowest weight.
5606 if (CallOperandVal == NULL)
5608 Type *type = CallOperandVal->getType();
5609 // Look at the constraint type.
5610 switch (*constraint) {
5612 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5615 if (type->isIntegerTy())
5616 weight = CW_Register;
5619 if (type->isFloatTy())
5620 weight = CW_Register;
5623 if (type->isDoubleTy())
5624 weight = CW_Register;
5627 if (type->isVectorTy())
5628 weight = CW_Register;
5631 weight = CW_Register;
5637 std::pair<unsigned, const TargetRegisterClass*>
5638 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5640 if (Constraint.size() == 1) {
5641 // GCC RS6000 Constraint Letters
5642 switch (Constraint[0]) {
5645 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5646 return std::make_pair(0U, &PPC::G8RCRegClass);
5647 return std::make_pair(0U, &PPC::GPRCRegClass);
5650 return std::make_pair(0U, &PPC::F4RCRegClass);
5652 return std::make_pair(0U, &PPC::F8RCRegClass);
5655 return std::make_pair(0U, &PPC::VRRCRegClass);
5657 return std::make_pair(0U, &PPC::CRRCRegClass);
5661 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5665 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5666 /// vector. If it is invalid, don't add anything to Ops.
5667 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5668 std::string &Constraint,
5669 std::vector<SDValue>&Ops,
5670 SelectionDAG &DAG) const {
5671 SDValue Result(0,0);
5673 // Only support length 1 constraints.
5674 if (Constraint.length() > 1) return;
5676 char Letter = Constraint[0];
5687 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5688 if (!CST) return; // Must be an immediate to match.
5689 unsigned Value = CST->getZExtValue();
5691 default: llvm_unreachable("Unknown constraint letter!");
5692 case 'I': // "I" is a signed 16-bit constant.
5693 if ((short)Value == (int)Value)
5694 Result = DAG.getTargetConstant(Value, Op.getValueType());
5696 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5697 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5698 if ((short)Value == 0)
5699 Result = DAG.getTargetConstant(Value, Op.getValueType());
5701 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5702 if ((Value >> 16) == 0)
5703 Result = DAG.getTargetConstant(Value, Op.getValueType());
5705 case 'M': // "M" is a constant that is greater than 31.
5707 Result = DAG.getTargetConstant(Value, Op.getValueType());
5709 case 'N': // "N" is a positive constant that is an exact power of two.
5710 if ((int)Value > 0 && isPowerOf2_32(Value))
5711 Result = DAG.getTargetConstant(Value, Op.getValueType());
5713 case 'O': // "O" is the constant zero.
5715 Result = DAG.getTargetConstant(Value, Op.getValueType());
5717 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5718 if ((short)-Value == (int)-Value)
5719 Result = DAG.getTargetConstant(Value, Op.getValueType());
5726 if (Result.getNode()) {
5727 Ops.push_back(Result);
5731 // Handle standard constraint letters.
5732 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5735 // isLegalAddressingMode - Return true if the addressing mode represented
5736 // by AM is legal for this target, for a load/store of the specified type.
5737 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5739 // FIXME: PPC does not allow r+i addressing modes for vectors!
5741 // PPC allows a sign-extended 16-bit immediate field.
5742 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5745 // No global is ever allowed as a base.
5749 // PPC only support r+r,
5751 case 0: // "r+i" or just "i", depending on HasBaseReg.
5754 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5756 // Otherwise we have r+r or r+i.
5759 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5761 // Allow 2*r as r+r.
5764 // No other scales are supported.
5771 /// isLegalAddressImmediate - Return true if the integer value can be used
5772 /// as the offset of the target addressing mode for load / store of the
5774 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5775 // PPC allows a sign-extended 16-bit immediate field.
5776 return (V > -(1 << 16) && V < (1 << 16)-1);
5779 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
5783 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5784 SelectionDAG &DAG) const {
5785 MachineFunction &MF = DAG.getMachineFunction();
5786 MachineFrameInfo *MFI = MF.getFrameInfo();
5787 MFI->setReturnAddressIsTaken(true);
5789 DebugLoc dl = Op.getDebugLoc();
5790 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5792 // Make sure the function does not optimize away the store of the RA to
5794 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5795 FuncInfo->setLRStoreRequired();
5796 bool isPPC64 = PPCSubTarget.isPPC64();
5797 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5800 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5803 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5804 isPPC64? MVT::i64 : MVT::i32);
5805 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5806 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5808 MachinePointerInfo(), false, false, false, 0);
5811 // Just load the return address off the stack.
5812 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5813 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5814 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5817 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5818 SelectionDAG &DAG) const {
5819 DebugLoc dl = Op.getDebugLoc();
5820 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5822 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5823 bool isPPC64 = PtrVT == MVT::i64;
5825 MachineFunction &MF = DAG.getMachineFunction();
5826 MachineFrameInfo *MFI = MF.getFrameInfo();
5827 MFI->setFrameAddressIsTaken(true);
5828 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
5829 MFI->hasVarSizedObjects()) &&
5830 MFI->getStackSize() &&
5831 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5832 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5833 (is31 ? PPC::R31 : PPC::R1);
5834 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5837 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5838 FrameAddr, MachinePointerInfo(), false, false,
5844 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5845 // The PowerPC target isn't yet aware of offsets.
5849 /// getOptimalMemOpType - Returns the target specific optimal type for load
5850 /// and store operations as a result of memset, memcpy, and memmove
5851 /// lowering. If DstAlign is zero that means it's safe to destination
5852 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5853 /// means there isn't a need to check it against alignment requirement,
5854 /// probably because the source does not need to be loaded. If
5855 /// 'IsZeroVal' is true, that means it's safe to return a
5856 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5857 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5858 /// constant so it does not need to be loaded.
5859 /// It returns EVT::Other if the type should be determined using generic
5860 /// target-independent logic.
5861 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5862 unsigned DstAlign, unsigned SrcAlign,
5865 MachineFunction &MF) const {
5866 if (this->PPCSubTarget.isPPC64()) {
5873 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
5874 unsigned Directive = PPCSubTarget.getDarwinDirective();
5875 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2)
5878 return TargetLowering::getSchedulingPreference(N);