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::TargetConstantPool ||
909 Disp.getOpcode() == ISD::TargetJumpTable);
910 Base = N.getOperand(0);
911 return true; // [&g+r]
913 } else if (N.getOpcode() == ISD::OR) {
915 if (isIntS16Immediate(N.getOperand(1), imm)) {
916 // If this is an or of disjoint bitfields, we can codegen this as an add
917 // (for better address arithmetic) if the LHS and RHS of the OR are
918 // provably disjoint.
919 APInt LHSKnownZero, LHSKnownOne;
920 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
922 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
923 // If all of the bits are known zero on the LHS or RHS, the add won't
925 Base = N.getOperand(0);
926 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
930 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
931 // Loading from a constant address.
933 // If this address fits entirely in a 16-bit sext immediate field, codegen
936 if (isIntS16Immediate(CN, Imm)) {
937 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
938 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
939 CN->getValueType(0));
943 // Handle 32-bit sext immediates with LIS + addr mode.
944 if (CN->getValueType(0) == MVT::i32 ||
945 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
946 int Addr = (int)CN->getZExtValue();
948 // Otherwise, break this down into an LIS + disp.
949 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
951 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
952 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
953 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
958 Disp = DAG.getTargetConstant(0, getPointerTy());
959 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
960 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
963 return true; // [r+0]
966 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
967 /// represented as an indexed [r+r] operation.
968 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
970 SelectionDAG &DAG) const {
971 // Check to see if we can easily represent this as an [r+r] address. This
972 // will fail if it thinks that the address is more profitably represented as
973 // reg+imm, e.g. where imm = 0.
974 if (SelectAddressRegReg(N, Base, Index, DAG))
977 // If the operand is an addition, always emit this as [r+r], since this is
978 // better (for code size, and execution, as the memop does the add for free)
979 // than emitting an explicit add.
980 if (N.getOpcode() == ISD::ADD) {
981 Base = N.getOperand(0);
982 Index = N.getOperand(1);
986 // Otherwise, do it the hard way, using R0 as the base register.
987 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
993 /// SelectAddressRegImmShift - Returns true if the address N can be
994 /// represented by a base register plus a signed 14-bit displacement
995 /// [r+imm*4]. Suitable for use by STD and friends.
996 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
998 SelectionDAG &DAG) const {
999 // FIXME dl should come from the parent load or store, not the address
1000 DebugLoc dl = N.getDebugLoc();
1001 // If this can be more profitably realized as r+r, fail.
1002 if (SelectAddressRegReg(N, Disp, Base, DAG))
1005 if (N.getOpcode() == ISD::ADD) {
1007 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1008 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1009 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1010 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1012 Base = N.getOperand(0);
1014 return true; // [r+i]
1015 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1016 // Match LOAD (ADD (X, Lo(G))).
1017 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1018 && "Cannot handle constant offsets yet!");
1019 Disp = N.getOperand(1).getOperand(0); // The global address.
1020 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1021 Disp.getOpcode() == ISD::TargetConstantPool ||
1022 Disp.getOpcode() == ISD::TargetJumpTable);
1023 Base = N.getOperand(0);
1024 return true; // [&g+r]
1026 } else if (N.getOpcode() == ISD::OR) {
1028 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1029 // If this is an or of disjoint bitfields, we can codegen this as an add
1030 // (for better address arithmetic) if the LHS and RHS of the OR are
1031 // provably disjoint.
1032 APInt LHSKnownZero, LHSKnownOne;
1033 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1034 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1035 // If all of the bits are known zero on the LHS or RHS, the add won't
1037 Base = N.getOperand(0);
1038 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1042 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1043 // Loading from a constant address. Verify low two bits are clear.
1044 if ((CN->getZExtValue() & 3) == 0) {
1045 // If this address fits entirely in a 14-bit sext immediate field, codegen
1048 if (isIntS16Immediate(CN, Imm)) {
1049 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1050 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1051 CN->getValueType(0));
1055 // Fold the low-part of 32-bit absolute addresses into addr mode.
1056 if (CN->getValueType(0) == MVT::i32 ||
1057 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1058 int Addr = (int)CN->getZExtValue();
1060 // Otherwise, break this down into an LIS + disp.
1061 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1062 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1063 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1064 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1070 Disp = DAG.getTargetConstant(0, getPointerTy());
1071 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1072 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1075 return true; // [r+0]
1079 /// getPreIndexedAddressParts - returns true by value, base pointer and
1080 /// offset pointer and addressing mode by reference if the node's address
1081 /// can be legally represented as pre-indexed load / store address.
1082 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1084 ISD::MemIndexedMode &AM,
1085 SelectionDAG &DAG) const {
1086 if (DisablePPCPreinc) return false;
1090 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1091 Ptr = LD->getBasePtr();
1092 VT = LD->getMemoryVT();
1094 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1095 Ptr = ST->getBasePtr();
1096 VT = ST->getMemoryVT();
1100 // PowerPC doesn't have preinc load/store instructions for vectors.
1104 // TODO: Check reg+reg first.
1106 // LDU/STU use reg+imm*4, others use reg+imm.
1107 if (VT != MVT::i64) {
1109 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1113 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1117 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1118 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1119 // sext i32 to i64 when addr mode is r+i.
1120 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1121 LD->getExtensionType() == ISD::SEXTLOAD &&
1122 isa<ConstantSDNode>(Offset))
1130 //===----------------------------------------------------------------------===//
1131 // LowerOperation implementation
1132 //===----------------------------------------------------------------------===//
1134 /// GetLabelAccessInfo - Return true if we should reference labels using a
1135 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1136 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1137 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1138 HiOpFlags = PPCII::MO_HA16;
1139 LoOpFlags = PPCII::MO_LO16;
1141 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1142 // non-darwin platform. We don't support PIC on other platforms yet.
1143 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1144 TM.getSubtarget<PPCSubtarget>().isDarwin();
1146 HiOpFlags |= PPCII::MO_PIC_FLAG;
1147 LoOpFlags |= PPCII::MO_PIC_FLAG;
1150 // If this is a reference to a global value that requires a non-lazy-ptr, make
1151 // sure that instruction lowering adds it.
1152 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1153 HiOpFlags |= PPCII::MO_NLP_FLAG;
1154 LoOpFlags |= PPCII::MO_NLP_FLAG;
1156 if (GV->hasHiddenVisibility()) {
1157 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1158 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1165 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1166 SelectionDAG &DAG) {
1167 EVT PtrVT = HiPart.getValueType();
1168 SDValue Zero = DAG.getConstant(0, PtrVT);
1169 DebugLoc DL = HiPart.getDebugLoc();
1171 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1172 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1174 // With PIC, the first instruction is actually "GR+hi(&G)".
1176 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1177 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1179 // Generate non-pic code that has direct accesses to the constant pool.
1180 // The address of the global is just (hi(&g)+lo(&g)).
1181 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1184 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1185 SelectionDAG &DAG) const {
1186 EVT PtrVT = Op.getValueType();
1187 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1188 const Constant *C = CP->getConstVal();
1190 unsigned MOHiFlag, MOLoFlag;
1191 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1193 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1195 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1196 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1199 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1200 EVT PtrVT = Op.getValueType();
1201 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1203 unsigned MOHiFlag, MOLoFlag;
1204 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1205 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1206 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1207 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1210 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1211 SelectionDAG &DAG) const {
1212 EVT PtrVT = Op.getValueType();
1214 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1216 unsigned MOHiFlag, MOLoFlag;
1217 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1218 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1219 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1220 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1223 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1224 SelectionDAG &DAG) const {
1225 EVT PtrVT = Op.getValueType();
1226 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1227 DebugLoc DL = GSDN->getDebugLoc();
1228 const GlobalValue *GV = GSDN->getGlobal();
1230 // 64-bit SVR4 ABI code is always position-independent.
1231 // The actual address of the GlobalValue is stored in the TOC.
1232 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1233 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1234 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1235 DAG.getRegister(PPC::X2, MVT::i64));
1238 unsigned MOHiFlag, MOLoFlag;
1239 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1242 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1244 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1246 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1248 // If the global reference is actually to a non-lazy-pointer, we have to do an
1249 // extra load to get the address of the global.
1250 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1251 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1252 false, false, false, 0);
1256 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1257 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1258 DebugLoc dl = Op.getDebugLoc();
1260 // If we're comparing for equality to zero, expose the fact that this is
1261 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1262 // fold the new nodes.
1263 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1264 if (C->isNullValue() && CC == ISD::SETEQ) {
1265 EVT VT = Op.getOperand(0).getValueType();
1266 SDValue Zext = Op.getOperand(0);
1267 if (VT.bitsLT(MVT::i32)) {
1269 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1271 unsigned Log2b = Log2_32(VT.getSizeInBits());
1272 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1273 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1274 DAG.getConstant(Log2b, MVT::i32));
1275 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1277 // Leave comparisons against 0 and -1 alone for now, since they're usually
1278 // optimized. FIXME: revisit this when we can custom lower all setcc
1280 if (C->isAllOnesValue() || C->isNullValue())
1284 // If we have an integer seteq/setne, turn it into a compare against zero
1285 // by xor'ing the rhs with the lhs, which is faster than setting a
1286 // condition register, reading it back out, and masking the correct bit. The
1287 // normal approach here uses sub to do this instead of xor. Using xor exposes
1288 // the result to other bit-twiddling opportunities.
1289 EVT LHSVT = Op.getOperand(0).getValueType();
1290 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1291 EVT VT = Op.getValueType();
1292 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1294 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1299 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1300 const PPCSubtarget &Subtarget) const {
1301 SDNode *Node = Op.getNode();
1302 EVT VT = Node->getValueType(0);
1303 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1304 SDValue InChain = Node->getOperand(0);
1305 SDValue VAListPtr = Node->getOperand(1);
1306 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1307 DebugLoc dl = Node->getDebugLoc();
1309 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1312 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1313 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1315 InChain = GprIndex.getValue(1);
1317 if (VT == MVT::i64) {
1318 // Check if GprIndex is even
1319 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1320 DAG.getConstant(1, MVT::i32));
1321 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1322 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1323 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1324 DAG.getConstant(1, MVT::i32));
1325 // Align GprIndex to be even if it isn't
1326 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1330 // fpr index is 1 byte after gpr
1331 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1332 DAG.getConstant(1, MVT::i32));
1335 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1336 FprPtr, MachinePointerInfo(SV), MVT::i8,
1338 InChain = FprIndex.getValue(1);
1340 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1341 DAG.getConstant(8, MVT::i32));
1343 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1344 DAG.getConstant(4, MVT::i32));
1347 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1348 MachinePointerInfo(), false, false,
1350 InChain = OverflowArea.getValue(1);
1352 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1353 MachinePointerInfo(), false, false,
1355 InChain = RegSaveArea.getValue(1);
1357 // select overflow_area if index > 8
1358 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1359 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1361 // adjustment constant gpr_index * 4/8
1362 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1363 VT.isInteger() ? GprIndex : FprIndex,
1364 DAG.getConstant(VT.isInteger() ? 4 : 8,
1367 // OurReg = RegSaveArea + RegConstant
1368 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1371 // Floating types are 32 bytes into RegSaveArea
1372 if (VT.isFloatingPoint())
1373 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1374 DAG.getConstant(32, MVT::i32));
1376 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1377 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1378 VT.isInteger() ? GprIndex : FprIndex,
1379 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1382 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1383 VT.isInteger() ? VAListPtr : FprPtr,
1384 MachinePointerInfo(SV),
1385 MVT::i8, false, false, 0);
1387 // determine if we should load from reg_save_area or overflow_area
1388 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1390 // increase overflow_area by 4/8 if gpr/fpr > 8
1391 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1392 DAG.getConstant(VT.isInteger() ? 4 : 8,
1395 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1398 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1400 MachinePointerInfo(),
1401 MVT::i32, false, false, 0);
1403 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1404 false, false, false, 0);
1407 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1408 SelectionDAG &DAG) const {
1409 return Op.getOperand(0);
1412 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1413 SelectionDAG &DAG) const {
1414 SDValue Chain = Op.getOperand(0);
1415 SDValue Trmp = Op.getOperand(1); // trampoline
1416 SDValue FPtr = Op.getOperand(2); // nested function
1417 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1418 DebugLoc dl = Op.getDebugLoc();
1420 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1421 bool isPPC64 = (PtrVT == MVT::i64);
1423 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1426 TargetLowering::ArgListTy Args;
1427 TargetLowering::ArgListEntry Entry;
1429 Entry.Ty = IntPtrTy;
1430 Entry.Node = Trmp; Args.push_back(Entry);
1432 // TrampSize == (isPPC64 ? 48 : 40);
1433 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1434 isPPC64 ? MVT::i64 : MVT::i32);
1435 Args.push_back(Entry);
1437 Entry.Node = FPtr; Args.push_back(Entry);
1438 Entry.Node = Nest; Args.push_back(Entry);
1440 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1441 TargetLowering::CallLoweringInfo CLI(Chain,
1442 Type::getVoidTy(*DAG.getContext()),
1443 false, false, false, false, 0,
1445 /*isTailCall=*/false,
1446 /*doesNotRet=*/false,
1447 /*isReturnValueUsed=*/true,
1448 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1450 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1452 return CallResult.second;
1455 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1456 const PPCSubtarget &Subtarget) const {
1457 MachineFunction &MF = DAG.getMachineFunction();
1458 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1460 DebugLoc dl = Op.getDebugLoc();
1462 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1463 // vastart just stores the address of the VarArgsFrameIndex slot into the
1464 // memory location argument.
1465 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1466 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1467 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1468 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1469 MachinePointerInfo(SV),
1473 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1474 // We suppose the given va_list is already allocated.
1477 // char gpr; /* index into the array of 8 GPRs
1478 // * stored in the register save area
1479 // * gpr=0 corresponds to r3,
1480 // * gpr=1 to r4, etc.
1482 // char fpr; /* index into the array of 8 FPRs
1483 // * stored in the register save area
1484 // * fpr=0 corresponds to f1,
1485 // * fpr=1 to f2, etc.
1487 // char *overflow_arg_area;
1488 // /* location on stack that holds
1489 // * the next overflow argument
1491 // char *reg_save_area;
1492 // /* where r3:r10 and f1:f8 (if saved)
1498 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1499 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1502 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1504 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1506 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1509 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1510 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1512 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1513 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1515 uint64_t FPROffset = 1;
1516 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1518 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1520 // Store first byte : number of int regs
1521 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1523 MachinePointerInfo(SV),
1524 MVT::i8, false, false, 0);
1525 uint64_t nextOffset = FPROffset;
1526 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1529 // Store second byte : number of float regs
1530 SDValue secondStore =
1531 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1532 MachinePointerInfo(SV, nextOffset), MVT::i8,
1534 nextOffset += StackOffset;
1535 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1537 // Store second word : arguments given on stack
1538 SDValue thirdStore =
1539 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1540 MachinePointerInfo(SV, nextOffset),
1542 nextOffset += FrameOffset;
1543 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1545 // Store third word : arguments given in registers
1546 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1547 MachinePointerInfo(SV, nextOffset),
1552 #include "PPCGenCallingConv.inc"
1554 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1555 CCValAssign::LocInfo &LocInfo,
1556 ISD::ArgFlagsTy &ArgFlags,
1561 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1563 CCValAssign::LocInfo &LocInfo,
1564 ISD::ArgFlagsTy &ArgFlags,
1566 static const uint16_t ArgRegs[] = {
1567 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1568 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1570 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1572 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1574 // Skip one register if the first unallocated register has an even register
1575 // number and there are still argument registers available which have not been
1576 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1577 // need to skip a register if RegNum is odd.
1578 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1579 State.AllocateReg(ArgRegs[RegNum]);
1582 // Always return false here, as this function only makes sure that the first
1583 // unallocated register has an odd register number and does not actually
1584 // allocate a register for the current argument.
1588 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1590 CCValAssign::LocInfo &LocInfo,
1591 ISD::ArgFlagsTy &ArgFlags,
1593 static const uint16_t ArgRegs[] = {
1594 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1598 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1600 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1602 // If there is only one Floating-point register left we need to put both f64
1603 // values of a split ppc_fp128 value on the stack.
1604 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1605 State.AllocateReg(ArgRegs[RegNum]);
1608 // Always return false here, as this function only makes sure that the two f64
1609 // values a ppc_fp128 value is split into are both passed in registers or both
1610 // passed on the stack and does not actually allocate a register for the
1611 // current argument.
1615 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1617 static const uint16_t *GetFPR() {
1618 static const uint16_t FPR[] = {
1619 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1620 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1626 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1628 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1629 unsigned PtrByteSize) {
1630 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1631 if (Flags.isByVal())
1632 ArgSize = Flags.getByValSize();
1633 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1639 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1640 CallingConv::ID CallConv, bool isVarArg,
1641 const SmallVectorImpl<ISD::InputArg>
1643 DebugLoc dl, SelectionDAG &DAG,
1644 SmallVectorImpl<SDValue> &InVals)
1646 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1647 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1650 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1656 PPCTargetLowering::LowerFormalArguments_SVR4(
1658 CallingConv::ID CallConv, bool isVarArg,
1659 const SmallVectorImpl<ISD::InputArg>
1661 DebugLoc dl, SelectionDAG &DAG,
1662 SmallVectorImpl<SDValue> &InVals) const {
1664 // 32-bit SVR4 ABI Stack Frame Layout:
1665 // +-----------------------------------+
1666 // +--> | Back chain |
1667 // | +-----------------------------------+
1668 // | | Floating-point register save area |
1669 // | +-----------------------------------+
1670 // | | General register save area |
1671 // | +-----------------------------------+
1672 // | | CR save word |
1673 // | +-----------------------------------+
1674 // | | VRSAVE save word |
1675 // | +-----------------------------------+
1676 // | | Alignment padding |
1677 // | +-----------------------------------+
1678 // | | Vector register save area |
1679 // | +-----------------------------------+
1680 // | | Local variable space |
1681 // | +-----------------------------------+
1682 // | | Parameter list area |
1683 // | +-----------------------------------+
1684 // | | LR save word |
1685 // | +-----------------------------------+
1686 // SP--> +--- | Back chain |
1687 // +-----------------------------------+
1690 // System V Application Binary Interface PowerPC Processor Supplement
1691 // AltiVec Technology Programming Interface Manual
1693 MachineFunction &MF = DAG.getMachineFunction();
1694 MachineFrameInfo *MFI = MF.getFrameInfo();
1695 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1697 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1698 // Potential tail calls could cause overwriting of argument stack slots.
1699 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1700 (CallConv == CallingConv::Fast));
1701 unsigned PtrByteSize = 4;
1703 // Assign locations to all of the incoming arguments.
1704 SmallVector<CCValAssign, 16> ArgLocs;
1705 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1706 getTargetMachine(), ArgLocs, *DAG.getContext());
1708 // Reserve space for the linkage area on the stack.
1709 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1711 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1713 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1714 CCValAssign &VA = ArgLocs[i];
1716 // Arguments stored in registers.
1717 if (VA.isRegLoc()) {
1718 const TargetRegisterClass *RC;
1719 EVT ValVT = VA.getValVT();
1721 switch (ValVT.getSimpleVT().SimpleTy) {
1723 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1725 RC = &PPC::GPRCRegClass;
1728 RC = &PPC::F4RCRegClass;
1731 RC = &PPC::F8RCRegClass;
1737 RC = &PPC::VRRCRegClass;
1741 // Transform the arguments stored in physical registers into virtual ones.
1742 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1743 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1745 InVals.push_back(ArgValue);
1747 // Argument stored in memory.
1748 assert(VA.isMemLoc());
1750 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1751 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1754 // Create load nodes to retrieve arguments from the stack.
1755 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1756 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1757 MachinePointerInfo(),
1758 false, false, false, 0));
1762 // Assign locations to all of the incoming aggregate by value arguments.
1763 // Aggregates passed by value are stored in the local variable space of the
1764 // caller's stack frame, right above the parameter list area.
1765 SmallVector<CCValAssign, 16> ByValArgLocs;
1766 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1767 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1769 // Reserve stack space for the allocations in CCInfo.
1770 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1772 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1774 // Area that is at least reserved in the caller of this function.
1775 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1777 // Set the size that is at least reserved in caller of this function. Tail
1778 // call optimized function's reserved stack space needs to be aligned so that
1779 // taking the difference between two stack areas will result in an aligned
1781 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1784 std::max(MinReservedArea,
1785 PPCFrameLowering::getMinCallFrameSize(false, false));
1787 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1788 getStackAlignment();
1789 unsigned AlignMask = TargetAlign-1;
1790 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1792 FI->setMinReservedArea(MinReservedArea);
1794 SmallVector<SDValue, 8> MemOps;
1796 // If the function takes variable number of arguments, make a frame index for
1797 // the start of the first vararg value... for expansion of llvm.va_start.
1799 static const uint16_t GPArgRegs[] = {
1800 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1801 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1803 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1805 static const uint16_t FPArgRegs[] = {
1806 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1809 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1811 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1813 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1816 // Make room for NumGPArgRegs and NumFPArgRegs.
1817 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1818 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1820 FuncInfo->setVarArgsStackOffset(
1821 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1822 CCInfo.getNextStackOffset(), true));
1824 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1825 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1827 // The fixed integer arguments of a variadic function are stored to the
1828 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1829 // the result of va_next.
1830 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1831 // Get an existing live-in vreg, or add a new one.
1832 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1834 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1836 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1837 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1838 MachinePointerInfo(), false, false, 0);
1839 MemOps.push_back(Store);
1840 // Increment the address by four for the next argument to store
1841 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1842 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1845 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1847 // The double arguments are stored to the VarArgsFrameIndex
1849 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1850 // Get an existing live-in vreg, or add a new one.
1851 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1853 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1855 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1856 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1857 MachinePointerInfo(), false, false, 0);
1858 MemOps.push_back(Store);
1859 // Increment the address by eight for the next argument to store
1860 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1862 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1866 if (!MemOps.empty())
1867 Chain = DAG.getNode(ISD::TokenFactor, dl,
1868 MVT::Other, &MemOps[0], MemOps.size());
1874 PPCTargetLowering::LowerFormalArguments_Darwin(
1876 CallingConv::ID CallConv, bool isVarArg,
1877 const SmallVectorImpl<ISD::InputArg>
1879 DebugLoc dl, SelectionDAG &DAG,
1880 SmallVectorImpl<SDValue> &InVals) const {
1881 // TODO: add description of PPC stack frame format, or at least some docs.
1883 MachineFunction &MF = DAG.getMachineFunction();
1884 MachineFrameInfo *MFI = MF.getFrameInfo();
1885 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1887 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1888 bool isPPC64 = PtrVT == MVT::i64;
1889 // Potential tail calls could cause overwriting of argument stack slots.
1890 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1891 (CallConv == CallingConv::Fast));
1892 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1894 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1895 // Area that is at least reserved in caller of this function.
1896 unsigned MinReservedArea = ArgOffset;
1898 static const uint16_t GPR_32[] = { // 32-bit registers.
1899 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1900 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1902 static const uint16_t GPR_64[] = { // 64-bit registers.
1903 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1904 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1907 static const uint16_t *FPR = GetFPR();
1909 static const uint16_t VR[] = {
1910 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1911 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1914 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1915 const unsigned Num_FPR_Regs = 13;
1916 const unsigned Num_VR_Regs = array_lengthof( VR);
1918 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1920 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
1922 // In 32-bit non-varargs functions, the stack space for vectors is after the
1923 // stack space for non-vectors. We do not use this space unless we have
1924 // too many vectors to fit in registers, something that only occurs in
1925 // constructed examples:), but we have to walk the arglist to figure
1926 // that out...for the pathological case, compute VecArgOffset as the
1927 // start of the vector parameter area. Computing VecArgOffset is the
1928 // entire point of the following loop.
1929 unsigned VecArgOffset = ArgOffset;
1930 if (!isVarArg && !isPPC64) {
1931 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1933 EVT ObjectVT = Ins[ArgNo].VT;
1934 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1936 if (Flags.isByVal()) {
1937 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1938 unsigned ObjSize = Flags.getByValSize();
1940 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1941 VecArgOffset += ArgSize;
1945 switch(ObjectVT.getSimpleVT().SimpleTy) {
1946 default: llvm_unreachable("Unhandled argument type!");
1949 VecArgOffset += isPPC64 ? 8 : 4;
1951 case MVT::i64: // PPC64
1959 // Nothing to do, we're only looking at Nonvector args here.
1964 // We've found where the vector parameter area in memory is. Skip the
1965 // first 12 parameters; these don't use that memory.
1966 VecArgOffset = ((VecArgOffset+15)/16)*16;
1967 VecArgOffset += 12*16;
1969 // Add DAG nodes to load the arguments or copy them out of registers. On
1970 // entry to a function on PPC, the arguments start after the linkage area,
1971 // although the first ones are often in registers.
1973 SmallVector<SDValue, 8> MemOps;
1974 unsigned nAltivecParamsAtEnd = 0;
1975 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1977 bool needsLoad = false;
1978 EVT ObjectVT = Ins[ArgNo].VT;
1979 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1980 unsigned ArgSize = ObjSize;
1981 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1983 unsigned CurArgOffset = ArgOffset;
1985 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1986 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1987 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1988 if (isVarArg || isPPC64) {
1989 MinReservedArea = ((MinReservedArea+15)/16)*16;
1990 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1993 } else nAltivecParamsAtEnd++;
1995 // Calculate min reserved area.
1996 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2000 // FIXME the codegen can be much improved in some cases.
2001 // We do not have to keep everything in memory.
2002 if (Flags.isByVal()) {
2003 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2004 ObjSize = Flags.getByValSize();
2005 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2006 // Objects of size 1 and 2 are right justified, everything else is
2007 // left justified. This means the memory address is adjusted forwards.
2008 if (ObjSize==1 || ObjSize==2) {
2009 CurArgOffset = CurArgOffset + (4 - ObjSize);
2011 // The value of the object is its address.
2012 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2013 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2014 InVals.push_back(FIN);
2015 if (ObjSize==1 || ObjSize==2) {
2016 if (GPR_idx != Num_GPR_Regs) {
2019 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2021 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2022 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2023 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2024 MachinePointerInfo(),
2025 ObjSize==1 ? MVT::i8 : MVT::i16,
2027 MemOps.push_back(Store);
2031 ArgOffset += PtrByteSize;
2035 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2036 // Store whatever pieces of the object are in registers
2037 // to memory. ArgVal will be address of the beginning of
2039 if (GPR_idx != Num_GPR_Regs) {
2042 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2044 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2045 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2046 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2047 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2048 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2049 MachinePointerInfo(),
2051 MemOps.push_back(Store);
2053 ArgOffset += PtrByteSize;
2055 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2062 switch (ObjectVT.getSimpleVT().SimpleTy) {
2063 default: llvm_unreachable("Unhandled argument type!");
2066 if (GPR_idx != Num_GPR_Regs) {
2067 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2068 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2072 ArgSize = PtrByteSize;
2074 // All int arguments reserve stack space in the Darwin ABI.
2075 ArgOffset += PtrByteSize;
2079 case MVT::i64: // PPC64
2080 if (GPR_idx != Num_GPR_Regs) {
2081 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2082 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2084 if (ObjectVT == MVT::i32) {
2085 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2086 // value to MVT::i64 and then truncate to the correct register size.
2088 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2089 DAG.getValueType(ObjectVT));
2090 else if (Flags.isZExt())
2091 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2092 DAG.getValueType(ObjectVT));
2094 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2100 ArgSize = PtrByteSize;
2102 // All int arguments reserve stack space in the Darwin ABI.
2108 // Every 4 bytes of argument space consumes one of the GPRs available for
2109 // argument passing.
2110 if (GPR_idx != Num_GPR_Regs) {
2112 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2115 if (FPR_idx != Num_FPR_Regs) {
2118 if (ObjectVT == MVT::f32)
2119 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2121 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2123 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2129 // All FP arguments reserve stack space in the Darwin ABI.
2130 ArgOffset += isPPC64 ? 8 : ObjSize;
2136 // Note that vector arguments in registers don't reserve stack space,
2137 // except in varargs functions.
2138 if (VR_idx != Num_VR_Regs) {
2139 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2140 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2142 while ((ArgOffset % 16) != 0) {
2143 ArgOffset += PtrByteSize;
2144 if (GPR_idx != Num_GPR_Regs)
2148 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2152 if (!isVarArg && !isPPC64) {
2153 // Vectors go after all the nonvectors.
2154 CurArgOffset = VecArgOffset;
2157 // Vectors are aligned.
2158 ArgOffset = ((ArgOffset+15)/16)*16;
2159 CurArgOffset = ArgOffset;
2167 // We need to load the argument to a virtual register if we determined above
2168 // that we ran out of physical registers of the appropriate type.
2170 int FI = MFI->CreateFixedObject(ObjSize,
2171 CurArgOffset + (ArgSize - ObjSize),
2173 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2174 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2175 false, false, false, 0);
2178 InVals.push_back(ArgVal);
2181 // Set the size that is at least reserved in caller of this function. Tail
2182 // call optimized function's reserved stack space needs to be aligned so that
2183 // taking the difference between two stack areas will result in an aligned
2185 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2186 // Add the Altivec parameters at the end, if needed.
2187 if (nAltivecParamsAtEnd) {
2188 MinReservedArea = ((MinReservedArea+15)/16)*16;
2189 MinReservedArea += 16*nAltivecParamsAtEnd;
2192 std::max(MinReservedArea,
2193 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2194 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2195 getStackAlignment();
2196 unsigned AlignMask = TargetAlign-1;
2197 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2198 FI->setMinReservedArea(MinReservedArea);
2200 // If the function takes variable number of arguments, make a frame index for
2201 // the start of the first vararg value... for expansion of llvm.va_start.
2203 int Depth = ArgOffset;
2205 FuncInfo->setVarArgsFrameIndex(
2206 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2208 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2210 // If this function is vararg, store any remaining integer argument regs
2211 // to their spots on the stack so that they may be loaded by deferencing the
2212 // result of va_next.
2213 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2217 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2219 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2221 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2222 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2223 MachinePointerInfo(), false, false, 0);
2224 MemOps.push_back(Store);
2225 // Increment the address by four for the next argument to store
2226 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2227 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2231 if (!MemOps.empty())
2232 Chain = DAG.getNode(ISD::TokenFactor, dl,
2233 MVT::Other, &MemOps[0], MemOps.size());
2238 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2239 /// linkage area for the Darwin ABI.
2241 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2245 const SmallVectorImpl<ISD::OutputArg>
2247 const SmallVectorImpl<SDValue> &OutVals,
2248 unsigned &nAltivecParamsAtEnd) {
2249 // Count how many bytes are to be pushed on the stack, including the linkage
2250 // area, and parameter passing area. We start with 24/48 bytes, which is
2251 // prereserved space for [SP][CR][LR][3 x unused].
2252 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2253 unsigned NumOps = Outs.size();
2254 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2256 // Add up all the space actually used.
2257 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2258 // they all go in registers, but we must reserve stack space for them for
2259 // possible use by the caller. In varargs or 64-bit calls, parameters are
2260 // assigned stack space in order, with padding so Altivec parameters are
2262 nAltivecParamsAtEnd = 0;
2263 for (unsigned i = 0; i != NumOps; ++i) {
2264 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2265 EVT ArgVT = Outs[i].VT;
2266 // Varargs Altivec parameters are padded to a 16 byte boundary.
2267 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2268 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2269 if (!isVarArg && !isPPC64) {
2270 // Non-varargs Altivec parameters go after all the non-Altivec
2271 // parameters; handle those later so we know how much padding we need.
2272 nAltivecParamsAtEnd++;
2275 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2276 NumBytes = ((NumBytes+15)/16)*16;
2278 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2281 // Allow for Altivec parameters at the end, if needed.
2282 if (nAltivecParamsAtEnd) {
2283 NumBytes = ((NumBytes+15)/16)*16;
2284 NumBytes += 16*nAltivecParamsAtEnd;
2287 // The prolog code of the callee may store up to 8 GPR argument registers to
2288 // the stack, allowing va_start to index over them in memory if its varargs.
2289 // Because we cannot tell if this is needed on the caller side, we have to
2290 // conservatively assume that it is needed. As such, make sure we have at
2291 // least enough stack space for the caller to store the 8 GPRs.
2292 NumBytes = std::max(NumBytes,
2293 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2295 // Tail call needs the stack to be aligned.
2296 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2297 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2298 getFrameLowering()->getStackAlignment();
2299 unsigned AlignMask = TargetAlign-1;
2300 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2306 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2307 /// adjusted to accommodate the arguments for the tailcall.
2308 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2309 unsigned ParamSize) {
2311 if (!isTailCall) return 0;
2313 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2314 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2315 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2316 // Remember only if the new adjustement is bigger.
2317 if (SPDiff < FI->getTailCallSPDelta())
2318 FI->setTailCallSPDelta(SPDiff);
2323 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2324 /// for tail call optimization. Targets which want to do tail call
2325 /// optimization should implement this function.
2327 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2328 CallingConv::ID CalleeCC,
2330 const SmallVectorImpl<ISD::InputArg> &Ins,
2331 SelectionDAG& DAG) const {
2332 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2335 // Variable argument functions are not supported.
2339 MachineFunction &MF = DAG.getMachineFunction();
2340 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2341 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2342 // Functions containing by val parameters are not supported.
2343 for (unsigned i = 0; i != Ins.size(); i++) {
2344 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2345 if (Flags.isByVal()) return false;
2348 // Non PIC/GOT tail calls are supported.
2349 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2352 // At the moment we can only do local tail calls (in same module, hidden
2353 // or protected) if we are generating PIC.
2354 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2355 return G->getGlobal()->hasHiddenVisibility()
2356 || G->getGlobal()->hasProtectedVisibility();
2362 /// isCallCompatibleAddress - Return the immediate to use if the specified
2363 /// 32-bit value is representable in the immediate field of a BxA instruction.
2364 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2365 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2368 int Addr = C->getZExtValue();
2369 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2370 (Addr << 6 >> 6) != Addr)
2371 return 0; // Top 6 bits have to be sext of immediate.
2373 return DAG.getConstant((int)C->getZExtValue() >> 2,
2374 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2379 struct TailCallArgumentInfo {
2384 TailCallArgumentInfo() : FrameIdx(0) {}
2389 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2391 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2393 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2394 SmallVector<SDValue, 8> &MemOpChains,
2396 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2397 SDValue Arg = TailCallArgs[i].Arg;
2398 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2399 int FI = TailCallArgs[i].FrameIdx;
2400 // Store relative to framepointer.
2401 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2402 MachinePointerInfo::getFixedStack(FI),
2407 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2408 /// the appropriate stack slot for the tail call optimized function call.
2409 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2410 MachineFunction &MF,
2419 // Calculate the new stack slot for the return address.
2420 int SlotSize = isPPC64 ? 8 : 4;
2421 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2423 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2424 NewRetAddrLoc, true);
2425 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2426 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2427 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2428 MachinePointerInfo::getFixedStack(NewRetAddr),
2431 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2432 // slot as the FP is never overwritten.
2435 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2436 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2438 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2439 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2440 MachinePointerInfo::getFixedStack(NewFPIdx),
2447 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2448 /// the position of the argument.
2450 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2451 SDValue Arg, int SPDiff, unsigned ArgOffset,
2452 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2453 int Offset = ArgOffset + SPDiff;
2454 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2455 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2456 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2457 SDValue FIN = DAG.getFrameIndex(FI, VT);
2458 TailCallArgumentInfo Info;
2460 Info.FrameIdxOp = FIN;
2462 TailCallArguments.push_back(Info);
2465 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2466 /// stack slot. Returns the chain as result and the loaded frame pointers in
2467 /// LROpOut/FPOpout. Used when tail calling.
2468 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2474 DebugLoc dl) const {
2476 // Load the LR and FP stack slot for later adjusting.
2477 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2478 LROpOut = getReturnAddrFrameIndex(DAG);
2479 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2480 false, false, false, 0);
2481 Chain = SDValue(LROpOut.getNode(), 1);
2483 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2484 // slot as the FP is never overwritten.
2486 FPOpOut = getFramePointerFrameIndex(DAG);
2487 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2488 false, false, false, 0);
2489 Chain = SDValue(FPOpOut.getNode(), 1);
2495 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2496 /// by "Src" to address "Dst" of size "Size". Alignment information is
2497 /// specified by the specific parameter attribute. The copy will be passed as
2498 /// a byval function parameter.
2499 /// Sometimes what we are copying is the end of a larger object, the part that
2500 /// does not fit in registers.
2502 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2503 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2505 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2506 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2507 false, false, MachinePointerInfo(0),
2508 MachinePointerInfo(0));
2511 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2514 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2515 SDValue Arg, SDValue PtrOff, int SPDiff,
2516 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2517 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2518 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2520 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2525 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2527 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2528 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2529 DAG.getConstant(ArgOffset, PtrVT));
2531 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2532 MachinePointerInfo(), false, false, 0));
2533 // Calculate and remember argument location.
2534 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2539 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2540 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2541 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2542 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2543 MachineFunction &MF = DAG.getMachineFunction();
2545 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2546 // might overwrite each other in case of tail call optimization.
2547 SmallVector<SDValue, 8> MemOpChains2;
2548 // Do not flag preceding copytoreg stuff together with the following stuff.
2550 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2552 if (!MemOpChains2.empty())
2553 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2554 &MemOpChains2[0], MemOpChains2.size());
2556 // Store the return address to the appropriate stack slot.
2557 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2558 isPPC64, isDarwinABI, dl);
2560 // Emit callseq_end just before tailcall node.
2561 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2562 DAG.getIntPtrConstant(0, true), InFlag);
2563 InFlag = Chain.getValue(1);
2567 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2568 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2569 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2570 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2571 const PPCSubtarget &PPCSubTarget) {
2573 bool isPPC64 = PPCSubTarget.isPPC64();
2574 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2576 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2577 NodeTys.push_back(MVT::Other); // Returns a chain
2578 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2580 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2582 bool needIndirectCall = true;
2583 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2584 // If this is an absolute destination address, use the munged value.
2585 Callee = SDValue(Dest, 0);
2586 needIndirectCall = false;
2589 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2590 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2591 // Use indirect calls for ALL functions calls in JIT mode, since the
2592 // far-call stubs may be outside relocation limits for a BL instruction.
2593 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2594 unsigned OpFlags = 0;
2595 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2596 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2597 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2598 (G->getGlobal()->isDeclaration() ||
2599 G->getGlobal()->isWeakForLinker())) {
2600 // PC-relative references to external symbols should go through $stub,
2601 // unless we're building with the leopard linker or later, which
2602 // automatically synthesizes these stubs.
2603 OpFlags = PPCII::MO_DARWIN_STUB;
2606 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2607 // every direct call is) turn it into a TargetGlobalAddress /
2608 // TargetExternalSymbol node so that legalize doesn't hack it.
2609 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2610 Callee.getValueType(),
2612 needIndirectCall = false;
2616 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2617 unsigned char OpFlags = 0;
2619 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2620 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2621 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2622 // PC-relative references to external symbols should go through $stub,
2623 // unless we're building with the leopard linker or later, which
2624 // automatically synthesizes these stubs.
2625 OpFlags = PPCII::MO_DARWIN_STUB;
2628 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2630 needIndirectCall = false;
2633 if (needIndirectCall) {
2634 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2635 // to do the call, we can't use PPCISD::CALL.
2636 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2638 if (isSVR4ABI && isPPC64) {
2639 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2640 // entry point, but to the function descriptor (the function entry point
2641 // address is part of the function descriptor though).
2642 // The function descriptor is a three doubleword structure with the
2643 // following fields: function entry point, TOC base address and
2644 // environment pointer.
2645 // Thus for a call through a function pointer, the following actions need
2647 // 1. Save the TOC of the caller in the TOC save area of its stack
2648 // frame (this is done in LowerCall_Darwin()).
2649 // 2. Load the address of the function entry point from the function
2651 // 3. Load the TOC of the callee from the function descriptor into r2.
2652 // 4. Load the environment pointer from the function descriptor into
2654 // 5. Branch to the function entry point address.
2655 // 6. On return of the callee, the TOC of the caller needs to be
2656 // restored (this is done in FinishCall()).
2658 // All those operations are flagged together to ensure that no other
2659 // operations can be scheduled in between. E.g. without flagging the
2660 // operations together, a TOC access in the caller could be scheduled
2661 // between the load of the callee TOC and the branch to the callee, which
2662 // results in the TOC access going through the TOC of the callee instead
2663 // of going through the TOC of the caller, which leads to incorrect code.
2665 // Load the address of the function entry point from the function
2667 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2668 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2669 InFlag.getNode() ? 3 : 2);
2670 Chain = LoadFuncPtr.getValue(1);
2671 InFlag = LoadFuncPtr.getValue(2);
2673 // Load environment pointer into r11.
2674 // Offset of the environment pointer within the function descriptor.
2675 SDValue PtrOff = DAG.getIntPtrConstant(16);
2677 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2678 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2680 Chain = LoadEnvPtr.getValue(1);
2681 InFlag = LoadEnvPtr.getValue(2);
2683 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2685 Chain = EnvVal.getValue(0);
2686 InFlag = EnvVal.getValue(1);
2688 // Load TOC of the callee into r2. We are using a target-specific load
2689 // with r2 hard coded, because the result of a target-independent load
2690 // would never go directly into r2, since r2 is a reserved register (which
2691 // prevents the register allocator from allocating it), resulting in an
2692 // additional register being allocated and an unnecessary move instruction
2694 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2695 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2697 Chain = LoadTOCPtr.getValue(0);
2698 InFlag = LoadTOCPtr.getValue(1);
2700 MTCTROps[0] = Chain;
2701 MTCTROps[1] = LoadFuncPtr;
2702 MTCTROps[2] = InFlag;
2705 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2706 2 + (InFlag.getNode() != 0));
2707 InFlag = Chain.getValue(1);
2710 NodeTys.push_back(MVT::Other);
2711 NodeTys.push_back(MVT::Glue);
2712 Ops.push_back(Chain);
2713 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2715 // Add CTR register as callee so a bctr can be emitted later.
2717 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2720 // If this is a direct call, pass the chain and the callee.
2721 if (Callee.getNode()) {
2722 Ops.push_back(Chain);
2723 Ops.push_back(Callee);
2725 // If this is a tail call add stack pointer delta.
2727 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2729 // Add argument registers to the end of the list so that they are known live
2731 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2732 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2733 RegsToPass[i].second.getValueType()));
2739 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2740 CallingConv::ID CallConv, bool isVarArg,
2741 const SmallVectorImpl<ISD::InputArg> &Ins,
2742 DebugLoc dl, SelectionDAG &DAG,
2743 SmallVectorImpl<SDValue> &InVals) const {
2745 SmallVector<CCValAssign, 16> RVLocs;
2746 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2747 getTargetMachine(), RVLocs, *DAG.getContext());
2748 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2750 // Copy all of the result registers out of their specified physreg.
2751 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2752 CCValAssign &VA = RVLocs[i];
2753 EVT VT = VA.getValVT();
2754 assert(VA.isRegLoc() && "Can only return in registers!");
2755 Chain = DAG.getCopyFromReg(Chain, dl,
2756 VA.getLocReg(), VT, InFlag).getValue(1);
2757 InVals.push_back(Chain.getValue(0));
2758 InFlag = Chain.getValue(2);
2765 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2766 bool isTailCall, bool isVarArg,
2768 SmallVector<std::pair<unsigned, SDValue>, 8>
2770 SDValue InFlag, SDValue Chain,
2772 int SPDiff, unsigned NumBytes,
2773 const SmallVectorImpl<ISD::InputArg> &Ins,
2774 SmallVectorImpl<SDValue> &InVals) const {
2775 std::vector<EVT> NodeTys;
2776 SmallVector<SDValue, 8> Ops;
2777 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2778 isTailCall, RegsToPass, Ops, NodeTys,
2781 // When performing tail call optimization the callee pops its arguments off
2782 // the stack. Account for this here so these bytes can be pushed back on in
2783 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2784 int BytesCalleePops =
2785 (CallConv == CallingConv::Fast &&
2786 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
2788 // Add a register mask operand representing the call-preserved registers.
2789 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2790 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2791 assert(Mask && "Missing call preserved mask for calling convention");
2792 Ops.push_back(DAG.getRegisterMask(Mask));
2794 if (InFlag.getNode())
2795 Ops.push_back(InFlag);
2799 // If this is the first return lowered for this function, add the regs
2800 // to the liveout set for the function.
2801 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2802 SmallVector<CCValAssign, 16> RVLocs;
2803 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2804 getTargetMachine(), RVLocs, *DAG.getContext());
2805 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2806 for (unsigned i = 0; i != RVLocs.size(); ++i)
2807 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2810 assert(((Callee.getOpcode() == ISD::Register &&
2811 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2812 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2813 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2814 isa<ConstantSDNode>(Callee)) &&
2815 "Expecting an global address, external symbol, absolute value or register");
2817 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2820 // Add a NOP immediately after the branch instruction when using the 64-bit
2821 // SVR4 ABI. At link time, if caller and callee are in a different module and
2822 // thus have a different TOC, the call will be replaced with a call to a stub
2823 // function which saves the current TOC, loads the TOC of the callee and
2824 // branches to the callee. The NOP will be replaced with a load instruction
2825 // which restores the TOC of the caller from the TOC save slot of the current
2826 // stack frame. If caller and callee belong to the same module (and have the
2827 // same TOC), the NOP will remain unchanged.
2829 bool needsTOCRestore = false;
2830 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2831 if (CallOpc == PPCISD::BCTRL_SVR4) {
2832 // This is a call through a function pointer.
2833 // Restore the caller TOC from the save area into R2.
2834 // See PrepareCall() for more information about calls through function
2835 // pointers in the 64-bit SVR4 ABI.
2836 // We are using a target-specific load with r2 hard coded, because the
2837 // result of a target-independent load would never go directly into r2,
2838 // since r2 is a reserved register (which prevents the register allocator
2839 // from allocating it), resulting in an additional register being
2840 // allocated and an unnecessary move instruction being generated.
2841 needsTOCRestore = true;
2842 } else if (CallOpc == PPCISD::CALL_SVR4) {
2843 // Otherwise insert NOP.
2844 CallOpc = PPCISD::CALL_NOP_SVR4;
2848 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2849 InFlag = Chain.getValue(1);
2851 if (needsTOCRestore) {
2852 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2853 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2854 InFlag = Chain.getValue(1);
2857 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2858 DAG.getIntPtrConstant(BytesCalleePops, true),
2861 InFlag = Chain.getValue(1);
2863 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2864 Ins, dl, DAG, InVals);
2868 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2869 SmallVectorImpl<SDValue> &InVals) const {
2870 SelectionDAG &DAG = CLI.DAG;
2871 DebugLoc &dl = CLI.DL;
2872 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2873 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2874 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2875 SDValue Chain = CLI.Chain;
2876 SDValue Callee = CLI.Callee;
2877 bool &isTailCall = CLI.IsTailCall;
2878 CallingConv::ID CallConv = CLI.CallConv;
2879 bool isVarArg = CLI.IsVarArg;
2882 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2885 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2886 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2887 isTailCall, Outs, OutVals, Ins,
2890 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2891 isTailCall, Outs, OutVals, Ins,
2896 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2897 CallingConv::ID CallConv, bool isVarArg,
2899 const SmallVectorImpl<ISD::OutputArg> &Outs,
2900 const SmallVectorImpl<SDValue> &OutVals,
2901 const SmallVectorImpl<ISD::InputArg> &Ins,
2902 DebugLoc dl, SelectionDAG &DAG,
2903 SmallVectorImpl<SDValue> &InVals) const {
2904 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2905 // of the 32-bit SVR4 ABI stack frame layout.
2907 assert((CallConv == CallingConv::C ||
2908 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2910 unsigned PtrByteSize = 4;
2912 MachineFunction &MF = DAG.getMachineFunction();
2914 // Mark this function as potentially containing a function that contains a
2915 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2916 // and restoring the callers stack pointer in this functions epilog. This is
2917 // done because by tail calling the called function might overwrite the value
2918 // in this function's (MF) stack pointer stack slot 0(SP).
2919 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2920 CallConv == CallingConv::Fast)
2921 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2923 // Count how many bytes are to be pushed on the stack, including the linkage
2924 // area, parameter list area and the part of the local variable space which
2925 // contains copies of aggregates which are passed by value.
2927 // Assign locations to all of the outgoing arguments.
2928 SmallVector<CCValAssign, 16> ArgLocs;
2929 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2930 getTargetMachine(), ArgLocs, *DAG.getContext());
2932 // Reserve space for the linkage area on the stack.
2933 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2936 // Handle fixed and variable vector arguments differently.
2937 // Fixed vector arguments go into registers as long as registers are
2938 // available. Variable vector arguments always go into memory.
2939 unsigned NumArgs = Outs.size();
2941 for (unsigned i = 0; i != NumArgs; ++i) {
2942 MVT ArgVT = Outs[i].VT;
2943 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2946 if (Outs[i].IsFixed) {
2947 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2950 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2956 errs() << "Call operand #" << i << " has unhandled type "
2957 << EVT(ArgVT).getEVTString() << "\n";
2959 llvm_unreachable(0);
2963 // All arguments are treated the same.
2964 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2967 // Assign locations to all of the outgoing aggregate by value arguments.
2968 SmallVector<CCValAssign, 16> ByValArgLocs;
2969 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2970 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2972 // Reserve stack space for the allocations in CCInfo.
2973 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2975 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2977 // Size of the linkage area, parameter list area and the part of the local
2978 // space variable where copies of aggregates which are passed by value are
2980 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2982 // Calculate by how many bytes the stack has to be adjusted in case of tail
2983 // call optimization.
2984 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2986 // Adjust the stack pointer for the new arguments...
2987 // These operations are automatically eliminated by the prolog/epilog pass
2988 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2989 SDValue CallSeqStart = Chain;
2991 // Load the return address and frame pointer so it can be moved somewhere else
2994 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2997 // Set up a copy of the stack pointer for use loading and storing any
2998 // arguments that may not fit in the registers available for argument
3000 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3002 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3003 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3004 SmallVector<SDValue, 8> MemOpChains;
3006 bool seenFloatArg = false;
3007 // Walk the register/memloc assignments, inserting copies/loads.
3008 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3011 CCValAssign &VA = ArgLocs[i];
3012 SDValue Arg = OutVals[i];
3013 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3015 if (Flags.isByVal()) {
3016 // Argument is an aggregate which is passed by value, thus we need to
3017 // create a copy of it in the local variable space of the current stack
3018 // frame (which is the stack frame of the caller) and pass the address of
3019 // this copy to the callee.
3020 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3021 CCValAssign &ByValVA = ByValArgLocs[j++];
3022 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3024 // Memory reserved in the local variable space of the callers stack frame.
3025 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3027 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3028 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3030 // Create a copy of the argument in the local area of the current
3032 SDValue MemcpyCall =
3033 CreateCopyOfByValArgument(Arg, PtrOff,
3034 CallSeqStart.getNode()->getOperand(0),
3037 // This must go outside the CALLSEQ_START..END.
3038 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3039 CallSeqStart.getNode()->getOperand(1));
3040 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3041 NewCallSeqStart.getNode());
3042 Chain = CallSeqStart = NewCallSeqStart;
3044 // Pass the address of the aggregate copy on the stack either in a
3045 // physical register or in the parameter list area of the current stack
3046 // frame to the callee.
3050 if (VA.isRegLoc()) {
3051 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3052 // Put argument in a physical register.
3053 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3055 // Put argument in the parameter list area of the current stack frame.
3056 assert(VA.isMemLoc());
3057 unsigned LocMemOffset = VA.getLocMemOffset();
3060 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3061 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3063 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3064 MachinePointerInfo(),
3067 // Calculate and remember argument location.
3068 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3074 if (!MemOpChains.empty())
3075 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3076 &MemOpChains[0], MemOpChains.size());
3078 // Set CR6 to true if this is a vararg call with floating args passed in
3081 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3083 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3086 // Build a sequence of copy-to-reg nodes chained together with token chain
3087 // and flag operands which copy the outgoing args into the appropriate regs.
3089 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3090 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3091 RegsToPass[i].second, InFlag);
3092 InFlag = Chain.getValue(1);
3096 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3097 false, TailCallArguments);
3099 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3100 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3105 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3106 CallingConv::ID CallConv, bool isVarArg,
3108 const SmallVectorImpl<ISD::OutputArg> &Outs,
3109 const SmallVectorImpl<SDValue> &OutVals,
3110 const SmallVectorImpl<ISD::InputArg> &Ins,
3111 DebugLoc dl, SelectionDAG &DAG,
3112 SmallVectorImpl<SDValue> &InVals) const {
3114 unsigned NumOps = Outs.size();
3116 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3117 bool isPPC64 = PtrVT == MVT::i64;
3118 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3120 MachineFunction &MF = DAG.getMachineFunction();
3122 // Mark this function as potentially containing a function that contains a
3123 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3124 // and restoring the callers stack pointer in this functions epilog. This is
3125 // done because by tail calling the called function might overwrite the value
3126 // in this function's (MF) stack pointer stack slot 0(SP).
3127 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3128 CallConv == CallingConv::Fast)
3129 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3131 unsigned nAltivecParamsAtEnd = 0;
3133 // Count how many bytes are to be pushed on the stack, including the linkage
3134 // area, and parameter passing area. We start with 24/48 bytes, which is
3135 // prereserved space for [SP][CR][LR][3 x unused].
3137 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3139 nAltivecParamsAtEnd);
3141 // Calculate by how many bytes the stack has to be adjusted in case of tail
3142 // call optimization.
3143 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3145 // To protect arguments on the stack from being clobbered in a tail call,
3146 // force all the loads to happen before doing any other lowering.
3148 Chain = DAG.getStackArgumentTokenFactor(Chain);
3150 // Adjust the stack pointer for the new arguments...
3151 // These operations are automatically eliminated by the prolog/epilog pass
3152 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3153 SDValue CallSeqStart = Chain;
3155 // Load the return address and frame pointer so it can be move somewhere else
3158 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3161 // Set up a copy of the stack pointer for use loading and storing any
3162 // arguments that may not fit in the registers available for argument
3166 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3168 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3170 // Figure out which arguments are going to go in registers, and which in
3171 // memory. Also, if this is a vararg function, floating point operations
3172 // must be stored to our stack, and loaded into integer regs as well, if
3173 // any integer regs are available for argument passing.
3174 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3175 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3177 static const uint16_t GPR_32[] = { // 32-bit registers.
3178 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3179 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3181 static const uint16_t GPR_64[] = { // 64-bit registers.
3182 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3183 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3185 static const uint16_t *FPR = GetFPR();
3187 static const uint16_t VR[] = {
3188 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3189 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3191 const unsigned NumGPRs = array_lengthof(GPR_32);
3192 const unsigned NumFPRs = 13;
3193 const unsigned NumVRs = array_lengthof(VR);
3195 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3197 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3198 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3200 SmallVector<SDValue, 8> MemOpChains;
3201 for (unsigned i = 0; i != NumOps; ++i) {
3202 SDValue Arg = OutVals[i];
3203 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3205 // PtrOff will be used to store the current argument to the stack if a
3206 // register cannot be found for it.
3209 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3211 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3213 // On PPC64, promote integers to 64-bit values.
3214 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3215 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3216 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3217 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3220 // FIXME memcpy is used way more than necessary. Correctness first.
3221 if (Flags.isByVal()) {
3222 unsigned Size = Flags.getByValSize();
3223 if (Size==1 || Size==2) {
3224 // Very small objects are passed right-justified.
3225 // Everything else is passed left-justified.
3226 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3227 if (GPR_idx != NumGPRs) {
3228 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3229 MachinePointerInfo(), VT,
3231 MemOpChains.push_back(Load.getValue(1));
3232 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3234 ArgOffset += PtrByteSize;
3236 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3237 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3238 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3239 CallSeqStart.getNode()->getOperand(0),
3241 // This must go outside the CALLSEQ_START..END.
3242 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3243 CallSeqStart.getNode()->getOperand(1));
3244 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3245 NewCallSeqStart.getNode());
3246 Chain = CallSeqStart = NewCallSeqStart;
3247 ArgOffset += PtrByteSize;
3251 // Copy entire object into memory. There are cases where gcc-generated
3252 // code assumes it is there, even if it could be put entirely into
3253 // registers. (This is not what the doc says.)
3254 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3255 CallSeqStart.getNode()->getOperand(0),
3257 // This must go outside the CALLSEQ_START..END.
3258 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3259 CallSeqStart.getNode()->getOperand(1));
3260 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3261 Chain = CallSeqStart = NewCallSeqStart;
3262 // And copy the pieces of it that fit into registers.
3263 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3264 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3265 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3266 if (GPR_idx != NumGPRs) {
3267 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3268 MachinePointerInfo(),
3269 false, false, false, 0);
3270 MemOpChains.push_back(Load.getValue(1));
3271 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3272 ArgOffset += PtrByteSize;
3274 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3281 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3282 default: llvm_unreachable("Unexpected ValueType for argument!");
3285 if (GPR_idx != NumGPRs) {
3286 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3288 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3289 isPPC64, isTailCall, false, MemOpChains,
3290 TailCallArguments, dl);
3292 ArgOffset += PtrByteSize;
3296 if (FPR_idx != NumFPRs) {
3297 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3300 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3301 MachinePointerInfo(), false, false, 0);
3302 MemOpChains.push_back(Store);
3304 // Float varargs are always shadowed in available integer registers
3305 if (GPR_idx != NumGPRs) {
3306 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3307 MachinePointerInfo(), false, false,
3309 MemOpChains.push_back(Load.getValue(1));
3310 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3312 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3313 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3314 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3315 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3316 MachinePointerInfo(),
3317 false, false, false, 0);
3318 MemOpChains.push_back(Load.getValue(1));
3319 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3322 // If we have any FPRs remaining, we may also have GPRs remaining.
3323 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3325 if (GPR_idx != NumGPRs)
3327 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3328 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3332 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3333 isPPC64, isTailCall, false, MemOpChains,
3334 TailCallArguments, dl);
3339 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3346 // These go aligned on the stack, or in the corresponding R registers
3347 // when within range. The Darwin PPC ABI doc claims they also go in
3348 // V registers; in fact gcc does this only for arguments that are
3349 // prototyped, not for those that match the ... We do it for all
3350 // arguments, seems to work.
3351 while (ArgOffset % 16 !=0) {
3352 ArgOffset += PtrByteSize;
3353 if (GPR_idx != NumGPRs)
3356 // We could elide this store in the case where the object fits
3357 // entirely in R registers. Maybe later.
3358 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3359 DAG.getConstant(ArgOffset, PtrVT));
3360 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3361 MachinePointerInfo(), false, false, 0);
3362 MemOpChains.push_back(Store);
3363 if (VR_idx != NumVRs) {
3364 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3365 MachinePointerInfo(),
3366 false, false, false, 0);
3367 MemOpChains.push_back(Load.getValue(1));
3368 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3371 for (unsigned i=0; i<16; i+=PtrByteSize) {
3372 if (GPR_idx == NumGPRs)
3374 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3375 DAG.getConstant(i, PtrVT));
3376 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3377 false, false, false, 0);
3378 MemOpChains.push_back(Load.getValue(1));
3379 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3384 // Non-varargs Altivec params generally go in registers, but have
3385 // stack space allocated at the end.
3386 if (VR_idx != NumVRs) {
3387 // Doesn't have GPR space allocated.
3388 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3389 } else if (nAltivecParamsAtEnd==0) {
3390 // We are emitting Altivec params in order.
3391 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3392 isPPC64, isTailCall, true, MemOpChains,
3393 TailCallArguments, dl);
3399 // If all Altivec parameters fit in registers, as they usually do,
3400 // they get stack space following the non-Altivec parameters. We
3401 // don't track this here because nobody below needs it.
3402 // If there are more Altivec parameters than fit in registers emit
3404 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3406 // Offset is aligned; skip 1st 12 params which go in V registers.
3407 ArgOffset = ((ArgOffset+15)/16)*16;
3409 for (unsigned i = 0; i != NumOps; ++i) {
3410 SDValue Arg = OutVals[i];
3411 EVT ArgType = Outs[i].VT;
3412 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3413 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3416 // We are emitting Altivec params in order.
3417 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3418 isPPC64, isTailCall, true, MemOpChains,
3419 TailCallArguments, dl);
3426 if (!MemOpChains.empty())
3427 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3428 &MemOpChains[0], MemOpChains.size());
3430 // Check if this is an indirect call (MTCTR/BCTRL).
3431 // See PrepareCall() for more information about calls through function
3432 // pointers in the 64-bit SVR4 ABI.
3433 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3434 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3435 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3436 !isBLACompatibleAddress(Callee, DAG)) {
3437 // Load r2 into a virtual register and store it to the TOC save area.
3438 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3439 // TOC save area offset.
3440 SDValue PtrOff = DAG.getIntPtrConstant(40);
3441 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3442 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3446 // On Darwin, R12 must contain the address of an indirect callee. This does
3447 // not mean the MTCTR instruction must use R12; it's easier to model this as
3448 // an extra parameter, so do that.
3450 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3451 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3452 !isBLACompatibleAddress(Callee, DAG))
3453 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3454 PPC::R12), Callee));
3456 // Build a sequence of copy-to-reg nodes chained together with token chain
3457 // and flag operands which copy the outgoing args into the appropriate regs.
3459 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3460 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3461 RegsToPass[i].second, InFlag);
3462 InFlag = Chain.getValue(1);
3466 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3467 FPOp, true, TailCallArguments);
3469 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3470 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3475 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3476 MachineFunction &MF, bool isVarArg,
3477 const SmallVectorImpl<ISD::OutputArg> &Outs,
3478 LLVMContext &Context) const {
3479 SmallVector<CCValAssign, 16> RVLocs;
3480 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3482 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3486 PPCTargetLowering::LowerReturn(SDValue Chain,
3487 CallingConv::ID CallConv, bool isVarArg,
3488 const SmallVectorImpl<ISD::OutputArg> &Outs,
3489 const SmallVectorImpl<SDValue> &OutVals,
3490 DebugLoc dl, SelectionDAG &DAG) const {
3492 SmallVector<CCValAssign, 16> RVLocs;
3493 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3494 getTargetMachine(), RVLocs, *DAG.getContext());
3495 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3497 // If this is the first return lowered for this function, add the regs to the
3498 // liveout set for the function.
3499 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3500 for (unsigned i = 0; i != RVLocs.size(); ++i)
3501 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3506 // Copy the result values into the output registers.
3507 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3508 CCValAssign &VA = RVLocs[i];
3509 assert(VA.isRegLoc() && "Can only return in registers!");
3510 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3512 Flag = Chain.getValue(1);
3516 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3518 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3521 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3522 const PPCSubtarget &Subtarget) const {
3523 // When we pop the dynamic allocation we need to restore the SP link.
3524 DebugLoc dl = Op.getDebugLoc();
3526 // Get the corect type for pointers.
3527 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3529 // Construct the stack pointer operand.
3530 bool isPPC64 = Subtarget.isPPC64();
3531 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3532 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3534 // Get the operands for the STACKRESTORE.
3535 SDValue Chain = Op.getOperand(0);
3536 SDValue SaveSP = Op.getOperand(1);
3538 // Load the old link SP.
3539 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3540 MachinePointerInfo(),
3541 false, false, false, 0);
3543 // Restore the stack pointer.
3544 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3546 // Store the old link SP.
3547 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3554 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3555 MachineFunction &MF = DAG.getMachineFunction();
3556 bool isPPC64 = PPCSubTarget.isPPC64();
3557 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3558 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3560 // Get current frame pointer save index. The users of this index will be
3561 // primarily DYNALLOC instructions.
3562 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3563 int RASI = FI->getReturnAddrSaveIndex();
3565 // If the frame pointer save index hasn't been defined yet.
3567 // Find out what the fix offset of the frame pointer save area.
3568 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3569 // Allocate the frame index for frame pointer save area.
3570 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3572 FI->setReturnAddrSaveIndex(RASI);
3574 return DAG.getFrameIndex(RASI, PtrVT);
3578 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3579 MachineFunction &MF = DAG.getMachineFunction();
3580 bool isPPC64 = PPCSubTarget.isPPC64();
3581 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3582 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3584 // Get current frame pointer save index. The users of this index will be
3585 // primarily DYNALLOC instructions.
3586 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3587 int FPSI = FI->getFramePointerSaveIndex();
3589 // If the frame pointer save index hasn't been defined yet.
3591 // Find out what the fix offset of the frame pointer save area.
3592 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3595 // Allocate the frame index for frame pointer save area.
3596 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3598 FI->setFramePointerSaveIndex(FPSI);
3600 return DAG.getFrameIndex(FPSI, PtrVT);
3603 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3605 const PPCSubtarget &Subtarget) const {
3607 SDValue Chain = Op.getOperand(0);
3608 SDValue Size = Op.getOperand(1);
3609 DebugLoc dl = Op.getDebugLoc();
3611 // Get the corect type for pointers.
3612 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3614 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3615 DAG.getConstant(0, PtrVT), Size);
3616 // Construct a node for the frame pointer save index.
3617 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3618 // Build a DYNALLOC node.
3619 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3620 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3621 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3624 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3626 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3627 // Not FP? Not a fsel.
3628 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3629 !Op.getOperand(2).getValueType().isFloatingPoint())
3632 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3634 // Cannot handle SETEQ/SETNE.
3635 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3637 EVT ResVT = Op.getValueType();
3638 EVT CmpVT = Op.getOperand(0).getValueType();
3639 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3640 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3641 DebugLoc dl = Op.getDebugLoc();
3643 // If the RHS of the comparison is a 0.0, we don't need to do the
3644 // subtraction at all.
3645 if (isFloatingPointZero(RHS))
3647 default: break; // SETUO etc aren't handled by fsel.
3650 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3653 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3654 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3655 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3658 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3661 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3662 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3663 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3664 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3669 default: break; // SETUO etc aren't handled by fsel.
3672 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3673 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3674 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3675 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3678 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3679 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3680 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3681 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3684 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3685 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3686 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3687 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3690 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3691 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3692 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3693 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3698 // FIXME: Split this code up when LegalizeDAGTypes lands.
3699 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3700 DebugLoc dl) const {
3701 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3702 SDValue Src = Op.getOperand(0);
3703 if (Src.getValueType() == MVT::f32)
3704 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3707 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3708 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3710 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3715 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3719 // Convert the FP value to an int value through memory.
3720 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3722 // Emit a store to the stack slot.
3723 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3724 MachinePointerInfo(), false, false, 0);
3726 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3728 if (Op.getValueType() == MVT::i32)
3729 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3730 DAG.getConstant(4, FIPtr.getValueType()));
3731 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3732 false, false, false, 0);
3735 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3736 SelectionDAG &DAG) const {
3737 DebugLoc dl = Op.getDebugLoc();
3738 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3739 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3742 if (Op.getOperand(0).getValueType() == MVT::i64) {
3743 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3744 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3745 if (Op.getValueType() == MVT::f32)
3746 FP = DAG.getNode(ISD::FP_ROUND, dl,
3747 MVT::f32, FP, DAG.getIntPtrConstant(0));
3751 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3752 "Unhandled SINT_TO_FP type in custom expander!");
3753 // Since we only generate this in 64-bit mode, we can take advantage of
3754 // 64-bit registers. In particular, sign extend the input value into the
3755 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3756 // then lfd it and fcfid it.
3757 MachineFunction &MF = DAG.getMachineFunction();
3758 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3759 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3760 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3761 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3763 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3766 // STD the extended value into the stack slot.
3767 MachineMemOperand *MMO =
3768 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3769 MachineMemOperand::MOStore, 8, 8);
3770 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3772 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3773 Ops, 4, MVT::i64, MMO);
3774 // Load the value as a double.
3775 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3776 false, false, false, 0);
3778 // FCFID it and return it.
3779 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3780 if (Op.getValueType() == MVT::f32)
3781 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3785 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3786 SelectionDAG &DAG) const {
3787 DebugLoc dl = Op.getDebugLoc();
3789 The rounding mode is in bits 30:31 of FPSR, and has the following
3796 FLT_ROUNDS, on the other hand, expects the following:
3803 To perform the conversion, we do:
3804 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3807 MachineFunction &MF = DAG.getMachineFunction();
3808 EVT VT = Op.getValueType();
3809 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3810 std::vector<EVT> NodeTys;
3811 SDValue MFFSreg, InFlag;
3813 // Save FP Control Word to register
3814 NodeTys.push_back(MVT::f64); // return register
3815 NodeTys.push_back(MVT::Glue); // unused in this context
3816 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3818 // Save FP register to stack slot
3819 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3820 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3821 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3822 StackSlot, MachinePointerInfo(), false, false,0);
3824 // Load FP Control Word from low 32 bits of stack slot.
3825 SDValue Four = DAG.getConstant(4, PtrVT);
3826 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3827 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3828 false, false, false, 0);
3830 // Transform as necessary
3832 DAG.getNode(ISD::AND, dl, MVT::i32,
3833 CWD, DAG.getConstant(3, MVT::i32));
3835 DAG.getNode(ISD::SRL, dl, MVT::i32,
3836 DAG.getNode(ISD::AND, dl, MVT::i32,
3837 DAG.getNode(ISD::XOR, dl, MVT::i32,
3838 CWD, DAG.getConstant(3, MVT::i32)),
3839 DAG.getConstant(3, MVT::i32)),
3840 DAG.getConstant(1, MVT::i32));
3843 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3845 return DAG.getNode((VT.getSizeInBits() < 16 ?
3846 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3849 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3850 EVT VT = Op.getValueType();
3851 unsigned BitWidth = VT.getSizeInBits();
3852 DebugLoc dl = Op.getDebugLoc();
3853 assert(Op.getNumOperands() == 3 &&
3854 VT == Op.getOperand(1).getValueType() &&
3857 // Expand into a bunch of logical ops. Note that these ops
3858 // depend on the PPC behavior for oversized shift amounts.
3859 SDValue Lo = Op.getOperand(0);
3860 SDValue Hi = Op.getOperand(1);
3861 SDValue Amt = Op.getOperand(2);
3862 EVT AmtVT = Amt.getValueType();
3864 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3865 DAG.getConstant(BitWidth, AmtVT), Amt);
3866 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3867 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3868 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3869 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3870 DAG.getConstant(-BitWidth, AmtVT));
3871 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3872 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3873 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3874 SDValue OutOps[] = { OutLo, OutHi };
3875 return DAG.getMergeValues(OutOps, 2, dl);
3878 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3879 EVT VT = Op.getValueType();
3880 DebugLoc dl = Op.getDebugLoc();
3881 unsigned BitWidth = VT.getSizeInBits();
3882 assert(Op.getNumOperands() == 3 &&
3883 VT == Op.getOperand(1).getValueType() &&
3886 // Expand into a bunch of logical ops. Note that these ops
3887 // depend on the PPC behavior for oversized shift amounts.
3888 SDValue Lo = Op.getOperand(0);
3889 SDValue Hi = Op.getOperand(1);
3890 SDValue Amt = Op.getOperand(2);
3891 EVT AmtVT = Amt.getValueType();
3893 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3894 DAG.getConstant(BitWidth, AmtVT), Amt);
3895 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3896 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3897 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3898 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3899 DAG.getConstant(-BitWidth, AmtVT));
3900 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3901 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3902 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3903 SDValue OutOps[] = { OutLo, OutHi };
3904 return DAG.getMergeValues(OutOps, 2, dl);
3907 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3908 DebugLoc dl = Op.getDebugLoc();
3909 EVT VT = Op.getValueType();
3910 unsigned BitWidth = VT.getSizeInBits();
3911 assert(Op.getNumOperands() == 3 &&
3912 VT == Op.getOperand(1).getValueType() &&
3915 // Expand into a bunch of logical ops, followed by a select_cc.
3916 SDValue Lo = Op.getOperand(0);
3917 SDValue Hi = Op.getOperand(1);
3918 SDValue Amt = Op.getOperand(2);
3919 EVT AmtVT = Amt.getValueType();
3921 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3922 DAG.getConstant(BitWidth, AmtVT), Amt);
3923 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3924 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3925 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3926 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3927 DAG.getConstant(-BitWidth, AmtVT));
3928 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3929 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3930 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3931 Tmp4, Tmp6, ISD::SETLE);
3932 SDValue OutOps[] = { OutLo, OutHi };
3933 return DAG.getMergeValues(OutOps, 2, dl);
3936 //===----------------------------------------------------------------------===//
3937 // Vector related lowering.
3940 /// BuildSplatI - Build a canonical splati of Val with an element size of
3941 /// SplatSize. Cast the result to VT.
3942 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3943 SelectionDAG &DAG, DebugLoc dl) {
3944 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3946 static const EVT VTys[] = { // canonical VT to use for each size.
3947 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3950 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3952 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3956 EVT CanonicalVT = VTys[SplatSize-1];
3958 // Build a canonical splat for this value.
3959 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3960 SmallVector<SDValue, 8> Ops;
3961 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3962 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3963 &Ops[0], Ops.size());
3964 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3967 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3968 /// specified intrinsic ID.
3969 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3970 SelectionDAG &DAG, DebugLoc dl,
3971 EVT DestVT = MVT::Other) {
3972 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3973 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3974 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3977 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3978 /// specified intrinsic ID.
3979 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3980 SDValue Op2, SelectionDAG &DAG,
3981 DebugLoc dl, EVT DestVT = MVT::Other) {
3982 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3983 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3984 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3988 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3989 /// amount. The result has the specified value type.
3990 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3991 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3992 // Force LHS/RHS to be the right type.
3993 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3994 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3997 for (unsigned i = 0; i != 16; ++i)
3999 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4000 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4003 // If this is a case we can't handle, return null and let the default
4004 // expansion code take care of it. If we CAN select this case, and if it
4005 // selects to a single instruction, return Op. Otherwise, if we can codegen
4006 // this case more efficiently than a constant pool load, lower it to the
4007 // sequence of ops that should be used.
4008 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4009 SelectionDAG &DAG) const {
4010 DebugLoc dl = Op.getDebugLoc();
4011 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4012 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4014 // Check if this is a splat of a constant value.
4015 APInt APSplatBits, APSplatUndef;
4016 unsigned SplatBitSize;
4018 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4019 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4022 unsigned SplatBits = APSplatBits.getZExtValue();
4023 unsigned SplatUndef = APSplatUndef.getZExtValue();
4024 unsigned SplatSize = SplatBitSize / 8;
4026 // First, handle single instruction cases.
4029 if (SplatBits == 0) {
4030 // Canonicalize all zero vectors to be v4i32.
4031 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4032 SDValue Z = DAG.getConstant(0, MVT::i32);
4033 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4034 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4039 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4040 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4042 if (SextVal >= -16 && SextVal <= 15)
4043 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4046 // Two instruction sequences.
4048 // If this value is in the range [-32,30] and is even, use:
4049 // tmp = VSPLTI[bhw], result = add tmp, tmp
4050 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4051 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4052 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4053 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4056 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4057 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4059 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4060 // Make -1 and vspltisw -1:
4061 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4063 // Make the VSLW intrinsic, computing 0x8000_0000.
4064 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4067 // xor by OnesV to invert it.
4068 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4069 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4072 // Check to see if this is a wide variety of vsplti*, binop self cases.
4073 static const signed char SplatCsts[] = {
4074 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4075 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4078 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4079 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4080 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4081 int i = SplatCsts[idx];
4083 // Figure out what shift amount will be used by altivec if shifted by i in
4085 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4087 // vsplti + shl self.
4088 if (SextVal == (i << (int)TypeShiftAmt)) {
4089 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4090 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4091 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4092 Intrinsic::ppc_altivec_vslw
4094 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4095 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4098 // vsplti + srl self.
4099 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4100 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4101 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4102 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4103 Intrinsic::ppc_altivec_vsrw
4105 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4106 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4109 // vsplti + sra self.
4110 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4111 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4112 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4113 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4114 Intrinsic::ppc_altivec_vsraw
4116 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4117 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4120 // vsplti + rol self.
4121 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4122 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4123 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4124 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4125 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4126 Intrinsic::ppc_altivec_vrlw
4128 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4129 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4132 // t = vsplti c, result = vsldoi t, t, 1
4133 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4134 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4135 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4137 // t = vsplti c, result = vsldoi t, t, 2
4138 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4139 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4140 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4142 // t = vsplti c, result = vsldoi t, t, 3
4143 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4144 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4145 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4149 // Three instruction sequences.
4151 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4152 if (SextVal >= 0 && SextVal <= 31) {
4153 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4154 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4155 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4156 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4158 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4159 if (SextVal >= -31 && SextVal <= 0) {
4160 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4161 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4162 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4163 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4169 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4170 /// the specified operations to build the shuffle.
4171 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4172 SDValue RHS, SelectionDAG &DAG,
4174 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4175 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4176 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4179 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4191 if (OpNum == OP_COPY) {
4192 if (LHSID == (1*9+2)*9+3) return LHS;
4193 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4197 SDValue OpLHS, OpRHS;
4198 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4199 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4203 default: llvm_unreachable("Unknown i32 permute!");
4205 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4206 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4207 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4208 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4211 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4212 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4213 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4214 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4217 for (unsigned i = 0; i != 16; ++i)
4218 ShufIdxs[i] = (i&3)+0;
4221 for (unsigned i = 0; i != 16; ++i)
4222 ShufIdxs[i] = (i&3)+4;
4225 for (unsigned i = 0; i != 16; ++i)
4226 ShufIdxs[i] = (i&3)+8;
4229 for (unsigned i = 0; i != 16; ++i)
4230 ShufIdxs[i] = (i&3)+12;
4233 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4235 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4237 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4239 EVT VT = OpLHS.getValueType();
4240 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4241 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4242 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4243 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4246 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4247 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4248 /// return the code it can be lowered into. Worst case, it can always be
4249 /// lowered into a vperm.
4250 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4251 SelectionDAG &DAG) const {
4252 DebugLoc dl = Op.getDebugLoc();
4253 SDValue V1 = Op.getOperand(0);
4254 SDValue V2 = Op.getOperand(1);
4255 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4256 EVT VT = Op.getValueType();
4258 // Cases that are handled by instructions that take permute immediates
4259 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4260 // selected by the instruction selector.
4261 if (V2.getOpcode() == ISD::UNDEF) {
4262 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4263 PPC::isSplatShuffleMask(SVOp, 2) ||
4264 PPC::isSplatShuffleMask(SVOp, 4) ||
4265 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4266 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4267 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4268 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4269 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4270 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4271 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4272 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4273 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4278 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4279 // and produce a fixed permutation. If any of these match, do not lower to
4281 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4282 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4283 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4284 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4285 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4286 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4287 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4288 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4289 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4292 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4293 // perfect shuffle table to emit an optimal matching sequence.
4294 ArrayRef<int> PermMask = SVOp->getMask();
4296 unsigned PFIndexes[4];
4297 bool isFourElementShuffle = true;
4298 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4299 unsigned EltNo = 8; // Start out undef.
4300 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4301 if (PermMask[i*4+j] < 0)
4302 continue; // Undef, ignore it.
4304 unsigned ByteSource = PermMask[i*4+j];
4305 if ((ByteSource & 3) != j) {
4306 isFourElementShuffle = false;
4311 EltNo = ByteSource/4;
4312 } else if (EltNo != ByteSource/4) {
4313 isFourElementShuffle = false;
4317 PFIndexes[i] = EltNo;
4320 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4321 // perfect shuffle vector to determine if it is cost effective to do this as
4322 // discrete instructions, or whether we should use a vperm.
4323 if (isFourElementShuffle) {
4324 // Compute the index in the perfect shuffle table.
4325 unsigned PFTableIndex =
4326 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4328 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4329 unsigned Cost = (PFEntry >> 30);
4331 // Determining when to avoid vperm is tricky. Many things affect the cost
4332 // of vperm, particularly how many times the perm mask needs to be computed.
4333 // For example, if the perm mask can be hoisted out of a loop or is already
4334 // used (perhaps because there are multiple permutes with the same shuffle
4335 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4336 // the loop requires an extra register.
4338 // As a compromise, we only emit discrete instructions if the shuffle can be
4339 // generated in 3 or fewer operations. When we have loop information
4340 // available, if this block is within a loop, we should avoid using vperm
4341 // for 3-operation perms and use a constant pool load instead.
4343 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4346 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4347 // vector that will get spilled to the constant pool.
4348 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4350 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4351 // that it is in input element units, not in bytes. Convert now.
4352 EVT EltVT = V1.getValueType().getVectorElementType();
4353 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4355 SmallVector<SDValue, 16> ResultMask;
4356 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4357 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4359 for (unsigned j = 0; j != BytesPerElement; ++j)
4360 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4364 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4365 &ResultMask[0], ResultMask.size());
4366 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4369 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4370 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4371 /// information about the intrinsic.
4372 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4374 unsigned IntrinsicID =
4375 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4378 switch (IntrinsicID) {
4379 default: return false;
4380 // Comparison predicates.
4381 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4382 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4383 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4384 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4385 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4386 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4387 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4388 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4389 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4390 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4391 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4392 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4393 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4395 // Normal Comparisons.
4396 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4397 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4398 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4399 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4400 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4401 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4402 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4403 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4404 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4405 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4406 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4407 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4408 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4413 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4414 /// lower, do it, otherwise return null.
4415 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4416 SelectionDAG &DAG) const {
4417 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4418 // opcode number of the comparison.
4419 DebugLoc dl = Op.getDebugLoc();
4422 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4423 return SDValue(); // Don't custom lower most intrinsics.
4425 // If this is a non-dot comparison, make the VCMP node and we are done.
4427 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4428 Op.getOperand(1), Op.getOperand(2),
4429 DAG.getConstant(CompareOpc, MVT::i32));
4430 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4433 // Create the PPCISD altivec 'dot' comparison node.
4435 Op.getOperand(2), // LHS
4436 Op.getOperand(3), // RHS
4437 DAG.getConstant(CompareOpc, MVT::i32)
4439 std::vector<EVT> VTs;
4440 VTs.push_back(Op.getOperand(2).getValueType());
4441 VTs.push_back(MVT::Glue);
4442 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4444 // Now that we have the comparison, emit a copy from the CR to a GPR.
4445 // This is flagged to the above dot comparison.
4446 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4447 DAG.getRegister(PPC::CR6, MVT::i32),
4448 CompNode.getValue(1));
4450 // Unpack the result based on how the target uses it.
4451 unsigned BitNo; // Bit # of CR6.
4452 bool InvertBit; // Invert result?
4453 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4454 default: // Can't happen, don't crash on invalid number though.
4455 case 0: // Return the value of the EQ bit of CR6.
4456 BitNo = 0; InvertBit = false;
4458 case 1: // Return the inverted value of the EQ bit of CR6.
4459 BitNo = 0; InvertBit = true;
4461 case 2: // Return the value of the LT bit of CR6.
4462 BitNo = 2; InvertBit = false;
4464 case 3: // Return the inverted value of the LT bit of CR6.
4465 BitNo = 2; InvertBit = true;
4469 // Shift the bit into the low position.
4470 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4471 DAG.getConstant(8-(3-BitNo), MVT::i32));
4473 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4474 DAG.getConstant(1, MVT::i32));
4476 // If we are supposed to, toggle the bit.
4478 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4479 DAG.getConstant(1, MVT::i32));
4483 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4484 SelectionDAG &DAG) const {
4485 DebugLoc dl = Op.getDebugLoc();
4486 // Create a stack slot that is 16-byte aligned.
4487 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4488 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4489 EVT PtrVT = getPointerTy();
4490 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4492 // Store the input value into Value#0 of the stack slot.
4493 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4494 Op.getOperand(0), FIdx, MachinePointerInfo(),
4497 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4498 false, false, false, 0);
4501 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4502 DebugLoc dl = Op.getDebugLoc();
4503 if (Op.getValueType() == MVT::v4i32) {
4504 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4506 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4507 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4509 SDValue RHSSwap = // = vrlw RHS, 16
4510 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4512 // Shrinkify inputs to v8i16.
4513 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4514 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4515 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4517 // Low parts multiplied together, generating 32-bit results (we ignore the
4519 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4520 LHS, RHS, DAG, dl, MVT::v4i32);
4522 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4523 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4524 // Shift the high parts up 16 bits.
4525 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4527 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4528 } else if (Op.getValueType() == MVT::v8i16) {
4529 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4531 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4533 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4534 LHS, RHS, Zero, DAG, dl);
4535 } else if (Op.getValueType() == MVT::v16i8) {
4536 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4538 // Multiply the even 8-bit parts, producing 16-bit sums.
4539 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4540 LHS, RHS, DAG, dl, MVT::v8i16);
4541 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4543 // Multiply the odd 8-bit parts, producing 16-bit sums.
4544 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4545 LHS, RHS, DAG, dl, MVT::v8i16);
4546 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4548 // Merge the results together.
4550 for (unsigned i = 0; i != 8; ++i) {
4552 Ops[i*2+1] = 2*i+1+16;
4554 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4556 llvm_unreachable("Unknown mul to lower!");
4560 /// LowerOperation - Provide custom lowering hooks for some operations.
4562 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4563 switch (Op.getOpcode()) {
4564 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4565 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4566 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4567 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4568 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4569 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4570 case ISD::SETCC: return LowerSETCC(Op, DAG);
4571 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4572 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4574 return LowerVASTART(Op, DAG, PPCSubTarget);
4577 return LowerVAARG(Op, DAG, PPCSubTarget);
4579 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4580 case ISD::DYNAMIC_STACKALLOC:
4581 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4583 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4584 case ISD::FP_TO_UINT:
4585 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4587 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4588 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4590 // Lower 64-bit shifts.
4591 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4592 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4593 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4595 // Vector-related lowering.
4596 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4597 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4598 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4599 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4600 case ISD::MUL: return LowerMUL(Op, DAG);
4602 // Frame & Return address.
4603 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4604 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4608 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4609 SmallVectorImpl<SDValue>&Results,
4610 SelectionDAG &DAG) const {
4611 const TargetMachine &TM = getTargetMachine();
4612 DebugLoc dl = N->getDebugLoc();
4613 switch (N->getOpcode()) {
4615 llvm_unreachable("Do not know how to custom type legalize this operation!");
4617 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4618 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4621 EVT VT = N->getValueType(0);
4623 if (VT == MVT::i64) {
4624 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4626 Results.push_back(NewNode);
4627 Results.push_back(NewNode.getValue(1));
4631 case ISD::FP_ROUND_INREG: {
4632 assert(N->getValueType(0) == MVT::ppcf128);
4633 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4634 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4635 MVT::f64, N->getOperand(0),
4636 DAG.getIntPtrConstant(0));
4637 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4638 MVT::f64, N->getOperand(0),
4639 DAG.getIntPtrConstant(1));
4641 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4642 // of the long double, and puts FPSCR back the way it was. We do not
4643 // actually model FPSCR.
4644 std::vector<EVT> NodeTys;
4645 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4647 NodeTys.push_back(MVT::f64); // Return register
4648 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4649 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4650 MFFSreg = Result.getValue(0);
4651 InFlag = Result.getValue(1);
4654 NodeTys.push_back(MVT::Glue); // Returns a flag
4655 Ops[0] = DAG.getConstant(31, MVT::i32);
4657 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4658 InFlag = Result.getValue(0);
4661 NodeTys.push_back(MVT::Glue); // Returns a flag
4662 Ops[0] = DAG.getConstant(30, MVT::i32);
4664 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4665 InFlag = Result.getValue(0);
4668 NodeTys.push_back(MVT::f64); // result of add
4669 NodeTys.push_back(MVT::Glue); // Returns a flag
4673 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4674 FPreg = Result.getValue(0);
4675 InFlag = Result.getValue(1);
4678 NodeTys.push_back(MVT::f64);
4679 Ops[0] = DAG.getConstant(1, MVT::i32);
4683 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4684 FPreg = Result.getValue(0);
4686 // We know the low half is about to be thrown away, so just use something
4688 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4692 case ISD::FP_TO_SINT:
4693 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4699 //===----------------------------------------------------------------------===//
4700 // Other Lowering Code
4701 //===----------------------------------------------------------------------===//
4704 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4705 bool is64bit, unsigned BinOpcode) const {
4706 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4707 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4709 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4710 MachineFunction *F = BB->getParent();
4711 MachineFunction::iterator It = BB;
4714 unsigned dest = MI->getOperand(0).getReg();
4715 unsigned ptrA = MI->getOperand(1).getReg();
4716 unsigned ptrB = MI->getOperand(2).getReg();
4717 unsigned incr = MI->getOperand(3).getReg();
4718 DebugLoc dl = MI->getDebugLoc();
4720 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4721 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4722 F->insert(It, loopMBB);
4723 F->insert(It, exitMBB);
4724 exitMBB->splice(exitMBB->begin(), BB,
4725 llvm::next(MachineBasicBlock::iterator(MI)),
4727 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4729 MachineRegisterInfo &RegInfo = F->getRegInfo();
4730 unsigned TmpReg = (!BinOpcode) ? incr :
4731 RegInfo.createVirtualRegister(
4732 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4733 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4737 // fallthrough --> loopMBB
4738 BB->addSuccessor(loopMBB);
4741 // l[wd]arx dest, ptr
4742 // add r0, dest, incr
4743 // st[wd]cx. r0, ptr
4745 // fallthrough --> exitMBB
4747 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4748 .addReg(ptrA).addReg(ptrB);
4750 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4751 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4752 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4753 BuildMI(BB, dl, TII->get(PPC::BCC))
4754 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4755 BB->addSuccessor(loopMBB);
4756 BB->addSuccessor(exitMBB);
4765 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4766 MachineBasicBlock *BB,
4767 bool is8bit, // operation
4768 unsigned BinOpcode) const {
4769 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4770 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4771 // In 64 bit mode we have to use 64 bits for addresses, even though the
4772 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4773 // registers without caring whether they're 32 or 64, but here we're
4774 // doing actual arithmetic on the addresses.
4775 bool is64bit = PPCSubTarget.isPPC64();
4776 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4778 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4779 MachineFunction *F = BB->getParent();
4780 MachineFunction::iterator It = BB;
4783 unsigned dest = MI->getOperand(0).getReg();
4784 unsigned ptrA = MI->getOperand(1).getReg();
4785 unsigned ptrB = MI->getOperand(2).getReg();
4786 unsigned incr = MI->getOperand(3).getReg();
4787 DebugLoc dl = MI->getDebugLoc();
4789 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4790 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4791 F->insert(It, loopMBB);
4792 F->insert(It, exitMBB);
4793 exitMBB->splice(exitMBB->begin(), BB,
4794 llvm::next(MachineBasicBlock::iterator(MI)),
4796 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4798 MachineRegisterInfo &RegInfo = F->getRegInfo();
4799 const TargetRegisterClass *RC =
4800 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4801 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4802 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4803 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4804 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4805 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4806 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4807 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4808 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4809 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4810 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4811 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4812 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4814 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4818 // fallthrough --> loopMBB
4819 BB->addSuccessor(loopMBB);
4821 // The 4-byte load must be aligned, while a char or short may be
4822 // anywhere in the word. Hence all this nasty bookkeeping code.
4823 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4824 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4825 // xori shift, shift1, 24 [16]
4826 // rlwinm ptr, ptr1, 0, 0, 29
4827 // slw incr2, incr, shift
4828 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4829 // slw mask, mask2, shift
4831 // lwarx tmpDest, ptr
4832 // add tmp, tmpDest, incr2
4833 // andc tmp2, tmpDest, mask
4834 // and tmp3, tmp, mask
4835 // or tmp4, tmp3, tmp2
4838 // fallthrough --> exitMBB
4839 // srw dest, tmpDest, shift
4840 if (ptrA != ZeroReg) {
4841 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4842 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4843 .addReg(ptrA).addReg(ptrB);
4847 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4848 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4849 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4850 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4852 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4853 .addReg(Ptr1Reg).addImm(0).addImm(61);
4855 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4856 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4857 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4858 .addReg(incr).addReg(ShiftReg);
4860 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4862 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4863 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4865 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4866 .addReg(Mask2Reg).addReg(ShiftReg);
4869 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4870 .addReg(ZeroReg).addReg(PtrReg);
4872 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4873 .addReg(Incr2Reg).addReg(TmpDestReg);
4874 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4875 .addReg(TmpDestReg).addReg(MaskReg);
4876 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4877 .addReg(TmpReg).addReg(MaskReg);
4878 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4879 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4880 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4881 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4882 BuildMI(BB, dl, TII->get(PPC::BCC))
4883 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4884 BB->addSuccessor(loopMBB);
4885 BB->addSuccessor(exitMBB);
4890 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4896 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4897 MachineBasicBlock *BB) const {
4898 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4900 // To "insert" these instructions we actually have to insert their
4901 // control-flow patterns.
4902 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4903 MachineFunction::iterator It = BB;
4906 MachineFunction *F = BB->getParent();
4908 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4909 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4910 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4911 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4912 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4914 // The incoming instruction knows the destination vreg to set, the
4915 // condition code register to branch on, the true/false values to
4916 // select between, and a branch opcode to use.
4921 // cmpTY ccX, r1, r2
4923 // fallthrough --> copy0MBB
4924 MachineBasicBlock *thisMBB = BB;
4925 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4926 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4927 unsigned SelectPred = MI->getOperand(4).getImm();
4928 DebugLoc dl = MI->getDebugLoc();
4929 F->insert(It, copy0MBB);
4930 F->insert(It, sinkMBB);
4932 // Transfer the remainder of BB and its successor edges to sinkMBB.
4933 sinkMBB->splice(sinkMBB->begin(), BB,
4934 llvm::next(MachineBasicBlock::iterator(MI)),
4936 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4938 // Next, add the true and fallthrough blocks as its successors.
4939 BB->addSuccessor(copy0MBB);
4940 BB->addSuccessor(sinkMBB);
4942 BuildMI(BB, dl, TII->get(PPC::BCC))
4943 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4946 // %FalseValue = ...
4947 // # fallthrough to sinkMBB
4950 // Update machine-CFG edges
4951 BB->addSuccessor(sinkMBB);
4954 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4957 BuildMI(*BB, BB->begin(), dl,
4958 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4959 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4960 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4962 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4963 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4964 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4965 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4966 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4967 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4968 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4969 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4971 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4972 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4973 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4974 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4975 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4976 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4977 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4978 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4980 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4981 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4982 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4983 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4984 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4985 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4986 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4987 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4989 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4990 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4991 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4992 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4993 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4994 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4995 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4996 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4998 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4999 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5000 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5001 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5002 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5003 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5004 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5005 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5007 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5008 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5009 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5010 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5011 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5012 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5013 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5014 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5016 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5017 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5018 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5019 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5020 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5021 BB = EmitAtomicBinary(MI, BB, false, 0);
5022 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5023 BB = EmitAtomicBinary(MI, BB, true, 0);
5025 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5026 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5027 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5029 unsigned dest = MI->getOperand(0).getReg();
5030 unsigned ptrA = MI->getOperand(1).getReg();
5031 unsigned ptrB = MI->getOperand(2).getReg();
5032 unsigned oldval = MI->getOperand(3).getReg();
5033 unsigned newval = MI->getOperand(4).getReg();
5034 DebugLoc dl = MI->getDebugLoc();
5036 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5037 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5038 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5039 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5040 F->insert(It, loop1MBB);
5041 F->insert(It, loop2MBB);
5042 F->insert(It, midMBB);
5043 F->insert(It, exitMBB);
5044 exitMBB->splice(exitMBB->begin(), BB,
5045 llvm::next(MachineBasicBlock::iterator(MI)),
5047 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5051 // fallthrough --> loopMBB
5052 BB->addSuccessor(loop1MBB);
5055 // l[wd]arx dest, ptr
5056 // cmp[wd] dest, oldval
5059 // st[wd]cx. newval, ptr
5063 // st[wd]cx. dest, ptr
5066 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5067 .addReg(ptrA).addReg(ptrB);
5068 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5069 .addReg(oldval).addReg(dest);
5070 BuildMI(BB, dl, TII->get(PPC::BCC))
5071 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5072 BB->addSuccessor(loop2MBB);
5073 BB->addSuccessor(midMBB);
5076 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5077 .addReg(newval).addReg(ptrA).addReg(ptrB);
5078 BuildMI(BB, dl, TII->get(PPC::BCC))
5079 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5080 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5081 BB->addSuccessor(loop1MBB);
5082 BB->addSuccessor(exitMBB);
5085 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5086 .addReg(dest).addReg(ptrA).addReg(ptrB);
5087 BB->addSuccessor(exitMBB);
5092 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5093 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5094 // We must use 64-bit registers for addresses when targeting 64-bit,
5095 // since we're actually doing arithmetic on them. Other registers
5097 bool is64bit = PPCSubTarget.isPPC64();
5098 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5100 unsigned dest = MI->getOperand(0).getReg();
5101 unsigned ptrA = MI->getOperand(1).getReg();
5102 unsigned ptrB = MI->getOperand(2).getReg();
5103 unsigned oldval = MI->getOperand(3).getReg();
5104 unsigned newval = MI->getOperand(4).getReg();
5105 DebugLoc dl = MI->getDebugLoc();
5107 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5108 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5109 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5110 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5111 F->insert(It, loop1MBB);
5112 F->insert(It, loop2MBB);
5113 F->insert(It, midMBB);
5114 F->insert(It, exitMBB);
5115 exitMBB->splice(exitMBB->begin(), BB,
5116 llvm::next(MachineBasicBlock::iterator(MI)),
5118 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5120 MachineRegisterInfo &RegInfo = F->getRegInfo();
5121 const TargetRegisterClass *RC =
5122 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5123 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5124 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5125 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5126 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5127 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5128 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5129 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5130 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5131 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5132 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5133 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5134 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5135 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5136 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5138 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5139 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5142 // fallthrough --> loopMBB
5143 BB->addSuccessor(loop1MBB);
5145 // The 4-byte load must be aligned, while a char or short may be
5146 // anywhere in the word. Hence all this nasty bookkeeping code.
5147 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5148 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5149 // xori shift, shift1, 24 [16]
5150 // rlwinm ptr, ptr1, 0, 0, 29
5151 // slw newval2, newval, shift
5152 // slw oldval2, oldval,shift
5153 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5154 // slw mask, mask2, shift
5155 // and newval3, newval2, mask
5156 // and oldval3, oldval2, mask
5158 // lwarx tmpDest, ptr
5159 // and tmp, tmpDest, mask
5160 // cmpw tmp, oldval3
5163 // andc tmp2, tmpDest, mask
5164 // or tmp4, tmp2, newval3
5169 // stwcx. tmpDest, ptr
5171 // srw dest, tmpDest, shift
5172 if (ptrA != ZeroReg) {
5173 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5174 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5175 .addReg(ptrA).addReg(ptrB);
5179 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5180 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5181 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5182 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5184 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5185 .addReg(Ptr1Reg).addImm(0).addImm(61);
5187 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5188 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5189 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5190 .addReg(newval).addReg(ShiftReg);
5191 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5192 .addReg(oldval).addReg(ShiftReg);
5194 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5196 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5197 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5198 .addReg(Mask3Reg).addImm(65535);
5200 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5201 .addReg(Mask2Reg).addReg(ShiftReg);
5202 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5203 .addReg(NewVal2Reg).addReg(MaskReg);
5204 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5205 .addReg(OldVal2Reg).addReg(MaskReg);
5208 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5209 .addReg(ZeroReg).addReg(PtrReg);
5210 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5211 .addReg(TmpDestReg).addReg(MaskReg);
5212 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5213 .addReg(TmpReg).addReg(OldVal3Reg);
5214 BuildMI(BB, dl, TII->get(PPC::BCC))
5215 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5216 BB->addSuccessor(loop2MBB);
5217 BB->addSuccessor(midMBB);
5220 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5221 .addReg(TmpDestReg).addReg(MaskReg);
5222 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5223 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5224 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5225 .addReg(ZeroReg).addReg(PtrReg);
5226 BuildMI(BB, dl, TII->get(PPC::BCC))
5227 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5228 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5229 BB->addSuccessor(loop1MBB);
5230 BB->addSuccessor(exitMBB);
5233 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5234 .addReg(ZeroReg).addReg(PtrReg);
5235 BB->addSuccessor(exitMBB);
5240 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5243 llvm_unreachable("Unexpected instr type to insert");
5246 MI->eraseFromParent(); // The pseudo instruction is gone now.
5250 //===----------------------------------------------------------------------===//
5251 // Target Optimization Hooks
5252 //===----------------------------------------------------------------------===//
5254 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5255 DAGCombinerInfo &DCI) const {
5256 const TargetMachine &TM = getTargetMachine();
5257 SelectionDAG &DAG = DCI.DAG;
5258 DebugLoc dl = N->getDebugLoc();
5259 switch (N->getOpcode()) {
5262 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5263 if (C->isNullValue()) // 0 << V -> 0.
5264 return N->getOperand(0);
5268 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5269 if (C->isNullValue()) // 0 >>u V -> 0.
5270 return N->getOperand(0);
5274 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5275 if (C->isNullValue() || // 0 >>s V -> 0.
5276 C->isAllOnesValue()) // -1 >>s V -> -1.
5277 return N->getOperand(0);
5281 case ISD::SINT_TO_FP:
5282 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5283 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5284 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5285 // We allow the src/dst to be either f32/f64, but the intermediate
5286 // type must be i64.
5287 if (N->getOperand(0).getValueType() == MVT::i64 &&
5288 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5289 SDValue Val = N->getOperand(0).getOperand(0);
5290 if (Val.getValueType() == MVT::f32) {
5291 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5292 DCI.AddToWorklist(Val.getNode());
5295 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5296 DCI.AddToWorklist(Val.getNode());
5297 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5298 DCI.AddToWorklist(Val.getNode());
5299 if (N->getValueType(0) == MVT::f32) {
5300 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5301 DAG.getIntPtrConstant(0));
5302 DCI.AddToWorklist(Val.getNode());
5305 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5306 // If the intermediate type is i32, we can avoid the load/store here
5313 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5314 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5315 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5316 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5317 N->getOperand(1).getValueType() == MVT::i32 &&
5318 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5319 SDValue Val = N->getOperand(1).getOperand(0);
5320 if (Val.getValueType() == MVT::f32) {
5321 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5322 DCI.AddToWorklist(Val.getNode());
5324 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5325 DCI.AddToWorklist(Val.getNode());
5327 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5328 N->getOperand(2), N->getOperand(3));
5329 DCI.AddToWorklist(Val.getNode());
5333 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5334 if (cast<StoreSDNode>(N)->isUnindexed() &&
5335 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5336 N->getOperand(1).getNode()->hasOneUse() &&
5337 (N->getOperand(1).getValueType() == MVT::i32 ||
5338 N->getOperand(1).getValueType() == MVT::i16)) {
5339 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5340 // Do an any-extend to 32-bits if this is a half-word input.
5341 if (BSwapOp.getValueType() == MVT::i16)
5342 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5345 N->getOperand(0), BSwapOp, N->getOperand(2),
5346 DAG.getValueType(N->getOperand(1).getValueType())
5349 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5350 Ops, array_lengthof(Ops),
5351 cast<StoreSDNode>(N)->getMemoryVT(),
5352 cast<StoreSDNode>(N)->getMemOperand());
5356 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5357 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5358 N->getOperand(0).hasOneUse() &&
5359 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5360 SDValue Load = N->getOperand(0);
5361 LoadSDNode *LD = cast<LoadSDNode>(Load);
5362 // Create the byte-swapping load.
5364 LD->getChain(), // Chain
5365 LD->getBasePtr(), // Ptr
5366 DAG.getValueType(N->getValueType(0)) // VT
5369 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5370 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5371 LD->getMemoryVT(), LD->getMemOperand());
5373 // If this is an i16 load, insert the truncate.
5374 SDValue ResVal = BSLoad;
5375 if (N->getValueType(0) == MVT::i16)
5376 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5378 // First, combine the bswap away. This makes the value produced by the
5380 DCI.CombineTo(N, ResVal);
5382 // Next, combine the load away, we give it a bogus result value but a real
5383 // chain result. The result value is dead because the bswap is dead.
5384 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5386 // Return N so it doesn't get rechecked!
5387 return SDValue(N, 0);
5391 case PPCISD::VCMP: {
5392 // If a VCMPo node already exists with exactly the same operands as this
5393 // node, use its result instead of this node (VCMPo computes both a CR6 and
5394 // a normal output).
5396 if (!N->getOperand(0).hasOneUse() &&
5397 !N->getOperand(1).hasOneUse() &&
5398 !N->getOperand(2).hasOneUse()) {
5400 // Scan all of the users of the LHS, looking for VCMPo's that match.
5401 SDNode *VCMPoNode = 0;
5403 SDNode *LHSN = N->getOperand(0).getNode();
5404 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5406 if (UI->getOpcode() == PPCISD::VCMPo &&
5407 UI->getOperand(1) == N->getOperand(1) &&
5408 UI->getOperand(2) == N->getOperand(2) &&
5409 UI->getOperand(0) == N->getOperand(0)) {
5414 // If there is no VCMPo node, or if the flag value has a single use, don't
5416 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5419 // Look at the (necessarily single) use of the flag value. If it has a
5420 // chain, this transformation is more complex. Note that multiple things
5421 // could use the value result, which we should ignore.
5422 SDNode *FlagUser = 0;
5423 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5424 FlagUser == 0; ++UI) {
5425 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5427 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5428 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5435 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5436 // give up for right now.
5437 if (FlagUser->getOpcode() == PPCISD::MFCR)
5438 return SDValue(VCMPoNode, 0);
5443 // If this is a branch on an altivec predicate comparison, lower this so
5444 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5445 // lowering is done pre-legalize, because the legalizer lowers the predicate
5446 // compare down to code that is difficult to reassemble.
5447 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5448 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5452 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5453 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5454 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5455 assert(isDot && "Can't compare against a vector result!");
5457 // If this is a comparison against something other than 0/1, then we know
5458 // that the condition is never/always true.
5459 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5460 if (Val != 0 && Val != 1) {
5461 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5462 return N->getOperand(0);
5463 // Always !=, turn it into an unconditional branch.
5464 return DAG.getNode(ISD::BR, dl, MVT::Other,
5465 N->getOperand(0), N->getOperand(4));
5468 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5470 // Create the PPCISD altivec 'dot' comparison node.
5471 std::vector<EVT> VTs;
5473 LHS.getOperand(2), // LHS of compare
5474 LHS.getOperand(3), // RHS of compare
5475 DAG.getConstant(CompareOpc, MVT::i32)
5477 VTs.push_back(LHS.getOperand(2).getValueType());
5478 VTs.push_back(MVT::Glue);
5479 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5481 // Unpack the result based on how the target uses it.
5482 PPC::Predicate CompOpc;
5483 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5484 default: // Can't happen, don't crash on invalid number though.
5485 case 0: // Branch on the value of the EQ bit of CR6.
5486 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5488 case 1: // Branch on the inverted value of the EQ bit of CR6.
5489 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5491 case 2: // Branch on the value of the LT bit of CR6.
5492 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5494 case 3: // Branch on the inverted value of the LT bit of CR6.
5495 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5499 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5500 DAG.getConstant(CompOpc, MVT::i32),
5501 DAG.getRegister(PPC::CR6, MVT::i32),
5502 N->getOperand(4), CompNode.getValue(1));
5511 //===----------------------------------------------------------------------===//
5512 // Inline Assembly Support
5513 //===----------------------------------------------------------------------===//
5515 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5518 const SelectionDAG &DAG,
5519 unsigned Depth) const {
5520 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
5521 switch (Op.getOpcode()) {
5523 case PPCISD::LBRX: {
5524 // lhbrx is known to have the top bits cleared out.
5525 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5526 KnownZero = 0xFFFF0000;
5529 case ISD::INTRINSIC_WO_CHAIN: {
5530 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5532 case Intrinsic::ppc_altivec_vcmpbfp_p:
5533 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5534 case Intrinsic::ppc_altivec_vcmpequb_p:
5535 case Intrinsic::ppc_altivec_vcmpequh_p:
5536 case Intrinsic::ppc_altivec_vcmpequw_p:
5537 case Intrinsic::ppc_altivec_vcmpgefp_p:
5538 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5539 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5540 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5541 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5542 case Intrinsic::ppc_altivec_vcmpgtub_p:
5543 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5544 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5545 KnownZero = ~1U; // All bits but the low one are known to be zero.
5553 /// getConstraintType - Given a constraint, return the type of
5554 /// constraint it is for this target.
5555 PPCTargetLowering::ConstraintType
5556 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5557 if (Constraint.size() == 1) {
5558 switch (Constraint[0]) {
5565 return C_RegisterClass;
5568 return TargetLowering::getConstraintType(Constraint);
5571 /// Examine constraint type and operand type and determine a weight value.
5572 /// This object must already have been set up with the operand type
5573 /// and the current alternative constraint selected.
5574 TargetLowering::ConstraintWeight
5575 PPCTargetLowering::getSingleConstraintMatchWeight(
5576 AsmOperandInfo &info, const char *constraint) const {
5577 ConstraintWeight weight = CW_Invalid;
5578 Value *CallOperandVal = info.CallOperandVal;
5579 // If we don't have a value, we can't do a match,
5580 // but allow it at the lowest weight.
5581 if (CallOperandVal == NULL)
5583 Type *type = CallOperandVal->getType();
5584 // Look at the constraint type.
5585 switch (*constraint) {
5587 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5590 if (type->isIntegerTy())
5591 weight = CW_Register;
5594 if (type->isFloatTy())
5595 weight = CW_Register;
5598 if (type->isDoubleTy())
5599 weight = CW_Register;
5602 if (type->isVectorTy())
5603 weight = CW_Register;
5606 weight = CW_Register;
5612 std::pair<unsigned, const TargetRegisterClass*>
5613 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5615 if (Constraint.size() == 1) {
5616 // GCC RS6000 Constraint Letters
5617 switch (Constraint[0]) {
5620 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5621 return std::make_pair(0U, &PPC::G8RCRegClass);
5622 return std::make_pair(0U, &PPC::GPRCRegClass);
5625 return std::make_pair(0U, &PPC::F4RCRegClass);
5627 return std::make_pair(0U, &PPC::F8RCRegClass);
5630 return std::make_pair(0U, &PPC::VRRCRegClass);
5632 return std::make_pair(0U, &PPC::CRRCRegClass);
5636 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5640 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5641 /// vector. If it is invalid, don't add anything to Ops.
5642 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5643 std::string &Constraint,
5644 std::vector<SDValue>&Ops,
5645 SelectionDAG &DAG) const {
5646 SDValue Result(0,0);
5648 // Only support length 1 constraints.
5649 if (Constraint.length() > 1) return;
5651 char Letter = Constraint[0];
5662 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5663 if (!CST) return; // Must be an immediate to match.
5664 unsigned Value = CST->getZExtValue();
5666 default: llvm_unreachable("Unknown constraint letter!");
5667 case 'I': // "I" is a signed 16-bit constant.
5668 if ((short)Value == (int)Value)
5669 Result = DAG.getTargetConstant(Value, Op.getValueType());
5671 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5672 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5673 if ((short)Value == 0)
5674 Result = DAG.getTargetConstant(Value, Op.getValueType());
5676 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5677 if ((Value >> 16) == 0)
5678 Result = DAG.getTargetConstant(Value, Op.getValueType());
5680 case 'M': // "M" is a constant that is greater than 31.
5682 Result = DAG.getTargetConstant(Value, Op.getValueType());
5684 case 'N': // "N" is a positive constant that is an exact power of two.
5685 if ((int)Value > 0 && isPowerOf2_32(Value))
5686 Result = DAG.getTargetConstant(Value, Op.getValueType());
5688 case 'O': // "O" is the constant zero.
5690 Result = DAG.getTargetConstant(Value, Op.getValueType());
5692 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5693 if ((short)-Value == (int)-Value)
5694 Result = DAG.getTargetConstant(Value, Op.getValueType());
5701 if (Result.getNode()) {
5702 Ops.push_back(Result);
5706 // Handle standard constraint letters.
5707 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5710 // isLegalAddressingMode - Return true if the addressing mode represented
5711 // by AM is legal for this target, for a load/store of the specified type.
5712 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5714 // FIXME: PPC does not allow r+i addressing modes for vectors!
5716 // PPC allows a sign-extended 16-bit immediate field.
5717 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5720 // No global is ever allowed as a base.
5724 // PPC only support r+r,
5726 case 0: // "r+i" or just "i", depending on HasBaseReg.
5729 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5731 // Otherwise we have r+r or r+i.
5734 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5736 // Allow 2*r as r+r.
5739 // No other scales are supported.
5746 /// isLegalAddressImmediate - Return true if the integer value can be used
5747 /// as the offset of the target addressing mode for load / store of the
5749 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5750 // PPC allows a sign-extended 16-bit immediate field.
5751 return (V > -(1 << 16) && V < (1 << 16)-1);
5754 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
5758 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5759 SelectionDAG &DAG) const {
5760 MachineFunction &MF = DAG.getMachineFunction();
5761 MachineFrameInfo *MFI = MF.getFrameInfo();
5762 MFI->setReturnAddressIsTaken(true);
5764 DebugLoc dl = Op.getDebugLoc();
5765 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5767 // Make sure the function does not optimize away the store of the RA to
5769 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5770 FuncInfo->setLRStoreRequired();
5771 bool isPPC64 = PPCSubTarget.isPPC64();
5772 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5775 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5778 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5779 isPPC64? MVT::i64 : MVT::i32);
5780 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5781 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5783 MachinePointerInfo(), false, false, false, 0);
5786 // Just load the return address off the stack.
5787 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5788 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5789 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5792 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5793 SelectionDAG &DAG) const {
5794 DebugLoc dl = Op.getDebugLoc();
5795 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5797 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5798 bool isPPC64 = PtrVT == MVT::i64;
5800 MachineFunction &MF = DAG.getMachineFunction();
5801 MachineFrameInfo *MFI = MF.getFrameInfo();
5802 MFI->setFrameAddressIsTaken(true);
5803 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
5804 MFI->hasVarSizedObjects()) &&
5805 MFI->getStackSize() &&
5806 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5807 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5808 (is31 ? PPC::R31 : PPC::R1);
5809 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5812 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5813 FrameAddr, MachinePointerInfo(), false, false,
5819 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5820 // The PowerPC target isn't yet aware of offsets.
5824 /// getOptimalMemOpType - Returns the target specific optimal type for load
5825 /// and store operations as a result of memset, memcpy, and memmove
5826 /// lowering. If DstAlign is zero that means it's safe to destination
5827 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5828 /// means there isn't a need to check it against alignment requirement,
5829 /// probably because the source does not need to be loaded. If
5830 /// 'IsZeroVal' is true, that means it's safe to return a
5831 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5832 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5833 /// constant so it does not need to be loaded.
5834 /// It returns EVT::Other if the type should be determined using generic
5835 /// target-independent logic.
5836 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5837 unsigned DstAlign, unsigned SrcAlign,
5840 MachineFunction &MF) const {
5841 if (this->PPCSubTarget.isPPC64()) {
5848 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
5849 unsigned Directive = PPCSubTarget.getDarwinDirective();
5850 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2)
5853 return TargetLowering::getSchedulingPreference(N);