1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCMachineFunctionInfo.h"
16 #include "PPCPerfectShuffle.h"
17 #include "PPCTargetMachine.h"
18 #include "MCTargetDesc/PPCPredicates.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/VectorExtras.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
29 #include "llvm/CallingConv.h"
30 #include "llvm/Constants.h"
31 #include "llvm/Function.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/DerivedTypes.h"
41 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
42 CCValAssign::LocInfo &LocInfo,
43 ISD::ArgFlagsTy &ArgFlags,
45 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
47 CCValAssign::LocInfo &LocInfo,
48 ISD::ArgFlagsTy &ArgFlags,
50 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
52 CCValAssign::LocInfo &LocInfo,
53 ISD::ArgFlagsTy &ArgFlags,
56 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
57 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
60 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
61 if (TM.getSubtargetImpl()->isDarwin())
62 return new TargetLoweringObjectFileMachO();
64 return new TargetLoweringObjectFileELF();
67 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
68 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
72 // Use _setjmp/_longjmp instead of setjmp/longjmp.
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(true);
76 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
77 // arguments are at least 4/8 bytes aligned.
78 setMinStackArgumentAlignment(TM.getSubtarget<PPCSubtarget>().isPPC64() ? 8:4);
80 // Set up the register classes.
81 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
82 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
83 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
85 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
86 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
87 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
89 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
91 // PowerPC has pre-inc load and store's.
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
93 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
96 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
98 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
101 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
103 // This is used in the ppcf128->int sequence. Note it has different semantics
104 // from FP_ROUND: that rounds to nearest, this rounds to zero.
105 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
107 // PowerPC has no SREM/UREM instructions
108 setOperationAction(ISD::SREM, MVT::i32, Expand);
109 setOperationAction(ISD::UREM, MVT::i32, Expand);
110 setOperationAction(ISD::SREM, MVT::i64, Expand);
111 setOperationAction(ISD::UREM, MVT::i64, Expand);
113 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
114 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
115 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
116 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
117 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
118 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
119 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
120 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
121 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
123 // We don't support sin/cos/sqrt/fmod/pow
124 setOperationAction(ISD::FSIN , MVT::f64, Expand);
125 setOperationAction(ISD::FCOS , MVT::f64, Expand);
126 setOperationAction(ISD::FREM , MVT::f64, Expand);
127 setOperationAction(ISD::FPOW , MVT::f64, Expand);
128 setOperationAction(ISD::FMA , MVT::f64, Expand);
129 setOperationAction(ISD::FSIN , MVT::f32, Expand);
130 setOperationAction(ISD::FCOS , MVT::f32, Expand);
131 setOperationAction(ISD::FREM , MVT::f32, Expand);
132 setOperationAction(ISD::FPOW , MVT::f32, Expand);
133 setOperationAction(ISD::FMA , MVT::f32, Expand);
135 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
137 // If we're enabling GP optimizations, use hardware square root
138 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
139 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
140 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
143 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
144 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
146 // PowerPC does not have BSWAP, CTPOP or CTTZ
147 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
148 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
149 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
150 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
151 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
152 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
154 // PowerPC does not have ROTR
155 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
156 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
158 // PowerPC does not have Select
159 setOperationAction(ISD::SELECT, MVT::i32, Expand);
160 setOperationAction(ISD::SELECT, MVT::i64, Expand);
161 setOperationAction(ISD::SELECT, MVT::f32, Expand);
162 setOperationAction(ISD::SELECT, MVT::f64, Expand);
164 // PowerPC wants to turn select_cc of FP into fsel when possible.
165 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
166 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
168 // PowerPC wants to optimize integer setcc a bit
169 setOperationAction(ISD::SETCC, MVT::i32, Custom);
171 // PowerPC does not have BRCOND which requires SetCC
172 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
174 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
176 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
177 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
179 // PowerPC does not have [U|S]INT_TO_FP
180 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
181 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
183 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
184 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
185 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
186 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
188 // We cannot sextinreg(i1). Expand to shifts.
189 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
191 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
192 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
193 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
194 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
197 // We want to legalize GlobalAddress and ConstantPool nodes into the
198 // appropriate instructions to materialize the address.
199 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
200 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
201 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
202 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
203 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
204 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
205 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
206 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
207 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
208 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
211 setOperationAction(ISD::TRAP, MVT::Other, Legal);
213 // TRAMPOLINE is custom lowered.
214 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
216 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
217 setOperationAction(ISD::VASTART , MVT::Other, Custom);
219 // VAARG is custom lowered with the 32-bit SVR4 ABI.
220 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
221 && !TM.getSubtarget<PPCSubtarget>().isPPC64()) {
222 setOperationAction(ISD::VAARG, MVT::Other, Custom);
223 setOperationAction(ISD::VAARG, MVT::i64, Custom);
225 setOperationAction(ISD::VAARG, MVT::Other, Expand);
227 // Use the default implementation.
228 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
229 setOperationAction(ISD::VAEND , MVT::Other, Expand);
230 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
231 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
232 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
233 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
235 // We want to custom lower some of our intrinsics.
236 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
238 // Comparisons that require checking two conditions.
239 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
240 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
241 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
242 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
243 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
244 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
245 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
246 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
247 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
248 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
249 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
250 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
252 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
253 // They also have instructions for converting between i64 and fp.
254 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
255 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
256 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
257 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
258 // This is just the low 32 bits of a (signed) fp->i64 conversion.
259 // We cannot do this with Promote because i64 is not a legal type.
260 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
262 // FIXME: disable this lowered code. This generates 64-bit register values,
263 // and we don't model the fact that the top part is clobbered by calls. We
264 // need to flag these together so that the value isn't live across a call.
265 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
267 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
268 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
271 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
272 // 64-bit PowerPC implementations can support i64 types directly
273 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
274 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
275 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
276 // 64-bit PowerPC wants to expand i128 shifts itself.
277 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
278 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
279 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
281 // 32-bit PowerPC wants to expand i64 shifts itself.
282 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
283 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
284 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
287 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
288 // First set operation action for all vector types to expand. Then we
289 // will selectively turn on ones that can be effectively codegen'd.
290 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
291 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
292 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
294 // add/sub are legal for all supported vector VT's.
295 setOperationAction(ISD::ADD , VT, Legal);
296 setOperationAction(ISD::SUB , VT, Legal);
298 // We promote all shuffles to v16i8.
299 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
300 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
302 // We promote all non-typed operations to v4i32.
303 setOperationAction(ISD::AND , VT, Promote);
304 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
305 setOperationAction(ISD::OR , VT, Promote);
306 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
307 setOperationAction(ISD::XOR , VT, Promote);
308 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
309 setOperationAction(ISD::LOAD , VT, Promote);
310 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
311 setOperationAction(ISD::SELECT, VT, Promote);
312 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
313 setOperationAction(ISD::STORE, VT, Promote);
314 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
316 // No other operations are legal.
317 setOperationAction(ISD::MUL , VT, Expand);
318 setOperationAction(ISD::SDIV, VT, Expand);
319 setOperationAction(ISD::SREM, VT, Expand);
320 setOperationAction(ISD::UDIV, VT, Expand);
321 setOperationAction(ISD::UREM, VT, Expand);
322 setOperationAction(ISD::FDIV, VT, Expand);
323 setOperationAction(ISD::FNEG, VT, Expand);
324 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
325 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
326 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
327 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
328 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
329 setOperationAction(ISD::UDIVREM, VT, Expand);
330 setOperationAction(ISD::SDIVREM, VT, Expand);
331 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
332 setOperationAction(ISD::FPOW, VT, Expand);
333 setOperationAction(ISD::CTPOP, VT, Expand);
334 setOperationAction(ISD::CTLZ, VT, Expand);
335 setOperationAction(ISD::CTTZ, VT, Expand);
338 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
339 // with merges, splats, etc.
340 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
342 setOperationAction(ISD::AND , MVT::v4i32, Legal);
343 setOperationAction(ISD::OR , MVT::v4i32, Legal);
344 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
345 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
346 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
347 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
349 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
350 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
351 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
352 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
354 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
355 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
356 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
357 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
359 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
360 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
362 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
363 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
364 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
365 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
368 setBooleanContents(ZeroOrOneBooleanContent);
370 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
371 setStackPointerRegisterToSaveRestore(PPC::X1);
372 setExceptionPointerRegister(PPC::X3);
373 setExceptionSelectorRegister(PPC::X4);
375 setStackPointerRegisterToSaveRestore(PPC::R1);
376 setExceptionPointerRegister(PPC::R3);
377 setExceptionSelectorRegister(PPC::R4);
380 // We have target-specific dag combine patterns for the following nodes:
381 setTargetDAGCombine(ISD::SINT_TO_FP);
382 setTargetDAGCombine(ISD::STORE);
383 setTargetDAGCombine(ISD::BR_CC);
384 setTargetDAGCombine(ISD::BSWAP);
386 // Darwin long double math library functions have $LDBL128 appended.
387 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
388 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
389 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
390 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
391 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
392 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
393 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
394 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
395 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
396 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
397 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
400 setMinFunctionAlignment(2);
401 if (PPCSubTarget.isDarwin())
402 setPrefFunctionAlignment(4);
404 setInsertFencesForAtomic(true);
406 computeRegisterProperties();
409 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
410 /// function arguments in the caller parameter area.
411 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
412 const TargetMachine &TM = getTargetMachine();
413 // Darwin passes everything on 4 byte boundary.
414 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
420 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
423 case PPCISD::FSEL: return "PPCISD::FSEL";
424 case PPCISD::FCFID: return "PPCISD::FCFID";
425 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
426 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
427 case PPCISD::STFIWX: return "PPCISD::STFIWX";
428 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
429 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
430 case PPCISD::VPERM: return "PPCISD::VPERM";
431 case PPCISD::Hi: return "PPCISD::Hi";
432 case PPCISD::Lo: return "PPCISD::Lo";
433 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
434 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
435 case PPCISD::LOAD: return "PPCISD::LOAD";
436 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
437 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
438 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
439 case PPCISD::SRL: return "PPCISD::SRL";
440 case PPCISD::SRA: return "PPCISD::SRA";
441 case PPCISD::SHL: return "PPCISD::SHL";
442 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
443 case PPCISD::STD_32: return "PPCISD::STD_32";
444 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
445 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
446 case PPCISD::NOP: return "PPCISD::NOP";
447 case PPCISD::MTCTR: return "PPCISD::MTCTR";
448 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
449 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
450 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
451 case PPCISD::MFCR: return "PPCISD::MFCR";
452 case PPCISD::VCMP: return "PPCISD::VCMP";
453 case PPCISD::VCMPo: return "PPCISD::VCMPo";
454 case PPCISD::LBRX: return "PPCISD::LBRX";
455 case PPCISD::STBRX: return "PPCISD::STBRX";
456 case PPCISD::LARX: return "PPCISD::LARX";
457 case PPCISD::STCX: return "PPCISD::STCX";
458 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
459 case PPCISD::MFFS: return "PPCISD::MFFS";
460 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
461 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
462 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
463 case PPCISD::MTFSF: return "PPCISD::MTFSF";
464 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
468 MVT::SimpleValueType PPCTargetLowering::getSetCCResultType(EVT VT) const {
472 //===----------------------------------------------------------------------===//
473 // Node matching predicates, for use by the tblgen matching code.
474 //===----------------------------------------------------------------------===//
476 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
477 static bool isFloatingPointZero(SDValue Op) {
478 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
479 return CFP->getValueAPF().isZero();
480 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
481 // Maybe this has already been legalized into the constant pool?
482 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
483 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
484 return CFP->getValueAPF().isZero();
489 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
490 /// true if Op is undef or if it matches the specified value.
491 static bool isConstantOrUndef(int Op, int Val) {
492 return Op < 0 || Op == Val;
495 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
496 /// VPKUHUM instruction.
497 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
499 for (unsigned i = 0; i != 16; ++i)
500 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
503 for (unsigned i = 0; i != 8; ++i)
504 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
505 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
511 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
512 /// VPKUWUM instruction.
513 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
515 for (unsigned i = 0; i != 16; i += 2)
516 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
517 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
520 for (unsigned i = 0; i != 8; i += 2)
521 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
522 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
523 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
524 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
530 /// isVMerge - Common function, used to match vmrg* shuffles.
532 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
533 unsigned LHSStart, unsigned RHSStart) {
534 assert(N->getValueType(0) == MVT::v16i8 &&
535 "PPC only supports shuffles by bytes!");
536 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
537 "Unsupported merge size!");
539 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
540 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
541 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
542 LHSStart+j+i*UnitSize) ||
543 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
544 RHSStart+j+i*UnitSize))
550 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
551 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
552 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
555 return isVMerge(N, UnitSize, 8, 24);
556 return isVMerge(N, UnitSize, 8, 8);
559 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
560 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
561 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
564 return isVMerge(N, UnitSize, 0, 16);
565 return isVMerge(N, UnitSize, 0, 0);
569 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
570 /// amount, otherwise return -1.
571 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
572 assert(N->getValueType(0) == MVT::v16i8 &&
573 "PPC only supports shuffles by bytes!");
575 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
577 // Find the first non-undef value in the shuffle mask.
579 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
582 if (i == 16) return -1; // all undef.
584 // Otherwise, check to see if the rest of the elements are consecutively
585 // numbered from this value.
586 unsigned ShiftAmt = SVOp->getMaskElt(i);
587 if (ShiftAmt < i) return -1;
591 // Check the rest of the elements to see if they are consecutive.
592 for (++i; i != 16; ++i)
593 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
596 // Check the rest of the elements to see if they are consecutive.
597 for (++i; i != 16; ++i)
598 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
604 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
605 /// specifies a splat of a single element that is suitable for input to
606 /// VSPLTB/VSPLTH/VSPLTW.
607 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
608 assert(N->getValueType(0) == MVT::v16i8 &&
609 (EltSize == 1 || EltSize == 2 || EltSize == 4));
611 // This is a splat operation if each element of the permute is the same, and
612 // if the value doesn't reference the second vector.
613 unsigned ElementBase = N->getMaskElt(0);
615 // FIXME: Handle UNDEF elements too!
616 if (ElementBase >= 16)
619 // Check that the indices are consecutive, in the case of a multi-byte element
620 // splatted with a v16i8 mask.
621 for (unsigned i = 1; i != EltSize; ++i)
622 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
625 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
626 if (N->getMaskElt(i) < 0) continue;
627 for (unsigned j = 0; j != EltSize; ++j)
628 if (N->getMaskElt(i+j) != N->getMaskElt(j))
634 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
636 bool PPC::isAllNegativeZeroVector(SDNode *N) {
637 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
639 APInt APVal, APUndef;
643 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
644 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
645 return CFP->getValueAPF().isNegZero();
650 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
651 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
652 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
653 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
654 assert(isSplatShuffleMask(SVOp, EltSize));
655 return SVOp->getMaskElt(0) / EltSize;
658 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
659 /// by using a vspltis[bhw] instruction of the specified element size, return
660 /// the constant being splatted. The ByteSize field indicates the number of
661 /// bytes of each element [124] -> [bhw].
662 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
665 // If ByteSize of the splat is bigger than the element size of the
666 // build_vector, then we have a case where we are checking for a splat where
667 // multiple elements of the buildvector are folded together into a single
668 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
669 unsigned EltSize = 16/N->getNumOperands();
670 if (EltSize < ByteSize) {
671 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
672 SDValue UniquedVals[4];
673 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
675 // See if all of the elements in the buildvector agree across.
676 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
677 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
678 // If the element isn't a constant, bail fully out.
679 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
682 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
683 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
684 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
685 return SDValue(); // no match.
688 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
689 // either constant or undef values that are identical for each chunk. See
690 // if these chunks can form into a larger vspltis*.
692 // Check to see if all of the leading entries are either 0 or -1. If
693 // neither, then this won't fit into the immediate field.
694 bool LeadingZero = true;
695 bool LeadingOnes = true;
696 for (unsigned i = 0; i != Multiple-1; ++i) {
697 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
699 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
700 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
702 // Finally, check the least significant entry.
704 if (UniquedVals[Multiple-1].getNode() == 0)
705 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
706 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
708 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
711 if (UniquedVals[Multiple-1].getNode() == 0)
712 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
713 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
714 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
715 return DAG.getTargetConstant(Val, MVT::i32);
721 // Check to see if this buildvec has a single non-undef value in its elements.
722 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
723 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
724 if (OpVal.getNode() == 0)
725 OpVal = N->getOperand(i);
726 else if (OpVal != N->getOperand(i))
730 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
732 unsigned ValSizeInBytes = EltSize;
734 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
735 Value = CN->getZExtValue();
736 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
737 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
738 Value = FloatToBits(CN->getValueAPF().convertToFloat());
741 // If the splat value is larger than the element value, then we can never do
742 // this splat. The only case that we could fit the replicated bits into our
743 // immediate field for would be zero, and we prefer to use vxor for it.
744 if (ValSizeInBytes < ByteSize) return SDValue();
746 // If the element value is larger than the splat value, cut it in half and
747 // check to see if the two halves are equal. Continue doing this until we
748 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
749 while (ValSizeInBytes > ByteSize) {
750 ValSizeInBytes >>= 1;
752 // If the top half equals the bottom half, we're still ok.
753 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
754 (Value & ((1 << (8*ValSizeInBytes))-1)))
758 // Properly sign extend the value.
759 int ShAmt = (4-ByteSize)*8;
760 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
762 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
763 if (MaskVal == 0) return SDValue();
765 // Finally, if this value fits in a 5 bit sext field, return it
766 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
767 return DAG.getTargetConstant(MaskVal, MVT::i32);
771 //===----------------------------------------------------------------------===//
772 // Addressing Mode Selection
773 //===----------------------------------------------------------------------===//
775 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
776 /// or 64-bit immediate, and if the value can be accurately represented as a
777 /// sign extension from a 16-bit value. If so, this returns true and the
779 static bool isIntS16Immediate(SDNode *N, short &Imm) {
780 if (N->getOpcode() != ISD::Constant)
783 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
784 if (N->getValueType(0) == MVT::i32)
785 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
787 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
789 static bool isIntS16Immediate(SDValue Op, short &Imm) {
790 return isIntS16Immediate(Op.getNode(), Imm);
794 /// SelectAddressRegReg - Given the specified addressed, check to see if it
795 /// can be represented as an indexed [r+r] operation. Returns false if it
796 /// can be more efficiently represented with [r+imm].
797 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
799 SelectionDAG &DAG) const {
801 if (N.getOpcode() == ISD::ADD) {
802 if (isIntS16Immediate(N.getOperand(1), imm))
804 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
807 Base = N.getOperand(0);
808 Index = N.getOperand(1);
810 } else if (N.getOpcode() == ISD::OR) {
811 if (isIntS16Immediate(N.getOperand(1), imm))
812 return false; // r+i can fold it if we can.
814 // If this is an or of disjoint bitfields, we can codegen this as an add
815 // (for better address arithmetic) if the LHS and RHS of the OR are provably
817 APInt LHSKnownZero, LHSKnownOne;
818 APInt RHSKnownZero, RHSKnownOne;
819 DAG.ComputeMaskedBits(N.getOperand(0),
820 APInt::getAllOnesValue(N.getOperand(0)
821 .getValueSizeInBits()),
822 LHSKnownZero, LHSKnownOne);
824 if (LHSKnownZero.getBoolValue()) {
825 DAG.ComputeMaskedBits(N.getOperand(1),
826 APInt::getAllOnesValue(N.getOperand(1)
827 .getValueSizeInBits()),
828 RHSKnownZero, RHSKnownOne);
829 // If all of the bits are known zero on the LHS or RHS, the add won't
831 if (~(LHSKnownZero | RHSKnownZero) == 0) {
832 Base = N.getOperand(0);
833 Index = N.getOperand(1);
842 /// Returns true if the address N can be represented by a base register plus
843 /// a signed 16-bit displacement [r+imm], and if it is not better
844 /// represented as reg+reg.
845 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
847 SelectionDAG &DAG) const {
848 // FIXME dl should come from parent load or store, not from address
849 DebugLoc dl = N.getDebugLoc();
850 // If this can be more profitably realized as r+r, fail.
851 if (SelectAddressRegReg(N, Disp, Base, DAG))
854 if (N.getOpcode() == ISD::ADD) {
856 if (isIntS16Immediate(N.getOperand(1), imm)) {
857 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
858 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
859 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
861 Base = N.getOperand(0);
863 return true; // [r+i]
864 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
865 // Match LOAD (ADD (X, Lo(G))).
866 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
867 && "Cannot handle constant offsets yet!");
868 Disp = N.getOperand(1).getOperand(0); // The global address.
869 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
870 Disp.getOpcode() == ISD::TargetConstantPool ||
871 Disp.getOpcode() == ISD::TargetJumpTable);
872 Base = N.getOperand(0);
873 return true; // [&g+r]
875 } else if (N.getOpcode() == ISD::OR) {
877 if (isIntS16Immediate(N.getOperand(1), imm)) {
878 // If this is an or of disjoint bitfields, we can codegen this as an add
879 // (for better address arithmetic) if the LHS and RHS of the OR are
880 // provably disjoint.
881 APInt LHSKnownZero, LHSKnownOne;
882 DAG.ComputeMaskedBits(N.getOperand(0),
883 APInt::getAllOnesValue(N.getOperand(0)
884 .getValueSizeInBits()),
885 LHSKnownZero, LHSKnownOne);
887 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
888 // If all of the bits are known zero on the LHS or RHS, the add won't
890 Base = N.getOperand(0);
891 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
895 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
896 // Loading from a constant address.
898 // If this address fits entirely in a 16-bit sext immediate field, codegen
901 if (isIntS16Immediate(CN, Imm)) {
902 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
903 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
904 CN->getValueType(0));
908 // Handle 32-bit sext immediates with LIS + addr mode.
909 if (CN->getValueType(0) == MVT::i32 ||
910 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
911 int Addr = (int)CN->getZExtValue();
913 // Otherwise, break this down into an LIS + disp.
914 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
916 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
917 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
918 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
923 Disp = DAG.getTargetConstant(0, getPointerTy());
924 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
925 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
928 return true; // [r+0]
931 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
932 /// represented as an indexed [r+r] operation.
933 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
935 SelectionDAG &DAG) const {
936 // Check to see if we can easily represent this as an [r+r] address. This
937 // will fail if it thinks that the address is more profitably represented as
938 // reg+imm, e.g. where imm = 0.
939 if (SelectAddressRegReg(N, Base, Index, DAG))
942 // If the operand is an addition, always emit this as [r+r], since this is
943 // better (for code size, and execution, as the memop does the add for free)
944 // than emitting an explicit add.
945 if (N.getOpcode() == ISD::ADD) {
946 Base = N.getOperand(0);
947 Index = N.getOperand(1);
951 // Otherwise, do it the hard way, using R0 as the base register.
952 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
958 /// SelectAddressRegImmShift - Returns true if the address N can be
959 /// represented by a base register plus a signed 14-bit displacement
960 /// [r+imm*4]. Suitable for use by STD and friends.
961 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
963 SelectionDAG &DAG) const {
964 // FIXME dl should come from the parent load or store, not the address
965 DebugLoc dl = N.getDebugLoc();
966 // If this can be more profitably realized as r+r, fail.
967 if (SelectAddressRegReg(N, Disp, Base, DAG))
970 if (N.getOpcode() == ISD::ADD) {
972 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
973 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
974 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
975 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
977 Base = N.getOperand(0);
979 return true; // [r+i]
980 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
981 // Match LOAD (ADD (X, Lo(G))).
982 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
983 && "Cannot handle constant offsets yet!");
984 Disp = N.getOperand(1).getOperand(0); // The global address.
985 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
986 Disp.getOpcode() == ISD::TargetConstantPool ||
987 Disp.getOpcode() == ISD::TargetJumpTable);
988 Base = N.getOperand(0);
989 return true; // [&g+r]
991 } else if (N.getOpcode() == ISD::OR) {
993 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
994 // If this is an or of disjoint bitfields, we can codegen this as an add
995 // (for better address arithmetic) if the LHS and RHS of the OR are
996 // provably disjoint.
997 APInt LHSKnownZero, LHSKnownOne;
998 DAG.ComputeMaskedBits(N.getOperand(0),
999 APInt::getAllOnesValue(N.getOperand(0)
1000 .getValueSizeInBits()),
1001 LHSKnownZero, LHSKnownOne);
1002 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1003 // If all of the bits are known zero on the LHS or RHS, the add won't
1005 Base = N.getOperand(0);
1006 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1010 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1011 // Loading from a constant address. Verify low two bits are clear.
1012 if ((CN->getZExtValue() & 3) == 0) {
1013 // If this address fits entirely in a 14-bit sext immediate field, codegen
1016 if (isIntS16Immediate(CN, Imm)) {
1017 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1018 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1019 CN->getValueType(0));
1023 // Fold the low-part of 32-bit absolute addresses into addr mode.
1024 if (CN->getValueType(0) == MVT::i32 ||
1025 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1026 int Addr = (int)CN->getZExtValue();
1028 // Otherwise, break this down into an LIS + disp.
1029 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1030 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1031 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1032 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1038 Disp = DAG.getTargetConstant(0, getPointerTy());
1039 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1040 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1043 return true; // [r+0]
1047 /// getPreIndexedAddressParts - returns true by value, base pointer and
1048 /// offset pointer and addressing mode by reference if the node's address
1049 /// can be legally represented as pre-indexed load / store address.
1050 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1052 ISD::MemIndexedMode &AM,
1053 SelectionDAG &DAG) const {
1054 // Disabled by default for now.
1055 if (!EnablePPCPreinc) return false;
1059 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1060 Ptr = LD->getBasePtr();
1061 VT = LD->getMemoryVT();
1063 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1064 Ptr = ST->getBasePtr();
1065 VT = ST->getMemoryVT();
1069 // PowerPC doesn't have preinc load/store instructions for vectors.
1073 // TODO: Check reg+reg first.
1075 // LDU/STU use reg+imm*4, others use reg+imm.
1076 if (VT != MVT::i64) {
1078 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1082 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1086 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1087 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1088 // sext i32 to i64 when addr mode is r+i.
1089 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1090 LD->getExtensionType() == ISD::SEXTLOAD &&
1091 isa<ConstantSDNode>(Offset))
1099 //===----------------------------------------------------------------------===//
1100 // LowerOperation implementation
1101 //===----------------------------------------------------------------------===//
1103 /// GetLabelAccessInfo - Return true if we should reference labels using a
1104 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1105 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1106 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1107 HiOpFlags = PPCII::MO_HA16;
1108 LoOpFlags = PPCII::MO_LO16;
1110 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1111 // non-darwin platform. We don't support PIC on other platforms yet.
1112 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1113 TM.getSubtarget<PPCSubtarget>().isDarwin();
1115 HiOpFlags |= PPCII::MO_PIC_FLAG;
1116 LoOpFlags |= PPCII::MO_PIC_FLAG;
1119 // If this is a reference to a global value that requires a non-lazy-ptr, make
1120 // sure that instruction lowering adds it.
1121 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1122 HiOpFlags |= PPCII::MO_NLP_FLAG;
1123 LoOpFlags |= PPCII::MO_NLP_FLAG;
1125 if (GV->hasHiddenVisibility()) {
1126 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1127 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1134 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1135 SelectionDAG &DAG) {
1136 EVT PtrVT = HiPart.getValueType();
1137 SDValue Zero = DAG.getConstant(0, PtrVT);
1138 DebugLoc DL = HiPart.getDebugLoc();
1140 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1141 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1143 // With PIC, the first instruction is actually "GR+hi(&G)".
1145 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1146 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1148 // Generate non-pic code that has direct accesses to the constant pool.
1149 // The address of the global is just (hi(&g)+lo(&g)).
1150 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1153 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1154 SelectionDAG &DAG) const {
1155 EVT PtrVT = Op.getValueType();
1156 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1157 const Constant *C = CP->getConstVal();
1159 unsigned MOHiFlag, MOLoFlag;
1160 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1162 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1164 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1165 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1168 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1169 EVT PtrVT = Op.getValueType();
1170 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1172 unsigned MOHiFlag, MOLoFlag;
1173 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1174 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1175 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1176 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1179 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1180 SelectionDAG &DAG) const {
1181 EVT PtrVT = Op.getValueType();
1183 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1185 unsigned MOHiFlag, MOLoFlag;
1186 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1187 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1188 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1189 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1192 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1193 SelectionDAG &DAG) const {
1194 EVT PtrVT = Op.getValueType();
1195 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1196 DebugLoc DL = GSDN->getDebugLoc();
1197 const GlobalValue *GV = GSDN->getGlobal();
1199 // 64-bit SVR4 ABI code is always position-independent.
1200 // The actual address of the GlobalValue is stored in the TOC.
1201 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1202 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1203 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1204 DAG.getRegister(PPC::X2, MVT::i64));
1207 unsigned MOHiFlag, MOLoFlag;
1208 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1211 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1213 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1215 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1217 // If the global reference is actually to a non-lazy-pointer, we have to do an
1218 // extra load to get the address of the global.
1219 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1220 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1225 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1226 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1227 DebugLoc dl = Op.getDebugLoc();
1229 // If we're comparing for equality to zero, expose the fact that this is
1230 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1231 // fold the new nodes.
1232 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1233 if (C->isNullValue() && CC == ISD::SETEQ) {
1234 EVT VT = Op.getOperand(0).getValueType();
1235 SDValue Zext = Op.getOperand(0);
1236 if (VT.bitsLT(MVT::i32)) {
1238 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1240 unsigned Log2b = Log2_32(VT.getSizeInBits());
1241 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1242 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1243 DAG.getConstant(Log2b, MVT::i32));
1244 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1246 // Leave comparisons against 0 and -1 alone for now, since they're usually
1247 // optimized. FIXME: revisit this when we can custom lower all setcc
1249 if (C->isAllOnesValue() || C->isNullValue())
1253 // If we have an integer seteq/setne, turn it into a compare against zero
1254 // by xor'ing the rhs with the lhs, which is faster than setting a
1255 // condition register, reading it back out, and masking the correct bit. The
1256 // normal approach here uses sub to do this instead of xor. Using xor exposes
1257 // the result to other bit-twiddling opportunities.
1258 EVT LHSVT = Op.getOperand(0).getValueType();
1259 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1260 EVT VT = Op.getValueType();
1261 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1263 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1268 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1269 const PPCSubtarget &Subtarget) const {
1270 SDNode *Node = Op.getNode();
1271 EVT VT = Node->getValueType(0);
1272 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1273 SDValue InChain = Node->getOperand(0);
1274 SDValue VAListPtr = Node->getOperand(1);
1275 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1276 DebugLoc dl = Node->getDebugLoc();
1278 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1281 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1282 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1284 InChain = GprIndex.getValue(1);
1286 if (VT == MVT::i64) {
1287 // Check if GprIndex is even
1288 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1289 DAG.getConstant(1, MVT::i32));
1290 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1291 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1292 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1293 DAG.getConstant(1, MVT::i32));
1294 // Align GprIndex to be even if it isn't
1295 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1299 // fpr index is 1 byte after gpr
1300 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1301 DAG.getConstant(1, MVT::i32));
1304 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1305 FprPtr, MachinePointerInfo(SV), MVT::i8,
1307 InChain = FprIndex.getValue(1);
1309 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1310 DAG.getConstant(8, MVT::i32));
1312 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1313 DAG.getConstant(4, MVT::i32));
1316 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1317 MachinePointerInfo(), false, false, 0);
1318 InChain = OverflowArea.getValue(1);
1320 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1321 MachinePointerInfo(), false, false, 0);
1322 InChain = RegSaveArea.getValue(1);
1324 // select overflow_area if index > 8
1325 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1326 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1328 // adjustment constant gpr_index * 4/8
1329 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1330 VT.isInteger() ? GprIndex : FprIndex,
1331 DAG.getConstant(VT.isInteger() ? 4 : 8,
1334 // OurReg = RegSaveArea + RegConstant
1335 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1338 // Floating types are 32 bytes into RegSaveArea
1339 if (VT.isFloatingPoint())
1340 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1341 DAG.getConstant(32, MVT::i32));
1343 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1344 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1345 VT.isInteger() ? GprIndex : FprIndex,
1346 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1349 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1350 VT.isInteger() ? VAListPtr : FprPtr,
1351 MachinePointerInfo(SV),
1352 MVT::i8, false, false, 0);
1354 // determine if we should load from reg_save_area or overflow_area
1355 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1357 // increase overflow_area by 4/8 if gpr/fpr > 8
1358 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1359 DAG.getConstant(VT.isInteger() ? 4 : 8,
1362 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1365 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1367 MachinePointerInfo(),
1368 MVT::i32, false, false, 0);
1370 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(), false, false, 0);
1373 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op,
1374 SelectionDAG &DAG) const {
1375 SDValue Chain = Op.getOperand(0);
1376 SDValue Trmp = Op.getOperand(1); // trampoline
1377 SDValue FPtr = Op.getOperand(2); // nested function
1378 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1379 DebugLoc dl = Op.getDebugLoc();
1381 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1382 bool isPPC64 = (PtrVT == MVT::i64);
1384 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1387 TargetLowering::ArgListTy Args;
1388 TargetLowering::ArgListEntry Entry;
1390 Entry.Ty = IntPtrTy;
1391 Entry.Node = Trmp; Args.push_back(Entry);
1393 // TrampSize == (isPPC64 ? 48 : 40);
1394 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1395 isPPC64 ? MVT::i64 : MVT::i32);
1396 Args.push_back(Entry);
1398 Entry.Node = FPtr; Args.push_back(Entry);
1399 Entry.Node = Nest; Args.push_back(Entry);
1401 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1402 std::pair<SDValue, SDValue> CallResult =
1403 LowerCallTo(Chain, Op.getValueType().getTypeForEVT(*DAG.getContext()),
1404 false, false, false, false, 0, CallingConv::C, false,
1405 /*isReturnValueUsed=*/true,
1406 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1410 { CallResult.first, CallResult.second };
1412 return DAG.getMergeValues(Ops, 2, dl);
1415 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1416 const PPCSubtarget &Subtarget) const {
1417 MachineFunction &MF = DAG.getMachineFunction();
1418 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1420 DebugLoc dl = Op.getDebugLoc();
1422 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1423 // vastart just stores the address of the VarArgsFrameIndex slot into the
1424 // memory location argument.
1425 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1426 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1427 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1428 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1429 MachinePointerInfo(SV),
1433 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1434 // We suppose the given va_list is already allocated.
1437 // char gpr; /* index into the array of 8 GPRs
1438 // * stored in the register save area
1439 // * gpr=0 corresponds to r3,
1440 // * gpr=1 to r4, etc.
1442 // char fpr; /* index into the array of 8 FPRs
1443 // * stored in the register save area
1444 // * fpr=0 corresponds to f1,
1445 // * fpr=1 to f2, etc.
1447 // char *overflow_arg_area;
1448 // /* location on stack that holds
1449 // * the next overflow argument
1451 // char *reg_save_area;
1452 // /* where r3:r10 and f1:f8 (if saved)
1458 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1459 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1462 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1464 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1466 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1469 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1470 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1472 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1473 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1475 uint64_t FPROffset = 1;
1476 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1478 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1480 // Store first byte : number of int regs
1481 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1483 MachinePointerInfo(SV),
1484 MVT::i8, false, false, 0);
1485 uint64_t nextOffset = FPROffset;
1486 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1489 // Store second byte : number of float regs
1490 SDValue secondStore =
1491 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1492 MachinePointerInfo(SV, nextOffset), MVT::i8,
1494 nextOffset += StackOffset;
1495 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1497 // Store second word : arguments given on stack
1498 SDValue thirdStore =
1499 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1500 MachinePointerInfo(SV, nextOffset),
1502 nextOffset += FrameOffset;
1503 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1505 // Store third word : arguments given in registers
1506 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1507 MachinePointerInfo(SV, nextOffset),
1512 #include "PPCGenCallingConv.inc"
1514 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1515 CCValAssign::LocInfo &LocInfo,
1516 ISD::ArgFlagsTy &ArgFlags,
1521 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1523 CCValAssign::LocInfo &LocInfo,
1524 ISD::ArgFlagsTy &ArgFlags,
1526 static const unsigned ArgRegs[] = {
1527 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1528 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1530 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1532 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1534 // Skip one register if the first unallocated register has an even register
1535 // number and there are still argument registers available which have not been
1536 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1537 // need to skip a register if RegNum is odd.
1538 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1539 State.AllocateReg(ArgRegs[RegNum]);
1542 // Always return false here, as this function only makes sure that the first
1543 // unallocated register has an odd register number and does not actually
1544 // allocate a register for the current argument.
1548 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1550 CCValAssign::LocInfo &LocInfo,
1551 ISD::ArgFlagsTy &ArgFlags,
1553 static const unsigned ArgRegs[] = {
1554 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1558 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1560 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1562 // If there is only one Floating-point register left we need to put both f64
1563 // values of a split ppc_fp128 value on the stack.
1564 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1565 State.AllocateReg(ArgRegs[RegNum]);
1568 // Always return false here, as this function only makes sure that the two f64
1569 // values a ppc_fp128 value is split into are both passed in registers or both
1570 // passed on the stack and does not actually allocate a register for the
1571 // current argument.
1575 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1577 static const unsigned *GetFPR() {
1578 static const unsigned FPR[] = {
1579 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1580 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1586 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1588 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1589 unsigned PtrByteSize) {
1590 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1591 if (Flags.isByVal())
1592 ArgSize = Flags.getByValSize();
1593 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1599 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1600 CallingConv::ID CallConv, bool isVarArg,
1601 const SmallVectorImpl<ISD::InputArg>
1603 DebugLoc dl, SelectionDAG &DAG,
1604 SmallVectorImpl<SDValue> &InVals)
1606 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1607 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1610 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1616 PPCTargetLowering::LowerFormalArguments_SVR4(
1618 CallingConv::ID CallConv, bool isVarArg,
1619 const SmallVectorImpl<ISD::InputArg>
1621 DebugLoc dl, SelectionDAG &DAG,
1622 SmallVectorImpl<SDValue> &InVals) const {
1624 // 32-bit SVR4 ABI Stack Frame Layout:
1625 // +-----------------------------------+
1626 // +--> | Back chain |
1627 // | +-----------------------------------+
1628 // | | Floating-point register save area |
1629 // | +-----------------------------------+
1630 // | | General register save area |
1631 // | +-----------------------------------+
1632 // | | CR save word |
1633 // | +-----------------------------------+
1634 // | | VRSAVE save word |
1635 // | +-----------------------------------+
1636 // | | Alignment padding |
1637 // | +-----------------------------------+
1638 // | | Vector register save area |
1639 // | +-----------------------------------+
1640 // | | Local variable space |
1641 // | +-----------------------------------+
1642 // | | Parameter list area |
1643 // | +-----------------------------------+
1644 // | | LR save word |
1645 // | +-----------------------------------+
1646 // SP--> +--- | Back chain |
1647 // +-----------------------------------+
1650 // System V Application Binary Interface PowerPC Processor Supplement
1651 // AltiVec Technology Programming Interface Manual
1653 MachineFunction &MF = DAG.getMachineFunction();
1654 MachineFrameInfo *MFI = MF.getFrameInfo();
1655 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1657 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1658 // Potential tail calls could cause overwriting of argument stack slots.
1659 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1660 unsigned PtrByteSize = 4;
1662 // Assign locations to all of the incoming arguments.
1663 SmallVector<CCValAssign, 16> ArgLocs;
1664 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1665 getTargetMachine(), ArgLocs, *DAG.getContext());
1667 // Reserve space for the linkage area on the stack.
1668 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1670 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1672 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1673 CCValAssign &VA = ArgLocs[i];
1675 // Arguments stored in registers.
1676 if (VA.isRegLoc()) {
1677 TargetRegisterClass *RC;
1678 EVT ValVT = VA.getValVT();
1680 switch (ValVT.getSimpleVT().SimpleTy) {
1682 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1684 RC = PPC::GPRCRegisterClass;
1687 RC = PPC::F4RCRegisterClass;
1690 RC = PPC::F8RCRegisterClass;
1696 RC = PPC::VRRCRegisterClass;
1700 // Transform the arguments stored in physical registers into virtual ones.
1701 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1702 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1704 InVals.push_back(ArgValue);
1706 // Argument stored in memory.
1707 assert(VA.isMemLoc());
1709 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1710 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1713 // Create load nodes to retrieve arguments from the stack.
1714 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1715 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1716 MachinePointerInfo(),
1721 // Assign locations to all of the incoming aggregate by value arguments.
1722 // Aggregates passed by value are stored in the local variable space of the
1723 // caller's stack frame, right above the parameter list area.
1724 SmallVector<CCValAssign, 16> ByValArgLocs;
1725 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1726 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1728 // Reserve stack space for the allocations in CCInfo.
1729 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1731 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1733 // Area that is at least reserved in the caller of this function.
1734 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1736 // Set the size that is at least reserved in caller of this function. Tail
1737 // call optimized function's reserved stack space needs to be aligned so that
1738 // taking the difference between two stack areas will result in an aligned
1740 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1743 std::max(MinReservedArea,
1744 PPCFrameLowering::getMinCallFrameSize(false, false));
1746 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1747 getStackAlignment();
1748 unsigned AlignMask = TargetAlign-1;
1749 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1751 FI->setMinReservedArea(MinReservedArea);
1753 SmallVector<SDValue, 8> MemOps;
1755 // If the function takes variable number of arguments, make a frame index for
1756 // the start of the first vararg value... for expansion of llvm.va_start.
1758 static const unsigned GPArgRegs[] = {
1759 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1760 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1762 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1764 static const unsigned FPArgRegs[] = {
1765 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1768 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1770 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1772 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1775 // Make room for NumGPArgRegs and NumFPArgRegs.
1776 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1777 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1779 FuncInfo->setVarArgsStackOffset(
1780 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1781 CCInfo.getNextStackOffset(), true));
1783 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1784 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1786 // The fixed integer arguments of a variadic function are stored to the
1787 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1788 // the result of va_next.
1789 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1790 // Get an existing live-in vreg, or add a new one.
1791 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1793 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1795 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1796 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1797 MachinePointerInfo(), false, false, 0);
1798 MemOps.push_back(Store);
1799 // Increment the address by four for the next argument to store
1800 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1801 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1804 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1806 // The double arguments are stored to the VarArgsFrameIndex
1808 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1809 // Get an existing live-in vreg, or add a new one.
1810 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1812 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1814 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1815 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1816 MachinePointerInfo(), false, false, 0);
1817 MemOps.push_back(Store);
1818 // Increment the address by eight for the next argument to store
1819 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1821 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1825 if (!MemOps.empty())
1826 Chain = DAG.getNode(ISD::TokenFactor, dl,
1827 MVT::Other, &MemOps[0], MemOps.size());
1833 PPCTargetLowering::LowerFormalArguments_Darwin(
1835 CallingConv::ID CallConv, bool isVarArg,
1836 const SmallVectorImpl<ISD::InputArg>
1838 DebugLoc dl, SelectionDAG &DAG,
1839 SmallVectorImpl<SDValue> &InVals) const {
1840 // TODO: add description of PPC stack frame format, or at least some docs.
1842 MachineFunction &MF = DAG.getMachineFunction();
1843 MachineFrameInfo *MFI = MF.getFrameInfo();
1844 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1846 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1847 bool isPPC64 = PtrVT == MVT::i64;
1848 // Potential tail calls could cause overwriting of argument stack slots.
1849 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1850 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1852 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1853 // Area that is at least reserved in caller of this function.
1854 unsigned MinReservedArea = ArgOffset;
1856 static const unsigned GPR_32[] = { // 32-bit registers.
1857 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1858 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1860 static const unsigned GPR_64[] = { // 64-bit registers.
1861 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1862 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1865 static const unsigned *FPR = GetFPR();
1867 static const unsigned VR[] = {
1868 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1869 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1872 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1873 const unsigned Num_FPR_Regs = 13;
1874 const unsigned Num_VR_Regs = array_lengthof( VR);
1876 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1878 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1880 // In 32-bit non-varargs functions, the stack space for vectors is after the
1881 // stack space for non-vectors. We do not use this space unless we have
1882 // too many vectors to fit in registers, something that only occurs in
1883 // constructed examples:), but we have to walk the arglist to figure
1884 // that out...for the pathological case, compute VecArgOffset as the
1885 // start of the vector parameter area. Computing VecArgOffset is the
1886 // entire point of the following loop.
1887 unsigned VecArgOffset = ArgOffset;
1888 if (!isVarArg && !isPPC64) {
1889 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1891 EVT ObjectVT = Ins[ArgNo].VT;
1892 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1893 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1895 if (Flags.isByVal()) {
1896 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1897 ObjSize = Flags.getByValSize();
1899 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1900 VecArgOffset += ArgSize;
1904 switch(ObjectVT.getSimpleVT().SimpleTy) {
1905 default: llvm_unreachable("Unhandled argument type!");
1908 VecArgOffset += isPPC64 ? 8 : 4;
1910 case MVT::i64: // PPC64
1918 // Nothing to do, we're only looking at Nonvector args here.
1923 // We've found where the vector parameter area in memory is. Skip the
1924 // first 12 parameters; these don't use that memory.
1925 VecArgOffset = ((VecArgOffset+15)/16)*16;
1926 VecArgOffset += 12*16;
1928 // Add DAG nodes to load the arguments or copy them out of registers. On
1929 // entry to a function on PPC, the arguments start after the linkage area,
1930 // although the first ones are often in registers.
1932 SmallVector<SDValue, 8> MemOps;
1933 unsigned nAltivecParamsAtEnd = 0;
1934 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1936 bool needsLoad = false;
1937 EVT ObjectVT = Ins[ArgNo].VT;
1938 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1939 unsigned ArgSize = ObjSize;
1940 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1942 unsigned CurArgOffset = ArgOffset;
1944 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1945 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1946 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1947 if (isVarArg || isPPC64) {
1948 MinReservedArea = ((MinReservedArea+15)/16)*16;
1949 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1952 } else nAltivecParamsAtEnd++;
1954 // Calculate min reserved area.
1955 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1959 // FIXME the codegen can be much improved in some cases.
1960 // We do not have to keep everything in memory.
1961 if (Flags.isByVal()) {
1962 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1963 ObjSize = Flags.getByValSize();
1964 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1965 // Objects of size 1 and 2 are right justified, everything else is
1966 // left justified. This means the memory address is adjusted forwards.
1967 if (ObjSize==1 || ObjSize==2) {
1968 CurArgOffset = CurArgOffset + (4 - ObjSize);
1970 // The value of the object is its address.
1971 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
1972 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1973 InVals.push_back(FIN);
1974 if (ObjSize==1 || ObjSize==2) {
1975 if (GPR_idx != Num_GPR_Regs) {
1978 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1980 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1981 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1982 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1983 MachinePointerInfo(),
1984 ObjSize==1 ? MVT::i8 : MVT::i16,
1986 MemOps.push_back(Store);
1990 ArgOffset += PtrByteSize;
1994 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1995 // Store whatever pieces of the object are in registers
1996 // to memory. ArgVal will be address of the beginning of
1998 if (GPR_idx != Num_GPR_Regs) {
2001 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2003 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2004 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2005 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2006 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2007 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2008 MachinePointerInfo(),
2010 MemOps.push_back(Store);
2012 ArgOffset += PtrByteSize;
2014 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2021 switch (ObjectVT.getSimpleVT().SimpleTy) {
2022 default: llvm_unreachable("Unhandled argument type!");
2025 if (GPR_idx != Num_GPR_Regs) {
2026 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2027 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2031 ArgSize = PtrByteSize;
2033 // All int arguments reserve stack space in the Darwin ABI.
2034 ArgOffset += PtrByteSize;
2038 case MVT::i64: // PPC64
2039 if (GPR_idx != Num_GPR_Regs) {
2040 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2041 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2043 if (ObjectVT == MVT::i32) {
2044 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2045 // value to MVT::i64 and then truncate to the correct register size.
2047 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2048 DAG.getValueType(ObjectVT));
2049 else if (Flags.isZExt())
2050 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2051 DAG.getValueType(ObjectVT));
2053 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2059 ArgSize = PtrByteSize;
2061 // All int arguments reserve stack space in the Darwin ABI.
2067 // Every 4 bytes of argument space consumes one of the GPRs available for
2068 // argument passing.
2069 if (GPR_idx != Num_GPR_Regs) {
2071 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2074 if (FPR_idx != Num_FPR_Regs) {
2077 if (ObjectVT == MVT::f32)
2078 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2080 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2082 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2088 // All FP arguments reserve stack space in the Darwin ABI.
2089 ArgOffset += isPPC64 ? 8 : ObjSize;
2095 // Note that vector arguments in registers don't reserve stack space,
2096 // except in varargs functions.
2097 if (VR_idx != Num_VR_Regs) {
2098 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2099 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2101 while ((ArgOffset % 16) != 0) {
2102 ArgOffset += PtrByteSize;
2103 if (GPR_idx != Num_GPR_Regs)
2107 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2111 if (!isVarArg && !isPPC64) {
2112 // Vectors go after all the nonvectors.
2113 CurArgOffset = VecArgOffset;
2116 // Vectors are aligned.
2117 ArgOffset = ((ArgOffset+15)/16)*16;
2118 CurArgOffset = ArgOffset;
2126 // We need to load the argument to a virtual register if we determined above
2127 // that we ran out of physical registers of the appropriate type.
2129 int FI = MFI->CreateFixedObject(ObjSize,
2130 CurArgOffset + (ArgSize - ObjSize),
2132 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2133 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2137 InVals.push_back(ArgVal);
2140 // Set the size that is at least reserved in caller of this function. Tail
2141 // call optimized function's reserved stack space needs to be aligned so that
2142 // taking the difference between two stack areas will result in an aligned
2144 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2145 // Add the Altivec parameters at the end, if needed.
2146 if (nAltivecParamsAtEnd) {
2147 MinReservedArea = ((MinReservedArea+15)/16)*16;
2148 MinReservedArea += 16*nAltivecParamsAtEnd;
2151 std::max(MinReservedArea,
2152 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2153 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2154 getStackAlignment();
2155 unsigned AlignMask = TargetAlign-1;
2156 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2157 FI->setMinReservedArea(MinReservedArea);
2159 // If the function takes variable number of arguments, make a frame index for
2160 // the start of the first vararg value... for expansion of llvm.va_start.
2162 int Depth = ArgOffset;
2164 FuncInfo->setVarArgsFrameIndex(
2165 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2167 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2169 // If this function is vararg, store any remaining integer argument regs
2170 // to their spots on the stack so that they may be loaded by deferencing the
2171 // result of va_next.
2172 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2176 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2178 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2180 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2181 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2182 MachinePointerInfo(), false, false, 0);
2183 MemOps.push_back(Store);
2184 // Increment the address by four for the next argument to store
2185 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2186 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2190 if (!MemOps.empty())
2191 Chain = DAG.getNode(ISD::TokenFactor, dl,
2192 MVT::Other, &MemOps[0], MemOps.size());
2197 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2198 /// linkage area for the Darwin ABI.
2200 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2204 const SmallVectorImpl<ISD::OutputArg>
2206 const SmallVectorImpl<SDValue> &OutVals,
2207 unsigned &nAltivecParamsAtEnd) {
2208 // Count how many bytes are to be pushed on the stack, including the linkage
2209 // area, and parameter passing area. We start with 24/48 bytes, which is
2210 // prereserved space for [SP][CR][LR][3 x unused].
2211 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2212 unsigned NumOps = Outs.size();
2213 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2215 // Add up all the space actually used.
2216 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2217 // they all go in registers, but we must reserve stack space for them for
2218 // possible use by the caller. In varargs or 64-bit calls, parameters are
2219 // assigned stack space in order, with padding so Altivec parameters are
2221 nAltivecParamsAtEnd = 0;
2222 for (unsigned i = 0; i != NumOps; ++i) {
2223 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2224 EVT ArgVT = Outs[i].VT;
2225 // Varargs Altivec parameters are padded to a 16 byte boundary.
2226 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2227 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2228 if (!isVarArg && !isPPC64) {
2229 // Non-varargs Altivec parameters go after all the non-Altivec
2230 // parameters; handle those later so we know how much padding we need.
2231 nAltivecParamsAtEnd++;
2234 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2235 NumBytes = ((NumBytes+15)/16)*16;
2237 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2240 // Allow for Altivec parameters at the end, if needed.
2241 if (nAltivecParamsAtEnd) {
2242 NumBytes = ((NumBytes+15)/16)*16;
2243 NumBytes += 16*nAltivecParamsAtEnd;
2246 // The prolog code of the callee may store up to 8 GPR argument registers to
2247 // the stack, allowing va_start to index over them in memory if its varargs.
2248 // Because we cannot tell if this is needed on the caller side, we have to
2249 // conservatively assume that it is needed. As such, make sure we have at
2250 // least enough stack space for the caller to store the 8 GPRs.
2251 NumBytes = std::max(NumBytes,
2252 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2254 // Tail call needs the stack to be aligned.
2255 if (CC==CallingConv::Fast && GuaranteedTailCallOpt) {
2256 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2257 getStackAlignment();
2258 unsigned AlignMask = TargetAlign-1;
2259 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2265 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2266 /// adjusted to accommodate the arguments for the tailcall.
2267 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2268 unsigned ParamSize) {
2270 if (!isTailCall) return 0;
2272 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2273 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2274 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2275 // Remember only if the new adjustement is bigger.
2276 if (SPDiff < FI->getTailCallSPDelta())
2277 FI->setTailCallSPDelta(SPDiff);
2282 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2283 /// for tail call optimization. Targets which want to do tail call
2284 /// optimization should implement this function.
2286 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2287 CallingConv::ID CalleeCC,
2289 const SmallVectorImpl<ISD::InputArg> &Ins,
2290 SelectionDAG& DAG) const {
2291 if (!GuaranteedTailCallOpt)
2294 // Variable argument functions are not supported.
2298 MachineFunction &MF = DAG.getMachineFunction();
2299 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2300 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2301 // Functions containing by val parameters are not supported.
2302 for (unsigned i = 0; i != Ins.size(); i++) {
2303 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2304 if (Flags.isByVal()) return false;
2307 // Non PIC/GOT tail calls are supported.
2308 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2311 // At the moment we can only do local tail calls (in same module, hidden
2312 // or protected) if we are generating PIC.
2313 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2314 return G->getGlobal()->hasHiddenVisibility()
2315 || G->getGlobal()->hasProtectedVisibility();
2321 /// isCallCompatibleAddress - Return the immediate to use if the specified
2322 /// 32-bit value is representable in the immediate field of a BxA instruction.
2323 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2324 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2327 int Addr = C->getZExtValue();
2328 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2329 (Addr << 6 >> 6) != Addr)
2330 return 0; // Top 6 bits have to be sext of immediate.
2332 return DAG.getConstant((int)C->getZExtValue() >> 2,
2333 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2338 struct TailCallArgumentInfo {
2343 TailCallArgumentInfo() : FrameIdx(0) {}
2348 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2350 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2352 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2353 SmallVector<SDValue, 8> &MemOpChains,
2355 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2356 SDValue Arg = TailCallArgs[i].Arg;
2357 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2358 int FI = TailCallArgs[i].FrameIdx;
2359 // Store relative to framepointer.
2360 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2361 MachinePointerInfo::getFixedStack(FI),
2366 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2367 /// the appropriate stack slot for the tail call optimized function call.
2368 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2369 MachineFunction &MF,
2378 // Calculate the new stack slot for the return address.
2379 int SlotSize = isPPC64 ? 8 : 4;
2380 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2382 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2383 NewRetAddrLoc, true);
2384 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2385 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2386 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2387 MachinePointerInfo::getFixedStack(NewRetAddr),
2390 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2391 // slot as the FP is never overwritten.
2394 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2395 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2397 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2398 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2399 MachinePointerInfo::getFixedStack(NewFPIdx),
2406 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2407 /// the position of the argument.
2409 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2410 SDValue Arg, int SPDiff, unsigned ArgOffset,
2411 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2412 int Offset = ArgOffset + SPDiff;
2413 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2414 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2415 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2416 SDValue FIN = DAG.getFrameIndex(FI, VT);
2417 TailCallArgumentInfo Info;
2419 Info.FrameIdxOp = FIN;
2421 TailCallArguments.push_back(Info);
2424 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2425 /// stack slot. Returns the chain as result and the loaded frame pointers in
2426 /// LROpOut/FPOpout. Used when tail calling.
2427 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2433 DebugLoc dl) const {
2435 // Load the LR and FP stack slot for later adjusting.
2436 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2437 LROpOut = getReturnAddrFrameIndex(DAG);
2438 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2440 Chain = SDValue(LROpOut.getNode(), 1);
2442 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2443 // slot as the FP is never overwritten.
2445 FPOpOut = getFramePointerFrameIndex(DAG);
2446 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2448 Chain = SDValue(FPOpOut.getNode(), 1);
2454 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2455 /// by "Src" to address "Dst" of size "Size". Alignment information is
2456 /// specified by the specific parameter attribute. The copy will be passed as
2457 /// a byval function parameter.
2458 /// Sometimes what we are copying is the end of a larger object, the part that
2459 /// does not fit in registers.
2461 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2462 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2464 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2465 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2466 false, false, MachinePointerInfo(0),
2467 MachinePointerInfo(0));
2470 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2473 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2474 SDValue Arg, SDValue PtrOff, int SPDiff,
2475 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2476 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2477 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2479 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2484 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2486 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2487 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2488 DAG.getConstant(ArgOffset, PtrVT));
2490 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2491 MachinePointerInfo(), false, false, 0));
2492 // Calculate and remember argument location.
2493 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2498 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2499 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2500 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2501 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2502 MachineFunction &MF = DAG.getMachineFunction();
2504 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2505 // might overwrite each other in case of tail call optimization.
2506 SmallVector<SDValue, 8> MemOpChains2;
2507 // Do not flag preceding copytoreg stuff together with the following stuff.
2509 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2511 if (!MemOpChains2.empty())
2512 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2513 &MemOpChains2[0], MemOpChains2.size());
2515 // Store the return address to the appropriate stack slot.
2516 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2517 isPPC64, isDarwinABI, dl);
2519 // Emit callseq_end just before tailcall node.
2520 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2521 DAG.getIntPtrConstant(0, true), InFlag);
2522 InFlag = Chain.getValue(1);
2526 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2527 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2528 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2529 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2530 const PPCSubtarget &PPCSubTarget) {
2532 bool isPPC64 = PPCSubTarget.isPPC64();
2533 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2535 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2536 NodeTys.push_back(MVT::Other); // Returns a chain
2537 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2539 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2541 bool needIndirectCall = true;
2542 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2543 // If this is an absolute destination address, use the munged value.
2544 Callee = SDValue(Dest, 0);
2545 needIndirectCall = false;
2548 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2549 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2550 // Use indirect calls for ALL functions calls in JIT mode, since the
2551 // far-call stubs may be outside relocation limits for a BL instruction.
2552 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2553 unsigned OpFlags = 0;
2554 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2555 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2556 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2557 (G->getGlobal()->isDeclaration() ||
2558 G->getGlobal()->isWeakForLinker())) {
2559 // PC-relative references to external symbols should go through $stub,
2560 // unless we're building with the leopard linker or later, which
2561 // automatically synthesizes these stubs.
2562 OpFlags = PPCII::MO_DARWIN_STUB;
2565 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2566 // every direct call is) turn it into a TargetGlobalAddress /
2567 // TargetExternalSymbol node so that legalize doesn't hack it.
2568 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2569 Callee.getValueType(),
2571 needIndirectCall = false;
2575 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2576 unsigned char OpFlags = 0;
2578 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2579 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2580 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2581 // PC-relative references to external symbols should go through $stub,
2582 // unless we're building with the leopard linker or later, which
2583 // automatically synthesizes these stubs.
2584 OpFlags = PPCII::MO_DARWIN_STUB;
2587 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2589 needIndirectCall = false;
2592 if (needIndirectCall) {
2593 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2594 // to do the call, we can't use PPCISD::CALL.
2595 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2597 if (isSVR4ABI && isPPC64) {
2598 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2599 // entry point, but to the function descriptor (the function entry point
2600 // address is part of the function descriptor though).
2601 // The function descriptor is a three doubleword structure with the
2602 // following fields: function entry point, TOC base address and
2603 // environment pointer.
2604 // Thus for a call through a function pointer, the following actions need
2606 // 1. Save the TOC of the caller in the TOC save area of its stack
2607 // frame (this is done in LowerCall_Darwin()).
2608 // 2. Load the address of the function entry point from the function
2610 // 3. Load the TOC of the callee from the function descriptor into r2.
2611 // 4. Load the environment pointer from the function descriptor into
2613 // 5. Branch to the function entry point address.
2614 // 6. On return of the callee, the TOC of the caller needs to be
2615 // restored (this is done in FinishCall()).
2617 // All those operations are flagged together to ensure that no other
2618 // operations can be scheduled in between. E.g. without flagging the
2619 // operations together, a TOC access in the caller could be scheduled
2620 // between the load of the callee TOC and the branch to the callee, which
2621 // results in the TOC access going through the TOC of the callee instead
2622 // of going through the TOC of the caller, which leads to incorrect code.
2624 // Load the address of the function entry point from the function
2626 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2627 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2628 InFlag.getNode() ? 3 : 2);
2629 Chain = LoadFuncPtr.getValue(1);
2630 InFlag = LoadFuncPtr.getValue(2);
2632 // Load environment pointer into r11.
2633 // Offset of the environment pointer within the function descriptor.
2634 SDValue PtrOff = DAG.getIntPtrConstant(16);
2636 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2637 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2639 Chain = LoadEnvPtr.getValue(1);
2640 InFlag = LoadEnvPtr.getValue(2);
2642 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2644 Chain = EnvVal.getValue(0);
2645 InFlag = EnvVal.getValue(1);
2647 // Load TOC of the callee into r2. We are using a target-specific load
2648 // with r2 hard coded, because the result of a target-independent load
2649 // would never go directly into r2, since r2 is a reserved register (which
2650 // prevents the register allocator from allocating it), resulting in an
2651 // additional register being allocated and an unnecessary move instruction
2653 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2654 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2656 Chain = LoadTOCPtr.getValue(0);
2657 InFlag = LoadTOCPtr.getValue(1);
2659 MTCTROps[0] = Chain;
2660 MTCTROps[1] = LoadFuncPtr;
2661 MTCTROps[2] = InFlag;
2664 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2665 2 + (InFlag.getNode() != 0));
2666 InFlag = Chain.getValue(1);
2669 NodeTys.push_back(MVT::Other);
2670 NodeTys.push_back(MVT::Glue);
2671 Ops.push_back(Chain);
2672 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2674 // Add CTR register as callee so a bctr can be emitted later.
2676 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2679 // If this is a direct call, pass the chain and the callee.
2680 if (Callee.getNode()) {
2681 Ops.push_back(Chain);
2682 Ops.push_back(Callee);
2684 // If this is a tail call add stack pointer delta.
2686 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2688 // Add argument registers to the end of the list so that they are known live
2690 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2691 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2692 RegsToPass[i].second.getValueType()));
2698 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2699 CallingConv::ID CallConv, bool isVarArg,
2700 const SmallVectorImpl<ISD::InputArg> &Ins,
2701 DebugLoc dl, SelectionDAG &DAG,
2702 SmallVectorImpl<SDValue> &InVals) const {
2704 SmallVector<CCValAssign, 16> RVLocs;
2705 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2706 getTargetMachine(), RVLocs, *DAG.getContext());
2707 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2709 // Copy all of the result registers out of their specified physreg.
2710 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2711 CCValAssign &VA = RVLocs[i];
2712 EVT VT = VA.getValVT();
2713 assert(VA.isRegLoc() && "Can only return in registers!");
2714 Chain = DAG.getCopyFromReg(Chain, dl,
2715 VA.getLocReg(), VT, InFlag).getValue(1);
2716 InVals.push_back(Chain.getValue(0));
2717 InFlag = Chain.getValue(2);
2724 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2725 bool isTailCall, bool isVarArg,
2727 SmallVector<std::pair<unsigned, SDValue>, 8>
2729 SDValue InFlag, SDValue Chain,
2731 int SPDiff, unsigned NumBytes,
2732 const SmallVectorImpl<ISD::InputArg> &Ins,
2733 SmallVectorImpl<SDValue> &InVals) const {
2734 std::vector<EVT> NodeTys;
2735 SmallVector<SDValue, 8> Ops;
2736 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2737 isTailCall, RegsToPass, Ops, NodeTys,
2740 // When performing tail call optimization the callee pops its arguments off
2741 // the stack. Account for this here so these bytes can be pushed back on in
2742 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2743 int BytesCalleePops =
2744 (CallConv==CallingConv::Fast && GuaranteedTailCallOpt) ? NumBytes : 0;
2746 if (InFlag.getNode())
2747 Ops.push_back(InFlag);
2751 // If this is the first return lowered for this function, add the regs
2752 // to the liveout set for the function.
2753 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2754 SmallVector<CCValAssign, 16> RVLocs;
2755 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2756 getTargetMachine(), RVLocs, *DAG.getContext());
2757 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2758 for (unsigned i = 0; i != RVLocs.size(); ++i)
2759 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2762 assert(((Callee.getOpcode() == ISD::Register &&
2763 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2764 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2765 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2766 isa<ConstantSDNode>(Callee)) &&
2767 "Expecting an global address, external symbol, absolute value or register");
2769 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2772 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2773 InFlag = Chain.getValue(1);
2775 // Add a NOP immediately after the branch instruction when using the 64-bit
2776 // SVR4 ABI. At link time, if caller and callee are in a different module and
2777 // thus have a different TOC, the call will be replaced with a call to a stub
2778 // function which saves the current TOC, loads the TOC of the callee and
2779 // branches to the callee. The NOP will be replaced with a load instruction
2780 // which restores the TOC of the caller from the TOC save slot of the current
2781 // stack frame. If caller and callee belong to the same module (and have the
2782 // same TOC), the NOP will remain unchanged.
2783 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2784 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2785 if (CallOpc == PPCISD::BCTRL_SVR4) {
2786 // This is a call through a function pointer.
2787 // Restore the caller TOC from the save area into R2.
2788 // See PrepareCall() for more information about calls through function
2789 // pointers in the 64-bit SVR4 ABI.
2790 // We are using a target-specific load with r2 hard coded, because the
2791 // result of a target-independent load would never go directly into r2,
2792 // since r2 is a reserved register (which prevents the register allocator
2793 // from allocating it), resulting in an additional register being
2794 // allocated and an unnecessary move instruction being generated.
2795 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2796 InFlag = Chain.getValue(1);
2798 // Otherwise insert NOP.
2799 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Glue, InFlag);
2803 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2804 DAG.getIntPtrConstant(BytesCalleePops, true),
2807 InFlag = Chain.getValue(1);
2809 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2810 Ins, dl, DAG, InVals);
2814 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2815 CallingConv::ID CallConv, bool isVarArg,
2817 const SmallVectorImpl<ISD::OutputArg> &Outs,
2818 const SmallVectorImpl<SDValue> &OutVals,
2819 const SmallVectorImpl<ISD::InputArg> &Ins,
2820 DebugLoc dl, SelectionDAG &DAG,
2821 SmallVectorImpl<SDValue> &InVals) const {
2823 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2826 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2827 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2828 isTailCall, Outs, OutVals, Ins,
2831 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2832 isTailCall, Outs, OutVals, Ins,
2837 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2838 CallingConv::ID CallConv, bool isVarArg,
2840 const SmallVectorImpl<ISD::OutputArg> &Outs,
2841 const SmallVectorImpl<SDValue> &OutVals,
2842 const SmallVectorImpl<ISD::InputArg> &Ins,
2843 DebugLoc dl, SelectionDAG &DAG,
2844 SmallVectorImpl<SDValue> &InVals) const {
2845 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2846 // of the 32-bit SVR4 ABI stack frame layout.
2848 assert((CallConv == CallingConv::C ||
2849 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2851 unsigned PtrByteSize = 4;
2853 MachineFunction &MF = DAG.getMachineFunction();
2855 // Mark this function as potentially containing a function that contains a
2856 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2857 // and restoring the callers stack pointer in this functions epilog. This is
2858 // done because by tail calling the called function might overwrite the value
2859 // in this function's (MF) stack pointer stack slot 0(SP).
2860 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2861 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2863 // Count how many bytes are to be pushed on the stack, including the linkage
2864 // area, parameter list area and the part of the local variable space which
2865 // contains copies of aggregates which are passed by value.
2867 // Assign locations to all of the outgoing arguments.
2868 SmallVector<CCValAssign, 16> ArgLocs;
2869 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2870 getTargetMachine(), ArgLocs, *DAG.getContext());
2872 // Reserve space for the linkage area on the stack.
2873 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2876 // Handle fixed and variable vector arguments differently.
2877 // Fixed vector arguments go into registers as long as registers are
2878 // available. Variable vector arguments always go into memory.
2879 unsigned NumArgs = Outs.size();
2881 for (unsigned i = 0; i != NumArgs; ++i) {
2882 MVT ArgVT = Outs[i].VT;
2883 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2886 if (Outs[i].IsFixed) {
2887 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2890 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2896 errs() << "Call operand #" << i << " has unhandled type "
2897 << EVT(ArgVT).getEVTString() << "\n";
2899 llvm_unreachable(0);
2903 // All arguments are treated the same.
2904 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2907 // Assign locations to all of the outgoing aggregate by value arguments.
2908 SmallVector<CCValAssign, 16> ByValArgLocs;
2909 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2910 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2912 // Reserve stack space for the allocations in CCInfo.
2913 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2915 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2917 // Size of the linkage area, parameter list area and the part of the local
2918 // space variable where copies of aggregates which are passed by value are
2920 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2922 // Calculate by how many bytes the stack has to be adjusted in case of tail
2923 // call optimization.
2924 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2926 // Adjust the stack pointer for the new arguments...
2927 // These operations are automatically eliminated by the prolog/epilog pass
2928 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2929 SDValue CallSeqStart = Chain;
2931 // Load the return address and frame pointer so it can be moved somewhere else
2934 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2937 // Set up a copy of the stack pointer for use loading and storing any
2938 // arguments that may not fit in the registers available for argument
2940 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2942 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2943 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2944 SmallVector<SDValue, 8> MemOpChains;
2946 // Walk the register/memloc assignments, inserting copies/loads.
2947 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2950 CCValAssign &VA = ArgLocs[i];
2951 SDValue Arg = OutVals[i];
2952 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2954 if (Flags.isByVal()) {
2955 // Argument is an aggregate which is passed by value, thus we need to
2956 // create a copy of it in the local variable space of the current stack
2957 // frame (which is the stack frame of the caller) and pass the address of
2958 // this copy to the callee.
2959 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2960 CCValAssign &ByValVA = ByValArgLocs[j++];
2961 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2963 // Memory reserved in the local variable space of the callers stack frame.
2964 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2966 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2967 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2969 // Create a copy of the argument in the local area of the current
2971 SDValue MemcpyCall =
2972 CreateCopyOfByValArgument(Arg, PtrOff,
2973 CallSeqStart.getNode()->getOperand(0),
2976 // This must go outside the CALLSEQ_START..END.
2977 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2978 CallSeqStart.getNode()->getOperand(1));
2979 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2980 NewCallSeqStart.getNode());
2981 Chain = CallSeqStart = NewCallSeqStart;
2983 // Pass the address of the aggregate copy on the stack either in a
2984 // physical register or in the parameter list area of the current stack
2985 // frame to the callee.
2989 if (VA.isRegLoc()) {
2990 // Put argument in a physical register.
2991 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2993 // Put argument in the parameter list area of the current stack frame.
2994 assert(VA.isMemLoc());
2995 unsigned LocMemOffset = VA.getLocMemOffset();
2998 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2999 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3001 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3002 MachinePointerInfo(),
3005 // Calculate and remember argument location.
3006 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3012 if (!MemOpChains.empty())
3013 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3014 &MemOpChains[0], MemOpChains.size());
3016 // Set CR6 to true if this is a vararg call.
3018 SDValue SetCR(DAG.getMachineNode(PPC::CRSET, dl, MVT::i32), 0);
3019 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3022 // Build a sequence of copy-to-reg nodes chained together with token chain
3023 // and flag operands which copy the outgoing args into the appropriate regs.
3025 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3026 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3027 RegsToPass[i].second, InFlag);
3028 InFlag = Chain.getValue(1);
3032 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3033 false, TailCallArguments);
3035 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3036 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3041 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3042 CallingConv::ID CallConv, bool isVarArg,
3044 const SmallVectorImpl<ISD::OutputArg> &Outs,
3045 const SmallVectorImpl<SDValue> &OutVals,
3046 const SmallVectorImpl<ISD::InputArg> &Ins,
3047 DebugLoc dl, SelectionDAG &DAG,
3048 SmallVectorImpl<SDValue> &InVals) const {
3050 unsigned NumOps = Outs.size();
3052 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3053 bool isPPC64 = PtrVT == MVT::i64;
3054 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3056 MachineFunction &MF = DAG.getMachineFunction();
3058 // Mark this function as potentially containing a function that contains a
3059 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3060 // and restoring the callers stack pointer in this functions epilog. This is
3061 // done because by tail calling the called function might overwrite the value
3062 // in this function's (MF) stack pointer stack slot 0(SP).
3063 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
3064 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3066 unsigned nAltivecParamsAtEnd = 0;
3068 // Count how many bytes are to be pushed on the stack, including the linkage
3069 // area, and parameter passing area. We start with 24/48 bytes, which is
3070 // prereserved space for [SP][CR][LR][3 x unused].
3072 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3074 nAltivecParamsAtEnd);
3076 // Calculate by how many bytes the stack has to be adjusted in case of tail
3077 // call optimization.
3078 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3080 // To protect arguments on the stack from being clobbered in a tail call,
3081 // force all the loads to happen before doing any other lowering.
3083 Chain = DAG.getStackArgumentTokenFactor(Chain);
3085 // Adjust the stack pointer for the new arguments...
3086 // These operations are automatically eliminated by the prolog/epilog pass
3087 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3088 SDValue CallSeqStart = Chain;
3090 // Load the return address and frame pointer so it can be move somewhere else
3093 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3096 // Set up a copy of the stack pointer for use loading and storing any
3097 // arguments that may not fit in the registers available for argument
3101 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3103 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3105 // Figure out which arguments are going to go in registers, and which in
3106 // memory. Also, if this is a vararg function, floating point operations
3107 // must be stored to our stack, and loaded into integer regs as well, if
3108 // any integer regs are available for argument passing.
3109 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3110 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3112 static const unsigned GPR_32[] = { // 32-bit registers.
3113 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3114 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3116 static const unsigned GPR_64[] = { // 64-bit registers.
3117 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3118 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3120 static const unsigned *FPR = GetFPR();
3122 static const unsigned VR[] = {
3123 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3124 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3126 const unsigned NumGPRs = array_lengthof(GPR_32);
3127 const unsigned NumFPRs = 13;
3128 const unsigned NumVRs = array_lengthof(VR);
3130 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
3132 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3133 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3135 SmallVector<SDValue, 8> MemOpChains;
3136 for (unsigned i = 0; i != NumOps; ++i) {
3137 SDValue Arg = OutVals[i];
3138 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3140 // PtrOff will be used to store the current argument to the stack if a
3141 // register cannot be found for it.
3144 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3146 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3148 // On PPC64, promote integers to 64-bit values.
3149 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3150 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3151 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3152 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3155 // FIXME memcpy is used way more than necessary. Correctness first.
3156 if (Flags.isByVal()) {
3157 unsigned Size = Flags.getByValSize();
3158 if (Size==1 || Size==2) {
3159 // Very small objects are passed right-justified.
3160 // Everything else is passed left-justified.
3161 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3162 if (GPR_idx != NumGPRs) {
3163 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3164 MachinePointerInfo(), VT,
3166 MemOpChains.push_back(Load.getValue(1));
3167 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3169 ArgOffset += PtrByteSize;
3171 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3172 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3173 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3174 CallSeqStart.getNode()->getOperand(0),
3176 // This must go outside the CALLSEQ_START..END.
3177 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3178 CallSeqStart.getNode()->getOperand(1));
3179 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3180 NewCallSeqStart.getNode());
3181 Chain = CallSeqStart = NewCallSeqStart;
3182 ArgOffset += PtrByteSize;
3186 // Copy entire object into memory. There are cases where gcc-generated
3187 // code assumes it is there, even if it could be put entirely into
3188 // registers. (This is not what the doc says.)
3189 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3190 CallSeqStart.getNode()->getOperand(0),
3192 // This must go outside the CALLSEQ_START..END.
3193 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3194 CallSeqStart.getNode()->getOperand(1));
3195 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3196 Chain = CallSeqStart = NewCallSeqStart;
3197 // And copy the pieces of it that fit into registers.
3198 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3199 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3200 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3201 if (GPR_idx != NumGPRs) {
3202 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3203 MachinePointerInfo(),
3205 MemOpChains.push_back(Load.getValue(1));
3206 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3207 ArgOffset += PtrByteSize;
3209 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3216 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3217 default: llvm_unreachable("Unexpected ValueType for argument!");
3220 if (GPR_idx != NumGPRs) {
3221 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3223 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3224 isPPC64, isTailCall, false, MemOpChains,
3225 TailCallArguments, dl);
3227 ArgOffset += PtrByteSize;
3231 if (FPR_idx != NumFPRs) {
3232 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3235 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3236 MachinePointerInfo(), false, false, 0);
3237 MemOpChains.push_back(Store);
3239 // Float varargs are always shadowed in available integer registers
3240 if (GPR_idx != NumGPRs) {
3241 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3242 MachinePointerInfo(), false, false, 0);
3243 MemOpChains.push_back(Load.getValue(1));
3244 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3246 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3247 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3248 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3249 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3250 MachinePointerInfo(),
3252 MemOpChains.push_back(Load.getValue(1));
3253 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3256 // If we have any FPRs remaining, we may also have GPRs remaining.
3257 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3259 if (GPR_idx != NumGPRs)
3261 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3262 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3266 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3267 isPPC64, isTailCall, false, MemOpChains,
3268 TailCallArguments, dl);
3273 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3280 // These go aligned on the stack, or in the corresponding R registers
3281 // when within range. The Darwin PPC ABI doc claims they also go in
3282 // V registers; in fact gcc does this only for arguments that are
3283 // prototyped, not for those that match the ... We do it for all
3284 // arguments, seems to work.
3285 while (ArgOffset % 16 !=0) {
3286 ArgOffset += PtrByteSize;
3287 if (GPR_idx != NumGPRs)
3290 // We could elide this store in the case where the object fits
3291 // entirely in R registers. Maybe later.
3292 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3293 DAG.getConstant(ArgOffset, PtrVT));
3294 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3295 MachinePointerInfo(), false, false, 0);
3296 MemOpChains.push_back(Store);
3297 if (VR_idx != NumVRs) {
3298 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3299 MachinePointerInfo(),
3301 MemOpChains.push_back(Load.getValue(1));
3302 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3305 for (unsigned i=0; i<16; i+=PtrByteSize) {
3306 if (GPR_idx == NumGPRs)
3308 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3309 DAG.getConstant(i, PtrVT));
3310 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3312 MemOpChains.push_back(Load.getValue(1));
3313 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3318 // Non-varargs Altivec params generally go in registers, but have
3319 // stack space allocated at the end.
3320 if (VR_idx != NumVRs) {
3321 // Doesn't have GPR space allocated.
3322 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3323 } else if (nAltivecParamsAtEnd==0) {
3324 // We are emitting Altivec params in order.
3325 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3326 isPPC64, isTailCall, true, MemOpChains,
3327 TailCallArguments, dl);
3333 // If all Altivec parameters fit in registers, as they usually do,
3334 // they get stack space following the non-Altivec parameters. We
3335 // don't track this here because nobody below needs it.
3336 // If there are more Altivec parameters than fit in registers emit
3338 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3340 // Offset is aligned; skip 1st 12 params which go in V registers.
3341 ArgOffset = ((ArgOffset+15)/16)*16;
3343 for (unsigned i = 0; i != NumOps; ++i) {
3344 SDValue Arg = OutVals[i];
3345 EVT ArgType = Outs[i].VT;
3346 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3347 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3350 // We are emitting Altivec params in order.
3351 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3352 isPPC64, isTailCall, true, MemOpChains,
3353 TailCallArguments, dl);
3360 if (!MemOpChains.empty())
3361 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3362 &MemOpChains[0], MemOpChains.size());
3364 // Check if this is an indirect call (MTCTR/BCTRL).
3365 // See PrepareCall() for more information about calls through function
3366 // pointers in the 64-bit SVR4 ABI.
3367 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3368 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3369 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3370 !isBLACompatibleAddress(Callee, DAG)) {
3371 // Load r2 into a virtual register and store it to the TOC save area.
3372 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3373 // TOC save area offset.
3374 SDValue PtrOff = DAG.getIntPtrConstant(40);
3375 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3376 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3380 // On Darwin, R12 must contain the address of an indirect callee. This does
3381 // not mean the MTCTR instruction must use R12; it's easier to model this as
3382 // an extra parameter, so do that.
3384 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3385 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3386 !isBLACompatibleAddress(Callee, DAG))
3387 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3388 PPC::R12), Callee));
3390 // Build a sequence of copy-to-reg nodes chained together with token chain
3391 // and flag operands which copy the outgoing args into the appropriate regs.
3393 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3394 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3395 RegsToPass[i].second, InFlag);
3396 InFlag = Chain.getValue(1);
3400 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3401 FPOp, true, TailCallArguments);
3403 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3404 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3409 PPCTargetLowering::LowerReturn(SDValue Chain,
3410 CallingConv::ID CallConv, bool isVarArg,
3411 const SmallVectorImpl<ISD::OutputArg> &Outs,
3412 const SmallVectorImpl<SDValue> &OutVals,
3413 DebugLoc dl, SelectionDAG &DAG) const {
3415 SmallVector<CCValAssign, 16> RVLocs;
3416 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3417 getTargetMachine(), RVLocs, *DAG.getContext());
3418 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3420 // If this is the first return lowered for this function, add the regs to the
3421 // liveout set for the function.
3422 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3423 for (unsigned i = 0; i != RVLocs.size(); ++i)
3424 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3429 // Copy the result values into the output registers.
3430 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3431 CCValAssign &VA = RVLocs[i];
3432 assert(VA.isRegLoc() && "Can only return in registers!");
3433 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3435 Flag = Chain.getValue(1);
3439 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3441 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3444 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3445 const PPCSubtarget &Subtarget) const {
3446 // When we pop the dynamic allocation we need to restore the SP link.
3447 DebugLoc dl = Op.getDebugLoc();
3449 // Get the corect type for pointers.
3450 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3452 // Construct the stack pointer operand.
3453 bool isPPC64 = Subtarget.isPPC64();
3454 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3455 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3457 // Get the operands for the STACKRESTORE.
3458 SDValue Chain = Op.getOperand(0);
3459 SDValue SaveSP = Op.getOperand(1);
3461 // Load the old link SP.
3462 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3463 MachinePointerInfo(),
3466 // Restore the stack pointer.
3467 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3469 // Store the old link SP.
3470 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3477 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3478 MachineFunction &MF = DAG.getMachineFunction();
3479 bool isPPC64 = PPCSubTarget.isPPC64();
3480 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3481 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3483 // Get current frame pointer save index. The users of this index will be
3484 // primarily DYNALLOC instructions.
3485 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3486 int RASI = FI->getReturnAddrSaveIndex();
3488 // If the frame pointer save index hasn't been defined yet.
3490 // Find out what the fix offset of the frame pointer save area.
3491 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3492 // Allocate the frame index for frame pointer save area.
3493 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3495 FI->setReturnAddrSaveIndex(RASI);
3497 return DAG.getFrameIndex(RASI, PtrVT);
3501 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3502 MachineFunction &MF = DAG.getMachineFunction();
3503 bool isPPC64 = PPCSubTarget.isPPC64();
3504 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3505 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3507 // Get current frame pointer save index. The users of this index will be
3508 // primarily DYNALLOC instructions.
3509 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3510 int FPSI = FI->getFramePointerSaveIndex();
3512 // If the frame pointer save index hasn't been defined yet.
3514 // Find out what the fix offset of the frame pointer save area.
3515 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3518 // Allocate the frame index for frame pointer save area.
3519 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3521 FI->setFramePointerSaveIndex(FPSI);
3523 return DAG.getFrameIndex(FPSI, PtrVT);
3526 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3528 const PPCSubtarget &Subtarget) const {
3530 SDValue Chain = Op.getOperand(0);
3531 SDValue Size = Op.getOperand(1);
3532 DebugLoc dl = Op.getDebugLoc();
3534 // Get the corect type for pointers.
3535 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3537 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3538 DAG.getConstant(0, PtrVT), Size);
3539 // Construct a node for the frame pointer save index.
3540 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3541 // Build a DYNALLOC node.
3542 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3543 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3544 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3547 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3549 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3550 // Not FP? Not a fsel.
3551 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3552 !Op.getOperand(2).getValueType().isFloatingPoint())
3555 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3557 // Cannot handle SETEQ/SETNE.
3558 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3560 EVT ResVT = Op.getValueType();
3561 EVT CmpVT = Op.getOperand(0).getValueType();
3562 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3563 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3564 DebugLoc dl = Op.getDebugLoc();
3566 // If the RHS of the comparison is a 0.0, we don't need to do the
3567 // subtraction at all.
3568 if (isFloatingPointZero(RHS))
3570 default: break; // SETUO etc aren't handled by fsel.
3573 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3576 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3577 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3578 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3581 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3584 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3585 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3586 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3587 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3592 default: break; // SETUO etc aren't handled by fsel.
3595 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3596 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3597 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3598 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3601 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3602 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3603 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3604 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3607 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3608 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3609 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3610 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3613 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3614 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3615 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3616 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3621 // FIXME: Split this code up when LegalizeDAGTypes lands.
3622 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3623 DebugLoc dl) const {
3624 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3625 SDValue Src = Op.getOperand(0);
3626 if (Src.getValueType() == MVT::f32)
3627 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3630 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3631 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3633 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3638 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3642 // Convert the FP value to an int value through memory.
3643 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3645 // Emit a store to the stack slot.
3646 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3647 MachinePointerInfo(), false, false, 0);
3649 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3651 if (Op.getValueType() == MVT::i32)
3652 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3653 DAG.getConstant(4, FIPtr.getValueType()));
3654 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3658 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3659 SelectionDAG &DAG) const {
3660 DebugLoc dl = Op.getDebugLoc();
3661 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3662 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3665 if (Op.getOperand(0).getValueType() == MVT::i64) {
3666 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3667 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3668 if (Op.getValueType() == MVT::f32)
3669 FP = DAG.getNode(ISD::FP_ROUND, dl,
3670 MVT::f32, FP, DAG.getIntPtrConstant(0));
3674 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3675 "Unhandled SINT_TO_FP type in custom expander!");
3676 // Since we only generate this in 64-bit mode, we can take advantage of
3677 // 64-bit registers. In particular, sign extend the input value into the
3678 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3679 // then lfd it and fcfid it.
3680 MachineFunction &MF = DAG.getMachineFunction();
3681 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3682 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3683 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3684 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3686 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3689 // STD the extended value into the stack slot.
3690 MachineMemOperand *MMO =
3691 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3692 MachineMemOperand::MOStore, 8, 8);
3693 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3695 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3696 Ops, 4, MVT::i64, MMO);
3697 // Load the value as a double.
3698 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3701 // FCFID it and return it.
3702 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3703 if (Op.getValueType() == MVT::f32)
3704 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3708 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3709 SelectionDAG &DAG) const {
3710 DebugLoc dl = Op.getDebugLoc();
3712 The rounding mode is in bits 30:31 of FPSR, and has the following
3719 FLT_ROUNDS, on the other hand, expects the following:
3726 To perform the conversion, we do:
3727 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3730 MachineFunction &MF = DAG.getMachineFunction();
3731 EVT VT = Op.getValueType();
3732 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3733 std::vector<EVT> NodeTys;
3734 SDValue MFFSreg, InFlag;
3736 // Save FP Control Word to register
3737 NodeTys.push_back(MVT::f64); // return register
3738 NodeTys.push_back(MVT::Glue); // unused in this context
3739 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3741 // Save FP register to stack slot
3742 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3743 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3744 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3745 StackSlot, MachinePointerInfo(), false, false,0);
3747 // Load FP Control Word from low 32 bits of stack slot.
3748 SDValue Four = DAG.getConstant(4, PtrVT);
3749 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3750 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3753 // Transform as necessary
3755 DAG.getNode(ISD::AND, dl, MVT::i32,
3756 CWD, DAG.getConstant(3, MVT::i32));
3758 DAG.getNode(ISD::SRL, dl, MVT::i32,
3759 DAG.getNode(ISD::AND, dl, MVT::i32,
3760 DAG.getNode(ISD::XOR, dl, MVT::i32,
3761 CWD, DAG.getConstant(3, MVT::i32)),
3762 DAG.getConstant(3, MVT::i32)),
3763 DAG.getConstant(1, MVT::i32));
3766 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3768 return DAG.getNode((VT.getSizeInBits() < 16 ?
3769 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3772 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3773 EVT VT = Op.getValueType();
3774 unsigned BitWidth = VT.getSizeInBits();
3775 DebugLoc dl = Op.getDebugLoc();
3776 assert(Op.getNumOperands() == 3 &&
3777 VT == Op.getOperand(1).getValueType() &&
3780 // Expand into a bunch of logical ops. Note that these ops
3781 // depend on the PPC behavior for oversized shift amounts.
3782 SDValue Lo = Op.getOperand(0);
3783 SDValue Hi = Op.getOperand(1);
3784 SDValue Amt = Op.getOperand(2);
3785 EVT AmtVT = Amt.getValueType();
3787 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3788 DAG.getConstant(BitWidth, AmtVT), Amt);
3789 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3790 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3791 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3792 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3793 DAG.getConstant(-BitWidth, AmtVT));
3794 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3795 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3796 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3797 SDValue OutOps[] = { OutLo, OutHi };
3798 return DAG.getMergeValues(OutOps, 2, dl);
3801 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3802 EVT VT = Op.getValueType();
3803 DebugLoc dl = Op.getDebugLoc();
3804 unsigned BitWidth = VT.getSizeInBits();
3805 assert(Op.getNumOperands() == 3 &&
3806 VT == Op.getOperand(1).getValueType() &&
3809 // Expand into a bunch of logical ops. Note that these ops
3810 // depend on the PPC behavior for oversized shift amounts.
3811 SDValue Lo = Op.getOperand(0);
3812 SDValue Hi = Op.getOperand(1);
3813 SDValue Amt = Op.getOperand(2);
3814 EVT AmtVT = Amt.getValueType();
3816 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3817 DAG.getConstant(BitWidth, AmtVT), Amt);
3818 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3819 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3820 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3821 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3822 DAG.getConstant(-BitWidth, AmtVT));
3823 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3824 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3825 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3826 SDValue OutOps[] = { OutLo, OutHi };
3827 return DAG.getMergeValues(OutOps, 2, dl);
3830 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3831 DebugLoc dl = Op.getDebugLoc();
3832 EVT VT = Op.getValueType();
3833 unsigned BitWidth = VT.getSizeInBits();
3834 assert(Op.getNumOperands() == 3 &&
3835 VT == Op.getOperand(1).getValueType() &&
3838 // Expand into a bunch of logical ops, followed by a select_cc.
3839 SDValue Lo = Op.getOperand(0);
3840 SDValue Hi = Op.getOperand(1);
3841 SDValue Amt = Op.getOperand(2);
3842 EVT AmtVT = Amt.getValueType();
3844 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3845 DAG.getConstant(BitWidth, AmtVT), Amt);
3846 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3847 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3848 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3849 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3850 DAG.getConstant(-BitWidth, AmtVT));
3851 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3852 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3853 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3854 Tmp4, Tmp6, ISD::SETLE);
3855 SDValue OutOps[] = { OutLo, OutHi };
3856 return DAG.getMergeValues(OutOps, 2, dl);
3859 //===----------------------------------------------------------------------===//
3860 // Vector related lowering.
3863 /// BuildSplatI - Build a canonical splati of Val with an element size of
3864 /// SplatSize. Cast the result to VT.
3865 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3866 SelectionDAG &DAG, DebugLoc dl) {
3867 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3869 static const EVT VTys[] = { // canonical VT to use for each size.
3870 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3873 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3875 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3879 EVT CanonicalVT = VTys[SplatSize-1];
3881 // Build a canonical splat for this value.
3882 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3883 SmallVector<SDValue, 8> Ops;
3884 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3885 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3886 &Ops[0], Ops.size());
3887 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3890 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3891 /// specified intrinsic ID.
3892 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3893 SelectionDAG &DAG, DebugLoc dl,
3894 EVT DestVT = MVT::Other) {
3895 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3896 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3897 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3900 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3901 /// specified intrinsic ID.
3902 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3903 SDValue Op2, SelectionDAG &DAG,
3904 DebugLoc dl, EVT DestVT = MVT::Other) {
3905 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3906 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3907 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3911 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3912 /// amount. The result has the specified value type.
3913 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3914 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3915 // Force LHS/RHS to be the right type.
3916 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3917 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3920 for (unsigned i = 0; i != 16; ++i)
3922 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3923 return DAG.getNode(ISD::BITCAST, dl, VT, T);
3926 // If this is a case we can't handle, return null and let the default
3927 // expansion code take care of it. If we CAN select this case, and if it
3928 // selects to a single instruction, return Op. Otherwise, if we can codegen
3929 // this case more efficiently than a constant pool load, lower it to the
3930 // sequence of ops that should be used.
3931 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3932 SelectionDAG &DAG) const {
3933 DebugLoc dl = Op.getDebugLoc();
3934 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3935 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3937 // Check if this is a splat of a constant value.
3938 APInt APSplatBits, APSplatUndef;
3939 unsigned SplatBitSize;
3941 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3942 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3945 unsigned SplatBits = APSplatBits.getZExtValue();
3946 unsigned SplatUndef = APSplatUndef.getZExtValue();
3947 unsigned SplatSize = SplatBitSize / 8;
3949 // First, handle single instruction cases.
3952 if (SplatBits == 0) {
3953 // Canonicalize all zero vectors to be v4i32.
3954 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3955 SDValue Z = DAG.getConstant(0, MVT::i32);
3956 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3957 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
3962 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3963 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3965 if (SextVal >= -16 && SextVal <= 15)
3966 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3969 // Two instruction sequences.
3971 // If this value is in the range [-32,30] and is even, use:
3972 // tmp = VSPLTI[bhw], result = add tmp, tmp
3973 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3974 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3975 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3976 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3979 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3980 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3982 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3983 // Make -1 and vspltisw -1:
3984 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3986 // Make the VSLW intrinsic, computing 0x8000_0000.
3987 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3990 // xor by OnesV to invert it.
3991 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3992 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3995 // Check to see if this is a wide variety of vsplti*, binop self cases.
3996 static const signed char SplatCsts[] = {
3997 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3998 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4001 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4002 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4003 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4004 int i = SplatCsts[idx];
4006 // Figure out what shift amount will be used by altivec if shifted by i in
4008 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4010 // vsplti + shl self.
4011 if (SextVal == (i << (int)TypeShiftAmt)) {
4012 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4013 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4014 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4015 Intrinsic::ppc_altivec_vslw
4017 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4018 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4021 // vsplti + srl self.
4022 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4023 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4024 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4025 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4026 Intrinsic::ppc_altivec_vsrw
4028 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4029 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4032 // vsplti + sra self.
4033 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4034 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4035 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4036 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4037 Intrinsic::ppc_altivec_vsraw
4039 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4040 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4043 // vsplti + rol self.
4044 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4045 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4046 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4047 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4048 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4049 Intrinsic::ppc_altivec_vrlw
4051 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4052 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4055 // t = vsplti c, result = vsldoi t, t, 1
4056 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4057 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4058 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4060 // t = vsplti c, result = vsldoi t, t, 2
4061 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4062 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4063 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4065 // t = vsplti c, result = vsldoi t, t, 3
4066 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4067 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4068 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4072 // Three instruction sequences.
4074 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4075 if (SextVal >= 0 && SextVal <= 31) {
4076 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4077 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4078 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4079 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4081 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4082 if (SextVal >= -31 && SextVal <= 0) {
4083 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4084 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4085 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4086 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4092 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4093 /// the specified operations to build the shuffle.
4094 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4095 SDValue RHS, SelectionDAG &DAG,
4097 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4098 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4099 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4102 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4114 if (OpNum == OP_COPY) {
4115 if (LHSID == (1*9+2)*9+3) return LHS;
4116 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4120 SDValue OpLHS, OpRHS;
4121 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4122 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4126 default: llvm_unreachable("Unknown i32 permute!");
4128 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4129 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4130 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4131 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4134 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4135 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4136 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4137 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4140 for (unsigned i = 0; i != 16; ++i)
4141 ShufIdxs[i] = (i&3)+0;
4144 for (unsigned i = 0; i != 16; ++i)
4145 ShufIdxs[i] = (i&3)+4;
4148 for (unsigned i = 0; i != 16; ++i)
4149 ShufIdxs[i] = (i&3)+8;
4152 for (unsigned i = 0; i != 16; ++i)
4153 ShufIdxs[i] = (i&3)+12;
4156 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4158 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4160 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4162 EVT VT = OpLHS.getValueType();
4163 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4164 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4165 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4166 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4169 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4170 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4171 /// return the code it can be lowered into. Worst case, it can always be
4172 /// lowered into a vperm.
4173 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4174 SelectionDAG &DAG) const {
4175 DebugLoc dl = Op.getDebugLoc();
4176 SDValue V1 = Op.getOperand(0);
4177 SDValue V2 = Op.getOperand(1);
4178 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4179 EVT VT = Op.getValueType();
4181 // Cases that are handled by instructions that take permute immediates
4182 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4183 // selected by the instruction selector.
4184 if (V2.getOpcode() == ISD::UNDEF) {
4185 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4186 PPC::isSplatShuffleMask(SVOp, 2) ||
4187 PPC::isSplatShuffleMask(SVOp, 4) ||
4188 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4189 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4190 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4191 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4192 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4193 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4194 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4195 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4196 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4201 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4202 // and produce a fixed permutation. If any of these match, do not lower to
4204 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4205 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4206 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4207 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4208 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4209 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4210 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4211 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4212 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4215 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4216 // perfect shuffle table to emit an optimal matching sequence.
4217 SmallVector<int, 16> PermMask;
4218 SVOp->getMask(PermMask);
4220 unsigned PFIndexes[4];
4221 bool isFourElementShuffle = true;
4222 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4223 unsigned EltNo = 8; // Start out undef.
4224 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4225 if (PermMask[i*4+j] < 0)
4226 continue; // Undef, ignore it.
4228 unsigned ByteSource = PermMask[i*4+j];
4229 if ((ByteSource & 3) != j) {
4230 isFourElementShuffle = false;
4235 EltNo = ByteSource/4;
4236 } else if (EltNo != ByteSource/4) {
4237 isFourElementShuffle = false;
4241 PFIndexes[i] = EltNo;
4244 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4245 // perfect shuffle vector to determine if it is cost effective to do this as
4246 // discrete instructions, or whether we should use a vperm.
4247 if (isFourElementShuffle) {
4248 // Compute the index in the perfect shuffle table.
4249 unsigned PFTableIndex =
4250 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4252 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4253 unsigned Cost = (PFEntry >> 30);
4255 // Determining when to avoid vperm is tricky. Many things affect the cost
4256 // of vperm, particularly how many times the perm mask needs to be computed.
4257 // For example, if the perm mask can be hoisted out of a loop or is already
4258 // used (perhaps because there are multiple permutes with the same shuffle
4259 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4260 // the loop requires an extra register.
4262 // As a compromise, we only emit discrete instructions if the shuffle can be
4263 // generated in 3 or fewer operations. When we have loop information
4264 // available, if this block is within a loop, we should avoid using vperm
4265 // for 3-operation perms and use a constant pool load instead.
4267 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4270 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4271 // vector that will get spilled to the constant pool.
4272 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4274 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4275 // that it is in input element units, not in bytes. Convert now.
4276 EVT EltVT = V1.getValueType().getVectorElementType();
4277 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4279 SmallVector<SDValue, 16> ResultMask;
4280 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4281 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4283 for (unsigned j = 0; j != BytesPerElement; ++j)
4284 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4288 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4289 &ResultMask[0], ResultMask.size());
4290 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4293 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4294 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4295 /// information about the intrinsic.
4296 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4298 unsigned IntrinsicID =
4299 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4302 switch (IntrinsicID) {
4303 default: return false;
4304 // Comparison predicates.
4305 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4306 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4307 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4308 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4309 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4310 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4311 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4312 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4313 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4314 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4315 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4316 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4317 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4319 // Normal Comparisons.
4320 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4321 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4322 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4323 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4324 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4325 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4326 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4327 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4328 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4329 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4330 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4331 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4332 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4337 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4338 /// lower, do it, otherwise return null.
4339 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4340 SelectionDAG &DAG) const {
4341 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4342 // opcode number of the comparison.
4343 DebugLoc dl = Op.getDebugLoc();
4346 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4347 return SDValue(); // Don't custom lower most intrinsics.
4349 // If this is a non-dot comparison, make the VCMP node and we are done.
4351 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4352 Op.getOperand(1), Op.getOperand(2),
4353 DAG.getConstant(CompareOpc, MVT::i32));
4354 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4357 // Create the PPCISD altivec 'dot' comparison node.
4359 Op.getOperand(2), // LHS
4360 Op.getOperand(3), // RHS
4361 DAG.getConstant(CompareOpc, MVT::i32)
4363 std::vector<EVT> VTs;
4364 VTs.push_back(Op.getOperand(2).getValueType());
4365 VTs.push_back(MVT::Glue);
4366 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4368 // Now that we have the comparison, emit a copy from the CR to a GPR.
4369 // This is flagged to the above dot comparison.
4370 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4371 DAG.getRegister(PPC::CR6, MVT::i32),
4372 CompNode.getValue(1));
4374 // Unpack the result based on how the target uses it.
4375 unsigned BitNo; // Bit # of CR6.
4376 bool InvertBit; // Invert result?
4377 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4378 default: // Can't happen, don't crash on invalid number though.
4379 case 0: // Return the value of the EQ bit of CR6.
4380 BitNo = 0; InvertBit = false;
4382 case 1: // Return the inverted value of the EQ bit of CR6.
4383 BitNo = 0; InvertBit = true;
4385 case 2: // Return the value of the LT bit of CR6.
4386 BitNo = 2; InvertBit = false;
4388 case 3: // Return the inverted value of the LT bit of CR6.
4389 BitNo = 2; InvertBit = true;
4393 // Shift the bit into the low position.
4394 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4395 DAG.getConstant(8-(3-BitNo), MVT::i32));
4397 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4398 DAG.getConstant(1, MVT::i32));
4400 // If we are supposed to, toggle the bit.
4402 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4403 DAG.getConstant(1, MVT::i32));
4407 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4408 SelectionDAG &DAG) const {
4409 DebugLoc dl = Op.getDebugLoc();
4410 // Create a stack slot that is 16-byte aligned.
4411 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4412 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4413 EVT PtrVT = getPointerTy();
4414 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4416 // Store the input value into Value#0 of the stack slot.
4417 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4418 Op.getOperand(0), FIdx, MachinePointerInfo(),
4421 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4425 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4426 DebugLoc dl = Op.getDebugLoc();
4427 if (Op.getValueType() == MVT::v4i32) {
4428 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4430 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4431 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4433 SDValue RHSSwap = // = vrlw RHS, 16
4434 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4436 // Shrinkify inputs to v8i16.
4437 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4438 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4439 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4441 // Low parts multiplied together, generating 32-bit results (we ignore the
4443 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4444 LHS, RHS, DAG, dl, MVT::v4i32);
4446 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4447 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4448 // Shift the high parts up 16 bits.
4449 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4451 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4452 } else if (Op.getValueType() == MVT::v8i16) {
4453 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4455 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4457 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4458 LHS, RHS, Zero, DAG, dl);
4459 } else if (Op.getValueType() == MVT::v16i8) {
4460 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4462 // Multiply the even 8-bit parts, producing 16-bit sums.
4463 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4464 LHS, RHS, DAG, dl, MVT::v8i16);
4465 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4467 // Multiply the odd 8-bit parts, producing 16-bit sums.
4468 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4469 LHS, RHS, DAG, dl, MVT::v8i16);
4470 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4472 // Merge the results together.
4474 for (unsigned i = 0; i != 8; ++i) {
4476 Ops[i*2+1] = 2*i+1+16;
4478 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4480 llvm_unreachable("Unknown mul to lower!");
4484 /// LowerOperation - Provide custom lowering hooks for some operations.
4486 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4487 switch (Op.getOpcode()) {
4488 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4489 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4490 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4491 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4492 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4493 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4494 case ISD::SETCC: return LowerSETCC(Op, DAG);
4495 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4497 return LowerVASTART(Op, DAG, PPCSubTarget);
4500 return LowerVAARG(Op, DAG, PPCSubTarget);
4502 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4503 case ISD::DYNAMIC_STACKALLOC:
4504 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4506 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4507 case ISD::FP_TO_UINT:
4508 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4510 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4511 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4513 // Lower 64-bit shifts.
4514 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4515 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4516 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4518 // Vector-related lowering.
4519 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4520 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4521 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4522 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4523 case ISD::MUL: return LowerMUL(Op, DAG);
4525 // Frame & Return address.
4526 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4527 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4532 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4533 SmallVectorImpl<SDValue>&Results,
4534 SelectionDAG &DAG) const {
4535 const TargetMachine &TM = getTargetMachine();
4536 DebugLoc dl = N->getDebugLoc();
4537 switch (N->getOpcode()) {
4539 assert(false && "Do not know how to custom type legalize this operation!");
4542 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4543 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4546 EVT VT = N->getValueType(0);
4548 if (VT == MVT::i64) {
4549 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4551 Results.push_back(NewNode);
4552 Results.push_back(NewNode.getValue(1));
4556 case ISD::FP_ROUND_INREG: {
4557 assert(N->getValueType(0) == MVT::ppcf128);
4558 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4559 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4560 MVT::f64, N->getOperand(0),
4561 DAG.getIntPtrConstant(0));
4562 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4563 MVT::f64, N->getOperand(0),
4564 DAG.getIntPtrConstant(1));
4566 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4567 // of the long double, and puts FPSCR back the way it was. We do not
4568 // actually model FPSCR.
4569 std::vector<EVT> NodeTys;
4570 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4572 NodeTys.push_back(MVT::f64); // Return register
4573 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4574 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4575 MFFSreg = Result.getValue(0);
4576 InFlag = Result.getValue(1);
4579 NodeTys.push_back(MVT::Glue); // Returns a flag
4580 Ops[0] = DAG.getConstant(31, MVT::i32);
4582 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4583 InFlag = Result.getValue(0);
4586 NodeTys.push_back(MVT::Glue); // Returns a flag
4587 Ops[0] = DAG.getConstant(30, MVT::i32);
4589 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4590 InFlag = Result.getValue(0);
4593 NodeTys.push_back(MVT::f64); // result of add
4594 NodeTys.push_back(MVT::Glue); // Returns a flag
4598 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4599 FPreg = Result.getValue(0);
4600 InFlag = Result.getValue(1);
4603 NodeTys.push_back(MVT::f64);
4604 Ops[0] = DAG.getConstant(1, MVT::i32);
4608 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4609 FPreg = Result.getValue(0);
4611 // We know the low half is about to be thrown away, so just use something
4613 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4617 case ISD::FP_TO_SINT:
4618 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4624 //===----------------------------------------------------------------------===//
4625 // Other Lowering Code
4626 //===----------------------------------------------------------------------===//
4629 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4630 bool is64bit, unsigned BinOpcode) const {
4631 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4632 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4634 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4635 MachineFunction *F = BB->getParent();
4636 MachineFunction::iterator It = BB;
4639 unsigned dest = MI->getOperand(0).getReg();
4640 unsigned ptrA = MI->getOperand(1).getReg();
4641 unsigned ptrB = MI->getOperand(2).getReg();
4642 unsigned incr = MI->getOperand(3).getReg();
4643 DebugLoc dl = MI->getDebugLoc();
4645 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4646 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4647 F->insert(It, loopMBB);
4648 F->insert(It, exitMBB);
4649 exitMBB->splice(exitMBB->begin(), BB,
4650 llvm::next(MachineBasicBlock::iterator(MI)),
4652 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4654 MachineRegisterInfo &RegInfo = F->getRegInfo();
4655 unsigned TmpReg = (!BinOpcode) ? incr :
4656 RegInfo.createVirtualRegister(
4657 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4658 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4662 // fallthrough --> loopMBB
4663 BB->addSuccessor(loopMBB);
4666 // l[wd]arx dest, ptr
4667 // add r0, dest, incr
4668 // st[wd]cx. r0, ptr
4670 // fallthrough --> exitMBB
4672 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4673 .addReg(ptrA).addReg(ptrB);
4675 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4676 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4677 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4678 BuildMI(BB, dl, TII->get(PPC::BCC))
4679 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4680 BB->addSuccessor(loopMBB);
4681 BB->addSuccessor(exitMBB);
4690 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4691 MachineBasicBlock *BB,
4692 bool is8bit, // operation
4693 unsigned BinOpcode) const {
4694 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4695 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4696 // In 64 bit mode we have to use 64 bits for addresses, even though the
4697 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4698 // registers without caring whether they're 32 or 64, but here we're
4699 // doing actual arithmetic on the addresses.
4700 bool is64bit = PPCSubTarget.isPPC64();
4701 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4703 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4704 MachineFunction *F = BB->getParent();
4705 MachineFunction::iterator It = BB;
4708 unsigned dest = MI->getOperand(0).getReg();
4709 unsigned ptrA = MI->getOperand(1).getReg();
4710 unsigned ptrB = MI->getOperand(2).getReg();
4711 unsigned incr = MI->getOperand(3).getReg();
4712 DebugLoc dl = MI->getDebugLoc();
4714 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4715 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4716 F->insert(It, loopMBB);
4717 F->insert(It, exitMBB);
4718 exitMBB->splice(exitMBB->begin(), BB,
4719 llvm::next(MachineBasicBlock::iterator(MI)),
4721 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4723 MachineRegisterInfo &RegInfo = F->getRegInfo();
4724 const TargetRegisterClass *RC =
4725 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4726 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4727 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4728 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4729 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4730 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4731 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4732 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4733 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4734 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4735 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4736 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4737 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4739 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4743 // fallthrough --> loopMBB
4744 BB->addSuccessor(loopMBB);
4746 // The 4-byte load must be aligned, while a char or short may be
4747 // anywhere in the word. Hence all this nasty bookkeeping code.
4748 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4749 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4750 // xori shift, shift1, 24 [16]
4751 // rlwinm ptr, ptr1, 0, 0, 29
4752 // slw incr2, incr, shift
4753 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4754 // slw mask, mask2, shift
4756 // lwarx tmpDest, ptr
4757 // add tmp, tmpDest, incr2
4758 // andc tmp2, tmpDest, mask
4759 // and tmp3, tmp, mask
4760 // or tmp4, tmp3, tmp2
4763 // fallthrough --> exitMBB
4764 // srw dest, tmpDest, shift
4765 if (ptrA != ZeroReg) {
4766 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4767 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4768 .addReg(ptrA).addReg(ptrB);
4772 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4773 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4774 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4775 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4777 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4778 .addReg(Ptr1Reg).addImm(0).addImm(61);
4780 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4781 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4782 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4783 .addReg(incr).addReg(ShiftReg);
4785 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4787 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4788 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4790 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4791 .addReg(Mask2Reg).addReg(ShiftReg);
4794 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4795 .addReg(ZeroReg).addReg(PtrReg);
4797 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4798 .addReg(Incr2Reg).addReg(TmpDestReg);
4799 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4800 .addReg(TmpDestReg).addReg(MaskReg);
4801 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4802 .addReg(TmpReg).addReg(MaskReg);
4803 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4804 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4805 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4806 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4807 BuildMI(BB, dl, TII->get(PPC::BCC))
4808 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4809 BB->addSuccessor(loopMBB);
4810 BB->addSuccessor(exitMBB);
4815 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4821 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4822 MachineBasicBlock *BB) const {
4823 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4825 // To "insert" these instructions we actually have to insert their
4826 // control-flow patterns.
4827 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4828 MachineFunction::iterator It = BB;
4831 MachineFunction *F = BB->getParent();
4833 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4834 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4835 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4836 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4837 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4839 // The incoming instruction knows the destination vreg to set, the
4840 // condition code register to branch on, the true/false values to
4841 // select between, and a branch opcode to use.
4846 // cmpTY ccX, r1, r2
4848 // fallthrough --> copy0MBB
4849 MachineBasicBlock *thisMBB = BB;
4850 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4851 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4852 unsigned SelectPred = MI->getOperand(4).getImm();
4853 DebugLoc dl = MI->getDebugLoc();
4854 F->insert(It, copy0MBB);
4855 F->insert(It, sinkMBB);
4857 // Transfer the remainder of BB and its successor edges to sinkMBB.
4858 sinkMBB->splice(sinkMBB->begin(), BB,
4859 llvm::next(MachineBasicBlock::iterator(MI)),
4861 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4863 // Next, add the true and fallthrough blocks as its successors.
4864 BB->addSuccessor(copy0MBB);
4865 BB->addSuccessor(sinkMBB);
4867 BuildMI(BB, dl, TII->get(PPC::BCC))
4868 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4871 // %FalseValue = ...
4872 // # fallthrough to sinkMBB
4875 // Update machine-CFG edges
4876 BB->addSuccessor(sinkMBB);
4879 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4882 BuildMI(*BB, BB->begin(), dl,
4883 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4884 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4885 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4887 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4888 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4889 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4890 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4891 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4892 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4893 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4894 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4896 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4897 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4898 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4899 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4900 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4901 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4902 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4903 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4905 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4906 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4907 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4908 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4909 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4910 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4911 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4912 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4914 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4915 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4916 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4917 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4918 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4919 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4920 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4921 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4923 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4924 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4925 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4926 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4927 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4928 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4929 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4930 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4932 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4933 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4934 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4935 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4936 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4937 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4938 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4939 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4941 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4942 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4943 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4944 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4945 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4946 BB = EmitAtomicBinary(MI, BB, false, 0);
4947 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4948 BB = EmitAtomicBinary(MI, BB, true, 0);
4950 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4951 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4952 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4954 unsigned dest = MI->getOperand(0).getReg();
4955 unsigned ptrA = MI->getOperand(1).getReg();
4956 unsigned ptrB = MI->getOperand(2).getReg();
4957 unsigned oldval = MI->getOperand(3).getReg();
4958 unsigned newval = MI->getOperand(4).getReg();
4959 DebugLoc dl = MI->getDebugLoc();
4961 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4962 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4963 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4964 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4965 F->insert(It, loop1MBB);
4966 F->insert(It, loop2MBB);
4967 F->insert(It, midMBB);
4968 F->insert(It, exitMBB);
4969 exitMBB->splice(exitMBB->begin(), BB,
4970 llvm::next(MachineBasicBlock::iterator(MI)),
4972 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4976 // fallthrough --> loopMBB
4977 BB->addSuccessor(loop1MBB);
4980 // l[wd]arx dest, ptr
4981 // cmp[wd] dest, oldval
4984 // st[wd]cx. newval, ptr
4988 // st[wd]cx. dest, ptr
4991 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4992 .addReg(ptrA).addReg(ptrB);
4993 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4994 .addReg(oldval).addReg(dest);
4995 BuildMI(BB, dl, TII->get(PPC::BCC))
4996 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4997 BB->addSuccessor(loop2MBB);
4998 BB->addSuccessor(midMBB);
5001 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5002 .addReg(newval).addReg(ptrA).addReg(ptrB);
5003 BuildMI(BB, dl, TII->get(PPC::BCC))
5004 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5005 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5006 BB->addSuccessor(loop1MBB);
5007 BB->addSuccessor(exitMBB);
5010 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5011 .addReg(dest).addReg(ptrA).addReg(ptrB);
5012 BB->addSuccessor(exitMBB);
5017 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5018 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5019 // We must use 64-bit registers for addresses when targeting 64-bit,
5020 // since we're actually doing arithmetic on them. Other registers
5022 bool is64bit = PPCSubTarget.isPPC64();
5023 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5025 unsigned dest = MI->getOperand(0).getReg();
5026 unsigned ptrA = MI->getOperand(1).getReg();
5027 unsigned ptrB = MI->getOperand(2).getReg();
5028 unsigned oldval = MI->getOperand(3).getReg();
5029 unsigned newval = MI->getOperand(4).getReg();
5030 DebugLoc dl = MI->getDebugLoc();
5032 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5033 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5034 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5035 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5036 F->insert(It, loop1MBB);
5037 F->insert(It, loop2MBB);
5038 F->insert(It, midMBB);
5039 F->insert(It, exitMBB);
5040 exitMBB->splice(exitMBB->begin(), BB,
5041 llvm::next(MachineBasicBlock::iterator(MI)),
5043 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5045 MachineRegisterInfo &RegInfo = F->getRegInfo();
5046 const TargetRegisterClass *RC =
5047 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5048 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5049 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5050 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5051 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5052 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5053 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5054 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5055 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5056 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5057 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5058 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5059 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5060 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5061 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5063 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5064 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5067 // fallthrough --> loopMBB
5068 BB->addSuccessor(loop1MBB);
5070 // The 4-byte load must be aligned, while a char or short may be
5071 // anywhere in the word. Hence all this nasty bookkeeping code.
5072 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5073 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5074 // xori shift, shift1, 24 [16]
5075 // rlwinm ptr, ptr1, 0, 0, 29
5076 // slw newval2, newval, shift
5077 // slw oldval2, oldval,shift
5078 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5079 // slw mask, mask2, shift
5080 // and newval3, newval2, mask
5081 // and oldval3, oldval2, mask
5083 // lwarx tmpDest, ptr
5084 // and tmp, tmpDest, mask
5085 // cmpw tmp, oldval3
5088 // andc tmp2, tmpDest, mask
5089 // or tmp4, tmp2, newval3
5094 // stwcx. tmpDest, ptr
5096 // srw dest, tmpDest, shift
5097 if (ptrA != ZeroReg) {
5098 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5099 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5100 .addReg(ptrA).addReg(ptrB);
5104 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5105 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5106 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5107 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5109 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5110 .addReg(Ptr1Reg).addImm(0).addImm(61);
5112 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5113 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5114 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5115 .addReg(newval).addReg(ShiftReg);
5116 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5117 .addReg(oldval).addReg(ShiftReg);
5119 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5121 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5122 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5123 .addReg(Mask3Reg).addImm(65535);
5125 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5126 .addReg(Mask2Reg).addReg(ShiftReg);
5127 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5128 .addReg(NewVal2Reg).addReg(MaskReg);
5129 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5130 .addReg(OldVal2Reg).addReg(MaskReg);
5133 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5134 .addReg(ZeroReg).addReg(PtrReg);
5135 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5136 .addReg(TmpDestReg).addReg(MaskReg);
5137 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5138 .addReg(TmpReg).addReg(OldVal3Reg);
5139 BuildMI(BB, dl, TII->get(PPC::BCC))
5140 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5141 BB->addSuccessor(loop2MBB);
5142 BB->addSuccessor(midMBB);
5145 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5146 .addReg(TmpDestReg).addReg(MaskReg);
5147 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5148 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5149 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5150 .addReg(ZeroReg).addReg(PtrReg);
5151 BuildMI(BB, dl, TII->get(PPC::BCC))
5152 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5153 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5154 BB->addSuccessor(loop1MBB);
5155 BB->addSuccessor(exitMBB);
5158 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5159 .addReg(ZeroReg).addReg(PtrReg);
5160 BB->addSuccessor(exitMBB);
5165 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5168 llvm_unreachable("Unexpected instr type to insert");
5171 MI->eraseFromParent(); // The pseudo instruction is gone now.
5175 //===----------------------------------------------------------------------===//
5176 // Target Optimization Hooks
5177 //===----------------------------------------------------------------------===//
5179 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5180 DAGCombinerInfo &DCI) const {
5181 const TargetMachine &TM = getTargetMachine();
5182 SelectionDAG &DAG = DCI.DAG;
5183 DebugLoc dl = N->getDebugLoc();
5184 switch (N->getOpcode()) {
5187 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5188 if (C->isNullValue()) // 0 << V -> 0.
5189 return N->getOperand(0);
5193 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5194 if (C->isNullValue()) // 0 >>u V -> 0.
5195 return N->getOperand(0);
5199 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5200 if (C->isNullValue() || // 0 >>s V -> 0.
5201 C->isAllOnesValue()) // -1 >>s V -> -1.
5202 return N->getOperand(0);
5206 case ISD::SINT_TO_FP:
5207 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5208 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5209 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5210 // We allow the src/dst to be either f32/f64, but the intermediate
5211 // type must be i64.
5212 if (N->getOperand(0).getValueType() == MVT::i64 &&
5213 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5214 SDValue Val = N->getOperand(0).getOperand(0);
5215 if (Val.getValueType() == MVT::f32) {
5216 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5217 DCI.AddToWorklist(Val.getNode());
5220 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5221 DCI.AddToWorklist(Val.getNode());
5222 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5223 DCI.AddToWorklist(Val.getNode());
5224 if (N->getValueType(0) == MVT::f32) {
5225 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5226 DAG.getIntPtrConstant(0));
5227 DCI.AddToWorklist(Val.getNode());
5230 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5231 // If the intermediate type is i32, we can avoid the load/store here
5238 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5239 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5240 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5241 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5242 N->getOperand(1).getValueType() == MVT::i32 &&
5243 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5244 SDValue Val = N->getOperand(1).getOperand(0);
5245 if (Val.getValueType() == MVT::f32) {
5246 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5247 DCI.AddToWorklist(Val.getNode());
5249 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5250 DCI.AddToWorklist(Val.getNode());
5252 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5253 N->getOperand(2), N->getOperand(3));
5254 DCI.AddToWorklist(Val.getNode());
5258 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5259 if (cast<StoreSDNode>(N)->isUnindexed() &&
5260 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5261 N->getOperand(1).getNode()->hasOneUse() &&
5262 (N->getOperand(1).getValueType() == MVT::i32 ||
5263 N->getOperand(1).getValueType() == MVT::i16)) {
5264 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5265 // Do an any-extend to 32-bits if this is a half-word input.
5266 if (BSwapOp.getValueType() == MVT::i16)
5267 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5270 N->getOperand(0), BSwapOp, N->getOperand(2),
5271 DAG.getValueType(N->getOperand(1).getValueType())
5274 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5275 Ops, array_lengthof(Ops),
5276 cast<StoreSDNode>(N)->getMemoryVT(),
5277 cast<StoreSDNode>(N)->getMemOperand());
5281 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5282 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5283 N->getOperand(0).hasOneUse() &&
5284 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5285 SDValue Load = N->getOperand(0);
5286 LoadSDNode *LD = cast<LoadSDNode>(Load);
5287 // Create the byte-swapping load.
5289 LD->getChain(), // Chain
5290 LD->getBasePtr(), // Ptr
5291 DAG.getValueType(N->getValueType(0)) // VT
5294 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5295 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5296 LD->getMemoryVT(), LD->getMemOperand());
5298 // If this is an i16 load, insert the truncate.
5299 SDValue ResVal = BSLoad;
5300 if (N->getValueType(0) == MVT::i16)
5301 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5303 // First, combine the bswap away. This makes the value produced by the
5305 DCI.CombineTo(N, ResVal);
5307 // Next, combine the load away, we give it a bogus result value but a real
5308 // chain result. The result value is dead because the bswap is dead.
5309 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5311 // Return N so it doesn't get rechecked!
5312 return SDValue(N, 0);
5316 case PPCISD::VCMP: {
5317 // If a VCMPo node already exists with exactly the same operands as this
5318 // node, use its result instead of this node (VCMPo computes both a CR6 and
5319 // a normal output).
5321 if (!N->getOperand(0).hasOneUse() &&
5322 !N->getOperand(1).hasOneUse() &&
5323 !N->getOperand(2).hasOneUse()) {
5325 // Scan all of the users of the LHS, looking for VCMPo's that match.
5326 SDNode *VCMPoNode = 0;
5328 SDNode *LHSN = N->getOperand(0).getNode();
5329 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5331 if (UI->getOpcode() == PPCISD::VCMPo &&
5332 UI->getOperand(1) == N->getOperand(1) &&
5333 UI->getOperand(2) == N->getOperand(2) &&
5334 UI->getOperand(0) == N->getOperand(0)) {
5339 // If there is no VCMPo node, or if the flag value has a single use, don't
5341 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5344 // Look at the (necessarily single) use of the flag value. If it has a
5345 // chain, this transformation is more complex. Note that multiple things
5346 // could use the value result, which we should ignore.
5347 SDNode *FlagUser = 0;
5348 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5349 FlagUser == 0; ++UI) {
5350 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5352 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5353 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5360 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5361 // give up for right now.
5362 if (FlagUser->getOpcode() == PPCISD::MFCR)
5363 return SDValue(VCMPoNode, 0);
5368 // If this is a branch on an altivec predicate comparison, lower this so
5369 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5370 // lowering is done pre-legalize, because the legalizer lowers the predicate
5371 // compare down to code that is difficult to reassemble.
5372 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5373 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5377 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5378 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5379 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5380 assert(isDot && "Can't compare against a vector result!");
5382 // If this is a comparison against something other than 0/1, then we know
5383 // that the condition is never/always true.
5384 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5385 if (Val != 0 && Val != 1) {
5386 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5387 return N->getOperand(0);
5388 // Always !=, turn it into an unconditional branch.
5389 return DAG.getNode(ISD::BR, dl, MVT::Other,
5390 N->getOperand(0), N->getOperand(4));
5393 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5395 // Create the PPCISD altivec 'dot' comparison node.
5396 std::vector<EVT> VTs;
5398 LHS.getOperand(2), // LHS of compare
5399 LHS.getOperand(3), // RHS of compare
5400 DAG.getConstant(CompareOpc, MVT::i32)
5402 VTs.push_back(LHS.getOperand(2).getValueType());
5403 VTs.push_back(MVT::Glue);
5404 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5406 // Unpack the result based on how the target uses it.
5407 PPC::Predicate CompOpc;
5408 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5409 default: // Can't happen, don't crash on invalid number though.
5410 case 0: // Branch on the value of the EQ bit of CR6.
5411 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5413 case 1: // Branch on the inverted value of the EQ bit of CR6.
5414 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5416 case 2: // Branch on the value of the LT bit of CR6.
5417 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5419 case 3: // Branch on the inverted value of the LT bit of CR6.
5420 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5424 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5425 DAG.getConstant(CompOpc, MVT::i32),
5426 DAG.getRegister(PPC::CR6, MVT::i32),
5427 N->getOperand(4), CompNode.getValue(1));
5436 //===----------------------------------------------------------------------===//
5437 // Inline Assembly Support
5438 //===----------------------------------------------------------------------===//
5440 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5444 const SelectionDAG &DAG,
5445 unsigned Depth) const {
5446 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5447 switch (Op.getOpcode()) {
5449 case PPCISD::LBRX: {
5450 // lhbrx is known to have the top bits cleared out.
5451 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5452 KnownZero = 0xFFFF0000;
5455 case ISD::INTRINSIC_WO_CHAIN: {
5456 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5458 case Intrinsic::ppc_altivec_vcmpbfp_p:
5459 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5460 case Intrinsic::ppc_altivec_vcmpequb_p:
5461 case Intrinsic::ppc_altivec_vcmpequh_p:
5462 case Intrinsic::ppc_altivec_vcmpequw_p:
5463 case Intrinsic::ppc_altivec_vcmpgefp_p:
5464 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5465 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5466 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5467 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5468 case Intrinsic::ppc_altivec_vcmpgtub_p:
5469 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5470 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5471 KnownZero = ~1U; // All bits but the low one are known to be zero.
5479 /// getConstraintType - Given a constraint, return the type of
5480 /// constraint it is for this target.
5481 PPCTargetLowering::ConstraintType
5482 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5483 if (Constraint.size() == 1) {
5484 switch (Constraint[0]) {
5491 return C_RegisterClass;
5494 return TargetLowering::getConstraintType(Constraint);
5497 /// Examine constraint type and operand type and determine a weight value.
5498 /// This object must already have been set up with the operand type
5499 /// and the current alternative constraint selected.
5500 TargetLowering::ConstraintWeight
5501 PPCTargetLowering::getSingleConstraintMatchWeight(
5502 AsmOperandInfo &info, const char *constraint) const {
5503 ConstraintWeight weight = CW_Invalid;
5504 Value *CallOperandVal = info.CallOperandVal;
5505 // If we don't have a value, we can't do a match,
5506 // but allow it at the lowest weight.
5507 if (CallOperandVal == NULL)
5509 Type *type = CallOperandVal->getType();
5510 // Look at the constraint type.
5511 switch (*constraint) {
5513 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5516 if (type->isIntegerTy())
5517 weight = CW_Register;
5520 if (type->isFloatTy())
5521 weight = CW_Register;
5524 if (type->isDoubleTy())
5525 weight = CW_Register;
5528 if (type->isVectorTy())
5529 weight = CW_Register;
5532 weight = CW_Register;
5538 std::pair<unsigned, const TargetRegisterClass*>
5539 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5541 if (Constraint.size() == 1) {
5542 // GCC RS6000 Constraint Letters
5543 switch (Constraint[0]) {
5546 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5547 return std::make_pair(0U, PPC::G8RCRegisterClass);
5548 return std::make_pair(0U, PPC::GPRCRegisterClass);
5551 return std::make_pair(0U, PPC::F4RCRegisterClass);
5552 else if (VT == MVT::f64)
5553 return std::make_pair(0U, PPC::F8RCRegisterClass);
5556 return std::make_pair(0U, PPC::VRRCRegisterClass);
5558 return std::make_pair(0U, PPC::CRRCRegisterClass);
5562 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5566 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5567 /// vector. If it is invalid, don't add anything to Ops.
5568 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5569 std::string &Constraint,
5570 std::vector<SDValue>&Ops,
5571 SelectionDAG &DAG) const {
5572 SDValue Result(0,0);
5574 // Only support length 1 constraints.
5575 if (Constraint.length() > 1) return;
5577 char Letter = Constraint[0];
5588 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5589 if (!CST) return; // Must be an immediate to match.
5590 unsigned Value = CST->getZExtValue();
5592 default: llvm_unreachable("Unknown constraint letter!");
5593 case 'I': // "I" is a signed 16-bit constant.
5594 if ((short)Value == (int)Value)
5595 Result = DAG.getTargetConstant(Value, Op.getValueType());
5597 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5598 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5599 if ((short)Value == 0)
5600 Result = DAG.getTargetConstant(Value, Op.getValueType());
5602 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5603 if ((Value >> 16) == 0)
5604 Result = DAG.getTargetConstant(Value, Op.getValueType());
5606 case 'M': // "M" is a constant that is greater than 31.
5608 Result = DAG.getTargetConstant(Value, Op.getValueType());
5610 case 'N': // "N" is a positive constant that is an exact power of two.
5611 if ((int)Value > 0 && isPowerOf2_32(Value))
5612 Result = DAG.getTargetConstant(Value, Op.getValueType());
5614 case 'O': // "O" is the constant zero.
5616 Result = DAG.getTargetConstant(Value, Op.getValueType());
5618 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5619 if ((short)-Value == (int)-Value)
5620 Result = DAG.getTargetConstant(Value, Op.getValueType());
5627 if (Result.getNode()) {
5628 Ops.push_back(Result);
5632 // Handle standard constraint letters.
5633 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5636 // isLegalAddressingMode - Return true if the addressing mode represented
5637 // by AM is legal for this target, for a load/store of the specified type.
5638 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5640 // FIXME: PPC does not allow r+i addressing modes for vectors!
5642 // PPC allows a sign-extended 16-bit immediate field.
5643 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5646 // No global is ever allowed as a base.
5650 // PPC only support r+r,
5652 case 0: // "r+i" or just "i", depending on HasBaseReg.
5655 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5657 // Otherwise we have r+r or r+i.
5660 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5662 // Allow 2*r as r+r.
5665 // No other scales are supported.
5672 /// isLegalAddressImmediate - Return true if the integer value can be used
5673 /// as the offset of the target addressing mode for load / store of the
5675 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5676 // PPC allows a sign-extended 16-bit immediate field.
5677 return (V > -(1 << 16) && V < (1 << 16)-1);
5680 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5684 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5685 SelectionDAG &DAG) const {
5686 MachineFunction &MF = DAG.getMachineFunction();
5687 MachineFrameInfo *MFI = MF.getFrameInfo();
5688 MFI->setReturnAddressIsTaken(true);
5690 DebugLoc dl = Op.getDebugLoc();
5691 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5693 // Make sure the function does not optimize away the store of the RA to
5695 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5696 FuncInfo->setLRStoreRequired();
5697 bool isPPC64 = PPCSubTarget.isPPC64();
5698 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5701 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5704 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5705 isPPC64? MVT::i64 : MVT::i32);
5706 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5707 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5709 MachinePointerInfo(), false, false, 0);
5712 // Just load the return address off the stack.
5713 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5714 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5715 RetAddrFI, MachinePointerInfo(), false, false, 0);
5718 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5719 SelectionDAG &DAG) const {
5720 DebugLoc dl = Op.getDebugLoc();
5721 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5723 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5724 bool isPPC64 = PtrVT == MVT::i64;
5726 MachineFunction &MF = DAG.getMachineFunction();
5727 MachineFrameInfo *MFI = MF.getFrameInfo();
5728 MFI->setFrameAddressIsTaken(true);
5729 bool is31 = (DisableFramePointerElim(MF) || MFI->hasVarSizedObjects()) &&
5730 MFI->getStackSize() &&
5731 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5732 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5733 (is31 ? PPC::R31 : PPC::R1);
5734 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5737 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5738 FrameAddr, MachinePointerInfo(), false, false, 0);
5743 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5744 // The PowerPC target isn't yet aware of offsets.
5748 /// getOptimalMemOpType - Returns the target specific optimal type for load
5749 /// and store operations as a result of memset, memcpy, and memmove
5750 /// lowering. If DstAlign is zero that means it's safe to destination
5751 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5752 /// means there isn't a need to check it against alignment requirement,
5753 /// probably because the source does not need to be loaded. If
5754 /// 'NonScalarIntSafe' is true, that means it's safe to return a
5755 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5756 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5757 /// constant so it does not need to be loaded.
5758 /// It returns EVT::Other if the type should be determined using generic
5759 /// target-independent logic.
5760 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5761 unsigned DstAlign, unsigned SrcAlign,
5762 bool NonScalarIntSafe,
5764 MachineFunction &MF) const {
5765 if (this->PPCSubTarget.isPPC64()) {