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::INIT_TRAMPOLINE, MVT::Other, Custom);
215 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
217 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
218 setOperationAction(ISD::VASTART , MVT::Other, Custom);
220 // VAARG is custom lowered with the 32-bit SVR4 ABI.
221 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
222 && !TM.getSubtarget<PPCSubtarget>().isPPC64()) {
223 setOperationAction(ISD::VAARG, MVT::Other, Custom);
224 setOperationAction(ISD::VAARG, MVT::i64, Custom);
226 setOperationAction(ISD::VAARG, MVT::Other, Expand);
228 // Use the default implementation.
229 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
230 setOperationAction(ISD::VAEND , MVT::Other, Expand);
231 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
232 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
233 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
234 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
236 // We want to custom lower some of our intrinsics.
237 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
239 // Comparisons that require checking two conditions.
240 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
241 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
242 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
243 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
244 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
245 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
246 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
247 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
248 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
249 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
250 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
251 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
253 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
254 // They also have instructions for converting between i64 and fp.
255 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
256 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
257 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
258 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
259 // This is just the low 32 bits of a (signed) fp->i64 conversion.
260 // We cannot do this with Promote because i64 is not a legal type.
261 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
263 // FIXME: disable this lowered code. This generates 64-bit register values,
264 // and we don't model the fact that the top part is clobbered by calls. We
265 // need to flag these together so that the value isn't live across a call.
266 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
268 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
269 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
272 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
273 // 64-bit PowerPC implementations can support i64 types directly
274 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
275 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
276 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
277 // 64-bit PowerPC wants to expand i128 shifts itself.
278 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
279 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
280 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
282 // 32-bit PowerPC wants to expand i64 shifts itself.
283 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
284 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
285 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
288 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
289 // First set operation action for all vector types to expand. Then we
290 // will selectively turn on ones that can be effectively codegen'd.
291 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
292 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
293 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
295 // add/sub are legal for all supported vector VT's.
296 setOperationAction(ISD::ADD , VT, Legal);
297 setOperationAction(ISD::SUB , VT, Legal);
299 // We promote all shuffles to v16i8.
300 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
301 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
303 // We promote all non-typed operations to v4i32.
304 setOperationAction(ISD::AND , VT, Promote);
305 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
306 setOperationAction(ISD::OR , VT, Promote);
307 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
308 setOperationAction(ISD::XOR , VT, Promote);
309 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
310 setOperationAction(ISD::LOAD , VT, Promote);
311 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
312 setOperationAction(ISD::SELECT, VT, Promote);
313 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
314 setOperationAction(ISD::STORE, VT, Promote);
315 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
317 // No other operations are legal.
318 setOperationAction(ISD::MUL , VT, Expand);
319 setOperationAction(ISD::SDIV, VT, Expand);
320 setOperationAction(ISD::SREM, VT, Expand);
321 setOperationAction(ISD::UDIV, VT, Expand);
322 setOperationAction(ISD::UREM, VT, Expand);
323 setOperationAction(ISD::FDIV, VT, Expand);
324 setOperationAction(ISD::FNEG, VT, Expand);
325 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
326 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
327 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
328 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
329 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
330 setOperationAction(ISD::UDIVREM, VT, Expand);
331 setOperationAction(ISD::SDIVREM, VT, Expand);
332 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
333 setOperationAction(ISD::FPOW, VT, Expand);
334 setOperationAction(ISD::CTPOP, VT, Expand);
335 setOperationAction(ISD::CTLZ, VT, Expand);
336 setOperationAction(ISD::CTTZ, VT, Expand);
339 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
340 // with merges, splats, etc.
341 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
343 setOperationAction(ISD::AND , MVT::v4i32, Legal);
344 setOperationAction(ISD::OR , MVT::v4i32, Legal);
345 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
346 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
347 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
348 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
350 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
351 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
352 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
353 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
355 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
356 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
357 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
358 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
360 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
361 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
363 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
364 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
365 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
366 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
370 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
372 setBooleanContents(ZeroOrOneBooleanContent);
374 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
375 setStackPointerRegisterToSaveRestore(PPC::X1);
376 setExceptionPointerRegister(PPC::X3);
377 setExceptionSelectorRegister(PPC::X4);
379 setStackPointerRegisterToSaveRestore(PPC::R1);
380 setExceptionPointerRegister(PPC::R3);
381 setExceptionSelectorRegister(PPC::R4);
384 // We have target-specific dag combine patterns for the following nodes:
385 setTargetDAGCombine(ISD::SINT_TO_FP);
386 setTargetDAGCombine(ISD::STORE);
387 setTargetDAGCombine(ISD::BR_CC);
388 setTargetDAGCombine(ISD::BSWAP);
390 // Darwin long double math library functions have $LDBL128 appended.
391 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
392 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
393 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
394 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
395 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
396 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
397 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
398 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
399 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
400 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
401 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
404 setMinFunctionAlignment(2);
405 if (PPCSubTarget.isDarwin())
406 setPrefFunctionAlignment(4);
408 setInsertFencesForAtomic(true);
410 computeRegisterProperties();
413 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
414 /// function arguments in the caller parameter area.
415 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
416 const TargetMachine &TM = getTargetMachine();
417 // Darwin passes everything on 4 byte boundary.
418 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
424 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
427 case PPCISD::FSEL: return "PPCISD::FSEL";
428 case PPCISD::FCFID: return "PPCISD::FCFID";
429 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
430 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
431 case PPCISD::STFIWX: return "PPCISD::STFIWX";
432 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
433 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
434 case PPCISD::VPERM: return "PPCISD::VPERM";
435 case PPCISD::Hi: return "PPCISD::Hi";
436 case PPCISD::Lo: return "PPCISD::Lo";
437 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
438 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
439 case PPCISD::LOAD: return "PPCISD::LOAD";
440 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
441 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
442 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
443 case PPCISD::SRL: return "PPCISD::SRL";
444 case PPCISD::SRA: return "PPCISD::SRA";
445 case PPCISD::SHL: return "PPCISD::SHL";
446 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
447 case PPCISD::STD_32: return "PPCISD::STD_32";
448 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
449 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
450 case PPCISD::NOP: return "PPCISD::NOP";
451 case PPCISD::MTCTR: return "PPCISD::MTCTR";
452 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
453 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
454 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
455 case PPCISD::MFCR: return "PPCISD::MFCR";
456 case PPCISD::VCMP: return "PPCISD::VCMP";
457 case PPCISD::VCMPo: return "PPCISD::VCMPo";
458 case PPCISD::LBRX: return "PPCISD::LBRX";
459 case PPCISD::STBRX: return "PPCISD::STBRX";
460 case PPCISD::LARX: return "PPCISD::LARX";
461 case PPCISD::STCX: return "PPCISD::STCX";
462 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
463 case PPCISD::MFFS: return "PPCISD::MFFS";
464 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
465 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
466 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
467 case PPCISD::MTFSF: return "PPCISD::MTFSF";
468 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
472 MVT::SimpleValueType PPCTargetLowering::getSetCCResultType(EVT VT) const {
476 //===----------------------------------------------------------------------===//
477 // Node matching predicates, for use by the tblgen matching code.
478 //===----------------------------------------------------------------------===//
480 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
481 static bool isFloatingPointZero(SDValue Op) {
482 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
483 return CFP->getValueAPF().isZero();
484 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
485 // Maybe this has already been legalized into the constant pool?
486 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
487 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
488 return CFP->getValueAPF().isZero();
493 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
494 /// true if Op is undef or if it matches the specified value.
495 static bool isConstantOrUndef(int Op, int Val) {
496 return Op < 0 || Op == Val;
499 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
500 /// VPKUHUM instruction.
501 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
503 for (unsigned i = 0; i != 16; ++i)
504 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
507 for (unsigned i = 0; i != 8; ++i)
508 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
509 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
515 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
516 /// VPKUWUM instruction.
517 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
519 for (unsigned i = 0; i != 16; i += 2)
520 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
521 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
524 for (unsigned i = 0; i != 8; i += 2)
525 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
526 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
527 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
528 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
534 /// isVMerge - Common function, used to match vmrg* shuffles.
536 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
537 unsigned LHSStart, unsigned RHSStart) {
538 assert(N->getValueType(0) == MVT::v16i8 &&
539 "PPC only supports shuffles by bytes!");
540 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
541 "Unsupported merge size!");
543 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
544 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
545 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
546 LHSStart+j+i*UnitSize) ||
547 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
548 RHSStart+j+i*UnitSize))
554 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
555 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
556 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
559 return isVMerge(N, UnitSize, 8, 24);
560 return isVMerge(N, UnitSize, 8, 8);
563 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
564 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
565 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
568 return isVMerge(N, UnitSize, 0, 16);
569 return isVMerge(N, UnitSize, 0, 0);
573 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
574 /// amount, otherwise return -1.
575 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
576 assert(N->getValueType(0) == MVT::v16i8 &&
577 "PPC only supports shuffles by bytes!");
579 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
581 // Find the first non-undef value in the shuffle mask.
583 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
586 if (i == 16) return -1; // all undef.
588 // Otherwise, check to see if the rest of the elements are consecutively
589 // numbered from this value.
590 unsigned ShiftAmt = SVOp->getMaskElt(i);
591 if (ShiftAmt < i) return -1;
595 // Check the rest of the elements to see if they are consecutive.
596 for (++i; i != 16; ++i)
597 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
600 // Check the rest of the elements to see if they are consecutive.
601 for (++i; i != 16; ++i)
602 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
608 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
609 /// specifies a splat of a single element that is suitable for input to
610 /// VSPLTB/VSPLTH/VSPLTW.
611 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
612 assert(N->getValueType(0) == MVT::v16i8 &&
613 (EltSize == 1 || EltSize == 2 || EltSize == 4));
615 // This is a splat operation if each element of the permute is the same, and
616 // if the value doesn't reference the second vector.
617 unsigned ElementBase = N->getMaskElt(0);
619 // FIXME: Handle UNDEF elements too!
620 if (ElementBase >= 16)
623 // Check that the indices are consecutive, in the case of a multi-byte element
624 // splatted with a v16i8 mask.
625 for (unsigned i = 1; i != EltSize; ++i)
626 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
629 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
630 if (N->getMaskElt(i) < 0) continue;
631 for (unsigned j = 0; j != EltSize; ++j)
632 if (N->getMaskElt(i+j) != N->getMaskElt(j))
638 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
640 bool PPC::isAllNegativeZeroVector(SDNode *N) {
641 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
643 APInt APVal, APUndef;
647 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
648 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
649 return CFP->getValueAPF().isNegZero();
654 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
655 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
656 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
657 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
658 assert(isSplatShuffleMask(SVOp, EltSize));
659 return SVOp->getMaskElt(0) / EltSize;
662 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
663 /// by using a vspltis[bhw] instruction of the specified element size, return
664 /// the constant being splatted. The ByteSize field indicates the number of
665 /// bytes of each element [124] -> [bhw].
666 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
669 // If ByteSize of the splat is bigger than the element size of the
670 // build_vector, then we have a case where we are checking for a splat where
671 // multiple elements of the buildvector are folded together into a single
672 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
673 unsigned EltSize = 16/N->getNumOperands();
674 if (EltSize < ByteSize) {
675 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
676 SDValue UniquedVals[4];
677 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
679 // See if all of the elements in the buildvector agree across.
680 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
681 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
682 // If the element isn't a constant, bail fully out.
683 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
686 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
687 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
688 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
689 return SDValue(); // no match.
692 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
693 // either constant or undef values that are identical for each chunk. See
694 // if these chunks can form into a larger vspltis*.
696 // Check to see if all of the leading entries are either 0 or -1. If
697 // neither, then this won't fit into the immediate field.
698 bool LeadingZero = true;
699 bool LeadingOnes = true;
700 for (unsigned i = 0; i != Multiple-1; ++i) {
701 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
703 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
704 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
706 // Finally, check the least significant entry.
708 if (UniquedVals[Multiple-1].getNode() == 0)
709 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
710 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
712 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
715 if (UniquedVals[Multiple-1].getNode() == 0)
716 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
717 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
718 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
719 return DAG.getTargetConstant(Val, MVT::i32);
725 // Check to see if this buildvec has a single non-undef value in its elements.
726 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
727 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
728 if (OpVal.getNode() == 0)
729 OpVal = N->getOperand(i);
730 else if (OpVal != N->getOperand(i))
734 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
736 unsigned ValSizeInBytes = EltSize;
738 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
739 Value = CN->getZExtValue();
740 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
741 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
742 Value = FloatToBits(CN->getValueAPF().convertToFloat());
745 // If the splat value is larger than the element value, then we can never do
746 // this splat. The only case that we could fit the replicated bits into our
747 // immediate field for would be zero, and we prefer to use vxor for it.
748 if (ValSizeInBytes < ByteSize) return SDValue();
750 // If the element value is larger than the splat value, cut it in half and
751 // check to see if the two halves are equal. Continue doing this until we
752 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
753 while (ValSizeInBytes > ByteSize) {
754 ValSizeInBytes >>= 1;
756 // If the top half equals the bottom half, we're still ok.
757 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
758 (Value & ((1 << (8*ValSizeInBytes))-1)))
762 // Properly sign extend the value.
763 int ShAmt = (4-ByteSize)*8;
764 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
766 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
767 if (MaskVal == 0) return SDValue();
769 // Finally, if this value fits in a 5 bit sext field, return it
770 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
771 return DAG.getTargetConstant(MaskVal, MVT::i32);
775 //===----------------------------------------------------------------------===//
776 // Addressing Mode Selection
777 //===----------------------------------------------------------------------===//
779 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
780 /// or 64-bit immediate, and if the value can be accurately represented as a
781 /// sign extension from a 16-bit value. If so, this returns true and the
783 static bool isIntS16Immediate(SDNode *N, short &Imm) {
784 if (N->getOpcode() != ISD::Constant)
787 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
788 if (N->getValueType(0) == MVT::i32)
789 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
791 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
793 static bool isIntS16Immediate(SDValue Op, short &Imm) {
794 return isIntS16Immediate(Op.getNode(), Imm);
798 /// SelectAddressRegReg - Given the specified addressed, check to see if it
799 /// can be represented as an indexed [r+r] operation. Returns false if it
800 /// can be more efficiently represented with [r+imm].
801 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
803 SelectionDAG &DAG) const {
805 if (N.getOpcode() == ISD::ADD) {
806 if (isIntS16Immediate(N.getOperand(1), imm))
808 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
811 Base = N.getOperand(0);
812 Index = N.getOperand(1);
814 } else if (N.getOpcode() == ISD::OR) {
815 if (isIntS16Immediate(N.getOperand(1), imm))
816 return false; // r+i can fold it if we can.
818 // If this is an or of disjoint bitfields, we can codegen this as an add
819 // (for better address arithmetic) if the LHS and RHS of the OR are provably
821 APInt LHSKnownZero, LHSKnownOne;
822 APInt RHSKnownZero, RHSKnownOne;
823 DAG.ComputeMaskedBits(N.getOperand(0),
824 APInt::getAllOnesValue(N.getOperand(0)
825 .getValueSizeInBits()),
826 LHSKnownZero, LHSKnownOne);
828 if (LHSKnownZero.getBoolValue()) {
829 DAG.ComputeMaskedBits(N.getOperand(1),
830 APInt::getAllOnesValue(N.getOperand(1)
831 .getValueSizeInBits()),
832 RHSKnownZero, RHSKnownOne);
833 // If all of the bits are known zero on the LHS or RHS, the add won't
835 if (~(LHSKnownZero | RHSKnownZero) == 0) {
836 Base = N.getOperand(0);
837 Index = N.getOperand(1);
846 /// Returns true if the address N can be represented by a base register plus
847 /// a signed 16-bit displacement [r+imm], and if it is not better
848 /// represented as reg+reg.
849 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
851 SelectionDAG &DAG) const {
852 // FIXME dl should come from parent load or store, not from address
853 DebugLoc dl = N.getDebugLoc();
854 // If this can be more profitably realized as r+r, fail.
855 if (SelectAddressRegReg(N, Disp, Base, DAG))
858 if (N.getOpcode() == ISD::ADD) {
860 if (isIntS16Immediate(N.getOperand(1), imm)) {
861 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
862 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
863 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
865 Base = N.getOperand(0);
867 return true; // [r+i]
868 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
869 // Match LOAD (ADD (X, Lo(G))).
870 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
871 && "Cannot handle constant offsets yet!");
872 Disp = N.getOperand(1).getOperand(0); // The global address.
873 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
874 Disp.getOpcode() == ISD::TargetConstantPool ||
875 Disp.getOpcode() == ISD::TargetJumpTable);
876 Base = N.getOperand(0);
877 return true; // [&g+r]
879 } else if (N.getOpcode() == ISD::OR) {
881 if (isIntS16Immediate(N.getOperand(1), imm)) {
882 // If this is an or of disjoint bitfields, we can codegen this as an add
883 // (for better address arithmetic) if the LHS and RHS of the OR are
884 // provably disjoint.
885 APInt LHSKnownZero, LHSKnownOne;
886 DAG.ComputeMaskedBits(N.getOperand(0),
887 APInt::getAllOnesValue(N.getOperand(0)
888 .getValueSizeInBits()),
889 LHSKnownZero, LHSKnownOne);
891 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
892 // If all of the bits are known zero on the LHS or RHS, the add won't
894 Base = N.getOperand(0);
895 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
899 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
900 // Loading from a constant address.
902 // If this address fits entirely in a 16-bit sext immediate field, codegen
905 if (isIntS16Immediate(CN, Imm)) {
906 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
907 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
908 CN->getValueType(0));
912 // Handle 32-bit sext immediates with LIS + addr mode.
913 if (CN->getValueType(0) == MVT::i32 ||
914 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
915 int Addr = (int)CN->getZExtValue();
917 // Otherwise, break this down into an LIS + disp.
918 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
920 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
921 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
922 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
927 Disp = DAG.getTargetConstant(0, getPointerTy());
928 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
929 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
932 return true; // [r+0]
935 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
936 /// represented as an indexed [r+r] operation.
937 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
939 SelectionDAG &DAG) const {
940 // Check to see if we can easily represent this as an [r+r] address. This
941 // will fail if it thinks that the address is more profitably represented as
942 // reg+imm, e.g. where imm = 0.
943 if (SelectAddressRegReg(N, Base, Index, DAG))
946 // If the operand is an addition, always emit this as [r+r], since this is
947 // better (for code size, and execution, as the memop does the add for free)
948 // than emitting an explicit add.
949 if (N.getOpcode() == ISD::ADD) {
950 Base = N.getOperand(0);
951 Index = N.getOperand(1);
955 // Otherwise, do it the hard way, using R0 as the base register.
956 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
962 /// SelectAddressRegImmShift - Returns true if the address N can be
963 /// represented by a base register plus a signed 14-bit displacement
964 /// [r+imm*4]. Suitable for use by STD and friends.
965 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
967 SelectionDAG &DAG) const {
968 // FIXME dl should come from the parent load or store, not the address
969 DebugLoc dl = N.getDebugLoc();
970 // If this can be more profitably realized as r+r, fail.
971 if (SelectAddressRegReg(N, Disp, Base, DAG))
974 if (N.getOpcode() == ISD::ADD) {
976 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
977 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
978 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
979 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
981 Base = N.getOperand(0);
983 return true; // [r+i]
984 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
985 // Match LOAD (ADD (X, Lo(G))).
986 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
987 && "Cannot handle constant offsets yet!");
988 Disp = N.getOperand(1).getOperand(0); // The global address.
989 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
990 Disp.getOpcode() == ISD::TargetConstantPool ||
991 Disp.getOpcode() == ISD::TargetJumpTable);
992 Base = N.getOperand(0);
993 return true; // [&g+r]
995 } else if (N.getOpcode() == ISD::OR) {
997 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
998 // If this is an or of disjoint bitfields, we can codegen this as an add
999 // (for better address arithmetic) if the LHS and RHS of the OR are
1000 // provably disjoint.
1001 APInt LHSKnownZero, LHSKnownOne;
1002 DAG.ComputeMaskedBits(N.getOperand(0),
1003 APInt::getAllOnesValue(N.getOperand(0)
1004 .getValueSizeInBits()),
1005 LHSKnownZero, LHSKnownOne);
1006 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1007 // If all of the bits are known zero on the LHS or RHS, the add won't
1009 Base = N.getOperand(0);
1010 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1014 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1015 // Loading from a constant address. Verify low two bits are clear.
1016 if ((CN->getZExtValue() & 3) == 0) {
1017 // If this address fits entirely in a 14-bit sext immediate field, codegen
1020 if (isIntS16Immediate(CN, Imm)) {
1021 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1022 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1023 CN->getValueType(0));
1027 // Fold the low-part of 32-bit absolute addresses into addr mode.
1028 if (CN->getValueType(0) == MVT::i32 ||
1029 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1030 int Addr = (int)CN->getZExtValue();
1032 // Otherwise, break this down into an LIS + disp.
1033 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1034 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1035 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1036 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1042 Disp = DAG.getTargetConstant(0, getPointerTy());
1043 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1044 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1047 return true; // [r+0]
1051 /// getPreIndexedAddressParts - returns true by value, base pointer and
1052 /// offset pointer and addressing mode by reference if the node's address
1053 /// can be legally represented as pre-indexed load / store address.
1054 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1056 ISD::MemIndexedMode &AM,
1057 SelectionDAG &DAG) const {
1058 // Disabled by default for now.
1059 if (!EnablePPCPreinc) return false;
1063 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1064 Ptr = LD->getBasePtr();
1065 VT = LD->getMemoryVT();
1067 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1068 Ptr = ST->getBasePtr();
1069 VT = ST->getMemoryVT();
1073 // PowerPC doesn't have preinc load/store instructions for vectors.
1077 // TODO: Check reg+reg first.
1079 // LDU/STU use reg+imm*4, others use reg+imm.
1080 if (VT != MVT::i64) {
1082 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1086 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1090 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1091 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1092 // sext i32 to i64 when addr mode is r+i.
1093 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1094 LD->getExtensionType() == ISD::SEXTLOAD &&
1095 isa<ConstantSDNode>(Offset))
1103 //===----------------------------------------------------------------------===//
1104 // LowerOperation implementation
1105 //===----------------------------------------------------------------------===//
1107 /// GetLabelAccessInfo - Return true if we should reference labels using a
1108 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1109 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1110 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1111 HiOpFlags = PPCII::MO_HA16;
1112 LoOpFlags = PPCII::MO_LO16;
1114 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1115 // non-darwin platform. We don't support PIC on other platforms yet.
1116 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1117 TM.getSubtarget<PPCSubtarget>().isDarwin();
1119 HiOpFlags |= PPCII::MO_PIC_FLAG;
1120 LoOpFlags |= PPCII::MO_PIC_FLAG;
1123 // If this is a reference to a global value that requires a non-lazy-ptr, make
1124 // sure that instruction lowering adds it.
1125 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1126 HiOpFlags |= PPCII::MO_NLP_FLAG;
1127 LoOpFlags |= PPCII::MO_NLP_FLAG;
1129 if (GV->hasHiddenVisibility()) {
1130 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1131 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1138 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1139 SelectionDAG &DAG) {
1140 EVT PtrVT = HiPart.getValueType();
1141 SDValue Zero = DAG.getConstant(0, PtrVT);
1142 DebugLoc DL = HiPart.getDebugLoc();
1144 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1145 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1147 // With PIC, the first instruction is actually "GR+hi(&G)".
1149 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1150 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1152 // Generate non-pic code that has direct accesses to the constant pool.
1153 // The address of the global is just (hi(&g)+lo(&g)).
1154 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1157 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1158 SelectionDAG &DAG) const {
1159 EVT PtrVT = Op.getValueType();
1160 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1161 const Constant *C = CP->getConstVal();
1163 unsigned MOHiFlag, MOLoFlag;
1164 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1166 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1168 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1169 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1172 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1173 EVT PtrVT = Op.getValueType();
1174 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1176 unsigned MOHiFlag, MOLoFlag;
1177 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1178 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1179 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1180 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1183 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1184 SelectionDAG &DAG) const {
1185 EVT PtrVT = Op.getValueType();
1187 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1189 unsigned MOHiFlag, MOLoFlag;
1190 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1191 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1192 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1193 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1196 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1197 SelectionDAG &DAG) const {
1198 EVT PtrVT = Op.getValueType();
1199 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1200 DebugLoc DL = GSDN->getDebugLoc();
1201 const GlobalValue *GV = GSDN->getGlobal();
1203 // 64-bit SVR4 ABI code is always position-independent.
1204 // The actual address of the GlobalValue is stored in the TOC.
1205 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1206 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1207 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1208 DAG.getRegister(PPC::X2, MVT::i64));
1211 unsigned MOHiFlag, MOLoFlag;
1212 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1215 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1217 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1219 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1221 // If the global reference is actually to a non-lazy-pointer, we have to do an
1222 // extra load to get the address of the global.
1223 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1224 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1229 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1230 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1231 DebugLoc dl = Op.getDebugLoc();
1233 // If we're comparing for equality to zero, expose the fact that this is
1234 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1235 // fold the new nodes.
1236 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1237 if (C->isNullValue() && CC == ISD::SETEQ) {
1238 EVT VT = Op.getOperand(0).getValueType();
1239 SDValue Zext = Op.getOperand(0);
1240 if (VT.bitsLT(MVT::i32)) {
1242 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1244 unsigned Log2b = Log2_32(VT.getSizeInBits());
1245 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1246 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1247 DAG.getConstant(Log2b, MVT::i32));
1248 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1250 // Leave comparisons against 0 and -1 alone for now, since they're usually
1251 // optimized. FIXME: revisit this when we can custom lower all setcc
1253 if (C->isAllOnesValue() || C->isNullValue())
1257 // If we have an integer seteq/setne, turn it into a compare against zero
1258 // by xor'ing the rhs with the lhs, which is faster than setting a
1259 // condition register, reading it back out, and masking the correct bit. The
1260 // normal approach here uses sub to do this instead of xor. Using xor exposes
1261 // the result to other bit-twiddling opportunities.
1262 EVT LHSVT = Op.getOperand(0).getValueType();
1263 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1264 EVT VT = Op.getValueType();
1265 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1267 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1272 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1273 const PPCSubtarget &Subtarget) const {
1274 SDNode *Node = Op.getNode();
1275 EVT VT = Node->getValueType(0);
1276 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1277 SDValue InChain = Node->getOperand(0);
1278 SDValue VAListPtr = Node->getOperand(1);
1279 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1280 DebugLoc dl = Node->getDebugLoc();
1282 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1285 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1286 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1288 InChain = GprIndex.getValue(1);
1290 if (VT == MVT::i64) {
1291 // Check if GprIndex is even
1292 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1293 DAG.getConstant(1, MVT::i32));
1294 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1295 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1296 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1297 DAG.getConstant(1, MVT::i32));
1298 // Align GprIndex to be even if it isn't
1299 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1303 // fpr index is 1 byte after gpr
1304 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1305 DAG.getConstant(1, MVT::i32));
1308 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1309 FprPtr, MachinePointerInfo(SV), MVT::i8,
1311 InChain = FprIndex.getValue(1);
1313 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1314 DAG.getConstant(8, MVT::i32));
1316 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1317 DAG.getConstant(4, MVT::i32));
1320 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1321 MachinePointerInfo(), false, false, 0);
1322 InChain = OverflowArea.getValue(1);
1324 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1325 MachinePointerInfo(), false, false, 0);
1326 InChain = RegSaveArea.getValue(1);
1328 // select overflow_area if index > 8
1329 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1330 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1332 // adjustment constant gpr_index * 4/8
1333 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1334 VT.isInteger() ? GprIndex : FprIndex,
1335 DAG.getConstant(VT.isInteger() ? 4 : 8,
1338 // OurReg = RegSaveArea + RegConstant
1339 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1342 // Floating types are 32 bytes into RegSaveArea
1343 if (VT.isFloatingPoint())
1344 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1345 DAG.getConstant(32, MVT::i32));
1347 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1348 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1349 VT.isInteger() ? GprIndex : FprIndex,
1350 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1353 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1354 VT.isInteger() ? VAListPtr : FprPtr,
1355 MachinePointerInfo(SV),
1356 MVT::i8, false, false, 0);
1358 // determine if we should load from reg_save_area or overflow_area
1359 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1361 // increase overflow_area by 4/8 if gpr/fpr > 8
1362 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1363 DAG.getConstant(VT.isInteger() ? 4 : 8,
1366 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1369 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1371 MachinePointerInfo(),
1372 MVT::i32, false, false, 0);
1374 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(), false, false, 0);
1377 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1378 SelectionDAG &DAG) const {
1379 return Op.getOperand(0);
1382 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1383 SelectionDAG &DAG) const {
1384 SDValue Chain = Op.getOperand(0);
1385 SDValue Trmp = Op.getOperand(1); // trampoline
1386 SDValue FPtr = Op.getOperand(2); // nested function
1387 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1388 DebugLoc dl = Op.getDebugLoc();
1390 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1391 bool isPPC64 = (PtrVT == MVT::i64);
1393 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1396 TargetLowering::ArgListTy Args;
1397 TargetLowering::ArgListEntry Entry;
1399 Entry.Ty = IntPtrTy;
1400 Entry.Node = Trmp; Args.push_back(Entry);
1402 // TrampSize == (isPPC64 ? 48 : 40);
1403 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1404 isPPC64 ? MVT::i64 : MVT::i32);
1405 Args.push_back(Entry);
1407 Entry.Node = FPtr; Args.push_back(Entry);
1408 Entry.Node = Nest; Args.push_back(Entry);
1410 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1411 std::pair<SDValue, SDValue> CallResult =
1412 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
1413 false, false, false, false, 0, CallingConv::C, false,
1414 /*isReturnValueUsed=*/true,
1415 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1418 return CallResult.second;
1421 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1422 const PPCSubtarget &Subtarget) const {
1423 MachineFunction &MF = DAG.getMachineFunction();
1424 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1426 DebugLoc dl = Op.getDebugLoc();
1428 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1429 // vastart just stores the address of the VarArgsFrameIndex slot into the
1430 // memory location argument.
1431 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1432 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1433 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1434 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1435 MachinePointerInfo(SV),
1439 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1440 // We suppose the given va_list is already allocated.
1443 // char gpr; /* index into the array of 8 GPRs
1444 // * stored in the register save area
1445 // * gpr=0 corresponds to r3,
1446 // * gpr=1 to r4, etc.
1448 // char fpr; /* index into the array of 8 FPRs
1449 // * stored in the register save area
1450 // * fpr=0 corresponds to f1,
1451 // * fpr=1 to f2, etc.
1453 // char *overflow_arg_area;
1454 // /* location on stack that holds
1455 // * the next overflow argument
1457 // char *reg_save_area;
1458 // /* where r3:r10 and f1:f8 (if saved)
1464 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1465 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1468 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1470 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1472 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1475 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1476 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1478 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1479 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1481 uint64_t FPROffset = 1;
1482 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1484 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1486 // Store first byte : number of int regs
1487 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1489 MachinePointerInfo(SV),
1490 MVT::i8, false, false, 0);
1491 uint64_t nextOffset = FPROffset;
1492 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1495 // Store second byte : number of float regs
1496 SDValue secondStore =
1497 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1498 MachinePointerInfo(SV, nextOffset), MVT::i8,
1500 nextOffset += StackOffset;
1501 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1503 // Store second word : arguments given on stack
1504 SDValue thirdStore =
1505 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1506 MachinePointerInfo(SV, nextOffset),
1508 nextOffset += FrameOffset;
1509 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1511 // Store third word : arguments given in registers
1512 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1513 MachinePointerInfo(SV, nextOffset),
1518 #include "PPCGenCallingConv.inc"
1520 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1521 CCValAssign::LocInfo &LocInfo,
1522 ISD::ArgFlagsTy &ArgFlags,
1527 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1529 CCValAssign::LocInfo &LocInfo,
1530 ISD::ArgFlagsTy &ArgFlags,
1532 static const unsigned ArgRegs[] = {
1533 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1534 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1536 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1538 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1540 // Skip one register if the first unallocated register has an even register
1541 // number and there are still argument registers available which have not been
1542 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1543 // need to skip a register if RegNum is odd.
1544 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1545 State.AllocateReg(ArgRegs[RegNum]);
1548 // Always return false here, as this function only makes sure that the first
1549 // unallocated register has an odd register number and does not actually
1550 // allocate a register for the current argument.
1554 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1556 CCValAssign::LocInfo &LocInfo,
1557 ISD::ArgFlagsTy &ArgFlags,
1559 static const unsigned ArgRegs[] = {
1560 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1564 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1566 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1568 // If there is only one Floating-point register left we need to put both f64
1569 // values of a split ppc_fp128 value on the stack.
1570 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1571 State.AllocateReg(ArgRegs[RegNum]);
1574 // Always return false here, as this function only makes sure that the two f64
1575 // values a ppc_fp128 value is split into are both passed in registers or both
1576 // passed on the stack and does not actually allocate a register for the
1577 // current argument.
1581 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1583 static const unsigned *GetFPR() {
1584 static const unsigned FPR[] = {
1585 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1586 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1592 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1594 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1595 unsigned PtrByteSize) {
1596 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1597 if (Flags.isByVal())
1598 ArgSize = Flags.getByValSize();
1599 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1605 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1606 CallingConv::ID CallConv, bool isVarArg,
1607 const SmallVectorImpl<ISD::InputArg>
1609 DebugLoc dl, SelectionDAG &DAG,
1610 SmallVectorImpl<SDValue> &InVals)
1612 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1613 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1616 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1622 PPCTargetLowering::LowerFormalArguments_SVR4(
1624 CallingConv::ID CallConv, bool isVarArg,
1625 const SmallVectorImpl<ISD::InputArg>
1627 DebugLoc dl, SelectionDAG &DAG,
1628 SmallVectorImpl<SDValue> &InVals) const {
1630 // 32-bit SVR4 ABI Stack Frame Layout:
1631 // +-----------------------------------+
1632 // +--> | Back chain |
1633 // | +-----------------------------------+
1634 // | | Floating-point register save area |
1635 // | +-----------------------------------+
1636 // | | General register save area |
1637 // | +-----------------------------------+
1638 // | | CR save word |
1639 // | +-----------------------------------+
1640 // | | VRSAVE save word |
1641 // | +-----------------------------------+
1642 // | | Alignment padding |
1643 // | +-----------------------------------+
1644 // | | Vector register save area |
1645 // | +-----------------------------------+
1646 // | | Local variable space |
1647 // | +-----------------------------------+
1648 // | | Parameter list area |
1649 // | +-----------------------------------+
1650 // | | LR save word |
1651 // | +-----------------------------------+
1652 // SP--> +--- | Back chain |
1653 // +-----------------------------------+
1656 // System V Application Binary Interface PowerPC Processor Supplement
1657 // AltiVec Technology Programming Interface Manual
1659 MachineFunction &MF = DAG.getMachineFunction();
1660 MachineFrameInfo *MFI = MF.getFrameInfo();
1661 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1663 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1664 // Potential tail calls could cause overwriting of argument stack slots.
1665 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1666 unsigned PtrByteSize = 4;
1668 // Assign locations to all of the incoming arguments.
1669 SmallVector<CCValAssign, 16> ArgLocs;
1670 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1671 getTargetMachine(), ArgLocs, *DAG.getContext());
1673 // Reserve space for the linkage area on the stack.
1674 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1676 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1678 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1679 CCValAssign &VA = ArgLocs[i];
1681 // Arguments stored in registers.
1682 if (VA.isRegLoc()) {
1683 TargetRegisterClass *RC;
1684 EVT ValVT = VA.getValVT();
1686 switch (ValVT.getSimpleVT().SimpleTy) {
1688 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1690 RC = PPC::GPRCRegisterClass;
1693 RC = PPC::F4RCRegisterClass;
1696 RC = PPC::F8RCRegisterClass;
1702 RC = PPC::VRRCRegisterClass;
1706 // Transform the arguments stored in physical registers into virtual ones.
1707 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1708 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1710 InVals.push_back(ArgValue);
1712 // Argument stored in memory.
1713 assert(VA.isMemLoc());
1715 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1716 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1719 // Create load nodes to retrieve arguments from the stack.
1720 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1721 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1722 MachinePointerInfo(),
1727 // Assign locations to all of the incoming aggregate by value arguments.
1728 // Aggregates passed by value are stored in the local variable space of the
1729 // caller's stack frame, right above the parameter list area.
1730 SmallVector<CCValAssign, 16> ByValArgLocs;
1731 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1732 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1734 // Reserve stack space for the allocations in CCInfo.
1735 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1737 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1739 // Area that is at least reserved in the caller of this function.
1740 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1742 // Set the size that is at least reserved in caller of this function. Tail
1743 // call optimized function's reserved stack space needs to be aligned so that
1744 // taking the difference between two stack areas will result in an aligned
1746 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1749 std::max(MinReservedArea,
1750 PPCFrameLowering::getMinCallFrameSize(false, false));
1752 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1753 getStackAlignment();
1754 unsigned AlignMask = TargetAlign-1;
1755 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1757 FI->setMinReservedArea(MinReservedArea);
1759 SmallVector<SDValue, 8> MemOps;
1761 // If the function takes variable number of arguments, make a frame index for
1762 // the start of the first vararg value... for expansion of llvm.va_start.
1764 static const unsigned GPArgRegs[] = {
1765 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1766 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1768 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1770 static const unsigned FPArgRegs[] = {
1771 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1774 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1776 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1778 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1781 // Make room for NumGPArgRegs and NumFPArgRegs.
1782 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1783 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1785 FuncInfo->setVarArgsStackOffset(
1786 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1787 CCInfo.getNextStackOffset(), true));
1789 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1790 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1792 // The fixed integer arguments of a variadic function are stored to the
1793 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1794 // the result of va_next.
1795 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1796 // Get an existing live-in vreg, or add a new one.
1797 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1799 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1801 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1802 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1803 MachinePointerInfo(), false, false, 0);
1804 MemOps.push_back(Store);
1805 // Increment the address by four for the next argument to store
1806 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1807 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1810 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1812 // The double arguments are stored to the VarArgsFrameIndex
1814 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1815 // Get an existing live-in vreg, or add a new one.
1816 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1818 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1820 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1821 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1822 MachinePointerInfo(), false, false, 0);
1823 MemOps.push_back(Store);
1824 // Increment the address by eight for the next argument to store
1825 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1827 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1831 if (!MemOps.empty())
1832 Chain = DAG.getNode(ISD::TokenFactor, dl,
1833 MVT::Other, &MemOps[0], MemOps.size());
1839 PPCTargetLowering::LowerFormalArguments_Darwin(
1841 CallingConv::ID CallConv, bool isVarArg,
1842 const SmallVectorImpl<ISD::InputArg>
1844 DebugLoc dl, SelectionDAG &DAG,
1845 SmallVectorImpl<SDValue> &InVals) const {
1846 // TODO: add description of PPC stack frame format, or at least some docs.
1848 MachineFunction &MF = DAG.getMachineFunction();
1849 MachineFrameInfo *MFI = MF.getFrameInfo();
1850 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1852 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1853 bool isPPC64 = PtrVT == MVT::i64;
1854 // Potential tail calls could cause overwriting of argument stack slots.
1855 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1856 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1858 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1859 // Area that is at least reserved in caller of this function.
1860 unsigned MinReservedArea = ArgOffset;
1862 static const unsigned GPR_32[] = { // 32-bit registers.
1863 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1864 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1866 static const unsigned GPR_64[] = { // 64-bit registers.
1867 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1868 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1871 static const unsigned *FPR = GetFPR();
1873 static const unsigned VR[] = {
1874 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1875 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1878 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1879 const unsigned Num_FPR_Regs = 13;
1880 const unsigned Num_VR_Regs = array_lengthof( VR);
1882 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1884 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1886 // In 32-bit non-varargs functions, the stack space for vectors is after the
1887 // stack space for non-vectors. We do not use this space unless we have
1888 // too many vectors to fit in registers, something that only occurs in
1889 // constructed examples:), but we have to walk the arglist to figure
1890 // that out...for the pathological case, compute VecArgOffset as the
1891 // start of the vector parameter area. Computing VecArgOffset is the
1892 // entire point of the following loop.
1893 unsigned VecArgOffset = ArgOffset;
1894 if (!isVarArg && !isPPC64) {
1895 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1897 EVT ObjectVT = Ins[ArgNo].VT;
1898 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1899 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1901 if (Flags.isByVal()) {
1902 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1903 ObjSize = Flags.getByValSize();
1905 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1906 VecArgOffset += ArgSize;
1910 switch(ObjectVT.getSimpleVT().SimpleTy) {
1911 default: llvm_unreachable("Unhandled argument type!");
1914 VecArgOffset += isPPC64 ? 8 : 4;
1916 case MVT::i64: // PPC64
1924 // Nothing to do, we're only looking at Nonvector args here.
1929 // We've found where the vector parameter area in memory is. Skip the
1930 // first 12 parameters; these don't use that memory.
1931 VecArgOffset = ((VecArgOffset+15)/16)*16;
1932 VecArgOffset += 12*16;
1934 // Add DAG nodes to load the arguments or copy them out of registers. On
1935 // entry to a function on PPC, the arguments start after the linkage area,
1936 // although the first ones are often in registers.
1938 SmallVector<SDValue, 8> MemOps;
1939 unsigned nAltivecParamsAtEnd = 0;
1940 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1942 bool needsLoad = false;
1943 EVT ObjectVT = Ins[ArgNo].VT;
1944 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1945 unsigned ArgSize = ObjSize;
1946 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1948 unsigned CurArgOffset = ArgOffset;
1950 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1951 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1952 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1953 if (isVarArg || isPPC64) {
1954 MinReservedArea = ((MinReservedArea+15)/16)*16;
1955 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1958 } else nAltivecParamsAtEnd++;
1960 // Calculate min reserved area.
1961 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1965 // FIXME the codegen can be much improved in some cases.
1966 // We do not have to keep everything in memory.
1967 if (Flags.isByVal()) {
1968 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1969 ObjSize = Flags.getByValSize();
1970 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1971 // Objects of size 1 and 2 are right justified, everything else is
1972 // left justified. This means the memory address is adjusted forwards.
1973 if (ObjSize==1 || ObjSize==2) {
1974 CurArgOffset = CurArgOffset + (4 - ObjSize);
1976 // The value of the object is its address.
1977 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
1978 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1979 InVals.push_back(FIN);
1980 if (ObjSize==1 || ObjSize==2) {
1981 if (GPR_idx != Num_GPR_Regs) {
1984 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1986 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1987 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1988 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1989 MachinePointerInfo(),
1990 ObjSize==1 ? MVT::i8 : MVT::i16,
1992 MemOps.push_back(Store);
1996 ArgOffset += PtrByteSize;
2000 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2001 // Store whatever pieces of the object are in registers
2002 // to memory. ArgVal will be address of the beginning of
2004 if (GPR_idx != Num_GPR_Regs) {
2007 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2009 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2010 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2011 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2012 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2013 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2014 MachinePointerInfo(),
2016 MemOps.push_back(Store);
2018 ArgOffset += PtrByteSize;
2020 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2027 switch (ObjectVT.getSimpleVT().SimpleTy) {
2028 default: llvm_unreachable("Unhandled argument type!");
2031 if (GPR_idx != Num_GPR_Regs) {
2032 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2033 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2037 ArgSize = PtrByteSize;
2039 // All int arguments reserve stack space in the Darwin ABI.
2040 ArgOffset += PtrByteSize;
2044 case MVT::i64: // PPC64
2045 if (GPR_idx != Num_GPR_Regs) {
2046 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2047 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2049 if (ObjectVT == MVT::i32) {
2050 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2051 // value to MVT::i64 and then truncate to the correct register size.
2053 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2054 DAG.getValueType(ObjectVT));
2055 else if (Flags.isZExt())
2056 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2057 DAG.getValueType(ObjectVT));
2059 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2065 ArgSize = PtrByteSize;
2067 // All int arguments reserve stack space in the Darwin ABI.
2073 // Every 4 bytes of argument space consumes one of the GPRs available for
2074 // argument passing.
2075 if (GPR_idx != Num_GPR_Regs) {
2077 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2080 if (FPR_idx != Num_FPR_Regs) {
2083 if (ObjectVT == MVT::f32)
2084 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2086 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2088 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2094 // All FP arguments reserve stack space in the Darwin ABI.
2095 ArgOffset += isPPC64 ? 8 : ObjSize;
2101 // Note that vector arguments in registers don't reserve stack space,
2102 // except in varargs functions.
2103 if (VR_idx != Num_VR_Regs) {
2104 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2105 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2107 while ((ArgOffset % 16) != 0) {
2108 ArgOffset += PtrByteSize;
2109 if (GPR_idx != Num_GPR_Regs)
2113 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2117 if (!isVarArg && !isPPC64) {
2118 // Vectors go after all the nonvectors.
2119 CurArgOffset = VecArgOffset;
2122 // Vectors are aligned.
2123 ArgOffset = ((ArgOffset+15)/16)*16;
2124 CurArgOffset = ArgOffset;
2132 // We need to load the argument to a virtual register if we determined above
2133 // that we ran out of physical registers of the appropriate type.
2135 int FI = MFI->CreateFixedObject(ObjSize,
2136 CurArgOffset + (ArgSize - ObjSize),
2138 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2139 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2143 InVals.push_back(ArgVal);
2146 // Set the size that is at least reserved in caller of this function. Tail
2147 // call optimized function's reserved stack space needs to be aligned so that
2148 // taking the difference between two stack areas will result in an aligned
2150 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2151 // Add the Altivec parameters at the end, if needed.
2152 if (nAltivecParamsAtEnd) {
2153 MinReservedArea = ((MinReservedArea+15)/16)*16;
2154 MinReservedArea += 16*nAltivecParamsAtEnd;
2157 std::max(MinReservedArea,
2158 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2159 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2160 getStackAlignment();
2161 unsigned AlignMask = TargetAlign-1;
2162 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2163 FI->setMinReservedArea(MinReservedArea);
2165 // If the function takes variable number of arguments, make a frame index for
2166 // the start of the first vararg value... for expansion of llvm.va_start.
2168 int Depth = ArgOffset;
2170 FuncInfo->setVarArgsFrameIndex(
2171 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2173 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2175 // If this function is vararg, store any remaining integer argument regs
2176 // to their spots on the stack so that they may be loaded by deferencing the
2177 // result of va_next.
2178 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2182 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2184 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2186 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2187 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2188 MachinePointerInfo(), false, false, 0);
2189 MemOps.push_back(Store);
2190 // Increment the address by four for the next argument to store
2191 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2192 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2196 if (!MemOps.empty())
2197 Chain = DAG.getNode(ISD::TokenFactor, dl,
2198 MVT::Other, &MemOps[0], MemOps.size());
2203 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2204 /// linkage area for the Darwin ABI.
2206 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2210 const SmallVectorImpl<ISD::OutputArg>
2212 const SmallVectorImpl<SDValue> &OutVals,
2213 unsigned &nAltivecParamsAtEnd) {
2214 // Count how many bytes are to be pushed on the stack, including the linkage
2215 // area, and parameter passing area. We start with 24/48 bytes, which is
2216 // prereserved space for [SP][CR][LR][3 x unused].
2217 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2218 unsigned NumOps = Outs.size();
2219 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2221 // Add up all the space actually used.
2222 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2223 // they all go in registers, but we must reserve stack space for them for
2224 // possible use by the caller. In varargs or 64-bit calls, parameters are
2225 // assigned stack space in order, with padding so Altivec parameters are
2227 nAltivecParamsAtEnd = 0;
2228 for (unsigned i = 0; i != NumOps; ++i) {
2229 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2230 EVT ArgVT = Outs[i].VT;
2231 // Varargs Altivec parameters are padded to a 16 byte boundary.
2232 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2233 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2234 if (!isVarArg && !isPPC64) {
2235 // Non-varargs Altivec parameters go after all the non-Altivec
2236 // parameters; handle those later so we know how much padding we need.
2237 nAltivecParamsAtEnd++;
2240 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2241 NumBytes = ((NumBytes+15)/16)*16;
2243 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2246 // Allow for Altivec parameters at the end, if needed.
2247 if (nAltivecParamsAtEnd) {
2248 NumBytes = ((NumBytes+15)/16)*16;
2249 NumBytes += 16*nAltivecParamsAtEnd;
2252 // The prolog code of the callee may store up to 8 GPR argument registers to
2253 // the stack, allowing va_start to index over them in memory if its varargs.
2254 // Because we cannot tell if this is needed on the caller side, we have to
2255 // conservatively assume that it is needed. As such, make sure we have at
2256 // least enough stack space for the caller to store the 8 GPRs.
2257 NumBytes = std::max(NumBytes,
2258 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2260 // Tail call needs the stack to be aligned.
2261 if (CC==CallingConv::Fast && GuaranteedTailCallOpt) {
2262 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2263 getStackAlignment();
2264 unsigned AlignMask = TargetAlign-1;
2265 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2271 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2272 /// adjusted to accommodate the arguments for the tailcall.
2273 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2274 unsigned ParamSize) {
2276 if (!isTailCall) return 0;
2278 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2279 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2280 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2281 // Remember only if the new adjustement is bigger.
2282 if (SPDiff < FI->getTailCallSPDelta())
2283 FI->setTailCallSPDelta(SPDiff);
2288 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2289 /// for tail call optimization. Targets which want to do tail call
2290 /// optimization should implement this function.
2292 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2293 CallingConv::ID CalleeCC,
2295 const SmallVectorImpl<ISD::InputArg> &Ins,
2296 SelectionDAG& DAG) const {
2297 if (!GuaranteedTailCallOpt)
2300 // Variable argument functions are not supported.
2304 MachineFunction &MF = DAG.getMachineFunction();
2305 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2306 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2307 // Functions containing by val parameters are not supported.
2308 for (unsigned i = 0; i != Ins.size(); i++) {
2309 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2310 if (Flags.isByVal()) return false;
2313 // Non PIC/GOT tail calls are supported.
2314 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2317 // At the moment we can only do local tail calls (in same module, hidden
2318 // or protected) if we are generating PIC.
2319 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2320 return G->getGlobal()->hasHiddenVisibility()
2321 || G->getGlobal()->hasProtectedVisibility();
2327 /// isCallCompatibleAddress - Return the immediate to use if the specified
2328 /// 32-bit value is representable in the immediate field of a BxA instruction.
2329 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2330 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2333 int Addr = C->getZExtValue();
2334 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2335 (Addr << 6 >> 6) != Addr)
2336 return 0; // Top 6 bits have to be sext of immediate.
2338 return DAG.getConstant((int)C->getZExtValue() >> 2,
2339 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2344 struct TailCallArgumentInfo {
2349 TailCallArgumentInfo() : FrameIdx(0) {}
2354 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2356 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2358 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2359 SmallVector<SDValue, 8> &MemOpChains,
2361 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2362 SDValue Arg = TailCallArgs[i].Arg;
2363 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2364 int FI = TailCallArgs[i].FrameIdx;
2365 // Store relative to framepointer.
2366 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2367 MachinePointerInfo::getFixedStack(FI),
2372 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2373 /// the appropriate stack slot for the tail call optimized function call.
2374 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2375 MachineFunction &MF,
2384 // Calculate the new stack slot for the return address.
2385 int SlotSize = isPPC64 ? 8 : 4;
2386 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2388 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2389 NewRetAddrLoc, true);
2390 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2391 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2392 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2393 MachinePointerInfo::getFixedStack(NewRetAddr),
2396 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2397 // slot as the FP is never overwritten.
2400 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2401 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2403 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2404 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2405 MachinePointerInfo::getFixedStack(NewFPIdx),
2412 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2413 /// the position of the argument.
2415 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2416 SDValue Arg, int SPDiff, unsigned ArgOffset,
2417 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2418 int Offset = ArgOffset + SPDiff;
2419 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2420 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2421 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2422 SDValue FIN = DAG.getFrameIndex(FI, VT);
2423 TailCallArgumentInfo Info;
2425 Info.FrameIdxOp = FIN;
2427 TailCallArguments.push_back(Info);
2430 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2431 /// stack slot. Returns the chain as result and the loaded frame pointers in
2432 /// LROpOut/FPOpout. Used when tail calling.
2433 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2439 DebugLoc dl) const {
2441 // Load the LR and FP stack slot for later adjusting.
2442 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2443 LROpOut = getReturnAddrFrameIndex(DAG);
2444 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2446 Chain = SDValue(LROpOut.getNode(), 1);
2448 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2449 // slot as the FP is never overwritten.
2451 FPOpOut = getFramePointerFrameIndex(DAG);
2452 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2454 Chain = SDValue(FPOpOut.getNode(), 1);
2460 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2461 /// by "Src" to address "Dst" of size "Size". Alignment information is
2462 /// specified by the specific parameter attribute. The copy will be passed as
2463 /// a byval function parameter.
2464 /// Sometimes what we are copying is the end of a larger object, the part that
2465 /// does not fit in registers.
2467 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2468 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2470 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2471 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2472 false, false, MachinePointerInfo(0),
2473 MachinePointerInfo(0));
2476 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2479 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2480 SDValue Arg, SDValue PtrOff, int SPDiff,
2481 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2482 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2483 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2485 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2490 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2492 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2493 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2494 DAG.getConstant(ArgOffset, PtrVT));
2496 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2497 MachinePointerInfo(), false, false, 0));
2498 // Calculate and remember argument location.
2499 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2504 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2505 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2506 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2507 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2508 MachineFunction &MF = DAG.getMachineFunction();
2510 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2511 // might overwrite each other in case of tail call optimization.
2512 SmallVector<SDValue, 8> MemOpChains2;
2513 // Do not flag preceding copytoreg stuff together with the following stuff.
2515 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2517 if (!MemOpChains2.empty())
2518 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2519 &MemOpChains2[0], MemOpChains2.size());
2521 // Store the return address to the appropriate stack slot.
2522 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2523 isPPC64, isDarwinABI, dl);
2525 // Emit callseq_end just before tailcall node.
2526 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2527 DAG.getIntPtrConstant(0, true), InFlag);
2528 InFlag = Chain.getValue(1);
2532 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2533 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2534 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2535 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2536 const PPCSubtarget &PPCSubTarget) {
2538 bool isPPC64 = PPCSubTarget.isPPC64();
2539 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2541 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2542 NodeTys.push_back(MVT::Other); // Returns a chain
2543 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2545 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2547 bool needIndirectCall = true;
2548 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2549 // If this is an absolute destination address, use the munged value.
2550 Callee = SDValue(Dest, 0);
2551 needIndirectCall = false;
2554 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2555 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2556 // Use indirect calls for ALL functions calls in JIT mode, since the
2557 // far-call stubs may be outside relocation limits for a BL instruction.
2558 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2559 unsigned OpFlags = 0;
2560 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2561 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2562 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2563 (G->getGlobal()->isDeclaration() ||
2564 G->getGlobal()->isWeakForLinker())) {
2565 // PC-relative references to external symbols should go through $stub,
2566 // unless we're building with the leopard linker or later, which
2567 // automatically synthesizes these stubs.
2568 OpFlags = PPCII::MO_DARWIN_STUB;
2571 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2572 // every direct call is) turn it into a TargetGlobalAddress /
2573 // TargetExternalSymbol node so that legalize doesn't hack it.
2574 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2575 Callee.getValueType(),
2577 needIndirectCall = false;
2581 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2582 unsigned char OpFlags = 0;
2584 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2585 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2586 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2587 // PC-relative references to external symbols should go through $stub,
2588 // unless we're building with the leopard linker or later, which
2589 // automatically synthesizes these stubs.
2590 OpFlags = PPCII::MO_DARWIN_STUB;
2593 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2595 needIndirectCall = false;
2598 if (needIndirectCall) {
2599 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2600 // to do the call, we can't use PPCISD::CALL.
2601 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2603 if (isSVR4ABI && isPPC64) {
2604 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2605 // entry point, but to the function descriptor (the function entry point
2606 // address is part of the function descriptor though).
2607 // The function descriptor is a three doubleword structure with the
2608 // following fields: function entry point, TOC base address and
2609 // environment pointer.
2610 // Thus for a call through a function pointer, the following actions need
2612 // 1. Save the TOC of the caller in the TOC save area of its stack
2613 // frame (this is done in LowerCall_Darwin()).
2614 // 2. Load the address of the function entry point from the function
2616 // 3. Load the TOC of the callee from the function descriptor into r2.
2617 // 4. Load the environment pointer from the function descriptor into
2619 // 5. Branch to the function entry point address.
2620 // 6. On return of the callee, the TOC of the caller needs to be
2621 // restored (this is done in FinishCall()).
2623 // All those operations are flagged together to ensure that no other
2624 // operations can be scheduled in between. E.g. without flagging the
2625 // operations together, a TOC access in the caller could be scheduled
2626 // between the load of the callee TOC and the branch to the callee, which
2627 // results in the TOC access going through the TOC of the callee instead
2628 // of going through the TOC of the caller, which leads to incorrect code.
2630 // Load the address of the function entry point from the function
2632 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2633 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2634 InFlag.getNode() ? 3 : 2);
2635 Chain = LoadFuncPtr.getValue(1);
2636 InFlag = LoadFuncPtr.getValue(2);
2638 // Load environment pointer into r11.
2639 // Offset of the environment pointer within the function descriptor.
2640 SDValue PtrOff = DAG.getIntPtrConstant(16);
2642 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2643 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2645 Chain = LoadEnvPtr.getValue(1);
2646 InFlag = LoadEnvPtr.getValue(2);
2648 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2650 Chain = EnvVal.getValue(0);
2651 InFlag = EnvVal.getValue(1);
2653 // Load TOC of the callee into r2. We are using a target-specific load
2654 // with r2 hard coded, because the result of a target-independent load
2655 // would never go directly into r2, since r2 is a reserved register (which
2656 // prevents the register allocator from allocating it), resulting in an
2657 // additional register being allocated and an unnecessary move instruction
2659 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2660 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2662 Chain = LoadTOCPtr.getValue(0);
2663 InFlag = LoadTOCPtr.getValue(1);
2665 MTCTROps[0] = Chain;
2666 MTCTROps[1] = LoadFuncPtr;
2667 MTCTROps[2] = InFlag;
2670 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2671 2 + (InFlag.getNode() != 0));
2672 InFlag = Chain.getValue(1);
2675 NodeTys.push_back(MVT::Other);
2676 NodeTys.push_back(MVT::Glue);
2677 Ops.push_back(Chain);
2678 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2680 // Add CTR register as callee so a bctr can be emitted later.
2682 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2685 // If this is a direct call, pass the chain and the callee.
2686 if (Callee.getNode()) {
2687 Ops.push_back(Chain);
2688 Ops.push_back(Callee);
2690 // If this is a tail call add stack pointer delta.
2692 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2694 // Add argument registers to the end of the list so that they are known live
2696 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2697 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2698 RegsToPass[i].second.getValueType()));
2704 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2705 CallingConv::ID CallConv, bool isVarArg,
2706 const SmallVectorImpl<ISD::InputArg> &Ins,
2707 DebugLoc dl, SelectionDAG &DAG,
2708 SmallVectorImpl<SDValue> &InVals) const {
2710 SmallVector<CCValAssign, 16> RVLocs;
2711 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2712 getTargetMachine(), RVLocs, *DAG.getContext());
2713 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2715 // Copy all of the result registers out of their specified physreg.
2716 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2717 CCValAssign &VA = RVLocs[i];
2718 EVT VT = VA.getValVT();
2719 assert(VA.isRegLoc() && "Can only return in registers!");
2720 Chain = DAG.getCopyFromReg(Chain, dl,
2721 VA.getLocReg(), VT, InFlag).getValue(1);
2722 InVals.push_back(Chain.getValue(0));
2723 InFlag = Chain.getValue(2);
2730 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2731 bool isTailCall, bool isVarArg,
2733 SmallVector<std::pair<unsigned, SDValue>, 8>
2735 SDValue InFlag, SDValue Chain,
2737 int SPDiff, unsigned NumBytes,
2738 const SmallVectorImpl<ISD::InputArg> &Ins,
2739 SmallVectorImpl<SDValue> &InVals) const {
2740 std::vector<EVT> NodeTys;
2741 SmallVector<SDValue, 8> Ops;
2742 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2743 isTailCall, RegsToPass, Ops, NodeTys,
2746 // When performing tail call optimization the callee pops its arguments off
2747 // the stack. Account for this here so these bytes can be pushed back on in
2748 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2749 int BytesCalleePops =
2750 (CallConv==CallingConv::Fast && GuaranteedTailCallOpt) ? NumBytes : 0;
2752 if (InFlag.getNode())
2753 Ops.push_back(InFlag);
2757 // If this is the first return lowered for this function, add the regs
2758 // to the liveout set for the function.
2759 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2760 SmallVector<CCValAssign, 16> RVLocs;
2761 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2762 getTargetMachine(), RVLocs, *DAG.getContext());
2763 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2764 for (unsigned i = 0; i != RVLocs.size(); ++i)
2765 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2768 assert(((Callee.getOpcode() == ISD::Register &&
2769 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2770 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2771 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2772 isa<ConstantSDNode>(Callee)) &&
2773 "Expecting an global address, external symbol, absolute value or register");
2775 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2778 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2779 InFlag = Chain.getValue(1);
2781 // Add a NOP immediately after the branch instruction when using the 64-bit
2782 // SVR4 ABI. At link time, if caller and callee are in a different module and
2783 // thus have a different TOC, the call will be replaced with a call to a stub
2784 // function which saves the current TOC, loads the TOC of the callee and
2785 // branches to the callee. The NOP will be replaced with a load instruction
2786 // which restores the TOC of the caller from the TOC save slot of the current
2787 // stack frame. If caller and callee belong to the same module (and have the
2788 // same TOC), the NOP will remain unchanged.
2789 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2790 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2791 if (CallOpc == PPCISD::BCTRL_SVR4) {
2792 // This is a call through a function pointer.
2793 // Restore the caller TOC from the save area into R2.
2794 // See PrepareCall() for more information about calls through function
2795 // pointers in the 64-bit SVR4 ABI.
2796 // We are using a target-specific load with r2 hard coded, because the
2797 // result of a target-independent load would never go directly into r2,
2798 // since r2 is a reserved register (which prevents the register allocator
2799 // from allocating it), resulting in an additional register being
2800 // allocated and an unnecessary move instruction being generated.
2801 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2802 InFlag = Chain.getValue(1);
2804 // Otherwise insert NOP.
2805 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Glue, InFlag);
2809 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2810 DAG.getIntPtrConstant(BytesCalleePops, true),
2813 InFlag = Chain.getValue(1);
2815 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2816 Ins, dl, DAG, InVals);
2820 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2821 CallingConv::ID CallConv, bool isVarArg,
2823 const SmallVectorImpl<ISD::OutputArg> &Outs,
2824 const SmallVectorImpl<SDValue> &OutVals,
2825 const SmallVectorImpl<ISD::InputArg> &Ins,
2826 DebugLoc dl, SelectionDAG &DAG,
2827 SmallVectorImpl<SDValue> &InVals) const {
2829 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2832 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2833 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2834 isTailCall, Outs, OutVals, Ins,
2837 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2838 isTailCall, Outs, OutVals, Ins,
2843 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2844 CallingConv::ID CallConv, bool isVarArg,
2846 const SmallVectorImpl<ISD::OutputArg> &Outs,
2847 const SmallVectorImpl<SDValue> &OutVals,
2848 const SmallVectorImpl<ISD::InputArg> &Ins,
2849 DebugLoc dl, SelectionDAG &DAG,
2850 SmallVectorImpl<SDValue> &InVals) const {
2851 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2852 // of the 32-bit SVR4 ABI stack frame layout.
2854 assert((CallConv == CallingConv::C ||
2855 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2857 unsigned PtrByteSize = 4;
2859 MachineFunction &MF = DAG.getMachineFunction();
2861 // Mark this function as potentially containing a function that contains a
2862 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2863 // and restoring the callers stack pointer in this functions epilog. This is
2864 // done because by tail calling the called function might overwrite the value
2865 // in this function's (MF) stack pointer stack slot 0(SP).
2866 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2867 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2869 // Count how many bytes are to be pushed on the stack, including the linkage
2870 // area, parameter list area and the part of the local variable space which
2871 // contains copies of aggregates which are passed by value.
2873 // Assign locations to all of the outgoing arguments.
2874 SmallVector<CCValAssign, 16> ArgLocs;
2875 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2876 getTargetMachine(), ArgLocs, *DAG.getContext());
2878 // Reserve space for the linkage area on the stack.
2879 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2882 // Handle fixed and variable vector arguments differently.
2883 // Fixed vector arguments go into registers as long as registers are
2884 // available. Variable vector arguments always go into memory.
2885 unsigned NumArgs = Outs.size();
2887 for (unsigned i = 0; i != NumArgs; ++i) {
2888 MVT ArgVT = Outs[i].VT;
2889 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2892 if (Outs[i].IsFixed) {
2893 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2896 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2902 errs() << "Call operand #" << i << " has unhandled type "
2903 << EVT(ArgVT).getEVTString() << "\n";
2905 llvm_unreachable(0);
2909 // All arguments are treated the same.
2910 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2913 // Assign locations to all of the outgoing aggregate by value arguments.
2914 SmallVector<CCValAssign, 16> ByValArgLocs;
2915 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2916 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2918 // Reserve stack space for the allocations in CCInfo.
2919 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2921 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2923 // Size of the linkage area, parameter list area and the part of the local
2924 // space variable where copies of aggregates which are passed by value are
2926 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2928 // Calculate by how many bytes the stack has to be adjusted in case of tail
2929 // call optimization.
2930 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2932 // Adjust the stack pointer for the new arguments...
2933 // These operations are automatically eliminated by the prolog/epilog pass
2934 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2935 SDValue CallSeqStart = Chain;
2937 // Load the return address and frame pointer so it can be moved somewhere else
2940 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2943 // Set up a copy of the stack pointer for use loading and storing any
2944 // arguments that may not fit in the registers available for argument
2946 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2948 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2949 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2950 SmallVector<SDValue, 8> MemOpChains;
2952 bool seenFloatArg = false;
2953 // Walk the register/memloc assignments, inserting copies/loads.
2954 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2957 CCValAssign &VA = ArgLocs[i];
2958 SDValue Arg = OutVals[i];
2959 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2961 if (Flags.isByVal()) {
2962 // Argument is an aggregate which is passed by value, thus we need to
2963 // create a copy of it in the local variable space of the current stack
2964 // frame (which is the stack frame of the caller) and pass the address of
2965 // this copy to the callee.
2966 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2967 CCValAssign &ByValVA = ByValArgLocs[j++];
2968 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2970 // Memory reserved in the local variable space of the callers stack frame.
2971 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2973 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2974 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2976 // Create a copy of the argument in the local area of the current
2978 SDValue MemcpyCall =
2979 CreateCopyOfByValArgument(Arg, PtrOff,
2980 CallSeqStart.getNode()->getOperand(0),
2983 // This must go outside the CALLSEQ_START..END.
2984 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2985 CallSeqStart.getNode()->getOperand(1));
2986 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2987 NewCallSeqStart.getNode());
2988 Chain = CallSeqStart = NewCallSeqStart;
2990 // Pass the address of the aggregate copy on the stack either in a
2991 // physical register or in the parameter list area of the current stack
2992 // frame to the callee.
2996 if (VA.isRegLoc()) {
2997 seenFloatArg |= VA.getLocVT().isFloatingPoint();
2998 // Put argument in a physical register.
2999 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3001 // Put argument in the parameter list area of the current stack frame.
3002 assert(VA.isMemLoc());
3003 unsigned LocMemOffset = VA.getLocMemOffset();
3006 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3007 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3009 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3010 MachinePointerInfo(),
3013 // Calculate and remember argument location.
3014 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3020 if (!MemOpChains.empty())
3021 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3022 &MemOpChains[0], MemOpChains.size());
3024 // Set CR6 to true if this is a vararg call with floating args passed in
3027 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3029 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3032 // Build a sequence of copy-to-reg nodes chained together with token chain
3033 // and flag operands which copy the outgoing args into the appropriate regs.
3035 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3036 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3037 RegsToPass[i].second, InFlag);
3038 InFlag = Chain.getValue(1);
3042 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3043 false, TailCallArguments);
3045 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3046 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3051 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3052 CallingConv::ID CallConv, bool isVarArg,
3054 const SmallVectorImpl<ISD::OutputArg> &Outs,
3055 const SmallVectorImpl<SDValue> &OutVals,
3056 const SmallVectorImpl<ISD::InputArg> &Ins,
3057 DebugLoc dl, SelectionDAG &DAG,
3058 SmallVectorImpl<SDValue> &InVals) const {
3060 unsigned NumOps = Outs.size();
3062 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3063 bool isPPC64 = PtrVT == MVT::i64;
3064 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3066 MachineFunction &MF = DAG.getMachineFunction();
3068 // Mark this function as potentially containing a function that contains a
3069 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3070 // and restoring the callers stack pointer in this functions epilog. This is
3071 // done because by tail calling the called function might overwrite the value
3072 // in this function's (MF) stack pointer stack slot 0(SP).
3073 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
3074 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3076 unsigned nAltivecParamsAtEnd = 0;
3078 // Count how many bytes are to be pushed on the stack, including the linkage
3079 // area, and parameter passing area. We start with 24/48 bytes, which is
3080 // prereserved space for [SP][CR][LR][3 x unused].
3082 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3084 nAltivecParamsAtEnd);
3086 // Calculate by how many bytes the stack has to be adjusted in case of tail
3087 // call optimization.
3088 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3090 // To protect arguments on the stack from being clobbered in a tail call,
3091 // force all the loads to happen before doing any other lowering.
3093 Chain = DAG.getStackArgumentTokenFactor(Chain);
3095 // Adjust the stack pointer for the new arguments...
3096 // These operations are automatically eliminated by the prolog/epilog pass
3097 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3098 SDValue CallSeqStart = Chain;
3100 // Load the return address and frame pointer so it can be move somewhere else
3103 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3106 // Set up a copy of the stack pointer for use loading and storing any
3107 // arguments that may not fit in the registers available for argument
3111 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3113 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3115 // Figure out which arguments are going to go in registers, and which in
3116 // memory. Also, if this is a vararg function, floating point operations
3117 // must be stored to our stack, and loaded into integer regs as well, if
3118 // any integer regs are available for argument passing.
3119 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3120 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3122 static const unsigned GPR_32[] = { // 32-bit registers.
3123 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3124 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3126 static const unsigned GPR_64[] = { // 64-bit registers.
3127 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3128 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3130 static const unsigned *FPR = GetFPR();
3132 static const unsigned VR[] = {
3133 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3134 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3136 const unsigned NumGPRs = array_lengthof(GPR_32);
3137 const unsigned NumFPRs = 13;
3138 const unsigned NumVRs = array_lengthof(VR);
3140 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
3142 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3143 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3145 SmallVector<SDValue, 8> MemOpChains;
3146 for (unsigned i = 0; i != NumOps; ++i) {
3147 SDValue Arg = OutVals[i];
3148 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3150 // PtrOff will be used to store the current argument to the stack if a
3151 // register cannot be found for it.
3154 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3156 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3158 // On PPC64, promote integers to 64-bit values.
3159 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3160 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3161 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3162 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3165 // FIXME memcpy is used way more than necessary. Correctness first.
3166 if (Flags.isByVal()) {
3167 unsigned Size = Flags.getByValSize();
3168 if (Size==1 || Size==2) {
3169 // Very small objects are passed right-justified.
3170 // Everything else is passed left-justified.
3171 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3172 if (GPR_idx != NumGPRs) {
3173 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3174 MachinePointerInfo(), VT,
3176 MemOpChains.push_back(Load.getValue(1));
3177 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3179 ArgOffset += PtrByteSize;
3181 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3182 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3183 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3184 CallSeqStart.getNode()->getOperand(0),
3186 // This must go outside the CALLSEQ_START..END.
3187 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3188 CallSeqStart.getNode()->getOperand(1));
3189 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3190 NewCallSeqStart.getNode());
3191 Chain = CallSeqStart = NewCallSeqStart;
3192 ArgOffset += PtrByteSize;
3196 // Copy entire object into memory. There are cases where gcc-generated
3197 // code assumes it is there, even if it could be put entirely into
3198 // registers. (This is not what the doc says.)
3199 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3200 CallSeqStart.getNode()->getOperand(0),
3202 // This must go outside the CALLSEQ_START..END.
3203 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3204 CallSeqStart.getNode()->getOperand(1));
3205 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3206 Chain = CallSeqStart = NewCallSeqStart;
3207 // And copy the pieces of it that fit into registers.
3208 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3209 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3210 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3211 if (GPR_idx != NumGPRs) {
3212 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3213 MachinePointerInfo(),
3215 MemOpChains.push_back(Load.getValue(1));
3216 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3217 ArgOffset += PtrByteSize;
3219 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3226 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3227 default: llvm_unreachable("Unexpected ValueType for argument!");
3230 if (GPR_idx != NumGPRs) {
3231 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3233 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3234 isPPC64, isTailCall, false, MemOpChains,
3235 TailCallArguments, dl);
3237 ArgOffset += PtrByteSize;
3241 if (FPR_idx != NumFPRs) {
3242 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3245 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3246 MachinePointerInfo(), false, false, 0);
3247 MemOpChains.push_back(Store);
3249 // Float varargs are always shadowed in available integer registers
3250 if (GPR_idx != NumGPRs) {
3251 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3252 MachinePointerInfo(), false, false, 0);
3253 MemOpChains.push_back(Load.getValue(1));
3254 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3256 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3257 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3258 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3259 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3260 MachinePointerInfo(),
3262 MemOpChains.push_back(Load.getValue(1));
3263 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3266 // If we have any FPRs remaining, we may also have GPRs remaining.
3267 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3269 if (GPR_idx != NumGPRs)
3271 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3272 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3276 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3277 isPPC64, isTailCall, false, MemOpChains,
3278 TailCallArguments, dl);
3283 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3290 // These go aligned on the stack, or in the corresponding R registers
3291 // when within range. The Darwin PPC ABI doc claims they also go in
3292 // V registers; in fact gcc does this only for arguments that are
3293 // prototyped, not for those that match the ... We do it for all
3294 // arguments, seems to work.
3295 while (ArgOffset % 16 !=0) {
3296 ArgOffset += PtrByteSize;
3297 if (GPR_idx != NumGPRs)
3300 // We could elide this store in the case where the object fits
3301 // entirely in R registers. Maybe later.
3302 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3303 DAG.getConstant(ArgOffset, PtrVT));
3304 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3305 MachinePointerInfo(), false, false, 0);
3306 MemOpChains.push_back(Store);
3307 if (VR_idx != NumVRs) {
3308 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3309 MachinePointerInfo(),
3311 MemOpChains.push_back(Load.getValue(1));
3312 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3315 for (unsigned i=0; i<16; i+=PtrByteSize) {
3316 if (GPR_idx == NumGPRs)
3318 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3319 DAG.getConstant(i, PtrVT));
3320 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3322 MemOpChains.push_back(Load.getValue(1));
3323 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3328 // Non-varargs Altivec params generally go in registers, but have
3329 // stack space allocated at the end.
3330 if (VR_idx != NumVRs) {
3331 // Doesn't have GPR space allocated.
3332 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3333 } else if (nAltivecParamsAtEnd==0) {
3334 // We are emitting Altivec params in order.
3335 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3336 isPPC64, isTailCall, true, MemOpChains,
3337 TailCallArguments, dl);
3343 // If all Altivec parameters fit in registers, as they usually do,
3344 // they get stack space following the non-Altivec parameters. We
3345 // don't track this here because nobody below needs it.
3346 // If there are more Altivec parameters than fit in registers emit
3348 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3350 // Offset is aligned; skip 1st 12 params which go in V registers.
3351 ArgOffset = ((ArgOffset+15)/16)*16;
3353 for (unsigned i = 0; i != NumOps; ++i) {
3354 SDValue Arg = OutVals[i];
3355 EVT ArgType = Outs[i].VT;
3356 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3357 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3360 // We are emitting Altivec params in order.
3361 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3362 isPPC64, isTailCall, true, MemOpChains,
3363 TailCallArguments, dl);
3370 if (!MemOpChains.empty())
3371 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3372 &MemOpChains[0], MemOpChains.size());
3374 // Check if this is an indirect call (MTCTR/BCTRL).
3375 // See PrepareCall() for more information about calls through function
3376 // pointers in the 64-bit SVR4 ABI.
3377 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3378 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3379 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3380 !isBLACompatibleAddress(Callee, DAG)) {
3381 // Load r2 into a virtual register and store it to the TOC save area.
3382 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3383 // TOC save area offset.
3384 SDValue PtrOff = DAG.getIntPtrConstant(40);
3385 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3386 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3390 // On Darwin, R12 must contain the address of an indirect callee. This does
3391 // not mean the MTCTR instruction must use R12; it's easier to model this as
3392 // an extra parameter, so do that.
3394 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3395 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3396 !isBLACompatibleAddress(Callee, DAG))
3397 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3398 PPC::R12), Callee));
3400 // Build a sequence of copy-to-reg nodes chained together with token chain
3401 // and flag operands which copy the outgoing args into the appropriate regs.
3403 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3404 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3405 RegsToPass[i].second, InFlag);
3406 InFlag = Chain.getValue(1);
3410 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3411 FPOp, true, TailCallArguments);
3413 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3414 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3419 PPCTargetLowering::LowerReturn(SDValue Chain,
3420 CallingConv::ID CallConv, bool isVarArg,
3421 const SmallVectorImpl<ISD::OutputArg> &Outs,
3422 const SmallVectorImpl<SDValue> &OutVals,
3423 DebugLoc dl, SelectionDAG &DAG) const {
3425 SmallVector<CCValAssign, 16> RVLocs;
3426 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3427 getTargetMachine(), RVLocs, *DAG.getContext());
3428 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3430 // If this is the first return lowered for this function, add the regs to the
3431 // liveout set for the function.
3432 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3433 for (unsigned i = 0; i != RVLocs.size(); ++i)
3434 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3439 // Copy the result values into the output registers.
3440 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3441 CCValAssign &VA = RVLocs[i];
3442 assert(VA.isRegLoc() && "Can only return in registers!");
3443 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3445 Flag = Chain.getValue(1);
3449 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3451 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3454 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3455 const PPCSubtarget &Subtarget) const {
3456 // When we pop the dynamic allocation we need to restore the SP link.
3457 DebugLoc dl = Op.getDebugLoc();
3459 // Get the corect type for pointers.
3460 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3462 // Construct the stack pointer operand.
3463 bool isPPC64 = Subtarget.isPPC64();
3464 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3465 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3467 // Get the operands for the STACKRESTORE.
3468 SDValue Chain = Op.getOperand(0);
3469 SDValue SaveSP = Op.getOperand(1);
3471 // Load the old link SP.
3472 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3473 MachinePointerInfo(),
3476 // Restore the stack pointer.
3477 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3479 // Store the old link SP.
3480 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3487 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3488 MachineFunction &MF = DAG.getMachineFunction();
3489 bool isPPC64 = PPCSubTarget.isPPC64();
3490 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3491 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3493 // Get current frame pointer save index. The users of this index will be
3494 // primarily DYNALLOC instructions.
3495 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3496 int RASI = FI->getReturnAddrSaveIndex();
3498 // If the frame pointer save index hasn't been defined yet.
3500 // Find out what the fix offset of the frame pointer save area.
3501 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3502 // Allocate the frame index for frame pointer save area.
3503 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3505 FI->setReturnAddrSaveIndex(RASI);
3507 return DAG.getFrameIndex(RASI, PtrVT);
3511 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3512 MachineFunction &MF = DAG.getMachineFunction();
3513 bool isPPC64 = PPCSubTarget.isPPC64();
3514 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3515 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3517 // Get current frame pointer save index. The users of this index will be
3518 // primarily DYNALLOC instructions.
3519 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3520 int FPSI = FI->getFramePointerSaveIndex();
3522 // If the frame pointer save index hasn't been defined yet.
3524 // Find out what the fix offset of the frame pointer save area.
3525 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3528 // Allocate the frame index for frame pointer save area.
3529 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3531 FI->setFramePointerSaveIndex(FPSI);
3533 return DAG.getFrameIndex(FPSI, PtrVT);
3536 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3538 const PPCSubtarget &Subtarget) const {
3540 SDValue Chain = Op.getOperand(0);
3541 SDValue Size = Op.getOperand(1);
3542 DebugLoc dl = Op.getDebugLoc();
3544 // Get the corect type for pointers.
3545 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3547 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3548 DAG.getConstant(0, PtrVT), Size);
3549 // Construct a node for the frame pointer save index.
3550 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3551 // Build a DYNALLOC node.
3552 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3553 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3554 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3557 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3559 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3560 // Not FP? Not a fsel.
3561 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3562 !Op.getOperand(2).getValueType().isFloatingPoint())
3565 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3567 // Cannot handle SETEQ/SETNE.
3568 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3570 EVT ResVT = Op.getValueType();
3571 EVT CmpVT = Op.getOperand(0).getValueType();
3572 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3573 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3574 DebugLoc dl = Op.getDebugLoc();
3576 // If the RHS of the comparison is a 0.0, we don't need to do the
3577 // subtraction at all.
3578 if (isFloatingPointZero(RHS))
3580 default: break; // SETUO etc aren't handled by fsel.
3583 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3586 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3587 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3588 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3591 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3594 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3595 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3596 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3597 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3602 default: break; // SETUO etc aren't handled by fsel.
3605 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3606 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3607 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3608 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3611 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3612 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3613 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3614 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3617 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3618 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3619 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3620 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3623 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3624 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3625 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3626 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3631 // FIXME: Split this code up when LegalizeDAGTypes lands.
3632 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3633 DebugLoc dl) const {
3634 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3635 SDValue Src = Op.getOperand(0);
3636 if (Src.getValueType() == MVT::f32)
3637 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3640 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3641 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3643 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3648 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3652 // Convert the FP value to an int value through memory.
3653 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3655 // Emit a store to the stack slot.
3656 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3657 MachinePointerInfo(), false, false, 0);
3659 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3661 if (Op.getValueType() == MVT::i32)
3662 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3663 DAG.getConstant(4, FIPtr.getValueType()));
3664 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3668 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3669 SelectionDAG &DAG) const {
3670 DebugLoc dl = Op.getDebugLoc();
3671 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3672 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3675 if (Op.getOperand(0).getValueType() == MVT::i64) {
3676 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3677 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3678 if (Op.getValueType() == MVT::f32)
3679 FP = DAG.getNode(ISD::FP_ROUND, dl,
3680 MVT::f32, FP, DAG.getIntPtrConstant(0));
3684 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3685 "Unhandled SINT_TO_FP type in custom expander!");
3686 // Since we only generate this in 64-bit mode, we can take advantage of
3687 // 64-bit registers. In particular, sign extend the input value into the
3688 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3689 // then lfd it and fcfid it.
3690 MachineFunction &MF = DAG.getMachineFunction();
3691 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3692 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3693 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3694 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3696 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3699 // STD the extended value into the stack slot.
3700 MachineMemOperand *MMO =
3701 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3702 MachineMemOperand::MOStore, 8, 8);
3703 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3705 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3706 Ops, 4, MVT::i64, MMO);
3707 // Load the value as a double.
3708 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3711 // FCFID it and return it.
3712 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3713 if (Op.getValueType() == MVT::f32)
3714 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3718 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3719 SelectionDAG &DAG) const {
3720 DebugLoc dl = Op.getDebugLoc();
3722 The rounding mode is in bits 30:31 of FPSR, and has the following
3729 FLT_ROUNDS, on the other hand, expects the following:
3736 To perform the conversion, we do:
3737 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3740 MachineFunction &MF = DAG.getMachineFunction();
3741 EVT VT = Op.getValueType();
3742 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3743 std::vector<EVT> NodeTys;
3744 SDValue MFFSreg, InFlag;
3746 // Save FP Control Word to register
3747 NodeTys.push_back(MVT::f64); // return register
3748 NodeTys.push_back(MVT::Glue); // unused in this context
3749 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3751 // Save FP register to stack slot
3752 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3753 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3754 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3755 StackSlot, MachinePointerInfo(), false, false,0);
3757 // Load FP Control Word from low 32 bits of stack slot.
3758 SDValue Four = DAG.getConstant(4, PtrVT);
3759 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3760 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3763 // Transform as necessary
3765 DAG.getNode(ISD::AND, dl, MVT::i32,
3766 CWD, DAG.getConstant(3, MVT::i32));
3768 DAG.getNode(ISD::SRL, dl, MVT::i32,
3769 DAG.getNode(ISD::AND, dl, MVT::i32,
3770 DAG.getNode(ISD::XOR, dl, MVT::i32,
3771 CWD, DAG.getConstant(3, MVT::i32)),
3772 DAG.getConstant(3, MVT::i32)),
3773 DAG.getConstant(1, MVT::i32));
3776 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3778 return DAG.getNode((VT.getSizeInBits() < 16 ?
3779 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3782 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3783 EVT VT = Op.getValueType();
3784 unsigned BitWidth = VT.getSizeInBits();
3785 DebugLoc dl = Op.getDebugLoc();
3786 assert(Op.getNumOperands() == 3 &&
3787 VT == Op.getOperand(1).getValueType() &&
3790 // Expand into a bunch of logical ops. Note that these ops
3791 // depend on the PPC behavior for oversized shift amounts.
3792 SDValue Lo = Op.getOperand(0);
3793 SDValue Hi = Op.getOperand(1);
3794 SDValue Amt = Op.getOperand(2);
3795 EVT AmtVT = Amt.getValueType();
3797 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3798 DAG.getConstant(BitWidth, AmtVT), Amt);
3799 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3800 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3801 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3802 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3803 DAG.getConstant(-BitWidth, AmtVT));
3804 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3805 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3806 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3807 SDValue OutOps[] = { OutLo, OutHi };
3808 return DAG.getMergeValues(OutOps, 2, dl);
3811 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3812 EVT VT = Op.getValueType();
3813 DebugLoc dl = Op.getDebugLoc();
3814 unsigned BitWidth = VT.getSizeInBits();
3815 assert(Op.getNumOperands() == 3 &&
3816 VT == Op.getOperand(1).getValueType() &&
3819 // Expand into a bunch of logical ops. Note that these ops
3820 // depend on the PPC behavior for oversized shift amounts.
3821 SDValue Lo = Op.getOperand(0);
3822 SDValue Hi = Op.getOperand(1);
3823 SDValue Amt = Op.getOperand(2);
3824 EVT AmtVT = Amt.getValueType();
3826 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3827 DAG.getConstant(BitWidth, AmtVT), Amt);
3828 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3829 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3830 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3831 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3832 DAG.getConstant(-BitWidth, AmtVT));
3833 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3834 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3835 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3836 SDValue OutOps[] = { OutLo, OutHi };
3837 return DAG.getMergeValues(OutOps, 2, dl);
3840 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3841 DebugLoc dl = Op.getDebugLoc();
3842 EVT VT = Op.getValueType();
3843 unsigned BitWidth = VT.getSizeInBits();
3844 assert(Op.getNumOperands() == 3 &&
3845 VT == Op.getOperand(1).getValueType() &&
3848 // Expand into a bunch of logical ops, followed by a select_cc.
3849 SDValue Lo = Op.getOperand(0);
3850 SDValue Hi = Op.getOperand(1);
3851 SDValue Amt = Op.getOperand(2);
3852 EVT AmtVT = Amt.getValueType();
3854 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3855 DAG.getConstant(BitWidth, AmtVT), Amt);
3856 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3857 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3858 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3859 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3860 DAG.getConstant(-BitWidth, AmtVT));
3861 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3862 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3863 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3864 Tmp4, Tmp6, ISD::SETLE);
3865 SDValue OutOps[] = { OutLo, OutHi };
3866 return DAG.getMergeValues(OutOps, 2, dl);
3869 //===----------------------------------------------------------------------===//
3870 // Vector related lowering.
3873 /// BuildSplatI - Build a canonical splati of Val with an element size of
3874 /// SplatSize. Cast the result to VT.
3875 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3876 SelectionDAG &DAG, DebugLoc dl) {
3877 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3879 static const EVT VTys[] = { // canonical VT to use for each size.
3880 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3883 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3885 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3889 EVT CanonicalVT = VTys[SplatSize-1];
3891 // Build a canonical splat for this value.
3892 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3893 SmallVector<SDValue, 8> Ops;
3894 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3895 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3896 &Ops[0], Ops.size());
3897 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3900 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3901 /// specified intrinsic ID.
3902 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3903 SelectionDAG &DAG, DebugLoc dl,
3904 EVT DestVT = MVT::Other) {
3905 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3906 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3907 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3910 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3911 /// specified intrinsic ID.
3912 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3913 SDValue Op2, SelectionDAG &DAG,
3914 DebugLoc dl, EVT DestVT = MVT::Other) {
3915 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3916 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3917 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3921 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3922 /// amount. The result has the specified value type.
3923 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3924 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3925 // Force LHS/RHS to be the right type.
3926 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3927 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3930 for (unsigned i = 0; i != 16; ++i)
3932 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3933 return DAG.getNode(ISD::BITCAST, dl, VT, T);
3936 // If this is a case we can't handle, return null and let the default
3937 // expansion code take care of it. If we CAN select this case, and if it
3938 // selects to a single instruction, return Op. Otherwise, if we can codegen
3939 // this case more efficiently than a constant pool load, lower it to the
3940 // sequence of ops that should be used.
3941 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3942 SelectionDAG &DAG) const {
3943 DebugLoc dl = Op.getDebugLoc();
3944 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3945 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3947 // Check if this is a splat of a constant value.
3948 APInt APSplatBits, APSplatUndef;
3949 unsigned SplatBitSize;
3951 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3952 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3955 unsigned SplatBits = APSplatBits.getZExtValue();
3956 unsigned SplatUndef = APSplatUndef.getZExtValue();
3957 unsigned SplatSize = SplatBitSize / 8;
3959 // First, handle single instruction cases.
3962 if (SplatBits == 0) {
3963 // Canonicalize all zero vectors to be v4i32.
3964 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3965 SDValue Z = DAG.getConstant(0, MVT::i32);
3966 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3967 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
3972 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3973 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3975 if (SextVal >= -16 && SextVal <= 15)
3976 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3979 // Two instruction sequences.
3981 // If this value is in the range [-32,30] and is even, use:
3982 // tmp = VSPLTI[bhw], result = add tmp, tmp
3983 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3984 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3985 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3986 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3989 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3990 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3992 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3993 // Make -1 and vspltisw -1:
3994 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3996 // Make the VSLW intrinsic, computing 0x8000_0000.
3997 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4000 // xor by OnesV to invert it.
4001 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4002 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4005 // Check to see if this is a wide variety of vsplti*, binop self cases.
4006 static const signed char SplatCsts[] = {
4007 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4008 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4011 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4012 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4013 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4014 int i = SplatCsts[idx];
4016 // Figure out what shift amount will be used by altivec if shifted by i in
4018 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4020 // vsplti + shl self.
4021 if (SextVal == (i << (int)TypeShiftAmt)) {
4022 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4023 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4024 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4025 Intrinsic::ppc_altivec_vslw
4027 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4028 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4031 // vsplti + srl self.
4032 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4033 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4034 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4035 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4036 Intrinsic::ppc_altivec_vsrw
4038 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4039 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4042 // vsplti + sra self.
4043 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4044 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4045 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4046 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4047 Intrinsic::ppc_altivec_vsraw
4049 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4050 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4053 // vsplti + rol self.
4054 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4055 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4056 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4057 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4058 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4059 Intrinsic::ppc_altivec_vrlw
4061 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4062 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4065 // t = vsplti c, result = vsldoi t, t, 1
4066 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4067 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4068 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4070 // t = vsplti c, result = vsldoi t, t, 2
4071 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4072 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4073 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4075 // t = vsplti c, result = vsldoi t, t, 3
4076 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4077 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4078 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4082 // Three instruction sequences.
4084 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4085 if (SextVal >= 0 && SextVal <= 31) {
4086 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4087 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4088 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4089 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4091 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4092 if (SextVal >= -31 && SextVal <= 0) {
4093 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4094 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4095 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4096 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4102 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4103 /// the specified operations to build the shuffle.
4104 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4105 SDValue RHS, SelectionDAG &DAG,
4107 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4108 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4109 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4112 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4124 if (OpNum == OP_COPY) {
4125 if (LHSID == (1*9+2)*9+3) return LHS;
4126 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4130 SDValue OpLHS, OpRHS;
4131 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4132 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4136 default: llvm_unreachable("Unknown i32 permute!");
4138 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4139 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4140 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4141 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4144 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4145 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4146 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4147 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4150 for (unsigned i = 0; i != 16; ++i)
4151 ShufIdxs[i] = (i&3)+0;
4154 for (unsigned i = 0; i != 16; ++i)
4155 ShufIdxs[i] = (i&3)+4;
4158 for (unsigned i = 0; i != 16; ++i)
4159 ShufIdxs[i] = (i&3)+8;
4162 for (unsigned i = 0; i != 16; ++i)
4163 ShufIdxs[i] = (i&3)+12;
4166 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4168 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4170 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4172 EVT VT = OpLHS.getValueType();
4173 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4174 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4175 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4176 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4179 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4180 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4181 /// return the code it can be lowered into. Worst case, it can always be
4182 /// lowered into a vperm.
4183 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4184 SelectionDAG &DAG) const {
4185 DebugLoc dl = Op.getDebugLoc();
4186 SDValue V1 = Op.getOperand(0);
4187 SDValue V2 = Op.getOperand(1);
4188 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4189 EVT VT = Op.getValueType();
4191 // Cases that are handled by instructions that take permute immediates
4192 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4193 // selected by the instruction selector.
4194 if (V2.getOpcode() == ISD::UNDEF) {
4195 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4196 PPC::isSplatShuffleMask(SVOp, 2) ||
4197 PPC::isSplatShuffleMask(SVOp, 4) ||
4198 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4199 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4200 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4201 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4202 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4203 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4204 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4205 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4206 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4211 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4212 // and produce a fixed permutation. If any of these match, do not lower to
4214 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4215 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4216 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4217 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4218 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4219 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4220 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4221 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4222 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4225 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4226 // perfect shuffle table to emit an optimal matching sequence.
4227 SmallVector<int, 16> PermMask;
4228 SVOp->getMask(PermMask);
4230 unsigned PFIndexes[4];
4231 bool isFourElementShuffle = true;
4232 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4233 unsigned EltNo = 8; // Start out undef.
4234 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4235 if (PermMask[i*4+j] < 0)
4236 continue; // Undef, ignore it.
4238 unsigned ByteSource = PermMask[i*4+j];
4239 if ((ByteSource & 3) != j) {
4240 isFourElementShuffle = false;
4245 EltNo = ByteSource/4;
4246 } else if (EltNo != ByteSource/4) {
4247 isFourElementShuffle = false;
4251 PFIndexes[i] = EltNo;
4254 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4255 // perfect shuffle vector to determine if it is cost effective to do this as
4256 // discrete instructions, or whether we should use a vperm.
4257 if (isFourElementShuffle) {
4258 // Compute the index in the perfect shuffle table.
4259 unsigned PFTableIndex =
4260 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4262 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4263 unsigned Cost = (PFEntry >> 30);
4265 // Determining when to avoid vperm is tricky. Many things affect the cost
4266 // of vperm, particularly how many times the perm mask needs to be computed.
4267 // For example, if the perm mask can be hoisted out of a loop or is already
4268 // used (perhaps because there are multiple permutes with the same shuffle
4269 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4270 // the loop requires an extra register.
4272 // As a compromise, we only emit discrete instructions if the shuffle can be
4273 // generated in 3 or fewer operations. When we have loop information
4274 // available, if this block is within a loop, we should avoid using vperm
4275 // for 3-operation perms and use a constant pool load instead.
4277 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4280 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4281 // vector that will get spilled to the constant pool.
4282 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4284 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4285 // that it is in input element units, not in bytes. Convert now.
4286 EVT EltVT = V1.getValueType().getVectorElementType();
4287 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4289 SmallVector<SDValue, 16> ResultMask;
4290 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4291 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4293 for (unsigned j = 0; j != BytesPerElement; ++j)
4294 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4298 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4299 &ResultMask[0], ResultMask.size());
4300 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4303 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4304 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4305 /// information about the intrinsic.
4306 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4308 unsigned IntrinsicID =
4309 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4312 switch (IntrinsicID) {
4313 default: return false;
4314 // Comparison predicates.
4315 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4316 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4317 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4318 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4319 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4320 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4321 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4322 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4323 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4324 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4325 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4326 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4327 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4329 // Normal Comparisons.
4330 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4331 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4332 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4333 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4334 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4335 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4336 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4337 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4338 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4339 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4340 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4341 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4342 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4347 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4348 /// lower, do it, otherwise return null.
4349 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4350 SelectionDAG &DAG) const {
4351 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4352 // opcode number of the comparison.
4353 DebugLoc dl = Op.getDebugLoc();
4356 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4357 return SDValue(); // Don't custom lower most intrinsics.
4359 // If this is a non-dot comparison, make the VCMP node and we are done.
4361 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4362 Op.getOperand(1), Op.getOperand(2),
4363 DAG.getConstant(CompareOpc, MVT::i32));
4364 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4367 // Create the PPCISD altivec 'dot' comparison node.
4369 Op.getOperand(2), // LHS
4370 Op.getOperand(3), // RHS
4371 DAG.getConstant(CompareOpc, MVT::i32)
4373 std::vector<EVT> VTs;
4374 VTs.push_back(Op.getOperand(2).getValueType());
4375 VTs.push_back(MVT::Glue);
4376 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4378 // Now that we have the comparison, emit a copy from the CR to a GPR.
4379 // This is flagged to the above dot comparison.
4380 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4381 DAG.getRegister(PPC::CR6, MVT::i32),
4382 CompNode.getValue(1));
4384 // Unpack the result based on how the target uses it.
4385 unsigned BitNo; // Bit # of CR6.
4386 bool InvertBit; // Invert result?
4387 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4388 default: // Can't happen, don't crash on invalid number though.
4389 case 0: // Return the value of the EQ bit of CR6.
4390 BitNo = 0; InvertBit = false;
4392 case 1: // Return the inverted value of the EQ bit of CR6.
4393 BitNo = 0; InvertBit = true;
4395 case 2: // Return the value of the LT bit of CR6.
4396 BitNo = 2; InvertBit = false;
4398 case 3: // Return the inverted value of the LT bit of CR6.
4399 BitNo = 2; InvertBit = true;
4403 // Shift the bit into the low position.
4404 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4405 DAG.getConstant(8-(3-BitNo), MVT::i32));
4407 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4408 DAG.getConstant(1, MVT::i32));
4410 // If we are supposed to, toggle the bit.
4412 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4413 DAG.getConstant(1, MVT::i32));
4417 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4418 SelectionDAG &DAG) const {
4419 DebugLoc dl = Op.getDebugLoc();
4420 // Create a stack slot that is 16-byte aligned.
4421 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4422 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4423 EVT PtrVT = getPointerTy();
4424 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4426 // Store the input value into Value#0 of the stack slot.
4427 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4428 Op.getOperand(0), FIdx, MachinePointerInfo(),
4431 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4435 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4436 DebugLoc dl = Op.getDebugLoc();
4437 if (Op.getValueType() == MVT::v4i32) {
4438 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4440 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4441 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4443 SDValue RHSSwap = // = vrlw RHS, 16
4444 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4446 // Shrinkify inputs to v8i16.
4447 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4448 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4449 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4451 // Low parts multiplied together, generating 32-bit results (we ignore the
4453 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4454 LHS, RHS, DAG, dl, MVT::v4i32);
4456 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4457 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4458 // Shift the high parts up 16 bits.
4459 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4461 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4462 } else if (Op.getValueType() == MVT::v8i16) {
4463 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4465 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4467 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4468 LHS, RHS, Zero, DAG, dl);
4469 } else if (Op.getValueType() == MVT::v16i8) {
4470 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4472 // Multiply the even 8-bit parts, producing 16-bit sums.
4473 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4474 LHS, RHS, DAG, dl, MVT::v8i16);
4475 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4477 // Multiply the odd 8-bit parts, producing 16-bit sums.
4478 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4479 LHS, RHS, DAG, dl, MVT::v8i16);
4480 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4482 // Merge the results together.
4484 for (unsigned i = 0; i != 8; ++i) {
4486 Ops[i*2+1] = 2*i+1+16;
4488 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4490 llvm_unreachable("Unknown mul to lower!");
4494 /// LowerOperation - Provide custom lowering hooks for some operations.
4496 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4497 switch (Op.getOpcode()) {
4498 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4499 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4500 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4501 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4502 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4503 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4504 case ISD::SETCC: return LowerSETCC(Op, DAG);
4505 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4506 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4508 return LowerVASTART(Op, DAG, PPCSubTarget);
4511 return LowerVAARG(Op, DAG, PPCSubTarget);
4513 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4514 case ISD::DYNAMIC_STACKALLOC:
4515 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4517 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4518 case ISD::FP_TO_UINT:
4519 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4521 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4522 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4524 // Lower 64-bit shifts.
4525 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4526 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4527 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4529 // Vector-related lowering.
4530 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4531 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4532 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4533 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4534 case ISD::MUL: return LowerMUL(Op, DAG);
4536 // Frame & Return address.
4537 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4538 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4543 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4544 SmallVectorImpl<SDValue>&Results,
4545 SelectionDAG &DAG) const {
4546 const TargetMachine &TM = getTargetMachine();
4547 DebugLoc dl = N->getDebugLoc();
4548 switch (N->getOpcode()) {
4550 assert(false && "Do not know how to custom type legalize this operation!");
4553 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4554 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4557 EVT VT = N->getValueType(0);
4559 if (VT == MVT::i64) {
4560 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4562 Results.push_back(NewNode);
4563 Results.push_back(NewNode.getValue(1));
4567 case ISD::FP_ROUND_INREG: {
4568 assert(N->getValueType(0) == MVT::ppcf128);
4569 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4570 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4571 MVT::f64, N->getOperand(0),
4572 DAG.getIntPtrConstant(0));
4573 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4574 MVT::f64, N->getOperand(0),
4575 DAG.getIntPtrConstant(1));
4577 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4578 // of the long double, and puts FPSCR back the way it was. We do not
4579 // actually model FPSCR.
4580 std::vector<EVT> NodeTys;
4581 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4583 NodeTys.push_back(MVT::f64); // Return register
4584 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4585 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4586 MFFSreg = Result.getValue(0);
4587 InFlag = Result.getValue(1);
4590 NodeTys.push_back(MVT::Glue); // Returns a flag
4591 Ops[0] = DAG.getConstant(31, MVT::i32);
4593 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4594 InFlag = Result.getValue(0);
4597 NodeTys.push_back(MVT::Glue); // Returns a flag
4598 Ops[0] = DAG.getConstant(30, MVT::i32);
4600 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4601 InFlag = Result.getValue(0);
4604 NodeTys.push_back(MVT::f64); // result of add
4605 NodeTys.push_back(MVT::Glue); // Returns a flag
4609 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4610 FPreg = Result.getValue(0);
4611 InFlag = Result.getValue(1);
4614 NodeTys.push_back(MVT::f64);
4615 Ops[0] = DAG.getConstant(1, MVT::i32);
4619 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4620 FPreg = Result.getValue(0);
4622 // We know the low half is about to be thrown away, so just use something
4624 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4628 case ISD::FP_TO_SINT:
4629 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4635 //===----------------------------------------------------------------------===//
4636 // Other Lowering Code
4637 //===----------------------------------------------------------------------===//
4640 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4641 bool is64bit, unsigned BinOpcode) const {
4642 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4643 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4645 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4646 MachineFunction *F = BB->getParent();
4647 MachineFunction::iterator It = BB;
4650 unsigned dest = MI->getOperand(0).getReg();
4651 unsigned ptrA = MI->getOperand(1).getReg();
4652 unsigned ptrB = MI->getOperand(2).getReg();
4653 unsigned incr = MI->getOperand(3).getReg();
4654 DebugLoc dl = MI->getDebugLoc();
4656 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4657 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4658 F->insert(It, loopMBB);
4659 F->insert(It, exitMBB);
4660 exitMBB->splice(exitMBB->begin(), BB,
4661 llvm::next(MachineBasicBlock::iterator(MI)),
4663 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4665 MachineRegisterInfo &RegInfo = F->getRegInfo();
4666 unsigned TmpReg = (!BinOpcode) ? incr :
4667 RegInfo.createVirtualRegister(
4668 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4669 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4673 // fallthrough --> loopMBB
4674 BB->addSuccessor(loopMBB);
4677 // l[wd]arx dest, ptr
4678 // add r0, dest, incr
4679 // st[wd]cx. r0, ptr
4681 // fallthrough --> exitMBB
4683 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4684 .addReg(ptrA).addReg(ptrB);
4686 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4687 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4688 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4689 BuildMI(BB, dl, TII->get(PPC::BCC))
4690 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4691 BB->addSuccessor(loopMBB);
4692 BB->addSuccessor(exitMBB);
4701 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4702 MachineBasicBlock *BB,
4703 bool is8bit, // operation
4704 unsigned BinOpcode) const {
4705 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4706 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4707 // In 64 bit mode we have to use 64 bits for addresses, even though the
4708 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4709 // registers without caring whether they're 32 or 64, but here we're
4710 // doing actual arithmetic on the addresses.
4711 bool is64bit = PPCSubTarget.isPPC64();
4712 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4714 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4715 MachineFunction *F = BB->getParent();
4716 MachineFunction::iterator It = BB;
4719 unsigned dest = MI->getOperand(0).getReg();
4720 unsigned ptrA = MI->getOperand(1).getReg();
4721 unsigned ptrB = MI->getOperand(2).getReg();
4722 unsigned incr = MI->getOperand(3).getReg();
4723 DebugLoc dl = MI->getDebugLoc();
4725 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4726 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4727 F->insert(It, loopMBB);
4728 F->insert(It, exitMBB);
4729 exitMBB->splice(exitMBB->begin(), BB,
4730 llvm::next(MachineBasicBlock::iterator(MI)),
4732 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4734 MachineRegisterInfo &RegInfo = F->getRegInfo();
4735 const TargetRegisterClass *RC =
4736 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4737 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4738 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4739 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4740 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4741 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4742 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4743 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4744 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4745 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4746 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4747 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4748 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4750 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4754 // fallthrough --> loopMBB
4755 BB->addSuccessor(loopMBB);
4757 // The 4-byte load must be aligned, while a char or short may be
4758 // anywhere in the word. Hence all this nasty bookkeeping code.
4759 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4760 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4761 // xori shift, shift1, 24 [16]
4762 // rlwinm ptr, ptr1, 0, 0, 29
4763 // slw incr2, incr, shift
4764 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4765 // slw mask, mask2, shift
4767 // lwarx tmpDest, ptr
4768 // add tmp, tmpDest, incr2
4769 // andc tmp2, tmpDest, mask
4770 // and tmp3, tmp, mask
4771 // or tmp4, tmp3, tmp2
4774 // fallthrough --> exitMBB
4775 // srw dest, tmpDest, shift
4776 if (ptrA != ZeroReg) {
4777 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4778 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4779 .addReg(ptrA).addReg(ptrB);
4783 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4784 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4785 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4786 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4788 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4789 .addReg(Ptr1Reg).addImm(0).addImm(61);
4791 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4792 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4793 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4794 .addReg(incr).addReg(ShiftReg);
4796 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4798 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4799 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4801 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4802 .addReg(Mask2Reg).addReg(ShiftReg);
4805 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4806 .addReg(ZeroReg).addReg(PtrReg);
4808 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4809 .addReg(Incr2Reg).addReg(TmpDestReg);
4810 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4811 .addReg(TmpDestReg).addReg(MaskReg);
4812 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4813 .addReg(TmpReg).addReg(MaskReg);
4814 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4815 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4816 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4817 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4818 BuildMI(BB, dl, TII->get(PPC::BCC))
4819 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4820 BB->addSuccessor(loopMBB);
4821 BB->addSuccessor(exitMBB);
4826 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4832 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4833 MachineBasicBlock *BB) const {
4834 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4836 // To "insert" these instructions we actually have to insert their
4837 // control-flow patterns.
4838 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4839 MachineFunction::iterator It = BB;
4842 MachineFunction *F = BB->getParent();
4844 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4845 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4846 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4847 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4848 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4850 // The incoming instruction knows the destination vreg to set, the
4851 // condition code register to branch on, the true/false values to
4852 // select between, and a branch opcode to use.
4857 // cmpTY ccX, r1, r2
4859 // fallthrough --> copy0MBB
4860 MachineBasicBlock *thisMBB = BB;
4861 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4862 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4863 unsigned SelectPred = MI->getOperand(4).getImm();
4864 DebugLoc dl = MI->getDebugLoc();
4865 F->insert(It, copy0MBB);
4866 F->insert(It, sinkMBB);
4868 // Transfer the remainder of BB and its successor edges to sinkMBB.
4869 sinkMBB->splice(sinkMBB->begin(), BB,
4870 llvm::next(MachineBasicBlock::iterator(MI)),
4872 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4874 // Next, add the true and fallthrough blocks as its successors.
4875 BB->addSuccessor(copy0MBB);
4876 BB->addSuccessor(sinkMBB);
4878 BuildMI(BB, dl, TII->get(PPC::BCC))
4879 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4882 // %FalseValue = ...
4883 // # fallthrough to sinkMBB
4886 // Update machine-CFG edges
4887 BB->addSuccessor(sinkMBB);
4890 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4893 BuildMI(*BB, BB->begin(), dl,
4894 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4895 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4896 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4898 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4899 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4900 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4901 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4902 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4903 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4904 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4905 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4907 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4908 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4909 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4910 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4911 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4912 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4913 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4914 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4916 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4917 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4918 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4919 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4920 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4921 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4922 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4923 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4925 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4926 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4927 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4928 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4929 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4930 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4931 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4932 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4934 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4935 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4936 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4937 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4938 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4939 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4940 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4941 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4943 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4944 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4945 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4946 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4947 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4948 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4949 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4950 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4952 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4953 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4954 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4955 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4956 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4957 BB = EmitAtomicBinary(MI, BB, false, 0);
4958 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4959 BB = EmitAtomicBinary(MI, BB, true, 0);
4961 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4962 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4963 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4965 unsigned dest = MI->getOperand(0).getReg();
4966 unsigned ptrA = MI->getOperand(1).getReg();
4967 unsigned ptrB = MI->getOperand(2).getReg();
4968 unsigned oldval = MI->getOperand(3).getReg();
4969 unsigned newval = MI->getOperand(4).getReg();
4970 DebugLoc dl = MI->getDebugLoc();
4972 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4973 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4974 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4975 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4976 F->insert(It, loop1MBB);
4977 F->insert(It, loop2MBB);
4978 F->insert(It, midMBB);
4979 F->insert(It, exitMBB);
4980 exitMBB->splice(exitMBB->begin(), BB,
4981 llvm::next(MachineBasicBlock::iterator(MI)),
4983 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4987 // fallthrough --> loopMBB
4988 BB->addSuccessor(loop1MBB);
4991 // l[wd]arx dest, ptr
4992 // cmp[wd] dest, oldval
4995 // st[wd]cx. newval, ptr
4999 // st[wd]cx. dest, ptr
5002 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5003 .addReg(ptrA).addReg(ptrB);
5004 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5005 .addReg(oldval).addReg(dest);
5006 BuildMI(BB, dl, TII->get(PPC::BCC))
5007 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5008 BB->addSuccessor(loop2MBB);
5009 BB->addSuccessor(midMBB);
5012 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5013 .addReg(newval).addReg(ptrA).addReg(ptrB);
5014 BuildMI(BB, dl, TII->get(PPC::BCC))
5015 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5016 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5017 BB->addSuccessor(loop1MBB);
5018 BB->addSuccessor(exitMBB);
5021 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5022 .addReg(dest).addReg(ptrA).addReg(ptrB);
5023 BB->addSuccessor(exitMBB);
5028 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5029 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5030 // We must use 64-bit registers for addresses when targeting 64-bit,
5031 // since we're actually doing arithmetic on them. Other registers
5033 bool is64bit = PPCSubTarget.isPPC64();
5034 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5036 unsigned dest = MI->getOperand(0).getReg();
5037 unsigned ptrA = MI->getOperand(1).getReg();
5038 unsigned ptrB = MI->getOperand(2).getReg();
5039 unsigned oldval = MI->getOperand(3).getReg();
5040 unsigned newval = MI->getOperand(4).getReg();
5041 DebugLoc dl = MI->getDebugLoc();
5043 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5044 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5045 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5046 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5047 F->insert(It, loop1MBB);
5048 F->insert(It, loop2MBB);
5049 F->insert(It, midMBB);
5050 F->insert(It, exitMBB);
5051 exitMBB->splice(exitMBB->begin(), BB,
5052 llvm::next(MachineBasicBlock::iterator(MI)),
5054 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5056 MachineRegisterInfo &RegInfo = F->getRegInfo();
5057 const TargetRegisterClass *RC =
5058 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5059 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5060 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5061 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5062 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5063 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5064 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5065 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5066 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5067 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5068 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5069 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5070 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5071 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5072 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5074 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5075 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5078 // fallthrough --> loopMBB
5079 BB->addSuccessor(loop1MBB);
5081 // The 4-byte load must be aligned, while a char or short may be
5082 // anywhere in the word. Hence all this nasty bookkeeping code.
5083 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5084 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5085 // xori shift, shift1, 24 [16]
5086 // rlwinm ptr, ptr1, 0, 0, 29
5087 // slw newval2, newval, shift
5088 // slw oldval2, oldval,shift
5089 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5090 // slw mask, mask2, shift
5091 // and newval3, newval2, mask
5092 // and oldval3, oldval2, mask
5094 // lwarx tmpDest, ptr
5095 // and tmp, tmpDest, mask
5096 // cmpw tmp, oldval3
5099 // andc tmp2, tmpDest, mask
5100 // or tmp4, tmp2, newval3
5105 // stwcx. tmpDest, ptr
5107 // srw dest, tmpDest, shift
5108 if (ptrA != ZeroReg) {
5109 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5110 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5111 .addReg(ptrA).addReg(ptrB);
5115 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5116 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5117 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5118 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5120 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5121 .addReg(Ptr1Reg).addImm(0).addImm(61);
5123 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5124 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5125 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5126 .addReg(newval).addReg(ShiftReg);
5127 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5128 .addReg(oldval).addReg(ShiftReg);
5130 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5132 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5133 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5134 .addReg(Mask3Reg).addImm(65535);
5136 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5137 .addReg(Mask2Reg).addReg(ShiftReg);
5138 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5139 .addReg(NewVal2Reg).addReg(MaskReg);
5140 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5141 .addReg(OldVal2Reg).addReg(MaskReg);
5144 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5145 .addReg(ZeroReg).addReg(PtrReg);
5146 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5147 .addReg(TmpDestReg).addReg(MaskReg);
5148 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5149 .addReg(TmpReg).addReg(OldVal3Reg);
5150 BuildMI(BB, dl, TII->get(PPC::BCC))
5151 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5152 BB->addSuccessor(loop2MBB);
5153 BB->addSuccessor(midMBB);
5156 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5157 .addReg(TmpDestReg).addReg(MaskReg);
5158 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5159 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5160 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5161 .addReg(ZeroReg).addReg(PtrReg);
5162 BuildMI(BB, dl, TII->get(PPC::BCC))
5163 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5164 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5165 BB->addSuccessor(loop1MBB);
5166 BB->addSuccessor(exitMBB);
5169 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5170 .addReg(ZeroReg).addReg(PtrReg);
5171 BB->addSuccessor(exitMBB);
5176 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5179 llvm_unreachable("Unexpected instr type to insert");
5182 MI->eraseFromParent(); // The pseudo instruction is gone now.
5186 //===----------------------------------------------------------------------===//
5187 // Target Optimization Hooks
5188 //===----------------------------------------------------------------------===//
5190 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5191 DAGCombinerInfo &DCI) const {
5192 const TargetMachine &TM = getTargetMachine();
5193 SelectionDAG &DAG = DCI.DAG;
5194 DebugLoc dl = N->getDebugLoc();
5195 switch (N->getOpcode()) {
5198 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5199 if (C->isNullValue()) // 0 << V -> 0.
5200 return N->getOperand(0);
5204 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5205 if (C->isNullValue()) // 0 >>u V -> 0.
5206 return N->getOperand(0);
5210 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5211 if (C->isNullValue() || // 0 >>s V -> 0.
5212 C->isAllOnesValue()) // -1 >>s V -> -1.
5213 return N->getOperand(0);
5217 case ISD::SINT_TO_FP:
5218 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5219 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5220 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5221 // We allow the src/dst to be either f32/f64, but the intermediate
5222 // type must be i64.
5223 if (N->getOperand(0).getValueType() == MVT::i64 &&
5224 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5225 SDValue Val = N->getOperand(0).getOperand(0);
5226 if (Val.getValueType() == MVT::f32) {
5227 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5228 DCI.AddToWorklist(Val.getNode());
5231 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5232 DCI.AddToWorklist(Val.getNode());
5233 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5234 DCI.AddToWorklist(Val.getNode());
5235 if (N->getValueType(0) == MVT::f32) {
5236 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5237 DAG.getIntPtrConstant(0));
5238 DCI.AddToWorklist(Val.getNode());
5241 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5242 // If the intermediate type is i32, we can avoid the load/store here
5249 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5250 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5251 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5252 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5253 N->getOperand(1).getValueType() == MVT::i32 &&
5254 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5255 SDValue Val = N->getOperand(1).getOperand(0);
5256 if (Val.getValueType() == MVT::f32) {
5257 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5258 DCI.AddToWorklist(Val.getNode());
5260 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5261 DCI.AddToWorklist(Val.getNode());
5263 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5264 N->getOperand(2), N->getOperand(3));
5265 DCI.AddToWorklist(Val.getNode());
5269 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5270 if (cast<StoreSDNode>(N)->isUnindexed() &&
5271 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5272 N->getOperand(1).getNode()->hasOneUse() &&
5273 (N->getOperand(1).getValueType() == MVT::i32 ||
5274 N->getOperand(1).getValueType() == MVT::i16)) {
5275 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5276 // Do an any-extend to 32-bits if this is a half-word input.
5277 if (BSwapOp.getValueType() == MVT::i16)
5278 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5281 N->getOperand(0), BSwapOp, N->getOperand(2),
5282 DAG.getValueType(N->getOperand(1).getValueType())
5285 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5286 Ops, array_lengthof(Ops),
5287 cast<StoreSDNode>(N)->getMemoryVT(),
5288 cast<StoreSDNode>(N)->getMemOperand());
5292 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5293 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5294 N->getOperand(0).hasOneUse() &&
5295 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5296 SDValue Load = N->getOperand(0);
5297 LoadSDNode *LD = cast<LoadSDNode>(Load);
5298 // Create the byte-swapping load.
5300 LD->getChain(), // Chain
5301 LD->getBasePtr(), // Ptr
5302 DAG.getValueType(N->getValueType(0)) // VT
5305 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5306 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5307 LD->getMemoryVT(), LD->getMemOperand());
5309 // If this is an i16 load, insert the truncate.
5310 SDValue ResVal = BSLoad;
5311 if (N->getValueType(0) == MVT::i16)
5312 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5314 // First, combine the bswap away. This makes the value produced by the
5316 DCI.CombineTo(N, ResVal);
5318 // Next, combine the load away, we give it a bogus result value but a real
5319 // chain result. The result value is dead because the bswap is dead.
5320 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5322 // Return N so it doesn't get rechecked!
5323 return SDValue(N, 0);
5327 case PPCISD::VCMP: {
5328 // If a VCMPo node already exists with exactly the same operands as this
5329 // node, use its result instead of this node (VCMPo computes both a CR6 and
5330 // a normal output).
5332 if (!N->getOperand(0).hasOneUse() &&
5333 !N->getOperand(1).hasOneUse() &&
5334 !N->getOperand(2).hasOneUse()) {
5336 // Scan all of the users of the LHS, looking for VCMPo's that match.
5337 SDNode *VCMPoNode = 0;
5339 SDNode *LHSN = N->getOperand(0).getNode();
5340 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5342 if (UI->getOpcode() == PPCISD::VCMPo &&
5343 UI->getOperand(1) == N->getOperand(1) &&
5344 UI->getOperand(2) == N->getOperand(2) &&
5345 UI->getOperand(0) == N->getOperand(0)) {
5350 // If there is no VCMPo node, or if the flag value has a single use, don't
5352 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5355 // Look at the (necessarily single) use of the flag value. If it has a
5356 // chain, this transformation is more complex. Note that multiple things
5357 // could use the value result, which we should ignore.
5358 SDNode *FlagUser = 0;
5359 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5360 FlagUser == 0; ++UI) {
5361 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5363 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5364 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5371 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5372 // give up for right now.
5373 if (FlagUser->getOpcode() == PPCISD::MFCR)
5374 return SDValue(VCMPoNode, 0);
5379 // If this is a branch on an altivec predicate comparison, lower this so
5380 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5381 // lowering is done pre-legalize, because the legalizer lowers the predicate
5382 // compare down to code that is difficult to reassemble.
5383 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5384 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5388 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5389 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5390 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5391 assert(isDot && "Can't compare against a vector result!");
5393 // If this is a comparison against something other than 0/1, then we know
5394 // that the condition is never/always true.
5395 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5396 if (Val != 0 && Val != 1) {
5397 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5398 return N->getOperand(0);
5399 // Always !=, turn it into an unconditional branch.
5400 return DAG.getNode(ISD::BR, dl, MVT::Other,
5401 N->getOperand(0), N->getOperand(4));
5404 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5406 // Create the PPCISD altivec 'dot' comparison node.
5407 std::vector<EVT> VTs;
5409 LHS.getOperand(2), // LHS of compare
5410 LHS.getOperand(3), // RHS of compare
5411 DAG.getConstant(CompareOpc, MVT::i32)
5413 VTs.push_back(LHS.getOperand(2).getValueType());
5414 VTs.push_back(MVT::Glue);
5415 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5417 // Unpack the result based on how the target uses it.
5418 PPC::Predicate CompOpc;
5419 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5420 default: // Can't happen, don't crash on invalid number though.
5421 case 0: // Branch on the value of the EQ bit of CR6.
5422 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5424 case 1: // Branch on the inverted value of the EQ bit of CR6.
5425 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5427 case 2: // Branch on the value of the LT bit of CR6.
5428 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5430 case 3: // Branch on the inverted value of the LT bit of CR6.
5431 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5435 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5436 DAG.getConstant(CompOpc, MVT::i32),
5437 DAG.getRegister(PPC::CR6, MVT::i32),
5438 N->getOperand(4), CompNode.getValue(1));
5447 //===----------------------------------------------------------------------===//
5448 // Inline Assembly Support
5449 //===----------------------------------------------------------------------===//
5451 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5455 const SelectionDAG &DAG,
5456 unsigned Depth) const {
5457 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5458 switch (Op.getOpcode()) {
5460 case PPCISD::LBRX: {
5461 // lhbrx is known to have the top bits cleared out.
5462 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5463 KnownZero = 0xFFFF0000;
5466 case ISD::INTRINSIC_WO_CHAIN: {
5467 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5469 case Intrinsic::ppc_altivec_vcmpbfp_p:
5470 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5471 case Intrinsic::ppc_altivec_vcmpequb_p:
5472 case Intrinsic::ppc_altivec_vcmpequh_p:
5473 case Intrinsic::ppc_altivec_vcmpequw_p:
5474 case Intrinsic::ppc_altivec_vcmpgefp_p:
5475 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5476 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5477 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5478 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5479 case Intrinsic::ppc_altivec_vcmpgtub_p:
5480 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5481 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5482 KnownZero = ~1U; // All bits but the low one are known to be zero.
5490 /// getConstraintType - Given a constraint, return the type of
5491 /// constraint it is for this target.
5492 PPCTargetLowering::ConstraintType
5493 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5494 if (Constraint.size() == 1) {
5495 switch (Constraint[0]) {
5502 return C_RegisterClass;
5505 return TargetLowering::getConstraintType(Constraint);
5508 /// Examine constraint type and operand type and determine a weight value.
5509 /// This object must already have been set up with the operand type
5510 /// and the current alternative constraint selected.
5511 TargetLowering::ConstraintWeight
5512 PPCTargetLowering::getSingleConstraintMatchWeight(
5513 AsmOperandInfo &info, const char *constraint) const {
5514 ConstraintWeight weight = CW_Invalid;
5515 Value *CallOperandVal = info.CallOperandVal;
5516 // If we don't have a value, we can't do a match,
5517 // but allow it at the lowest weight.
5518 if (CallOperandVal == NULL)
5520 Type *type = CallOperandVal->getType();
5521 // Look at the constraint type.
5522 switch (*constraint) {
5524 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5527 if (type->isIntegerTy())
5528 weight = CW_Register;
5531 if (type->isFloatTy())
5532 weight = CW_Register;
5535 if (type->isDoubleTy())
5536 weight = CW_Register;
5539 if (type->isVectorTy())
5540 weight = CW_Register;
5543 weight = CW_Register;
5549 std::pair<unsigned, const TargetRegisterClass*>
5550 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5552 if (Constraint.size() == 1) {
5553 // GCC RS6000 Constraint Letters
5554 switch (Constraint[0]) {
5557 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5558 return std::make_pair(0U, PPC::G8RCRegisterClass);
5559 return std::make_pair(0U, PPC::GPRCRegisterClass);
5562 return std::make_pair(0U, PPC::F4RCRegisterClass);
5563 else if (VT == MVT::f64)
5564 return std::make_pair(0U, PPC::F8RCRegisterClass);
5567 return std::make_pair(0U, PPC::VRRCRegisterClass);
5569 return std::make_pair(0U, PPC::CRRCRegisterClass);
5573 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5577 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5578 /// vector. If it is invalid, don't add anything to Ops.
5579 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5580 std::string &Constraint,
5581 std::vector<SDValue>&Ops,
5582 SelectionDAG &DAG) const {
5583 SDValue Result(0,0);
5585 // Only support length 1 constraints.
5586 if (Constraint.length() > 1) return;
5588 char Letter = Constraint[0];
5599 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5600 if (!CST) return; // Must be an immediate to match.
5601 unsigned Value = CST->getZExtValue();
5603 default: llvm_unreachable("Unknown constraint letter!");
5604 case 'I': // "I" is a signed 16-bit constant.
5605 if ((short)Value == (int)Value)
5606 Result = DAG.getTargetConstant(Value, Op.getValueType());
5608 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5609 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5610 if ((short)Value == 0)
5611 Result = DAG.getTargetConstant(Value, Op.getValueType());
5613 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5614 if ((Value >> 16) == 0)
5615 Result = DAG.getTargetConstant(Value, Op.getValueType());
5617 case 'M': // "M" is a constant that is greater than 31.
5619 Result = DAG.getTargetConstant(Value, Op.getValueType());
5621 case 'N': // "N" is a positive constant that is an exact power of two.
5622 if ((int)Value > 0 && isPowerOf2_32(Value))
5623 Result = DAG.getTargetConstant(Value, Op.getValueType());
5625 case 'O': // "O" is the constant zero.
5627 Result = DAG.getTargetConstant(Value, Op.getValueType());
5629 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5630 if ((short)-Value == (int)-Value)
5631 Result = DAG.getTargetConstant(Value, Op.getValueType());
5638 if (Result.getNode()) {
5639 Ops.push_back(Result);
5643 // Handle standard constraint letters.
5644 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5647 // isLegalAddressingMode - Return true if the addressing mode represented
5648 // by AM is legal for this target, for a load/store of the specified type.
5649 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5651 // FIXME: PPC does not allow r+i addressing modes for vectors!
5653 // PPC allows a sign-extended 16-bit immediate field.
5654 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5657 // No global is ever allowed as a base.
5661 // PPC only support r+r,
5663 case 0: // "r+i" or just "i", depending on HasBaseReg.
5666 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5668 // Otherwise we have r+r or r+i.
5671 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5673 // Allow 2*r as r+r.
5676 // No other scales are supported.
5683 /// isLegalAddressImmediate - Return true if the integer value can be used
5684 /// as the offset of the target addressing mode for load / store of the
5686 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5687 // PPC allows a sign-extended 16-bit immediate field.
5688 return (V > -(1 << 16) && V < (1 << 16)-1);
5691 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5695 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5696 SelectionDAG &DAG) const {
5697 MachineFunction &MF = DAG.getMachineFunction();
5698 MachineFrameInfo *MFI = MF.getFrameInfo();
5699 MFI->setReturnAddressIsTaken(true);
5701 DebugLoc dl = Op.getDebugLoc();
5702 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5704 // Make sure the function does not optimize away the store of the RA to
5706 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5707 FuncInfo->setLRStoreRequired();
5708 bool isPPC64 = PPCSubTarget.isPPC64();
5709 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5712 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5715 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5716 isPPC64? MVT::i64 : MVT::i32);
5717 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5718 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5720 MachinePointerInfo(), false, false, 0);
5723 // Just load the return address off the stack.
5724 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5725 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5726 RetAddrFI, MachinePointerInfo(), false, false, 0);
5729 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5730 SelectionDAG &DAG) const {
5731 DebugLoc dl = Op.getDebugLoc();
5732 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5734 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5735 bool isPPC64 = PtrVT == MVT::i64;
5737 MachineFunction &MF = DAG.getMachineFunction();
5738 MachineFrameInfo *MFI = MF.getFrameInfo();
5739 MFI->setFrameAddressIsTaken(true);
5740 bool is31 = (DisableFramePointerElim(MF) || MFI->hasVarSizedObjects()) &&
5741 MFI->getStackSize() &&
5742 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5743 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5744 (is31 ? PPC::R31 : PPC::R1);
5745 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5748 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5749 FrameAddr, MachinePointerInfo(), false, false, 0);
5754 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5755 // The PowerPC target isn't yet aware of offsets.
5759 /// getOptimalMemOpType - Returns the target specific optimal type for load
5760 /// and store operations as a result of memset, memcpy, and memmove
5761 /// lowering. If DstAlign is zero that means it's safe to destination
5762 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5763 /// means there isn't a need to check it against alignment requirement,
5764 /// probably because the source does not need to be loaded. If
5765 /// 'NonScalarIntSafe' is true, that means it's safe to return a
5766 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5767 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5768 /// constant so it does not need to be loaded.
5769 /// It returns EVT::Other if the type should be determined using generic
5770 /// target-independent logic.
5771 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5772 unsigned DstAlign, unsigned SrcAlign,
5773 bool NonScalarIntSafe,
5775 MachineFunction &MF) const {
5776 if (this->PPCSubTarget.isPPC64()) {