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/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
27 #include "llvm/CallingConv.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/DerivedTypes.h"
39 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
40 CCValAssign::LocInfo &LocInfo,
41 ISD::ArgFlagsTy &ArgFlags,
43 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
54 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
55 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
58 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
59 if (TM.getSubtargetImpl()->isDarwin())
60 return new TargetLoweringObjectFileMachO();
62 return new TargetLoweringObjectFileELF();
65 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
66 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
70 // Use _setjmp/_longjmp instead of setjmp/longjmp.
71 setUseUnderscoreSetJmp(true);
72 setUseUnderscoreLongJmp(true);
74 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
75 // arguments are at least 4/8 bytes aligned.
76 setMinStackArgumentAlignment(TM.getSubtarget<PPCSubtarget>().isPPC64() ? 8:4);
78 // Set up the register classes.
79 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
80 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
81 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
83 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
84 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
85 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
87 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
89 // PowerPC has pre-inc load and store's.
90 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
91 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
93 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
95 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
96 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
98 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
101 // This is used in the ppcf128->int sequence. Note it has different semantics
102 // from FP_ROUND: that rounds to nearest, this rounds to zero.
103 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
105 // We do not currently implment this libm ops for PowerPC.
106 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
107 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
108 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
109 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
110 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
112 // PowerPC has no SREM/UREM instructions
113 setOperationAction(ISD::SREM, MVT::i32, Expand);
114 setOperationAction(ISD::UREM, MVT::i32, Expand);
115 setOperationAction(ISD::SREM, MVT::i64, Expand);
116 setOperationAction(ISD::UREM, MVT::i64, Expand);
118 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
119 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
120 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
121 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
122 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
123 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
124 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
125 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
126 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
128 // We don't support sin/cos/sqrt/fmod/pow
129 setOperationAction(ISD::FSIN , MVT::f64, Expand);
130 setOperationAction(ISD::FCOS , MVT::f64, Expand);
131 setOperationAction(ISD::FREM , MVT::f64, Expand);
132 setOperationAction(ISD::FPOW , MVT::f64, Expand);
133 setOperationAction(ISD::FMA , MVT::f64, Expand);
134 setOperationAction(ISD::FSIN , MVT::f32, Expand);
135 setOperationAction(ISD::FCOS , MVT::f32, Expand);
136 setOperationAction(ISD::FREM , MVT::f32, Expand);
137 setOperationAction(ISD::FPOW , MVT::f32, Expand);
138 setOperationAction(ISD::FMA , MVT::f32, Expand);
140 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
142 // If we're enabling GP optimizations, use hardware square root
143 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
144 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
145 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
148 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
149 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
151 // PowerPC does not have BSWAP, CTPOP or CTTZ
152 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
153 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
154 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
155 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
156 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
157 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
158 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
159 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
161 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
163 // PowerPC does not have ROTR
164 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
165 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
167 // PowerPC does not have Select
168 setOperationAction(ISD::SELECT, MVT::i32, Expand);
169 setOperationAction(ISD::SELECT, MVT::i64, Expand);
170 setOperationAction(ISD::SELECT, MVT::f32, Expand);
171 setOperationAction(ISD::SELECT, MVT::f64, Expand);
173 // PowerPC wants to turn select_cc of FP into fsel when possible.
174 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
175 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
177 // PowerPC wants to optimize integer setcc a bit
178 setOperationAction(ISD::SETCC, MVT::i32, Custom);
180 // PowerPC does not have BRCOND which requires SetCC
181 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
183 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
185 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
186 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
188 // PowerPC does not have [U|S]INT_TO_FP
189 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
190 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
192 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
193 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
194 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
195 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
197 // We cannot sextinreg(i1). Expand to shifts.
198 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
200 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
201 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
202 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
203 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
206 // We want to legalize GlobalAddress and ConstantPool nodes into the
207 // appropriate instructions to materialize the address.
208 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
209 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
210 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
211 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
212 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
213 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
214 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
215 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
216 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
217 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
220 setOperationAction(ISD::TRAP, MVT::Other, Legal);
222 // TRAMPOLINE is custom lowered.
223 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
224 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
226 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
227 setOperationAction(ISD::VASTART , MVT::Other, Custom);
229 // VAARG is custom lowered with the 32-bit SVR4 ABI.
230 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
231 && !TM.getSubtarget<PPCSubtarget>().isPPC64()) {
232 setOperationAction(ISD::VAARG, MVT::Other, Custom);
233 setOperationAction(ISD::VAARG, MVT::i64, Custom);
235 setOperationAction(ISD::VAARG, MVT::Other, Expand);
237 // Use the default implementation.
238 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
239 setOperationAction(ISD::VAEND , MVT::Other, Expand);
240 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
241 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
242 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
243 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
245 // We want to custom lower some of our intrinsics.
246 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
248 // Comparisons that require checking two conditions.
249 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
250 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
251 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
252 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
253 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
254 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
255 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
256 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
257 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
258 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
259 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
260 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
262 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
263 // They also have instructions for converting between i64 and fp.
264 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
265 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
266 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
267 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
268 // This is just the low 32 bits of a (signed) fp->i64 conversion.
269 // We cannot do this with Promote because i64 is not a legal type.
270 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
272 // FIXME: disable this lowered code. This generates 64-bit register values,
273 // and we don't model the fact that the top part is clobbered by calls. We
274 // need to flag these together so that the value isn't live across a call.
275 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
277 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
278 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
281 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
282 // 64-bit PowerPC implementations can support i64 types directly
283 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
284 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
285 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
286 // 64-bit PowerPC wants to expand i128 shifts itself.
287 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
288 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
289 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
291 // 32-bit PowerPC wants to expand i64 shifts itself.
292 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
293 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
294 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
297 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
298 // First set operation action for all vector types to expand. Then we
299 // will selectively turn on ones that can be effectively codegen'd.
300 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
301 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
302 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
304 // add/sub are legal for all supported vector VT's.
305 setOperationAction(ISD::ADD , VT, Legal);
306 setOperationAction(ISD::SUB , VT, Legal);
308 // We promote all shuffles to v16i8.
309 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
310 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
312 // We promote all non-typed operations to v4i32.
313 setOperationAction(ISD::AND , VT, Promote);
314 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
315 setOperationAction(ISD::OR , VT, Promote);
316 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
317 setOperationAction(ISD::XOR , VT, Promote);
318 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
319 setOperationAction(ISD::LOAD , VT, Promote);
320 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
321 setOperationAction(ISD::SELECT, VT, Promote);
322 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
323 setOperationAction(ISD::STORE, VT, Promote);
324 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
326 // No other operations are legal.
327 setOperationAction(ISD::MUL , VT, Expand);
328 setOperationAction(ISD::SDIV, VT, Expand);
329 setOperationAction(ISD::SREM, VT, Expand);
330 setOperationAction(ISD::UDIV, VT, Expand);
331 setOperationAction(ISD::UREM, VT, Expand);
332 setOperationAction(ISD::FDIV, VT, Expand);
333 setOperationAction(ISD::FNEG, VT, Expand);
334 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
335 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
336 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
337 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
338 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
339 setOperationAction(ISD::UDIVREM, VT, Expand);
340 setOperationAction(ISD::SDIVREM, VT, Expand);
341 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
342 setOperationAction(ISD::FPOW, VT, Expand);
343 setOperationAction(ISD::CTPOP, VT, Expand);
344 setOperationAction(ISD::CTLZ, VT, Expand);
345 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
346 setOperationAction(ISD::CTTZ, VT, Expand);
347 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
350 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
351 // with merges, splats, etc.
352 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
354 setOperationAction(ISD::AND , MVT::v4i32, Legal);
355 setOperationAction(ISD::OR , MVT::v4i32, Legal);
356 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
357 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
358 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
359 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
361 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
362 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
363 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
364 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
366 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
367 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
368 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
369 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
371 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
372 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
374 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
375 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
376 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
377 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
381 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
383 setBooleanContents(ZeroOrOneBooleanContent);
384 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
386 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
387 setStackPointerRegisterToSaveRestore(PPC::X1);
388 setExceptionPointerRegister(PPC::X3);
389 setExceptionSelectorRegister(PPC::X4);
391 setStackPointerRegisterToSaveRestore(PPC::R1);
392 setExceptionPointerRegister(PPC::R3);
393 setExceptionSelectorRegister(PPC::R4);
396 // We have target-specific dag combine patterns for the following nodes:
397 setTargetDAGCombine(ISD::SINT_TO_FP);
398 setTargetDAGCombine(ISD::STORE);
399 setTargetDAGCombine(ISD::BR_CC);
400 setTargetDAGCombine(ISD::BSWAP);
402 // Darwin long double math library functions have $LDBL128 appended.
403 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
404 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
405 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
406 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
407 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
408 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
409 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
410 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
411 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
412 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
413 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
416 setMinFunctionAlignment(2);
417 if (PPCSubTarget.isDarwin())
418 setPrefFunctionAlignment(4);
420 setInsertFencesForAtomic(true);
422 setSchedulingPreference(Sched::Hybrid);
424 computeRegisterProperties();
427 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
428 /// function arguments in the caller parameter area.
429 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
430 const TargetMachine &TM = getTargetMachine();
431 // Darwin passes everything on 4 byte boundary.
432 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
438 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
441 case PPCISD::FSEL: return "PPCISD::FSEL";
442 case PPCISD::FCFID: return "PPCISD::FCFID";
443 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
444 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
445 case PPCISD::STFIWX: return "PPCISD::STFIWX";
446 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
447 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
448 case PPCISD::VPERM: return "PPCISD::VPERM";
449 case PPCISD::Hi: return "PPCISD::Hi";
450 case PPCISD::Lo: return "PPCISD::Lo";
451 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
452 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
453 case PPCISD::LOAD: return "PPCISD::LOAD";
454 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
455 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
456 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
457 case PPCISD::SRL: return "PPCISD::SRL";
458 case PPCISD::SRA: return "PPCISD::SRA";
459 case PPCISD::SHL: return "PPCISD::SHL";
460 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
461 case PPCISD::STD_32: return "PPCISD::STD_32";
462 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
463 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
464 case PPCISD::NOP: return "PPCISD::NOP";
465 case PPCISD::MTCTR: return "PPCISD::MTCTR";
466 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
467 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
468 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
469 case PPCISD::MFCR: return "PPCISD::MFCR";
470 case PPCISD::VCMP: return "PPCISD::VCMP";
471 case PPCISD::VCMPo: return "PPCISD::VCMPo";
472 case PPCISD::LBRX: return "PPCISD::LBRX";
473 case PPCISD::STBRX: return "PPCISD::STBRX";
474 case PPCISD::LARX: return "PPCISD::LARX";
475 case PPCISD::STCX: return "PPCISD::STCX";
476 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
477 case PPCISD::MFFS: return "PPCISD::MFFS";
478 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
479 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
480 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
481 case PPCISD::MTFSF: return "PPCISD::MTFSF";
482 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
486 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
490 //===----------------------------------------------------------------------===//
491 // Node matching predicates, for use by the tblgen matching code.
492 //===----------------------------------------------------------------------===//
494 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
495 static bool isFloatingPointZero(SDValue Op) {
496 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
497 return CFP->getValueAPF().isZero();
498 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
499 // Maybe this has already been legalized into the constant pool?
500 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
501 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
502 return CFP->getValueAPF().isZero();
507 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
508 /// true if Op is undef or if it matches the specified value.
509 static bool isConstantOrUndef(int Op, int Val) {
510 return Op < 0 || Op == Val;
513 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
514 /// VPKUHUM instruction.
515 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
517 for (unsigned i = 0; i != 16; ++i)
518 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
521 for (unsigned i = 0; i != 8; ++i)
522 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
523 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
529 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
530 /// VPKUWUM instruction.
531 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
533 for (unsigned i = 0; i != 16; i += 2)
534 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
535 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
538 for (unsigned i = 0; i != 8; i += 2)
539 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
540 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
541 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
542 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
548 /// isVMerge - Common function, used to match vmrg* shuffles.
550 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
551 unsigned LHSStart, unsigned RHSStart) {
552 assert(N->getValueType(0) == MVT::v16i8 &&
553 "PPC only supports shuffles by bytes!");
554 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
555 "Unsupported merge size!");
557 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
558 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
559 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
560 LHSStart+j+i*UnitSize) ||
561 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
562 RHSStart+j+i*UnitSize))
568 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
569 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
570 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
573 return isVMerge(N, UnitSize, 8, 24);
574 return isVMerge(N, UnitSize, 8, 8);
577 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
578 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
579 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
582 return isVMerge(N, UnitSize, 0, 16);
583 return isVMerge(N, UnitSize, 0, 0);
587 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
588 /// amount, otherwise return -1.
589 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
590 assert(N->getValueType(0) == MVT::v16i8 &&
591 "PPC only supports shuffles by bytes!");
593 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
595 // Find the first non-undef value in the shuffle mask.
597 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
600 if (i == 16) return -1; // all undef.
602 // Otherwise, check to see if the rest of the elements are consecutively
603 // numbered from this value.
604 unsigned ShiftAmt = SVOp->getMaskElt(i);
605 if (ShiftAmt < i) return -1;
609 // Check the rest of the elements to see if they are consecutive.
610 for (++i; i != 16; ++i)
611 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
614 // Check the rest of the elements to see if they are consecutive.
615 for (++i; i != 16; ++i)
616 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
622 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
623 /// specifies a splat of a single element that is suitable for input to
624 /// VSPLTB/VSPLTH/VSPLTW.
625 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
626 assert(N->getValueType(0) == MVT::v16i8 &&
627 (EltSize == 1 || EltSize == 2 || EltSize == 4));
629 // This is a splat operation if each element of the permute is the same, and
630 // if the value doesn't reference the second vector.
631 unsigned ElementBase = N->getMaskElt(0);
633 // FIXME: Handle UNDEF elements too!
634 if (ElementBase >= 16)
637 // Check that the indices are consecutive, in the case of a multi-byte element
638 // splatted with a v16i8 mask.
639 for (unsigned i = 1; i != EltSize; ++i)
640 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
643 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
644 if (N->getMaskElt(i) < 0) continue;
645 for (unsigned j = 0; j != EltSize; ++j)
646 if (N->getMaskElt(i+j) != N->getMaskElt(j))
652 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
654 bool PPC::isAllNegativeZeroVector(SDNode *N) {
655 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
657 APInt APVal, APUndef;
661 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
662 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
663 return CFP->getValueAPF().isNegZero();
668 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
669 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
670 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
671 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
672 assert(isSplatShuffleMask(SVOp, EltSize));
673 return SVOp->getMaskElt(0) / EltSize;
676 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
677 /// by using a vspltis[bhw] instruction of the specified element size, return
678 /// the constant being splatted. The ByteSize field indicates the number of
679 /// bytes of each element [124] -> [bhw].
680 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
683 // If ByteSize of the splat is bigger than the element size of the
684 // build_vector, then we have a case where we are checking for a splat where
685 // multiple elements of the buildvector are folded together into a single
686 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
687 unsigned EltSize = 16/N->getNumOperands();
688 if (EltSize < ByteSize) {
689 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
690 SDValue UniquedVals[4];
691 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
693 // See if all of the elements in the buildvector agree across.
694 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
695 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
696 // If the element isn't a constant, bail fully out.
697 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
700 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
701 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
702 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
703 return SDValue(); // no match.
706 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
707 // either constant or undef values that are identical for each chunk. See
708 // if these chunks can form into a larger vspltis*.
710 // Check to see if all of the leading entries are either 0 or -1. If
711 // neither, then this won't fit into the immediate field.
712 bool LeadingZero = true;
713 bool LeadingOnes = true;
714 for (unsigned i = 0; i != Multiple-1; ++i) {
715 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
717 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
718 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
720 // Finally, check the least significant entry.
722 if (UniquedVals[Multiple-1].getNode() == 0)
723 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
724 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
726 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
729 if (UniquedVals[Multiple-1].getNode() == 0)
730 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
731 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
732 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
733 return DAG.getTargetConstant(Val, MVT::i32);
739 // Check to see if this buildvec has a single non-undef value in its elements.
740 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
741 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
742 if (OpVal.getNode() == 0)
743 OpVal = N->getOperand(i);
744 else if (OpVal != N->getOperand(i))
748 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
750 unsigned ValSizeInBytes = EltSize;
752 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
753 Value = CN->getZExtValue();
754 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
755 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
756 Value = FloatToBits(CN->getValueAPF().convertToFloat());
759 // If the splat value is larger than the element value, then we can never do
760 // this splat. The only case that we could fit the replicated bits into our
761 // immediate field for would be zero, and we prefer to use vxor for it.
762 if (ValSizeInBytes < ByteSize) return SDValue();
764 // If the element value is larger than the splat value, cut it in half and
765 // check to see if the two halves are equal. Continue doing this until we
766 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
767 while (ValSizeInBytes > ByteSize) {
768 ValSizeInBytes >>= 1;
770 // If the top half equals the bottom half, we're still ok.
771 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
772 (Value & ((1 << (8*ValSizeInBytes))-1)))
776 // Properly sign extend the value.
777 int ShAmt = (4-ByteSize)*8;
778 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
780 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
781 if (MaskVal == 0) return SDValue();
783 // Finally, if this value fits in a 5 bit sext field, return it
784 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
785 return DAG.getTargetConstant(MaskVal, MVT::i32);
789 //===----------------------------------------------------------------------===//
790 // Addressing Mode Selection
791 //===----------------------------------------------------------------------===//
793 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
794 /// or 64-bit immediate, and if the value can be accurately represented as a
795 /// sign extension from a 16-bit value. If so, this returns true and the
797 static bool isIntS16Immediate(SDNode *N, short &Imm) {
798 if (N->getOpcode() != ISD::Constant)
801 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
802 if (N->getValueType(0) == MVT::i32)
803 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
805 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
807 static bool isIntS16Immediate(SDValue Op, short &Imm) {
808 return isIntS16Immediate(Op.getNode(), Imm);
812 /// SelectAddressRegReg - Given the specified addressed, check to see if it
813 /// can be represented as an indexed [r+r] operation. Returns false if it
814 /// can be more efficiently represented with [r+imm].
815 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
817 SelectionDAG &DAG) const {
819 if (N.getOpcode() == ISD::ADD) {
820 if (isIntS16Immediate(N.getOperand(1), imm))
822 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
825 Base = N.getOperand(0);
826 Index = N.getOperand(1);
828 } else if (N.getOpcode() == ISD::OR) {
829 if (isIntS16Immediate(N.getOperand(1), imm))
830 return false; // r+i can fold it if we can.
832 // If this is an or of disjoint bitfields, we can codegen this as an add
833 // (for better address arithmetic) if the LHS and RHS of the OR are provably
835 APInt LHSKnownZero, LHSKnownOne;
836 APInt RHSKnownZero, RHSKnownOne;
837 DAG.ComputeMaskedBits(N.getOperand(0),
838 APInt::getAllOnesValue(N.getOperand(0)
839 .getValueSizeInBits()),
840 LHSKnownZero, LHSKnownOne);
842 if (LHSKnownZero.getBoolValue()) {
843 DAG.ComputeMaskedBits(N.getOperand(1),
844 APInt::getAllOnesValue(N.getOperand(1)
845 .getValueSizeInBits()),
846 RHSKnownZero, RHSKnownOne);
847 // If all of the bits are known zero on the LHS or RHS, the add won't
849 if (~(LHSKnownZero | RHSKnownZero) == 0) {
850 Base = N.getOperand(0);
851 Index = N.getOperand(1);
860 /// Returns true if the address N can be represented by a base register plus
861 /// a signed 16-bit displacement [r+imm], and if it is not better
862 /// represented as reg+reg.
863 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
865 SelectionDAG &DAG) const {
866 // FIXME dl should come from parent load or store, not from address
867 DebugLoc dl = N.getDebugLoc();
868 // If this can be more profitably realized as r+r, fail.
869 if (SelectAddressRegReg(N, Disp, Base, DAG))
872 if (N.getOpcode() == ISD::ADD) {
874 if (isIntS16Immediate(N.getOperand(1), imm)) {
875 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
876 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
877 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
879 Base = N.getOperand(0);
881 return true; // [r+i]
882 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
883 // Match LOAD (ADD (X, Lo(G))).
884 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
885 && "Cannot handle constant offsets yet!");
886 Disp = N.getOperand(1).getOperand(0); // The global address.
887 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
888 Disp.getOpcode() == ISD::TargetConstantPool ||
889 Disp.getOpcode() == ISD::TargetJumpTable);
890 Base = N.getOperand(0);
891 return true; // [&g+r]
893 } else if (N.getOpcode() == ISD::OR) {
895 if (isIntS16Immediate(N.getOperand(1), imm)) {
896 // If this is an or of disjoint bitfields, we can codegen this as an add
897 // (for better address arithmetic) if the LHS and RHS of the OR are
898 // provably disjoint.
899 APInt LHSKnownZero, LHSKnownOne;
900 DAG.ComputeMaskedBits(N.getOperand(0),
901 APInt::getAllOnesValue(N.getOperand(0)
902 .getValueSizeInBits()),
903 LHSKnownZero, LHSKnownOne);
905 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
906 // If all of the bits are known zero on the LHS or RHS, the add won't
908 Base = N.getOperand(0);
909 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
913 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
914 // Loading from a constant address.
916 // If this address fits entirely in a 16-bit sext immediate field, codegen
919 if (isIntS16Immediate(CN, Imm)) {
920 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
921 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
922 CN->getValueType(0));
926 // Handle 32-bit sext immediates with LIS + addr mode.
927 if (CN->getValueType(0) == MVT::i32 ||
928 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
929 int Addr = (int)CN->getZExtValue();
931 // Otherwise, break this down into an LIS + disp.
932 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
934 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
935 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
936 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
941 Disp = DAG.getTargetConstant(0, getPointerTy());
942 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
943 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
946 return true; // [r+0]
949 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
950 /// represented as an indexed [r+r] operation.
951 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
953 SelectionDAG &DAG) const {
954 // Check to see if we can easily represent this as an [r+r] address. This
955 // will fail if it thinks that the address is more profitably represented as
956 // reg+imm, e.g. where imm = 0.
957 if (SelectAddressRegReg(N, Base, Index, DAG))
960 // If the operand is an addition, always emit this as [r+r], since this is
961 // better (for code size, and execution, as the memop does the add for free)
962 // than emitting an explicit add.
963 if (N.getOpcode() == ISD::ADD) {
964 Base = N.getOperand(0);
965 Index = N.getOperand(1);
969 // Otherwise, do it the hard way, using R0 as the base register.
970 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
976 /// SelectAddressRegImmShift - Returns true if the address N can be
977 /// represented by a base register plus a signed 14-bit displacement
978 /// [r+imm*4]. Suitable for use by STD and friends.
979 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
981 SelectionDAG &DAG) const {
982 // FIXME dl should come from the parent load or store, not the address
983 DebugLoc dl = N.getDebugLoc();
984 // If this can be more profitably realized as r+r, fail.
985 if (SelectAddressRegReg(N, Disp, Base, DAG))
988 if (N.getOpcode() == ISD::ADD) {
990 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
991 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
992 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
993 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
995 Base = N.getOperand(0);
997 return true; // [r+i]
998 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
999 // Match LOAD (ADD (X, Lo(G))).
1000 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1001 && "Cannot handle constant offsets yet!");
1002 Disp = N.getOperand(1).getOperand(0); // The global address.
1003 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1004 Disp.getOpcode() == ISD::TargetConstantPool ||
1005 Disp.getOpcode() == ISD::TargetJumpTable);
1006 Base = N.getOperand(0);
1007 return true; // [&g+r]
1009 } else if (N.getOpcode() == ISD::OR) {
1011 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1012 // If this is an or of disjoint bitfields, we can codegen this as an add
1013 // (for better address arithmetic) if the LHS and RHS of the OR are
1014 // provably disjoint.
1015 APInt LHSKnownZero, LHSKnownOne;
1016 DAG.ComputeMaskedBits(N.getOperand(0),
1017 APInt::getAllOnesValue(N.getOperand(0)
1018 .getValueSizeInBits()),
1019 LHSKnownZero, LHSKnownOne);
1020 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1021 // If all of the bits are known zero on the LHS or RHS, the add won't
1023 Base = N.getOperand(0);
1024 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1028 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1029 // Loading from a constant address. Verify low two bits are clear.
1030 if ((CN->getZExtValue() & 3) == 0) {
1031 // If this address fits entirely in a 14-bit sext immediate field, codegen
1034 if (isIntS16Immediate(CN, Imm)) {
1035 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1036 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1037 CN->getValueType(0));
1041 // Fold the low-part of 32-bit absolute addresses into addr mode.
1042 if (CN->getValueType(0) == MVT::i32 ||
1043 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1044 int Addr = (int)CN->getZExtValue();
1046 // Otherwise, break this down into an LIS + disp.
1047 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1048 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1049 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1050 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1056 Disp = DAG.getTargetConstant(0, getPointerTy());
1057 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1058 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1061 return true; // [r+0]
1065 /// getPreIndexedAddressParts - returns true by value, base pointer and
1066 /// offset pointer and addressing mode by reference if the node's address
1067 /// can be legally represented as pre-indexed load / store address.
1068 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1070 ISD::MemIndexedMode &AM,
1071 SelectionDAG &DAG) const {
1072 // Disabled by default for now.
1073 if (!EnablePPCPreinc) return false;
1077 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1078 Ptr = LD->getBasePtr();
1079 VT = LD->getMemoryVT();
1081 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1082 Ptr = ST->getBasePtr();
1083 VT = ST->getMemoryVT();
1087 // PowerPC doesn't have preinc load/store instructions for vectors.
1091 // TODO: Check reg+reg first.
1093 // LDU/STU use reg+imm*4, others use reg+imm.
1094 if (VT != MVT::i64) {
1096 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1100 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1104 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1105 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1106 // sext i32 to i64 when addr mode is r+i.
1107 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1108 LD->getExtensionType() == ISD::SEXTLOAD &&
1109 isa<ConstantSDNode>(Offset))
1117 //===----------------------------------------------------------------------===//
1118 // LowerOperation implementation
1119 //===----------------------------------------------------------------------===//
1121 /// GetLabelAccessInfo - Return true if we should reference labels using a
1122 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1123 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1124 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1125 HiOpFlags = PPCII::MO_HA16;
1126 LoOpFlags = PPCII::MO_LO16;
1128 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1129 // non-darwin platform. We don't support PIC on other platforms yet.
1130 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1131 TM.getSubtarget<PPCSubtarget>().isDarwin();
1133 HiOpFlags |= PPCII::MO_PIC_FLAG;
1134 LoOpFlags |= PPCII::MO_PIC_FLAG;
1137 // If this is a reference to a global value that requires a non-lazy-ptr, make
1138 // sure that instruction lowering adds it.
1139 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1140 HiOpFlags |= PPCII::MO_NLP_FLAG;
1141 LoOpFlags |= PPCII::MO_NLP_FLAG;
1143 if (GV->hasHiddenVisibility()) {
1144 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1145 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1152 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1153 SelectionDAG &DAG) {
1154 EVT PtrVT = HiPart.getValueType();
1155 SDValue Zero = DAG.getConstant(0, PtrVT);
1156 DebugLoc DL = HiPart.getDebugLoc();
1158 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1159 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1161 // With PIC, the first instruction is actually "GR+hi(&G)".
1163 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1164 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1166 // Generate non-pic code that has direct accesses to the constant pool.
1167 // The address of the global is just (hi(&g)+lo(&g)).
1168 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1171 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1172 SelectionDAG &DAG) const {
1173 EVT PtrVT = Op.getValueType();
1174 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1175 const Constant *C = CP->getConstVal();
1177 unsigned MOHiFlag, MOLoFlag;
1178 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1180 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1182 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1183 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1186 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1187 EVT PtrVT = Op.getValueType();
1188 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1190 unsigned MOHiFlag, MOLoFlag;
1191 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1192 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1193 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1194 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1197 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1198 SelectionDAG &DAG) const {
1199 EVT PtrVT = Op.getValueType();
1201 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1203 unsigned MOHiFlag, MOLoFlag;
1204 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1205 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1206 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1207 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1210 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1211 SelectionDAG &DAG) const {
1212 EVT PtrVT = Op.getValueType();
1213 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1214 DebugLoc DL = GSDN->getDebugLoc();
1215 const GlobalValue *GV = GSDN->getGlobal();
1217 // 64-bit SVR4 ABI code is always position-independent.
1218 // The actual address of the GlobalValue is stored in the TOC.
1219 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1220 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1221 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1222 DAG.getRegister(PPC::X2, MVT::i64));
1225 unsigned MOHiFlag, MOLoFlag;
1226 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1229 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1231 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1233 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1235 // If the global reference is actually to a non-lazy-pointer, we have to do an
1236 // extra load to get the address of the global.
1237 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1238 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1239 false, false, false, 0);
1243 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1244 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1245 DebugLoc dl = Op.getDebugLoc();
1247 // If we're comparing for equality to zero, expose the fact that this is
1248 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1249 // fold the new nodes.
1250 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1251 if (C->isNullValue() && CC == ISD::SETEQ) {
1252 EVT VT = Op.getOperand(0).getValueType();
1253 SDValue Zext = Op.getOperand(0);
1254 if (VT.bitsLT(MVT::i32)) {
1256 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1258 unsigned Log2b = Log2_32(VT.getSizeInBits());
1259 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1260 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1261 DAG.getConstant(Log2b, MVT::i32));
1262 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1264 // Leave comparisons against 0 and -1 alone for now, since they're usually
1265 // optimized. FIXME: revisit this when we can custom lower all setcc
1267 if (C->isAllOnesValue() || C->isNullValue())
1271 // If we have an integer seteq/setne, turn it into a compare against zero
1272 // by xor'ing the rhs with the lhs, which is faster than setting a
1273 // condition register, reading it back out, and masking the correct bit. The
1274 // normal approach here uses sub to do this instead of xor. Using xor exposes
1275 // the result to other bit-twiddling opportunities.
1276 EVT LHSVT = Op.getOperand(0).getValueType();
1277 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1278 EVT VT = Op.getValueType();
1279 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1281 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1286 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1287 const PPCSubtarget &Subtarget) const {
1288 SDNode *Node = Op.getNode();
1289 EVT VT = Node->getValueType(0);
1290 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1291 SDValue InChain = Node->getOperand(0);
1292 SDValue VAListPtr = Node->getOperand(1);
1293 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1294 DebugLoc dl = Node->getDebugLoc();
1296 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1299 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1300 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1302 InChain = GprIndex.getValue(1);
1304 if (VT == MVT::i64) {
1305 // Check if GprIndex is even
1306 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1307 DAG.getConstant(1, MVT::i32));
1308 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1309 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1310 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1311 DAG.getConstant(1, MVT::i32));
1312 // Align GprIndex to be even if it isn't
1313 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1317 // fpr index is 1 byte after gpr
1318 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1319 DAG.getConstant(1, MVT::i32));
1322 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1323 FprPtr, MachinePointerInfo(SV), MVT::i8,
1325 InChain = FprIndex.getValue(1);
1327 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1328 DAG.getConstant(8, MVT::i32));
1330 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1331 DAG.getConstant(4, MVT::i32));
1334 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1335 MachinePointerInfo(), false, false,
1337 InChain = OverflowArea.getValue(1);
1339 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1340 MachinePointerInfo(), false, false,
1342 InChain = RegSaveArea.getValue(1);
1344 // select overflow_area if index > 8
1345 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1346 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1348 // adjustment constant gpr_index * 4/8
1349 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1350 VT.isInteger() ? GprIndex : FprIndex,
1351 DAG.getConstant(VT.isInteger() ? 4 : 8,
1354 // OurReg = RegSaveArea + RegConstant
1355 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1358 // Floating types are 32 bytes into RegSaveArea
1359 if (VT.isFloatingPoint())
1360 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1361 DAG.getConstant(32, MVT::i32));
1363 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1364 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1365 VT.isInteger() ? GprIndex : FprIndex,
1366 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1369 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1370 VT.isInteger() ? VAListPtr : FprPtr,
1371 MachinePointerInfo(SV),
1372 MVT::i8, false, false, 0);
1374 // determine if we should load from reg_save_area or overflow_area
1375 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1377 // increase overflow_area by 4/8 if gpr/fpr > 8
1378 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1379 DAG.getConstant(VT.isInteger() ? 4 : 8,
1382 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1385 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1387 MachinePointerInfo(),
1388 MVT::i32, false, false, 0);
1390 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1391 false, false, false, 0);
1394 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1395 SelectionDAG &DAG) const {
1396 return Op.getOperand(0);
1399 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1400 SelectionDAG &DAG) const {
1401 SDValue Chain = Op.getOperand(0);
1402 SDValue Trmp = Op.getOperand(1); // trampoline
1403 SDValue FPtr = Op.getOperand(2); // nested function
1404 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1405 DebugLoc dl = Op.getDebugLoc();
1407 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1408 bool isPPC64 = (PtrVT == MVT::i64);
1410 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1413 TargetLowering::ArgListTy Args;
1414 TargetLowering::ArgListEntry Entry;
1416 Entry.Ty = IntPtrTy;
1417 Entry.Node = Trmp; Args.push_back(Entry);
1419 // TrampSize == (isPPC64 ? 48 : 40);
1420 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1421 isPPC64 ? MVT::i64 : MVT::i32);
1422 Args.push_back(Entry);
1424 Entry.Node = FPtr; Args.push_back(Entry);
1425 Entry.Node = Nest; Args.push_back(Entry);
1427 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1428 std::pair<SDValue, SDValue> CallResult =
1429 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
1430 false, false, false, false, 0, CallingConv::C,
1431 /*isTailCall=*/false,
1432 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
1433 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1436 return CallResult.second;
1439 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1440 const PPCSubtarget &Subtarget) const {
1441 MachineFunction &MF = DAG.getMachineFunction();
1442 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1444 DebugLoc dl = Op.getDebugLoc();
1446 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1447 // vastart just stores the address of the VarArgsFrameIndex slot into the
1448 // memory location argument.
1449 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1450 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1451 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1452 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1453 MachinePointerInfo(SV),
1457 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1458 // We suppose the given va_list is already allocated.
1461 // char gpr; /* index into the array of 8 GPRs
1462 // * stored in the register save area
1463 // * gpr=0 corresponds to r3,
1464 // * gpr=1 to r4, etc.
1466 // char fpr; /* index into the array of 8 FPRs
1467 // * stored in the register save area
1468 // * fpr=0 corresponds to f1,
1469 // * fpr=1 to f2, etc.
1471 // char *overflow_arg_area;
1472 // /* location on stack that holds
1473 // * the next overflow argument
1475 // char *reg_save_area;
1476 // /* where r3:r10 and f1:f8 (if saved)
1482 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1483 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1486 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1488 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1490 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1493 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1494 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1496 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1497 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1499 uint64_t FPROffset = 1;
1500 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1502 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1504 // Store first byte : number of int regs
1505 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1507 MachinePointerInfo(SV),
1508 MVT::i8, false, false, 0);
1509 uint64_t nextOffset = FPROffset;
1510 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1513 // Store second byte : number of float regs
1514 SDValue secondStore =
1515 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1516 MachinePointerInfo(SV, nextOffset), MVT::i8,
1518 nextOffset += StackOffset;
1519 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1521 // Store second word : arguments given on stack
1522 SDValue thirdStore =
1523 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1524 MachinePointerInfo(SV, nextOffset),
1526 nextOffset += FrameOffset;
1527 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1529 // Store third word : arguments given in registers
1530 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1531 MachinePointerInfo(SV, nextOffset),
1536 #include "PPCGenCallingConv.inc"
1538 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1539 CCValAssign::LocInfo &LocInfo,
1540 ISD::ArgFlagsTy &ArgFlags,
1545 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1547 CCValAssign::LocInfo &LocInfo,
1548 ISD::ArgFlagsTy &ArgFlags,
1550 static const unsigned ArgRegs[] = {
1551 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1552 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1554 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1556 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1558 // Skip one register if the first unallocated register has an even register
1559 // number and there are still argument registers available which have not been
1560 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1561 // need to skip a register if RegNum is odd.
1562 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1563 State.AllocateReg(ArgRegs[RegNum]);
1566 // Always return false here, as this function only makes sure that the first
1567 // unallocated register has an odd register number and does not actually
1568 // allocate a register for the current argument.
1572 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1574 CCValAssign::LocInfo &LocInfo,
1575 ISD::ArgFlagsTy &ArgFlags,
1577 static const unsigned ArgRegs[] = {
1578 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1582 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1584 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1586 // If there is only one Floating-point register left we need to put both f64
1587 // values of a split ppc_fp128 value on the stack.
1588 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1589 State.AllocateReg(ArgRegs[RegNum]);
1592 // Always return false here, as this function only makes sure that the two f64
1593 // values a ppc_fp128 value is split into are both passed in registers or both
1594 // passed on the stack and does not actually allocate a register for the
1595 // current argument.
1599 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1601 static const unsigned *GetFPR() {
1602 static const unsigned FPR[] = {
1603 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1604 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1610 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1612 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1613 unsigned PtrByteSize) {
1614 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1615 if (Flags.isByVal())
1616 ArgSize = Flags.getByValSize();
1617 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1623 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1624 CallingConv::ID CallConv, bool isVarArg,
1625 const SmallVectorImpl<ISD::InputArg>
1627 DebugLoc dl, SelectionDAG &DAG,
1628 SmallVectorImpl<SDValue> &InVals)
1630 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1631 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1634 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1640 PPCTargetLowering::LowerFormalArguments_SVR4(
1642 CallingConv::ID CallConv, bool isVarArg,
1643 const SmallVectorImpl<ISD::InputArg>
1645 DebugLoc dl, SelectionDAG &DAG,
1646 SmallVectorImpl<SDValue> &InVals) const {
1648 // 32-bit SVR4 ABI Stack Frame Layout:
1649 // +-----------------------------------+
1650 // +--> | Back chain |
1651 // | +-----------------------------------+
1652 // | | Floating-point register save area |
1653 // | +-----------------------------------+
1654 // | | General register save area |
1655 // | +-----------------------------------+
1656 // | | CR save word |
1657 // | +-----------------------------------+
1658 // | | VRSAVE save word |
1659 // | +-----------------------------------+
1660 // | | Alignment padding |
1661 // | +-----------------------------------+
1662 // | | Vector register save area |
1663 // | +-----------------------------------+
1664 // | | Local variable space |
1665 // | +-----------------------------------+
1666 // | | Parameter list area |
1667 // | +-----------------------------------+
1668 // | | LR save word |
1669 // | +-----------------------------------+
1670 // SP--> +--- | Back chain |
1671 // +-----------------------------------+
1674 // System V Application Binary Interface PowerPC Processor Supplement
1675 // AltiVec Technology Programming Interface Manual
1677 MachineFunction &MF = DAG.getMachineFunction();
1678 MachineFrameInfo *MFI = MF.getFrameInfo();
1679 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1681 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1682 // Potential tail calls could cause overwriting of argument stack slots.
1683 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1684 (CallConv == CallingConv::Fast));
1685 unsigned PtrByteSize = 4;
1687 // Assign locations to all of the incoming arguments.
1688 SmallVector<CCValAssign, 16> ArgLocs;
1689 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1690 getTargetMachine(), ArgLocs, *DAG.getContext());
1692 // Reserve space for the linkage area on the stack.
1693 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1695 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1697 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1698 CCValAssign &VA = ArgLocs[i];
1700 // Arguments stored in registers.
1701 if (VA.isRegLoc()) {
1702 const TargetRegisterClass *RC;
1703 EVT ValVT = VA.getValVT();
1705 switch (ValVT.getSimpleVT().SimpleTy) {
1707 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1709 RC = PPC::GPRCRegisterClass;
1712 RC = PPC::F4RCRegisterClass;
1715 RC = PPC::F8RCRegisterClass;
1721 RC = PPC::VRRCRegisterClass;
1725 // Transform the arguments stored in physical registers into virtual ones.
1726 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1727 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1729 InVals.push_back(ArgValue);
1731 // Argument stored in memory.
1732 assert(VA.isMemLoc());
1734 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1735 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1738 // Create load nodes to retrieve arguments from the stack.
1739 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1740 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1741 MachinePointerInfo(),
1742 false, false, false, 0));
1746 // Assign locations to all of the incoming aggregate by value arguments.
1747 // Aggregates passed by value are stored in the local variable space of the
1748 // caller's stack frame, right above the parameter list area.
1749 SmallVector<CCValAssign, 16> ByValArgLocs;
1750 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1751 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1753 // Reserve stack space for the allocations in CCInfo.
1754 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1756 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1758 // Area that is at least reserved in the caller of this function.
1759 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1761 // Set the size that is at least reserved in caller of this function. Tail
1762 // call optimized function's reserved stack space needs to be aligned so that
1763 // taking the difference between two stack areas will result in an aligned
1765 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1768 std::max(MinReservedArea,
1769 PPCFrameLowering::getMinCallFrameSize(false, false));
1771 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1772 getStackAlignment();
1773 unsigned AlignMask = TargetAlign-1;
1774 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1776 FI->setMinReservedArea(MinReservedArea);
1778 SmallVector<SDValue, 8> MemOps;
1780 // If the function takes variable number of arguments, make a frame index for
1781 // the start of the first vararg value... for expansion of llvm.va_start.
1783 static const unsigned GPArgRegs[] = {
1784 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1785 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1787 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1789 static const unsigned FPArgRegs[] = {
1790 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1793 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1795 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1797 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1800 // Make room for NumGPArgRegs and NumFPArgRegs.
1801 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1802 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1804 FuncInfo->setVarArgsStackOffset(
1805 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1806 CCInfo.getNextStackOffset(), true));
1808 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1809 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1811 // The fixed integer arguments of a variadic function are stored to the
1812 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1813 // the result of va_next.
1814 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1815 // Get an existing live-in vreg, or add a new one.
1816 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1818 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1820 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
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 four for the next argument to store
1825 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1826 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1829 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1831 // The double arguments are stored to the VarArgsFrameIndex
1833 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1834 // Get an existing live-in vreg, or add a new one.
1835 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1837 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1839 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1840 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1841 MachinePointerInfo(), false, false, 0);
1842 MemOps.push_back(Store);
1843 // Increment the address by eight for the next argument to store
1844 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1846 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1850 if (!MemOps.empty())
1851 Chain = DAG.getNode(ISD::TokenFactor, dl,
1852 MVT::Other, &MemOps[0], MemOps.size());
1858 PPCTargetLowering::LowerFormalArguments_Darwin(
1860 CallingConv::ID CallConv, bool isVarArg,
1861 const SmallVectorImpl<ISD::InputArg>
1863 DebugLoc dl, SelectionDAG &DAG,
1864 SmallVectorImpl<SDValue> &InVals) const {
1865 // TODO: add description of PPC stack frame format, or at least some docs.
1867 MachineFunction &MF = DAG.getMachineFunction();
1868 MachineFrameInfo *MFI = MF.getFrameInfo();
1869 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1871 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1872 bool isPPC64 = PtrVT == MVT::i64;
1873 // Potential tail calls could cause overwriting of argument stack slots.
1874 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1875 (CallConv == CallingConv::Fast));
1876 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1878 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1879 // Area that is at least reserved in caller of this function.
1880 unsigned MinReservedArea = ArgOffset;
1882 static const unsigned GPR_32[] = { // 32-bit registers.
1883 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1884 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1886 static const unsigned GPR_64[] = { // 64-bit registers.
1887 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1888 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1891 static const unsigned *FPR = GetFPR();
1893 static const unsigned VR[] = {
1894 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1895 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1898 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1899 const unsigned Num_FPR_Regs = 13;
1900 const unsigned Num_VR_Regs = array_lengthof( VR);
1902 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1904 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1906 // In 32-bit non-varargs functions, the stack space for vectors is after the
1907 // stack space for non-vectors. We do not use this space unless we have
1908 // too many vectors to fit in registers, something that only occurs in
1909 // constructed examples:), but we have to walk the arglist to figure
1910 // that out...for the pathological case, compute VecArgOffset as the
1911 // start of the vector parameter area. Computing VecArgOffset is the
1912 // entire point of the following loop.
1913 unsigned VecArgOffset = ArgOffset;
1914 if (!isVarArg && !isPPC64) {
1915 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1917 EVT ObjectVT = Ins[ArgNo].VT;
1918 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1920 if (Flags.isByVal()) {
1921 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1922 unsigned ObjSize = Flags.getByValSize();
1924 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1925 VecArgOffset += ArgSize;
1929 switch(ObjectVT.getSimpleVT().SimpleTy) {
1930 default: llvm_unreachable("Unhandled argument type!");
1933 VecArgOffset += isPPC64 ? 8 : 4;
1935 case MVT::i64: // PPC64
1943 // Nothing to do, we're only looking at Nonvector args here.
1948 // We've found where the vector parameter area in memory is. Skip the
1949 // first 12 parameters; these don't use that memory.
1950 VecArgOffset = ((VecArgOffset+15)/16)*16;
1951 VecArgOffset += 12*16;
1953 // Add DAG nodes to load the arguments or copy them out of registers. On
1954 // entry to a function on PPC, the arguments start after the linkage area,
1955 // although the first ones are often in registers.
1957 SmallVector<SDValue, 8> MemOps;
1958 unsigned nAltivecParamsAtEnd = 0;
1959 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1961 bool needsLoad = false;
1962 EVT ObjectVT = Ins[ArgNo].VT;
1963 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1964 unsigned ArgSize = ObjSize;
1965 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1967 unsigned CurArgOffset = ArgOffset;
1969 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1970 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1971 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1972 if (isVarArg || isPPC64) {
1973 MinReservedArea = ((MinReservedArea+15)/16)*16;
1974 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1977 } else nAltivecParamsAtEnd++;
1979 // Calculate min reserved area.
1980 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1984 // FIXME the codegen can be much improved in some cases.
1985 // We do not have to keep everything in memory.
1986 if (Flags.isByVal()) {
1987 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1988 ObjSize = Flags.getByValSize();
1989 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1990 // Objects of size 1 and 2 are right justified, everything else is
1991 // left justified. This means the memory address is adjusted forwards.
1992 if (ObjSize==1 || ObjSize==2) {
1993 CurArgOffset = CurArgOffset + (4 - ObjSize);
1995 // The value of the object is its address.
1996 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
1997 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1998 InVals.push_back(FIN);
1999 if (ObjSize==1 || ObjSize==2) {
2000 if (GPR_idx != Num_GPR_Regs) {
2003 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2005 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2006 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2007 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2008 MachinePointerInfo(),
2009 ObjSize==1 ? MVT::i8 : MVT::i16,
2011 MemOps.push_back(Store);
2015 ArgOffset += PtrByteSize;
2019 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2020 // Store whatever pieces of the object are in registers
2021 // to memory. ArgVal will be address of the beginning of
2023 if (GPR_idx != Num_GPR_Regs) {
2026 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2028 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2029 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2030 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2031 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2032 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2033 MachinePointerInfo(),
2035 MemOps.push_back(Store);
2037 ArgOffset += PtrByteSize;
2039 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2046 switch (ObjectVT.getSimpleVT().SimpleTy) {
2047 default: llvm_unreachable("Unhandled argument type!");
2050 if (GPR_idx != Num_GPR_Regs) {
2051 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2052 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2056 ArgSize = PtrByteSize;
2058 // All int arguments reserve stack space in the Darwin ABI.
2059 ArgOffset += PtrByteSize;
2063 case MVT::i64: // PPC64
2064 if (GPR_idx != Num_GPR_Regs) {
2065 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2066 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2068 if (ObjectVT == MVT::i32) {
2069 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2070 // value to MVT::i64 and then truncate to the correct register size.
2072 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2073 DAG.getValueType(ObjectVT));
2074 else if (Flags.isZExt())
2075 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2076 DAG.getValueType(ObjectVT));
2078 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2084 ArgSize = PtrByteSize;
2086 // All int arguments reserve stack space in the Darwin ABI.
2092 // Every 4 bytes of argument space consumes one of the GPRs available for
2093 // argument passing.
2094 if (GPR_idx != Num_GPR_Regs) {
2096 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2099 if (FPR_idx != Num_FPR_Regs) {
2102 if (ObjectVT == MVT::f32)
2103 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2105 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2107 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2113 // All FP arguments reserve stack space in the Darwin ABI.
2114 ArgOffset += isPPC64 ? 8 : ObjSize;
2120 // Note that vector arguments in registers don't reserve stack space,
2121 // except in varargs functions.
2122 if (VR_idx != Num_VR_Regs) {
2123 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2124 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2126 while ((ArgOffset % 16) != 0) {
2127 ArgOffset += PtrByteSize;
2128 if (GPR_idx != Num_GPR_Regs)
2132 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2136 if (!isVarArg && !isPPC64) {
2137 // Vectors go after all the nonvectors.
2138 CurArgOffset = VecArgOffset;
2141 // Vectors are aligned.
2142 ArgOffset = ((ArgOffset+15)/16)*16;
2143 CurArgOffset = ArgOffset;
2151 // We need to load the argument to a virtual register if we determined above
2152 // that we ran out of physical registers of the appropriate type.
2154 int FI = MFI->CreateFixedObject(ObjSize,
2155 CurArgOffset + (ArgSize - ObjSize),
2157 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2158 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2159 false, false, false, 0);
2162 InVals.push_back(ArgVal);
2165 // Set the size that is at least reserved in caller of this function. Tail
2166 // call optimized function's reserved stack space needs to be aligned so that
2167 // taking the difference between two stack areas will result in an aligned
2169 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2170 // Add the Altivec parameters at the end, if needed.
2171 if (nAltivecParamsAtEnd) {
2172 MinReservedArea = ((MinReservedArea+15)/16)*16;
2173 MinReservedArea += 16*nAltivecParamsAtEnd;
2176 std::max(MinReservedArea,
2177 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2178 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2179 getStackAlignment();
2180 unsigned AlignMask = TargetAlign-1;
2181 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2182 FI->setMinReservedArea(MinReservedArea);
2184 // If the function takes variable number of arguments, make a frame index for
2185 // the start of the first vararg value... for expansion of llvm.va_start.
2187 int Depth = ArgOffset;
2189 FuncInfo->setVarArgsFrameIndex(
2190 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2192 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2194 // If this function is vararg, store any remaining integer argument regs
2195 // to their spots on the stack so that they may be loaded by deferencing the
2196 // result of va_next.
2197 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2201 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2203 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2205 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2206 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2207 MachinePointerInfo(), false, false, 0);
2208 MemOps.push_back(Store);
2209 // Increment the address by four for the next argument to store
2210 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2211 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2215 if (!MemOps.empty())
2216 Chain = DAG.getNode(ISD::TokenFactor, dl,
2217 MVT::Other, &MemOps[0], MemOps.size());
2222 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2223 /// linkage area for the Darwin ABI.
2225 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2229 const SmallVectorImpl<ISD::OutputArg>
2231 const SmallVectorImpl<SDValue> &OutVals,
2232 unsigned &nAltivecParamsAtEnd) {
2233 // Count how many bytes are to be pushed on the stack, including the linkage
2234 // area, and parameter passing area. We start with 24/48 bytes, which is
2235 // prereserved space for [SP][CR][LR][3 x unused].
2236 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2237 unsigned NumOps = Outs.size();
2238 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2240 // Add up all the space actually used.
2241 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2242 // they all go in registers, but we must reserve stack space for them for
2243 // possible use by the caller. In varargs or 64-bit calls, parameters are
2244 // assigned stack space in order, with padding so Altivec parameters are
2246 nAltivecParamsAtEnd = 0;
2247 for (unsigned i = 0; i != NumOps; ++i) {
2248 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2249 EVT ArgVT = Outs[i].VT;
2250 // Varargs Altivec parameters are padded to a 16 byte boundary.
2251 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2252 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2253 if (!isVarArg && !isPPC64) {
2254 // Non-varargs Altivec parameters go after all the non-Altivec
2255 // parameters; handle those later so we know how much padding we need.
2256 nAltivecParamsAtEnd++;
2259 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2260 NumBytes = ((NumBytes+15)/16)*16;
2262 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2265 // Allow for Altivec parameters at the end, if needed.
2266 if (nAltivecParamsAtEnd) {
2267 NumBytes = ((NumBytes+15)/16)*16;
2268 NumBytes += 16*nAltivecParamsAtEnd;
2271 // The prolog code of the callee may store up to 8 GPR argument registers to
2272 // the stack, allowing va_start to index over them in memory if its varargs.
2273 // Because we cannot tell if this is needed on the caller side, we have to
2274 // conservatively assume that it is needed. As such, make sure we have at
2275 // least enough stack space for the caller to store the 8 GPRs.
2276 NumBytes = std::max(NumBytes,
2277 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2279 // Tail call needs the stack to be aligned.
2280 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2281 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2282 getFrameLowering()->getStackAlignment();
2283 unsigned AlignMask = TargetAlign-1;
2284 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2290 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2291 /// adjusted to accommodate the arguments for the tailcall.
2292 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2293 unsigned ParamSize) {
2295 if (!isTailCall) return 0;
2297 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2298 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2299 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2300 // Remember only if the new adjustement is bigger.
2301 if (SPDiff < FI->getTailCallSPDelta())
2302 FI->setTailCallSPDelta(SPDiff);
2307 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2308 /// for tail call optimization. Targets which want to do tail call
2309 /// optimization should implement this function.
2311 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2312 CallingConv::ID CalleeCC,
2314 const SmallVectorImpl<ISD::InputArg> &Ins,
2315 SelectionDAG& DAG) const {
2316 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2319 // Variable argument functions are not supported.
2323 MachineFunction &MF = DAG.getMachineFunction();
2324 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2325 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2326 // Functions containing by val parameters are not supported.
2327 for (unsigned i = 0; i != Ins.size(); i++) {
2328 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2329 if (Flags.isByVal()) return false;
2332 // Non PIC/GOT tail calls are supported.
2333 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2336 // At the moment we can only do local tail calls (in same module, hidden
2337 // or protected) if we are generating PIC.
2338 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2339 return G->getGlobal()->hasHiddenVisibility()
2340 || G->getGlobal()->hasProtectedVisibility();
2346 /// isCallCompatibleAddress - Return the immediate to use if the specified
2347 /// 32-bit value is representable in the immediate field of a BxA instruction.
2348 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2349 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2352 int Addr = C->getZExtValue();
2353 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2354 (Addr << 6 >> 6) != Addr)
2355 return 0; // Top 6 bits have to be sext of immediate.
2357 return DAG.getConstant((int)C->getZExtValue() >> 2,
2358 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2363 struct TailCallArgumentInfo {
2368 TailCallArgumentInfo() : FrameIdx(0) {}
2373 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2375 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2377 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2378 SmallVector<SDValue, 8> &MemOpChains,
2380 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2381 SDValue Arg = TailCallArgs[i].Arg;
2382 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2383 int FI = TailCallArgs[i].FrameIdx;
2384 // Store relative to framepointer.
2385 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2386 MachinePointerInfo::getFixedStack(FI),
2391 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2392 /// the appropriate stack slot for the tail call optimized function call.
2393 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2394 MachineFunction &MF,
2403 // Calculate the new stack slot for the return address.
2404 int SlotSize = isPPC64 ? 8 : 4;
2405 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2407 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2408 NewRetAddrLoc, true);
2409 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2410 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2411 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2412 MachinePointerInfo::getFixedStack(NewRetAddr),
2415 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2416 // slot as the FP is never overwritten.
2419 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2420 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2422 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2423 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2424 MachinePointerInfo::getFixedStack(NewFPIdx),
2431 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2432 /// the position of the argument.
2434 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2435 SDValue Arg, int SPDiff, unsigned ArgOffset,
2436 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2437 int Offset = ArgOffset + SPDiff;
2438 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2439 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2440 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2441 SDValue FIN = DAG.getFrameIndex(FI, VT);
2442 TailCallArgumentInfo Info;
2444 Info.FrameIdxOp = FIN;
2446 TailCallArguments.push_back(Info);
2449 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2450 /// stack slot. Returns the chain as result and the loaded frame pointers in
2451 /// LROpOut/FPOpout. Used when tail calling.
2452 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2458 DebugLoc dl) const {
2460 // Load the LR and FP stack slot for later adjusting.
2461 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2462 LROpOut = getReturnAddrFrameIndex(DAG);
2463 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2464 false, false, false, 0);
2465 Chain = SDValue(LROpOut.getNode(), 1);
2467 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2468 // slot as the FP is never overwritten.
2470 FPOpOut = getFramePointerFrameIndex(DAG);
2471 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2472 false, false, false, 0);
2473 Chain = SDValue(FPOpOut.getNode(), 1);
2479 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2480 /// by "Src" to address "Dst" of size "Size". Alignment information is
2481 /// specified by the specific parameter attribute. The copy will be passed as
2482 /// a byval function parameter.
2483 /// Sometimes what we are copying is the end of a larger object, the part that
2484 /// does not fit in registers.
2486 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2487 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2489 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2490 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2491 false, false, MachinePointerInfo(0),
2492 MachinePointerInfo(0));
2495 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2498 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2499 SDValue Arg, SDValue PtrOff, int SPDiff,
2500 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2501 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2502 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2504 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2509 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2511 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2512 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2513 DAG.getConstant(ArgOffset, PtrVT));
2515 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2516 MachinePointerInfo(), false, false, 0));
2517 // Calculate and remember argument location.
2518 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2523 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2524 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2525 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2526 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2527 MachineFunction &MF = DAG.getMachineFunction();
2529 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2530 // might overwrite each other in case of tail call optimization.
2531 SmallVector<SDValue, 8> MemOpChains2;
2532 // Do not flag preceding copytoreg stuff together with the following stuff.
2534 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2536 if (!MemOpChains2.empty())
2537 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2538 &MemOpChains2[0], MemOpChains2.size());
2540 // Store the return address to the appropriate stack slot.
2541 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2542 isPPC64, isDarwinABI, dl);
2544 // Emit callseq_end just before tailcall node.
2545 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2546 DAG.getIntPtrConstant(0, true), InFlag);
2547 InFlag = Chain.getValue(1);
2551 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2552 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2553 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2554 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2555 const PPCSubtarget &PPCSubTarget) {
2557 bool isPPC64 = PPCSubTarget.isPPC64();
2558 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2560 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2561 NodeTys.push_back(MVT::Other); // Returns a chain
2562 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2564 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2566 bool needIndirectCall = true;
2567 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2568 // If this is an absolute destination address, use the munged value.
2569 Callee = SDValue(Dest, 0);
2570 needIndirectCall = false;
2573 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2574 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2575 // Use indirect calls for ALL functions calls in JIT mode, since the
2576 // far-call stubs may be outside relocation limits for a BL instruction.
2577 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2578 unsigned OpFlags = 0;
2579 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2580 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2581 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2582 (G->getGlobal()->isDeclaration() ||
2583 G->getGlobal()->isWeakForLinker())) {
2584 // PC-relative references to external symbols should go through $stub,
2585 // unless we're building with the leopard linker or later, which
2586 // automatically synthesizes these stubs.
2587 OpFlags = PPCII::MO_DARWIN_STUB;
2590 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2591 // every direct call is) turn it into a TargetGlobalAddress /
2592 // TargetExternalSymbol node so that legalize doesn't hack it.
2593 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2594 Callee.getValueType(),
2596 needIndirectCall = false;
2600 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2601 unsigned char OpFlags = 0;
2603 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2604 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2605 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2606 // PC-relative references to external symbols should go through $stub,
2607 // unless we're building with the leopard linker or later, which
2608 // automatically synthesizes these stubs.
2609 OpFlags = PPCII::MO_DARWIN_STUB;
2612 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2614 needIndirectCall = false;
2617 if (needIndirectCall) {
2618 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2619 // to do the call, we can't use PPCISD::CALL.
2620 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2622 if (isSVR4ABI && isPPC64) {
2623 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2624 // entry point, but to the function descriptor (the function entry point
2625 // address is part of the function descriptor though).
2626 // The function descriptor is a three doubleword structure with the
2627 // following fields: function entry point, TOC base address and
2628 // environment pointer.
2629 // Thus for a call through a function pointer, the following actions need
2631 // 1. Save the TOC of the caller in the TOC save area of its stack
2632 // frame (this is done in LowerCall_Darwin()).
2633 // 2. Load the address of the function entry point from the function
2635 // 3. Load the TOC of the callee from the function descriptor into r2.
2636 // 4. Load the environment pointer from the function descriptor into
2638 // 5. Branch to the function entry point address.
2639 // 6. On return of the callee, the TOC of the caller needs to be
2640 // restored (this is done in FinishCall()).
2642 // All those operations are flagged together to ensure that no other
2643 // operations can be scheduled in between. E.g. without flagging the
2644 // operations together, a TOC access in the caller could be scheduled
2645 // between the load of the callee TOC and the branch to the callee, which
2646 // results in the TOC access going through the TOC of the callee instead
2647 // of going through the TOC of the caller, which leads to incorrect code.
2649 // Load the address of the function entry point from the function
2651 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2652 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2653 InFlag.getNode() ? 3 : 2);
2654 Chain = LoadFuncPtr.getValue(1);
2655 InFlag = LoadFuncPtr.getValue(2);
2657 // Load environment pointer into r11.
2658 // Offset of the environment pointer within the function descriptor.
2659 SDValue PtrOff = DAG.getIntPtrConstant(16);
2661 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2662 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2664 Chain = LoadEnvPtr.getValue(1);
2665 InFlag = LoadEnvPtr.getValue(2);
2667 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2669 Chain = EnvVal.getValue(0);
2670 InFlag = EnvVal.getValue(1);
2672 // Load TOC of the callee into r2. We are using a target-specific load
2673 // with r2 hard coded, because the result of a target-independent load
2674 // would never go directly into r2, since r2 is a reserved register (which
2675 // prevents the register allocator from allocating it), resulting in an
2676 // additional register being allocated and an unnecessary move instruction
2678 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2679 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2681 Chain = LoadTOCPtr.getValue(0);
2682 InFlag = LoadTOCPtr.getValue(1);
2684 MTCTROps[0] = Chain;
2685 MTCTROps[1] = LoadFuncPtr;
2686 MTCTROps[2] = InFlag;
2689 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2690 2 + (InFlag.getNode() != 0));
2691 InFlag = Chain.getValue(1);
2694 NodeTys.push_back(MVT::Other);
2695 NodeTys.push_back(MVT::Glue);
2696 Ops.push_back(Chain);
2697 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2699 // Add CTR register as callee so a bctr can be emitted later.
2701 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2704 // If this is a direct call, pass the chain and the callee.
2705 if (Callee.getNode()) {
2706 Ops.push_back(Chain);
2707 Ops.push_back(Callee);
2709 // If this is a tail call add stack pointer delta.
2711 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2713 // Add argument registers to the end of the list so that they are known live
2715 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2716 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2717 RegsToPass[i].second.getValueType()));
2723 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2724 CallingConv::ID CallConv, bool isVarArg,
2725 const SmallVectorImpl<ISD::InputArg> &Ins,
2726 DebugLoc dl, SelectionDAG &DAG,
2727 SmallVectorImpl<SDValue> &InVals) const {
2729 SmallVector<CCValAssign, 16> RVLocs;
2730 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2731 getTargetMachine(), RVLocs, *DAG.getContext());
2732 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2734 // Copy all of the result registers out of their specified physreg.
2735 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2736 CCValAssign &VA = RVLocs[i];
2737 EVT VT = VA.getValVT();
2738 assert(VA.isRegLoc() && "Can only return in registers!");
2739 Chain = DAG.getCopyFromReg(Chain, dl,
2740 VA.getLocReg(), VT, InFlag).getValue(1);
2741 InVals.push_back(Chain.getValue(0));
2742 InFlag = Chain.getValue(2);
2749 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2750 bool isTailCall, bool isVarArg,
2752 SmallVector<std::pair<unsigned, SDValue>, 8>
2754 SDValue InFlag, SDValue Chain,
2756 int SPDiff, unsigned NumBytes,
2757 const SmallVectorImpl<ISD::InputArg> &Ins,
2758 SmallVectorImpl<SDValue> &InVals) const {
2759 std::vector<EVT> NodeTys;
2760 SmallVector<SDValue, 8> Ops;
2761 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2762 isTailCall, RegsToPass, Ops, NodeTys,
2765 // When performing tail call optimization the callee pops its arguments off
2766 // the stack. Account for this here so these bytes can be pushed back on in
2767 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2768 int BytesCalleePops =
2769 (CallConv == CallingConv::Fast &&
2770 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
2772 if (InFlag.getNode())
2773 Ops.push_back(InFlag);
2777 // If this is the first return lowered for this function, add the regs
2778 // to the liveout set for the function.
2779 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2780 SmallVector<CCValAssign, 16> RVLocs;
2781 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2782 getTargetMachine(), RVLocs, *DAG.getContext());
2783 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2784 for (unsigned i = 0; i != RVLocs.size(); ++i)
2785 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2788 assert(((Callee.getOpcode() == ISD::Register &&
2789 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2790 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2791 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2792 isa<ConstantSDNode>(Callee)) &&
2793 "Expecting an global address, external symbol, absolute value or register");
2795 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2798 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2799 InFlag = Chain.getValue(1);
2801 // Add a NOP immediately after the branch instruction when using the 64-bit
2802 // SVR4 ABI. At link time, if caller and callee are in a different module and
2803 // thus have a different TOC, the call will be replaced with a call to a stub
2804 // function which saves the current TOC, loads the TOC of the callee and
2805 // branches to the callee. The NOP will be replaced with a load instruction
2806 // which restores the TOC of the caller from the TOC save slot of the current
2807 // stack frame. If caller and callee belong to the same module (and have the
2808 // same TOC), the NOP will remain unchanged.
2809 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2810 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2811 if (CallOpc == PPCISD::BCTRL_SVR4) {
2812 // This is a call through a function pointer.
2813 // Restore the caller TOC from the save area into R2.
2814 // See PrepareCall() for more information about calls through function
2815 // pointers in the 64-bit SVR4 ABI.
2816 // We are using a target-specific load with r2 hard coded, because the
2817 // result of a target-independent load would never go directly into r2,
2818 // since r2 is a reserved register (which prevents the register allocator
2819 // from allocating it), resulting in an additional register being
2820 // allocated and an unnecessary move instruction being generated.
2821 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2822 InFlag = Chain.getValue(1);
2824 // Otherwise insert NOP.
2825 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Glue, InFlag);
2829 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2830 DAG.getIntPtrConstant(BytesCalleePops, true),
2833 InFlag = Chain.getValue(1);
2835 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2836 Ins, dl, DAG, InVals);
2840 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2841 CallingConv::ID CallConv, bool isVarArg,
2842 bool doesNotRet, bool &isTailCall,
2843 const SmallVectorImpl<ISD::OutputArg> &Outs,
2844 const SmallVectorImpl<SDValue> &OutVals,
2845 const SmallVectorImpl<ISD::InputArg> &Ins,
2846 DebugLoc dl, SelectionDAG &DAG,
2847 SmallVectorImpl<SDValue> &InVals) const {
2849 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2852 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2853 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2854 isTailCall, Outs, OutVals, Ins,
2857 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2858 isTailCall, Outs, OutVals, Ins,
2863 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2864 CallingConv::ID CallConv, bool isVarArg,
2866 const SmallVectorImpl<ISD::OutputArg> &Outs,
2867 const SmallVectorImpl<SDValue> &OutVals,
2868 const SmallVectorImpl<ISD::InputArg> &Ins,
2869 DebugLoc dl, SelectionDAG &DAG,
2870 SmallVectorImpl<SDValue> &InVals) const {
2871 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2872 // of the 32-bit SVR4 ABI stack frame layout.
2874 assert((CallConv == CallingConv::C ||
2875 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2877 unsigned PtrByteSize = 4;
2879 MachineFunction &MF = DAG.getMachineFunction();
2881 // Mark this function as potentially containing a function that contains a
2882 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2883 // and restoring the callers stack pointer in this functions epilog. This is
2884 // done because by tail calling the called function might overwrite the value
2885 // in this function's (MF) stack pointer stack slot 0(SP).
2886 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2887 CallConv == CallingConv::Fast)
2888 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2890 // Count how many bytes are to be pushed on the stack, including the linkage
2891 // area, parameter list area and the part of the local variable space which
2892 // contains copies of aggregates which are passed by value.
2894 // Assign locations to all of the outgoing arguments.
2895 SmallVector<CCValAssign, 16> ArgLocs;
2896 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2897 getTargetMachine(), ArgLocs, *DAG.getContext());
2899 // Reserve space for the linkage area on the stack.
2900 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2903 // Handle fixed and variable vector arguments differently.
2904 // Fixed vector arguments go into registers as long as registers are
2905 // available. Variable vector arguments always go into memory.
2906 unsigned NumArgs = Outs.size();
2908 for (unsigned i = 0; i != NumArgs; ++i) {
2909 MVT ArgVT = Outs[i].VT;
2910 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2913 if (Outs[i].IsFixed) {
2914 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2917 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2923 errs() << "Call operand #" << i << " has unhandled type "
2924 << EVT(ArgVT).getEVTString() << "\n";
2926 llvm_unreachable(0);
2930 // All arguments are treated the same.
2931 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2934 // Assign locations to all of the outgoing aggregate by value arguments.
2935 SmallVector<CCValAssign, 16> ByValArgLocs;
2936 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2937 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2939 // Reserve stack space for the allocations in CCInfo.
2940 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2942 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2944 // Size of the linkage area, parameter list area and the part of the local
2945 // space variable where copies of aggregates which are passed by value are
2947 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2949 // Calculate by how many bytes the stack has to be adjusted in case of tail
2950 // call optimization.
2951 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2953 // Adjust the stack pointer for the new arguments...
2954 // These operations are automatically eliminated by the prolog/epilog pass
2955 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2956 SDValue CallSeqStart = Chain;
2958 // Load the return address and frame pointer so it can be moved somewhere else
2961 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2964 // Set up a copy of the stack pointer for use loading and storing any
2965 // arguments that may not fit in the registers available for argument
2967 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2969 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2970 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2971 SmallVector<SDValue, 8> MemOpChains;
2973 bool seenFloatArg = false;
2974 // Walk the register/memloc assignments, inserting copies/loads.
2975 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2978 CCValAssign &VA = ArgLocs[i];
2979 SDValue Arg = OutVals[i];
2980 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2982 if (Flags.isByVal()) {
2983 // Argument is an aggregate which is passed by value, thus we need to
2984 // create a copy of it in the local variable space of the current stack
2985 // frame (which is the stack frame of the caller) and pass the address of
2986 // this copy to the callee.
2987 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2988 CCValAssign &ByValVA = ByValArgLocs[j++];
2989 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2991 // Memory reserved in the local variable space of the callers stack frame.
2992 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2994 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2995 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2997 // Create a copy of the argument in the local area of the current
2999 SDValue MemcpyCall =
3000 CreateCopyOfByValArgument(Arg, PtrOff,
3001 CallSeqStart.getNode()->getOperand(0),
3004 // This must go outside the CALLSEQ_START..END.
3005 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3006 CallSeqStart.getNode()->getOperand(1));
3007 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3008 NewCallSeqStart.getNode());
3009 Chain = CallSeqStart = NewCallSeqStart;
3011 // Pass the address of the aggregate copy on the stack either in a
3012 // physical register or in the parameter list area of the current stack
3013 // frame to the callee.
3017 if (VA.isRegLoc()) {
3018 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3019 // Put argument in a physical register.
3020 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3022 // Put argument in the parameter list area of the current stack frame.
3023 assert(VA.isMemLoc());
3024 unsigned LocMemOffset = VA.getLocMemOffset();
3027 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3028 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3030 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3031 MachinePointerInfo(),
3034 // Calculate and remember argument location.
3035 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3041 if (!MemOpChains.empty())
3042 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3043 &MemOpChains[0], MemOpChains.size());
3045 // Set CR6 to true if this is a vararg call with floating args passed in
3048 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3050 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3053 // Build a sequence of copy-to-reg nodes chained together with token chain
3054 // and flag operands which copy the outgoing args into the appropriate regs.
3056 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3057 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3058 RegsToPass[i].second, InFlag);
3059 InFlag = Chain.getValue(1);
3063 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3064 false, TailCallArguments);
3066 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3067 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3072 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3073 CallingConv::ID CallConv, bool isVarArg,
3075 const SmallVectorImpl<ISD::OutputArg> &Outs,
3076 const SmallVectorImpl<SDValue> &OutVals,
3077 const SmallVectorImpl<ISD::InputArg> &Ins,
3078 DebugLoc dl, SelectionDAG &DAG,
3079 SmallVectorImpl<SDValue> &InVals) const {
3081 unsigned NumOps = Outs.size();
3083 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3084 bool isPPC64 = PtrVT == MVT::i64;
3085 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3087 MachineFunction &MF = DAG.getMachineFunction();
3089 // Mark this function as potentially containing a function that contains a
3090 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3091 // and restoring the callers stack pointer in this functions epilog. This is
3092 // done because by tail calling the called function might overwrite the value
3093 // in this function's (MF) stack pointer stack slot 0(SP).
3094 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3095 CallConv == CallingConv::Fast)
3096 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3098 unsigned nAltivecParamsAtEnd = 0;
3100 // Count how many bytes are to be pushed on the stack, including the linkage
3101 // area, and parameter passing area. We start with 24/48 bytes, which is
3102 // prereserved space for [SP][CR][LR][3 x unused].
3104 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3106 nAltivecParamsAtEnd);
3108 // Calculate by how many bytes the stack has to be adjusted in case of tail
3109 // call optimization.
3110 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3112 // To protect arguments on the stack from being clobbered in a tail call,
3113 // force all the loads to happen before doing any other lowering.
3115 Chain = DAG.getStackArgumentTokenFactor(Chain);
3117 // Adjust the stack pointer for the new arguments...
3118 // These operations are automatically eliminated by the prolog/epilog pass
3119 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3120 SDValue CallSeqStart = Chain;
3122 // Load the return address and frame pointer so it can be move somewhere else
3125 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3128 // Set up a copy of the stack pointer for use loading and storing any
3129 // arguments that may not fit in the registers available for argument
3133 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3135 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3137 // Figure out which arguments are going to go in registers, and which in
3138 // memory. Also, if this is a vararg function, floating point operations
3139 // must be stored to our stack, and loaded into integer regs as well, if
3140 // any integer regs are available for argument passing.
3141 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3142 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3144 static const unsigned GPR_32[] = { // 32-bit registers.
3145 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3146 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3148 static const unsigned GPR_64[] = { // 64-bit registers.
3149 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3150 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3152 static const unsigned *FPR = GetFPR();
3154 static const unsigned VR[] = {
3155 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3156 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3158 const unsigned NumGPRs = array_lengthof(GPR_32);
3159 const unsigned NumFPRs = 13;
3160 const unsigned NumVRs = array_lengthof(VR);
3162 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
3164 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3165 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3167 SmallVector<SDValue, 8> MemOpChains;
3168 for (unsigned i = 0; i != NumOps; ++i) {
3169 SDValue Arg = OutVals[i];
3170 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3172 // PtrOff will be used to store the current argument to the stack if a
3173 // register cannot be found for it.
3176 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3178 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3180 // On PPC64, promote integers to 64-bit values.
3181 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3182 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3183 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3184 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3187 // FIXME memcpy is used way more than necessary. Correctness first.
3188 if (Flags.isByVal()) {
3189 unsigned Size = Flags.getByValSize();
3190 if (Size==1 || Size==2) {
3191 // Very small objects are passed right-justified.
3192 // Everything else is passed left-justified.
3193 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3194 if (GPR_idx != NumGPRs) {
3195 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3196 MachinePointerInfo(), VT,
3198 MemOpChains.push_back(Load.getValue(1));
3199 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3201 ArgOffset += PtrByteSize;
3203 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3204 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3205 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3206 CallSeqStart.getNode()->getOperand(0),
3208 // This must go outside the CALLSEQ_START..END.
3209 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3210 CallSeqStart.getNode()->getOperand(1));
3211 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3212 NewCallSeqStart.getNode());
3213 Chain = CallSeqStart = NewCallSeqStart;
3214 ArgOffset += PtrByteSize;
3218 // Copy entire object into memory. There are cases where gcc-generated
3219 // code assumes it is there, even if it could be put entirely into
3220 // registers. (This is not what the doc says.)
3221 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3222 CallSeqStart.getNode()->getOperand(0),
3224 // This must go outside the CALLSEQ_START..END.
3225 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3226 CallSeqStart.getNode()->getOperand(1));
3227 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3228 Chain = CallSeqStart = NewCallSeqStart;
3229 // And copy the pieces of it that fit into registers.
3230 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3231 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3232 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3233 if (GPR_idx != NumGPRs) {
3234 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3235 MachinePointerInfo(),
3236 false, false, false, 0);
3237 MemOpChains.push_back(Load.getValue(1));
3238 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3239 ArgOffset += PtrByteSize;
3241 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3248 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3249 default: llvm_unreachable("Unexpected ValueType for argument!");
3252 if (GPR_idx != NumGPRs) {
3253 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3255 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3256 isPPC64, isTailCall, false, MemOpChains,
3257 TailCallArguments, dl);
3259 ArgOffset += PtrByteSize;
3263 if (FPR_idx != NumFPRs) {
3264 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3267 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3268 MachinePointerInfo(), false, false, 0);
3269 MemOpChains.push_back(Store);
3271 // Float varargs are always shadowed in available integer registers
3272 if (GPR_idx != NumGPRs) {
3273 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3274 MachinePointerInfo(), false, false,
3276 MemOpChains.push_back(Load.getValue(1));
3277 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3279 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3280 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3281 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3282 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3283 MachinePointerInfo(),
3284 false, false, false, 0);
3285 MemOpChains.push_back(Load.getValue(1));
3286 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3289 // If we have any FPRs remaining, we may also have GPRs remaining.
3290 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3292 if (GPR_idx != NumGPRs)
3294 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3295 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3299 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3300 isPPC64, isTailCall, false, MemOpChains,
3301 TailCallArguments, dl);
3306 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3313 // These go aligned on the stack, or in the corresponding R registers
3314 // when within range. The Darwin PPC ABI doc claims they also go in
3315 // V registers; in fact gcc does this only for arguments that are
3316 // prototyped, not for those that match the ... We do it for all
3317 // arguments, seems to work.
3318 while (ArgOffset % 16 !=0) {
3319 ArgOffset += PtrByteSize;
3320 if (GPR_idx != NumGPRs)
3323 // We could elide this store in the case where the object fits
3324 // entirely in R registers. Maybe later.
3325 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3326 DAG.getConstant(ArgOffset, PtrVT));
3327 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3328 MachinePointerInfo(), false, false, 0);
3329 MemOpChains.push_back(Store);
3330 if (VR_idx != NumVRs) {
3331 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3332 MachinePointerInfo(),
3333 false, false, false, 0);
3334 MemOpChains.push_back(Load.getValue(1));
3335 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3338 for (unsigned i=0; i<16; i+=PtrByteSize) {
3339 if (GPR_idx == NumGPRs)
3341 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3342 DAG.getConstant(i, PtrVT));
3343 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3344 false, false, false, 0);
3345 MemOpChains.push_back(Load.getValue(1));
3346 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3351 // Non-varargs Altivec params generally go in registers, but have
3352 // stack space allocated at the end.
3353 if (VR_idx != NumVRs) {
3354 // Doesn't have GPR space allocated.
3355 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3356 } else if (nAltivecParamsAtEnd==0) {
3357 // We are emitting Altivec params in order.
3358 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3359 isPPC64, isTailCall, true, MemOpChains,
3360 TailCallArguments, dl);
3366 // If all Altivec parameters fit in registers, as they usually do,
3367 // they get stack space following the non-Altivec parameters. We
3368 // don't track this here because nobody below needs it.
3369 // If there are more Altivec parameters than fit in registers emit
3371 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3373 // Offset is aligned; skip 1st 12 params which go in V registers.
3374 ArgOffset = ((ArgOffset+15)/16)*16;
3376 for (unsigned i = 0; i != NumOps; ++i) {
3377 SDValue Arg = OutVals[i];
3378 EVT ArgType = Outs[i].VT;
3379 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3380 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3383 // We are emitting Altivec params in order.
3384 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3385 isPPC64, isTailCall, true, MemOpChains,
3386 TailCallArguments, dl);
3393 if (!MemOpChains.empty())
3394 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3395 &MemOpChains[0], MemOpChains.size());
3397 // Check if this is an indirect call (MTCTR/BCTRL).
3398 // See PrepareCall() for more information about calls through function
3399 // pointers in the 64-bit SVR4 ABI.
3400 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3401 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3402 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3403 !isBLACompatibleAddress(Callee, DAG)) {
3404 // Load r2 into a virtual register and store it to the TOC save area.
3405 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3406 // TOC save area offset.
3407 SDValue PtrOff = DAG.getIntPtrConstant(40);
3408 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3409 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3413 // On Darwin, R12 must contain the address of an indirect callee. This does
3414 // not mean the MTCTR instruction must use R12; it's easier to model this as
3415 // an extra parameter, so do that.
3417 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3418 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3419 !isBLACompatibleAddress(Callee, DAG))
3420 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3421 PPC::R12), Callee));
3423 // Build a sequence of copy-to-reg nodes chained together with token chain
3424 // and flag operands which copy the outgoing args into the appropriate regs.
3426 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3427 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3428 RegsToPass[i].second, InFlag);
3429 InFlag = Chain.getValue(1);
3433 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3434 FPOp, true, TailCallArguments);
3436 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3437 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3442 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3443 MachineFunction &MF, bool isVarArg,
3444 const SmallVectorImpl<ISD::OutputArg> &Outs,
3445 LLVMContext &Context) const {
3446 SmallVector<CCValAssign, 16> RVLocs;
3447 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3449 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3453 PPCTargetLowering::LowerReturn(SDValue Chain,
3454 CallingConv::ID CallConv, bool isVarArg,
3455 const SmallVectorImpl<ISD::OutputArg> &Outs,
3456 const SmallVectorImpl<SDValue> &OutVals,
3457 DebugLoc dl, SelectionDAG &DAG) const {
3459 SmallVector<CCValAssign, 16> RVLocs;
3460 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3461 getTargetMachine(), RVLocs, *DAG.getContext());
3462 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3464 // If this is the first return lowered for this function, add the regs to the
3465 // liveout set for the function.
3466 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3467 for (unsigned i = 0; i != RVLocs.size(); ++i)
3468 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3473 // Copy the result values into the output registers.
3474 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3475 CCValAssign &VA = RVLocs[i];
3476 assert(VA.isRegLoc() && "Can only return in registers!");
3477 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3479 Flag = Chain.getValue(1);
3483 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3485 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3488 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3489 const PPCSubtarget &Subtarget) const {
3490 // When we pop the dynamic allocation we need to restore the SP link.
3491 DebugLoc dl = Op.getDebugLoc();
3493 // Get the corect type for pointers.
3494 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3496 // Construct the stack pointer operand.
3497 bool isPPC64 = Subtarget.isPPC64();
3498 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3499 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3501 // Get the operands for the STACKRESTORE.
3502 SDValue Chain = Op.getOperand(0);
3503 SDValue SaveSP = Op.getOperand(1);
3505 // Load the old link SP.
3506 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3507 MachinePointerInfo(),
3508 false, false, false, 0);
3510 // Restore the stack pointer.
3511 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3513 // Store the old link SP.
3514 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3521 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3522 MachineFunction &MF = DAG.getMachineFunction();
3523 bool isPPC64 = PPCSubTarget.isPPC64();
3524 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3525 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3527 // Get current frame pointer save index. The users of this index will be
3528 // primarily DYNALLOC instructions.
3529 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3530 int RASI = FI->getReturnAddrSaveIndex();
3532 // If the frame pointer save index hasn't been defined yet.
3534 // Find out what the fix offset of the frame pointer save area.
3535 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3536 // Allocate the frame index for frame pointer save area.
3537 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3539 FI->setReturnAddrSaveIndex(RASI);
3541 return DAG.getFrameIndex(RASI, PtrVT);
3545 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3546 MachineFunction &MF = DAG.getMachineFunction();
3547 bool isPPC64 = PPCSubTarget.isPPC64();
3548 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3549 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3551 // Get current frame pointer save index. The users of this index will be
3552 // primarily DYNALLOC instructions.
3553 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3554 int FPSI = FI->getFramePointerSaveIndex();
3556 // If the frame pointer save index hasn't been defined yet.
3558 // Find out what the fix offset of the frame pointer save area.
3559 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3562 // Allocate the frame index for frame pointer save area.
3563 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3565 FI->setFramePointerSaveIndex(FPSI);
3567 return DAG.getFrameIndex(FPSI, PtrVT);
3570 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3572 const PPCSubtarget &Subtarget) const {
3574 SDValue Chain = Op.getOperand(0);
3575 SDValue Size = Op.getOperand(1);
3576 DebugLoc dl = Op.getDebugLoc();
3578 // Get the corect type for pointers.
3579 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3581 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3582 DAG.getConstant(0, PtrVT), Size);
3583 // Construct a node for the frame pointer save index.
3584 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3585 // Build a DYNALLOC node.
3586 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3587 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3588 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3591 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3593 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3594 // Not FP? Not a fsel.
3595 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3596 !Op.getOperand(2).getValueType().isFloatingPoint())
3599 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3601 // Cannot handle SETEQ/SETNE.
3602 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3604 EVT ResVT = Op.getValueType();
3605 EVT CmpVT = Op.getOperand(0).getValueType();
3606 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3607 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3608 DebugLoc dl = Op.getDebugLoc();
3610 // If the RHS of the comparison is a 0.0, we don't need to do the
3611 // subtraction at all.
3612 if (isFloatingPointZero(RHS))
3614 default: break; // SETUO etc aren't handled by fsel.
3617 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3620 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3621 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3622 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3625 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3628 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3629 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3630 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3631 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3636 default: break; // SETUO etc aren't handled by fsel.
3639 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3640 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3641 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3642 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3645 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3646 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3647 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3648 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3651 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3652 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3653 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3654 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3657 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3658 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3659 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3660 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3665 // FIXME: Split this code up when LegalizeDAGTypes lands.
3666 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3667 DebugLoc dl) const {
3668 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3669 SDValue Src = Op.getOperand(0);
3670 if (Src.getValueType() == MVT::f32)
3671 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3674 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3675 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3677 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3682 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3686 // Convert the FP value to an int value through memory.
3687 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3689 // Emit a store to the stack slot.
3690 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3691 MachinePointerInfo(), false, false, 0);
3693 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3695 if (Op.getValueType() == MVT::i32)
3696 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3697 DAG.getConstant(4, FIPtr.getValueType()));
3698 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3699 false, false, false, 0);
3702 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3703 SelectionDAG &DAG) const {
3704 DebugLoc dl = Op.getDebugLoc();
3705 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3706 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3709 if (Op.getOperand(0).getValueType() == MVT::i64) {
3710 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3711 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3712 if (Op.getValueType() == MVT::f32)
3713 FP = DAG.getNode(ISD::FP_ROUND, dl,
3714 MVT::f32, FP, DAG.getIntPtrConstant(0));
3718 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3719 "Unhandled SINT_TO_FP type in custom expander!");
3720 // Since we only generate this in 64-bit mode, we can take advantage of
3721 // 64-bit registers. In particular, sign extend the input value into the
3722 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3723 // then lfd it and fcfid it.
3724 MachineFunction &MF = DAG.getMachineFunction();
3725 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3726 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3727 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3728 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3730 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3733 // STD the extended value into the stack slot.
3734 MachineMemOperand *MMO =
3735 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3736 MachineMemOperand::MOStore, 8, 8);
3737 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3739 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3740 Ops, 4, MVT::i64, MMO);
3741 // Load the value as a double.
3742 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3743 false, false, false, 0);
3745 // FCFID it and return it.
3746 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3747 if (Op.getValueType() == MVT::f32)
3748 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3752 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3753 SelectionDAG &DAG) const {
3754 DebugLoc dl = Op.getDebugLoc();
3756 The rounding mode is in bits 30:31 of FPSR, and has the following
3763 FLT_ROUNDS, on the other hand, expects the following:
3770 To perform the conversion, we do:
3771 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3774 MachineFunction &MF = DAG.getMachineFunction();
3775 EVT VT = Op.getValueType();
3776 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3777 std::vector<EVT> NodeTys;
3778 SDValue MFFSreg, InFlag;
3780 // Save FP Control Word to register
3781 NodeTys.push_back(MVT::f64); // return register
3782 NodeTys.push_back(MVT::Glue); // unused in this context
3783 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3785 // Save FP register to stack slot
3786 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3787 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3788 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3789 StackSlot, MachinePointerInfo(), false, false,0);
3791 // Load FP Control Word from low 32 bits of stack slot.
3792 SDValue Four = DAG.getConstant(4, PtrVT);
3793 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3794 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3795 false, false, false, 0);
3797 // Transform as necessary
3799 DAG.getNode(ISD::AND, dl, MVT::i32,
3800 CWD, DAG.getConstant(3, MVT::i32));
3802 DAG.getNode(ISD::SRL, dl, MVT::i32,
3803 DAG.getNode(ISD::AND, dl, MVT::i32,
3804 DAG.getNode(ISD::XOR, dl, MVT::i32,
3805 CWD, DAG.getConstant(3, MVT::i32)),
3806 DAG.getConstant(3, MVT::i32)),
3807 DAG.getConstant(1, MVT::i32));
3810 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3812 return DAG.getNode((VT.getSizeInBits() < 16 ?
3813 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3816 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3817 EVT VT = Op.getValueType();
3818 unsigned BitWidth = VT.getSizeInBits();
3819 DebugLoc dl = Op.getDebugLoc();
3820 assert(Op.getNumOperands() == 3 &&
3821 VT == Op.getOperand(1).getValueType() &&
3824 // Expand into a bunch of logical ops. Note that these ops
3825 // depend on the PPC behavior for oversized shift amounts.
3826 SDValue Lo = Op.getOperand(0);
3827 SDValue Hi = Op.getOperand(1);
3828 SDValue Amt = Op.getOperand(2);
3829 EVT AmtVT = Amt.getValueType();
3831 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3832 DAG.getConstant(BitWidth, AmtVT), Amt);
3833 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3834 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3835 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3836 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3837 DAG.getConstant(-BitWidth, AmtVT));
3838 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3839 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3840 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3841 SDValue OutOps[] = { OutLo, OutHi };
3842 return DAG.getMergeValues(OutOps, 2, dl);
3845 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3846 EVT VT = Op.getValueType();
3847 DebugLoc dl = Op.getDebugLoc();
3848 unsigned BitWidth = VT.getSizeInBits();
3849 assert(Op.getNumOperands() == 3 &&
3850 VT == Op.getOperand(1).getValueType() &&
3853 // Expand into a bunch of logical ops. Note that these ops
3854 // depend on the PPC behavior for oversized shift amounts.
3855 SDValue Lo = Op.getOperand(0);
3856 SDValue Hi = Op.getOperand(1);
3857 SDValue Amt = Op.getOperand(2);
3858 EVT AmtVT = Amt.getValueType();
3860 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3861 DAG.getConstant(BitWidth, AmtVT), Amt);
3862 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3863 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3864 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3865 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3866 DAG.getConstant(-BitWidth, AmtVT));
3867 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3868 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3869 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3870 SDValue OutOps[] = { OutLo, OutHi };
3871 return DAG.getMergeValues(OutOps, 2, dl);
3874 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3875 DebugLoc dl = Op.getDebugLoc();
3876 EVT VT = Op.getValueType();
3877 unsigned BitWidth = VT.getSizeInBits();
3878 assert(Op.getNumOperands() == 3 &&
3879 VT == Op.getOperand(1).getValueType() &&
3882 // Expand into a bunch of logical ops, followed by a select_cc.
3883 SDValue Lo = Op.getOperand(0);
3884 SDValue Hi = Op.getOperand(1);
3885 SDValue Amt = Op.getOperand(2);
3886 EVT AmtVT = Amt.getValueType();
3888 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3889 DAG.getConstant(BitWidth, AmtVT), Amt);
3890 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3891 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3892 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3893 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3894 DAG.getConstant(-BitWidth, AmtVT));
3895 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3896 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3897 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3898 Tmp4, Tmp6, ISD::SETLE);
3899 SDValue OutOps[] = { OutLo, OutHi };
3900 return DAG.getMergeValues(OutOps, 2, dl);
3903 //===----------------------------------------------------------------------===//
3904 // Vector related lowering.
3907 /// BuildSplatI - Build a canonical splati of Val with an element size of
3908 /// SplatSize. Cast the result to VT.
3909 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3910 SelectionDAG &DAG, DebugLoc dl) {
3911 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3913 static const EVT VTys[] = { // canonical VT to use for each size.
3914 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3917 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3919 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3923 EVT CanonicalVT = VTys[SplatSize-1];
3925 // Build a canonical splat for this value.
3926 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3927 SmallVector<SDValue, 8> Ops;
3928 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3929 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3930 &Ops[0], Ops.size());
3931 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3934 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3935 /// specified intrinsic ID.
3936 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3937 SelectionDAG &DAG, DebugLoc dl,
3938 EVT DestVT = MVT::Other) {
3939 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3940 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3941 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3944 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3945 /// specified intrinsic ID.
3946 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3947 SDValue Op2, SelectionDAG &DAG,
3948 DebugLoc dl, EVT DestVT = MVT::Other) {
3949 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3950 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3951 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3955 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3956 /// amount. The result has the specified value type.
3957 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3958 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3959 // Force LHS/RHS to be the right type.
3960 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3961 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3964 for (unsigned i = 0; i != 16; ++i)
3966 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3967 return DAG.getNode(ISD::BITCAST, dl, VT, T);
3970 // If this is a case we can't handle, return null and let the default
3971 // expansion code take care of it. If we CAN select this case, and if it
3972 // selects to a single instruction, return Op. Otherwise, if we can codegen
3973 // this case more efficiently than a constant pool load, lower it to the
3974 // sequence of ops that should be used.
3975 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3976 SelectionDAG &DAG) const {
3977 DebugLoc dl = Op.getDebugLoc();
3978 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3979 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3981 // Check if this is a splat of a constant value.
3982 APInt APSplatBits, APSplatUndef;
3983 unsigned SplatBitSize;
3985 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3986 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3989 unsigned SplatBits = APSplatBits.getZExtValue();
3990 unsigned SplatUndef = APSplatUndef.getZExtValue();
3991 unsigned SplatSize = SplatBitSize / 8;
3993 // First, handle single instruction cases.
3996 if (SplatBits == 0) {
3997 // Canonicalize all zero vectors to be v4i32.
3998 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3999 SDValue Z = DAG.getConstant(0, MVT::i32);
4000 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4001 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4006 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4007 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4009 if (SextVal >= -16 && SextVal <= 15)
4010 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4013 // Two instruction sequences.
4015 // If this value is in the range [-32,30] and is even, use:
4016 // tmp = VSPLTI[bhw], result = add tmp, tmp
4017 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4018 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4019 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4020 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4023 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4024 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4026 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4027 // Make -1 and vspltisw -1:
4028 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4030 // Make the VSLW intrinsic, computing 0x8000_0000.
4031 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4034 // xor by OnesV to invert it.
4035 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4036 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4039 // Check to see if this is a wide variety of vsplti*, binop self cases.
4040 static const signed char SplatCsts[] = {
4041 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4042 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4045 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4046 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4047 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4048 int i = SplatCsts[idx];
4050 // Figure out what shift amount will be used by altivec if shifted by i in
4052 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4054 // vsplti + shl self.
4055 if (SextVal == (i << (int)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_vslb, Intrinsic::ppc_altivec_vslh, 0,
4059 Intrinsic::ppc_altivec_vslw
4061 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4062 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4065 // vsplti + srl self.
4066 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4067 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4068 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4069 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4070 Intrinsic::ppc_altivec_vsrw
4072 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4073 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4076 // vsplti + sra self.
4077 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4078 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4079 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4080 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4081 Intrinsic::ppc_altivec_vsraw
4083 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4084 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4087 // vsplti + rol self.
4088 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4089 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4090 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4091 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4092 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4093 Intrinsic::ppc_altivec_vrlw
4095 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4096 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4099 // t = vsplti c, result = vsldoi t, t, 1
4100 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4101 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4102 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4104 // t = vsplti c, result = vsldoi t, t, 2
4105 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4106 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4107 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4109 // t = vsplti c, result = vsldoi t, t, 3
4110 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4111 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4112 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4116 // Three instruction sequences.
4118 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4119 if (SextVal >= 0 && SextVal <= 31) {
4120 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4121 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4122 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4123 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4125 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4126 if (SextVal >= -31 && SextVal <= 0) {
4127 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4128 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4129 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4130 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4136 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4137 /// the specified operations to build the shuffle.
4138 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4139 SDValue RHS, SelectionDAG &DAG,
4141 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4142 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4143 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4146 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4158 if (OpNum == OP_COPY) {
4159 if (LHSID == (1*9+2)*9+3) return LHS;
4160 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4164 SDValue OpLHS, OpRHS;
4165 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4166 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4170 default: llvm_unreachable("Unknown i32 permute!");
4172 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4173 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4174 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4175 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4178 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4179 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4180 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4181 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4184 for (unsigned i = 0; i != 16; ++i)
4185 ShufIdxs[i] = (i&3)+0;
4188 for (unsigned i = 0; i != 16; ++i)
4189 ShufIdxs[i] = (i&3)+4;
4192 for (unsigned i = 0; i != 16; ++i)
4193 ShufIdxs[i] = (i&3)+8;
4196 for (unsigned i = 0; i != 16; ++i)
4197 ShufIdxs[i] = (i&3)+12;
4200 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4202 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4204 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4206 EVT VT = OpLHS.getValueType();
4207 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4208 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4209 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4210 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4213 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4214 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4215 /// return the code it can be lowered into. Worst case, it can always be
4216 /// lowered into a vperm.
4217 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4218 SelectionDAG &DAG) const {
4219 DebugLoc dl = Op.getDebugLoc();
4220 SDValue V1 = Op.getOperand(0);
4221 SDValue V2 = Op.getOperand(1);
4222 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4223 EVT VT = Op.getValueType();
4225 // Cases that are handled by instructions that take permute immediates
4226 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4227 // selected by the instruction selector.
4228 if (V2.getOpcode() == ISD::UNDEF) {
4229 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4230 PPC::isSplatShuffleMask(SVOp, 2) ||
4231 PPC::isSplatShuffleMask(SVOp, 4) ||
4232 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4233 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4234 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4235 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4236 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4237 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4238 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4239 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4240 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4245 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4246 // and produce a fixed permutation. If any of these match, do not lower to
4248 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4249 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4250 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4251 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4252 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4253 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4254 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4255 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4256 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4259 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4260 // perfect shuffle table to emit an optimal matching sequence.
4261 ArrayRef<int> PermMask = SVOp->getMask();
4263 unsigned PFIndexes[4];
4264 bool isFourElementShuffle = true;
4265 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4266 unsigned EltNo = 8; // Start out undef.
4267 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4268 if (PermMask[i*4+j] < 0)
4269 continue; // Undef, ignore it.
4271 unsigned ByteSource = PermMask[i*4+j];
4272 if ((ByteSource & 3) != j) {
4273 isFourElementShuffle = false;
4278 EltNo = ByteSource/4;
4279 } else if (EltNo != ByteSource/4) {
4280 isFourElementShuffle = false;
4284 PFIndexes[i] = EltNo;
4287 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4288 // perfect shuffle vector to determine if it is cost effective to do this as
4289 // discrete instructions, or whether we should use a vperm.
4290 if (isFourElementShuffle) {
4291 // Compute the index in the perfect shuffle table.
4292 unsigned PFTableIndex =
4293 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4295 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4296 unsigned Cost = (PFEntry >> 30);
4298 // Determining when to avoid vperm is tricky. Many things affect the cost
4299 // of vperm, particularly how many times the perm mask needs to be computed.
4300 // For example, if the perm mask can be hoisted out of a loop or is already
4301 // used (perhaps because there are multiple permutes with the same shuffle
4302 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4303 // the loop requires an extra register.
4305 // As a compromise, we only emit discrete instructions if the shuffle can be
4306 // generated in 3 or fewer operations. When we have loop information
4307 // available, if this block is within a loop, we should avoid using vperm
4308 // for 3-operation perms and use a constant pool load instead.
4310 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4313 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4314 // vector that will get spilled to the constant pool.
4315 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4317 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4318 // that it is in input element units, not in bytes. Convert now.
4319 EVT EltVT = V1.getValueType().getVectorElementType();
4320 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4322 SmallVector<SDValue, 16> ResultMask;
4323 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4324 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4326 for (unsigned j = 0; j != BytesPerElement; ++j)
4327 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4331 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4332 &ResultMask[0], ResultMask.size());
4333 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4336 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4337 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4338 /// information about the intrinsic.
4339 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4341 unsigned IntrinsicID =
4342 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4345 switch (IntrinsicID) {
4346 default: return false;
4347 // Comparison predicates.
4348 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4349 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4350 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4351 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4352 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4353 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4354 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4355 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4356 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4357 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4358 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4359 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4360 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4362 // Normal Comparisons.
4363 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4364 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4365 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4366 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4367 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4368 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4369 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4370 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4371 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4372 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4373 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4374 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4375 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4380 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4381 /// lower, do it, otherwise return null.
4382 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4383 SelectionDAG &DAG) const {
4384 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4385 // opcode number of the comparison.
4386 DebugLoc dl = Op.getDebugLoc();
4389 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4390 return SDValue(); // Don't custom lower most intrinsics.
4392 // If this is a non-dot comparison, make the VCMP node and we are done.
4394 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4395 Op.getOperand(1), Op.getOperand(2),
4396 DAG.getConstant(CompareOpc, MVT::i32));
4397 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4400 // Create the PPCISD altivec 'dot' comparison node.
4402 Op.getOperand(2), // LHS
4403 Op.getOperand(3), // RHS
4404 DAG.getConstant(CompareOpc, MVT::i32)
4406 std::vector<EVT> VTs;
4407 VTs.push_back(Op.getOperand(2).getValueType());
4408 VTs.push_back(MVT::Glue);
4409 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4411 // Now that we have the comparison, emit a copy from the CR to a GPR.
4412 // This is flagged to the above dot comparison.
4413 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4414 DAG.getRegister(PPC::CR6, MVT::i32),
4415 CompNode.getValue(1));
4417 // Unpack the result based on how the target uses it.
4418 unsigned BitNo; // Bit # of CR6.
4419 bool InvertBit; // Invert result?
4420 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4421 default: // Can't happen, don't crash on invalid number though.
4422 case 0: // Return the value of the EQ bit of CR6.
4423 BitNo = 0; InvertBit = false;
4425 case 1: // Return the inverted value of the EQ bit of CR6.
4426 BitNo = 0; InvertBit = true;
4428 case 2: // Return the value of the LT bit of CR6.
4429 BitNo = 2; InvertBit = false;
4431 case 3: // Return the inverted value of the LT bit of CR6.
4432 BitNo = 2; InvertBit = true;
4436 // Shift the bit into the low position.
4437 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4438 DAG.getConstant(8-(3-BitNo), MVT::i32));
4440 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4441 DAG.getConstant(1, MVT::i32));
4443 // If we are supposed to, toggle the bit.
4445 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4446 DAG.getConstant(1, MVT::i32));
4450 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4451 SelectionDAG &DAG) const {
4452 DebugLoc dl = Op.getDebugLoc();
4453 // Create a stack slot that is 16-byte aligned.
4454 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4455 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4456 EVT PtrVT = getPointerTy();
4457 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4459 // Store the input value into Value#0 of the stack slot.
4460 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4461 Op.getOperand(0), FIdx, MachinePointerInfo(),
4464 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4465 false, false, false, 0);
4468 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4469 DebugLoc dl = Op.getDebugLoc();
4470 if (Op.getValueType() == MVT::v4i32) {
4471 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4473 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4474 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4476 SDValue RHSSwap = // = vrlw RHS, 16
4477 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4479 // Shrinkify inputs to v8i16.
4480 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4481 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4482 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4484 // Low parts multiplied together, generating 32-bit results (we ignore the
4486 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4487 LHS, RHS, DAG, dl, MVT::v4i32);
4489 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4490 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4491 // Shift the high parts up 16 bits.
4492 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4494 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4495 } else if (Op.getValueType() == MVT::v8i16) {
4496 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4498 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4500 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4501 LHS, RHS, Zero, DAG, dl);
4502 } else if (Op.getValueType() == MVT::v16i8) {
4503 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4505 // Multiply the even 8-bit parts, producing 16-bit sums.
4506 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4507 LHS, RHS, DAG, dl, MVT::v8i16);
4508 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4510 // Multiply the odd 8-bit parts, producing 16-bit sums.
4511 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4512 LHS, RHS, DAG, dl, MVT::v8i16);
4513 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4515 // Merge the results together.
4517 for (unsigned i = 0; i != 8; ++i) {
4519 Ops[i*2+1] = 2*i+1+16;
4521 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4523 llvm_unreachable("Unknown mul to lower!");
4527 /// LowerOperation - Provide custom lowering hooks for some operations.
4529 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4530 switch (Op.getOpcode()) {
4531 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4532 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4533 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4534 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4535 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4536 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4537 case ISD::SETCC: return LowerSETCC(Op, DAG);
4538 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4539 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4541 return LowerVASTART(Op, DAG, PPCSubTarget);
4544 return LowerVAARG(Op, DAG, PPCSubTarget);
4546 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4547 case ISD::DYNAMIC_STACKALLOC:
4548 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4550 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4551 case ISD::FP_TO_UINT:
4552 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4554 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4555 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4557 // Lower 64-bit shifts.
4558 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4559 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4560 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4562 // Vector-related lowering.
4563 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4564 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4565 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4566 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4567 case ISD::MUL: return LowerMUL(Op, DAG);
4569 // Frame & Return address.
4570 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4571 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4575 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4576 SmallVectorImpl<SDValue>&Results,
4577 SelectionDAG &DAG) const {
4578 const TargetMachine &TM = getTargetMachine();
4579 DebugLoc dl = N->getDebugLoc();
4580 switch (N->getOpcode()) {
4582 llvm_unreachable("Do not know how to custom type legalize this operation!");
4584 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4585 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4588 EVT VT = N->getValueType(0);
4590 if (VT == MVT::i64) {
4591 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4593 Results.push_back(NewNode);
4594 Results.push_back(NewNode.getValue(1));
4598 case ISD::FP_ROUND_INREG: {
4599 assert(N->getValueType(0) == MVT::ppcf128);
4600 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4601 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4602 MVT::f64, N->getOperand(0),
4603 DAG.getIntPtrConstant(0));
4604 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4605 MVT::f64, N->getOperand(0),
4606 DAG.getIntPtrConstant(1));
4608 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4609 // of the long double, and puts FPSCR back the way it was. We do not
4610 // actually model FPSCR.
4611 std::vector<EVT> NodeTys;
4612 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4614 NodeTys.push_back(MVT::f64); // Return register
4615 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4616 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4617 MFFSreg = Result.getValue(0);
4618 InFlag = Result.getValue(1);
4621 NodeTys.push_back(MVT::Glue); // Returns a flag
4622 Ops[0] = DAG.getConstant(31, MVT::i32);
4624 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4625 InFlag = Result.getValue(0);
4628 NodeTys.push_back(MVT::Glue); // Returns a flag
4629 Ops[0] = DAG.getConstant(30, MVT::i32);
4631 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4632 InFlag = Result.getValue(0);
4635 NodeTys.push_back(MVT::f64); // result of add
4636 NodeTys.push_back(MVT::Glue); // Returns a flag
4640 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4641 FPreg = Result.getValue(0);
4642 InFlag = Result.getValue(1);
4645 NodeTys.push_back(MVT::f64);
4646 Ops[0] = DAG.getConstant(1, MVT::i32);
4650 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4651 FPreg = Result.getValue(0);
4653 // We know the low half is about to be thrown away, so just use something
4655 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4659 case ISD::FP_TO_SINT:
4660 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4666 //===----------------------------------------------------------------------===//
4667 // Other Lowering Code
4668 //===----------------------------------------------------------------------===//
4671 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4672 bool is64bit, unsigned BinOpcode) const {
4673 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4674 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4676 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4677 MachineFunction *F = BB->getParent();
4678 MachineFunction::iterator It = BB;
4681 unsigned dest = MI->getOperand(0).getReg();
4682 unsigned ptrA = MI->getOperand(1).getReg();
4683 unsigned ptrB = MI->getOperand(2).getReg();
4684 unsigned incr = MI->getOperand(3).getReg();
4685 DebugLoc dl = MI->getDebugLoc();
4687 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4688 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4689 F->insert(It, loopMBB);
4690 F->insert(It, exitMBB);
4691 exitMBB->splice(exitMBB->begin(), BB,
4692 llvm::next(MachineBasicBlock::iterator(MI)),
4694 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4696 MachineRegisterInfo &RegInfo = F->getRegInfo();
4697 unsigned TmpReg = (!BinOpcode) ? incr :
4698 RegInfo.createVirtualRegister(
4699 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4700 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4704 // fallthrough --> loopMBB
4705 BB->addSuccessor(loopMBB);
4708 // l[wd]arx dest, ptr
4709 // add r0, dest, incr
4710 // st[wd]cx. r0, ptr
4712 // fallthrough --> exitMBB
4714 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4715 .addReg(ptrA).addReg(ptrB);
4717 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4718 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4719 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4720 BuildMI(BB, dl, TII->get(PPC::BCC))
4721 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4722 BB->addSuccessor(loopMBB);
4723 BB->addSuccessor(exitMBB);
4732 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4733 MachineBasicBlock *BB,
4734 bool is8bit, // operation
4735 unsigned BinOpcode) const {
4736 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4737 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4738 // In 64 bit mode we have to use 64 bits for addresses, even though the
4739 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4740 // registers without caring whether they're 32 or 64, but here we're
4741 // doing actual arithmetic on the addresses.
4742 bool is64bit = PPCSubTarget.isPPC64();
4743 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4745 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4746 MachineFunction *F = BB->getParent();
4747 MachineFunction::iterator It = BB;
4750 unsigned dest = MI->getOperand(0).getReg();
4751 unsigned ptrA = MI->getOperand(1).getReg();
4752 unsigned ptrB = MI->getOperand(2).getReg();
4753 unsigned incr = MI->getOperand(3).getReg();
4754 DebugLoc dl = MI->getDebugLoc();
4756 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4757 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4758 F->insert(It, loopMBB);
4759 F->insert(It, exitMBB);
4760 exitMBB->splice(exitMBB->begin(), BB,
4761 llvm::next(MachineBasicBlock::iterator(MI)),
4763 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4765 MachineRegisterInfo &RegInfo = F->getRegInfo();
4766 const TargetRegisterClass *RC =
4767 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4768 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4769 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4770 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4771 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4772 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4773 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4774 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4775 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4776 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4777 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4778 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4779 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4781 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4785 // fallthrough --> loopMBB
4786 BB->addSuccessor(loopMBB);
4788 // The 4-byte load must be aligned, while a char or short may be
4789 // anywhere in the word. Hence all this nasty bookkeeping code.
4790 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4791 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4792 // xori shift, shift1, 24 [16]
4793 // rlwinm ptr, ptr1, 0, 0, 29
4794 // slw incr2, incr, shift
4795 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4796 // slw mask, mask2, shift
4798 // lwarx tmpDest, ptr
4799 // add tmp, tmpDest, incr2
4800 // andc tmp2, tmpDest, mask
4801 // and tmp3, tmp, mask
4802 // or tmp4, tmp3, tmp2
4805 // fallthrough --> exitMBB
4806 // srw dest, tmpDest, shift
4807 if (ptrA != ZeroReg) {
4808 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4809 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4810 .addReg(ptrA).addReg(ptrB);
4814 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4815 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4816 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4817 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4819 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4820 .addReg(Ptr1Reg).addImm(0).addImm(61);
4822 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4823 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4824 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4825 .addReg(incr).addReg(ShiftReg);
4827 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4829 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4830 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4832 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4833 .addReg(Mask2Reg).addReg(ShiftReg);
4836 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4837 .addReg(ZeroReg).addReg(PtrReg);
4839 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4840 .addReg(Incr2Reg).addReg(TmpDestReg);
4841 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4842 .addReg(TmpDestReg).addReg(MaskReg);
4843 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4844 .addReg(TmpReg).addReg(MaskReg);
4845 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4846 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4847 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4848 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4849 BuildMI(BB, dl, TII->get(PPC::BCC))
4850 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4851 BB->addSuccessor(loopMBB);
4852 BB->addSuccessor(exitMBB);
4857 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4863 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4864 MachineBasicBlock *BB) const {
4865 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4867 // To "insert" these instructions we actually have to insert their
4868 // control-flow patterns.
4869 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4870 MachineFunction::iterator It = BB;
4873 MachineFunction *F = BB->getParent();
4875 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4876 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4877 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4878 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4879 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4881 // The incoming instruction knows the destination vreg to set, the
4882 // condition code register to branch on, the true/false values to
4883 // select between, and a branch opcode to use.
4888 // cmpTY ccX, r1, r2
4890 // fallthrough --> copy0MBB
4891 MachineBasicBlock *thisMBB = BB;
4892 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4893 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4894 unsigned SelectPred = MI->getOperand(4).getImm();
4895 DebugLoc dl = MI->getDebugLoc();
4896 F->insert(It, copy0MBB);
4897 F->insert(It, sinkMBB);
4899 // Transfer the remainder of BB and its successor edges to sinkMBB.
4900 sinkMBB->splice(sinkMBB->begin(), BB,
4901 llvm::next(MachineBasicBlock::iterator(MI)),
4903 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4905 // Next, add the true and fallthrough blocks as its successors.
4906 BB->addSuccessor(copy0MBB);
4907 BB->addSuccessor(sinkMBB);
4909 BuildMI(BB, dl, TII->get(PPC::BCC))
4910 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4913 // %FalseValue = ...
4914 // # fallthrough to sinkMBB
4917 // Update machine-CFG edges
4918 BB->addSuccessor(sinkMBB);
4921 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4924 BuildMI(*BB, BB->begin(), dl,
4925 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4926 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4927 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4929 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4930 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4931 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4932 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4933 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4934 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4935 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4936 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4938 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4939 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4940 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4941 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4942 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4943 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4944 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4945 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4947 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4948 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4949 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4950 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4951 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4952 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4953 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4954 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4956 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4957 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4958 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4959 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4960 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4961 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4962 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4963 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4965 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4966 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4967 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4968 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4969 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4970 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4971 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4972 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4974 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4975 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4976 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4977 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4978 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4979 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4980 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4981 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4983 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4984 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4985 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4986 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4987 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4988 BB = EmitAtomicBinary(MI, BB, false, 0);
4989 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4990 BB = EmitAtomicBinary(MI, BB, true, 0);
4992 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4993 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4994 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4996 unsigned dest = MI->getOperand(0).getReg();
4997 unsigned ptrA = MI->getOperand(1).getReg();
4998 unsigned ptrB = MI->getOperand(2).getReg();
4999 unsigned oldval = MI->getOperand(3).getReg();
5000 unsigned newval = MI->getOperand(4).getReg();
5001 DebugLoc dl = MI->getDebugLoc();
5003 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5004 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5005 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5006 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5007 F->insert(It, loop1MBB);
5008 F->insert(It, loop2MBB);
5009 F->insert(It, midMBB);
5010 F->insert(It, exitMBB);
5011 exitMBB->splice(exitMBB->begin(), BB,
5012 llvm::next(MachineBasicBlock::iterator(MI)),
5014 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5018 // fallthrough --> loopMBB
5019 BB->addSuccessor(loop1MBB);
5022 // l[wd]arx dest, ptr
5023 // cmp[wd] dest, oldval
5026 // st[wd]cx. newval, ptr
5030 // st[wd]cx. dest, ptr
5033 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5034 .addReg(ptrA).addReg(ptrB);
5035 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5036 .addReg(oldval).addReg(dest);
5037 BuildMI(BB, dl, TII->get(PPC::BCC))
5038 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5039 BB->addSuccessor(loop2MBB);
5040 BB->addSuccessor(midMBB);
5043 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5044 .addReg(newval).addReg(ptrA).addReg(ptrB);
5045 BuildMI(BB, dl, TII->get(PPC::BCC))
5046 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5047 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5048 BB->addSuccessor(loop1MBB);
5049 BB->addSuccessor(exitMBB);
5052 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5053 .addReg(dest).addReg(ptrA).addReg(ptrB);
5054 BB->addSuccessor(exitMBB);
5059 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5060 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5061 // We must use 64-bit registers for addresses when targeting 64-bit,
5062 // since we're actually doing arithmetic on them. Other registers
5064 bool is64bit = PPCSubTarget.isPPC64();
5065 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5067 unsigned dest = MI->getOperand(0).getReg();
5068 unsigned ptrA = MI->getOperand(1).getReg();
5069 unsigned ptrB = MI->getOperand(2).getReg();
5070 unsigned oldval = MI->getOperand(3).getReg();
5071 unsigned newval = MI->getOperand(4).getReg();
5072 DebugLoc dl = MI->getDebugLoc();
5074 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5075 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5076 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5077 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5078 F->insert(It, loop1MBB);
5079 F->insert(It, loop2MBB);
5080 F->insert(It, midMBB);
5081 F->insert(It, exitMBB);
5082 exitMBB->splice(exitMBB->begin(), BB,
5083 llvm::next(MachineBasicBlock::iterator(MI)),
5085 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5087 MachineRegisterInfo &RegInfo = F->getRegInfo();
5088 const TargetRegisterClass *RC =
5089 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5090 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5091 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5092 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5093 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5094 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5095 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5096 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5097 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5098 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5099 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5100 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5101 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5102 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5103 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5105 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5106 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5109 // fallthrough --> loopMBB
5110 BB->addSuccessor(loop1MBB);
5112 // The 4-byte load must be aligned, while a char or short may be
5113 // anywhere in the word. Hence all this nasty bookkeeping code.
5114 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5115 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5116 // xori shift, shift1, 24 [16]
5117 // rlwinm ptr, ptr1, 0, 0, 29
5118 // slw newval2, newval, shift
5119 // slw oldval2, oldval,shift
5120 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5121 // slw mask, mask2, shift
5122 // and newval3, newval2, mask
5123 // and oldval3, oldval2, mask
5125 // lwarx tmpDest, ptr
5126 // and tmp, tmpDest, mask
5127 // cmpw tmp, oldval3
5130 // andc tmp2, tmpDest, mask
5131 // or tmp4, tmp2, newval3
5136 // stwcx. tmpDest, ptr
5138 // srw dest, tmpDest, shift
5139 if (ptrA != ZeroReg) {
5140 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5141 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5142 .addReg(ptrA).addReg(ptrB);
5146 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5147 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5148 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5149 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5151 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5152 .addReg(Ptr1Reg).addImm(0).addImm(61);
5154 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5155 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5156 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5157 .addReg(newval).addReg(ShiftReg);
5158 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5159 .addReg(oldval).addReg(ShiftReg);
5161 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5163 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5164 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5165 .addReg(Mask3Reg).addImm(65535);
5167 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5168 .addReg(Mask2Reg).addReg(ShiftReg);
5169 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5170 .addReg(NewVal2Reg).addReg(MaskReg);
5171 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5172 .addReg(OldVal2Reg).addReg(MaskReg);
5175 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5176 .addReg(ZeroReg).addReg(PtrReg);
5177 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5178 .addReg(TmpDestReg).addReg(MaskReg);
5179 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5180 .addReg(TmpReg).addReg(OldVal3Reg);
5181 BuildMI(BB, dl, TII->get(PPC::BCC))
5182 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5183 BB->addSuccessor(loop2MBB);
5184 BB->addSuccessor(midMBB);
5187 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5188 .addReg(TmpDestReg).addReg(MaskReg);
5189 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5190 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5191 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5192 .addReg(ZeroReg).addReg(PtrReg);
5193 BuildMI(BB, dl, TII->get(PPC::BCC))
5194 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5195 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5196 BB->addSuccessor(loop1MBB);
5197 BB->addSuccessor(exitMBB);
5200 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5201 .addReg(ZeroReg).addReg(PtrReg);
5202 BB->addSuccessor(exitMBB);
5207 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5210 llvm_unreachable("Unexpected instr type to insert");
5213 MI->eraseFromParent(); // The pseudo instruction is gone now.
5217 //===----------------------------------------------------------------------===//
5218 // Target Optimization Hooks
5219 //===----------------------------------------------------------------------===//
5221 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5222 DAGCombinerInfo &DCI) const {
5223 const TargetMachine &TM = getTargetMachine();
5224 SelectionDAG &DAG = DCI.DAG;
5225 DebugLoc dl = N->getDebugLoc();
5226 switch (N->getOpcode()) {
5229 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5230 if (C->isNullValue()) // 0 << V -> 0.
5231 return N->getOperand(0);
5235 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5236 if (C->isNullValue()) // 0 >>u V -> 0.
5237 return N->getOperand(0);
5241 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5242 if (C->isNullValue() || // 0 >>s V -> 0.
5243 C->isAllOnesValue()) // -1 >>s V -> -1.
5244 return N->getOperand(0);
5248 case ISD::SINT_TO_FP:
5249 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5250 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5251 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5252 // We allow the src/dst to be either f32/f64, but the intermediate
5253 // type must be i64.
5254 if (N->getOperand(0).getValueType() == MVT::i64 &&
5255 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5256 SDValue Val = N->getOperand(0).getOperand(0);
5257 if (Val.getValueType() == MVT::f32) {
5258 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5259 DCI.AddToWorklist(Val.getNode());
5262 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5263 DCI.AddToWorklist(Val.getNode());
5264 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5265 DCI.AddToWorklist(Val.getNode());
5266 if (N->getValueType(0) == MVT::f32) {
5267 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5268 DAG.getIntPtrConstant(0));
5269 DCI.AddToWorklist(Val.getNode());
5272 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5273 // If the intermediate type is i32, we can avoid the load/store here
5280 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5281 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5282 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5283 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5284 N->getOperand(1).getValueType() == MVT::i32 &&
5285 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5286 SDValue Val = N->getOperand(1).getOperand(0);
5287 if (Val.getValueType() == MVT::f32) {
5288 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5289 DCI.AddToWorklist(Val.getNode());
5291 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5292 DCI.AddToWorklist(Val.getNode());
5294 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5295 N->getOperand(2), N->getOperand(3));
5296 DCI.AddToWorklist(Val.getNode());
5300 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5301 if (cast<StoreSDNode>(N)->isUnindexed() &&
5302 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5303 N->getOperand(1).getNode()->hasOneUse() &&
5304 (N->getOperand(1).getValueType() == MVT::i32 ||
5305 N->getOperand(1).getValueType() == MVT::i16)) {
5306 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5307 // Do an any-extend to 32-bits if this is a half-word input.
5308 if (BSwapOp.getValueType() == MVT::i16)
5309 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5312 N->getOperand(0), BSwapOp, N->getOperand(2),
5313 DAG.getValueType(N->getOperand(1).getValueType())
5316 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5317 Ops, array_lengthof(Ops),
5318 cast<StoreSDNode>(N)->getMemoryVT(),
5319 cast<StoreSDNode>(N)->getMemOperand());
5323 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5324 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5325 N->getOperand(0).hasOneUse() &&
5326 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5327 SDValue Load = N->getOperand(0);
5328 LoadSDNode *LD = cast<LoadSDNode>(Load);
5329 // Create the byte-swapping load.
5331 LD->getChain(), // Chain
5332 LD->getBasePtr(), // Ptr
5333 DAG.getValueType(N->getValueType(0)) // VT
5336 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5337 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5338 LD->getMemoryVT(), LD->getMemOperand());
5340 // If this is an i16 load, insert the truncate.
5341 SDValue ResVal = BSLoad;
5342 if (N->getValueType(0) == MVT::i16)
5343 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5345 // First, combine the bswap away. This makes the value produced by the
5347 DCI.CombineTo(N, ResVal);
5349 // Next, combine the load away, we give it a bogus result value but a real
5350 // chain result. The result value is dead because the bswap is dead.
5351 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5353 // Return N so it doesn't get rechecked!
5354 return SDValue(N, 0);
5358 case PPCISD::VCMP: {
5359 // If a VCMPo node already exists with exactly the same operands as this
5360 // node, use its result instead of this node (VCMPo computes both a CR6 and
5361 // a normal output).
5363 if (!N->getOperand(0).hasOneUse() &&
5364 !N->getOperand(1).hasOneUse() &&
5365 !N->getOperand(2).hasOneUse()) {
5367 // Scan all of the users of the LHS, looking for VCMPo's that match.
5368 SDNode *VCMPoNode = 0;
5370 SDNode *LHSN = N->getOperand(0).getNode();
5371 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5373 if (UI->getOpcode() == PPCISD::VCMPo &&
5374 UI->getOperand(1) == N->getOperand(1) &&
5375 UI->getOperand(2) == N->getOperand(2) &&
5376 UI->getOperand(0) == N->getOperand(0)) {
5381 // If there is no VCMPo node, or if the flag value has a single use, don't
5383 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5386 // Look at the (necessarily single) use of the flag value. If it has a
5387 // chain, this transformation is more complex. Note that multiple things
5388 // could use the value result, which we should ignore.
5389 SDNode *FlagUser = 0;
5390 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5391 FlagUser == 0; ++UI) {
5392 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5394 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5395 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5402 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5403 // give up for right now.
5404 if (FlagUser->getOpcode() == PPCISD::MFCR)
5405 return SDValue(VCMPoNode, 0);
5410 // If this is a branch on an altivec predicate comparison, lower this so
5411 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5412 // lowering is done pre-legalize, because the legalizer lowers the predicate
5413 // compare down to code that is difficult to reassemble.
5414 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5415 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5419 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5420 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5421 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5422 assert(isDot && "Can't compare against a vector result!");
5424 // If this is a comparison against something other than 0/1, then we know
5425 // that the condition is never/always true.
5426 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5427 if (Val != 0 && Val != 1) {
5428 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5429 return N->getOperand(0);
5430 // Always !=, turn it into an unconditional branch.
5431 return DAG.getNode(ISD::BR, dl, MVT::Other,
5432 N->getOperand(0), N->getOperand(4));
5435 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5437 // Create the PPCISD altivec 'dot' comparison node.
5438 std::vector<EVT> VTs;
5440 LHS.getOperand(2), // LHS of compare
5441 LHS.getOperand(3), // RHS of compare
5442 DAG.getConstant(CompareOpc, MVT::i32)
5444 VTs.push_back(LHS.getOperand(2).getValueType());
5445 VTs.push_back(MVT::Glue);
5446 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5448 // Unpack the result based on how the target uses it.
5449 PPC::Predicate CompOpc;
5450 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5451 default: // Can't happen, don't crash on invalid number though.
5452 case 0: // Branch on the value of the EQ bit of CR6.
5453 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5455 case 1: // Branch on the inverted value of the EQ bit of CR6.
5456 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5458 case 2: // Branch on the value of the LT bit of CR6.
5459 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5461 case 3: // Branch on the inverted value of the LT bit of CR6.
5462 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5466 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5467 DAG.getConstant(CompOpc, MVT::i32),
5468 DAG.getRegister(PPC::CR6, MVT::i32),
5469 N->getOperand(4), CompNode.getValue(1));
5478 //===----------------------------------------------------------------------===//
5479 // Inline Assembly Support
5480 //===----------------------------------------------------------------------===//
5482 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5486 const SelectionDAG &DAG,
5487 unsigned Depth) const {
5488 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5489 switch (Op.getOpcode()) {
5491 case PPCISD::LBRX: {
5492 // lhbrx is known to have the top bits cleared out.
5493 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5494 KnownZero = 0xFFFF0000;
5497 case ISD::INTRINSIC_WO_CHAIN: {
5498 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5500 case Intrinsic::ppc_altivec_vcmpbfp_p:
5501 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5502 case Intrinsic::ppc_altivec_vcmpequb_p:
5503 case Intrinsic::ppc_altivec_vcmpequh_p:
5504 case Intrinsic::ppc_altivec_vcmpequw_p:
5505 case Intrinsic::ppc_altivec_vcmpgefp_p:
5506 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5507 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5508 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5509 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5510 case Intrinsic::ppc_altivec_vcmpgtub_p:
5511 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5512 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5513 KnownZero = ~1U; // All bits but the low one are known to be zero.
5521 /// getConstraintType - Given a constraint, return the type of
5522 /// constraint it is for this target.
5523 PPCTargetLowering::ConstraintType
5524 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5525 if (Constraint.size() == 1) {
5526 switch (Constraint[0]) {
5533 return C_RegisterClass;
5536 return TargetLowering::getConstraintType(Constraint);
5539 /// Examine constraint type and operand type and determine a weight value.
5540 /// This object must already have been set up with the operand type
5541 /// and the current alternative constraint selected.
5542 TargetLowering::ConstraintWeight
5543 PPCTargetLowering::getSingleConstraintMatchWeight(
5544 AsmOperandInfo &info, const char *constraint) const {
5545 ConstraintWeight weight = CW_Invalid;
5546 Value *CallOperandVal = info.CallOperandVal;
5547 // If we don't have a value, we can't do a match,
5548 // but allow it at the lowest weight.
5549 if (CallOperandVal == NULL)
5551 Type *type = CallOperandVal->getType();
5552 // Look at the constraint type.
5553 switch (*constraint) {
5555 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5558 if (type->isIntegerTy())
5559 weight = CW_Register;
5562 if (type->isFloatTy())
5563 weight = CW_Register;
5566 if (type->isDoubleTy())
5567 weight = CW_Register;
5570 if (type->isVectorTy())
5571 weight = CW_Register;
5574 weight = CW_Register;
5580 std::pair<unsigned, const TargetRegisterClass*>
5581 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5583 if (Constraint.size() == 1) {
5584 // GCC RS6000 Constraint Letters
5585 switch (Constraint[0]) {
5588 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5589 return std::make_pair(0U, PPC::G8RCRegisterClass);
5590 return std::make_pair(0U, PPC::GPRCRegisterClass);
5593 return std::make_pair(0U, PPC::F4RCRegisterClass);
5594 else if (VT == MVT::f64)
5595 return std::make_pair(0U, PPC::F8RCRegisterClass);
5598 return std::make_pair(0U, PPC::VRRCRegisterClass);
5600 return std::make_pair(0U, PPC::CRRCRegisterClass);
5604 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5608 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5609 /// vector. If it is invalid, don't add anything to Ops.
5610 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5611 std::string &Constraint,
5612 std::vector<SDValue>&Ops,
5613 SelectionDAG &DAG) const {
5614 SDValue Result(0,0);
5616 // Only support length 1 constraints.
5617 if (Constraint.length() > 1) return;
5619 char Letter = Constraint[0];
5630 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5631 if (!CST) return; // Must be an immediate to match.
5632 unsigned Value = CST->getZExtValue();
5634 default: llvm_unreachable("Unknown constraint letter!");
5635 case 'I': // "I" is a signed 16-bit constant.
5636 if ((short)Value == (int)Value)
5637 Result = DAG.getTargetConstant(Value, Op.getValueType());
5639 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5640 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5641 if ((short)Value == 0)
5642 Result = DAG.getTargetConstant(Value, Op.getValueType());
5644 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5645 if ((Value >> 16) == 0)
5646 Result = DAG.getTargetConstant(Value, Op.getValueType());
5648 case 'M': // "M" is a constant that is greater than 31.
5650 Result = DAG.getTargetConstant(Value, Op.getValueType());
5652 case 'N': // "N" is a positive constant that is an exact power of two.
5653 if ((int)Value > 0 && isPowerOf2_32(Value))
5654 Result = DAG.getTargetConstant(Value, Op.getValueType());
5656 case 'O': // "O" is the constant zero.
5658 Result = DAG.getTargetConstant(Value, Op.getValueType());
5660 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5661 if ((short)-Value == (int)-Value)
5662 Result = DAG.getTargetConstant(Value, Op.getValueType());
5669 if (Result.getNode()) {
5670 Ops.push_back(Result);
5674 // Handle standard constraint letters.
5675 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5678 // isLegalAddressingMode - Return true if the addressing mode represented
5679 // by AM is legal for this target, for a load/store of the specified type.
5680 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5682 // FIXME: PPC does not allow r+i addressing modes for vectors!
5684 // PPC allows a sign-extended 16-bit immediate field.
5685 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5688 // No global is ever allowed as a base.
5692 // PPC only support r+r,
5694 case 0: // "r+i" or just "i", depending on HasBaseReg.
5697 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5699 // Otherwise we have r+r or r+i.
5702 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5704 // Allow 2*r as r+r.
5707 // No other scales are supported.
5714 /// isLegalAddressImmediate - Return true if the integer value can be used
5715 /// as the offset of the target addressing mode for load / store of the
5717 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5718 // PPC allows a sign-extended 16-bit immediate field.
5719 return (V > -(1 << 16) && V < (1 << 16)-1);
5722 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5726 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5727 SelectionDAG &DAG) const {
5728 MachineFunction &MF = DAG.getMachineFunction();
5729 MachineFrameInfo *MFI = MF.getFrameInfo();
5730 MFI->setReturnAddressIsTaken(true);
5732 DebugLoc dl = Op.getDebugLoc();
5733 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5735 // Make sure the function does not optimize away the store of the RA to
5737 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5738 FuncInfo->setLRStoreRequired();
5739 bool isPPC64 = PPCSubTarget.isPPC64();
5740 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5743 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5746 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5747 isPPC64? MVT::i64 : MVT::i32);
5748 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5749 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5751 MachinePointerInfo(), false, false, false, 0);
5754 // Just load the return address off the stack.
5755 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5756 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5757 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5760 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5761 SelectionDAG &DAG) const {
5762 DebugLoc dl = Op.getDebugLoc();
5763 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5765 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5766 bool isPPC64 = PtrVT == MVT::i64;
5768 MachineFunction &MF = DAG.getMachineFunction();
5769 MachineFrameInfo *MFI = MF.getFrameInfo();
5770 MFI->setFrameAddressIsTaken(true);
5771 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
5772 MFI->hasVarSizedObjects()) &&
5773 MFI->getStackSize() &&
5774 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5775 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5776 (is31 ? PPC::R31 : PPC::R1);
5777 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5780 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5781 FrameAddr, MachinePointerInfo(), false, false,
5787 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5788 // The PowerPC target isn't yet aware of offsets.
5792 /// getOptimalMemOpType - Returns the target specific optimal type for load
5793 /// and store operations as a result of memset, memcpy, and memmove
5794 /// lowering. If DstAlign is zero that means it's safe to destination
5795 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5796 /// means there isn't a need to check it against alignment requirement,
5797 /// probably because the source does not need to be loaded. If
5798 /// 'IsZeroVal' is true, that means it's safe to return a
5799 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5800 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5801 /// constant so it does not need to be loaded.
5802 /// It returns EVT::Other if the type should be determined using generic
5803 /// target-independent logic.
5804 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5805 unsigned DstAlign, unsigned SrcAlign,
5808 MachineFunction &MF) const {
5809 if (this->PPCSubTarget.isPPC64()) {