1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCMachineFunctionInfo.h"
16 #include "PPCPerfectShuffle.h"
17 #include "PPCTargetMachine.h"
18 #include "MCTargetDesc/PPCPredicates.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
40 CCValAssign::LocInfo &LocInfo,
41 ISD::ArgFlagsTy &ArgFlags,
43 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
54 static cl::opt<bool> 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::GPRCRegClass);
80 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
81 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
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 if (TM.getSubtarget<PPCSubtarget>().isSVR4ABI()) {
230 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
231 // VAARG always uses double-word chunks, so promote anything smaller.
232 setOperationAction(ISD::VAARG, MVT::i1, Promote);
233 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
234 setOperationAction(ISD::VAARG, MVT::i8, Promote);
235 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
236 setOperationAction(ISD::VAARG, MVT::i16, Promote);
237 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
238 setOperationAction(ISD::VAARG, MVT::i32, Promote);
239 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
240 setOperationAction(ISD::VAARG, MVT::Other, Expand);
242 // VAARG is custom lowered with the 32-bit SVR4 ABI.
243 setOperationAction(ISD::VAARG, MVT::Other, Custom);
244 setOperationAction(ISD::VAARG, MVT::i64, Custom);
247 setOperationAction(ISD::VAARG, MVT::Other, Expand);
249 // Use the default implementation.
250 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
251 setOperationAction(ISD::VAEND , MVT::Other, Expand);
252 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
253 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
254 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
255 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
257 // We want to custom lower some of our intrinsics.
258 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
260 // Comparisons that require checking two conditions.
261 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
262 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
263 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
264 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
265 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
266 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
267 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
270 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
271 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
272 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
274 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
275 // They also have instructions for converting between i64 and fp.
276 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
277 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
278 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
279 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
280 // This is just the low 32 bits of a (signed) fp->i64 conversion.
281 // We cannot do this with Promote because i64 is not a legal type.
282 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
284 // FIXME: disable this lowered code. This generates 64-bit register values,
285 // and we don't model the fact that the top part is clobbered by calls. We
286 // need to flag these together so that the value isn't live across a call.
287 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
289 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
290 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
293 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
294 // 64-bit PowerPC implementations can support i64 types directly
295 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
296 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
297 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
298 // 64-bit PowerPC wants to expand i128 shifts itself.
299 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
300 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
301 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
303 // 32-bit PowerPC wants to expand i64 shifts itself.
304 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
305 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
306 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
309 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
310 // First set operation action for all vector types to expand. Then we
311 // will selectively turn on ones that can be effectively codegen'd.
312 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
313 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
314 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
316 // add/sub are legal for all supported vector VT's.
317 setOperationAction(ISD::ADD , VT, Legal);
318 setOperationAction(ISD::SUB , VT, Legal);
320 // We promote all shuffles to v16i8.
321 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
322 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
324 // We promote all non-typed operations to v4i32.
325 setOperationAction(ISD::AND , VT, Promote);
326 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
327 setOperationAction(ISD::OR , VT, Promote);
328 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
329 setOperationAction(ISD::XOR , VT, Promote);
330 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
331 setOperationAction(ISD::LOAD , VT, Promote);
332 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
333 setOperationAction(ISD::SELECT, VT, Promote);
334 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
335 setOperationAction(ISD::STORE, VT, Promote);
336 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
338 // No other operations are legal.
339 setOperationAction(ISD::MUL , VT, Expand);
340 setOperationAction(ISD::SDIV, VT, Expand);
341 setOperationAction(ISD::SREM, VT, Expand);
342 setOperationAction(ISD::UDIV, VT, Expand);
343 setOperationAction(ISD::UREM, VT, Expand);
344 setOperationAction(ISD::FDIV, VT, Expand);
345 setOperationAction(ISD::FNEG, VT, Expand);
346 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
347 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
348 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
349 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
350 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
351 setOperationAction(ISD::UDIVREM, VT, Expand);
352 setOperationAction(ISD::SDIVREM, VT, Expand);
353 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
354 setOperationAction(ISD::FPOW, VT, Expand);
355 setOperationAction(ISD::CTPOP, VT, Expand);
356 setOperationAction(ISD::CTLZ, VT, Expand);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
358 setOperationAction(ISD::CTTZ, VT, Expand);
359 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
362 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
363 // with merges, splats, etc.
364 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
366 setOperationAction(ISD::AND , MVT::v4i32, Legal);
367 setOperationAction(ISD::OR , MVT::v4i32, Legal);
368 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
369 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
370 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
371 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
373 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
374 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
375 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
376 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
378 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
379 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
380 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
381 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
383 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
384 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
386 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
387 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
388 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
389 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
392 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport())
393 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
395 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
396 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
398 setBooleanContents(ZeroOrOneBooleanContent);
399 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
401 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
402 setStackPointerRegisterToSaveRestore(PPC::X1);
403 setExceptionPointerRegister(PPC::X3);
404 setExceptionSelectorRegister(PPC::X4);
406 setStackPointerRegisterToSaveRestore(PPC::R1);
407 setExceptionPointerRegister(PPC::R3);
408 setExceptionSelectorRegister(PPC::R4);
411 // We have target-specific dag combine patterns for the following nodes:
412 setTargetDAGCombine(ISD::SINT_TO_FP);
413 setTargetDAGCombine(ISD::STORE);
414 setTargetDAGCombine(ISD::BR_CC);
415 setTargetDAGCombine(ISD::BSWAP);
417 // Darwin long double math library functions have $LDBL128 appended.
418 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
419 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
420 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
421 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
422 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
423 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
424 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
425 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
426 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
427 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
428 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
431 setMinFunctionAlignment(2);
432 if (PPCSubTarget.isDarwin())
433 setPrefFunctionAlignment(4);
435 setInsertFencesForAtomic(true);
437 setSchedulingPreference(Sched::Hybrid);
439 computeRegisterProperties();
442 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
443 /// function arguments in the caller parameter area.
444 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
445 const TargetMachine &TM = getTargetMachine();
446 // Darwin passes everything on 4 byte boundary.
447 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
450 // 16byte and wider vectors are passed on 16byte boundary.
451 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
452 if (VTy->getBitWidth() >= 128)
455 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
456 if (PPCSubTarget.isPPC64())
462 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
465 case PPCISD::FSEL: return "PPCISD::FSEL";
466 case PPCISD::FCFID: return "PPCISD::FCFID";
467 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
468 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
469 case PPCISD::STFIWX: return "PPCISD::STFIWX";
470 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
471 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
472 case PPCISD::VPERM: return "PPCISD::VPERM";
473 case PPCISD::Hi: return "PPCISD::Hi";
474 case PPCISD::Lo: return "PPCISD::Lo";
475 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
476 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
477 case PPCISD::LOAD: return "PPCISD::LOAD";
478 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
479 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
480 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
481 case PPCISD::SRL: return "PPCISD::SRL";
482 case PPCISD::SRA: return "PPCISD::SRA";
483 case PPCISD::SHL: return "PPCISD::SHL";
484 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
485 case PPCISD::STD_32: return "PPCISD::STD_32";
486 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
487 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
488 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
489 case PPCISD::NOP: return "PPCISD::NOP";
490 case PPCISD::MTCTR: return "PPCISD::MTCTR";
491 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
492 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
493 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
494 case PPCISD::MFCR: return "PPCISD::MFCR";
495 case PPCISD::VCMP: return "PPCISD::VCMP";
496 case PPCISD::VCMPo: return "PPCISD::VCMPo";
497 case PPCISD::LBRX: return "PPCISD::LBRX";
498 case PPCISD::STBRX: return "PPCISD::STBRX";
499 case PPCISD::LARX: return "PPCISD::LARX";
500 case PPCISD::STCX: return "PPCISD::STCX";
501 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
502 case PPCISD::MFFS: return "PPCISD::MFFS";
503 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
504 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
505 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
506 case PPCISD::MTFSF: return "PPCISD::MTFSF";
507 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
511 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
515 //===----------------------------------------------------------------------===//
516 // Node matching predicates, for use by the tblgen matching code.
517 //===----------------------------------------------------------------------===//
519 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
520 static bool isFloatingPointZero(SDValue Op) {
521 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
522 return CFP->getValueAPF().isZero();
523 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
524 // Maybe this has already been legalized into the constant pool?
525 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
526 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
527 return CFP->getValueAPF().isZero();
532 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
533 /// true if Op is undef or if it matches the specified value.
534 static bool isConstantOrUndef(int Op, int Val) {
535 return Op < 0 || Op == Val;
538 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
539 /// VPKUHUM instruction.
540 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
542 for (unsigned i = 0; i != 16; ++i)
543 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
546 for (unsigned i = 0; i != 8; ++i)
547 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
548 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
554 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
555 /// VPKUWUM instruction.
556 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
558 for (unsigned i = 0; i != 16; i += 2)
559 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
560 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
563 for (unsigned i = 0; i != 8; i += 2)
564 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
565 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
566 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
567 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
573 /// isVMerge - Common function, used to match vmrg* shuffles.
575 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
576 unsigned LHSStart, unsigned RHSStart) {
577 assert(N->getValueType(0) == MVT::v16i8 &&
578 "PPC only supports shuffles by bytes!");
579 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
580 "Unsupported merge size!");
582 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
583 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
584 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
585 LHSStart+j+i*UnitSize) ||
586 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
587 RHSStart+j+i*UnitSize))
593 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
594 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
595 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
598 return isVMerge(N, UnitSize, 8, 24);
599 return isVMerge(N, UnitSize, 8, 8);
602 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
603 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
604 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
607 return isVMerge(N, UnitSize, 0, 16);
608 return isVMerge(N, UnitSize, 0, 0);
612 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
613 /// amount, otherwise return -1.
614 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
615 assert(N->getValueType(0) == MVT::v16i8 &&
616 "PPC only supports shuffles by bytes!");
618 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
620 // Find the first non-undef value in the shuffle mask.
622 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
625 if (i == 16) return -1; // all undef.
627 // Otherwise, check to see if the rest of the elements are consecutively
628 // numbered from this value.
629 unsigned ShiftAmt = SVOp->getMaskElt(i);
630 if (ShiftAmt < i) return -1;
634 // Check the rest of the elements to see if they are consecutive.
635 for (++i; i != 16; ++i)
636 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
639 // Check the rest of the elements to see if they are consecutive.
640 for (++i; i != 16; ++i)
641 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
647 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
648 /// specifies a splat of a single element that is suitable for input to
649 /// VSPLTB/VSPLTH/VSPLTW.
650 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
651 assert(N->getValueType(0) == MVT::v16i8 &&
652 (EltSize == 1 || EltSize == 2 || EltSize == 4));
654 // This is a splat operation if each element of the permute is the same, and
655 // if the value doesn't reference the second vector.
656 unsigned ElementBase = N->getMaskElt(0);
658 // FIXME: Handle UNDEF elements too!
659 if (ElementBase >= 16)
662 // Check that the indices are consecutive, in the case of a multi-byte element
663 // splatted with a v16i8 mask.
664 for (unsigned i = 1; i != EltSize; ++i)
665 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
668 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
669 if (N->getMaskElt(i) < 0) continue;
670 for (unsigned j = 0; j != EltSize; ++j)
671 if (N->getMaskElt(i+j) != N->getMaskElt(j))
677 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
679 bool PPC::isAllNegativeZeroVector(SDNode *N) {
680 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
682 APInt APVal, APUndef;
686 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
687 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
688 return CFP->getValueAPF().isNegZero();
693 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
694 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
695 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
696 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
697 assert(isSplatShuffleMask(SVOp, EltSize));
698 return SVOp->getMaskElt(0) / EltSize;
701 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
702 /// by using a vspltis[bhw] instruction of the specified element size, return
703 /// the constant being splatted. The ByteSize field indicates the number of
704 /// bytes of each element [124] -> [bhw].
705 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
708 // If ByteSize of the splat is bigger than the element size of the
709 // build_vector, then we have a case where we are checking for a splat where
710 // multiple elements of the buildvector are folded together into a single
711 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
712 unsigned EltSize = 16/N->getNumOperands();
713 if (EltSize < ByteSize) {
714 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
715 SDValue UniquedVals[4];
716 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
718 // See if all of the elements in the buildvector agree across.
719 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
720 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
721 // If the element isn't a constant, bail fully out.
722 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
725 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
726 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
727 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
728 return SDValue(); // no match.
731 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
732 // either constant or undef values that are identical for each chunk. See
733 // if these chunks can form into a larger vspltis*.
735 // Check to see if all of the leading entries are either 0 or -1. If
736 // neither, then this won't fit into the immediate field.
737 bool LeadingZero = true;
738 bool LeadingOnes = true;
739 for (unsigned i = 0; i != Multiple-1; ++i) {
740 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
742 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
743 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
745 // Finally, check the least significant entry.
747 if (UniquedVals[Multiple-1].getNode() == 0)
748 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
749 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
751 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
754 if (UniquedVals[Multiple-1].getNode() == 0)
755 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
756 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
757 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
758 return DAG.getTargetConstant(Val, MVT::i32);
764 // Check to see if this buildvec has a single non-undef value in its elements.
765 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
766 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
767 if (OpVal.getNode() == 0)
768 OpVal = N->getOperand(i);
769 else if (OpVal != N->getOperand(i))
773 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
775 unsigned ValSizeInBytes = EltSize;
777 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
778 Value = CN->getZExtValue();
779 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
780 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
781 Value = FloatToBits(CN->getValueAPF().convertToFloat());
784 // If the splat value is larger than the element value, then we can never do
785 // this splat. The only case that we could fit the replicated bits into our
786 // immediate field for would be zero, and we prefer to use vxor for it.
787 if (ValSizeInBytes < ByteSize) return SDValue();
789 // If the element value is larger than the splat value, cut it in half and
790 // check to see if the two halves are equal. Continue doing this until we
791 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
792 while (ValSizeInBytes > ByteSize) {
793 ValSizeInBytes >>= 1;
795 // If the top half equals the bottom half, we're still ok.
796 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
797 (Value & ((1 << (8*ValSizeInBytes))-1)))
801 // Properly sign extend the value.
802 int ShAmt = (4-ByteSize)*8;
803 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
805 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
806 if (MaskVal == 0) return SDValue();
808 // Finally, if this value fits in a 5 bit sext field, return it
809 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
810 return DAG.getTargetConstant(MaskVal, MVT::i32);
814 //===----------------------------------------------------------------------===//
815 // Addressing Mode Selection
816 //===----------------------------------------------------------------------===//
818 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
819 /// or 64-bit immediate, and if the value can be accurately represented as a
820 /// sign extension from a 16-bit value. If so, this returns true and the
822 static bool isIntS16Immediate(SDNode *N, short &Imm) {
823 if (N->getOpcode() != ISD::Constant)
826 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
827 if (N->getValueType(0) == MVT::i32)
828 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
830 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
832 static bool isIntS16Immediate(SDValue Op, short &Imm) {
833 return isIntS16Immediate(Op.getNode(), Imm);
837 /// SelectAddressRegReg - Given the specified addressed, check to see if it
838 /// can be represented as an indexed [r+r] operation. Returns false if it
839 /// can be more efficiently represented with [r+imm].
840 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
842 SelectionDAG &DAG) const {
844 if (N.getOpcode() == ISD::ADD) {
845 if (isIntS16Immediate(N.getOperand(1), imm))
847 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
850 Base = N.getOperand(0);
851 Index = N.getOperand(1);
853 } else if (N.getOpcode() == ISD::OR) {
854 if (isIntS16Immediate(N.getOperand(1), imm))
855 return false; // r+i can fold it if we can.
857 // If this is an or of disjoint bitfields, we can codegen this as an add
858 // (for better address arithmetic) if the LHS and RHS of the OR are provably
860 APInt LHSKnownZero, LHSKnownOne;
861 APInt RHSKnownZero, RHSKnownOne;
862 DAG.ComputeMaskedBits(N.getOperand(0),
863 LHSKnownZero, LHSKnownOne);
865 if (LHSKnownZero.getBoolValue()) {
866 DAG.ComputeMaskedBits(N.getOperand(1),
867 RHSKnownZero, RHSKnownOne);
868 // If all of the bits are known zero on the LHS or RHS, the add won't
870 if (~(LHSKnownZero | RHSKnownZero) == 0) {
871 Base = N.getOperand(0);
872 Index = N.getOperand(1);
881 /// Returns true if the address N can be represented by a base register plus
882 /// a signed 16-bit displacement [r+imm], and if it is not better
883 /// represented as reg+reg.
884 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
886 SelectionDAG &DAG) const {
887 // FIXME dl should come from parent load or store, not from address
888 DebugLoc dl = N.getDebugLoc();
889 // If this can be more profitably realized as r+r, fail.
890 if (SelectAddressRegReg(N, Disp, Base, DAG))
893 if (N.getOpcode() == ISD::ADD) {
895 if (isIntS16Immediate(N.getOperand(1), imm)) {
896 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
897 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
898 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
900 Base = N.getOperand(0);
902 return true; // [r+i]
903 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
904 // Match LOAD (ADD (X, Lo(G))).
905 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
906 && "Cannot handle constant offsets yet!");
907 Disp = N.getOperand(1).getOperand(0); // The global address.
908 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
909 Disp.getOpcode() == ISD::TargetConstantPool ||
910 Disp.getOpcode() == ISD::TargetJumpTable);
911 Base = N.getOperand(0);
912 return true; // [&g+r]
914 } else if (N.getOpcode() == ISD::OR) {
916 if (isIntS16Immediate(N.getOperand(1), imm)) {
917 // If this is an or of disjoint bitfields, we can codegen this as an add
918 // (for better address arithmetic) if the LHS and RHS of the OR are
919 // provably disjoint.
920 APInt LHSKnownZero, LHSKnownOne;
921 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
923 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
924 // If all of the bits are known zero on the LHS or RHS, the add won't
926 Base = N.getOperand(0);
927 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
931 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
932 // Loading from a constant address.
934 // If this address fits entirely in a 16-bit sext immediate field, codegen
937 if (isIntS16Immediate(CN, Imm)) {
938 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
939 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
940 CN->getValueType(0));
944 // Handle 32-bit sext immediates with LIS + addr mode.
945 if (CN->getValueType(0) == MVT::i32 ||
946 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
947 int Addr = (int)CN->getZExtValue();
949 // Otherwise, break this down into an LIS + disp.
950 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
952 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
953 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
954 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
959 Disp = DAG.getTargetConstant(0, getPointerTy());
960 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
961 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
964 return true; // [r+0]
967 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
968 /// represented as an indexed [r+r] operation.
969 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
971 SelectionDAG &DAG) const {
972 // Check to see if we can easily represent this as an [r+r] address. This
973 // will fail if it thinks that the address is more profitably represented as
974 // reg+imm, e.g. where imm = 0.
975 if (SelectAddressRegReg(N, Base, Index, DAG))
978 // If the operand is an addition, always emit this as [r+r], since this is
979 // better (for code size, and execution, as the memop does the add for free)
980 // than emitting an explicit add.
981 if (N.getOpcode() == ISD::ADD) {
982 Base = N.getOperand(0);
983 Index = N.getOperand(1);
987 // Otherwise, do it the hard way, using R0 as the base register.
988 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
994 /// SelectAddressRegImmShift - Returns true if the address N can be
995 /// represented by a base register plus a signed 14-bit displacement
996 /// [r+imm*4]. Suitable for use by STD and friends.
997 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
999 SelectionDAG &DAG) const {
1000 // FIXME dl should come from the parent load or store, not the address
1001 DebugLoc dl = N.getDebugLoc();
1002 // If this can be more profitably realized as r+r, fail.
1003 if (SelectAddressRegReg(N, Disp, Base, DAG))
1006 if (N.getOpcode() == ISD::ADD) {
1008 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1009 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1010 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1011 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1013 Base = N.getOperand(0);
1015 return true; // [r+i]
1016 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1017 // Match LOAD (ADD (X, Lo(G))).
1018 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1019 && "Cannot handle constant offsets yet!");
1020 Disp = N.getOperand(1).getOperand(0); // The global address.
1021 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1022 Disp.getOpcode() == ISD::TargetConstantPool ||
1023 Disp.getOpcode() == ISD::TargetJumpTable);
1024 Base = N.getOperand(0);
1025 return true; // [&g+r]
1027 } else if (N.getOpcode() == ISD::OR) {
1029 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1030 // If this is an or of disjoint bitfields, we can codegen this as an add
1031 // (for better address arithmetic) if the LHS and RHS of the OR are
1032 // provably disjoint.
1033 APInt LHSKnownZero, LHSKnownOne;
1034 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1035 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1036 // If all of the bits are known zero on the LHS or RHS, the add won't
1038 Base = N.getOperand(0);
1039 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1043 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1044 // Loading from a constant address. Verify low two bits are clear.
1045 if ((CN->getZExtValue() & 3) == 0) {
1046 // If this address fits entirely in a 14-bit sext immediate field, codegen
1049 if (isIntS16Immediate(CN, Imm)) {
1050 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1051 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1052 CN->getValueType(0));
1056 // Fold the low-part of 32-bit absolute addresses into addr mode.
1057 if (CN->getValueType(0) == MVT::i32 ||
1058 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1059 int Addr = (int)CN->getZExtValue();
1061 // Otherwise, break this down into an LIS + disp.
1062 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1063 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1064 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1065 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1071 Disp = DAG.getTargetConstant(0, getPointerTy());
1072 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1073 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1076 return true; // [r+0]
1080 /// getPreIndexedAddressParts - returns true by value, base pointer and
1081 /// offset pointer and addressing mode by reference if the node's address
1082 /// can be legally represented as pre-indexed load / store address.
1083 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1085 ISD::MemIndexedMode &AM,
1086 SelectionDAG &DAG) const {
1087 // Disabled by default for now.
1088 if (!EnablePPCPreinc) return false;
1092 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1093 Ptr = LD->getBasePtr();
1094 VT = LD->getMemoryVT();
1096 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1097 Ptr = ST->getBasePtr();
1098 VT = ST->getMemoryVT();
1102 // PowerPC doesn't have preinc load/store instructions for vectors.
1106 // TODO: Check reg+reg first.
1108 // LDU/STU use reg+imm*4, others use reg+imm.
1109 if (VT != MVT::i64) {
1111 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1115 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1119 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1120 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1121 // sext i32 to i64 when addr mode is r+i.
1122 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1123 LD->getExtensionType() == ISD::SEXTLOAD &&
1124 isa<ConstantSDNode>(Offset))
1132 //===----------------------------------------------------------------------===//
1133 // LowerOperation implementation
1134 //===----------------------------------------------------------------------===//
1136 /// GetLabelAccessInfo - Return true if we should reference labels using a
1137 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1138 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1139 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1140 HiOpFlags = PPCII::MO_HA16;
1141 LoOpFlags = PPCII::MO_LO16;
1143 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1144 // non-darwin platform. We don't support PIC on other platforms yet.
1145 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1146 TM.getSubtarget<PPCSubtarget>().isDarwin();
1148 HiOpFlags |= PPCII::MO_PIC_FLAG;
1149 LoOpFlags |= PPCII::MO_PIC_FLAG;
1152 // If this is a reference to a global value that requires a non-lazy-ptr, make
1153 // sure that instruction lowering adds it.
1154 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1155 HiOpFlags |= PPCII::MO_NLP_FLAG;
1156 LoOpFlags |= PPCII::MO_NLP_FLAG;
1158 if (GV->hasHiddenVisibility()) {
1159 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1160 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1167 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1168 SelectionDAG &DAG) {
1169 EVT PtrVT = HiPart.getValueType();
1170 SDValue Zero = DAG.getConstant(0, PtrVT);
1171 DebugLoc DL = HiPart.getDebugLoc();
1173 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1174 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1176 // With PIC, the first instruction is actually "GR+hi(&G)".
1178 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1179 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1181 // Generate non-pic code that has direct accesses to the constant pool.
1182 // The address of the global is just (hi(&g)+lo(&g)).
1183 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1186 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1187 SelectionDAG &DAG) const {
1188 EVT PtrVT = Op.getValueType();
1189 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1190 const Constant *C = CP->getConstVal();
1192 unsigned MOHiFlag, MOLoFlag;
1193 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1195 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1197 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1198 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1201 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1202 EVT PtrVT = Op.getValueType();
1203 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1205 unsigned MOHiFlag, MOLoFlag;
1206 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1207 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1208 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1209 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1212 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1213 SelectionDAG &DAG) const {
1214 EVT PtrVT = Op.getValueType();
1216 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1218 unsigned MOHiFlag, MOLoFlag;
1219 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1220 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1221 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1222 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1225 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1226 SelectionDAG &DAG) const {
1227 EVT PtrVT = Op.getValueType();
1228 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1229 DebugLoc DL = GSDN->getDebugLoc();
1230 const GlobalValue *GV = GSDN->getGlobal();
1232 // 64-bit SVR4 ABI code is always position-independent.
1233 // The actual address of the GlobalValue is stored in the TOC.
1234 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1235 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1236 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1237 DAG.getRegister(PPC::X2, MVT::i64));
1240 unsigned MOHiFlag, MOLoFlag;
1241 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1244 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1246 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1248 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1250 // If the global reference is actually to a non-lazy-pointer, we have to do an
1251 // extra load to get the address of the global.
1252 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1253 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1254 false, false, false, 0);
1258 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1259 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1260 DebugLoc dl = Op.getDebugLoc();
1262 // If we're comparing for equality to zero, expose the fact that this is
1263 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1264 // fold the new nodes.
1265 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1266 if (C->isNullValue() && CC == ISD::SETEQ) {
1267 EVT VT = Op.getOperand(0).getValueType();
1268 SDValue Zext = Op.getOperand(0);
1269 if (VT.bitsLT(MVT::i32)) {
1271 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1273 unsigned Log2b = Log2_32(VT.getSizeInBits());
1274 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1275 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1276 DAG.getConstant(Log2b, MVT::i32));
1277 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1279 // Leave comparisons against 0 and -1 alone for now, since they're usually
1280 // optimized. FIXME: revisit this when we can custom lower all setcc
1282 if (C->isAllOnesValue() || C->isNullValue())
1286 // If we have an integer seteq/setne, turn it into a compare against zero
1287 // by xor'ing the rhs with the lhs, which is faster than setting a
1288 // condition register, reading it back out, and masking the correct bit. The
1289 // normal approach here uses sub to do this instead of xor. Using xor exposes
1290 // the result to other bit-twiddling opportunities.
1291 EVT LHSVT = Op.getOperand(0).getValueType();
1292 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1293 EVT VT = Op.getValueType();
1294 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1296 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1301 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1302 const PPCSubtarget &Subtarget) const {
1303 SDNode *Node = Op.getNode();
1304 EVT VT = Node->getValueType(0);
1305 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1306 SDValue InChain = Node->getOperand(0);
1307 SDValue VAListPtr = Node->getOperand(1);
1308 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1309 DebugLoc dl = Node->getDebugLoc();
1311 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1314 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1315 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1317 InChain = GprIndex.getValue(1);
1319 if (VT == MVT::i64) {
1320 // Check if GprIndex is even
1321 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1322 DAG.getConstant(1, MVT::i32));
1323 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1324 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1325 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1326 DAG.getConstant(1, MVT::i32));
1327 // Align GprIndex to be even if it isn't
1328 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1332 // fpr index is 1 byte after gpr
1333 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1334 DAG.getConstant(1, MVT::i32));
1337 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1338 FprPtr, MachinePointerInfo(SV), MVT::i8,
1340 InChain = FprIndex.getValue(1);
1342 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1343 DAG.getConstant(8, MVT::i32));
1345 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1346 DAG.getConstant(4, MVT::i32));
1349 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1350 MachinePointerInfo(), false, false,
1352 InChain = OverflowArea.getValue(1);
1354 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1355 MachinePointerInfo(), false, false,
1357 InChain = RegSaveArea.getValue(1);
1359 // select overflow_area if index > 8
1360 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1361 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1363 // adjustment constant gpr_index * 4/8
1364 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1365 VT.isInteger() ? GprIndex : FprIndex,
1366 DAG.getConstant(VT.isInteger() ? 4 : 8,
1369 // OurReg = RegSaveArea + RegConstant
1370 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1373 // Floating types are 32 bytes into RegSaveArea
1374 if (VT.isFloatingPoint())
1375 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1376 DAG.getConstant(32, MVT::i32));
1378 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1379 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1380 VT.isInteger() ? GprIndex : FprIndex,
1381 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1384 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1385 VT.isInteger() ? VAListPtr : FprPtr,
1386 MachinePointerInfo(SV),
1387 MVT::i8, false, false, 0);
1389 // determine if we should load from reg_save_area or overflow_area
1390 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1392 // increase overflow_area by 4/8 if gpr/fpr > 8
1393 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1394 DAG.getConstant(VT.isInteger() ? 4 : 8,
1397 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1400 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1402 MachinePointerInfo(),
1403 MVT::i32, false, false, 0);
1405 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1406 false, false, false, 0);
1409 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1410 SelectionDAG &DAG) const {
1411 return Op.getOperand(0);
1414 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1415 SelectionDAG &DAG) const {
1416 SDValue Chain = Op.getOperand(0);
1417 SDValue Trmp = Op.getOperand(1); // trampoline
1418 SDValue FPtr = Op.getOperand(2); // nested function
1419 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1420 DebugLoc dl = Op.getDebugLoc();
1422 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1423 bool isPPC64 = (PtrVT == MVT::i64);
1425 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1428 TargetLowering::ArgListTy Args;
1429 TargetLowering::ArgListEntry Entry;
1431 Entry.Ty = IntPtrTy;
1432 Entry.Node = Trmp; Args.push_back(Entry);
1434 // TrampSize == (isPPC64 ? 48 : 40);
1435 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1436 isPPC64 ? MVT::i64 : MVT::i32);
1437 Args.push_back(Entry);
1439 Entry.Node = FPtr; Args.push_back(Entry);
1440 Entry.Node = Nest; Args.push_back(Entry);
1442 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1443 TargetLowering::CallLoweringInfo CLI(Chain,
1444 Type::getVoidTy(*DAG.getContext()),
1445 false, false, false, false, 0,
1447 /*isTailCall=*/false,
1448 /*doesNotRet=*/false,
1449 /*isReturnValueUsed=*/true,
1450 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1452 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1454 return CallResult.second;
1457 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1458 const PPCSubtarget &Subtarget) const {
1459 MachineFunction &MF = DAG.getMachineFunction();
1460 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1462 DebugLoc dl = Op.getDebugLoc();
1464 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1465 // vastart just stores the address of the VarArgsFrameIndex slot into the
1466 // memory location argument.
1467 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1468 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1469 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1470 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1471 MachinePointerInfo(SV),
1475 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1476 // We suppose the given va_list is already allocated.
1479 // char gpr; /* index into the array of 8 GPRs
1480 // * stored in the register save area
1481 // * gpr=0 corresponds to r3,
1482 // * gpr=1 to r4, etc.
1484 // char fpr; /* index into the array of 8 FPRs
1485 // * stored in the register save area
1486 // * fpr=0 corresponds to f1,
1487 // * fpr=1 to f2, etc.
1489 // char *overflow_arg_area;
1490 // /* location on stack that holds
1491 // * the next overflow argument
1493 // char *reg_save_area;
1494 // /* where r3:r10 and f1:f8 (if saved)
1500 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1501 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1504 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1506 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1508 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1511 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1512 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1514 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1515 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1517 uint64_t FPROffset = 1;
1518 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1520 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1522 // Store first byte : number of int regs
1523 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1525 MachinePointerInfo(SV),
1526 MVT::i8, false, false, 0);
1527 uint64_t nextOffset = FPROffset;
1528 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1531 // Store second byte : number of float regs
1532 SDValue secondStore =
1533 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1534 MachinePointerInfo(SV, nextOffset), MVT::i8,
1536 nextOffset += StackOffset;
1537 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1539 // Store second word : arguments given on stack
1540 SDValue thirdStore =
1541 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1542 MachinePointerInfo(SV, nextOffset),
1544 nextOffset += FrameOffset;
1545 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1547 // Store third word : arguments given in registers
1548 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1549 MachinePointerInfo(SV, nextOffset),
1554 #include "PPCGenCallingConv.inc"
1556 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1557 CCValAssign::LocInfo &LocInfo,
1558 ISD::ArgFlagsTy &ArgFlags,
1563 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1565 CCValAssign::LocInfo &LocInfo,
1566 ISD::ArgFlagsTy &ArgFlags,
1568 static const uint16_t ArgRegs[] = {
1569 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1570 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1572 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1574 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1576 // Skip one register if the first unallocated register has an even register
1577 // number and there are still argument registers available which have not been
1578 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1579 // need to skip a register if RegNum is odd.
1580 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1581 State.AllocateReg(ArgRegs[RegNum]);
1584 // Always return false here, as this function only makes sure that the first
1585 // unallocated register has an odd register number and does not actually
1586 // allocate a register for the current argument.
1590 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1592 CCValAssign::LocInfo &LocInfo,
1593 ISD::ArgFlagsTy &ArgFlags,
1595 static const uint16_t ArgRegs[] = {
1596 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1600 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1602 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1604 // If there is only one Floating-point register left we need to put both f64
1605 // values of a split ppc_fp128 value on the stack.
1606 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1607 State.AllocateReg(ArgRegs[RegNum]);
1610 // Always return false here, as this function only makes sure that the two f64
1611 // values a ppc_fp128 value is split into are both passed in registers or both
1612 // passed on the stack and does not actually allocate a register for the
1613 // current argument.
1617 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1619 static const uint16_t *GetFPR() {
1620 static const uint16_t FPR[] = {
1621 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1622 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1628 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1630 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1631 unsigned PtrByteSize) {
1632 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1633 if (Flags.isByVal())
1634 ArgSize = Flags.getByValSize();
1635 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1641 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1642 CallingConv::ID CallConv, bool isVarArg,
1643 const SmallVectorImpl<ISD::InputArg>
1645 DebugLoc dl, SelectionDAG &DAG,
1646 SmallVectorImpl<SDValue> &InVals)
1648 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1649 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1652 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1658 PPCTargetLowering::LowerFormalArguments_SVR4(
1660 CallingConv::ID CallConv, bool isVarArg,
1661 const SmallVectorImpl<ISD::InputArg>
1663 DebugLoc dl, SelectionDAG &DAG,
1664 SmallVectorImpl<SDValue> &InVals) const {
1666 // 32-bit SVR4 ABI Stack Frame Layout:
1667 // +-----------------------------------+
1668 // +--> | Back chain |
1669 // | +-----------------------------------+
1670 // | | Floating-point register save area |
1671 // | +-----------------------------------+
1672 // | | General register save area |
1673 // | +-----------------------------------+
1674 // | | CR save word |
1675 // | +-----------------------------------+
1676 // | | VRSAVE save word |
1677 // | +-----------------------------------+
1678 // | | Alignment padding |
1679 // | +-----------------------------------+
1680 // | | Vector register save area |
1681 // | +-----------------------------------+
1682 // | | Local variable space |
1683 // | +-----------------------------------+
1684 // | | Parameter list area |
1685 // | +-----------------------------------+
1686 // | | LR save word |
1687 // | +-----------------------------------+
1688 // SP--> +--- | Back chain |
1689 // +-----------------------------------+
1692 // System V Application Binary Interface PowerPC Processor Supplement
1693 // AltiVec Technology Programming Interface Manual
1695 MachineFunction &MF = DAG.getMachineFunction();
1696 MachineFrameInfo *MFI = MF.getFrameInfo();
1697 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1699 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1700 // Potential tail calls could cause overwriting of argument stack slots.
1701 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1702 (CallConv == CallingConv::Fast));
1703 unsigned PtrByteSize = 4;
1705 // Assign locations to all of the incoming arguments.
1706 SmallVector<CCValAssign, 16> ArgLocs;
1707 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1708 getTargetMachine(), ArgLocs, *DAG.getContext());
1710 // Reserve space for the linkage area on the stack.
1711 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1713 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1715 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1716 CCValAssign &VA = ArgLocs[i];
1718 // Arguments stored in registers.
1719 if (VA.isRegLoc()) {
1720 const TargetRegisterClass *RC;
1721 EVT ValVT = VA.getValVT();
1723 switch (ValVT.getSimpleVT().SimpleTy) {
1725 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1727 RC = &PPC::GPRCRegClass;
1730 RC = &PPC::F4RCRegClass;
1733 RC = &PPC::F8RCRegClass;
1739 RC = &PPC::VRRCRegClass;
1743 // Transform the arguments stored in physical registers into virtual ones.
1744 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1745 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1747 InVals.push_back(ArgValue);
1749 // Argument stored in memory.
1750 assert(VA.isMemLoc());
1752 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1753 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1756 // Create load nodes to retrieve arguments from the stack.
1757 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1758 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1759 MachinePointerInfo(),
1760 false, false, false, 0));
1764 // Assign locations to all of the incoming aggregate by value arguments.
1765 // Aggregates passed by value are stored in the local variable space of the
1766 // caller's stack frame, right above the parameter list area.
1767 SmallVector<CCValAssign, 16> ByValArgLocs;
1768 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1769 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1771 // Reserve stack space for the allocations in CCInfo.
1772 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1774 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1776 // Area that is at least reserved in the caller of this function.
1777 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1779 // Set the size that is at least reserved in caller of this function. Tail
1780 // call optimized function's reserved stack space needs to be aligned so that
1781 // taking the difference between two stack areas will result in an aligned
1783 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1786 std::max(MinReservedArea,
1787 PPCFrameLowering::getMinCallFrameSize(false, false));
1789 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1790 getStackAlignment();
1791 unsigned AlignMask = TargetAlign-1;
1792 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1794 FI->setMinReservedArea(MinReservedArea);
1796 SmallVector<SDValue, 8> MemOps;
1798 // If the function takes variable number of arguments, make a frame index for
1799 // the start of the first vararg value... for expansion of llvm.va_start.
1801 static const uint16_t GPArgRegs[] = {
1802 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1803 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1805 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1807 static const uint16_t FPArgRegs[] = {
1808 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1811 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1813 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1815 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1818 // Make room for NumGPArgRegs and NumFPArgRegs.
1819 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1820 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1822 FuncInfo->setVarArgsStackOffset(
1823 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1824 CCInfo.getNextStackOffset(), true));
1826 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1827 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1829 // The fixed integer arguments of a variadic function are stored to the
1830 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1831 // the result of va_next.
1832 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1833 // Get an existing live-in vreg, or add a new one.
1834 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1836 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1838 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1839 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1840 MachinePointerInfo(), false, false, 0);
1841 MemOps.push_back(Store);
1842 // Increment the address by four for the next argument to store
1843 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1844 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1847 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1849 // The double arguments are stored to the VarArgsFrameIndex
1851 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1852 // Get an existing live-in vreg, or add a new one.
1853 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1855 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1857 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1858 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1859 MachinePointerInfo(), false, false, 0);
1860 MemOps.push_back(Store);
1861 // Increment the address by eight for the next argument to store
1862 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1864 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1868 if (!MemOps.empty())
1869 Chain = DAG.getNode(ISD::TokenFactor, dl,
1870 MVT::Other, &MemOps[0], MemOps.size());
1876 PPCTargetLowering::LowerFormalArguments_Darwin(
1878 CallingConv::ID CallConv, bool isVarArg,
1879 const SmallVectorImpl<ISD::InputArg>
1881 DebugLoc dl, SelectionDAG &DAG,
1882 SmallVectorImpl<SDValue> &InVals) const {
1883 // TODO: add description of PPC stack frame format, or at least some docs.
1885 MachineFunction &MF = DAG.getMachineFunction();
1886 MachineFrameInfo *MFI = MF.getFrameInfo();
1887 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1889 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1890 bool isPPC64 = PtrVT == MVT::i64;
1891 // Potential tail calls could cause overwriting of argument stack slots.
1892 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1893 (CallConv == CallingConv::Fast));
1894 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1896 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1897 // Area that is at least reserved in caller of this function.
1898 unsigned MinReservedArea = ArgOffset;
1900 static const uint16_t GPR_32[] = { // 32-bit registers.
1901 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1902 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1904 static const uint16_t GPR_64[] = { // 64-bit registers.
1905 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1906 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1909 static const uint16_t *FPR = GetFPR();
1911 static const uint16_t VR[] = {
1912 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1913 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1916 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1917 const unsigned Num_FPR_Regs = 13;
1918 const unsigned Num_VR_Regs = array_lengthof( VR);
1920 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1922 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
1924 // In 32-bit non-varargs functions, the stack space for vectors is after the
1925 // stack space for non-vectors. We do not use this space unless we have
1926 // too many vectors to fit in registers, something that only occurs in
1927 // constructed examples:), but we have to walk the arglist to figure
1928 // that out...for the pathological case, compute VecArgOffset as the
1929 // start of the vector parameter area. Computing VecArgOffset is the
1930 // entire point of the following loop.
1931 unsigned VecArgOffset = ArgOffset;
1932 if (!isVarArg && !isPPC64) {
1933 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1935 EVT ObjectVT = Ins[ArgNo].VT;
1936 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1938 if (Flags.isByVal()) {
1939 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1940 unsigned ObjSize = Flags.getByValSize();
1942 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1943 VecArgOffset += ArgSize;
1947 switch(ObjectVT.getSimpleVT().SimpleTy) {
1948 default: llvm_unreachable("Unhandled argument type!");
1951 VecArgOffset += isPPC64 ? 8 : 4;
1953 case MVT::i64: // PPC64
1961 // Nothing to do, we're only looking at Nonvector args here.
1966 // We've found where the vector parameter area in memory is. Skip the
1967 // first 12 parameters; these don't use that memory.
1968 VecArgOffset = ((VecArgOffset+15)/16)*16;
1969 VecArgOffset += 12*16;
1971 // Add DAG nodes to load the arguments or copy them out of registers. On
1972 // entry to a function on PPC, the arguments start after the linkage area,
1973 // although the first ones are often in registers.
1975 SmallVector<SDValue, 8> MemOps;
1976 unsigned nAltivecParamsAtEnd = 0;
1977 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1979 bool needsLoad = false;
1980 EVT ObjectVT = Ins[ArgNo].VT;
1981 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1982 unsigned ArgSize = ObjSize;
1983 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1985 unsigned CurArgOffset = ArgOffset;
1987 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1988 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1989 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1990 if (isVarArg || isPPC64) {
1991 MinReservedArea = ((MinReservedArea+15)/16)*16;
1992 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1995 } else nAltivecParamsAtEnd++;
1997 // Calculate min reserved area.
1998 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2002 // FIXME the codegen can be much improved in some cases.
2003 // We do not have to keep everything in memory.
2004 if (Flags.isByVal()) {
2005 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2006 ObjSize = Flags.getByValSize();
2007 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2008 // Objects of size 1 and 2 are right justified, everything else is
2009 // left justified. This means the memory address is adjusted forwards.
2010 if (ObjSize==1 || ObjSize==2) {
2011 CurArgOffset = CurArgOffset + (4 - ObjSize);
2013 // The value of the object is its address.
2014 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2015 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2016 InVals.push_back(FIN);
2017 if (ObjSize==1 || ObjSize==2) {
2018 if (GPR_idx != Num_GPR_Regs) {
2021 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2023 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2024 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2025 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2026 MachinePointerInfo(),
2027 ObjSize==1 ? MVT::i8 : MVT::i16,
2029 MemOps.push_back(Store);
2033 ArgOffset += PtrByteSize;
2037 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2038 // Store whatever pieces of the object are in registers
2039 // to memory. ArgVal will be address of the beginning of
2041 if (GPR_idx != Num_GPR_Regs) {
2044 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2046 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2047 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2048 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2049 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2050 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2051 MachinePointerInfo(),
2053 MemOps.push_back(Store);
2055 ArgOffset += PtrByteSize;
2057 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2064 switch (ObjectVT.getSimpleVT().SimpleTy) {
2065 default: llvm_unreachable("Unhandled argument type!");
2068 if (GPR_idx != Num_GPR_Regs) {
2069 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2070 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2074 ArgSize = PtrByteSize;
2076 // All int arguments reserve stack space in the Darwin ABI.
2077 ArgOffset += PtrByteSize;
2081 case MVT::i64: // PPC64
2082 if (GPR_idx != Num_GPR_Regs) {
2083 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2084 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2086 if (ObjectVT == MVT::i32) {
2087 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2088 // value to MVT::i64 and then truncate to the correct register size.
2090 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2091 DAG.getValueType(ObjectVT));
2092 else if (Flags.isZExt())
2093 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2094 DAG.getValueType(ObjectVT));
2096 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2102 ArgSize = PtrByteSize;
2104 // All int arguments reserve stack space in the Darwin ABI.
2110 // Every 4 bytes of argument space consumes one of the GPRs available for
2111 // argument passing.
2112 if (GPR_idx != Num_GPR_Regs) {
2114 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2117 if (FPR_idx != Num_FPR_Regs) {
2120 if (ObjectVT == MVT::f32)
2121 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2123 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2125 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2131 // All FP arguments reserve stack space in the Darwin ABI.
2132 ArgOffset += isPPC64 ? 8 : ObjSize;
2138 // Note that vector arguments in registers don't reserve stack space,
2139 // except in varargs functions.
2140 if (VR_idx != Num_VR_Regs) {
2141 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2142 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2144 while ((ArgOffset % 16) != 0) {
2145 ArgOffset += PtrByteSize;
2146 if (GPR_idx != Num_GPR_Regs)
2150 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2154 if (!isVarArg && !isPPC64) {
2155 // Vectors go after all the nonvectors.
2156 CurArgOffset = VecArgOffset;
2159 // Vectors are aligned.
2160 ArgOffset = ((ArgOffset+15)/16)*16;
2161 CurArgOffset = ArgOffset;
2169 // We need to load the argument to a virtual register if we determined above
2170 // that we ran out of physical registers of the appropriate type.
2172 int FI = MFI->CreateFixedObject(ObjSize,
2173 CurArgOffset + (ArgSize - ObjSize),
2175 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2176 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2177 false, false, false, 0);
2180 InVals.push_back(ArgVal);
2183 // Set the size that is at least reserved in caller of this function. Tail
2184 // call optimized function's reserved stack space needs to be aligned so that
2185 // taking the difference between two stack areas will result in an aligned
2187 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2188 // Add the Altivec parameters at the end, if needed.
2189 if (nAltivecParamsAtEnd) {
2190 MinReservedArea = ((MinReservedArea+15)/16)*16;
2191 MinReservedArea += 16*nAltivecParamsAtEnd;
2194 std::max(MinReservedArea,
2195 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2196 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2197 getStackAlignment();
2198 unsigned AlignMask = TargetAlign-1;
2199 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2200 FI->setMinReservedArea(MinReservedArea);
2202 // If the function takes variable number of arguments, make a frame index for
2203 // the start of the first vararg value... for expansion of llvm.va_start.
2205 int Depth = ArgOffset;
2207 FuncInfo->setVarArgsFrameIndex(
2208 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2210 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2212 // If this function is vararg, store any remaining integer argument regs
2213 // to their spots on the stack so that they may be loaded by deferencing the
2214 // result of va_next.
2215 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2219 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2221 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2223 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2224 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2225 MachinePointerInfo(), false, false, 0);
2226 MemOps.push_back(Store);
2227 // Increment the address by four for the next argument to store
2228 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2229 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2233 if (!MemOps.empty())
2234 Chain = DAG.getNode(ISD::TokenFactor, dl,
2235 MVT::Other, &MemOps[0], MemOps.size());
2240 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2241 /// linkage area for the Darwin ABI.
2243 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2247 const SmallVectorImpl<ISD::OutputArg>
2249 const SmallVectorImpl<SDValue> &OutVals,
2250 unsigned &nAltivecParamsAtEnd) {
2251 // Count how many bytes are to be pushed on the stack, including the linkage
2252 // area, and parameter passing area. We start with 24/48 bytes, which is
2253 // prereserved space for [SP][CR][LR][3 x unused].
2254 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2255 unsigned NumOps = Outs.size();
2256 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2258 // Add up all the space actually used.
2259 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2260 // they all go in registers, but we must reserve stack space for them for
2261 // possible use by the caller. In varargs or 64-bit calls, parameters are
2262 // assigned stack space in order, with padding so Altivec parameters are
2264 nAltivecParamsAtEnd = 0;
2265 for (unsigned i = 0; i != NumOps; ++i) {
2266 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2267 EVT ArgVT = Outs[i].VT;
2268 // Varargs Altivec parameters are padded to a 16 byte boundary.
2269 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2270 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2271 if (!isVarArg && !isPPC64) {
2272 // Non-varargs Altivec parameters go after all the non-Altivec
2273 // parameters; handle those later so we know how much padding we need.
2274 nAltivecParamsAtEnd++;
2277 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2278 NumBytes = ((NumBytes+15)/16)*16;
2280 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2283 // Allow for Altivec parameters at the end, if needed.
2284 if (nAltivecParamsAtEnd) {
2285 NumBytes = ((NumBytes+15)/16)*16;
2286 NumBytes += 16*nAltivecParamsAtEnd;
2289 // The prolog code of the callee may store up to 8 GPR argument registers to
2290 // the stack, allowing va_start to index over them in memory if its varargs.
2291 // Because we cannot tell if this is needed on the caller side, we have to
2292 // conservatively assume that it is needed. As such, make sure we have at
2293 // least enough stack space for the caller to store the 8 GPRs.
2294 NumBytes = std::max(NumBytes,
2295 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2297 // Tail call needs the stack to be aligned.
2298 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2299 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2300 getFrameLowering()->getStackAlignment();
2301 unsigned AlignMask = TargetAlign-1;
2302 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2308 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2309 /// adjusted to accommodate the arguments for the tailcall.
2310 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2311 unsigned ParamSize) {
2313 if (!isTailCall) return 0;
2315 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2316 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2317 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2318 // Remember only if the new adjustement is bigger.
2319 if (SPDiff < FI->getTailCallSPDelta())
2320 FI->setTailCallSPDelta(SPDiff);
2325 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2326 /// for tail call optimization. Targets which want to do tail call
2327 /// optimization should implement this function.
2329 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2330 CallingConv::ID CalleeCC,
2332 const SmallVectorImpl<ISD::InputArg> &Ins,
2333 SelectionDAG& DAG) const {
2334 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2337 // Variable argument functions are not supported.
2341 MachineFunction &MF = DAG.getMachineFunction();
2342 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2343 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2344 // Functions containing by val parameters are not supported.
2345 for (unsigned i = 0; i != Ins.size(); i++) {
2346 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2347 if (Flags.isByVal()) return false;
2350 // Non PIC/GOT tail calls are supported.
2351 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2354 // At the moment we can only do local tail calls (in same module, hidden
2355 // or protected) if we are generating PIC.
2356 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2357 return G->getGlobal()->hasHiddenVisibility()
2358 || G->getGlobal()->hasProtectedVisibility();
2364 /// isCallCompatibleAddress - Return the immediate to use if the specified
2365 /// 32-bit value is representable in the immediate field of a BxA instruction.
2366 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2367 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2370 int Addr = C->getZExtValue();
2371 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2372 (Addr << 6 >> 6) != Addr)
2373 return 0; // Top 6 bits have to be sext of immediate.
2375 return DAG.getConstant((int)C->getZExtValue() >> 2,
2376 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2381 struct TailCallArgumentInfo {
2386 TailCallArgumentInfo() : FrameIdx(0) {}
2391 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2393 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2395 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2396 SmallVector<SDValue, 8> &MemOpChains,
2398 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2399 SDValue Arg = TailCallArgs[i].Arg;
2400 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2401 int FI = TailCallArgs[i].FrameIdx;
2402 // Store relative to framepointer.
2403 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2404 MachinePointerInfo::getFixedStack(FI),
2409 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2410 /// the appropriate stack slot for the tail call optimized function call.
2411 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2412 MachineFunction &MF,
2421 // Calculate the new stack slot for the return address.
2422 int SlotSize = isPPC64 ? 8 : 4;
2423 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2425 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2426 NewRetAddrLoc, true);
2427 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2428 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2429 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2430 MachinePointerInfo::getFixedStack(NewRetAddr),
2433 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2434 // slot as the FP is never overwritten.
2437 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2438 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2440 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2441 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2442 MachinePointerInfo::getFixedStack(NewFPIdx),
2449 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2450 /// the position of the argument.
2452 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2453 SDValue Arg, int SPDiff, unsigned ArgOffset,
2454 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2455 int Offset = ArgOffset + SPDiff;
2456 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2457 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2458 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2459 SDValue FIN = DAG.getFrameIndex(FI, VT);
2460 TailCallArgumentInfo Info;
2462 Info.FrameIdxOp = FIN;
2464 TailCallArguments.push_back(Info);
2467 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2468 /// stack slot. Returns the chain as result and the loaded frame pointers in
2469 /// LROpOut/FPOpout. Used when tail calling.
2470 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2476 DebugLoc dl) const {
2478 // Load the LR and FP stack slot for later adjusting.
2479 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2480 LROpOut = getReturnAddrFrameIndex(DAG);
2481 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2482 false, false, false, 0);
2483 Chain = SDValue(LROpOut.getNode(), 1);
2485 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2486 // slot as the FP is never overwritten.
2488 FPOpOut = getFramePointerFrameIndex(DAG);
2489 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2490 false, false, false, 0);
2491 Chain = SDValue(FPOpOut.getNode(), 1);
2497 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2498 /// by "Src" to address "Dst" of size "Size". Alignment information is
2499 /// specified by the specific parameter attribute. The copy will be passed as
2500 /// a byval function parameter.
2501 /// Sometimes what we are copying is the end of a larger object, the part that
2502 /// does not fit in registers.
2504 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2505 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2507 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2508 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2509 false, false, MachinePointerInfo(0),
2510 MachinePointerInfo(0));
2513 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2516 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2517 SDValue Arg, SDValue PtrOff, int SPDiff,
2518 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2519 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2520 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2522 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2527 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2529 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2530 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2531 DAG.getConstant(ArgOffset, PtrVT));
2533 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2534 MachinePointerInfo(), false, false, 0));
2535 // Calculate and remember argument location.
2536 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2541 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2542 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2543 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2544 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2545 MachineFunction &MF = DAG.getMachineFunction();
2547 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2548 // might overwrite each other in case of tail call optimization.
2549 SmallVector<SDValue, 8> MemOpChains2;
2550 // Do not flag preceding copytoreg stuff together with the following stuff.
2552 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2554 if (!MemOpChains2.empty())
2555 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2556 &MemOpChains2[0], MemOpChains2.size());
2558 // Store the return address to the appropriate stack slot.
2559 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2560 isPPC64, isDarwinABI, dl);
2562 // Emit callseq_end just before tailcall node.
2563 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2564 DAG.getIntPtrConstant(0, true), InFlag);
2565 InFlag = Chain.getValue(1);
2569 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2570 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2571 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2572 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2573 const PPCSubtarget &PPCSubTarget) {
2575 bool isPPC64 = PPCSubTarget.isPPC64();
2576 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2578 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2579 NodeTys.push_back(MVT::Other); // Returns a chain
2580 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2582 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2584 bool needIndirectCall = true;
2585 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2586 // If this is an absolute destination address, use the munged value.
2587 Callee = SDValue(Dest, 0);
2588 needIndirectCall = false;
2591 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2592 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2593 // Use indirect calls for ALL functions calls in JIT mode, since the
2594 // far-call stubs may be outside relocation limits for a BL instruction.
2595 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2596 unsigned OpFlags = 0;
2597 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2598 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2599 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2600 (G->getGlobal()->isDeclaration() ||
2601 G->getGlobal()->isWeakForLinker())) {
2602 // PC-relative references to external symbols should go through $stub,
2603 // unless we're building with the leopard linker or later, which
2604 // automatically synthesizes these stubs.
2605 OpFlags = PPCII::MO_DARWIN_STUB;
2608 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2609 // every direct call is) turn it into a TargetGlobalAddress /
2610 // TargetExternalSymbol node so that legalize doesn't hack it.
2611 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2612 Callee.getValueType(),
2614 needIndirectCall = false;
2618 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2619 unsigned char OpFlags = 0;
2621 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2622 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2623 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2624 // PC-relative references to external symbols should go through $stub,
2625 // unless we're building with the leopard linker or later, which
2626 // automatically synthesizes these stubs.
2627 OpFlags = PPCII::MO_DARWIN_STUB;
2630 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2632 needIndirectCall = false;
2635 if (needIndirectCall) {
2636 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2637 // to do the call, we can't use PPCISD::CALL.
2638 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2640 if (isSVR4ABI && isPPC64) {
2641 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2642 // entry point, but to the function descriptor (the function entry point
2643 // address is part of the function descriptor though).
2644 // The function descriptor is a three doubleword structure with the
2645 // following fields: function entry point, TOC base address and
2646 // environment pointer.
2647 // Thus for a call through a function pointer, the following actions need
2649 // 1. Save the TOC of the caller in the TOC save area of its stack
2650 // frame (this is done in LowerCall_Darwin()).
2651 // 2. Load the address of the function entry point from the function
2653 // 3. Load the TOC of the callee from the function descriptor into r2.
2654 // 4. Load the environment pointer from the function descriptor into
2656 // 5. Branch to the function entry point address.
2657 // 6. On return of the callee, the TOC of the caller needs to be
2658 // restored (this is done in FinishCall()).
2660 // All those operations are flagged together to ensure that no other
2661 // operations can be scheduled in between. E.g. without flagging the
2662 // operations together, a TOC access in the caller could be scheduled
2663 // between the load of the callee TOC and the branch to the callee, which
2664 // results in the TOC access going through the TOC of the callee instead
2665 // of going through the TOC of the caller, which leads to incorrect code.
2667 // Load the address of the function entry point from the function
2669 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2670 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2671 InFlag.getNode() ? 3 : 2);
2672 Chain = LoadFuncPtr.getValue(1);
2673 InFlag = LoadFuncPtr.getValue(2);
2675 // Load environment pointer into r11.
2676 // Offset of the environment pointer within the function descriptor.
2677 SDValue PtrOff = DAG.getIntPtrConstant(16);
2679 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2680 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2682 Chain = LoadEnvPtr.getValue(1);
2683 InFlag = LoadEnvPtr.getValue(2);
2685 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2687 Chain = EnvVal.getValue(0);
2688 InFlag = EnvVal.getValue(1);
2690 // Load TOC of the callee into r2. We are using a target-specific load
2691 // with r2 hard coded, because the result of a target-independent load
2692 // would never go directly into r2, since r2 is a reserved register (which
2693 // prevents the register allocator from allocating it), resulting in an
2694 // additional register being allocated and an unnecessary move instruction
2696 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2697 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2699 Chain = LoadTOCPtr.getValue(0);
2700 InFlag = LoadTOCPtr.getValue(1);
2702 MTCTROps[0] = Chain;
2703 MTCTROps[1] = LoadFuncPtr;
2704 MTCTROps[2] = InFlag;
2707 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2708 2 + (InFlag.getNode() != 0));
2709 InFlag = Chain.getValue(1);
2712 NodeTys.push_back(MVT::Other);
2713 NodeTys.push_back(MVT::Glue);
2714 Ops.push_back(Chain);
2715 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2717 // Add CTR register as callee so a bctr can be emitted later.
2719 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2722 // If this is a direct call, pass the chain and the callee.
2723 if (Callee.getNode()) {
2724 Ops.push_back(Chain);
2725 Ops.push_back(Callee);
2727 // If this is a tail call add stack pointer delta.
2729 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2731 // Add argument registers to the end of the list so that they are known live
2733 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2734 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2735 RegsToPass[i].second.getValueType()));
2741 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2742 CallingConv::ID CallConv, bool isVarArg,
2743 const SmallVectorImpl<ISD::InputArg> &Ins,
2744 DebugLoc dl, SelectionDAG &DAG,
2745 SmallVectorImpl<SDValue> &InVals) const {
2747 SmallVector<CCValAssign, 16> RVLocs;
2748 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2749 getTargetMachine(), RVLocs, *DAG.getContext());
2750 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2752 // Copy all of the result registers out of their specified physreg.
2753 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2754 CCValAssign &VA = RVLocs[i];
2755 EVT VT = VA.getValVT();
2756 assert(VA.isRegLoc() && "Can only return in registers!");
2757 Chain = DAG.getCopyFromReg(Chain, dl,
2758 VA.getLocReg(), VT, InFlag).getValue(1);
2759 InVals.push_back(Chain.getValue(0));
2760 InFlag = Chain.getValue(2);
2767 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2768 bool isTailCall, bool isVarArg,
2770 SmallVector<std::pair<unsigned, SDValue>, 8>
2772 SDValue InFlag, SDValue Chain,
2774 int SPDiff, unsigned NumBytes,
2775 const SmallVectorImpl<ISD::InputArg> &Ins,
2776 SmallVectorImpl<SDValue> &InVals) const {
2777 std::vector<EVT> NodeTys;
2778 SmallVector<SDValue, 8> Ops;
2779 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2780 isTailCall, RegsToPass, Ops, NodeTys,
2783 // When performing tail call optimization the callee pops its arguments off
2784 // the stack. Account for this here so these bytes can be pushed back on in
2785 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2786 int BytesCalleePops =
2787 (CallConv == CallingConv::Fast &&
2788 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
2790 // Add a register mask operand representing the call-preserved registers.
2791 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2792 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
2793 assert(Mask && "Missing call preserved mask for calling convention");
2794 Ops.push_back(DAG.getRegisterMask(Mask));
2796 if (InFlag.getNode())
2797 Ops.push_back(InFlag);
2801 // If this is the first return lowered for this function, add the regs
2802 // to the liveout set for the function.
2803 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2804 SmallVector<CCValAssign, 16> RVLocs;
2805 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2806 getTargetMachine(), RVLocs, *DAG.getContext());
2807 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2808 for (unsigned i = 0; i != RVLocs.size(); ++i)
2809 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2812 assert(((Callee.getOpcode() == ISD::Register &&
2813 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2814 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2815 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2816 isa<ConstantSDNode>(Callee)) &&
2817 "Expecting an global address, external symbol, absolute value or register");
2819 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2822 // Add a NOP immediately after the branch instruction when using the 64-bit
2823 // SVR4 ABI. At link time, if caller and callee are in a different module and
2824 // thus have a different TOC, the call will be replaced with a call to a stub
2825 // function which saves the current TOC, loads the TOC of the callee and
2826 // branches to the callee. The NOP will be replaced with a load instruction
2827 // which restores the TOC of the caller from the TOC save slot of the current
2828 // stack frame. If caller and callee belong to the same module (and have the
2829 // same TOC), the NOP will remain unchanged.
2831 bool needsTOCRestore = false;
2832 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2833 if (CallOpc == PPCISD::BCTRL_SVR4) {
2834 // This is a call through a function pointer.
2835 // Restore the caller TOC from the save area into R2.
2836 // See PrepareCall() for more information about calls through function
2837 // pointers in the 64-bit SVR4 ABI.
2838 // We are using a target-specific load with r2 hard coded, because the
2839 // result of a target-independent load would never go directly into r2,
2840 // since r2 is a reserved register (which prevents the register allocator
2841 // from allocating it), resulting in an additional register being
2842 // allocated and an unnecessary move instruction being generated.
2843 needsTOCRestore = true;
2844 } else if (CallOpc == PPCISD::CALL_SVR4) {
2845 // Otherwise insert NOP.
2846 CallOpc = PPCISD::CALL_NOP_SVR4;
2850 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2851 InFlag = Chain.getValue(1);
2853 if (needsTOCRestore) {
2854 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2855 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2856 InFlag = Chain.getValue(1);
2859 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2860 DAG.getIntPtrConstant(BytesCalleePops, true),
2863 InFlag = Chain.getValue(1);
2865 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2866 Ins, dl, DAG, InVals);
2870 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2871 SmallVectorImpl<SDValue> &InVals) const {
2872 SelectionDAG &DAG = CLI.DAG;
2873 DebugLoc &dl = CLI.DL;
2874 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
2875 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
2876 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
2877 SDValue Chain = CLI.Chain;
2878 SDValue Callee = CLI.Callee;
2879 bool &isTailCall = CLI.IsTailCall;
2880 CallingConv::ID CallConv = CLI.CallConv;
2881 bool isVarArg = CLI.IsVarArg;
2884 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2887 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2888 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2889 isTailCall, Outs, OutVals, Ins,
2892 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2893 isTailCall, Outs, OutVals, Ins,
2898 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2899 CallingConv::ID CallConv, bool isVarArg,
2901 const SmallVectorImpl<ISD::OutputArg> &Outs,
2902 const SmallVectorImpl<SDValue> &OutVals,
2903 const SmallVectorImpl<ISD::InputArg> &Ins,
2904 DebugLoc dl, SelectionDAG &DAG,
2905 SmallVectorImpl<SDValue> &InVals) const {
2906 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2907 // of the 32-bit SVR4 ABI stack frame layout.
2909 assert((CallConv == CallingConv::C ||
2910 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2912 unsigned PtrByteSize = 4;
2914 MachineFunction &MF = DAG.getMachineFunction();
2916 // Mark this function as potentially containing a function that contains a
2917 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2918 // and restoring the callers stack pointer in this functions epilog. This is
2919 // done because by tail calling the called function might overwrite the value
2920 // in this function's (MF) stack pointer stack slot 0(SP).
2921 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
2922 CallConv == CallingConv::Fast)
2923 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2925 // Count how many bytes are to be pushed on the stack, including the linkage
2926 // area, parameter list area and the part of the local variable space which
2927 // contains copies of aggregates which are passed by value.
2929 // Assign locations to all of the outgoing arguments.
2930 SmallVector<CCValAssign, 16> ArgLocs;
2931 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2932 getTargetMachine(), ArgLocs, *DAG.getContext());
2934 // Reserve space for the linkage area on the stack.
2935 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2938 // Handle fixed and variable vector arguments differently.
2939 // Fixed vector arguments go into registers as long as registers are
2940 // available. Variable vector arguments always go into memory.
2941 unsigned NumArgs = Outs.size();
2943 for (unsigned i = 0; i != NumArgs; ++i) {
2944 MVT ArgVT = Outs[i].VT;
2945 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2948 if (Outs[i].IsFixed) {
2949 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2952 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2958 errs() << "Call operand #" << i << " has unhandled type "
2959 << EVT(ArgVT).getEVTString() << "\n";
2961 llvm_unreachable(0);
2965 // All arguments are treated the same.
2966 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2969 // Assign locations to all of the outgoing aggregate by value arguments.
2970 SmallVector<CCValAssign, 16> ByValArgLocs;
2971 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2972 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2974 // Reserve stack space for the allocations in CCInfo.
2975 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2977 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2979 // Size of the linkage area, parameter list area and the part of the local
2980 // space variable where copies of aggregates which are passed by value are
2982 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2984 // Calculate by how many bytes the stack has to be adjusted in case of tail
2985 // call optimization.
2986 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2988 // Adjust the stack pointer for the new arguments...
2989 // These operations are automatically eliminated by the prolog/epilog pass
2990 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2991 SDValue CallSeqStart = Chain;
2993 // Load the return address and frame pointer so it can be moved somewhere else
2996 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2999 // Set up a copy of the stack pointer for use loading and storing any
3000 // arguments that may not fit in the registers available for argument
3002 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3004 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3005 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3006 SmallVector<SDValue, 8> MemOpChains;
3008 bool seenFloatArg = false;
3009 // Walk the register/memloc assignments, inserting copies/loads.
3010 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3013 CCValAssign &VA = ArgLocs[i];
3014 SDValue Arg = OutVals[i];
3015 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3017 if (Flags.isByVal()) {
3018 // Argument is an aggregate which is passed by value, thus we need to
3019 // create a copy of it in the local variable space of the current stack
3020 // frame (which is the stack frame of the caller) and pass the address of
3021 // this copy to the callee.
3022 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3023 CCValAssign &ByValVA = ByValArgLocs[j++];
3024 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3026 // Memory reserved in the local variable space of the callers stack frame.
3027 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3029 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3030 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3032 // Create a copy of the argument in the local area of the current
3034 SDValue MemcpyCall =
3035 CreateCopyOfByValArgument(Arg, PtrOff,
3036 CallSeqStart.getNode()->getOperand(0),
3039 // This must go outside the CALLSEQ_START..END.
3040 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3041 CallSeqStart.getNode()->getOperand(1));
3042 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3043 NewCallSeqStart.getNode());
3044 Chain = CallSeqStart = NewCallSeqStart;
3046 // Pass the address of the aggregate copy on the stack either in a
3047 // physical register or in the parameter list area of the current stack
3048 // frame to the callee.
3052 if (VA.isRegLoc()) {
3053 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3054 // Put argument in a physical register.
3055 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3057 // Put argument in the parameter list area of the current stack frame.
3058 assert(VA.isMemLoc());
3059 unsigned LocMemOffset = VA.getLocMemOffset();
3062 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3063 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3065 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3066 MachinePointerInfo(),
3069 // Calculate and remember argument location.
3070 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3076 if (!MemOpChains.empty())
3077 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3078 &MemOpChains[0], MemOpChains.size());
3080 // Set CR6 to true if this is a vararg call with floating args passed in
3083 SDValue SetCR(DAG.getMachineNode(seenFloatArg ? PPC::CRSET : PPC::CRUNSET,
3085 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
3088 // Build a sequence of copy-to-reg nodes chained together with token chain
3089 // and flag operands which copy the outgoing args into the appropriate regs.
3091 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3092 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3093 RegsToPass[i].second, InFlag);
3094 InFlag = Chain.getValue(1);
3098 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3099 false, TailCallArguments);
3101 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3102 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3107 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3108 CallingConv::ID CallConv, bool isVarArg,
3110 const SmallVectorImpl<ISD::OutputArg> &Outs,
3111 const SmallVectorImpl<SDValue> &OutVals,
3112 const SmallVectorImpl<ISD::InputArg> &Ins,
3113 DebugLoc dl, SelectionDAG &DAG,
3114 SmallVectorImpl<SDValue> &InVals) const {
3116 unsigned NumOps = Outs.size();
3118 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3119 bool isPPC64 = PtrVT == MVT::i64;
3120 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3122 MachineFunction &MF = DAG.getMachineFunction();
3124 // Mark this function as potentially containing a function that contains a
3125 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3126 // and restoring the callers stack pointer in this functions epilog. This is
3127 // done because by tail calling the called function might overwrite the value
3128 // in this function's (MF) stack pointer stack slot 0(SP).
3129 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3130 CallConv == CallingConv::Fast)
3131 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3133 unsigned nAltivecParamsAtEnd = 0;
3135 // Count how many bytes are to be pushed on the stack, including the linkage
3136 // area, and parameter passing area. We start with 24/48 bytes, which is
3137 // prereserved space for [SP][CR][LR][3 x unused].
3139 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3141 nAltivecParamsAtEnd);
3143 // Calculate by how many bytes the stack has to be adjusted in case of tail
3144 // call optimization.
3145 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3147 // To protect arguments on the stack from being clobbered in a tail call,
3148 // force all the loads to happen before doing any other lowering.
3150 Chain = DAG.getStackArgumentTokenFactor(Chain);
3152 // Adjust the stack pointer for the new arguments...
3153 // These operations are automatically eliminated by the prolog/epilog pass
3154 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3155 SDValue CallSeqStart = Chain;
3157 // Load the return address and frame pointer so it can be move somewhere else
3160 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3163 // Set up a copy of the stack pointer for use loading and storing any
3164 // arguments that may not fit in the registers available for argument
3168 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3170 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3172 // Figure out which arguments are going to go in registers, and which in
3173 // memory. Also, if this is a vararg function, floating point operations
3174 // must be stored to our stack, and loaded into integer regs as well, if
3175 // any integer regs are available for argument passing.
3176 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3177 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3179 static const uint16_t GPR_32[] = { // 32-bit registers.
3180 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3181 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3183 static const uint16_t GPR_64[] = { // 64-bit registers.
3184 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3185 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3187 static const uint16_t *FPR = GetFPR();
3189 static const uint16_t VR[] = {
3190 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3191 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3193 const unsigned NumGPRs = array_lengthof(GPR_32);
3194 const unsigned NumFPRs = 13;
3195 const unsigned NumVRs = array_lengthof(VR);
3197 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3199 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3200 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3202 SmallVector<SDValue, 8> MemOpChains;
3203 for (unsigned i = 0; i != NumOps; ++i) {
3204 SDValue Arg = OutVals[i];
3205 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3207 // PtrOff will be used to store the current argument to the stack if a
3208 // register cannot be found for it.
3211 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3213 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3215 // On PPC64, promote integers to 64-bit values.
3216 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3217 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3218 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3219 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3222 // FIXME memcpy is used way more than necessary. Correctness first.
3223 if (Flags.isByVal()) {
3224 unsigned Size = Flags.getByValSize();
3225 if (Size==1 || Size==2) {
3226 // Very small objects are passed right-justified.
3227 // Everything else is passed left-justified.
3228 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3229 if (GPR_idx != NumGPRs) {
3230 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3231 MachinePointerInfo(), VT,
3233 MemOpChains.push_back(Load.getValue(1));
3234 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3236 ArgOffset += PtrByteSize;
3238 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3239 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3240 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3241 CallSeqStart.getNode()->getOperand(0),
3243 // This must go outside the CALLSEQ_START..END.
3244 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3245 CallSeqStart.getNode()->getOperand(1));
3246 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3247 NewCallSeqStart.getNode());
3248 Chain = CallSeqStart = NewCallSeqStart;
3249 ArgOffset += PtrByteSize;
3253 // Copy entire object into memory. There are cases where gcc-generated
3254 // code assumes it is there, even if it could be put entirely into
3255 // registers. (This is not what the doc says.)
3256 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3257 CallSeqStart.getNode()->getOperand(0),
3259 // This must go outside the CALLSEQ_START..END.
3260 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3261 CallSeqStart.getNode()->getOperand(1));
3262 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3263 Chain = CallSeqStart = NewCallSeqStart;
3264 // And copy the pieces of it that fit into registers.
3265 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3266 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3267 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3268 if (GPR_idx != NumGPRs) {
3269 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3270 MachinePointerInfo(),
3271 false, false, false, 0);
3272 MemOpChains.push_back(Load.getValue(1));
3273 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3274 ArgOffset += PtrByteSize;
3276 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3283 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3284 default: llvm_unreachable("Unexpected ValueType for argument!");
3287 if (GPR_idx != NumGPRs) {
3288 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3290 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3291 isPPC64, isTailCall, false, MemOpChains,
3292 TailCallArguments, dl);
3294 ArgOffset += PtrByteSize;
3298 if (FPR_idx != NumFPRs) {
3299 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3302 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3303 MachinePointerInfo(), false, false, 0);
3304 MemOpChains.push_back(Store);
3306 // Float varargs are always shadowed in available integer registers
3307 if (GPR_idx != NumGPRs) {
3308 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3309 MachinePointerInfo(), false, false,
3311 MemOpChains.push_back(Load.getValue(1));
3312 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3314 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3315 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3316 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3317 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3318 MachinePointerInfo(),
3319 false, false, false, 0);
3320 MemOpChains.push_back(Load.getValue(1));
3321 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3324 // If we have any FPRs remaining, we may also have GPRs remaining.
3325 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3327 if (GPR_idx != NumGPRs)
3329 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3330 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3334 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3335 isPPC64, isTailCall, false, MemOpChains,
3336 TailCallArguments, dl);
3341 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3348 // These go aligned on the stack, or in the corresponding R registers
3349 // when within range. The Darwin PPC ABI doc claims they also go in
3350 // V registers; in fact gcc does this only for arguments that are
3351 // prototyped, not for those that match the ... We do it for all
3352 // arguments, seems to work.
3353 while (ArgOffset % 16 !=0) {
3354 ArgOffset += PtrByteSize;
3355 if (GPR_idx != NumGPRs)
3358 // We could elide this store in the case where the object fits
3359 // entirely in R registers. Maybe later.
3360 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3361 DAG.getConstant(ArgOffset, PtrVT));
3362 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3363 MachinePointerInfo(), false, false, 0);
3364 MemOpChains.push_back(Store);
3365 if (VR_idx != NumVRs) {
3366 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3367 MachinePointerInfo(),
3368 false, false, false, 0);
3369 MemOpChains.push_back(Load.getValue(1));
3370 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3373 for (unsigned i=0; i<16; i+=PtrByteSize) {
3374 if (GPR_idx == NumGPRs)
3376 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3377 DAG.getConstant(i, PtrVT));
3378 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3379 false, false, false, 0);
3380 MemOpChains.push_back(Load.getValue(1));
3381 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3386 // Non-varargs Altivec params generally go in registers, but have
3387 // stack space allocated at the end.
3388 if (VR_idx != NumVRs) {
3389 // Doesn't have GPR space allocated.
3390 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3391 } else if (nAltivecParamsAtEnd==0) {
3392 // We are emitting Altivec params in order.
3393 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3394 isPPC64, isTailCall, true, MemOpChains,
3395 TailCallArguments, dl);
3401 // If all Altivec parameters fit in registers, as they usually do,
3402 // they get stack space following the non-Altivec parameters. We
3403 // don't track this here because nobody below needs it.
3404 // If there are more Altivec parameters than fit in registers emit
3406 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3408 // Offset is aligned; skip 1st 12 params which go in V registers.
3409 ArgOffset = ((ArgOffset+15)/16)*16;
3411 for (unsigned i = 0; i != NumOps; ++i) {
3412 SDValue Arg = OutVals[i];
3413 EVT ArgType = Outs[i].VT;
3414 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3415 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3418 // We are emitting Altivec params in order.
3419 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3420 isPPC64, isTailCall, true, MemOpChains,
3421 TailCallArguments, dl);
3428 if (!MemOpChains.empty())
3429 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3430 &MemOpChains[0], MemOpChains.size());
3432 // Check if this is an indirect call (MTCTR/BCTRL).
3433 // See PrepareCall() for more information about calls through function
3434 // pointers in the 64-bit SVR4 ABI.
3435 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3436 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3437 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3438 !isBLACompatibleAddress(Callee, DAG)) {
3439 // Load r2 into a virtual register and store it to the TOC save area.
3440 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3441 // TOC save area offset.
3442 SDValue PtrOff = DAG.getIntPtrConstant(40);
3443 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3444 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3448 // On Darwin, R12 must contain the address of an indirect callee. This does
3449 // not mean the MTCTR instruction must use R12; it's easier to model this as
3450 // an extra parameter, so do that.
3452 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3453 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3454 !isBLACompatibleAddress(Callee, DAG))
3455 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3456 PPC::R12), Callee));
3458 // Build a sequence of copy-to-reg nodes chained together with token chain
3459 // and flag operands which copy the outgoing args into the appropriate regs.
3461 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3462 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3463 RegsToPass[i].second, InFlag);
3464 InFlag = Chain.getValue(1);
3468 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3469 FPOp, true, TailCallArguments);
3471 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3472 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3477 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3478 MachineFunction &MF, bool isVarArg,
3479 const SmallVectorImpl<ISD::OutputArg> &Outs,
3480 LLVMContext &Context) const {
3481 SmallVector<CCValAssign, 16> RVLocs;
3482 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3484 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3488 PPCTargetLowering::LowerReturn(SDValue Chain,
3489 CallingConv::ID CallConv, bool isVarArg,
3490 const SmallVectorImpl<ISD::OutputArg> &Outs,
3491 const SmallVectorImpl<SDValue> &OutVals,
3492 DebugLoc dl, SelectionDAG &DAG) const {
3494 SmallVector<CCValAssign, 16> RVLocs;
3495 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3496 getTargetMachine(), RVLocs, *DAG.getContext());
3497 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3499 // If this is the first return lowered for this function, add the regs to the
3500 // liveout set for the function.
3501 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3502 for (unsigned i = 0; i != RVLocs.size(); ++i)
3503 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3508 // Copy the result values into the output registers.
3509 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3510 CCValAssign &VA = RVLocs[i];
3511 assert(VA.isRegLoc() && "Can only return in registers!");
3512 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3514 Flag = Chain.getValue(1);
3518 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3520 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3523 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3524 const PPCSubtarget &Subtarget) const {
3525 // When we pop the dynamic allocation we need to restore the SP link.
3526 DebugLoc dl = Op.getDebugLoc();
3528 // Get the corect type for pointers.
3529 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3531 // Construct the stack pointer operand.
3532 bool isPPC64 = Subtarget.isPPC64();
3533 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3534 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3536 // Get the operands for the STACKRESTORE.
3537 SDValue Chain = Op.getOperand(0);
3538 SDValue SaveSP = Op.getOperand(1);
3540 // Load the old link SP.
3541 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3542 MachinePointerInfo(),
3543 false, false, false, 0);
3545 // Restore the stack pointer.
3546 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3548 // Store the old link SP.
3549 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3556 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3557 MachineFunction &MF = DAG.getMachineFunction();
3558 bool isPPC64 = PPCSubTarget.isPPC64();
3559 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3560 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3562 // Get current frame pointer save index. The users of this index will be
3563 // primarily DYNALLOC instructions.
3564 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3565 int RASI = FI->getReturnAddrSaveIndex();
3567 // If the frame pointer save index hasn't been defined yet.
3569 // Find out what the fix offset of the frame pointer save area.
3570 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3571 // Allocate the frame index for frame pointer save area.
3572 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3574 FI->setReturnAddrSaveIndex(RASI);
3576 return DAG.getFrameIndex(RASI, PtrVT);
3580 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3581 MachineFunction &MF = DAG.getMachineFunction();
3582 bool isPPC64 = PPCSubTarget.isPPC64();
3583 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3584 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3586 // Get current frame pointer save index. The users of this index will be
3587 // primarily DYNALLOC instructions.
3588 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3589 int FPSI = FI->getFramePointerSaveIndex();
3591 // If the frame pointer save index hasn't been defined yet.
3593 // Find out what the fix offset of the frame pointer save area.
3594 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3597 // Allocate the frame index for frame pointer save area.
3598 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3600 FI->setFramePointerSaveIndex(FPSI);
3602 return DAG.getFrameIndex(FPSI, PtrVT);
3605 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3607 const PPCSubtarget &Subtarget) const {
3609 SDValue Chain = Op.getOperand(0);
3610 SDValue Size = Op.getOperand(1);
3611 DebugLoc dl = Op.getDebugLoc();
3613 // Get the corect type for pointers.
3614 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3616 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3617 DAG.getConstant(0, PtrVT), Size);
3618 // Construct a node for the frame pointer save index.
3619 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3620 // Build a DYNALLOC node.
3621 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3622 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3623 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3626 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3628 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3629 // Not FP? Not a fsel.
3630 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3631 !Op.getOperand(2).getValueType().isFloatingPoint())
3634 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3636 // Cannot handle SETEQ/SETNE.
3637 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3639 EVT ResVT = Op.getValueType();
3640 EVT CmpVT = Op.getOperand(0).getValueType();
3641 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3642 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3643 DebugLoc dl = Op.getDebugLoc();
3645 // If the RHS of the comparison is a 0.0, we don't need to do the
3646 // subtraction at all.
3647 if (isFloatingPointZero(RHS))
3649 default: break; // SETUO etc aren't handled by fsel.
3652 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3655 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3656 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3657 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3660 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3663 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3664 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3665 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3666 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3671 default: break; // SETUO etc aren't handled by fsel.
3674 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3675 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3676 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3677 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3680 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3681 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3682 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3683 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3686 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3687 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3688 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3689 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3692 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3693 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3694 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3695 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3700 // FIXME: Split this code up when LegalizeDAGTypes lands.
3701 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3702 DebugLoc dl) const {
3703 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3704 SDValue Src = Op.getOperand(0);
3705 if (Src.getValueType() == MVT::f32)
3706 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3709 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3710 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3712 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3717 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3721 // Convert the FP value to an int value through memory.
3722 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3724 // Emit a store to the stack slot.
3725 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3726 MachinePointerInfo(), false, false, 0);
3728 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3730 if (Op.getValueType() == MVT::i32)
3731 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3732 DAG.getConstant(4, FIPtr.getValueType()));
3733 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3734 false, false, false, 0);
3737 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3738 SelectionDAG &DAG) const {
3739 DebugLoc dl = Op.getDebugLoc();
3740 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3741 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3744 if (Op.getOperand(0).getValueType() == MVT::i64) {
3745 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3746 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3747 if (Op.getValueType() == MVT::f32)
3748 FP = DAG.getNode(ISD::FP_ROUND, dl,
3749 MVT::f32, FP, DAG.getIntPtrConstant(0));
3753 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3754 "Unhandled SINT_TO_FP type in custom expander!");
3755 // Since we only generate this in 64-bit mode, we can take advantage of
3756 // 64-bit registers. In particular, sign extend the input value into the
3757 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3758 // then lfd it and fcfid it.
3759 MachineFunction &MF = DAG.getMachineFunction();
3760 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3761 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3762 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3763 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3765 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3768 // STD the extended value into the stack slot.
3769 MachineMemOperand *MMO =
3770 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3771 MachineMemOperand::MOStore, 8, 8);
3772 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3774 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3775 Ops, 4, MVT::i64, MMO);
3776 // Load the value as a double.
3777 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3778 false, false, false, 0);
3780 // FCFID it and return it.
3781 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3782 if (Op.getValueType() == MVT::f32)
3783 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3787 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3788 SelectionDAG &DAG) const {
3789 DebugLoc dl = Op.getDebugLoc();
3791 The rounding mode is in bits 30:31 of FPSR, and has the following
3798 FLT_ROUNDS, on the other hand, expects the following:
3805 To perform the conversion, we do:
3806 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3809 MachineFunction &MF = DAG.getMachineFunction();
3810 EVT VT = Op.getValueType();
3811 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3812 std::vector<EVT> NodeTys;
3813 SDValue MFFSreg, InFlag;
3815 // Save FP Control Word to register
3816 NodeTys.push_back(MVT::f64); // return register
3817 NodeTys.push_back(MVT::Glue); // unused in this context
3818 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3820 // Save FP register to stack slot
3821 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3822 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3823 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3824 StackSlot, MachinePointerInfo(), false, false,0);
3826 // Load FP Control Word from low 32 bits of stack slot.
3827 SDValue Four = DAG.getConstant(4, PtrVT);
3828 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3829 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3830 false, false, false, 0);
3832 // Transform as necessary
3834 DAG.getNode(ISD::AND, dl, MVT::i32,
3835 CWD, DAG.getConstant(3, MVT::i32));
3837 DAG.getNode(ISD::SRL, dl, MVT::i32,
3838 DAG.getNode(ISD::AND, dl, MVT::i32,
3839 DAG.getNode(ISD::XOR, dl, MVT::i32,
3840 CWD, DAG.getConstant(3, MVT::i32)),
3841 DAG.getConstant(3, MVT::i32)),
3842 DAG.getConstant(1, MVT::i32));
3845 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3847 return DAG.getNode((VT.getSizeInBits() < 16 ?
3848 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3851 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3852 EVT VT = Op.getValueType();
3853 unsigned BitWidth = VT.getSizeInBits();
3854 DebugLoc dl = Op.getDebugLoc();
3855 assert(Op.getNumOperands() == 3 &&
3856 VT == Op.getOperand(1).getValueType() &&
3859 // Expand into a bunch of logical ops. Note that these ops
3860 // depend on the PPC behavior for oversized shift amounts.
3861 SDValue Lo = Op.getOperand(0);
3862 SDValue Hi = Op.getOperand(1);
3863 SDValue Amt = Op.getOperand(2);
3864 EVT AmtVT = Amt.getValueType();
3866 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3867 DAG.getConstant(BitWidth, AmtVT), Amt);
3868 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3869 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3870 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3871 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3872 DAG.getConstant(-BitWidth, AmtVT));
3873 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3874 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3875 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3876 SDValue OutOps[] = { OutLo, OutHi };
3877 return DAG.getMergeValues(OutOps, 2, dl);
3880 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3881 EVT VT = Op.getValueType();
3882 DebugLoc dl = Op.getDebugLoc();
3883 unsigned BitWidth = VT.getSizeInBits();
3884 assert(Op.getNumOperands() == 3 &&
3885 VT == Op.getOperand(1).getValueType() &&
3888 // Expand into a bunch of logical ops. Note that these ops
3889 // depend on the PPC behavior for oversized shift amounts.
3890 SDValue Lo = Op.getOperand(0);
3891 SDValue Hi = Op.getOperand(1);
3892 SDValue Amt = Op.getOperand(2);
3893 EVT AmtVT = Amt.getValueType();
3895 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3896 DAG.getConstant(BitWidth, AmtVT), Amt);
3897 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3898 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3899 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3900 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3901 DAG.getConstant(-BitWidth, AmtVT));
3902 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3903 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3904 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3905 SDValue OutOps[] = { OutLo, OutHi };
3906 return DAG.getMergeValues(OutOps, 2, dl);
3909 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3910 DebugLoc dl = Op.getDebugLoc();
3911 EVT VT = Op.getValueType();
3912 unsigned BitWidth = VT.getSizeInBits();
3913 assert(Op.getNumOperands() == 3 &&
3914 VT == Op.getOperand(1).getValueType() &&
3917 // Expand into a bunch of logical ops, followed by a select_cc.
3918 SDValue Lo = Op.getOperand(0);
3919 SDValue Hi = Op.getOperand(1);
3920 SDValue Amt = Op.getOperand(2);
3921 EVT AmtVT = Amt.getValueType();
3923 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3924 DAG.getConstant(BitWidth, AmtVT), Amt);
3925 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3926 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3927 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3928 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3929 DAG.getConstant(-BitWidth, AmtVT));
3930 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3931 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3932 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3933 Tmp4, Tmp6, ISD::SETLE);
3934 SDValue OutOps[] = { OutLo, OutHi };
3935 return DAG.getMergeValues(OutOps, 2, dl);
3938 //===----------------------------------------------------------------------===//
3939 // Vector related lowering.
3942 /// BuildSplatI - Build a canonical splati of Val with an element size of
3943 /// SplatSize. Cast the result to VT.
3944 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3945 SelectionDAG &DAG, DebugLoc dl) {
3946 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3948 static const EVT VTys[] = { // canonical VT to use for each size.
3949 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3952 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3954 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3958 EVT CanonicalVT = VTys[SplatSize-1];
3960 // Build a canonical splat for this value.
3961 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3962 SmallVector<SDValue, 8> Ops;
3963 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3964 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3965 &Ops[0], Ops.size());
3966 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3969 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3970 /// specified intrinsic ID.
3971 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3972 SelectionDAG &DAG, DebugLoc dl,
3973 EVT DestVT = MVT::Other) {
3974 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3975 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3976 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3979 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3980 /// specified intrinsic ID.
3981 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3982 SDValue Op2, SelectionDAG &DAG,
3983 DebugLoc dl, EVT DestVT = MVT::Other) {
3984 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3985 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3986 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3990 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3991 /// amount. The result has the specified value type.
3992 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3993 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3994 // Force LHS/RHS to be the right type.
3995 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3996 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3999 for (unsigned i = 0; i != 16; ++i)
4001 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4002 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4005 // If this is a case we can't handle, return null and let the default
4006 // expansion code take care of it. If we CAN select this case, and if it
4007 // selects to a single instruction, return Op. Otherwise, if we can codegen
4008 // this case more efficiently than a constant pool load, lower it to the
4009 // sequence of ops that should be used.
4010 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4011 SelectionDAG &DAG) const {
4012 DebugLoc dl = Op.getDebugLoc();
4013 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4014 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4016 // Check if this is a splat of a constant value.
4017 APInt APSplatBits, APSplatUndef;
4018 unsigned SplatBitSize;
4020 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4021 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4024 unsigned SplatBits = APSplatBits.getZExtValue();
4025 unsigned SplatUndef = APSplatUndef.getZExtValue();
4026 unsigned SplatSize = SplatBitSize / 8;
4028 // First, handle single instruction cases.
4031 if (SplatBits == 0) {
4032 // Canonicalize all zero vectors to be v4i32.
4033 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4034 SDValue Z = DAG.getConstant(0, MVT::i32);
4035 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4036 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4041 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4042 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4044 if (SextVal >= -16 && SextVal <= 15)
4045 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4048 // Two instruction sequences.
4050 // If this value is in the range [-32,30] and is even, use:
4051 // tmp = VSPLTI[bhw], result = add tmp, tmp
4052 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4053 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4054 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4055 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4058 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4059 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4061 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4062 // Make -1 and vspltisw -1:
4063 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4065 // Make the VSLW intrinsic, computing 0x8000_0000.
4066 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4069 // xor by OnesV to invert it.
4070 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4071 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4074 // Check to see if this is a wide variety of vsplti*, binop self cases.
4075 static const signed char SplatCsts[] = {
4076 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4077 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4080 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4081 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4082 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4083 int i = SplatCsts[idx];
4085 // Figure out what shift amount will be used by altivec if shifted by i in
4087 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4089 // vsplti + shl self.
4090 if (SextVal == (i << (int)TypeShiftAmt)) {
4091 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4092 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4093 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4094 Intrinsic::ppc_altivec_vslw
4096 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4097 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4100 // vsplti + srl self.
4101 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4102 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4103 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4104 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4105 Intrinsic::ppc_altivec_vsrw
4107 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4108 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4111 // vsplti + sra self.
4112 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4113 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4114 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4115 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4116 Intrinsic::ppc_altivec_vsraw
4118 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4119 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4122 // vsplti + rol self.
4123 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4124 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4125 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4126 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4127 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4128 Intrinsic::ppc_altivec_vrlw
4130 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4131 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4134 // t = vsplti c, result = vsldoi t, t, 1
4135 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
4136 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4137 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4139 // t = vsplti c, result = vsldoi t, t, 2
4140 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
4141 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4142 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4144 // t = vsplti c, result = vsldoi t, t, 3
4145 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4146 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4147 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4151 // Three instruction sequences.
4153 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4154 if (SextVal >= 0 && SextVal <= 31) {
4155 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4156 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4157 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4158 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4160 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4161 if (SextVal >= -31 && SextVal <= 0) {
4162 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4163 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4164 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4165 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4171 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4172 /// the specified operations to build the shuffle.
4173 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4174 SDValue RHS, SelectionDAG &DAG,
4176 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4177 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4178 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4181 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4193 if (OpNum == OP_COPY) {
4194 if (LHSID == (1*9+2)*9+3) return LHS;
4195 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4199 SDValue OpLHS, OpRHS;
4200 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4201 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4205 default: llvm_unreachable("Unknown i32 permute!");
4207 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4208 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4209 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4210 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4213 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4214 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4215 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4216 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4219 for (unsigned i = 0; i != 16; ++i)
4220 ShufIdxs[i] = (i&3)+0;
4223 for (unsigned i = 0; i != 16; ++i)
4224 ShufIdxs[i] = (i&3)+4;
4227 for (unsigned i = 0; i != 16; ++i)
4228 ShufIdxs[i] = (i&3)+8;
4231 for (unsigned i = 0; i != 16; ++i)
4232 ShufIdxs[i] = (i&3)+12;
4235 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4237 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4239 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4241 EVT VT = OpLHS.getValueType();
4242 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4243 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4244 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4245 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4248 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4249 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4250 /// return the code it can be lowered into. Worst case, it can always be
4251 /// lowered into a vperm.
4252 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4253 SelectionDAG &DAG) const {
4254 DebugLoc dl = Op.getDebugLoc();
4255 SDValue V1 = Op.getOperand(0);
4256 SDValue V2 = Op.getOperand(1);
4257 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4258 EVT VT = Op.getValueType();
4260 // Cases that are handled by instructions that take permute immediates
4261 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4262 // selected by the instruction selector.
4263 if (V2.getOpcode() == ISD::UNDEF) {
4264 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4265 PPC::isSplatShuffleMask(SVOp, 2) ||
4266 PPC::isSplatShuffleMask(SVOp, 4) ||
4267 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4268 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4269 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4270 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4271 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4272 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4273 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4274 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4275 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4280 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4281 // and produce a fixed permutation. If any of these match, do not lower to
4283 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4284 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4285 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4286 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4287 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4288 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4289 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4290 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4291 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4294 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4295 // perfect shuffle table to emit an optimal matching sequence.
4296 ArrayRef<int> PermMask = SVOp->getMask();
4298 unsigned PFIndexes[4];
4299 bool isFourElementShuffle = true;
4300 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4301 unsigned EltNo = 8; // Start out undef.
4302 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4303 if (PermMask[i*4+j] < 0)
4304 continue; // Undef, ignore it.
4306 unsigned ByteSource = PermMask[i*4+j];
4307 if ((ByteSource & 3) != j) {
4308 isFourElementShuffle = false;
4313 EltNo = ByteSource/4;
4314 } else if (EltNo != ByteSource/4) {
4315 isFourElementShuffle = false;
4319 PFIndexes[i] = EltNo;
4322 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4323 // perfect shuffle vector to determine if it is cost effective to do this as
4324 // discrete instructions, or whether we should use a vperm.
4325 if (isFourElementShuffle) {
4326 // Compute the index in the perfect shuffle table.
4327 unsigned PFTableIndex =
4328 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4330 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4331 unsigned Cost = (PFEntry >> 30);
4333 // Determining when to avoid vperm is tricky. Many things affect the cost
4334 // of vperm, particularly how many times the perm mask needs to be computed.
4335 // For example, if the perm mask can be hoisted out of a loop or is already
4336 // used (perhaps because there are multiple permutes with the same shuffle
4337 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4338 // the loop requires an extra register.
4340 // As a compromise, we only emit discrete instructions if the shuffle can be
4341 // generated in 3 or fewer operations. When we have loop information
4342 // available, if this block is within a loop, we should avoid using vperm
4343 // for 3-operation perms and use a constant pool load instead.
4345 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4348 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4349 // vector that will get spilled to the constant pool.
4350 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4352 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4353 // that it is in input element units, not in bytes. Convert now.
4354 EVT EltVT = V1.getValueType().getVectorElementType();
4355 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4357 SmallVector<SDValue, 16> ResultMask;
4358 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4359 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4361 for (unsigned j = 0; j != BytesPerElement; ++j)
4362 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4366 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4367 &ResultMask[0], ResultMask.size());
4368 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4371 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4372 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4373 /// information about the intrinsic.
4374 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4376 unsigned IntrinsicID =
4377 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4380 switch (IntrinsicID) {
4381 default: return false;
4382 // Comparison predicates.
4383 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4384 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4385 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4386 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4387 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4388 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4389 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4390 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4391 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4392 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4393 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4394 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4395 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4397 // Normal Comparisons.
4398 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4399 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4400 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4401 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4402 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4403 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4404 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4405 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4406 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4407 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4408 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4409 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4410 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4415 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4416 /// lower, do it, otherwise return null.
4417 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4418 SelectionDAG &DAG) const {
4419 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4420 // opcode number of the comparison.
4421 DebugLoc dl = Op.getDebugLoc();
4424 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4425 return SDValue(); // Don't custom lower most intrinsics.
4427 // If this is a non-dot comparison, make the VCMP node and we are done.
4429 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4430 Op.getOperand(1), Op.getOperand(2),
4431 DAG.getConstant(CompareOpc, MVT::i32));
4432 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4435 // Create the PPCISD altivec 'dot' comparison node.
4437 Op.getOperand(2), // LHS
4438 Op.getOperand(3), // RHS
4439 DAG.getConstant(CompareOpc, MVT::i32)
4441 std::vector<EVT> VTs;
4442 VTs.push_back(Op.getOperand(2).getValueType());
4443 VTs.push_back(MVT::Glue);
4444 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4446 // Now that we have the comparison, emit a copy from the CR to a GPR.
4447 // This is flagged to the above dot comparison.
4448 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4449 DAG.getRegister(PPC::CR6, MVT::i32),
4450 CompNode.getValue(1));
4452 // Unpack the result based on how the target uses it.
4453 unsigned BitNo; // Bit # of CR6.
4454 bool InvertBit; // Invert result?
4455 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4456 default: // Can't happen, don't crash on invalid number though.
4457 case 0: // Return the value of the EQ bit of CR6.
4458 BitNo = 0; InvertBit = false;
4460 case 1: // Return the inverted value of the EQ bit of CR6.
4461 BitNo = 0; InvertBit = true;
4463 case 2: // Return the value of the LT bit of CR6.
4464 BitNo = 2; InvertBit = false;
4466 case 3: // Return the inverted value of the LT bit of CR6.
4467 BitNo = 2; InvertBit = true;
4471 // Shift the bit into the low position.
4472 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4473 DAG.getConstant(8-(3-BitNo), MVT::i32));
4475 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4476 DAG.getConstant(1, MVT::i32));
4478 // If we are supposed to, toggle the bit.
4480 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4481 DAG.getConstant(1, MVT::i32));
4485 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4486 SelectionDAG &DAG) const {
4487 DebugLoc dl = Op.getDebugLoc();
4488 // Create a stack slot that is 16-byte aligned.
4489 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4490 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4491 EVT PtrVT = getPointerTy();
4492 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4494 // Store the input value into Value#0 of the stack slot.
4495 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4496 Op.getOperand(0), FIdx, MachinePointerInfo(),
4499 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4500 false, false, false, 0);
4503 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4504 DebugLoc dl = Op.getDebugLoc();
4505 if (Op.getValueType() == MVT::v4i32) {
4506 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4508 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4509 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4511 SDValue RHSSwap = // = vrlw RHS, 16
4512 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4514 // Shrinkify inputs to v8i16.
4515 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4516 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4517 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4519 // Low parts multiplied together, generating 32-bit results (we ignore the
4521 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4522 LHS, RHS, DAG, dl, MVT::v4i32);
4524 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4525 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4526 // Shift the high parts up 16 bits.
4527 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4529 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4530 } else if (Op.getValueType() == MVT::v8i16) {
4531 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4533 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4535 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4536 LHS, RHS, Zero, DAG, dl);
4537 } else if (Op.getValueType() == MVT::v16i8) {
4538 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4540 // Multiply the even 8-bit parts, producing 16-bit sums.
4541 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4542 LHS, RHS, DAG, dl, MVT::v8i16);
4543 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4545 // Multiply the odd 8-bit parts, producing 16-bit sums.
4546 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4547 LHS, RHS, DAG, dl, MVT::v8i16);
4548 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4550 // Merge the results together.
4552 for (unsigned i = 0; i != 8; ++i) {
4554 Ops[i*2+1] = 2*i+1+16;
4556 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4558 llvm_unreachable("Unknown mul to lower!");
4562 /// LowerOperation - Provide custom lowering hooks for some operations.
4564 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4565 switch (Op.getOpcode()) {
4566 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4567 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4568 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4569 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4570 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4571 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4572 case ISD::SETCC: return LowerSETCC(Op, DAG);
4573 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
4574 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
4576 return LowerVASTART(Op, DAG, PPCSubTarget);
4579 return LowerVAARG(Op, DAG, PPCSubTarget);
4581 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4582 case ISD::DYNAMIC_STACKALLOC:
4583 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4585 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4586 case ISD::FP_TO_UINT:
4587 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4589 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4590 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4592 // Lower 64-bit shifts.
4593 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4594 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4595 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4597 // Vector-related lowering.
4598 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4599 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4600 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4601 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4602 case ISD::MUL: return LowerMUL(Op, DAG);
4604 // Frame & Return address.
4605 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4606 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4610 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4611 SmallVectorImpl<SDValue>&Results,
4612 SelectionDAG &DAG) const {
4613 const TargetMachine &TM = getTargetMachine();
4614 DebugLoc dl = N->getDebugLoc();
4615 switch (N->getOpcode()) {
4617 llvm_unreachable("Do not know how to custom type legalize this operation!");
4619 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
4620 || TM.getSubtarget<PPCSubtarget>().isPPC64())
4623 EVT VT = N->getValueType(0);
4625 if (VT == MVT::i64) {
4626 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
4628 Results.push_back(NewNode);
4629 Results.push_back(NewNode.getValue(1));
4633 case ISD::FP_ROUND_INREG: {
4634 assert(N->getValueType(0) == MVT::ppcf128);
4635 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4636 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4637 MVT::f64, N->getOperand(0),
4638 DAG.getIntPtrConstant(0));
4639 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4640 MVT::f64, N->getOperand(0),
4641 DAG.getIntPtrConstant(1));
4643 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4644 // of the long double, and puts FPSCR back the way it was. We do not
4645 // actually model FPSCR.
4646 std::vector<EVT> NodeTys;
4647 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4649 NodeTys.push_back(MVT::f64); // Return register
4650 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4651 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4652 MFFSreg = Result.getValue(0);
4653 InFlag = Result.getValue(1);
4656 NodeTys.push_back(MVT::Glue); // Returns a flag
4657 Ops[0] = DAG.getConstant(31, MVT::i32);
4659 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4660 InFlag = Result.getValue(0);
4663 NodeTys.push_back(MVT::Glue); // Returns a flag
4664 Ops[0] = DAG.getConstant(30, MVT::i32);
4666 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4667 InFlag = Result.getValue(0);
4670 NodeTys.push_back(MVT::f64); // result of add
4671 NodeTys.push_back(MVT::Glue); // Returns a flag
4675 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4676 FPreg = Result.getValue(0);
4677 InFlag = Result.getValue(1);
4680 NodeTys.push_back(MVT::f64);
4681 Ops[0] = DAG.getConstant(1, MVT::i32);
4685 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4686 FPreg = Result.getValue(0);
4688 // We know the low half is about to be thrown away, so just use something
4690 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4694 case ISD::FP_TO_SINT:
4695 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4701 //===----------------------------------------------------------------------===//
4702 // Other Lowering Code
4703 //===----------------------------------------------------------------------===//
4706 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4707 bool is64bit, unsigned BinOpcode) const {
4708 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4709 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4711 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4712 MachineFunction *F = BB->getParent();
4713 MachineFunction::iterator It = BB;
4716 unsigned dest = MI->getOperand(0).getReg();
4717 unsigned ptrA = MI->getOperand(1).getReg();
4718 unsigned ptrB = MI->getOperand(2).getReg();
4719 unsigned incr = MI->getOperand(3).getReg();
4720 DebugLoc dl = MI->getDebugLoc();
4722 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4723 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4724 F->insert(It, loopMBB);
4725 F->insert(It, exitMBB);
4726 exitMBB->splice(exitMBB->begin(), BB,
4727 llvm::next(MachineBasicBlock::iterator(MI)),
4729 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4731 MachineRegisterInfo &RegInfo = F->getRegInfo();
4732 unsigned TmpReg = (!BinOpcode) ? incr :
4733 RegInfo.createVirtualRegister(
4734 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4735 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4739 // fallthrough --> loopMBB
4740 BB->addSuccessor(loopMBB);
4743 // l[wd]arx dest, ptr
4744 // add r0, dest, incr
4745 // st[wd]cx. r0, ptr
4747 // fallthrough --> exitMBB
4749 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4750 .addReg(ptrA).addReg(ptrB);
4752 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4753 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4754 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4755 BuildMI(BB, dl, TII->get(PPC::BCC))
4756 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4757 BB->addSuccessor(loopMBB);
4758 BB->addSuccessor(exitMBB);
4767 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4768 MachineBasicBlock *BB,
4769 bool is8bit, // operation
4770 unsigned BinOpcode) const {
4771 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4772 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4773 // In 64 bit mode we have to use 64 bits for addresses, even though the
4774 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4775 // registers without caring whether they're 32 or 64, but here we're
4776 // doing actual arithmetic on the addresses.
4777 bool is64bit = PPCSubTarget.isPPC64();
4778 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4780 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4781 MachineFunction *F = BB->getParent();
4782 MachineFunction::iterator It = BB;
4785 unsigned dest = MI->getOperand(0).getReg();
4786 unsigned ptrA = MI->getOperand(1).getReg();
4787 unsigned ptrB = MI->getOperand(2).getReg();
4788 unsigned incr = MI->getOperand(3).getReg();
4789 DebugLoc dl = MI->getDebugLoc();
4791 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4792 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4793 F->insert(It, loopMBB);
4794 F->insert(It, exitMBB);
4795 exitMBB->splice(exitMBB->begin(), BB,
4796 llvm::next(MachineBasicBlock::iterator(MI)),
4798 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4800 MachineRegisterInfo &RegInfo = F->getRegInfo();
4801 const TargetRegisterClass *RC =
4802 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4803 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4804 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4805 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4806 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4807 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4808 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4809 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4810 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4811 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4812 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4813 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4814 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4816 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4820 // fallthrough --> loopMBB
4821 BB->addSuccessor(loopMBB);
4823 // The 4-byte load must be aligned, while a char or short may be
4824 // anywhere in the word. Hence all this nasty bookkeeping code.
4825 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4826 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4827 // xori shift, shift1, 24 [16]
4828 // rlwinm ptr, ptr1, 0, 0, 29
4829 // slw incr2, incr, shift
4830 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4831 // slw mask, mask2, shift
4833 // lwarx tmpDest, ptr
4834 // add tmp, tmpDest, incr2
4835 // andc tmp2, tmpDest, mask
4836 // and tmp3, tmp, mask
4837 // or tmp4, tmp3, tmp2
4840 // fallthrough --> exitMBB
4841 // srw dest, tmpDest, shift
4842 if (ptrA != ZeroReg) {
4843 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4844 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4845 .addReg(ptrA).addReg(ptrB);
4849 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4850 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4851 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4852 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4854 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4855 .addReg(Ptr1Reg).addImm(0).addImm(61);
4857 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4858 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4859 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4860 .addReg(incr).addReg(ShiftReg);
4862 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4864 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4865 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4867 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4868 .addReg(Mask2Reg).addReg(ShiftReg);
4871 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4872 .addReg(ZeroReg).addReg(PtrReg);
4874 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4875 .addReg(Incr2Reg).addReg(TmpDestReg);
4876 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4877 .addReg(TmpDestReg).addReg(MaskReg);
4878 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4879 .addReg(TmpReg).addReg(MaskReg);
4880 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4881 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4882 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4883 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4884 BuildMI(BB, dl, TII->get(PPC::BCC))
4885 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4886 BB->addSuccessor(loopMBB);
4887 BB->addSuccessor(exitMBB);
4892 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4898 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4899 MachineBasicBlock *BB) const {
4900 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4902 // To "insert" these instructions we actually have to insert their
4903 // control-flow patterns.
4904 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4905 MachineFunction::iterator It = BB;
4908 MachineFunction *F = BB->getParent();
4910 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4911 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4912 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4913 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4914 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4916 // The incoming instruction knows the destination vreg to set, the
4917 // condition code register to branch on, the true/false values to
4918 // select between, and a branch opcode to use.
4923 // cmpTY ccX, r1, r2
4925 // fallthrough --> copy0MBB
4926 MachineBasicBlock *thisMBB = BB;
4927 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4928 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4929 unsigned SelectPred = MI->getOperand(4).getImm();
4930 DebugLoc dl = MI->getDebugLoc();
4931 F->insert(It, copy0MBB);
4932 F->insert(It, sinkMBB);
4934 // Transfer the remainder of BB and its successor edges to sinkMBB.
4935 sinkMBB->splice(sinkMBB->begin(), BB,
4936 llvm::next(MachineBasicBlock::iterator(MI)),
4938 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4940 // Next, add the true and fallthrough blocks as its successors.
4941 BB->addSuccessor(copy0MBB);
4942 BB->addSuccessor(sinkMBB);
4944 BuildMI(BB, dl, TII->get(PPC::BCC))
4945 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4948 // %FalseValue = ...
4949 // # fallthrough to sinkMBB
4952 // Update machine-CFG edges
4953 BB->addSuccessor(sinkMBB);
4956 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4959 BuildMI(*BB, BB->begin(), dl,
4960 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4961 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4962 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4964 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4965 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4966 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4967 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4968 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4969 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4970 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4971 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4973 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4974 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4975 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4976 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4977 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4978 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4979 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4980 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4982 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4983 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4984 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4985 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4986 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4987 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4988 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4989 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4991 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4992 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4993 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4994 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4995 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4996 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4997 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4998 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5000 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5001 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5002 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5003 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5004 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5005 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5006 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5007 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5009 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5010 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5011 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5012 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5013 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5014 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5015 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5016 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5018 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5019 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5020 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5021 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5022 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5023 BB = EmitAtomicBinary(MI, BB, false, 0);
5024 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5025 BB = EmitAtomicBinary(MI, BB, true, 0);
5027 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5028 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5029 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5031 unsigned dest = MI->getOperand(0).getReg();
5032 unsigned ptrA = MI->getOperand(1).getReg();
5033 unsigned ptrB = MI->getOperand(2).getReg();
5034 unsigned oldval = MI->getOperand(3).getReg();
5035 unsigned newval = MI->getOperand(4).getReg();
5036 DebugLoc dl = MI->getDebugLoc();
5038 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5039 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5040 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5041 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5042 F->insert(It, loop1MBB);
5043 F->insert(It, loop2MBB);
5044 F->insert(It, midMBB);
5045 F->insert(It, exitMBB);
5046 exitMBB->splice(exitMBB->begin(), BB,
5047 llvm::next(MachineBasicBlock::iterator(MI)),
5049 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5053 // fallthrough --> loopMBB
5054 BB->addSuccessor(loop1MBB);
5057 // l[wd]arx dest, ptr
5058 // cmp[wd] dest, oldval
5061 // st[wd]cx. newval, ptr
5065 // st[wd]cx. dest, ptr
5068 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5069 .addReg(ptrA).addReg(ptrB);
5070 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5071 .addReg(oldval).addReg(dest);
5072 BuildMI(BB, dl, TII->get(PPC::BCC))
5073 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5074 BB->addSuccessor(loop2MBB);
5075 BB->addSuccessor(midMBB);
5078 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5079 .addReg(newval).addReg(ptrA).addReg(ptrB);
5080 BuildMI(BB, dl, TII->get(PPC::BCC))
5081 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5082 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5083 BB->addSuccessor(loop1MBB);
5084 BB->addSuccessor(exitMBB);
5087 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5088 .addReg(dest).addReg(ptrA).addReg(ptrB);
5089 BB->addSuccessor(exitMBB);
5094 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5095 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5096 // We must use 64-bit registers for addresses when targeting 64-bit,
5097 // since we're actually doing arithmetic on them. Other registers
5099 bool is64bit = PPCSubTarget.isPPC64();
5100 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5102 unsigned dest = MI->getOperand(0).getReg();
5103 unsigned ptrA = MI->getOperand(1).getReg();
5104 unsigned ptrB = MI->getOperand(2).getReg();
5105 unsigned oldval = MI->getOperand(3).getReg();
5106 unsigned newval = MI->getOperand(4).getReg();
5107 DebugLoc dl = MI->getDebugLoc();
5109 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5110 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5111 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5112 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5113 F->insert(It, loop1MBB);
5114 F->insert(It, loop2MBB);
5115 F->insert(It, midMBB);
5116 F->insert(It, exitMBB);
5117 exitMBB->splice(exitMBB->begin(), BB,
5118 llvm::next(MachineBasicBlock::iterator(MI)),
5120 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5122 MachineRegisterInfo &RegInfo = F->getRegInfo();
5123 const TargetRegisterClass *RC =
5124 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5125 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5126 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5127 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5128 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5129 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5130 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5131 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5132 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5133 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5134 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5135 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5136 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5137 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5138 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5140 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5141 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5144 // fallthrough --> loopMBB
5145 BB->addSuccessor(loop1MBB);
5147 // The 4-byte load must be aligned, while a char or short may be
5148 // anywhere in the word. Hence all this nasty bookkeeping code.
5149 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5150 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5151 // xori shift, shift1, 24 [16]
5152 // rlwinm ptr, ptr1, 0, 0, 29
5153 // slw newval2, newval, shift
5154 // slw oldval2, oldval,shift
5155 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5156 // slw mask, mask2, shift
5157 // and newval3, newval2, mask
5158 // and oldval3, oldval2, mask
5160 // lwarx tmpDest, ptr
5161 // and tmp, tmpDest, mask
5162 // cmpw tmp, oldval3
5165 // andc tmp2, tmpDest, mask
5166 // or tmp4, tmp2, newval3
5171 // stwcx. tmpDest, ptr
5173 // srw dest, tmpDest, shift
5174 if (ptrA != ZeroReg) {
5175 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5176 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5177 .addReg(ptrA).addReg(ptrB);
5181 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5182 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5183 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5184 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5186 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5187 .addReg(Ptr1Reg).addImm(0).addImm(61);
5189 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5190 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5191 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5192 .addReg(newval).addReg(ShiftReg);
5193 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5194 .addReg(oldval).addReg(ShiftReg);
5196 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5198 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5199 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5200 .addReg(Mask3Reg).addImm(65535);
5202 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5203 .addReg(Mask2Reg).addReg(ShiftReg);
5204 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5205 .addReg(NewVal2Reg).addReg(MaskReg);
5206 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5207 .addReg(OldVal2Reg).addReg(MaskReg);
5210 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5211 .addReg(ZeroReg).addReg(PtrReg);
5212 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5213 .addReg(TmpDestReg).addReg(MaskReg);
5214 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5215 .addReg(TmpReg).addReg(OldVal3Reg);
5216 BuildMI(BB, dl, TII->get(PPC::BCC))
5217 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5218 BB->addSuccessor(loop2MBB);
5219 BB->addSuccessor(midMBB);
5222 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5223 .addReg(TmpDestReg).addReg(MaskReg);
5224 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5225 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5226 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5227 .addReg(ZeroReg).addReg(PtrReg);
5228 BuildMI(BB, dl, TII->get(PPC::BCC))
5229 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5230 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5231 BB->addSuccessor(loop1MBB);
5232 BB->addSuccessor(exitMBB);
5235 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5236 .addReg(ZeroReg).addReg(PtrReg);
5237 BB->addSuccessor(exitMBB);
5242 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5245 llvm_unreachable("Unexpected instr type to insert");
5248 MI->eraseFromParent(); // The pseudo instruction is gone now.
5252 //===----------------------------------------------------------------------===//
5253 // Target Optimization Hooks
5254 //===----------------------------------------------------------------------===//
5256 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5257 DAGCombinerInfo &DCI) const {
5258 const TargetMachine &TM = getTargetMachine();
5259 SelectionDAG &DAG = DCI.DAG;
5260 DebugLoc dl = N->getDebugLoc();
5261 switch (N->getOpcode()) {
5264 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5265 if (C->isNullValue()) // 0 << V -> 0.
5266 return N->getOperand(0);
5270 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5271 if (C->isNullValue()) // 0 >>u V -> 0.
5272 return N->getOperand(0);
5276 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5277 if (C->isNullValue() || // 0 >>s V -> 0.
5278 C->isAllOnesValue()) // -1 >>s V -> -1.
5279 return N->getOperand(0);
5283 case ISD::SINT_TO_FP:
5284 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5285 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5286 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5287 // We allow the src/dst to be either f32/f64, but the intermediate
5288 // type must be i64.
5289 if (N->getOperand(0).getValueType() == MVT::i64 &&
5290 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5291 SDValue Val = N->getOperand(0).getOperand(0);
5292 if (Val.getValueType() == MVT::f32) {
5293 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5294 DCI.AddToWorklist(Val.getNode());
5297 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5298 DCI.AddToWorklist(Val.getNode());
5299 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5300 DCI.AddToWorklist(Val.getNode());
5301 if (N->getValueType(0) == MVT::f32) {
5302 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5303 DAG.getIntPtrConstant(0));
5304 DCI.AddToWorklist(Val.getNode());
5307 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5308 // If the intermediate type is i32, we can avoid the load/store here
5315 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5316 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5317 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5318 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5319 N->getOperand(1).getValueType() == MVT::i32 &&
5320 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5321 SDValue Val = N->getOperand(1).getOperand(0);
5322 if (Val.getValueType() == MVT::f32) {
5323 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5324 DCI.AddToWorklist(Val.getNode());
5326 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5327 DCI.AddToWorklist(Val.getNode());
5329 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5330 N->getOperand(2), N->getOperand(3));
5331 DCI.AddToWorklist(Val.getNode());
5335 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5336 if (cast<StoreSDNode>(N)->isUnindexed() &&
5337 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5338 N->getOperand(1).getNode()->hasOneUse() &&
5339 (N->getOperand(1).getValueType() == MVT::i32 ||
5340 N->getOperand(1).getValueType() == MVT::i16)) {
5341 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5342 // Do an any-extend to 32-bits if this is a half-word input.
5343 if (BSwapOp.getValueType() == MVT::i16)
5344 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5347 N->getOperand(0), BSwapOp, N->getOperand(2),
5348 DAG.getValueType(N->getOperand(1).getValueType())
5351 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5352 Ops, array_lengthof(Ops),
5353 cast<StoreSDNode>(N)->getMemoryVT(),
5354 cast<StoreSDNode>(N)->getMemOperand());
5358 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5359 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5360 N->getOperand(0).hasOneUse() &&
5361 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5362 SDValue Load = N->getOperand(0);
5363 LoadSDNode *LD = cast<LoadSDNode>(Load);
5364 // Create the byte-swapping load.
5366 LD->getChain(), // Chain
5367 LD->getBasePtr(), // Ptr
5368 DAG.getValueType(N->getValueType(0)) // VT
5371 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5372 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5373 LD->getMemoryVT(), LD->getMemOperand());
5375 // If this is an i16 load, insert the truncate.
5376 SDValue ResVal = BSLoad;
5377 if (N->getValueType(0) == MVT::i16)
5378 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5380 // First, combine the bswap away. This makes the value produced by the
5382 DCI.CombineTo(N, ResVal);
5384 // Next, combine the load away, we give it a bogus result value but a real
5385 // chain result. The result value is dead because the bswap is dead.
5386 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5388 // Return N so it doesn't get rechecked!
5389 return SDValue(N, 0);
5393 case PPCISD::VCMP: {
5394 // If a VCMPo node already exists with exactly the same operands as this
5395 // node, use its result instead of this node (VCMPo computes both a CR6 and
5396 // a normal output).
5398 if (!N->getOperand(0).hasOneUse() &&
5399 !N->getOperand(1).hasOneUse() &&
5400 !N->getOperand(2).hasOneUse()) {
5402 // Scan all of the users of the LHS, looking for VCMPo's that match.
5403 SDNode *VCMPoNode = 0;
5405 SDNode *LHSN = N->getOperand(0).getNode();
5406 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5408 if (UI->getOpcode() == PPCISD::VCMPo &&
5409 UI->getOperand(1) == N->getOperand(1) &&
5410 UI->getOperand(2) == N->getOperand(2) &&
5411 UI->getOperand(0) == N->getOperand(0)) {
5416 // If there is no VCMPo node, or if the flag value has a single use, don't
5418 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5421 // Look at the (necessarily single) use of the flag value. If it has a
5422 // chain, this transformation is more complex. Note that multiple things
5423 // could use the value result, which we should ignore.
5424 SDNode *FlagUser = 0;
5425 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5426 FlagUser == 0; ++UI) {
5427 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5429 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5430 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5437 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5438 // give up for right now.
5439 if (FlagUser->getOpcode() == PPCISD::MFCR)
5440 return SDValue(VCMPoNode, 0);
5445 // If this is a branch on an altivec predicate comparison, lower this so
5446 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5447 // lowering is done pre-legalize, because the legalizer lowers the predicate
5448 // compare down to code that is difficult to reassemble.
5449 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5450 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5454 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5455 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5456 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5457 assert(isDot && "Can't compare against a vector result!");
5459 // If this is a comparison against something other than 0/1, then we know
5460 // that the condition is never/always true.
5461 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5462 if (Val != 0 && Val != 1) {
5463 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5464 return N->getOperand(0);
5465 // Always !=, turn it into an unconditional branch.
5466 return DAG.getNode(ISD::BR, dl, MVT::Other,
5467 N->getOperand(0), N->getOperand(4));
5470 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5472 // Create the PPCISD altivec 'dot' comparison node.
5473 std::vector<EVT> VTs;
5475 LHS.getOperand(2), // LHS of compare
5476 LHS.getOperand(3), // RHS of compare
5477 DAG.getConstant(CompareOpc, MVT::i32)
5479 VTs.push_back(LHS.getOperand(2).getValueType());
5480 VTs.push_back(MVT::Glue);
5481 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5483 // Unpack the result based on how the target uses it.
5484 PPC::Predicate CompOpc;
5485 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5486 default: // Can't happen, don't crash on invalid number though.
5487 case 0: // Branch on the value of the EQ bit of CR6.
5488 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5490 case 1: // Branch on the inverted value of the EQ bit of CR6.
5491 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5493 case 2: // Branch on the value of the LT bit of CR6.
5494 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5496 case 3: // Branch on the inverted value of the LT bit of CR6.
5497 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5501 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5502 DAG.getConstant(CompOpc, MVT::i32),
5503 DAG.getRegister(PPC::CR6, MVT::i32),
5504 N->getOperand(4), CompNode.getValue(1));
5513 //===----------------------------------------------------------------------===//
5514 // Inline Assembly Support
5515 //===----------------------------------------------------------------------===//
5517 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5520 const SelectionDAG &DAG,
5521 unsigned Depth) const {
5522 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
5523 switch (Op.getOpcode()) {
5525 case PPCISD::LBRX: {
5526 // lhbrx is known to have the top bits cleared out.
5527 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5528 KnownZero = 0xFFFF0000;
5531 case ISD::INTRINSIC_WO_CHAIN: {
5532 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5534 case Intrinsic::ppc_altivec_vcmpbfp_p:
5535 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5536 case Intrinsic::ppc_altivec_vcmpequb_p:
5537 case Intrinsic::ppc_altivec_vcmpequh_p:
5538 case Intrinsic::ppc_altivec_vcmpequw_p:
5539 case Intrinsic::ppc_altivec_vcmpgefp_p:
5540 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5541 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5542 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5543 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5544 case Intrinsic::ppc_altivec_vcmpgtub_p:
5545 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5546 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5547 KnownZero = ~1U; // All bits but the low one are known to be zero.
5555 /// getConstraintType - Given a constraint, return the type of
5556 /// constraint it is for this target.
5557 PPCTargetLowering::ConstraintType
5558 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5559 if (Constraint.size() == 1) {
5560 switch (Constraint[0]) {
5567 return C_RegisterClass;
5570 return TargetLowering::getConstraintType(Constraint);
5573 /// Examine constraint type and operand type and determine a weight value.
5574 /// This object must already have been set up with the operand type
5575 /// and the current alternative constraint selected.
5576 TargetLowering::ConstraintWeight
5577 PPCTargetLowering::getSingleConstraintMatchWeight(
5578 AsmOperandInfo &info, const char *constraint) const {
5579 ConstraintWeight weight = CW_Invalid;
5580 Value *CallOperandVal = info.CallOperandVal;
5581 // If we don't have a value, we can't do a match,
5582 // but allow it at the lowest weight.
5583 if (CallOperandVal == NULL)
5585 Type *type = CallOperandVal->getType();
5586 // Look at the constraint type.
5587 switch (*constraint) {
5589 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5592 if (type->isIntegerTy())
5593 weight = CW_Register;
5596 if (type->isFloatTy())
5597 weight = CW_Register;
5600 if (type->isDoubleTy())
5601 weight = CW_Register;
5604 if (type->isVectorTy())
5605 weight = CW_Register;
5608 weight = CW_Register;
5614 std::pair<unsigned, const TargetRegisterClass*>
5615 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5617 if (Constraint.size() == 1) {
5618 // GCC RS6000 Constraint Letters
5619 switch (Constraint[0]) {
5622 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5623 return std::make_pair(0U, &PPC::G8RCRegClass);
5624 return std::make_pair(0U, &PPC::GPRCRegClass);
5627 return std::make_pair(0U, &PPC::F4RCRegClass);
5629 return std::make_pair(0U, &PPC::F8RCRegClass);
5632 return std::make_pair(0U, &PPC::VRRCRegClass);
5634 return std::make_pair(0U, &PPC::CRRCRegClass);
5638 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5642 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5643 /// vector. If it is invalid, don't add anything to Ops.
5644 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5645 std::string &Constraint,
5646 std::vector<SDValue>&Ops,
5647 SelectionDAG &DAG) const {
5648 SDValue Result(0,0);
5650 // Only support length 1 constraints.
5651 if (Constraint.length() > 1) return;
5653 char Letter = Constraint[0];
5664 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5665 if (!CST) return; // Must be an immediate to match.
5666 unsigned Value = CST->getZExtValue();
5668 default: llvm_unreachable("Unknown constraint letter!");
5669 case 'I': // "I" is a signed 16-bit constant.
5670 if ((short)Value == (int)Value)
5671 Result = DAG.getTargetConstant(Value, Op.getValueType());
5673 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5674 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5675 if ((short)Value == 0)
5676 Result = DAG.getTargetConstant(Value, Op.getValueType());
5678 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5679 if ((Value >> 16) == 0)
5680 Result = DAG.getTargetConstant(Value, Op.getValueType());
5682 case 'M': // "M" is a constant that is greater than 31.
5684 Result = DAG.getTargetConstant(Value, Op.getValueType());
5686 case 'N': // "N" is a positive constant that is an exact power of two.
5687 if ((int)Value > 0 && isPowerOf2_32(Value))
5688 Result = DAG.getTargetConstant(Value, Op.getValueType());
5690 case 'O': // "O" is the constant zero.
5692 Result = DAG.getTargetConstant(Value, Op.getValueType());
5694 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5695 if ((short)-Value == (int)-Value)
5696 Result = DAG.getTargetConstant(Value, Op.getValueType());
5703 if (Result.getNode()) {
5704 Ops.push_back(Result);
5708 // Handle standard constraint letters.
5709 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5712 // isLegalAddressingMode - Return true if the addressing mode represented
5713 // by AM is legal for this target, for a load/store of the specified type.
5714 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5716 // FIXME: PPC does not allow r+i addressing modes for vectors!
5718 // PPC allows a sign-extended 16-bit immediate field.
5719 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5722 // No global is ever allowed as a base.
5726 // PPC only support r+r,
5728 case 0: // "r+i" or just "i", depending on HasBaseReg.
5731 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5733 // Otherwise we have r+r or r+i.
5736 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5738 // Allow 2*r as r+r.
5741 // No other scales are supported.
5748 /// isLegalAddressImmediate - Return true if the integer value can be used
5749 /// as the offset of the target addressing mode for load / store of the
5751 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
5752 // PPC allows a sign-extended 16-bit immediate field.
5753 return (V > -(1 << 16) && V < (1 << 16)-1);
5756 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
5760 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5761 SelectionDAG &DAG) const {
5762 MachineFunction &MF = DAG.getMachineFunction();
5763 MachineFrameInfo *MFI = MF.getFrameInfo();
5764 MFI->setReturnAddressIsTaken(true);
5766 DebugLoc dl = Op.getDebugLoc();
5767 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5769 // Make sure the function does not optimize away the store of the RA to
5771 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5772 FuncInfo->setLRStoreRequired();
5773 bool isPPC64 = PPCSubTarget.isPPC64();
5774 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5777 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5780 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5781 isPPC64? MVT::i64 : MVT::i32);
5782 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5783 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5785 MachinePointerInfo(), false, false, false, 0);
5788 // Just load the return address off the stack.
5789 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5790 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5791 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
5794 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5795 SelectionDAG &DAG) const {
5796 DebugLoc dl = Op.getDebugLoc();
5797 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5799 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5800 bool isPPC64 = PtrVT == MVT::i64;
5802 MachineFunction &MF = DAG.getMachineFunction();
5803 MachineFrameInfo *MFI = MF.getFrameInfo();
5804 MFI->setFrameAddressIsTaken(true);
5805 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
5806 MFI->hasVarSizedObjects()) &&
5807 MFI->getStackSize() &&
5808 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5809 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5810 (is31 ? PPC::R31 : PPC::R1);
5811 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5814 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5815 FrameAddr, MachinePointerInfo(), false, false,
5821 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5822 // The PowerPC target isn't yet aware of offsets.
5826 /// getOptimalMemOpType - Returns the target specific optimal type for load
5827 /// and store operations as a result of memset, memcpy, and memmove
5828 /// lowering. If DstAlign is zero that means it's safe to destination
5829 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5830 /// means there isn't a need to check it against alignment requirement,
5831 /// probably because the source does not need to be loaded. If
5832 /// 'IsZeroVal' is true, that means it's safe to return a
5833 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5834 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5835 /// constant so it does not need to be loaded.
5836 /// It returns EVT::Other if the type should be determined using generic
5837 /// target-independent logic.
5838 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5839 unsigned DstAlign, unsigned SrcAlign,
5842 MachineFunction &MF) const {
5843 if (this->PPCSubTarget.isPPC64()) {
5850 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
5851 unsigned Directive = PPCSubTarget.getDarwinDirective();
5852 if (Directive == PPC::DIR_440 || Directive == PPC::DIR_A2)
5855 return TargetLowering::getSchedulingPreference(N);