1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCMachineFunctionInfo.h"
16 #include "PPCPerfectShuffle.h"
17 #include "PPCTargetMachine.h"
18 #include "MCTargetDesc/PPCPredicates.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
40 CCValAssign::LocInfo &LocInfo,
41 ISD::ArgFlagsTy &ArgFlags,
43 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
54 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
55 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
57 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
58 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
60 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
61 if (TM.getSubtargetImpl()->isDarwin())
62 return new TargetLoweringObjectFileMachO();
64 return new TargetLoweringObjectFileELF();
67 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
68 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
69 const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
73 // Use _setjmp/_longjmp instead of setjmp/longjmp.
74 setUseUnderscoreSetJmp(true);
75 setUseUnderscoreLongJmp(true);
77 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
78 // arguments are at least 4/8 bytes aligned.
79 bool isPPC64 = Subtarget->isPPC64();
80 setMinStackArgumentAlignment(isPPC64 ? 8:4);
82 // Set up the register classes.
83 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
84 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
85 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
87 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
88 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
91 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
93 // PowerPC has pre-inc load and store's.
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
96 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
97 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
98 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
101 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
102 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
103 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
105 // This is used in the ppcf128->int sequence. Note it has different semantics
106 // from FP_ROUND: that rounds to nearest, this rounds to zero.
107 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
109 // We do not currently implement these libm ops for PowerPC.
110 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
111 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
112 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
113 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
114 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
116 // PowerPC has no SREM/UREM instructions
117 setOperationAction(ISD::SREM, MVT::i32, Expand);
118 setOperationAction(ISD::UREM, MVT::i32, Expand);
119 setOperationAction(ISD::SREM, MVT::i64, Expand);
120 setOperationAction(ISD::UREM, MVT::i64, Expand);
122 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
123 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
124 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
125 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
126 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
127 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
128 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
129 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
130 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
132 // We don't support sin/cos/sqrt/fmod/pow
133 setOperationAction(ISD::FSIN , MVT::f64, Expand);
134 setOperationAction(ISD::FCOS , MVT::f64, Expand);
135 setOperationAction(ISD::FREM , MVT::f64, Expand);
136 setOperationAction(ISD::FPOW , MVT::f64, Expand);
137 setOperationAction(ISD::FMA , MVT::f64, Legal);
138 setOperationAction(ISD::FSIN , MVT::f32, Expand);
139 setOperationAction(ISD::FCOS , MVT::f32, Expand);
140 setOperationAction(ISD::FREM , MVT::f32, Expand);
141 setOperationAction(ISD::FPOW , MVT::f32, Expand);
142 setOperationAction(ISD::FMA , MVT::f32, Legal);
144 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
146 // If we're enabling GP optimizations, use hardware square root
147 if (!Subtarget->hasFSQRT()) {
148 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
149 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
152 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
153 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
155 // PowerPC does not have BSWAP, CTPOP or CTTZ
156 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
157 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
158 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
160 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
161 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
162 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
163 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
164 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
165 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
167 // PowerPC does not have ROTR
168 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
169 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
171 // PowerPC does not have Select
172 setOperationAction(ISD::SELECT, MVT::i32, Expand);
173 setOperationAction(ISD::SELECT, MVT::i64, Expand);
174 setOperationAction(ISD::SELECT, MVT::f32, Expand);
175 setOperationAction(ISD::SELECT, MVT::f64, Expand);
177 // PowerPC wants to turn select_cc of FP into fsel when possible.
178 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
179 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
181 // PowerPC wants to optimize integer setcc a bit
182 setOperationAction(ISD::SETCC, MVT::i32, Custom);
184 // PowerPC does not have BRCOND which requires SetCC
185 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
187 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
189 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
190 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
192 // PowerPC does not have [U|S]INT_TO_FP
193 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
194 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
196 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
197 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
198 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
199 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
201 // We cannot sextinreg(i1). Expand to shifts.
202 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
204 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
205 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
206 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
207 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
210 // We want to legalize GlobalAddress and ConstantPool nodes into the
211 // appropriate instructions to materialize the address.
212 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
213 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
214 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
215 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
216 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
217 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
218 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
219 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
220 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
221 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
224 setOperationAction(ISD::TRAP, MVT::Other, Legal);
226 // TRAMPOLINE is custom lowered.
227 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
228 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
230 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
231 setOperationAction(ISD::VASTART , MVT::Other, Custom);
233 if (Subtarget->isSVR4ABI()) {
235 // VAARG always uses double-word chunks, so promote anything smaller.
236 setOperationAction(ISD::VAARG, MVT::i1, Promote);
237 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
238 setOperationAction(ISD::VAARG, MVT::i8, Promote);
239 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
240 setOperationAction(ISD::VAARG, MVT::i16, Promote);
241 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
242 setOperationAction(ISD::VAARG, MVT::i32, Promote);
243 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
244 setOperationAction(ISD::VAARG, MVT::Other, Expand);
246 // VAARG is custom lowered with the 32-bit SVR4 ABI.
247 setOperationAction(ISD::VAARG, MVT::Other, Custom);
248 setOperationAction(ISD::VAARG, MVT::i64, Custom);
251 setOperationAction(ISD::VAARG, MVT::Other, Expand);
253 // Use the default implementation.
254 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
255 setOperationAction(ISD::VAEND , MVT::Other, Expand);
256 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
257 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
258 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
259 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
261 // We want to custom lower some of our intrinsics.
262 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
264 // Comparisons that require checking two conditions.
265 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
266 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
267 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
270 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
271 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
272 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
273 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
274 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
275 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
276 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
278 if (Subtarget->has64BitSupport()) {
279 // They also have instructions for converting between i64 and fp.
280 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
281 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
282 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
283 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
284 // This is just the low 32 bits of a (signed) fp->i64 conversion.
285 // We cannot do this with Promote because i64 is not a legal type.
286 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
288 // FIXME: disable this lowered code. This generates 64-bit register values,
289 // and we don't model the fact that the top part is clobbered by calls. We
290 // need to flag these together so that the value isn't live across a call.
291 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
293 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
294 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
297 if (Subtarget->use64BitRegs()) {
298 // 64-bit PowerPC implementations can support i64 types directly
299 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
300 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
301 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
302 // 64-bit PowerPC wants to expand i128 shifts itself.
303 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
304 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
305 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
307 // 32-bit PowerPC wants to expand i64 shifts itself.
308 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
309 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
310 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
313 if (Subtarget->hasAltivec()) {
314 // First set operation action for all vector types to expand. Then we
315 // will selectively turn on ones that can be effectively codegen'd.
316 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
317 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
318 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
320 // add/sub are legal for all supported vector VT's.
321 setOperationAction(ISD::ADD , VT, Legal);
322 setOperationAction(ISD::SUB , VT, Legal);
324 // We promote all shuffles to v16i8.
325 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
326 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
328 // We promote all non-typed operations to v4i32.
329 setOperationAction(ISD::AND , VT, Promote);
330 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
331 setOperationAction(ISD::OR , VT, Promote);
332 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
333 setOperationAction(ISD::XOR , VT, Promote);
334 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
335 setOperationAction(ISD::LOAD , VT, Promote);
336 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
337 setOperationAction(ISD::SELECT, VT, Promote);
338 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
339 setOperationAction(ISD::STORE, VT, Promote);
340 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
342 // No other operations are legal.
343 setOperationAction(ISD::MUL , VT, Expand);
344 setOperationAction(ISD::SDIV, VT, Expand);
345 setOperationAction(ISD::SREM, VT, Expand);
346 setOperationAction(ISD::UDIV, VT, Expand);
347 setOperationAction(ISD::UREM, VT, Expand);
348 setOperationAction(ISD::FDIV, VT, Expand);
349 setOperationAction(ISD::FNEG, VT, Expand);
350 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
351 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
352 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
353 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
354 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
355 setOperationAction(ISD::UDIVREM, VT, Expand);
356 setOperationAction(ISD::SDIVREM, VT, Expand);
357 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
358 setOperationAction(ISD::FPOW, VT, Expand);
359 setOperationAction(ISD::CTPOP, VT, Expand);
360 setOperationAction(ISD::CTLZ, VT, Expand);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
362 setOperationAction(ISD::CTTZ, VT, Expand);
363 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
366 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
367 // with merges, splats, etc.
368 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
370 setOperationAction(ISD::AND , MVT::v4i32, Legal);
371 setOperationAction(ISD::OR , MVT::v4i32, Legal);
372 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
373 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
374 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
375 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
376 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
377 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
378 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
379 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
381 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
382 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
383 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
384 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
386 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
387 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
388 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
389 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
390 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
392 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
393 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
395 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
396 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
397 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
398 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
401 if (Subtarget->has64BitSupport()) {
402 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
403 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
406 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
407 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
409 setBooleanContents(ZeroOrOneBooleanContent);
410 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
413 setStackPointerRegisterToSaveRestore(PPC::X1);
414 setExceptionPointerRegister(PPC::X3);
415 setExceptionSelectorRegister(PPC::X4);
417 setStackPointerRegisterToSaveRestore(PPC::R1);
418 setExceptionPointerRegister(PPC::R3);
419 setExceptionSelectorRegister(PPC::R4);
422 // We have target-specific dag combine patterns for the following nodes:
423 setTargetDAGCombine(ISD::SINT_TO_FP);
424 setTargetDAGCombine(ISD::STORE);
425 setTargetDAGCombine(ISD::BR_CC);
426 setTargetDAGCombine(ISD::BSWAP);
428 // Darwin long double math library functions have $LDBL128 appended.
429 if (Subtarget->isDarwin()) {
430 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
431 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
432 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
433 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
434 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
435 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
436 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
437 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
438 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
439 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
442 setMinFunctionAlignment(2);
443 if (PPCSubTarget.isDarwin())
444 setPrefFunctionAlignment(4);
446 if (isPPC64 && Subtarget->isJITCodeModel())
447 // Temporary workaround for the inability of PPC64 JIT to handle jump
449 setSupportJumpTables(false);
451 setInsertFencesForAtomic(true);
453 setSchedulingPreference(Sched::Hybrid);
455 computeRegisterProperties();
457 // The Freescale cores does better with aggressive inlining of memcpy and
458 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
459 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
460 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
461 maxStoresPerMemset = 32;
462 maxStoresPerMemsetOptSize = 16;
463 maxStoresPerMemcpy = 32;
464 maxStoresPerMemcpyOptSize = 8;
465 maxStoresPerMemmove = 32;
466 maxStoresPerMemmoveOptSize = 8;
468 setPrefFunctionAlignment(4);
469 benefitFromCodePlacementOpt = true;
473 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
474 /// function arguments in the caller parameter area.
475 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
476 const TargetMachine &TM = getTargetMachine();
477 // Darwin passes everything on 4 byte boundary.
478 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
481 // 16byte and wider vectors are passed on 16byte boundary.
482 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
483 if (VTy->getBitWidth() >= 128)
486 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
487 if (PPCSubTarget.isPPC64())
493 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
496 case PPCISD::FSEL: return "PPCISD::FSEL";
497 case PPCISD::FCFID: return "PPCISD::FCFID";
498 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
499 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
500 case PPCISD::STFIWX: return "PPCISD::STFIWX";
501 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
502 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
503 case PPCISD::VPERM: return "PPCISD::VPERM";
504 case PPCISD::Hi: return "PPCISD::Hi";
505 case PPCISD::Lo: return "PPCISD::Lo";
506 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
507 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
508 case PPCISD::LOAD: return "PPCISD::LOAD";
509 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
510 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
511 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
512 case PPCISD::SRL: return "PPCISD::SRL";
513 case PPCISD::SRA: return "PPCISD::SRA";
514 case PPCISD::SHL: return "PPCISD::SHL";
515 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
516 case PPCISD::STD_32: return "PPCISD::STD_32";
517 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
518 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
519 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
520 case PPCISD::NOP: return "PPCISD::NOP";
521 case PPCISD::MTCTR: return "PPCISD::MTCTR";
522 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
523 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
524 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
525 case PPCISD::MFCR: return "PPCISD::MFCR";
526 case PPCISD::VCMP: return "PPCISD::VCMP";
527 case PPCISD::VCMPo: return "PPCISD::VCMPo";
528 case PPCISD::LBRX: return "PPCISD::LBRX";
529 case PPCISD::STBRX: return "PPCISD::STBRX";
530 case PPCISD::LARX: return "PPCISD::LARX";
531 case PPCISD::STCX: return "PPCISD::STCX";
532 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
533 case PPCISD::MFFS: return "PPCISD::MFFS";
534 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
535 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
536 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
537 case PPCISD::MTFSF: return "PPCISD::MTFSF";
538 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
539 case PPCISD::CR6SET: return "PPCISD::CR6SET";
540 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
544 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
547 return VT.changeVectorElementTypeToInteger();
550 //===----------------------------------------------------------------------===//
551 // Node matching predicates, for use by the tblgen matching code.
552 //===----------------------------------------------------------------------===//
554 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
555 static bool isFloatingPointZero(SDValue Op) {
556 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
557 return CFP->getValueAPF().isZero();
558 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
559 // Maybe this has already been legalized into the constant pool?
560 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
561 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
562 return CFP->getValueAPF().isZero();
567 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
568 /// true if Op is undef or if it matches the specified value.
569 static bool isConstantOrUndef(int Op, int Val) {
570 return Op < 0 || Op == Val;
573 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
574 /// VPKUHUM instruction.
575 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
577 for (unsigned i = 0; i != 16; ++i)
578 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
581 for (unsigned i = 0; i != 8; ++i)
582 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
583 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
589 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
590 /// VPKUWUM instruction.
591 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
593 for (unsigned i = 0; i != 16; i += 2)
594 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
595 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
598 for (unsigned i = 0; i != 8; i += 2)
599 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
600 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
601 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
602 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
608 /// isVMerge - Common function, used to match vmrg* shuffles.
610 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
611 unsigned LHSStart, unsigned RHSStart) {
612 assert(N->getValueType(0) == MVT::v16i8 &&
613 "PPC only supports shuffles by bytes!");
614 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
615 "Unsupported merge size!");
617 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
618 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
619 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
620 LHSStart+j+i*UnitSize) ||
621 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
622 RHSStart+j+i*UnitSize))
628 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
629 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
630 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
633 return isVMerge(N, UnitSize, 8, 24);
634 return isVMerge(N, UnitSize, 8, 8);
637 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
638 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
639 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
642 return isVMerge(N, UnitSize, 0, 16);
643 return isVMerge(N, UnitSize, 0, 0);
647 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
648 /// amount, otherwise return -1.
649 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
650 assert(N->getValueType(0) == MVT::v16i8 &&
651 "PPC only supports shuffles by bytes!");
653 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
655 // Find the first non-undef value in the shuffle mask.
657 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
660 if (i == 16) return -1; // all undef.
662 // Otherwise, check to see if the rest of the elements are consecutively
663 // numbered from this value.
664 unsigned ShiftAmt = SVOp->getMaskElt(i);
665 if (ShiftAmt < i) return -1;
669 // Check the rest of the elements to see if they are consecutive.
670 for (++i; i != 16; ++i)
671 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
674 // Check the rest of the elements to see if they are consecutive.
675 for (++i; i != 16; ++i)
676 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
682 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
683 /// specifies a splat of a single element that is suitable for input to
684 /// VSPLTB/VSPLTH/VSPLTW.
685 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
686 assert(N->getValueType(0) == MVT::v16i8 &&
687 (EltSize == 1 || EltSize == 2 || EltSize == 4));
689 // This is a splat operation if each element of the permute is the same, and
690 // if the value doesn't reference the second vector.
691 unsigned ElementBase = N->getMaskElt(0);
693 // FIXME: Handle UNDEF elements too!
694 if (ElementBase >= 16)
697 // Check that the indices are consecutive, in the case of a multi-byte element
698 // splatted with a v16i8 mask.
699 for (unsigned i = 1; i != EltSize; ++i)
700 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
703 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
704 if (N->getMaskElt(i) < 0) continue;
705 for (unsigned j = 0; j != EltSize; ++j)
706 if (N->getMaskElt(i+j) != N->getMaskElt(j))
712 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
714 bool PPC::isAllNegativeZeroVector(SDNode *N) {
715 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
717 APInt APVal, APUndef;
721 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
722 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
723 return CFP->getValueAPF().isNegZero();
728 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
729 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
730 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
731 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
732 assert(isSplatShuffleMask(SVOp, EltSize));
733 return SVOp->getMaskElt(0) / EltSize;
736 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
737 /// by using a vspltis[bhw] instruction of the specified element size, return
738 /// the constant being splatted. The ByteSize field indicates the number of
739 /// bytes of each element [124] -> [bhw].
740 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
743 // If ByteSize of the splat is bigger than the element size of the
744 // build_vector, then we have a case where we are checking for a splat where
745 // multiple elements of the buildvector are folded together into a single
746 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
747 unsigned EltSize = 16/N->getNumOperands();
748 if (EltSize < ByteSize) {
749 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
750 SDValue UniquedVals[4];
751 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
753 // See if all of the elements in the buildvector agree across.
754 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
755 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
756 // If the element isn't a constant, bail fully out.
757 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
760 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
761 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
762 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
763 return SDValue(); // no match.
766 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
767 // either constant or undef values that are identical for each chunk. See
768 // if these chunks can form into a larger vspltis*.
770 // Check to see if all of the leading entries are either 0 or -1. If
771 // neither, then this won't fit into the immediate field.
772 bool LeadingZero = true;
773 bool LeadingOnes = true;
774 for (unsigned i = 0; i != Multiple-1; ++i) {
775 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
777 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
778 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
780 // Finally, check the least significant entry.
782 if (UniquedVals[Multiple-1].getNode() == 0)
783 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
784 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
786 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
789 if (UniquedVals[Multiple-1].getNode() == 0)
790 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
791 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
792 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
793 return DAG.getTargetConstant(Val, MVT::i32);
799 // Check to see if this buildvec has a single non-undef value in its elements.
800 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
801 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
802 if (OpVal.getNode() == 0)
803 OpVal = N->getOperand(i);
804 else if (OpVal != N->getOperand(i))
808 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
810 unsigned ValSizeInBytes = EltSize;
812 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
813 Value = CN->getZExtValue();
814 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
815 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
816 Value = FloatToBits(CN->getValueAPF().convertToFloat());
819 // If the splat value is larger than the element value, then we can never do
820 // this splat. The only case that we could fit the replicated bits into our
821 // immediate field for would be zero, and we prefer to use vxor for it.
822 if (ValSizeInBytes < ByteSize) return SDValue();
824 // If the element value is larger than the splat value, cut it in half and
825 // check to see if the two halves are equal. Continue doing this until we
826 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
827 while (ValSizeInBytes > ByteSize) {
828 ValSizeInBytes >>= 1;
830 // If the top half equals the bottom half, we're still ok.
831 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
832 (Value & ((1 << (8*ValSizeInBytes))-1)))
836 // Properly sign extend the value.
837 int MaskVal = SignExtend32(Value, ByteSize * 8);
839 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
840 if (MaskVal == 0) return SDValue();
842 // Finally, if this value fits in a 5 bit sext field, return it
843 if (SignExtend32<5>(MaskVal) == MaskVal)
844 return DAG.getTargetConstant(MaskVal, MVT::i32);
848 //===----------------------------------------------------------------------===//
849 // Addressing Mode Selection
850 //===----------------------------------------------------------------------===//
852 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
853 /// or 64-bit immediate, and if the value can be accurately represented as a
854 /// sign extension from a 16-bit value. If so, this returns true and the
856 static bool isIntS16Immediate(SDNode *N, short &Imm) {
857 if (N->getOpcode() != ISD::Constant)
860 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
861 if (N->getValueType(0) == MVT::i32)
862 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
864 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
866 static bool isIntS16Immediate(SDValue Op, short &Imm) {
867 return isIntS16Immediate(Op.getNode(), Imm);
871 /// SelectAddressRegReg - Given the specified addressed, check to see if it
872 /// can be represented as an indexed [r+r] operation. Returns false if it
873 /// can be more efficiently represented with [r+imm].
874 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
876 SelectionDAG &DAG) const {
878 if (N.getOpcode() == ISD::ADD) {
879 if (isIntS16Immediate(N.getOperand(1), imm))
881 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
884 Base = N.getOperand(0);
885 Index = N.getOperand(1);
887 } else if (N.getOpcode() == ISD::OR) {
888 if (isIntS16Immediate(N.getOperand(1), imm))
889 return false; // r+i can fold it if we can.
891 // If this is an or of disjoint bitfields, we can codegen this as an add
892 // (for better address arithmetic) if the LHS and RHS of the OR are provably
894 APInt LHSKnownZero, LHSKnownOne;
895 APInt RHSKnownZero, RHSKnownOne;
896 DAG.ComputeMaskedBits(N.getOperand(0),
897 LHSKnownZero, LHSKnownOne);
899 if (LHSKnownZero.getBoolValue()) {
900 DAG.ComputeMaskedBits(N.getOperand(1),
901 RHSKnownZero, RHSKnownOne);
902 // If all of the bits are known zero on the LHS or RHS, the add won't
904 if (~(LHSKnownZero | RHSKnownZero) == 0) {
905 Base = N.getOperand(0);
906 Index = N.getOperand(1);
915 /// Returns true if the address N can be represented by a base register plus
916 /// a signed 16-bit displacement [r+imm], and if it is not better
917 /// represented as reg+reg.
918 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
920 SelectionDAG &DAG) const {
921 // FIXME dl should come from parent load or store, not from address
922 DebugLoc dl = N.getDebugLoc();
923 // If this can be more profitably realized as r+r, fail.
924 if (SelectAddressRegReg(N, Disp, Base, DAG))
927 if (N.getOpcode() == ISD::ADD) {
929 if (isIntS16Immediate(N.getOperand(1), imm)) {
930 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
931 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
932 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
934 Base = N.getOperand(0);
936 return true; // [r+i]
937 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
938 // Match LOAD (ADD (X, Lo(G))).
939 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
940 && "Cannot handle constant offsets yet!");
941 Disp = N.getOperand(1).getOperand(0); // The global address.
942 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
943 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
944 Disp.getOpcode() == ISD::TargetConstantPool ||
945 Disp.getOpcode() == ISD::TargetJumpTable);
946 Base = N.getOperand(0);
947 return true; // [&g+r]
949 } else if (N.getOpcode() == ISD::OR) {
951 if (isIntS16Immediate(N.getOperand(1), imm)) {
952 // If this is an or of disjoint bitfields, we can codegen this as an add
953 // (for better address arithmetic) if the LHS and RHS of the OR are
954 // provably disjoint.
955 APInt LHSKnownZero, LHSKnownOne;
956 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
958 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
959 // If all of the bits are known zero on the LHS or RHS, the add won't
961 Base = N.getOperand(0);
962 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
966 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
967 // Loading from a constant address.
969 // If this address fits entirely in a 16-bit sext immediate field, codegen
972 if (isIntS16Immediate(CN, Imm)) {
973 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
974 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
975 CN->getValueType(0));
979 // Handle 32-bit sext immediates with LIS + addr mode.
980 if (CN->getValueType(0) == MVT::i32 ||
981 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
982 int Addr = (int)CN->getZExtValue();
984 // Otherwise, break this down into an LIS + disp.
985 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
987 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
988 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
989 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
994 Disp = DAG.getTargetConstant(0, getPointerTy());
995 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
996 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
999 return true; // [r+0]
1002 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1003 /// represented as an indexed [r+r] operation.
1004 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1006 SelectionDAG &DAG) const {
1007 // Check to see if we can easily represent this as an [r+r] address. This
1008 // will fail if it thinks that the address is more profitably represented as
1009 // reg+imm, e.g. where imm = 0.
1010 if (SelectAddressRegReg(N, Base, Index, DAG))
1013 // If the operand is an addition, always emit this as [r+r], since this is
1014 // better (for code size, and execution, as the memop does the add for free)
1015 // than emitting an explicit add.
1016 if (N.getOpcode() == ISD::ADD) {
1017 Base = N.getOperand(0);
1018 Index = N.getOperand(1);
1022 // Otherwise, do it the hard way, using R0 as the base register.
1023 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1029 /// SelectAddressRegImmShift - Returns true if the address N can be
1030 /// represented by a base register plus a signed 14-bit displacement
1031 /// [r+imm*4]. Suitable for use by STD and friends.
1032 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1034 SelectionDAG &DAG) const {
1035 // FIXME dl should come from the parent load or store, not the address
1036 DebugLoc dl = N.getDebugLoc();
1037 // If this can be more profitably realized as r+r, fail.
1038 if (SelectAddressRegReg(N, Disp, Base, DAG))
1041 if (N.getOpcode() == ISD::ADD) {
1043 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1044 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1045 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1046 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1048 Base = N.getOperand(0);
1050 return true; // [r+i]
1051 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1052 // Match LOAD (ADD (X, Lo(G))).
1053 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1054 && "Cannot handle constant offsets yet!");
1055 Disp = N.getOperand(1).getOperand(0); // The global address.
1056 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1057 Disp.getOpcode() == ISD::TargetConstantPool ||
1058 Disp.getOpcode() == ISD::TargetJumpTable);
1059 Base = N.getOperand(0);
1060 return true; // [&g+r]
1062 } else if (N.getOpcode() == ISD::OR) {
1064 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1065 // If this is an or of disjoint bitfields, we can codegen this as an add
1066 // (for better address arithmetic) if the LHS and RHS of the OR are
1067 // provably disjoint.
1068 APInt LHSKnownZero, LHSKnownOne;
1069 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1070 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1071 // If all of the bits are known zero on the LHS or RHS, the add won't
1073 Base = N.getOperand(0);
1074 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1078 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1079 // Loading from a constant address. Verify low two bits are clear.
1080 if ((CN->getZExtValue() & 3) == 0) {
1081 // If this address fits entirely in a 14-bit sext immediate field, codegen
1084 if (isIntS16Immediate(CN, Imm)) {
1085 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1086 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1087 CN->getValueType(0));
1091 // Fold the low-part of 32-bit absolute addresses into addr mode.
1092 if (CN->getValueType(0) == MVT::i32 ||
1093 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1094 int Addr = (int)CN->getZExtValue();
1096 // Otherwise, break this down into an LIS + disp.
1097 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1098 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1099 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1100 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1106 Disp = DAG.getTargetConstant(0, getPointerTy());
1107 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1108 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1111 return true; // [r+0]
1115 /// getPreIndexedAddressParts - returns true by value, base pointer and
1116 /// offset pointer and addressing mode by reference if the node's address
1117 /// can be legally represented as pre-indexed load / store address.
1118 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1120 ISD::MemIndexedMode &AM,
1121 SelectionDAG &DAG) const {
1122 if (DisablePPCPreinc) return false;
1126 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1127 Ptr = LD->getBasePtr();
1128 VT = LD->getMemoryVT();
1130 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1131 Ptr = ST->getBasePtr();
1132 VT = ST->getMemoryVT();
1136 // PowerPC doesn't have preinc load/store instructions for vectors.
1140 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1145 // LDU/STU use reg+imm*4, others use reg+imm.
1146 if (VT != MVT::i64) {
1148 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1152 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1156 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1157 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1158 // sext i32 to i64 when addr mode is r+i.
1159 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1160 LD->getExtensionType() == ISD::SEXTLOAD &&
1161 isa<ConstantSDNode>(Offset))
1169 //===----------------------------------------------------------------------===//
1170 // LowerOperation implementation
1171 //===----------------------------------------------------------------------===//
1173 /// GetLabelAccessInfo - Return true if we should reference labels using a
1174 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1175 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1176 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1177 HiOpFlags = PPCII::MO_HA16;
1178 LoOpFlags = PPCII::MO_LO16;
1180 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1181 // non-darwin platform. We don't support PIC on other platforms yet.
1182 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1183 TM.getSubtarget<PPCSubtarget>().isDarwin();
1185 HiOpFlags |= PPCII::MO_PIC_FLAG;
1186 LoOpFlags |= PPCII::MO_PIC_FLAG;
1189 // If this is a reference to a global value that requires a non-lazy-ptr, make
1190 // sure that instruction lowering adds it.
1191 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1192 HiOpFlags |= PPCII::MO_NLP_FLAG;
1193 LoOpFlags |= PPCII::MO_NLP_FLAG;
1195 if (GV->hasHiddenVisibility()) {
1196 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1197 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1204 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1205 SelectionDAG &DAG) {
1206 EVT PtrVT = HiPart.getValueType();
1207 SDValue Zero = DAG.getConstant(0, PtrVT);
1208 DebugLoc DL = HiPart.getDebugLoc();
1210 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1211 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1213 // With PIC, the first instruction is actually "GR+hi(&G)".
1215 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1216 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1218 // Generate non-pic code that has direct accesses to the constant pool.
1219 // The address of the global is just (hi(&g)+lo(&g)).
1220 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1223 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1224 SelectionDAG &DAG) const {
1225 EVT PtrVT = Op.getValueType();
1226 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1227 const Constant *C = CP->getConstVal();
1229 // 64-bit SVR4 ABI code is always position-independent.
1230 // The actual address of the GlobalValue is stored in the TOC.
1231 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1232 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1233 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1234 DAG.getRegister(PPC::X2, MVT::i64));
1237 unsigned MOHiFlag, MOLoFlag;
1238 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1240 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1242 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1243 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1246 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1247 EVT PtrVT = Op.getValueType();
1248 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1250 // 64-bit SVR4 ABI code is always position-independent.
1251 // The actual address of the GlobalValue is stored in the TOC.
1252 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1253 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1254 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1255 DAG.getRegister(PPC::X2, MVT::i64));
1258 unsigned MOHiFlag, MOLoFlag;
1259 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1260 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1261 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1262 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1265 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1266 SelectionDAG &DAG) const {
1267 EVT PtrVT = Op.getValueType();
1269 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1271 unsigned MOHiFlag, MOLoFlag;
1272 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1273 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1274 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1275 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1278 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1279 SelectionDAG &DAG) const {
1281 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1282 DebugLoc dl = GA->getDebugLoc();
1283 const GlobalValue *GV = GA->getGlobal();
1284 EVT PtrVT = getPointerTy();
1285 bool is64bit = PPCSubTarget.isPPC64();
1287 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1289 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1290 PPCII::MO_TPREL16_HA);
1291 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1292 PPCII::MO_TPREL16_LO);
1294 if (model != TLSModel::LocalExec)
1295 llvm_unreachable("only local-exec TLS mode supported");
1296 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1297 is64bit ? MVT::i64 : MVT::i32);
1298 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1299 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1302 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1303 SelectionDAG &DAG) const {
1304 EVT PtrVT = Op.getValueType();
1305 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1306 DebugLoc DL = GSDN->getDebugLoc();
1307 const GlobalValue *GV = GSDN->getGlobal();
1309 // 64-bit SVR4 ABI code is always position-independent.
1310 // The actual address of the GlobalValue is stored in the TOC.
1311 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1312 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1313 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1314 DAG.getRegister(PPC::X2, MVT::i64));
1317 unsigned MOHiFlag, MOLoFlag;
1318 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1321 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1323 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1325 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1327 // If the global reference is actually to a non-lazy-pointer, we have to do an
1328 // extra load to get the address of the global.
1329 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1330 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1331 false, false, false, 0);
1335 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1336 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1337 DebugLoc dl = Op.getDebugLoc();
1339 // If we're comparing for equality to zero, expose the fact that this is
1340 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1341 // fold the new nodes.
1342 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1343 if (C->isNullValue() && CC == ISD::SETEQ) {
1344 EVT VT = Op.getOperand(0).getValueType();
1345 SDValue Zext = Op.getOperand(0);
1346 if (VT.bitsLT(MVT::i32)) {
1348 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1350 unsigned Log2b = Log2_32(VT.getSizeInBits());
1351 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1352 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1353 DAG.getConstant(Log2b, MVT::i32));
1354 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1356 // Leave comparisons against 0 and -1 alone for now, since they're usually
1357 // optimized. FIXME: revisit this when we can custom lower all setcc
1359 if (C->isAllOnesValue() || C->isNullValue())
1363 // If we have an integer seteq/setne, turn it into a compare against zero
1364 // by xor'ing the rhs with the lhs, which is faster than setting a
1365 // condition register, reading it back out, and masking the correct bit. The
1366 // normal approach here uses sub to do this instead of xor. Using xor exposes
1367 // the result to other bit-twiddling opportunities.
1368 EVT LHSVT = Op.getOperand(0).getValueType();
1369 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1370 EVT VT = Op.getValueType();
1371 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1373 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1378 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1379 const PPCSubtarget &Subtarget) const {
1380 SDNode *Node = Op.getNode();
1381 EVT VT = Node->getValueType(0);
1382 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1383 SDValue InChain = Node->getOperand(0);
1384 SDValue VAListPtr = Node->getOperand(1);
1385 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1386 DebugLoc dl = Node->getDebugLoc();
1388 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1391 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1392 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1394 InChain = GprIndex.getValue(1);
1396 if (VT == MVT::i64) {
1397 // Check if GprIndex is even
1398 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1399 DAG.getConstant(1, MVT::i32));
1400 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1401 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1402 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1403 DAG.getConstant(1, MVT::i32));
1404 // Align GprIndex to be even if it isn't
1405 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1409 // fpr index is 1 byte after gpr
1410 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1411 DAG.getConstant(1, MVT::i32));
1414 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1415 FprPtr, MachinePointerInfo(SV), MVT::i8,
1417 InChain = FprIndex.getValue(1);
1419 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1420 DAG.getConstant(8, MVT::i32));
1422 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1423 DAG.getConstant(4, MVT::i32));
1426 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1427 MachinePointerInfo(), false, false,
1429 InChain = OverflowArea.getValue(1);
1431 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1432 MachinePointerInfo(), false, false,
1434 InChain = RegSaveArea.getValue(1);
1436 // select overflow_area if index > 8
1437 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1438 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1440 // adjustment constant gpr_index * 4/8
1441 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1442 VT.isInteger() ? GprIndex : FprIndex,
1443 DAG.getConstant(VT.isInteger() ? 4 : 8,
1446 // OurReg = RegSaveArea + RegConstant
1447 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1450 // Floating types are 32 bytes into RegSaveArea
1451 if (VT.isFloatingPoint())
1452 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1453 DAG.getConstant(32, MVT::i32));
1455 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1456 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1457 VT.isInteger() ? GprIndex : FprIndex,
1458 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1461 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1462 VT.isInteger() ? VAListPtr : FprPtr,
1463 MachinePointerInfo(SV),
1464 MVT::i8, false, false, 0);
1466 // determine if we should load from reg_save_area or overflow_area
1467 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1469 // increase overflow_area by 4/8 if gpr/fpr > 8
1470 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1471 DAG.getConstant(VT.isInteger() ? 4 : 8,
1474 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1477 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1479 MachinePointerInfo(),
1480 MVT::i32, false, false, 0);
1482 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1483 false, false, false, 0);
1486 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1487 SelectionDAG &DAG) const {
1488 return Op.getOperand(0);
1491 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1492 SelectionDAG &DAG) const {
1493 SDValue Chain = Op.getOperand(0);
1494 SDValue Trmp = Op.getOperand(1); // trampoline
1495 SDValue FPtr = Op.getOperand(2); // nested function
1496 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1497 DebugLoc dl = Op.getDebugLoc();
1499 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1500 bool isPPC64 = (PtrVT == MVT::i64);
1502 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1505 TargetLowering::ArgListTy Args;
1506 TargetLowering::ArgListEntry Entry;
1508 Entry.Ty = IntPtrTy;
1509 Entry.Node = Trmp; Args.push_back(Entry);
1511 // TrampSize == (isPPC64 ? 48 : 40);
1512 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1513 isPPC64 ? MVT::i64 : MVT::i32);
1514 Args.push_back(Entry);
1516 Entry.Node = FPtr; Args.push_back(Entry);
1517 Entry.Node = Nest; Args.push_back(Entry);
1519 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1520 TargetLowering::CallLoweringInfo CLI(Chain,
1521 Type::getVoidTy(*DAG.getContext()),
1522 false, false, false, false, 0,
1524 /*isTailCall=*/false,
1525 /*doesNotRet=*/false,
1526 /*isReturnValueUsed=*/true,
1527 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1529 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1531 return CallResult.second;
1534 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1535 const PPCSubtarget &Subtarget) const {
1536 MachineFunction &MF = DAG.getMachineFunction();
1537 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1539 DebugLoc dl = Op.getDebugLoc();
1541 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1542 // vastart just stores the address of the VarArgsFrameIndex slot into the
1543 // memory location argument.
1544 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1545 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1546 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1547 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1548 MachinePointerInfo(SV),
1552 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1553 // We suppose the given va_list is already allocated.
1556 // char gpr; /* index into the array of 8 GPRs
1557 // * stored in the register save area
1558 // * gpr=0 corresponds to r3,
1559 // * gpr=1 to r4, etc.
1561 // char fpr; /* index into the array of 8 FPRs
1562 // * stored in the register save area
1563 // * fpr=0 corresponds to f1,
1564 // * fpr=1 to f2, etc.
1566 // char *overflow_arg_area;
1567 // /* location on stack that holds
1568 // * the next overflow argument
1570 // char *reg_save_area;
1571 // /* where r3:r10 and f1:f8 (if saved)
1577 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1578 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1581 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1583 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1585 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1588 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1589 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1591 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1592 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1594 uint64_t FPROffset = 1;
1595 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1597 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1599 // Store first byte : number of int regs
1600 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1602 MachinePointerInfo(SV),
1603 MVT::i8, false, false, 0);
1604 uint64_t nextOffset = FPROffset;
1605 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1608 // Store second byte : number of float regs
1609 SDValue secondStore =
1610 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1611 MachinePointerInfo(SV, nextOffset), MVT::i8,
1613 nextOffset += StackOffset;
1614 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1616 // Store second word : arguments given on stack
1617 SDValue thirdStore =
1618 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1619 MachinePointerInfo(SV, nextOffset),
1621 nextOffset += FrameOffset;
1622 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1624 // Store third word : arguments given in registers
1625 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1626 MachinePointerInfo(SV, nextOffset),
1631 #include "PPCGenCallingConv.inc"
1633 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1634 CCValAssign::LocInfo &LocInfo,
1635 ISD::ArgFlagsTy &ArgFlags,
1640 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1642 CCValAssign::LocInfo &LocInfo,
1643 ISD::ArgFlagsTy &ArgFlags,
1645 static const uint16_t ArgRegs[] = {
1646 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1647 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1649 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1651 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1653 // Skip one register if the first unallocated register has an even register
1654 // number and there are still argument registers available which have not been
1655 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1656 // need to skip a register if RegNum is odd.
1657 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1658 State.AllocateReg(ArgRegs[RegNum]);
1661 // Always return false here, as this function only makes sure that the first
1662 // unallocated register has an odd register number and does not actually
1663 // allocate a register for the current argument.
1667 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1669 CCValAssign::LocInfo &LocInfo,
1670 ISD::ArgFlagsTy &ArgFlags,
1672 static const uint16_t ArgRegs[] = {
1673 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1677 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1679 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1681 // If there is only one Floating-point register left we need to put both f64
1682 // values of a split ppc_fp128 value on the stack.
1683 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1684 State.AllocateReg(ArgRegs[RegNum]);
1687 // Always return false here, as this function only makes sure that the two f64
1688 // values a ppc_fp128 value is split into are both passed in registers or both
1689 // passed on the stack and does not actually allocate a register for the
1690 // current argument.
1694 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1696 static const uint16_t *GetFPR() {
1697 static const uint16_t FPR[] = {
1698 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1699 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1705 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1707 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1708 unsigned PtrByteSize) {
1709 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1710 if (Flags.isByVal())
1711 ArgSize = Flags.getByValSize();
1712 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1718 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1719 CallingConv::ID CallConv, bool isVarArg,
1720 const SmallVectorImpl<ISD::InputArg>
1722 DebugLoc dl, SelectionDAG &DAG,
1723 SmallVectorImpl<SDValue> &InVals)
1725 if (PPCSubTarget.isSVR4ABI()) {
1726 if (PPCSubTarget.isPPC64())
1727 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1730 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1733 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1739 PPCTargetLowering::LowerFormalArguments_32SVR4(
1741 CallingConv::ID CallConv, bool isVarArg,
1742 const SmallVectorImpl<ISD::InputArg>
1744 DebugLoc dl, SelectionDAG &DAG,
1745 SmallVectorImpl<SDValue> &InVals) const {
1747 // 32-bit SVR4 ABI Stack Frame Layout:
1748 // +-----------------------------------+
1749 // +--> | Back chain |
1750 // | +-----------------------------------+
1751 // | | Floating-point register save area |
1752 // | +-----------------------------------+
1753 // | | General register save area |
1754 // | +-----------------------------------+
1755 // | | CR save word |
1756 // | +-----------------------------------+
1757 // | | VRSAVE save word |
1758 // | +-----------------------------------+
1759 // | | Alignment padding |
1760 // | +-----------------------------------+
1761 // | | Vector register save area |
1762 // | +-----------------------------------+
1763 // | | Local variable space |
1764 // | +-----------------------------------+
1765 // | | Parameter list area |
1766 // | +-----------------------------------+
1767 // | | LR save word |
1768 // | +-----------------------------------+
1769 // SP--> +--- | Back chain |
1770 // +-----------------------------------+
1773 // System V Application Binary Interface PowerPC Processor Supplement
1774 // AltiVec Technology Programming Interface Manual
1776 MachineFunction &MF = DAG.getMachineFunction();
1777 MachineFrameInfo *MFI = MF.getFrameInfo();
1778 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1780 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1781 // Potential tail calls could cause overwriting of argument stack slots.
1782 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1783 (CallConv == CallingConv::Fast));
1784 unsigned PtrByteSize = 4;
1786 // Assign locations to all of the incoming arguments.
1787 SmallVector<CCValAssign, 16> ArgLocs;
1788 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1789 getTargetMachine(), ArgLocs, *DAG.getContext());
1791 // Reserve space for the linkage area on the stack.
1792 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1794 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1796 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1797 CCValAssign &VA = ArgLocs[i];
1799 // Arguments stored in registers.
1800 if (VA.isRegLoc()) {
1801 const TargetRegisterClass *RC;
1802 EVT ValVT = VA.getValVT();
1804 switch (ValVT.getSimpleVT().SimpleTy) {
1806 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1808 RC = &PPC::GPRCRegClass;
1811 RC = &PPC::F4RCRegClass;
1814 RC = &PPC::F8RCRegClass;
1820 RC = &PPC::VRRCRegClass;
1824 // Transform the arguments stored in physical registers into virtual ones.
1825 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1826 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1828 InVals.push_back(ArgValue);
1830 // Argument stored in memory.
1831 assert(VA.isMemLoc());
1833 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1834 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1837 // Create load nodes to retrieve arguments from the stack.
1838 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1839 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1840 MachinePointerInfo(),
1841 false, false, false, 0));
1845 // Assign locations to all of the incoming aggregate by value arguments.
1846 // Aggregates passed by value are stored in the local variable space of the
1847 // caller's stack frame, right above the parameter list area.
1848 SmallVector<CCValAssign, 16> ByValArgLocs;
1849 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1850 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1852 // Reserve stack space for the allocations in CCInfo.
1853 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1855 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1857 // Area that is at least reserved in the caller of this function.
1858 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1860 // Set the size that is at least reserved in caller of this function. Tail
1861 // call optimized function's reserved stack space needs to be aligned so that
1862 // taking the difference between two stack areas will result in an aligned
1864 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1867 std::max(MinReservedArea,
1868 PPCFrameLowering::getMinCallFrameSize(false, false));
1870 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1871 getStackAlignment();
1872 unsigned AlignMask = TargetAlign-1;
1873 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1875 FI->setMinReservedArea(MinReservedArea);
1877 SmallVector<SDValue, 8> MemOps;
1879 // If the function takes variable number of arguments, make a frame index for
1880 // the start of the first vararg value... for expansion of llvm.va_start.
1882 static const uint16_t GPArgRegs[] = {
1883 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1884 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1886 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1888 static const uint16_t FPArgRegs[] = {
1889 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1892 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1894 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1896 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1899 // Make room for NumGPArgRegs and NumFPArgRegs.
1900 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1901 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1903 FuncInfo->setVarArgsStackOffset(
1904 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1905 CCInfo.getNextStackOffset(), true));
1907 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1908 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1910 // The fixed integer arguments of a variadic function are stored to the
1911 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1912 // the result of va_next.
1913 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1914 // Get an existing live-in vreg, or add a new one.
1915 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1917 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1919 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1920 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1921 MachinePointerInfo(), false, false, 0);
1922 MemOps.push_back(Store);
1923 // Increment the address by four for the next argument to store
1924 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1925 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1928 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1930 // The double arguments are stored to the VarArgsFrameIndex
1932 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1933 // Get an existing live-in vreg, or add a new one.
1934 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1936 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1938 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1939 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1940 MachinePointerInfo(), false, false, 0);
1941 MemOps.push_back(Store);
1942 // Increment the address by eight for the next argument to store
1943 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1945 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1949 if (!MemOps.empty())
1950 Chain = DAG.getNode(ISD::TokenFactor, dl,
1951 MVT::Other, &MemOps[0], MemOps.size());
1957 PPCTargetLowering::LowerFormalArguments_64SVR4(
1959 CallingConv::ID CallConv, bool isVarArg,
1960 const SmallVectorImpl<ISD::InputArg>
1962 DebugLoc dl, SelectionDAG &DAG,
1963 SmallVectorImpl<SDValue> &InVals) const {
1964 // TODO: add description of PPC stack frame format, or at least some docs.
1966 MachineFunction &MF = DAG.getMachineFunction();
1967 MachineFrameInfo *MFI = MF.getFrameInfo();
1968 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1970 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1971 // Potential tail calls could cause overwriting of argument stack slots.
1972 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1973 (CallConv == CallingConv::Fast));
1974 unsigned PtrByteSize = 8;
1976 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
1977 // Area that is at least reserved in caller of this function.
1978 unsigned MinReservedArea = ArgOffset;
1980 static const uint16_t GPR[] = {
1981 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1982 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1985 static const uint16_t *FPR = GetFPR();
1987 static const uint16_t VR[] = {
1988 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1989 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1992 const unsigned Num_GPR_Regs = array_lengthof(GPR);
1993 const unsigned Num_FPR_Regs = 13;
1994 const unsigned Num_VR_Regs = array_lengthof(VR);
1996 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1998 // Add DAG nodes to load the arguments or copy them out of registers. On
1999 // entry to a function on PPC, the arguments start after the linkage area,
2000 // although the first ones are often in registers.
2002 SmallVector<SDValue, 8> MemOps;
2003 unsigned nAltivecParamsAtEnd = 0;
2004 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2005 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2007 bool needsLoad = false;
2008 EVT ObjectVT = Ins[ArgNo].VT;
2009 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2010 unsigned ArgSize = ObjSize;
2011 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2013 unsigned CurArgOffset = ArgOffset;
2015 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2016 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2017 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2019 MinReservedArea = ((MinReservedArea+15)/16)*16;
2020 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2024 nAltivecParamsAtEnd++;
2026 // Calculate min reserved area.
2027 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2031 // FIXME the codegen can be much improved in some cases.
2032 // We do not have to keep everything in memory.
2033 if (Flags.isByVal()) {
2034 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2035 ObjSize = Flags.getByValSize();
2036 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2037 // All aggregates smaller than 8 bytes must be passed right-justified.
2038 if (ObjSize==1 || ObjSize==2) {
2039 CurArgOffset = CurArgOffset + (4 - ObjSize);
2041 // The value of the object is its address.
2042 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2043 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2044 InVals.push_back(FIN);
2045 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2046 if (GPR_idx != Num_GPR_Regs) {
2048 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2049 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2050 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2051 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2052 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2053 MachinePointerInfo(FuncArg,
2055 ObjType, false, false, 0);
2056 MemOps.push_back(Store);
2060 ArgOffset += PtrByteSize;
2064 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2065 // Store whatever pieces of the object are in registers
2066 // to memory. ArgOffset will be the address of the beginning
2068 if (GPR_idx != Num_GPR_Regs) {
2070 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2071 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2072 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2073 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2074 SDValue Shifted = Val;
2076 // For 64-bit SVR4, small structs come in right-adjusted.
2077 // Shift them left so the following logic works as expected.
2079 SDValue ShiftAmt = DAG.getConstant(64 - 8 * ObjSize, PtrVT);
2080 Shifted = DAG.getNode(ISD::SHL, dl, PtrVT, Val, ShiftAmt);
2083 SDValue Store = DAG.getStore(Val.getValue(1), dl, Shifted, FIN,
2084 MachinePointerInfo(FuncArg, ArgOffset),
2086 MemOps.push_back(Store);
2088 ArgOffset += PtrByteSize;
2090 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2097 switch (ObjectVT.getSimpleVT().SimpleTy) {
2098 default: llvm_unreachable("Unhandled argument type!");
2101 if (GPR_idx != Num_GPR_Regs) {
2102 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2103 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2105 if (ObjectVT == MVT::i32) {
2106 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2107 // value to MVT::i64 and then truncate to the correct register size.
2109 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2110 DAG.getValueType(ObjectVT));
2111 else if (Flags.isZExt())
2112 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2113 DAG.getValueType(ObjectVT));
2115 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2121 ArgSize = PtrByteSize;
2128 // Every 8 bytes of argument space consumes one of the GPRs available for
2129 // argument passing.
2130 if (GPR_idx != Num_GPR_Regs) {
2133 if (FPR_idx != Num_FPR_Regs) {
2136 if (ObjectVT == MVT::f32)
2137 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2139 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2141 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2145 ArgSize = PtrByteSize;
2154 // Note that vector arguments in registers don't reserve stack space,
2155 // except in varargs functions.
2156 if (VR_idx != Num_VR_Regs) {
2157 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2158 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2160 while ((ArgOffset % 16) != 0) {
2161 ArgOffset += PtrByteSize;
2162 if (GPR_idx != Num_GPR_Regs)
2166 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2170 // Vectors are aligned.
2171 ArgOffset = ((ArgOffset+15)/16)*16;
2172 CurArgOffset = ArgOffset;
2179 // We need to load the argument to a virtual register if we determined
2180 // above that we ran out of physical registers of the appropriate type.
2182 int FI = MFI->CreateFixedObject(ObjSize,
2183 CurArgOffset + (ArgSize - ObjSize),
2185 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2186 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2187 false, false, false, 0);
2190 InVals.push_back(ArgVal);
2193 // Set the size that is at least reserved in caller of this function. Tail
2194 // call optimized function's reserved stack space needs to be aligned so that
2195 // taking the difference between two stack areas will result in an aligned
2197 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2198 // Add the Altivec parameters at the end, if needed.
2199 if (nAltivecParamsAtEnd) {
2200 MinReservedArea = ((MinReservedArea+15)/16)*16;
2201 MinReservedArea += 16*nAltivecParamsAtEnd;
2204 std::max(MinReservedArea,
2205 PPCFrameLowering::getMinCallFrameSize(true, true));
2206 unsigned TargetAlign
2207 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2208 getStackAlignment();
2209 unsigned AlignMask = TargetAlign-1;
2210 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2211 FI->setMinReservedArea(MinReservedArea);
2213 // If the function takes variable number of arguments, make a frame index for
2214 // the start of the first vararg value... for expansion of llvm.va_start.
2216 int Depth = ArgOffset;
2218 FuncInfo->setVarArgsFrameIndex(
2219 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2221 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2223 // If this function is vararg, store any remaining integer argument regs
2224 // to their spots on the stack so that they may be loaded by deferencing the
2225 // result of va_next.
2226 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2227 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2228 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2229 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2230 MachinePointerInfo(), false, false, 0);
2231 MemOps.push_back(Store);
2232 // Increment the address by four for the next argument to store
2233 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2234 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2238 if (!MemOps.empty())
2239 Chain = DAG.getNode(ISD::TokenFactor, dl,
2240 MVT::Other, &MemOps[0], MemOps.size());
2246 PPCTargetLowering::LowerFormalArguments_Darwin(
2248 CallingConv::ID CallConv, bool isVarArg,
2249 const SmallVectorImpl<ISD::InputArg>
2251 DebugLoc dl, SelectionDAG &DAG,
2252 SmallVectorImpl<SDValue> &InVals) const {
2253 // TODO: add description of PPC stack frame format, or at least some docs.
2255 MachineFunction &MF = DAG.getMachineFunction();
2256 MachineFrameInfo *MFI = MF.getFrameInfo();
2257 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2259 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2260 bool isPPC64 = PtrVT == MVT::i64;
2261 // Potential tail calls could cause overwriting of argument stack slots.
2262 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2263 (CallConv == CallingConv::Fast));
2264 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2266 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2267 // Area that is at least reserved in caller of this function.
2268 unsigned MinReservedArea = ArgOffset;
2270 static const uint16_t GPR_32[] = { // 32-bit registers.
2271 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2272 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2274 static const uint16_t GPR_64[] = { // 64-bit registers.
2275 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2276 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2279 static const uint16_t *FPR = GetFPR();
2281 static const uint16_t VR[] = {
2282 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2283 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2286 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2287 const unsigned Num_FPR_Regs = 13;
2288 const unsigned Num_VR_Regs = array_lengthof( VR);
2290 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2292 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2294 // In 32-bit non-varargs functions, the stack space for vectors is after the
2295 // stack space for non-vectors. We do not use this space unless we have
2296 // too many vectors to fit in registers, something that only occurs in
2297 // constructed examples:), but we have to walk the arglist to figure
2298 // that out...for the pathological case, compute VecArgOffset as the
2299 // start of the vector parameter area. Computing VecArgOffset is the
2300 // entire point of the following loop.
2301 unsigned VecArgOffset = ArgOffset;
2302 if (!isVarArg && !isPPC64) {
2303 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2305 EVT ObjectVT = Ins[ArgNo].VT;
2306 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2308 if (Flags.isByVal()) {
2309 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2310 unsigned ObjSize = Flags.getByValSize();
2312 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2313 VecArgOffset += ArgSize;
2317 switch(ObjectVT.getSimpleVT().SimpleTy) {
2318 default: llvm_unreachable("Unhandled argument type!");
2323 case MVT::i64: // PPC64
2325 // FIXME: We are guaranteed to be !isPPC64 at this point.
2326 // Does MVT::i64 apply?
2333 // Nothing to do, we're only looking at Nonvector args here.
2338 // We've found where the vector parameter area in memory is. Skip the
2339 // first 12 parameters; these don't use that memory.
2340 VecArgOffset = ((VecArgOffset+15)/16)*16;
2341 VecArgOffset += 12*16;
2343 // Add DAG nodes to load the arguments or copy them out of registers. On
2344 // entry to a function on PPC, the arguments start after the linkage area,
2345 // although the first ones are often in registers.
2347 SmallVector<SDValue, 8> MemOps;
2348 unsigned nAltivecParamsAtEnd = 0;
2349 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2350 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2352 bool needsLoad = false;
2353 EVT ObjectVT = Ins[ArgNo].VT;
2354 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2355 unsigned ArgSize = ObjSize;
2356 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2358 unsigned CurArgOffset = ArgOffset;
2360 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2361 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2362 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2363 if (isVarArg || isPPC64) {
2364 MinReservedArea = ((MinReservedArea+15)/16)*16;
2365 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2368 } else nAltivecParamsAtEnd++;
2370 // Calculate min reserved area.
2371 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2375 // FIXME the codegen can be much improved in some cases.
2376 // We do not have to keep everything in memory.
2377 if (Flags.isByVal()) {
2378 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2379 ObjSize = Flags.getByValSize();
2380 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2381 // Objects of size 1 and 2 are right justified, everything else is
2382 // left justified. This means the memory address is adjusted forwards.
2383 if (ObjSize==1 || ObjSize==2) {
2384 CurArgOffset = CurArgOffset + (4 - ObjSize);
2386 // The value of the object is its address.
2387 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2388 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2389 InVals.push_back(FIN);
2390 if (ObjSize==1 || ObjSize==2) {
2391 if (GPR_idx != Num_GPR_Regs) {
2394 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2396 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2397 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2398 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2399 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2400 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2401 MachinePointerInfo(FuncArg,
2403 ObjType, false, false, 0);
2404 MemOps.push_back(Store);
2408 ArgOffset += PtrByteSize;
2412 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2413 // Store whatever pieces of the object are in registers
2414 // to memory. ArgOffset will be the address of the beginning
2416 if (GPR_idx != Num_GPR_Regs) {
2419 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2421 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2422 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2423 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2424 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2425 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2426 MachinePointerInfo(FuncArg, ArgOffset),
2428 MemOps.push_back(Store);
2430 ArgOffset += PtrByteSize;
2432 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2439 switch (ObjectVT.getSimpleVT().SimpleTy) {
2440 default: llvm_unreachable("Unhandled argument type!");
2443 if (GPR_idx != Num_GPR_Regs) {
2444 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2445 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2449 ArgSize = PtrByteSize;
2451 // All int arguments reserve stack space in the Darwin ABI.
2452 ArgOffset += PtrByteSize;
2456 case MVT::i64: // PPC64
2457 if (GPR_idx != Num_GPR_Regs) {
2458 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2459 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2461 if (ObjectVT == MVT::i32) {
2462 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2463 // value to MVT::i64 and then truncate to the correct register size.
2465 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2466 DAG.getValueType(ObjectVT));
2467 else if (Flags.isZExt())
2468 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2469 DAG.getValueType(ObjectVT));
2471 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2477 ArgSize = PtrByteSize;
2479 // All int arguments reserve stack space in the Darwin ABI.
2485 // Every 4 bytes of argument space consumes one of the GPRs available for
2486 // argument passing.
2487 if (GPR_idx != Num_GPR_Regs) {
2489 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2492 if (FPR_idx != Num_FPR_Regs) {
2495 if (ObjectVT == MVT::f32)
2496 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2498 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2500 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2506 // All FP arguments reserve stack space in the Darwin ABI.
2507 ArgOffset += isPPC64 ? 8 : ObjSize;
2513 // Note that vector arguments in registers don't reserve stack space,
2514 // except in varargs functions.
2515 if (VR_idx != Num_VR_Regs) {
2516 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2517 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2519 while ((ArgOffset % 16) != 0) {
2520 ArgOffset += PtrByteSize;
2521 if (GPR_idx != Num_GPR_Regs)
2525 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2529 if (!isVarArg && !isPPC64) {
2530 // Vectors go after all the nonvectors.
2531 CurArgOffset = VecArgOffset;
2534 // Vectors are aligned.
2535 ArgOffset = ((ArgOffset+15)/16)*16;
2536 CurArgOffset = ArgOffset;
2544 // We need to load the argument to a virtual register if we determined above
2545 // that we ran out of physical registers of the appropriate type.
2547 int FI = MFI->CreateFixedObject(ObjSize,
2548 CurArgOffset + (ArgSize - ObjSize),
2550 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2551 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2552 false, false, false, 0);
2555 InVals.push_back(ArgVal);
2558 // Set the size that is at least reserved in caller of this function. Tail
2559 // call optimized function's reserved stack space needs to be aligned so that
2560 // taking the difference between two stack areas will result in an aligned
2562 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2563 // Add the Altivec parameters at the end, if needed.
2564 if (nAltivecParamsAtEnd) {
2565 MinReservedArea = ((MinReservedArea+15)/16)*16;
2566 MinReservedArea += 16*nAltivecParamsAtEnd;
2569 std::max(MinReservedArea,
2570 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2571 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2572 getStackAlignment();
2573 unsigned AlignMask = TargetAlign-1;
2574 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2575 FI->setMinReservedArea(MinReservedArea);
2577 // If the function takes variable number of arguments, make a frame index for
2578 // the start of the first vararg value... for expansion of llvm.va_start.
2580 int Depth = ArgOffset;
2582 FuncInfo->setVarArgsFrameIndex(
2583 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2585 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2587 // If this function is vararg, store any remaining integer argument regs
2588 // to their spots on the stack so that they may be loaded by deferencing the
2589 // result of va_next.
2590 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2594 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2596 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2598 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2599 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2600 MachinePointerInfo(), false, false, 0);
2601 MemOps.push_back(Store);
2602 // Increment the address by four for the next argument to store
2603 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2604 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2608 if (!MemOps.empty())
2609 Chain = DAG.getNode(ISD::TokenFactor, dl,
2610 MVT::Other, &MemOps[0], MemOps.size());
2615 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2616 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2618 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2622 const SmallVectorImpl<ISD::OutputArg>
2624 const SmallVectorImpl<SDValue> &OutVals,
2625 unsigned &nAltivecParamsAtEnd) {
2626 // Count how many bytes are to be pushed on the stack, including the linkage
2627 // area, and parameter passing area. We start with 24/48 bytes, which is
2628 // prereserved space for [SP][CR][LR][3 x unused].
2629 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2630 unsigned NumOps = Outs.size();
2631 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2633 // Add up all the space actually used.
2634 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2635 // they all go in registers, but we must reserve stack space for them for
2636 // possible use by the caller. In varargs or 64-bit calls, parameters are
2637 // assigned stack space in order, with padding so Altivec parameters are
2639 nAltivecParamsAtEnd = 0;
2640 for (unsigned i = 0; i != NumOps; ++i) {
2641 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2642 EVT ArgVT = Outs[i].VT;
2643 // Varargs Altivec parameters are padded to a 16 byte boundary.
2644 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2645 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2646 if (!isVarArg && !isPPC64) {
2647 // Non-varargs Altivec parameters go after all the non-Altivec
2648 // parameters; handle those later so we know how much padding we need.
2649 nAltivecParamsAtEnd++;
2652 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2653 NumBytes = ((NumBytes+15)/16)*16;
2655 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2658 // Allow for Altivec parameters at the end, if needed.
2659 if (nAltivecParamsAtEnd) {
2660 NumBytes = ((NumBytes+15)/16)*16;
2661 NumBytes += 16*nAltivecParamsAtEnd;
2664 // The prolog code of the callee may store up to 8 GPR argument registers to
2665 // the stack, allowing va_start to index over them in memory if its varargs.
2666 // Because we cannot tell if this is needed on the caller side, we have to
2667 // conservatively assume that it is needed. As such, make sure we have at
2668 // least enough stack space for the caller to store the 8 GPRs.
2669 NumBytes = std::max(NumBytes,
2670 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2672 // Tail call needs the stack to be aligned.
2673 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2674 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2675 getFrameLowering()->getStackAlignment();
2676 unsigned AlignMask = TargetAlign-1;
2677 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2683 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2684 /// adjusted to accommodate the arguments for the tailcall.
2685 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2686 unsigned ParamSize) {
2688 if (!isTailCall) return 0;
2690 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2691 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2692 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2693 // Remember only if the new adjustement is bigger.
2694 if (SPDiff < FI->getTailCallSPDelta())
2695 FI->setTailCallSPDelta(SPDiff);
2700 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2701 /// for tail call optimization. Targets which want to do tail call
2702 /// optimization should implement this function.
2704 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2705 CallingConv::ID CalleeCC,
2707 const SmallVectorImpl<ISD::InputArg> &Ins,
2708 SelectionDAG& DAG) const {
2709 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2712 // Variable argument functions are not supported.
2716 MachineFunction &MF = DAG.getMachineFunction();
2717 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2718 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2719 // Functions containing by val parameters are not supported.
2720 for (unsigned i = 0; i != Ins.size(); i++) {
2721 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2722 if (Flags.isByVal()) return false;
2725 // Non PIC/GOT tail calls are supported.
2726 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2729 // At the moment we can only do local tail calls (in same module, hidden
2730 // or protected) if we are generating PIC.
2731 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2732 return G->getGlobal()->hasHiddenVisibility()
2733 || G->getGlobal()->hasProtectedVisibility();
2739 /// isCallCompatibleAddress - Return the immediate to use if the specified
2740 /// 32-bit value is representable in the immediate field of a BxA instruction.
2741 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2742 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2745 int Addr = C->getZExtValue();
2746 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2747 SignExtend32<26>(Addr) != Addr)
2748 return 0; // Top 6 bits have to be sext of immediate.
2750 return DAG.getConstant((int)C->getZExtValue() >> 2,
2751 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2756 struct TailCallArgumentInfo {
2761 TailCallArgumentInfo() : FrameIdx(0) {}
2766 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2768 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2770 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2771 SmallVector<SDValue, 8> &MemOpChains,
2773 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2774 SDValue Arg = TailCallArgs[i].Arg;
2775 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2776 int FI = TailCallArgs[i].FrameIdx;
2777 // Store relative to framepointer.
2778 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2779 MachinePointerInfo::getFixedStack(FI),
2784 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2785 /// the appropriate stack slot for the tail call optimized function call.
2786 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2787 MachineFunction &MF,
2796 // Calculate the new stack slot for the return address.
2797 int SlotSize = isPPC64 ? 8 : 4;
2798 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2800 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2801 NewRetAddrLoc, true);
2802 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2803 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2804 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2805 MachinePointerInfo::getFixedStack(NewRetAddr),
2808 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2809 // slot as the FP is never overwritten.
2812 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2813 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2815 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2816 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2817 MachinePointerInfo::getFixedStack(NewFPIdx),
2824 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2825 /// the position of the argument.
2827 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2828 SDValue Arg, int SPDiff, unsigned ArgOffset,
2829 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2830 int Offset = ArgOffset + SPDiff;
2831 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2832 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2833 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2834 SDValue FIN = DAG.getFrameIndex(FI, VT);
2835 TailCallArgumentInfo Info;
2837 Info.FrameIdxOp = FIN;
2839 TailCallArguments.push_back(Info);
2842 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2843 /// stack slot. Returns the chain as result and the loaded frame pointers in
2844 /// LROpOut/FPOpout. Used when tail calling.
2845 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2851 DebugLoc dl) const {
2853 // Load the LR and FP stack slot for later adjusting.
2854 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2855 LROpOut = getReturnAddrFrameIndex(DAG);
2856 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2857 false, false, false, 0);
2858 Chain = SDValue(LROpOut.getNode(), 1);
2860 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2861 // slot as the FP is never overwritten.
2863 FPOpOut = getFramePointerFrameIndex(DAG);
2864 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2865 false, false, false, 0);
2866 Chain = SDValue(FPOpOut.getNode(), 1);
2872 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2873 /// by "Src" to address "Dst" of size "Size". Alignment information is
2874 /// specified by the specific parameter attribute. The copy will be passed as
2875 /// a byval function parameter.
2876 /// Sometimes what we are copying is the end of a larger object, the part that
2877 /// does not fit in registers.
2879 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2880 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2882 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2883 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2884 false, false, MachinePointerInfo(0),
2885 MachinePointerInfo(0));
2888 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2891 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2892 SDValue Arg, SDValue PtrOff, int SPDiff,
2893 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2894 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2895 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2897 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2902 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2904 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2905 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2906 DAG.getConstant(ArgOffset, PtrVT));
2908 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2909 MachinePointerInfo(), false, false, 0));
2910 // Calculate and remember argument location.
2911 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2916 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2917 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2918 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2919 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2920 MachineFunction &MF = DAG.getMachineFunction();
2922 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2923 // might overwrite each other in case of tail call optimization.
2924 SmallVector<SDValue, 8> MemOpChains2;
2925 // Do not flag preceding copytoreg stuff together with the following stuff.
2927 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2929 if (!MemOpChains2.empty())
2930 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2931 &MemOpChains2[0], MemOpChains2.size());
2933 // Store the return address to the appropriate stack slot.
2934 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2935 isPPC64, isDarwinABI, dl);
2937 // Emit callseq_end just before tailcall node.
2938 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2939 DAG.getIntPtrConstant(0, true), InFlag);
2940 InFlag = Chain.getValue(1);
2944 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2945 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2946 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2947 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2948 const PPCSubtarget &PPCSubTarget) {
2950 bool isPPC64 = PPCSubTarget.isPPC64();
2951 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2953 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2954 NodeTys.push_back(MVT::Other); // Returns a chain
2955 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2957 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2959 bool needIndirectCall = true;
2960 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2961 // If this is an absolute destination address, use the munged value.
2962 Callee = SDValue(Dest, 0);
2963 needIndirectCall = false;
2966 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2967 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2968 // Use indirect calls for ALL functions calls in JIT mode, since the
2969 // far-call stubs may be outside relocation limits for a BL instruction.
2970 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2971 unsigned OpFlags = 0;
2972 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2973 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2974 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2975 (G->getGlobal()->isDeclaration() ||
2976 G->getGlobal()->isWeakForLinker())) {
2977 // PC-relative references to external symbols should go through $stub,
2978 // unless we're building with the leopard linker or later, which
2979 // automatically synthesizes these stubs.
2980 OpFlags = PPCII::MO_DARWIN_STUB;
2983 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2984 // every direct call is) turn it into a TargetGlobalAddress /
2985 // TargetExternalSymbol node so that legalize doesn't hack it.
2986 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2987 Callee.getValueType(),
2989 needIndirectCall = false;
2993 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2994 unsigned char OpFlags = 0;
2996 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2997 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2998 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2999 // PC-relative references to external symbols should go through $stub,
3000 // unless we're building with the leopard linker or later, which
3001 // automatically synthesizes these stubs.
3002 OpFlags = PPCII::MO_DARWIN_STUB;
3005 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3007 needIndirectCall = false;
3010 if (needIndirectCall) {
3011 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3012 // to do the call, we can't use PPCISD::CALL.
3013 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3015 if (isSVR4ABI && isPPC64) {
3016 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3017 // entry point, but to the function descriptor (the function entry point
3018 // address is part of the function descriptor though).
3019 // The function descriptor is a three doubleword structure with the
3020 // following fields: function entry point, TOC base address and
3021 // environment pointer.
3022 // Thus for a call through a function pointer, the following actions need
3024 // 1. Save the TOC of the caller in the TOC save area of its stack
3025 // frame (this is done in LowerCall_Darwin_Or_64SVR4()).
3026 // 2. Load the address of the function entry point from the function
3028 // 3. Load the TOC of the callee from the function descriptor into r2.
3029 // 4. Load the environment pointer from the function descriptor into
3031 // 5. Branch to the function entry point address.
3032 // 6. On return of the callee, the TOC of the caller needs to be
3033 // restored (this is done in FinishCall()).
3035 // All those operations are flagged together to ensure that no other
3036 // operations can be scheduled in between. E.g. without flagging the
3037 // operations together, a TOC access in the caller could be scheduled
3038 // between the load of the callee TOC and the branch to the callee, which
3039 // results in the TOC access going through the TOC of the callee instead
3040 // of going through the TOC of the caller, which leads to incorrect code.
3042 // Load the address of the function entry point from the function
3044 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3045 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3046 InFlag.getNode() ? 3 : 2);
3047 Chain = LoadFuncPtr.getValue(1);
3048 InFlag = LoadFuncPtr.getValue(2);
3050 // Load environment pointer into r11.
3051 // Offset of the environment pointer within the function descriptor.
3052 SDValue PtrOff = DAG.getIntPtrConstant(16);
3054 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3055 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3057 Chain = LoadEnvPtr.getValue(1);
3058 InFlag = LoadEnvPtr.getValue(2);
3060 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3062 Chain = EnvVal.getValue(0);
3063 InFlag = EnvVal.getValue(1);
3065 // Load TOC of the callee into r2. We are using a target-specific load
3066 // with r2 hard coded, because the result of a target-independent load
3067 // would never go directly into r2, since r2 is a reserved register (which
3068 // prevents the register allocator from allocating it), resulting in an
3069 // additional register being allocated and an unnecessary move instruction
3071 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3072 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3074 Chain = LoadTOCPtr.getValue(0);
3075 InFlag = LoadTOCPtr.getValue(1);
3077 MTCTROps[0] = Chain;
3078 MTCTROps[1] = LoadFuncPtr;
3079 MTCTROps[2] = InFlag;
3082 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3083 2 + (InFlag.getNode() != 0));
3084 InFlag = Chain.getValue(1);
3087 NodeTys.push_back(MVT::Other);
3088 NodeTys.push_back(MVT::Glue);
3089 Ops.push_back(Chain);
3090 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3092 // Add CTR register as callee so a bctr can be emitted later.
3094 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3097 // If this is a direct call, pass the chain and the callee.
3098 if (Callee.getNode()) {
3099 Ops.push_back(Chain);
3100 Ops.push_back(Callee);
3102 // If this is a tail call add stack pointer delta.
3104 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3106 // Add argument registers to the end of the list so that they are known live
3108 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3109 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3110 RegsToPass[i].second.getValueType()));
3116 bool isLocalCall(const SDValue &Callee)
3118 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3119 return !G->getGlobal()->isDeclaration() &&
3120 !G->getGlobal()->isWeakForLinker();
3125 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3126 CallingConv::ID CallConv, bool isVarArg,
3127 const SmallVectorImpl<ISD::InputArg> &Ins,
3128 DebugLoc dl, SelectionDAG &DAG,
3129 SmallVectorImpl<SDValue> &InVals) const {
3131 SmallVector<CCValAssign, 16> RVLocs;
3132 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3133 getTargetMachine(), RVLocs, *DAG.getContext());
3134 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3136 // Copy all of the result registers out of their specified physreg.
3137 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3138 CCValAssign &VA = RVLocs[i];
3139 EVT VT = VA.getValVT();
3140 assert(VA.isRegLoc() && "Can only return in registers!");
3141 Chain = DAG.getCopyFromReg(Chain, dl,
3142 VA.getLocReg(), VT, InFlag).getValue(1);
3143 InVals.push_back(Chain.getValue(0));
3144 InFlag = Chain.getValue(2);
3151 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3152 bool isTailCall, bool isVarArg,
3154 SmallVector<std::pair<unsigned, SDValue>, 8>
3156 SDValue InFlag, SDValue Chain,
3158 int SPDiff, unsigned NumBytes,
3159 const SmallVectorImpl<ISD::InputArg> &Ins,
3160 SmallVectorImpl<SDValue> &InVals) const {
3161 std::vector<EVT> NodeTys;
3162 SmallVector<SDValue, 8> Ops;
3163 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3164 isTailCall, RegsToPass, Ops, NodeTys,
3167 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3168 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3169 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3171 // When performing tail call optimization the callee pops its arguments off
3172 // the stack. Account for this here so these bytes can be pushed back on in
3173 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3174 int BytesCalleePops =
3175 (CallConv == CallingConv::Fast &&
3176 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3178 // Add a register mask operand representing the call-preserved registers.
3179 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3180 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3181 assert(Mask && "Missing call preserved mask for calling convention");
3182 Ops.push_back(DAG.getRegisterMask(Mask));
3184 if (InFlag.getNode())
3185 Ops.push_back(InFlag);
3189 // If this is the first return lowered for this function, add the regs
3190 // to the liveout set for the function.
3191 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3192 SmallVector<CCValAssign, 16> RVLocs;
3193 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3194 getTargetMachine(), RVLocs, *DAG.getContext());
3195 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3196 for (unsigned i = 0; i != RVLocs.size(); ++i)
3197 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3200 assert(((Callee.getOpcode() == ISD::Register &&
3201 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3202 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3203 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3204 isa<ConstantSDNode>(Callee)) &&
3205 "Expecting an global address, external symbol, absolute value or register");
3207 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3210 // Add a NOP immediately after the branch instruction when using the 64-bit
3211 // SVR4 ABI. At link time, if caller and callee are in a different module and
3212 // thus have a different TOC, the call will be replaced with a call to a stub
3213 // function which saves the current TOC, loads the TOC of the callee and
3214 // branches to the callee. The NOP will be replaced with a load instruction
3215 // which restores the TOC of the caller from the TOC save slot of the current
3216 // stack frame. If caller and callee belong to the same module (and have the
3217 // same TOC), the NOP will remain unchanged.
3219 bool needsTOCRestore = false;
3220 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3221 if (CallOpc == PPCISD::BCTRL_SVR4) {
3222 // This is a call through a function pointer.
3223 // Restore the caller TOC from the save area into R2.
3224 // See PrepareCall() for more information about calls through function
3225 // pointers in the 64-bit SVR4 ABI.
3226 // We are using a target-specific load with r2 hard coded, because the
3227 // result of a target-independent load would never go directly into r2,
3228 // since r2 is a reserved register (which prevents the register allocator
3229 // from allocating it), resulting in an additional register being
3230 // allocated and an unnecessary move instruction being generated.
3231 needsTOCRestore = true;
3232 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3233 // Otherwise insert NOP for non-local calls.
3234 CallOpc = PPCISD::CALL_NOP_SVR4;
3238 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3239 InFlag = Chain.getValue(1);
3241 if (needsTOCRestore) {
3242 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3243 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3244 InFlag = Chain.getValue(1);
3247 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3248 DAG.getIntPtrConstant(BytesCalleePops, true),
3251 InFlag = Chain.getValue(1);
3253 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3254 Ins, dl, DAG, InVals);
3258 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3259 SmallVectorImpl<SDValue> &InVals) const {
3260 SelectionDAG &DAG = CLI.DAG;
3261 DebugLoc &dl = CLI.DL;
3262 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3263 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3264 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3265 SDValue Chain = CLI.Chain;
3266 SDValue Callee = CLI.Callee;
3267 bool &isTailCall = CLI.IsTailCall;
3268 CallingConv::ID CallConv = CLI.CallConv;
3269 bool isVarArg = CLI.IsVarArg;
3272 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3275 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3276 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3277 isTailCall, Outs, OutVals, Ins,
3280 return LowerCall_Darwin_Or_64SVR4(Chain, Callee, CallConv, isVarArg,
3281 isTailCall, Outs, OutVals, Ins,
3286 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3287 CallingConv::ID CallConv, bool isVarArg,
3289 const SmallVectorImpl<ISD::OutputArg> &Outs,
3290 const SmallVectorImpl<SDValue> &OutVals,
3291 const SmallVectorImpl<ISD::InputArg> &Ins,
3292 DebugLoc dl, SelectionDAG &DAG,
3293 SmallVectorImpl<SDValue> &InVals) const {
3294 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3295 // of the 32-bit SVR4 ABI stack frame layout.
3297 assert((CallConv == CallingConv::C ||
3298 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3300 unsigned PtrByteSize = 4;
3302 MachineFunction &MF = DAG.getMachineFunction();
3304 // Mark this function as potentially containing a function that contains a
3305 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3306 // and restoring the callers stack pointer in this functions epilog. This is
3307 // done because by tail calling the called function might overwrite the value
3308 // in this function's (MF) stack pointer stack slot 0(SP).
3309 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3310 CallConv == CallingConv::Fast)
3311 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3313 // Count how many bytes are to be pushed on the stack, including the linkage
3314 // area, parameter list area and the part of the local variable space which
3315 // contains copies of aggregates which are passed by value.
3317 // Assign locations to all of the outgoing arguments.
3318 SmallVector<CCValAssign, 16> ArgLocs;
3319 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3320 getTargetMachine(), ArgLocs, *DAG.getContext());
3322 // Reserve space for the linkage area on the stack.
3323 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3326 // Handle fixed and variable vector arguments differently.
3327 // Fixed vector arguments go into registers as long as registers are
3328 // available. Variable vector arguments always go into memory.
3329 unsigned NumArgs = Outs.size();
3331 for (unsigned i = 0; i != NumArgs; ++i) {
3332 MVT ArgVT = Outs[i].VT;
3333 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3336 if (Outs[i].IsFixed) {
3337 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3340 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3346 errs() << "Call operand #" << i << " has unhandled type "
3347 << EVT(ArgVT).getEVTString() << "\n";
3349 llvm_unreachable(0);
3353 // All arguments are treated the same.
3354 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3357 // Assign locations to all of the outgoing aggregate by value arguments.
3358 SmallVector<CCValAssign, 16> ByValArgLocs;
3359 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3360 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3362 // Reserve stack space for the allocations in CCInfo.
3363 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3365 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3367 // Size of the linkage area, parameter list area and the part of the local
3368 // space variable where copies of aggregates which are passed by value are
3370 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3372 // Calculate by how many bytes the stack has to be adjusted in case of tail
3373 // call optimization.
3374 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3376 // Adjust the stack pointer for the new arguments...
3377 // These operations are automatically eliminated by the prolog/epilog pass
3378 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3379 SDValue CallSeqStart = Chain;
3381 // Load the return address and frame pointer so it can be moved somewhere else
3384 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3387 // Set up a copy of the stack pointer for use loading and storing any
3388 // arguments that may not fit in the registers available for argument
3390 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3392 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3393 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3394 SmallVector<SDValue, 8> MemOpChains;
3396 bool seenFloatArg = false;
3397 // Walk the register/memloc assignments, inserting copies/loads.
3398 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3401 CCValAssign &VA = ArgLocs[i];
3402 SDValue Arg = OutVals[i];
3403 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3405 if (Flags.isByVal()) {
3406 // Argument is an aggregate which is passed by value, thus we need to
3407 // create a copy of it in the local variable space of the current stack
3408 // frame (which is the stack frame of the caller) and pass the address of
3409 // this copy to the callee.
3410 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3411 CCValAssign &ByValVA = ByValArgLocs[j++];
3412 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3414 // Memory reserved in the local variable space of the callers stack frame.
3415 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3417 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3418 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3420 // Create a copy of the argument in the local area of the current
3422 SDValue MemcpyCall =
3423 CreateCopyOfByValArgument(Arg, PtrOff,
3424 CallSeqStart.getNode()->getOperand(0),
3427 // This must go outside the CALLSEQ_START..END.
3428 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3429 CallSeqStart.getNode()->getOperand(1));
3430 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3431 NewCallSeqStart.getNode());
3432 Chain = CallSeqStart = NewCallSeqStart;
3434 // Pass the address of the aggregate copy on the stack either in a
3435 // physical register or in the parameter list area of the current stack
3436 // frame to the callee.
3440 if (VA.isRegLoc()) {
3441 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3442 // Put argument in a physical register.
3443 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3445 // Put argument in the parameter list area of the current stack frame.
3446 assert(VA.isMemLoc());
3447 unsigned LocMemOffset = VA.getLocMemOffset();
3450 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3451 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3453 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3454 MachinePointerInfo(),
3457 // Calculate and remember argument location.
3458 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3464 if (!MemOpChains.empty())
3465 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3466 &MemOpChains[0], MemOpChains.size());
3468 // Build a sequence of copy-to-reg nodes chained together with token chain
3469 // and flag operands which copy the outgoing args into the appropriate regs.
3471 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3472 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3473 RegsToPass[i].second, InFlag);
3474 InFlag = Chain.getValue(1);
3477 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3480 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3481 SDValue Ops[] = { Chain, InFlag };
3483 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3484 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3486 InFlag = Chain.getValue(1);
3490 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3491 false, TailCallArguments);
3493 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3494 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3499 PPCTargetLowering::LowerCall_Darwin_Or_64SVR4(SDValue Chain, SDValue Callee,
3500 CallingConv::ID CallConv, bool isVarArg,
3502 const SmallVectorImpl<ISD::OutputArg> &Outs,
3503 const SmallVectorImpl<SDValue> &OutVals,
3504 const SmallVectorImpl<ISD::InputArg> &Ins,
3505 DebugLoc dl, SelectionDAG &DAG,
3506 SmallVectorImpl<SDValue> &InVals) const {
3508 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3510 unsigned NumOps = Outs.size();
3512 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3513 bool isPPC64 = PtrVT == MVT::i64;
3514 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3516 MachineFunction &MF = DAG.getMachineFunction();
3518 // Mark this function as potentially containing a function that contains a
3519 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3520 // and restoring the callers stack pointer in this functions epilog. This is
3521 // done because by tail calling the called function might overwrite the value
3522 // in this function's (MF) stack pointer stack slot 0(SP).
3523 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3524 CallConv == CallingConv::Fast)
3525 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3527 unsigned nAltivecParamsAtEnd = 0;
3529 // Count how many bytes are to be pushed on the stack, including the linkage
3530 // area, and parameter passing area. We start with 24/48 bytes, which is
3531 // prereserved space for [SP][CR][LR][3 x unused].
3533 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3535 nAltivecParamsAtEnd);
3537 // Calculate by how many bytes the stack has to be adjusted in case of tail
3538 // call optimization.
3539 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3541 // To protect arguments on the stack from being clobbered in a tail call,
3542 // force all the loads to happen before doing any other lowering.
3544 Chain = DAG.getStackArgumentTokenFactor(Chain);
3546 // Adjust the stack pointer for the new arguments...
3547 // These operations are automatically eliminated by the prolog/epilog pass
3548 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3549 SDValue CallSeqStart = Chain;
3551 // Load the return address and frame pointer so it can be move somewhere else
3554 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3557 // Set up a copy of the stack pointer for use loading and storing any
3558 // arguments that may not fit in the registers available for argument
3562 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3564 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3566 // Figure out which arguments are going to go in registers, and which in
3567 // memory. Also, if this is a vararg function, floating point operations
3568 // must be stored to our stack, and loaded into integer regs as well, if
3569 // any integer regs are available for argument passing.
3570 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3571 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3573 static const uint16_t GPR_32[] = { // 32-bit registers.
3574 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3575 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3577 static const uint16_t GPR_64[] = { // 64-bit registers.
3578 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3579 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3581 static const uint16_t *FPR = GetFPR();
3583 static const uint16_t VR[] = {
3584 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3585 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3587 const unsigned NumGPRs = array_lengthof(GPR_32);
3588 const unsigned NumFPRs = 13;
3589 const unsigned NumVRs = array_lengthof(VR);
3591 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3593 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3594 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3596 SmallVector<SDValue, 8> MemOpChains;
3597 for (unsigned i = 0; i != NumOps; ++i) {
3598 SDValue Arg = OutVals[i];
3599 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3601 // PtrOff will be used to store the current argument to the stack if a
3602 // register cannot be found for it.
3605 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3607 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3609 // On PPC64, promote integers to 64-bit values.
3610 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3611 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3612 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3613 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3616 // FIXME memcpy is used way more than necessary. Correctness first.
3617 // Note: "by value" is code for passing a structure by value, not
3619 if (Flags.isByVal()) {
3620 // Note: Size includes alignment padding, so
3621 // struct x { short a; char b; }
3622 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3623 // These are the proper values we need for right-justifying the
3624 // aggregate in a parameter register for 64-bit SVR4.
3625 unsigned Size = Flags.getByValSize();
3626 // FOR DARWIN ONLY: Very small objects are passed right-justified.
3627 // Everything else is passed left-justified.
3628 // FOR 64-BIT SVR4: All aggregates smaller than 8 bytes must
3629 // be passed right-justified.
3630 if (Size==1 || Size==2 ||
3631 (Size==4 && isSVR4ABI)) {
3632 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3633 if (GPR_idx != NumGPRs) {
3634 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3635 MachinePointerInfo(), VT,
3637 MemOpChains.push_back(Load.getValue(1));
3638 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3640 ArgOffset += PtrByteSize;
3642 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3643 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3644 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3645 CallSeqStart.getNode()->getOperand(0),
3647 // This must go outside the CALLSEQ_START..END.
3648 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3649 CallSeqStart.getNode()->getOperand(1));
3650 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3651 NewCallSeqStart.getNode());
3652 Chain = CallSeqStart = NewCallSeqStart;
3653 ArgOffset += PtrByteSize;
3657 // Copy entire object into memory. There are cases where gcc-generated
3658 // code assumes it is there, even if it could be put entirely into
3659 // registers. (This is not what the doc says.)
3661 // FIXME: The above statement is likely due to a misunderstanding of the
3662 // documents. At least for 64-bit SVR4, all arguments must be copied
3663 // into the parameter area BY THE CALLEE in the event that the callee
3664 // takes the address of any formal argument. That has not yet been
3665 // implemented. However, it is reasonable to use the stack area as a
3666 // staging area for the register load.
3668 // Skip this for small aggregates under 64-bit SVR4, as we will use
3669 // the same slot for a right-justified copy, below.
3670 if (Size >= 8 || !isSVR4ABI) {
3671 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3672 CallSeqStart.getNode()->getOperand(0),
3674 // This must go outside the CALLSEQ_START..END.
3675 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3676 CallSeqStart.getNode()->getOperand(1));
3677 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3678 NewCallSeqStart.getNode());
3679 Chain = CallSeqStart = NewCallSeqStart;
3682 // FOR 64-BIT SVR4: When a register is available, pass the
3683 // aggregate right-justified.
3684 if (isSVR4ABI && Size < 8 && GPR_idx != NumGPRs) {
3685 // The easiest way to get this right-justified in a register
3686 // is to copy the structure into the rightmost portion of a
3687 // local variable slot, then load the whole slot into the
3689 // FIXME: The memcpy seems to produce pretty awful code for
3690 // small aggregates, particularly for packed ones.
3691 // FIXME: It would be preferable to use the slot in the
3692 // parameter save area instead of a new local variable.
3693 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3694 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3695 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3696 CallSeqStart.getNode()->getOperand(0),
3699 // Place the memcpy outside the CALLSEQ_START..END.
3700 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3701 CallSeqStart.getNode()->getOperand(1));
3702 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3703 NewCallSeqStart.getNode());
3704 Chain = CallSeqStart = NewCallSeqStart;
3706 // Load the slot into the register.
3707 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3708 MachinePointerInfo(),
3709 false, false, false, 0);
3710 MemOpChains.push_back(Load.getValue(1));
3711 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3713 // Done with this argument.
3714 ArgOffset += PtrByteSize;
3718 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
3719 // copy the pieces of the object that fit into registers from the
3720 // parameter save area.
3721 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3722 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3723 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3724 if (GPR_idx != NumGPRs) {
3725 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3726 MachinePointerInfo(),
3727 false, false, false, 0);
3728 MemOpChains.push_back(Load.getValue(1));
3729 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3730 ArgOffset += PtrByteSize;
3732 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3739 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3740 default: llvm_unreachable("Unexpected ValueType for argument!");
3743 if (GPR_idx != NumGPRs) {
3744 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3746 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3747 isPPC64, isTailCall, false, MemOpChains,
3748 TailCallArguments, dl);
3750 ArgOffset += PtrByteSize;
3754 if (FPR_idx != NumFPRs) {
3755 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3758 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3759 MachinePointerInfo(), false, false, 0);
3760 MemOpChains.push_back(Store);
3762 // Float varargs are always shadowed in available integer registers
3763 if (GPR_idx != NumGPRs) {
3764 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3765 MachinePointerInfo(), false, false,
3767 MemOpChains.push_back(Load.getValue(1));
3768 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3770 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3771 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3772 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3773 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3774 MachinePointerInfo(),
3775 false, false, false, 0);
3776 MemOpChains.push_back(Load.getValue(1));
3777 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3780 // If we have any FPRs remaining, we may also have GPRs remaining.
3781 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3783 if (GPR_idx != NumGPRs)
3785 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3786 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3790 // Single-precision floating-point values are mapped to the
3791 // second (rightmost) word of the stack doubleword.
3792 if (Arg.getValueType() == MVT::f32 && isPPC64 && isSVR4ABI) {
3793 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3794 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3797 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3798 isPPC64, isTailCall, false, MemOpChains,
3799 TailCallArguments, dl);
3804 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3811 // These go aligned on the stack, or in the corresponding R registers
3812 // when within range. The Darwin PPC ABI doc claims they also go in
3813 // V registers; in fact gcc does this only for arguments that are
3814 // prototyped, not for those that match the ... We do it for all
3815 // arguments, seems to work.
3816 while (ArgOffset % 16 !=0) {
3817 ArgOffset += PtrByteSize;
3818 if (GPR_idx != NumGPRs)
3821 // We could elide this store in the case where the object fits
3822 // entirely in R registers. Maybe later.
3823 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3824 DAG.getConstant(ArgOffset, PtrVT));
3825 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3826 MachinePointerInfo(), false, false, 0);
3827 MemOpChains.push_back(Store);
3828 if (VR_idx != NumVRs) {
3829 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3830 MachinePointerInfo(),
3831 false, false, false, 0);
3832 MemOpChains.push_back(Load.getValue(1));
3833 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3836 for (unsigned i=0; i<16; i+=PtrByteSize) {
3837 if (GPR_idx == NumGPRs)
3839 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3840 DAG.getConstant(i, PtrVT));
3841 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3842 false, false, false, 0);
3843 MemOpChains.push_back(Load.getValue(1));
3844 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3849 // Non-varargs Altivec params generally go in registers, but have
3850 // stack space allocated at the end.
3851 if (VR_idx != NumVRs) {
3852 // Doesn't have GPR space allocated.
3853 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3854 } else if (nAltivecParamsAtEnd==0) {
3855 // We are emitting Altivec params in order.
3856 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3857 isPPC64, isTailCall, true, MemOpChains,
3858 TailCallArguments, dl);
3864 // If all Altivec parameters fit in registers, as they usually do,
3865 // they get stack space following the non-Altivec parameters. We
3866 // don't track this here because nobody below needs it.
3867 // If there are more Altivec parameters than fit in registers emit
3869 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3871 // Offset is aligned; skip 1st 12 params which go in V registers.
3872 ArgOffset = ((ArgOffset+15)/16)*16;
3874 for (unsigned i = 0; i != NumOps; ++i) {
3875 SDValue Arg = OutVals[i];
3876 EVT ArgType = Outs[i].VT;
3877 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3878 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3881 // We are emitting Altivec params in order.
3882 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3883 isPPC64, isTailCall, true, MemOpChains,
3884 TailCallArguments, dl);
3891 if (!MemOpChains.empty())
3892 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3893 &MemOpChains[0], MemOpChains.size());
3895 // Check if this is an indirect call (MTCTR/BCTRL).
3896 // See PrepareCall() for more information about calls through function
3897 // pointers in the 64-bit SVR4 ABI.
3898 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3899 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3900 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3901 !isBLACompatibleAddress(Callee, DAG)) {
3902 // Load r2 into a virtual register and store it to the TOC save area.
3903 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3904 // TOC save area offset.
3905 SDValue PtrOff = DAG.getIntPtrConstant(40);
3906 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3907 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3911 // On Darwin, R12 must contain the address of an indirect callee. This does
3912 // not mean the MTCTR instruction must use R12; it's easier to model this as
3913 // an extra parameter, so do that.
3915 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3916 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3917 !isBLACompatibleAddress(Callee, DAG))
3918 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3919 PPC::R12), Callee));
3921 // Build a sequence of copy-to-reg nodes chained together with token chain
3922 // and flag operands which copy the outgoing args into the appropriate regs.
3924 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3925 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3926 RegsToPass[i].second, InFlag);
3927 InFlag = Chain.getValue(1);
3931 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3932 FPOp, true, TailCallArguments);
3934 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3935 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3940 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3941 MachineFunction &MF, bool isVarArg,
3942 const SmallVectorImpl<ISD::OutputArg> &Outs,
3943 LLVMContext &Context) const {
3944 SmallVector<CCValAssign, 16> RVLocs;
3945 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3947 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3951 PPCTargetLowering::LowerReturn(SDValue Chain,
3952 CallingConv::ID CallConv, bool isVarArg,
3953 const SmallVectorImpl<ISD::OutputArg> &Outs,
3954 const SmallVectorImpl<SDValue> &OutVals,
3955 DebugLoc dl, SelectionDAG &DAG) const {
3957 SmallVector<CCValAssign, 16> RVLocs;
3958 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3959 getTargetMachine(), RVLocs, *DAG.getContext());
3960 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3962 // If this is the first return lowered for this function, add the regs to the
3963 // liveout set for the function.
3964 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3965 for (unsigned i = 0; i != RVLocs.size(); ++i)
3966 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3971 // Copy the result values into the output registers.
3972 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3973 CCValAssign &VA = RVLocs[i];
3974 assert(VA.isRegLoc() && "Can only return in registers!");
3975 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3977 Flag = Chain.getValue(1);
3981 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3983 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3986 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3987 const PPCSubtarget &Subtarget) const {
3988 // When we pop the dynamic allocation we need to restore the SP link.
3989 DebugLoc dl = Op.getDebugLoc();
3991 // Get the corect type for pointers.
3992 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3994 // Construct the stack pointer operand.
3995 bool isPPC64 = Subtarget.isPPC64();
3996 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3997 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3999 // Get the operands for the STACKRESTORE.
4000 SDValue Chain = Op.getOperand(0);
4001 SDValue SaveSP = Op.getOperand(1);
4003 // Load the old link SP.
4004 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4005 MachinePointerInfo(),
4006 false, false, false, 0);
4008 // Restore the stack pointer.
4009 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4011 // Store the old link SP.
4012 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4019 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4020 MachineFunction &MF = DAG.getMachineFunction();
4021 bool isPPC64 = PPCSubTarget.isPPC64();
4022 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4023 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4025 // Get current frame pointer save index. The users of this index will be
4026 // primarily DYNALLOC instructions.
4027 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4028 int RASI = FI->getReturnAddrSaveIndex();
4030 // If the frame pointer save index hasn't been defined yet.
4032 // Find out what the fix offset of the frame pointer save area.
4033 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4034 // Allocate the frame index for frame pointer save area.
4035 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4037 FI->setReturnAddrSaveIndex(RASI);
4039 return DAG.getFrameIndex(RASI, PtrVT);
4043 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4044 MachineFunction &MF = DAG.getMachineFunction();
4045 bool isPPC64 = PPCSubTarget.isPPC64();
4046 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4047 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4049 // Get current frame pointer save index. The users of this index will be
4050 // primarily DYNALLOC instructions.
4051 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4052 int FPSI = FI->getFramePointerSaveIndex();
4054 // If the frame pointer save index hasn't been defined yet.
4056 // Find out what the fix offset of the frame pointer save area.
4057 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4060 // Allocate the frame index for frame pointer save area.
4061 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4063 FI->setFramePointerSaveIndex(FPSI);
4065 return DAG.getFrameIndex(FPSI, PtrVT);
4068 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4070 const PPCSubtarget &Subtarget) const {
4072 SDValue Chain = Op.getOperand(0);
4073 SDValue Size = Op.getOperand(1);
4074 DebugLoc dl = Op.getDebugLoc();
4076 // Get the corect type for pointers.
4077 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4079 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4080 DAG.getConstant(0, PtrVT), Size);
4081 // Construct a node for the frame pointer save index.
4082 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4083 // Build a DYNALLOC node.
4084 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4085 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4086 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4089 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4091 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4092 // Not FP? Not a fsel.
4093 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4094 !Op.getOperand(2).getValueType().isFloatingPoint())
4097 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4099 // Cannot handle SETEQ/SETNE.
4100 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4102 EVT ResVT = Op.getValueType();
4103 EVT CmpVT = Op.getOperand(0).getValueType();
4104 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4105 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4106 DebugLoc dl = Op.getDebugLoc();
4108 // If the RHS of the comparison is a 0.0, we don't need to do the
4109 // subtraction at all.
4110 if (isFloatingPointZero(RHS))
4112 default: break; // SETUO etc aren't handled by fsel.
4115 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4118 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4119 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4120 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4123 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4126 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4127 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4128 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4129 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4134 default: break; // SETUO etc aren't handled by fsel.
4137 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4138 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4139 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4140 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4143 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4144 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4145 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4146 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4149 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4150 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4151 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4152 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4155 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4156 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4157 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4158 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4163 // FIXME: Split this code up when LegalizeDAGTypes lands.
4164 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4165 DebugLoc dl) const {
4166 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4167 SDValue Src = Op.getOperand(0);
4168 if (Src.getValueType() == MVT::f32)
4169 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4172 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4173 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4175 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4180 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4184 // Convert the FP value to an int value through memory.
4185 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4187 // Emit a store to the stack slot.
4188 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4189 MachinePointerInfo(), false, false, 0);
4191 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4193 if (Op.getValueType() == MVT::i32)
4194 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4195 DAG.getConstant(4, FIPtr.getValueType()));
4196 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4197 false, false, false, 0);
4200 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4201 SelectionDAG &DAG) const {
4202 DebugLoc dl = Op.getDebugLoc();
4203 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4204 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4207 if (Op.getOperand(0).getValueType() == MVT::i64) {
4208 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
4209 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4210 if (Op.getValueType() == MVT::f32)
4211 FP = DAG.getNode(ISD::FP_ROUND, dl,
4212 MVT::f32, FP, DAG.getIntPtrConstant(0));
4216 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4217 "Unhandled SINT_TO_FP type in custom expander!");
4218 // Since we only generate this in 64-bit mode, we can take advantage of
4219 // 64-bit registers. In particular, sign extend the input value into the
4220 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4221 // then lfd it and fcfid it.
4222 MachineFunction &MF = DAG.getMachineFunction();
4223 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4224 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4225 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4226 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4228 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4231 // STD the extended value into the stack slot.
4232 MachineMemOperand *MMO =
4233 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4234 MachineMemOperand::MOStore, 8, 8);
4235 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4237 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4238 Ops, 4, MVT::i64, MMO);
4239 // Load the value as a double.
4240 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4241 false, false, false, 0);
4243 // FCFID it and return it.
4244 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4245 if (Op.getValueType() == MVT::f32)
4246 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4250 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4251 SelectionDAG &DAG) const {
4252 DebugLoc dl = Op.getDebugLoc();
4254 The rounding mode is in bits 30:31 of FPSR, and has the following
4261 FLT_ROUNDS, on the other hand, expects the following:
4268 To perform the conversion, we do:
4269 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4272 MachineFunction &MF = DAG.getMachineFunction();
4273 EVT VT = Op.getValueType();
4274 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4275 std::vector<EVT> NodeTys;
4276 SDValue MFFSreg, InFlag;
4278 // Save FP Control Word to register
4279 NodeTys.push_back(MVT::f64); // return register
4280 NodeTys.push_back(MVT::Glue); // unused in this context
4281 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4283 // Save FP register to stack slot
4284 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4285 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4286 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4287 StackSlot, MachinePointerInfo(), false, false,0);
4289 // Load FP Control Word from low 32 bits of stack slot.
4290 SDValue Four = DAG.getConstant(4, PtrVT);
4291 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4292 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4293 false, false, false, 0);
4295 // Transform as necessary
4297 DAG.getNode(ISD::AND, dl, MVT::i32,
4298 CWD, DAG.getConstant(3, MVT::i32));
4300 DAG.getNode(ISD::SRL, dl, MVT::i32,
4301 DAG.getNode(ISD::AND, dl, MVT::i32,
4302 DAG.getNode(ISD::XOR, dl, MVT::i32,
4303 CWD, DAG.getConstant(3, MVT::i32)),
4304 DAG.getConstant(3, MVT::i32)),
4305 DAG.getConstant(1, MVT::i32));
4308 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4310 return DAG.getNode((VT.getSizeInBits() < 16 ?
4311 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4314 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4315 EVT VT = Op.getValueType();
4316 unsigned BitWidth = VT.getSizeInBits();
4317 DebugLoc dl = Op.getDebugLoc();
4318 assert(Op.getNumOperands() == 3 &&
4319 VT == Op.getOperand(1).getValueType() &&
4322 // Expand into a bunch of logical ops. Note that these ops
4323 // depend on the PPC behavior for oversized shift amounts.
4324 SDValue Lo = Op.getOperand(0);
4325 SDValue Hi = Op.getOperand(1);
4326 SDValue Amt = Op.getOperand(2);
4327 EVT AmtVT = Amt.getValueType();
4329 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4330 DAG.getConstant(BitWidth, AmtVT), Amt);
4331 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4332 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4333 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4334 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4335 DAG.getConstant(-BitWidth, AmtVT));
4336 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4337 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4338 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4339 SDValue OutOps[] = { OutLo, OutHi };
4340 return DAG.getMergeValues(OutOps, 2, dl);
4343 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4344 EVT VT = Op.getValueType();
4345 DebugLoc dl = Op.getDebugLoc();
4346 unsigned BitWidth = VT.getSizeInBits();
4347 assert(Op.getNumOperands() == 3 &&
4348 VT == Op.getOperand(1).getValueType() &&
4351 // Expand into a bunch of logical ops. Note that these ops
4352 // depend on the PPC behavior for oversized shift amounts.
4353 SDValue Lo = Op.getOperand(0);
4354 SDValue Hi = Op.getOperand(1);
4355 SDValue Amt = Op.getOperand(2);
4356 EVT AmtVT = Amt.getValueType();
4358 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4359 DAG.getConstant(BitWidth, AmtVT), Amt);
4360 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4361 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4362 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4363 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4364 DAG.getConstant(-BitWidth, AmtVT));
4365 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4366 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4367 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4368 SDValue OutOps[] = { OutLo, OutHi };
4369 return DAG.getMergeValues(OutOps, 2, dl);
4372 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4373 DebugLoc dl = Op.getDebugLoc();
4374 EVT VT = Op.getValueType();
4375 unsigned BitWidth = VT.getSizeInBits();
4376 assert(Op.getNumOperands() == 3 &&
4377 VT == Op.getOperand(1).getValueType() &&
4380 // Expand into a bunch of logical ops, followed by a select_cc.
4381 SDValue Lo = Op.getOperand(0);
4382 SDValue Hi = Op.getOperand(1);
4383 SDValue Amt = Op.getOperand(2);
4384 EVT AmtVT = Amt.getValueType();
4386 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4387 DAG.getConstant(BitWidth, AmtVT), Amt);
4388 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4389 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4390 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4391 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4392 DAG.getConstant(-BitWidth, AmtVT));
4393 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4394 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4395 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4396 Tmp4, Tmp6, ISD::SETLE);
4397 SDValue OutOps[] = { OutLo, OutHi };
4398 return DAG.getMergeValues(OutOps, 2, dl);
4401 //===----------------------------------------------------------------------===//
4402 // Vector related lowering.
4405 /// BuildSplatI - Build a canonical splati of Val with an element size of
4406 /// SplatSize. Cast the result to VT.
4407 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4408 SelectionDAG &DAG, DebugLoc dl) {
4409 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4411 static const EVT VTys[] = { // canonical VT to use for each size.
4412 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4415 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4417 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4421 EVT CanonicalVT = VTys[SplatSize-1];
4423 // Build a canonical splat for this value.
4424 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4425 SmallVector<SDValue, 8> Ops;
4426 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4427 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4428 &Ops[0], Ops.size());
4429 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4432 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4433 /// specified intrinsic ID.
4434 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4435 SelectionDAG &DAG, DebugLoc dl,
4436 EVT DestVT = MVT::Other) {
4437 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4438 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4439 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4442 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4443 /// specified intrinsic ID.
4444 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4445 SDValue Op2, SelectionDAG &DAG,
4446 DebugLoc dl, EVT DestVT = MVT::Other) {
4447 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4448 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4449 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4453 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4454 /// amount. The result has the specified value type.
4455 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4456 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4457 // Force LHS/RHS to be the right type.
4458 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4459 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4462 for (unsigned i = 0; i != 16; ++i)
4464 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4465 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4468 // If this is a case we can't handle, return null and let the default
4469 // expansion code take care of it. If we CAN select this case, and if it
4470 // selects to a single instruction, return Op. Otherwise, if we can codegen
4471 // this case more efficiently than a constant pool load, lower it to the
4472 // sequence of ops that should be used.
4473 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4474 SelectionDAG &DAG) const {
4475 DebugLoc dl = Op.getDebugLoc();
4476 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4477 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4479 // Check if this is a splat of a constant value.
4480 APInt APSplatBits, APSplatUndef;
4481 unsigned SplatBitSize;
4483 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4484 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4487 unsigned SplatBits = APSplatBits.getZExtValue();
4488 unsigned SplatUndef = APSplatUndef.getZExtValue();
4489 unsigned SplatSize = SplatBitSize / 8;
4491 // First, handle single instruction cases.
4494 if (SplatBits == 0) {
4495 // Canonicalize all zero vectors to be v4i32.
4496 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4497 SDValue Z = DAG.getConstant(0, MVT::i32);
4498 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4499 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4504 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4505 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4507 if (SextVal >= -16 && SextVal <= 15)
4508 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4511 // Two instruction sequences.
4513 // If this value is in the range [-32,30] and is even, use:
4514 // tmp = VSPLTI[bhw], result = add tmp, tmp
4515 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4516 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4517 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4518 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4521 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4522 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4524 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4525 // Make -1 and vspltisw -1:
4526 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4528 // Make the VSLW intrinsic, computing 0x8000_0000.
4529 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4532 // xor by OnesV to invert it.
4533 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4534 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4537 // Check to see if this is a wide variety of vsplti*, binop self cases.
4538 static const signed char SplatCsts[] = {
4539 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4540 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4543 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4544 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4545 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4546 int i = SplatCsts[idx];
4548 // Figure out what shift amount will be used by altivec if shifted by i in
4550 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4552 // vsplti + shl self.
4553 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4554 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4555 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4556 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4557 Intrinsic::ppc_altivec_vslw
4559 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4560 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4563 // vsplti + srl self.
4564 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4565 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4566 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4567 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4568 Intrinsic::ppc_altivec_vsrw
4570 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4571 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4574 // vsplti + sra self.
4575 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4576 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4577 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4578 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4579 Intrinsic::ppc_altivec_vsraw
4581 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4582 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4585 // vsplti + rol self.
4586 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4587 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4588 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4589 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4590 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4591 Intrinsic::ppc_altivec_vrlw
4593 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4594 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4597 // t = vsplti c, result = vsldoi t, t, 1
4598 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
4599 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4600 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4602 // t = vsplti c, result = vsldoi t, t, 2
4603 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
4604 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4605 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4607 // t = vsplti c, result = vsldoi t, t, 3
4608 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4609 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4610 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4614 // Three instruction sequences.
4616 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4617 if (SextVal >= 0 && SextVal <= 31) {
4618 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4619 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4620 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4621 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4623 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4624 if (SextVal >= -31 && SextVal <= 0) {
4625 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4626 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4627 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4628 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4634 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4635 /// the specified operations to build the shuffle.
4636 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4637 SDValue RHS, SelectionDAG &DAG,
4639 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4640 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4641 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4644 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4656 if (OpNum == OP_COPY) {
4657 if (LHSID == (1*9+2)*9+3) return LHS;
4658 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4662 SDValue OpLHS, OpRHS;
4663 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4664 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4668 default: llvm_unreachable("Unknown i32 permute!");
4670 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4671 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4672 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4673 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4676 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4677 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4678 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4679 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4682 for (unsigned i = 0; i != 16; ++i)
4683 ShufIdxs[i] = (i&3)+0;
4686 for (unsigned i = 0; i != 16; ++i)
4687 ShufIdxs[i] = (i&3)+4;
4690 for (unsigned i = 0; i != 16; ++i)
4691 ShufIdxs[i] = (i&3)+8;
4694 for (unsigned i = 0; i != 16; ++i)
4695 ShufIdxs[i] = (i&3)+12;
4698 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4700 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4702 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4704 EVT VT = OpLHS.getValueType();
4705 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4706 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4707 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4708 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4711 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4712 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4713 /// return the code it can be lowered into. Worst case, it can always be
4714 /// lowered into a vperm.
4715 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4716 SelectionDAG &DAG) const {
4717 DebugLoc dl = Op.getDebugLoc();
4718 SDValue V1 = Op.getOperand(0);
4719 SDValue V2 = Op.getOperand(1);
4720 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4721 EVT VT = Op.getValueType();
4723 // Cases that are handled by instructions that take permute immediates
4724 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4725 // selected by the instruction selector.
4726 if (V2.getOpcode() == ISD::UNDEF) {
4727 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4728 PPC::isSplatShuffleMask(SVOp, 2) ||
4729 PPC::isSplatShuffleMask(SVOp, 4) ||
4730 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4731 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4732 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4733 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4734 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4735 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4736 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4737 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4738 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4743 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4744 // and produce a fixed permutation. If any of these match, do not lower to
4746 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4747 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4748 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4749 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4750 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4751 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4752 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4753 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4754 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4757 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4758 // perfect shuffle table to emit an optimal matching sequence.
4759 ArrayRef<int> PermMask = SVOp->getMask();
4761 unsigned PFIndexes[4];
4762 bool isFourElementShuffle = true;
4763 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4764 unsigned EltNo = 8; // Start out undef.
4765 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4766 if (PermMask[i*4+j] < 0)
4767 continue; // Undef, ignore it.
4769 unsigned ByteSource = PermMask[i*4+j];
4770 if ((ByteSource & 3) != j) {
4771 isFourElementShuffle = false;
4776 EltNo = ByteSource/4;
4777 } else if (EltNo != ByteSource/4) {
4778 isFourElementShuffle = false;
4782 PFIndexes[i] = EltNo;
4785 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4786 // perfect shuffle vector to determine if it is cost effective to do this as
4787 // discrete instructions, or whether we should use a vperm.
4788 if (isFourElementShuffle) {
4789 // Compute the index in the perfect shuffle table.
4790 unsigned PFTableIndex =
4791 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4793 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4794 unsigned Cost = (PFEntry >> 30);
4796 // Determining when to avoid vperm is tricky. Many things affect the cost
4797 // of vperm, particularly how many times the perm mask needs to be computed.
4798 // For example, if the perm mask can be hoisted out of a loop or is already
4799 // used (perhaps because there are multiple permutes with the same shuffle
4800 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4801 // the loop requires an extra register.
4803 // As a compromise, we only emit discrete instructions if the shuffle can be
4804 // generated in 3 or fewer operations. When we have loop information
4805 // available, if this block is within a loop, we should avoid using vperm
4806 // for 3-operation perms and use a constant pool load instead.
4808 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4811 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4812 // vector that will get spilled to the constant pool.
4813 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4815 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4816 // that it is in input element units, not in bytes. Convert now.
4817 EVT EltVT = V1.getValueType().getVectorElementType();
4818 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4820 SmallVector<SDValue, 16> ResultMask;
4821 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4822 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4824 for (unsigned j = 0; j != BytesPerElement; ++j)
4825 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4829 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4830 &ResultMask[0], ResultMask.size());
4831 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4834 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4835 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4836 /// information about the intrinsic.
4837 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4839 unsigned IntrinsicID =
4840 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4843 switch (IntrinsicID) {
4844 default: return false;
4845 // Comparison predicates.
4846 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4847 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4848 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4849 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4850 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4851 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4852 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4853 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4854 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4855 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4856 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4857 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4858 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4860 // Normal Comparisons.
4861 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4862 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4863 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4864 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4865 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4866 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4867 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4868 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4869 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4870 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4871 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4872 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4873 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4878 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4879 /// lower, do it, otherwise return null.
4880 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4881 SelectionDAG &DAG) const {
4882 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4883 // opcode number of the comparison.
4884 DebugLoc dl = Op.getDebugLoc();
4887 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4888 return SDValue(); // Don't custom lower most intrinsics.
4890 // If this is a non-dot comparison, make the VCMP node and we are done.
4892 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4893 Op.getOperand(1), Op.getOperand(2),
4894 DAG.getConstant(CompareOpc, MVT::i32));
4895 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4898 // Create the PPCISD altivec 'dot' comparison node.
4900 Op.getOperand(2), // LHS
4901 Op.getOperand(3), // RHS
4902 DAG.getConstant(CompareOpc, MVT::i32)
4904 std::vector<EVT> VTs;
4905 VTs.push_back(Op.getOperand(2).getValueType());
4906 VTs.push_back(MVT::Glue);
4907 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4909 // Now that we have the comparison, emit a copy from the CR to a GPR.
4910 // This is flagged to the above dot comparison.
4911 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4912 DAG.getRegister(PPC::CR6, MVT::i32),
4913 CompNode.getValue(1));
4915 // Unpack the result based on how the target uses it.
4916 unsigned BitNo; // Bit # of CR6.
4917 bool InvertBit; // Invert result?
4918 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4919 default: // Can't happen, don't crash on invalid number though.
4920 case 0: // Return the value of the EQ bit of CR6.
4921 BitNo = 0; InvertBit = false;
4923 case 1: // Return the inverted value of the EQ bit of CR6.
4924 BitNo = 0; InvertBit = true;
4926 case 2: // Return the value of the LT bit of CR6.
4927 BitNo = 2; InvertBit = false;
4929 case 3: // Return the inverted value of the LT bit of CR6.
4930 BitNo = 2; InvertBit = true;
4934 // Shift the bit into the low position.
4935 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4936 DAG.getConstant(8-(3-BitNo), MVT::i32));
4938 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4939 DAG.getConstant(1, MVT::i32));
4941 // If we are supposed to, toggle the bit.
4943 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4944 DAG.getConstant(1, MVT::i32));
4948 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4949 SelectionDAG &DAG) const {
4950 DebugLoc dl = Op.getDebugLoc();
4951 // Create a stack slot that is 16-byte aligned.
4952 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4953 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4954 EVT PtrVT = getPointerTy();
4955 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4957 // Store the input value into Value#0 of the stack slot.
4958 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4959 Op.getOperand(0), FIdx, MachinePointerInfo(),
4962 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4963 false, false, false, 0);
4966 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4967 DebugLoc dl = Op.getDebugLoc();
4968 if (Op.getValueType() == MVT::v4i32) {
4969 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4971 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4972 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4974 SDValue RHSSwap = // = vrlw RHS, 16
4975 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4977 // Shrinkify inputs to v8i16.
4978 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4979 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4980 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4982 // Low parts multiplied together, generating 32-bit results (we ignore the
4984 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4985 LHS, RHS, DAG, dl, MVT::v4i32);
4987 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4988 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4989 // Shift the high parts up 16 bits.
4990 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4992 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4993 } else if (Op.getValueType() == MVT::v8i16) {
4994 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4996 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4998 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4999 LHS, RHS, Zero, DAG, dl);
5000 } else if (Op.getValueType() == MVT::v16i8) {
5001 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5003 // Multiply the even 8-bit parts, producing 16-bit sums.
5004 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5005 LHS, RHS, DAG, dl, MVT::v8i16);
5006 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5008 // Multiply the odd 8-bit parts, producing 16-bit sums.
5009 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5010 LHS, RHS, DAG, dl, MVT::v8i16);
5011 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5013 // Merge the results together.
5015 for (unsigned i = 0; i != 8; ++i) {
5017 Ops[i*2+1] = 2*i+1+16;
5019 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5021 llvm_unreachable("Unknown mul to lower!");
5025 /// LowerOperation - Provide custom lowering hooks for some operations.
5027 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5028 switch (Op.getOpcode()) {
5029 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5030 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5031 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5032 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5033 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5034 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5035 case ISD::SETCC: return LowerSETCC(Op, DAG);
5036 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5037 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5039 return LowerVASTART(Op, DAG, PPCSubTarget);
5042 return LowerVAARG(Op, DAG, PPCSubTarget);
5044 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5045 case ISD::DYNAMIC_STACKALLOC:
5046 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5048 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5049 case ISD::FP_TO_UINT:
5050 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5052 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5053 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5055 // Lower 64-bit shifts.
5056 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5057 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5058 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5060 // Vector-related lowering.
5061 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5062 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5063 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5064 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5065 case ISD::MUL: return LowerMUL(Op, DAG);
5067 // Frame & Return address.
5068 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5069 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5073 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5074 SmallVectorImpl<SDValue>&Results,
5075 SelectionDAG &DAG) const {
5076 const TargetMachine &TM = getTargetMachine();
5077 DebugLoc dl = N->getDebugLoc();
5078 switch (N->getOpcode()) {
5080 llvm_unreachable("Do not know how to custom type legalize this operation!");
5082 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5083 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5086 EVT VT = N->getValueType(0);
5088 if (VT == MVT::i64) {
5089 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5091 Results.push_back(NewNode);
5092 Results.push_back(NewNode.getValue(1));
5096 case ISD::FP_ROUND_INREG: {
5097 assert(N->getValueType(0) == MVT::ppcf128);
5098 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5099 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5100 MVT::f64, N->getOperand(0),
5101 DAG.getIntPtrConstant(0));
5102 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5103 MVT::f64, N->getOperand(0),
5104 DAG.getIntPtrConstant(1));
5106 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5107 // of the long double, and puts FPSCR back the way it was. We do not
5108 // actually model FPSCR.
5109 std::vector<EVT> NodeTys;
5110 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5112 NodeTys.push_back(MVT::f64); // Return register
5113 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5114 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5115 MFFSreg = Result.getValue(0);
5116 InFlag = Result.getValue(1);
5119 NodeTys.push_back(MVT::Glue); // Returns a flag
5120 Ops[0] = DAG.getConstant(31, MVT::i32);
5122 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5123 InFlag = Result.getValue(0);
5126 NodeTys.push_back(MVT::Glue); // Returns a flag
5127 Ops[0] = DAG.getConstant(30, MVT::i32);
5129 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5130 InFlag = Result.getValue(0);
5133 NodeTys.push_back(MVT::f64); // result of add
5134 NodeTys.push_back(MVT::Glue); // Returns a flag
5138 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5139 FPreg = Result.getValue(0);
5140 InFlag = Result.getValue(1);
5143 NodeTys.push_back(MVT::f64);
5144 Ops[0] = DAG.getConstant(1, MVT::i32);
5148 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5149 FPreg = Result.getValue(0);
5151 // We know the low half is about to be thrown away, so just use something
5153 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5157 case ISD::FP_TO_SINT:
5158 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5164 //===----------------------------------------------------------------------===//
5165 // Other Lowering Code
5166 //===----------------------------------------------------------------------===//
5169 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5170 bool is64bit, unsigned BinOpcode) const {
5171 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5172 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5174 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5175 MachineFunction *F = BB->getParent();
5176 MachineFunction::iterator It = BB;
5179 unsigned dest = MI->getOperand(0).getReg();
5180 unsigned ptrA = MI->getOperand(1).getReg();
5181 unsigned ptrB = MI->getOperand(2).getReg();
5182 unsigned incr = MI->getOperand(3).getReg();
5183 DebugLoc dl = MI->getDebugLoc();
5185 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5186 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5187 F->insert(It, loopMBB);
5188 F->insert(It, exitMBB);
5189 exitMBB->splice(exitMBB->begin(), BB,
5190 llvm::next(MachineBasicBlock::iterator(MI)),
5192 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5194 MachineRegisterInfo &RegInfo = F->getRegInfo();
5195 unsigned TmpReg = (!BinOpcode) ? incr :
5196 RegInfo.createVirtualRegister(
5197 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5198 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5202 // fallthrough --> loopMBB
5203 BB->addSuccessor(loopMBB);
5206 // l[wd]arx dest, ptr
5207 // add r0, dest, incr
5208 // st[wd]cx. r0, ptr
5210 // fallthrough --> exitMBB
5212 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5213 .addReg(ptrA).addReg(ptrB);
5215 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5216 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5217 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5218 BuildMI(BB, dl, TII->get(PPC::BCC))
5219 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5220 BB->addSuccessor(loopMBB);
5221 BB->addSuccessor(exitMBB);
5230 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5231 MachineBasicBlock *BB,
5232 bool is8bit, // operation
5233 unsigned BinOpcode) const {
5234 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5235 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5236 // In 64 bit mode we have to use 64 bits for addresses, even though the
5237 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5238 // registers without caring whether they're 32 or 64, but here we're
5239 // doing actual arithmetic on the addresses.
5240 bool is64bit = PPCSubTarget.isPPC64();
5241 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5243 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5244 MachineFunction *F = BB->getParent();
5245 MachineFunction::iterator It = BB;
5248 unsigned dest = MI->getOperand(0).getReg();
5249 unsigned ptrA = MI->getOperand(1).getReg();
5250 unsigned ptrB = MI->getOperand(2).getReg();
5251 unsigned incr = MI->getOperand(3).getReg();
5252 DebugLoc dl = MI->getDebugLoc();
5254 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5255 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5256 F->insert(It, loopMBB);
5257 F->insert(It, exitMBB);
5258 exitMBB->splice(exitMBB->begin(), BB,
5259 llvm::next(MachineBasicBlock::iterator(MI)),
5261 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5263 MachineRegisterInfo &RegInfo = F->getRegInfo();
5264 const TargetRegisterClass *RC =
5265 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5266 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5267 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5268 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5269 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5270 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5271 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5272 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5273 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5274 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5275 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5276 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5277 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5279 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5283 // fallthrough --> loopMBB
5284 BB->addSuccessor(loopMBB);
5286 // The 4-byte load must be aligned, while a char or short may be
5287 // anywhere in the word. Hence all this nasty bookkeeping code.
5288 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5289 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5290 // xori shift, shift1, 24 [16]
5291 // rlwinm ptr, ptr1, 0, 0, 29
5292 // slw incr2, incr, shift
5293 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5294 // slw mask, mask2, shift
5296 // lwarx tmpDest, ptr
5297 // add tmp, tmpDest, incr2
5298 // andc tmp2, tmpDest, mask
5299 // and tmp3, tmp, mask
5300 // or tmp4, tmp3, tmp2
5303 // fallthrough --> exitMBB
5304 // srw dest, tmpDest, shift
5305 if (ptrA != ZeroReg) {
5306 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5307 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5308 .addReg(ptrA).addReg(ptrB);
5312 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5313 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5314 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5315 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5317 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5318 .addReg(Ptr1Reg).addImm(0).addImm(61);
5320 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5321 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5322 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5323 .addReg(incr).addReg(ShiftReg);
5325 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5327 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5328 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5330 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5331 .addReg(Mask2Reg).addReg(ShiftReg);
5334 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5335 .addReg(ZeroReg).addReg(PtrReg);
5337 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5338 .addReg(Incr2Reg).addReg(TmpDestReg);
5339 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5340 .addReg(TmpDestReg).addReg(MaskReg);
5341 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5342 .addReg(TmpReg).addReg(MaskReg);
5343 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5344 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5345 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5346 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5347 BuildMI(BB, dl, TII->get(PPC::BCC))
5348 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5349 BB->addSuccessor(loopMBB);
5350 BB->addSuccessor(exitMBB);
5355 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5361 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5362 MachineBasicBlock *BB) const {
5363 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5365 // To "insert" these instructions we actually have to insert their
5366 // control-flow patterns.
5367 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5368 MachineFunction::iterator It = BB;
5371 MachineFunction *F = BB->getParent();
5373 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5374 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5375 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5376 PPC::ISEL8 : PPC::ISEL;
5377 unsigned SelectPred = MI->getOperand(4).getImm();
5378 DebugLoc dl = MI->getDebugLoc();
5380 // The SelectPred is ((BI << 5) | BO) for a BCC
5381 unsigned BO = SelectPred & 0xF;
5382 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5384 unsigned TrueOpNo, FalseOpNo;
5391 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5394 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5395 .addReg(MI->getOperand(TrueOpNo).getReg())
5396 .addReg(MI->getOperand(FalseOpNo).getReg())
5397 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5398 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5399 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5400 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5401 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5402 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5405 // The incoming instruction knows the destination vreg to set, the
5406 // condition code register to branch on, the true/false values to
5407 // select between, and a branch opcode to use.
5412 // cmpTY ccX, r1, r2
5414 // fallthrough --> copy0MBB
5415 MachineBasicBlock *thisMBB = BB;
5416 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5417 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5418 unsigned SelectPred = MI->getOperand(4).getImm();
5419 DebugLoc dl = MI->getDebugLoc();
5420 F->insert(It, copy0MBB);
5421 F->insert(It, sinkMBB);
5423 // Transfer the remainder of BB and its successor edges to sinkMBB.
5424 sinkMBB->splice(sinkMBB->begin(), BB,
5425 llvm::next(MachineBasicBlock::iterator(MI)),
5427 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5429 // Next, add the true and fallthrough blocks as its successors.
5430 BB->addSuccessor(copy0MBB);
5431 BB->addSuccessor(sinkMBB);
5433 BuildMI(BB, dl, TII->get(PPC::BCC))
5434 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5437 // %FalseValue = ...
5438 // # fallthrough to sinkMBB
5441 // Update machine-CFG edges
5442 BB->addSuccessor(sinkMBB);
5445 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5448 BuildMI(*BB, BB->begin(), dl,
5449 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5450 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5451 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5453 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5454 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5455 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5456 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5457 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5458 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5459 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5460 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5462 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5463 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5464 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5465 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5466 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5467 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5468 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5469 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5471 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5472 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5473 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5474 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5475 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5476 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5477 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5478 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5480 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5481 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5482 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5483 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5484 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5485 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5486 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5487 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5489 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5490 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5491 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5492 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5493 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5494 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5495 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5496 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5498 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5499 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5500 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5501 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5502 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5503 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5504 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5505 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5507 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5508 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5509 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5510 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5511 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5512 BB = EmitAtomicBinary(MI, BB, false, 0);
5513 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5514 BB = EmitAtomicBinary(MI, BB, true, 0);
5516 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5517 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5518 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5520 unsigned dest = MI->getOperand(0).getReg();
5521 unsigned ptrA = MI->getOperand(1).getReg();
5522 unsigned ptrB = MI->getOperand(2).getReg();
5523 unsigned oldval = MI->getOperand(3).getReg();
5524 unsigned newval = MI->getOperand(4).getReg();
5525 DebugLoc dl = MI->getDebugLoc();
5527 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5528 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5529 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5530 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5531 F->insert(It, loop1MBB);
5532 F->insert(It, loop2MBB);
5533 F->insert(It, midMBB);
5534 F->insert(It, exitMBB);
5535 exitMBB->splice(exitMBB->begin(), BB,
5536 llvm::next(MachineBasicBlock::iterator(MI)),
5538 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5542 // fallthrough --> loopMBB
5543 BB->addSuccessor(loop1MBB);
5546 // l[wd]arx dest, ptr
5547 // cmp[wd] dest, oldval
5550 // st[wd]cx. newval, ptr
5554 // st[wd]cx. dest, ptr
5557 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5558 .addReg(ptrA).addReg(ptrB);
5559 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5560 .addReg(oldval).addReg(dest);
5561 BuildMI(BB, dl, TII->get(PPC::BCC))
5562 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5563 BB->addSuccessor(loop2MBB);
5564 BB->addSuccessor(midMBB);
5567 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5568 .addReg(newval).addReg(ptrA).addReg(ptrB);
5569 BuildMI(BB, dl, TII->get(PPC::BCC))
5570 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5571 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5572 BB->addSuccessor(loop1MBB);
5573 BB->addSuccessor(exitMBB);
5576 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5577 .addReg(dest).addReg(ptrA).addReg(ptrB);
5578 BB->addSuccessor(exitMBB);
5583 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5584 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5585 // We must use 64-bit registers for addresses when targeting 64-bit,
5586 // since we're actually doing arithmetic on them. Other registers
5588 bool is64bit = PPCSubTarget.isPPC64();
5589 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5591 unsigned dest = MI->getOperand(0).getReg();
5592 unsigned ptrA = MI->getOperand(1).getReg();
5593 unsigned ptrB = MI->getOperand(2).getReg();
5594 unsigned oldval = MI->getOperand(3).getReg();
5595 unsigned newval = MI->getOperand(4).getReg();
5596 DebugLoc dl = MI->getDebugLoc();
5598 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5599 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5600 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5601 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5602 F->insert(It, loop1MBB);
5603 F->insert(It, loop2MBB);
5604 F->insert(It, midMBB);
5605 F->insert(It, exitMBB);
5606 exitMBB->splice(exitMBB->begin(), BB,
5607 llvm::next(MachineBasicBlock::iterator(MI)),
5609 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5611 MachineRegisterInfo &RegInfo = F->getRegInfo();
5612 const TargetRegisterClass *RC =
5613 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5614 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5615 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5616 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5617 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5618 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5619 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5620 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5621 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5622 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5623 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5624 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5625 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5626 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5627 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5629 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5630 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5633 // fallthrough --> loopMBB
5634 BB->addSuccessor(loop1MBB);
5636 // The 4-byte load must be aligned, while a char or short may be
5637 // anywhere in the word. Hence all this nasty bookkeeping code.
5638 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5639 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5640 // xori shift, shift1, 24 [16]
5641 // rlwinm ptr, ptr1, 0, 0, 29
5642 // slw newval2, newval, shift
5643 // slw oldval2, oldval,shift
5644 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5645 // slw mask, mask2, shift
5646 // and newval3, newval2, mask
5647 // and oldval3, oldval2, mask
5649 // lwarx tmpDest, ptr
5650 // and tmp, tmpDest, mask
5651 // cmpw tmp, oldval3
5654 // andc tmp2, tmpDest, mask
5655 // or tmp4, tmp2, newval3
5660 // stwcx. tmpDest, ptr
5662 // srw dest, tmpDest, shift
5663 if (ptrA != ZeroReg) {
5664 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5665 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5666 .addReg(ptrA).addReg(ptrB);
5670 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5671 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5672 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5673 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5675 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5676 .addReg(Ptr1Reg).addImm(0).addImm(61);
5678 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5679 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5680 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5681 .addReg(newval).addReg(ShiftReg);
5682 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5683 .addReg(oldval).addReg(ShiftReg);
5685 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5687 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5688 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5689 .addReg(Mask3Reg).addImm(65535);
5691 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5692 .addReg(Mask2Reg).addReg(ShiftReg);
5693 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5694 .addReg(NewVal2Reg).addReg(MaskReg);
5695 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5696 .addReg(OldVal2Reg).addReg(MaskReg);
5699 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5700 .addReg(ZeroReg).addReg(PtrReg);
5701 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5702 .addReg(TmpDestReg).addReg(MaskReg);
5703 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5704 .addReg(TmpReg).addReg(OldVal3Reg);
5705 BuildMI(BB, dl, TII->get(PPC::BCC))
5706 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5707 BB->addSuccessor(loop2MBB);
5708 BB->addSuccessor(midMBB);
5711 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5712 .addReg(TmpDestReg).addReg(MaskReg);
5713 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5714 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5715 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5716 .addReg(ZeroReg).addReg(PtrReg);
5717 BuildMI(BB, dl, TII->get(PPC::BCC))
5718 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5719 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5720 BB->addSuccessor(loop1MBB);
5721 BB->addSuccessor(exitMBB);
5724 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5725 .addReg(ZeroReg).addReg(PtrReg);
5726 BB->addSuccessor(exitMBB);
5731 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5734 llvm_unreachable("Unexpected instr type to insert");
5737 MI->eraseFromParent(); // The pseudo instruction is gone now.
5741 //===----------------------------------------------------------------------===//
5742 // Target Optimization Hooks
5743 //===----------------------------------------------------------------------===//
5745 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5746 DAGCombinerInfo &DCI) const {
5747 const TargetMachine &TM = getTargetMachine();
5748 SelectionDAG &DAG = DCI.DAG;
5749 DebugLoc dl = N->getDebugLoc();
5750 switch (N->getOpcode()) {
5753 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5754 if (C->isNullValue()) // 0 << V -> 0.
5755 return N->getOperand(0);
5759 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5760 if (C->isNullValue()) // 0 >>u V -> 0.
5761 return N->getOperand(0);
5765 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5766 if (C->isNullValue() || // 0 >>s V -> 0.
5767 C->isAllOnesValue()) // -1 >>s V -> -1.
5768 return N->getOperand(0);
5772 case ISD::SINT_TO_FP:
5773 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5774 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5775 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5776 // We allow the src/dst to be either f32/f64, but the intermediate
5777 // type must be i64.
5778 if (N->getOperand(0).getValueType() == MVT::i64 &&
5779 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5780 SDValue Val = N->getOperand(0).getOperand(0);
5781 if (Val.getValueType() == MVT::f32) {
5782 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5783 DCI.AddToWorklist(Val.getNode());
5786 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5787 DCI.AddToWorklist(Val.getNode());
5788 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5789 DCI.AddToWorklist(Val.getNode());
5790 if (N->getValueType(0) == MVT::f32) {
5791 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5792 DAG.getIntPtrConstant(0));
5793 DCI.AddToWorklist(Val.getNode());
5796 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5797 // If the intermediate type is i32, we can avoid the load/store here
5804 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5805 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5806 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5807 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5808 N->getOperand(1).getValueType() == MVT::i32 &&
5809 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5810 SDValue Val = N->getOperand(1).getOperand(0);
5811 if (Val.getValueType() == MVT::f32) {
5812 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5813 DCI.AddToWorklist(Val.getNode());
5815 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5816 DCI.AddToWorklist(Val.getNode());
5818 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5819 N->getOperand(2), N->getOperand(3));
5820 DCI.AddToWorklist(Val.getNode());
5824 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5825 if (cast<StoreSDNode>(N)->isUnindexed() &&
5826 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5827 N->getOperand(1).getNode()->hasOneUse() &&
5828 (N->getOperand(1).getValueType() == MVT::i32 ||
5829 N->getOperand(1).getValueType() == MVT::i16)) {
5830 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5831 // Do an any-extend to 32-bits if this is a half-word input.
5832 if (BSwapOp.getValueType() == MVT::i16)
5833 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5836 N->getOperand(0), BSwapOp, N->getOperand(2),
5837 DAG.getValueType(N->getOperand(1).getValueType())
5840 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5841 Ops, array_lengthof(Ops),
5842 cast<StoreSDNode>(N)->getMemoryVT(),
5843 cast<StoreSDNode>(N)->getMemOperand());
5847 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5848 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5849 N->getOperand(0).hasOneUse() &&
5850 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5851 SDValue Load = N->getOperand(0);
5852 LoadSDNode *LD = cast<LoadSDNode>(Load);
5853 // Create the byte-swapping load.
5855 LD->getChain(), // Chain
5856 LD->getBasePtr(), // Ptr
5857 DAG.getValueType(N->getValueType(0)) // VT
5860 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5861 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5862 LD->getMemoryVT(), LD->getMemOperand());
5864 // If this is an i16 load, insert the truncate.
5865 SDValue ResVal = BSLoad;
5866 if (N->getValueType(0) == MVT::i16)
5867 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5869 // First, combine the bswap away. This makes the value produced by the
5871 DCI.CombineTo(N, ResVal);
5873 // Next, combine the load away, we give it a bogus result value but a real
5874 // chain result. The result value is dead because the bswap is dead.
5875 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5877 // Return N so it doesn't get rechecked!
5878 return SDValue(N, 0);
5882 case PPCISD::VCMP: {
5883 // If a VCMPo node already exists with exactly the same operands as this
5884 // node, use its result instead of this node (VCMPo computes both a CR6 and
5885 // a normal output).
5887 if (!N->getOperand(0).hasOneUse() &&
5888 !N->getOperand(1).hasOneUse() &&
5889 !N->getOperand(2).hasOneUse()) {
5891 // Scan all of the users of the LHS, looking for VCMPo's that match.
5892 SDNode *VCMPoNode = 0;
5894 SDNode *LHSN = N->getOperand(0).getNode();
5895 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5897 if (UI->getOpcode() == PPCISD::VCMPo &&
5898 UI->getOperand(1) == N->getOperand(1) &&
5899 UI->getOperand(2) == N->getOperand(2) &&
5900 UI->getOperand(0) == N->getOperand(0)) {
5905 // If there is no VCMPo node, or if the flag value has a single use, don't
5907 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5910 // Look at the (necessarily single) use of the flag value. If it has a
5911 // chain, this transformation is more complex. Note that multiple things
5912 // could use the value result, which we should ignore.
5913 SDNode *FlagUser = 0;
5914 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5915 FlagUser == 0; ++UI) {
5916 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5918 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5919 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5926 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5927 // give up for right now.
5928 if (FlagUser->getOpcode() == PPCISD::MFCR)
5929 return SDValue(VCMPoNode, 0);
5934 // If this is a branch on an altivec predicate comparison, lower this so
5935 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5936 // lowering is done pre-legalize, because the legalizer lowers the predicate
5937 // compare down to code that is difficult to reassemble.
5938 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5939 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5943 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5944 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5945 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5946 assert(isDot && "Can't compare against a vector result!");
5948 // If this is a comparison against something other than 0/1, then we know
5949 // that the condition is never/always true.
5950 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5951 if (Val != 0 && Val != 1) {
5952 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5953 return N->getOperand(0);
5954 // Always !=, turn it into an unconditional branch.
5955 return DAG.getNode(ISD::BR, dl, MVT::Other,
5956 N->getOperand(0), N->getOperand(4));
5959 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5961 // Create the PPCISD altivec 'dot' comparison node.
5962 std::vector<EVT> VTs;
5964 LHS.getOperand(2), // LHS of compare
5965 LHS.getOperand(3), // RHS of compare
5966 DAG.getConstant(CompareOpc, MVT::i32)
5968 VTs.push_back(LHS.getOperand(2).getValueType());
5969 VTs.push_back(MVT::Glue);
5970 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5972 // Unpack the result based on how the target uses it.
5973 PPC::Predicate CompOpc;
5974 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5975 default: // Can't happen, don't crash on invalid number though.
5976 case 0: // Branch on the value of the EQ bit of CR6.
5977 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5979 case 1: // Branch on the inverted value of the EQ bit of CR6.
5980 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5982 case 2: // Branch on the value of the LT bit of CR6.
5983 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5985 case 3: // Branch on the inverted value of the LT bit of CR6.
5986 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5990 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5991 DAG.getConstant(CompOpc, MVT::i32),
5992 DAG.getRegister(PPC::CR6, MVT::i32),
5993 N->getOperand(4), CompNode.getValue(1));
6002 //===----------------------------------------------------------------------===//
6003 // Inline Assembly Support
6004 //===----------------------------------------------------------------------===//
6006 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6009 const SelectionDAG &DAG,
6010 unsigned Depth) const {
6011 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6012 switch (Op.getOpcode()) {
6014 case PPCISD::LBRX: {
6015 // lhbrx is known to have the top bits cleared out.
6016 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6017 KnownZero = 0xFFFF0000;
6020 case ISD::INTRINSIC_WO_CHAIN: {
6021 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6023 case Intrinsic::ppc_altivec_vcmpbfp_p:
6024 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6025 case Intrinsic::ppc_altivec_vcmpequb_p:
6026 case Intrinsic::ppc_altivec_vcmpequh_p:
6027 case Intrinsic::ppc_altivec_vcmpequw_p:
6028 case Intrinsic::ppc_altivec_vcmpgefp_p:
6029 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6030 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6031 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6032 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6033 case Intrinsic::ppc_altivec_vcmpgtub_p:
6034 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6035 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6036 KnownZero = ~1U; // All bits but the low one are known to be zero.
6044 /// getConstraintType - Given a constraint, return the type of
6045 /// constraint it is for this target.
6046 PPCTargetLowering::ConstraintType
6047 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6048 if (Constraint.size() == 1) {
6049 switch (Constraint[0]) {
6056 return C_RegisterClass;
6059 return TargetLowering::getConstraintType(Constraint);
6062 /// Examine constraint type and operand type and determine a weight value.
6063 /// This object must already have been set up with the operand type
6064 /// and the current alternative constraint selected.
6065 TargetLowering::ConstraintWeight
6066 PPCTargetLowering::getSingleConstraintMatchWeight(
6067 AsmOperandInfo &info, const char *constraint) const {
6068 ConstraintWeight weight = CW_Invalid;
6069 Value *CallOperandVal = info.CallOperandVal;
6070 // If we don't have a value, we can't do a match,
6071 // but allow it at the lowest weight.
6072 if (CallOperandVal == NULL)
6074 Type *type = CallOperandVal->getType();
6075 // Look at the constraint type.
6076 switch (*constraint) {
6078 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6081 if (type->isIntegerTy())
6082 weight = CW_Register;
6085 if (type->isFloatTy())
6086 weight = CW_Register;
6089 if (type->isDoubleTy())
6090 weight = CW_Register;
6093 if (type->isVectorTy())
6094 weight = CW_Register;
6097 weight = CW_Register;
6103 std::pair<unsigned, const TargetRegisterClass*>
6104 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6106 if (Constraint.size() == 1) {
6107 // GCC RS6000 Constraint Letters
6108 switch (Constraint[0]) {
6111 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6112 return std::make_pair(0U, &PPC::G8RCRegClass);
6113 return std::make_pair(0U, &PPC::GPRCRegClass);
6116 return std::make_pair(0U, &PPC::F4RCRegClass);
6118 return std::make_pair(0U, &PPC::F8RCRegClass);
6121 return std::make_pair(0U, &PPC::VRRCRegClass);
6123 return std::make_pair(0U, &PPC::CRRCRegClass);
6127 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6131 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6132 /// vector. If it is invalid, don't add anything to Ops.
6133 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6134 std::string &Constraint,
6135 std::vector<SDValue>&Ops,
6136 SelectionDAG &DAG) const {
6137 SDValue Result(0,0);
6139 // Only support length 1 constraints.
6140 if (Constraint.length() > 1) return;
6142 char Letter = Constraint[0];
6153 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6154 if (!CST) return; // Must be an immediate to match.
6155 unsigned Value = CST->getZExtValue();
6157 default: llvm_unreachable("Unknown constraint letter!");
6158 case 'I': // "I" is a signed 16-bit constant.
6159 if ((short)Value == (int)Value)
6160 Result = DAG.getTargetConstant(Value, Op.getValueType());
6162 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6163 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6164 if ((short)Value == 0)
6165 Result = DAG.getTargetConstant(Value, Op.getValueType());
6167 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6168 if ((Value >> 16) == 0)
6169 Result = DAG.getTargetConstant(Value, Op.getValueType());
6171 case 'M': // "M" is a constant that is greater than 31.
6173 Result = DAG.getTargetConstant(Value, Op.getValueType());
6175 case 'N': // "N" is a positive constant that is an exact power of two.
6176 if ((int)Value > 0 && isPowerOf2_32(Value))
6177 Result = DAG.getTargetConstant(Value, Op.getValueType());
6179 case 'O': // "O" is the constant zero.
6181 Result = DAG.getTargetConstant(Value, Op.getValueType());
6183 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6184 if ((short)-Value == (int)-Value)
6185 Result = DAG.getTargetConstant(Value, Op.getValueType());
6192 if (Result.getNode()) {
6193 Ops.push_back(Result);
6197 // Handle standard constraint letters.
6198 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6201 // isLegalAddressingMode - Return true if the addressing mode represented
6202 // by AM is legal for this target, for a load/store of the specified type.
6203 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6205 // FIXME: PPC does not allow r+i addressing modes for vectors!
6207 // PPC allows a sign-extended 16-bit immediate field.
6208 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6211 // No global is ever allowed as a base.
6215 // PPC only support r+r,
6217 case 0: // "r+i" or just "i", depending on HasBaseReg.
6220 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6222 // Otherwise we have r+r or r+i.
6225 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6227 // Allow 2*r as r+r.
6230 // No other scales are supported.
6237 /// isLegalAddressImmediate - Return true if the integer value can be used
6238 /// as the offset of the target addressing mode for load / store of the
6240 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6241 // PPC allows a sign-extended 16-bit immediate field.
6242 return (V > -(1 << 16) && V < (1 << 16)-1);
6245 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6249 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6250 SelectionDAG &DAG) const {
6251 MachineFunction &MF = DAG.getMachineFunction();
6252 MachineFrameInfo *MFI = MF.getFrameInfo();
6253 MFI->setReturnAddressIsTaken(true);
6255 DebugLoc dl = Op.getDebugLoc();
6256 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6258 // Make sure the function does not optimize away the store of the RA to
6260 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6261 FuncInfo->setLRStoreRequired();
6262 bool isPPC64 = PPCSubTarget.isPPC64();
6263 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6266 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6269 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6270 isPPC64? MVT::i64 : MVT::i32);
6271 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6272 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6274 MachinePointerInfo(), false, false, false, 0);
6277 // Just load the return address off the stack.
6278 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6279 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6280 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6283 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6284 SelectionDAG &DAG) const {
6285 DebugLoc dl = Op.getDebugLoc();
6286 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6288 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6289 bool isPPC64 = PtrVT == MVT::i64;
6291 MachineFunction &MF = DAG.getMachineFunction();
6292 MachineFrameInfo *MFI = MF.getFrameInfo();
6293 MFI->setFrameAddressIsTaken(true);
6294 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6295 MFI->hasVarSizedObjects()) &&
6296 MFI->getStackSize() &&
6297 !MF.getFunction()->getFnAttributes().
6298 hasAttribute(Attributes::Naked);
6299 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6300 (is31 ? PPC::R31 : PPC::R1);
6301 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6304 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6305 FrameAddr, MachinePointerInfo(), false, false,
6311 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6312 // The PowerPC target isn't yet aware of offsets.
6316 /// getOptimalMemOpType - Returns the target specific optimal type for load
6317 /// and store operations as a result of memset, memcpy, and memmove
6318 /// lowering. If DstAlign is zero that means it's safe to destination
6319 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6320 /// means there isn't a need to check it against alignment requirement,
6321 /// probably because the source does not need to be loaded. If
6322 /// 'IsZeroVal' is true, that means it's safe to return a
6323 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6324 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6325 /// constant so it does not need to be loaded.
6326 /// It returns EVT::Other if the type should be determined using generic
6327 /// target-independent logic.
6328 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6329 unsigned DstAlign, unsigned SrcAlign,
6332 MachineFunction &MF) const {
6333 if (this->PPCSubTarget.isPPC64()) {
6340 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6341 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6342 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6343 /// is expanded to mul + add.
6344 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6348 switch (VT.getSimpleVT().SimpleTy) {
6360 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6362 return TargetLowering::getSchedulingPreference(N);