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);
2153 // Note that vector arguments in registers don't reserve stack space,
2154 // except in varargs functions.
2155 if (VR_idx != Num_VR_Regs) {
2156 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2157 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2159 while ((ArgOffset % 16) != 0) {
2160 ArgOffset += PtrByteSize;
2161 if (GPR_idx != Num_GPR_Regs)
2165 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2169 // Vectors are aligned.
2170 ArgOffset = ((ArgOffset+15)/16)*16;
2171 CurArgOffset = ArgOffset;
2178 // We need to load the argument to a virtual register if we determined
2179 // above that we ran out of physical registers of the appropriate type.
2181 int FI = MFI->CreateFixedObject(ObjSize,
2182 CurArgOffset + (ArgSize - ObjSize),
2184 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2185 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2186 false, false, false, 0);
2189 InVals.push_back(ArgVal);
2192 // Set the size that is at least reserved in caller of this function. Tail
2193 // call optimized function's reserved stack space needs to be aligned so that
2194 // taking the difference between two stack areas will result in an aligned
2196 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2197 // Add the Altivec parameters at the end, if needed.
2198 if (nAltivecParamsAtEnd) {
2199 MinReservedArea = ((MinReservedArea+15)/16)*16;
2200 MinReservedArea += 16*nAltivecParamsAtEnd;
2203 std::max(MinReservedArea,
2204 PPCFrameLowering::getMinCallFrameSize(true, true));
2205 unsigned TargetAlign
2206 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2207 getStackAlignment();
2208 unsigned AlignMask = TargetAlign-1;
2209 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2210 FI->setMinReservedArea(MinReservedArea);
2212 // If the function takes variable number of arguments, make a frame index for
2213 // the start of the first vararg value... for expansion of llvm.va_start.
2215 int Depth = ArgOffset;
2217 FuncInfo->setVarArgsFrameIndex(
2218 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2220 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2222 // If this function is vararg, store any remaining integer argument regs
2223 // to their spots on the stack so that they may be loaded by deferencing the
2224 // result of va_next.
2225 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2226 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2227 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2228 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2229 MachinePointerInfo(), false, false, 0);
2230 MemOps.push_back(Store);
2231 // Increment the address by four for the next argument to store
2232 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2233 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2237 if (!MemOps.empty())
2238 Chain = DAG.getNode(ISD::TokenFactor, dl,
2239 MVT::Other, &MemOps[0], MemOps.size());
2245 PPCTargetLowering::LowerFormalArguments_Darwin(
2247 CallingConv::ID CallConv, bool isVarArg,
2248 const SmallVectorImpl<ISD::InputArg>
2250 DebugLoc dl, SelectionDAG &DAG,
2251 SmallVectorImpl<SDValue> &InVals) const {
2252 // TODO: add description of PPC stack frame format, or at least some docs.
2254 MachineFunction &MF = DAG.getMachineFunction();
2255 MachineFrameInfo *MFI = MF.getFrameInfo();
2256 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2258 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2259 bool isPPC64 = PtrVT == MVT::i64;
2260 // Potential tail calls could cause overwriting of argument stack slots.
2261 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2262 (CallConv == CallingConv::Fast));
2263 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2265 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2266 // Area that is at least reserved in caller of this function.
2267 unsigned MinReservedArea = ArgOffset;
2269 static const uint16_t GPR_32[] = { // 32-bit registers.
2270 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2271 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2273 static const uint16_t GPR_64[] = { // 64-bit registers.
2274 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2275 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2278 static const uint16_t *FPR = GetFPR();
2280 static const uint16_t VR[] = {
2281 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2282 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2285 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2286 const unsigned Num_FPR_Regs = 13;
2287 const unsigned Num_VR_Regs = array_lengthof( VR);
2289 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2291 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2293 // In 32-bit non-varargs functions, the stack space for vectors is after the
2294 // stack space for non-vectors. We do not use this space unless we have
2295 // too many vectors to fit in registers, something that only occurs in
2296 // constructed examples:), but we have to walk the arglist to figure
2297 // that out...for the pathological case, compute VecArgOffset as the
2298 // start of the vector parameter area. Computing VecArgOffset is the
2299 // entire point of the following loop.
2300 unsigned VecArgOffset = ArgOffset;
2301 if (!isVarArg && !isPPC64) {
2302 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2304 EVT ObjectVT = Ins[ArgNo].VT;
2305 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2307 if (Flags.isByVal()) {
2308 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2309 unsigned ObjSize = Flags.getByValSize();
2311 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2312 VecArgOffset += ArgSize;
2316 switch(ObjectVT.getSimpleVT().SimpleTy) {
2317 default: llvm_unreachable("Unhandled argument type!");
2322 case MVT::i64: // PPC64
2324 // FIXME: We are guaranteed to be !isPPC64 at this point.
2325 // Does MVT::i64 apply?
2332 // Nothing to do, we're only looking at Nonvector args here.
2337 // We've found where the vector parameter area in memory is. Skip the
2338 // first 12 parameters; these don't use that memory.
2339 VecArgOffset = ((VecArgOffset+15)/16)*16;
2340 VecArgOffset += 12*16;
2342 // Add DAG nodes to load the arguments or copy them out of registers. On
2343 // entry to a function on PPC, the arguments start after the linkage area,
2344 // although the first ones are often in registers.
2346 SmallVector<SDValue, 8> MemOps;
2347 unsigned nAltivecParamsAtEnd = 0;
2348 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2349 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2351 bool needsLoad = false;
2352 EVT ObjectVT = Ins[ArgNo].VT;
2353 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2354 unsigned ArgSize = ObjSize;
2355 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2357 unsigned CurArgOffset = ArgOffset;
2359 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2360 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2361 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2362 if (isVarArg || isPPC64) {
2363 MinReservedArea = ((MinReservedArea+15)/16)*16;
2364 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2367 } else nAltivecParamsAtEnd++;
2369 // Calculate min reserved area.
2370 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2374 // FIXME the codegen can be much improved in some cases.
2375 // We do not have to keep everything in memory.
2376 if (Flags.isByVal()) {
2377 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2378 ObjSize = Flags.getByValSize();
2379 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2380 // Objects of size 1 and 2 are right justified, everything else is
2381 // left justified. This means the memory address is adjusted forwards.
2382 if (ObjSize==1 || ObjSize==2) {
2383 CurArgOffset = CurArgOffset + (4 - ObjSize);
2385 // The value of the object is its address.
2386 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2387 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2388 InVals.push_back(FIN);
2389 if (ObjSize==1 || ObjSize==2) {
2390 if (GPR_idx != Num_GPR_Regs) {
2393 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2395 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2396 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2397 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2398 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2399 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2400 MachinePointerInfo(FuncArg,
2402 ObjType, false, false, 0);
2403 MemOps.push_back(Store);
2407 ArgOffset += PtrByteSize;
2411 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2412 // Store whatever pieces of the object are in registers
2413 // to memory. ArgOffset will be the address of the beginning
2415 if (GPR_idx != Num_GPR_Regs) {
2418 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2420 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2421 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2422 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2423 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2424 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2425 MachinePointerInfo(FuncArg, ArgOffset),
2427 MemOps.push_back(Store);
2429 ArgOffset += PtrByteSize;
2431 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2438 switch (ObjectVT.getSimpleVT().SimpleTy) {
2439 default: llvm_unreachable("Unhandled argument type!");
2442 if (GPR_idx != Num_GPR_Regs) {
2443 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2444 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2448 ArgSize = PtrByteSize;
2450 // All int arguments reserve stack space in the Darwin ABI.
2451 ArgOffset += PtrByteSize;
2455 case MVT::i64: // PPC64
2456 if (GPR_idx != Num_GPR_Regs) {
2457 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2458 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2460 if (ObjectVT == MVT::i32) {
2461 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2462 // value to MVT::i64 and then truncate to the correct register size.
2464 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2465 DAG.getValueType(ObjectVT));
2466 else if (Flags.isZExt())
2467 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2468 DAG.getValueType(ObjectVT));
2470 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2476 ArgSize = PtrByteSize;
2478 // All int arguments reserve stack space in the Darwin ABI.
2484 // Every 4 bytes of argument space consumes one of the GPRs available for
2485 // argument passing.
2486 if (GPR_idx != Num_GPR_Regs) {
2488 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2491 if (FPR_idx != Num_FPR_Regs) {
2494 if (ObjectVT == MVT::f32)
2495 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2497 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2499 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2505 // All FP arguments reserve stack space in the Darwin ABI.
2506 ArgOffset += isPPC64 ? 8 : ObjSize;
2512 // Note that vector arguments in registers don't reserve stack space,
2513 // except in varargs functions.
2514 if (VR_idx != Num_VR_Regs) {
2515 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2516 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2518 while ((ArgOffset % 16) != 0) {
2519 ArgOffset += PtrByteSize;
2520 if (GPR_idx != Num_GPR_Regs)
2524 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2528 if (!isVarArg && !isPPC64) {
2529 // Vectors go after all the nonvectors.
2530 CurArgOffset = VecArgOffset;
2533 // Vectors are aligned.
2534 ArgOffset = ((ArgOffset+15)/16)*16;
2535 CurArgOffset = ArgOffset;
2543 // We need to load the argument to a virtual register if we determined above
2544 // that we ran out of physical registers of the appropriate type.
2546 int FI = MFI->CreateFixedObject(ObjSize,
2547 CurArgOffset + (ArgSize - ObjSize),
2549 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2550 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2551 false, false, false, 0);
2554 InVals.push_back(ArgVal);
2557 // Set the size that is at least reserved in caller of this function. Tail
2558 // call optimized function's reserved stack space needs to be aligned so that
2559 // taking the difference between two stack areas will result in an aligned
2561 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2562 // Add the Altivec parameters at the end, if needed.
2563 if (nAltivecParamsAtEnd) {
2564 MinReservedArea = ((MinReservedArea+15)/16)*16;
2565 MinReservedArea += 16*nAltivecParamsAtEnd;
2568 std::max(MinReservedArea,
2569 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2570 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2571 getStackAlignment();
2572 unsigned AlignMask = TargetAlign-1;
2573 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2574 FI->setMinReservedArea(MinReservedArea);
2576 // If the function takes variable number of arguments, make a frame index for
2577 // the start of the first vararg value... for expansion of llvm.va_start.
2579 int Depth = ArgOffset;
2581 FuncInfo->setVarArgsFrameIndex(
2582 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2584 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2586 // If this function is vararg, store any remaining integer argument regs
2587 // to their spots on the stack so that they may be loaded by deferencing the
2588 // result of va_next.
2589 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2593 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2595 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2597 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2598 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2599 MachinePointerInfo(), false, false, 0);
2600 MemOps.push_back(Store);
2601 // Increment the address by four for the next argument to store
2602 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2603 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2607 if (!MemOps.empty())
2608 Chain = DAG.getNode(ISD::TokenFactor, dl,
2609 MVT::Other, &MemOps[0], MemOps.size());
2614 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2615 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2617 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2621 const SmallVectorImpl<ISD::OutputArg>
2623 const SmallVectorImpl<SDValue> &OutVals,
2624 unsigned &nAltivecParamsAtEnd) {
2625 // Count how many bytes are to be pushed on the stack, including the linkage
2626 // area, and parameter passing area. We start with 24/48 bytes, which is
2627 // prereserved space for [SP][CR][LR][3 x unused].
2628 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2629 unsigned NumOps = Outs.size();
2630 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2632 // Add up all the space actually used.
2633 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2634 // they all go in registers, but we must reserve stack space for them for
2635 // possible use by the caller. In varargs or 64-bit calls, parameters are
2636 // assigned stack space in order, with padding so Altivec parameters are
2638 nAltivecParamsAtEnd = 0;
2639 for (unsigned i = 0; i != NumOps; ++i) {
2640 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2641 EVT ArgVT = Outs[i].VT;
2642 // Varargs Altivec parameters are padded to a 16 byte boundary.
2643 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2644 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2645 if (!isVarArg && !isPPC64) {
2646 // Non-varargs Altivec parameters go after all the non-Altivec
2647 // parameters; handle those later so we know how much padding we need.
2648 nAltivecParamsAtEnd++;
2651 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2652 NumBytes = ((NumBytes+15)/16)*16;
2654 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2657 // Allow for Altivec parameters at the end, if needed.
2658 if (nAltivecParamsAtEnd) {
2659 NumBytes = ((NumBytes+15)/16)*16;
2660 NumBytes += 16*nAltivecParamsAtEnd;
2663 // The prolog code of the callee may store up to 8 GPR argument registers to
2664 // the stack, allowing va_start to index over them in memory if its varargs.
2665 // Because we cannot tell if this is needed on the caller side, we have to
2666 // conservatively assume that it is needed. As such, make sure we have at
2667 // least enough stack space for the caller to store the 8 GPRs.
2668 NumBytes = std::max(NumBytes,
2669 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2671 // Tail call needs the stack to be aligned.
2672 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2673 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2674 getFrameLowering()->getStackAlignment();
2675 unsigned AlignMask = TargetAlign-1;
2676 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2682 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2683 /// adjusted to accommodate the arguments for the tailcall.
2684 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2685 unsigned ParamSize) {
2687 if (!isTailCall) return 0;
2689 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2690 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2691 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2692 // Remember only if the new adjustement is bigger.
2693 if (SPDiff < FI->getTailCallSPDelta())
2694 FI->setTailCallSPDelta(SPDiff);
2699 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2700 /// for tail call optimization. Targets which want to do tail call
2701 /// optimization should implement this function.
2703 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2704 CallingConv::ID CalleeCC,
2706 const SmallVectorImpl<ISD::InputArg> &Ins,
2707 SelectionDAG& DAG) const {
2708 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2711 // Variable argument functions are not supported.
2715 MachineFunction &MF = DAG.getMachineFunction();
2716 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2717 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2718 // Functions containing by val parameters are not supported.
2719 for (unsigned i = 0; i != Ins.size(); i++) {
2720 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2721 if (Flags.isByVal()) return false;
2724 // Non PIC/GOT tail calls are supported.
2725 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2728 // At the moment we can only do local tail calls (in same module, hidden
2729 // or protected) if we are generating PIC.
2730 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2731 return G->getGlobal()->hasHiddenVisibility()
2732 || G->getGlobal()->hasProtectedVisibility();
2738 /// isCallCompatibleAddress - Return the immediate to use if the specified
2739 /// 32-bit value is representable in the immediate field of a BxA instruction.
2740 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2741 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2744 int Addr = C->getZExtValue();
2745 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2746 SignExtend32<26>(Addr) != Addr)
2747 return 0; // Top 6 bits have to be sext of immediate.
2749 return DAG.getConstant((int)C->getZExtValue() >> 2,
2750 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2755 struct TailCallArgumentInfo {
2760 TailCallArgumentInfo() : FrameIdx(0) {}
2765 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2767 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2769 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2770 SmallVector<SDValue, 8> &MemOpChains,
2772 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2773 SDValue Arg = TailCallArgs[i].Arg;
2774 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2775 int FI = TailCallArgs[i].FrameIdx;
2776 // Store relative to framepointer.
2777 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2778 MachinePointerInfo::getFixedStack(FI),
2783 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2784 /// the appropriate stack slot for the tail call optimized function call.
2785 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2786 MachineFunction &MF,
2795 // Calculate the new stack slot for the return address.
2796 int SlotSize = isPPC64 ? 8 : 4;
2797 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2799 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2800 NewRetAddrLoc, true);
2801 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2802 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2803 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2804 MachinePointerInfo::getFixedStack(NewRetAddr),
2807 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2808 // slot as the FP is never overwritten.
2811 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2812 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2814 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2815 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2816 MachinePointerInfo::getFixedStack(NewFPIdx),
2823 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2824 /// the position of the argument.
2826 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2827 SDValue Arg, int SPDiff, unsigned ArgOffset,
2828 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2829 int Offset = ArgOffset + SPDiff;
2830 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2831 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2832 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2833 SDValue FIN = DAG.getFrameIndex(FI, VT);
2834 TailCallArgumentInfo Info;
2836 Info.FrameIdxOp = FIN;
2838 TailCallArguments.push_back(Info);
2841 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2842 /// stack slot. Returns the chain as result and the loaded frame pointers in
2843 /// LROpOut/FPOpout. Used when tail calling.
2844 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2850 DebugLoc dl) const {
2852 // Load the LR and FP stack slot for later adjusting.
2853 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2854 LROpOut = getReturnAddrFrameIndex(DAG);
2855 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2856 false, false, false, 0);
2857 Chain = SDValue(LROpOut.getNode(), 1);
2859 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2860 // slot as the FP is never overwritten.
2862 FPOpOut = getFramePointerFrameIndex(DAG);
2863 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2864 false, false, false, 0);
2865 Chain = SDValue(FPOpOut.getNode(), 1);
2871 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2872 /// by "Src" to address "Dst" of size "Size". Alignment information is
2873 /// specified by the specific parameter attribute. The copy will be passed as
2874 /// a byval function parameter.
2875 /// Sometimes what we are copying is the end of a larger object, the part that
2876 /// does not fit in registers.
2878 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2879 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2881 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2882 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2883 false, false, MachinePointerInfo(0),
2884 MachinePointerInfo(0));
2887 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2890 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2891 SDValue Arg, SDValue PtrOff, int SPDiff,
2892 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2893 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2894 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2896 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2901 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2903 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2904 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2905 DAG.getConstant(ArgOffset, PtrVT));
2907 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2908 MachinePointerInfo(), false, false, 0));
2909 // Calculate and remember argument location.
2910 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2915 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2916 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2917 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2918 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2919 MachineFunction &MF = DAG.getMachineFunction();
2921 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2922 // might overwrite each other in case of tail call optimization.
2923 SmallVector<SDValue, 8> MemOpChains2;
2924 // Do not flag preceding copytoreg stuff together with the following stuff.
2926 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2928 if (!MemOpChains2.empty())
2929 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2930 &MemOpChains2[0], MemOpChains2.size());
2932 // Store the return address to the appropriate stack slot.
2933 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2934 isPPC64, isDarwinABI, dl);
2936 // Emit callseq_end just before tailcall node.
2937 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2938 DAG.getIntPtrConstant(0, true), InFlag);
2939 InFlag = Chain.getValue(1);
2943 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2944 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2945 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2946 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2947 const PPCSubtarget &PPCSubTarget) {
2949 bool isPPC64 = PPCSubTarget.isPPC64();
2950 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2952 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2953 NodeTys.push_back(MVT::Other); // Returns a chain
2954 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2956 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2958 bool needIndirectCall = true;
2959 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2960 // If this is an absolute destination address, use the munged value.
2961 Callee = SDValue(Dest, 0);
2962 needIndirectCall = false;
2965 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2966 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2967 // Use indirect calls for ALL functions calls in JIT mode, since the
2968 // far-call stubs may be outside relocation limits for a BL instruction.
2969 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2970 unsigned OpFlags = 0;
2971 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2972 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2973 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2974 (G->getGlobal()->isDeclaration() ||
2975 G->getGlobal()->isWeakForLinker())) {
2976 // PC-relative references to external symbols should go through $stub,
2977 // unless we're building with the leopard linker or later, which
2978 // automatically synthesizes these stubs.
2979 OpFlags = PPCII::MO_DARWIN_STUB;
2982 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2983 // every direct call is) turn it into a TargetGlobalAddress /
2984 // TargetExternalSymbol node so that legalize doesn't hack it.
2985 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2986 Callee.getValueType(),
2988 needIndirectCall = false;
2992 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2993 unsigned char OpFlags = 0;
2995 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2996 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2997 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2998 // PC-relative references to external symbols should go through $stub,
2999 // unless we're building with the leopard linker or later, which
3000 // automatically synthesizes these stubs.
3001 OpFlags = PPCII::MO_DARWIN_STUB;
3004 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3006 needIndirectCall = false;
3009 if (needIndirectCall) {
3010 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3011 // to do the call, we can't use PPCISD::CALL.
3012 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3014 if (isSVR4ABI && isPPC64) {
3015 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3016 // entry point, but to the function descriptor (the function entry point
3017 // address is part of the function descriptor though).
3018 // The function descriptor is a three doubleword structure with the
3019 // following fields: function entry point, TOC base address and
3020 // environment pointer.
3021 // Thus for a call through a function pointer, the following actions need
3023 // 1. Save the TOC of the caller in the TOC save area of its stack
3024 // frame (this is done in LowerCall_Darwin_Or_64SVR4()).
3025 // 2. Load the address of the function entry point from the function
3027 // 3. Load the TOC of the callee from the function descriptor into r2.
3028 // 4. Load the environment pointer from the function descriptor into
3030 // 5. Branch to the function entry point address.
3031 // 6. On return of the callee, the TOC of the caller needs to be
3032 // restored (this is done in FinishCall()).
3034 // All those operations are flagged together to ensure that no other
3035 // operations can be scheduled in between. E.g. without flagging the
3036 // operations together, a TOC access in the caller could be scheduled
3037 // between the load of the callee TOC and the branch to the callee, which
3038 // results in the TOC access going through the TOC of the callee instead
3039 // of going through the TOC of the caller, which leads to incorrect code.
3041 // Load the address of the function entry point from the function
3043 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3044 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3045 InFlag.getNode() ? 3 : 2);
3046 Chain = LoadFuncPtr.getValue(1);
3047 InFlag = LoadFuncPtr.getValue(2);
3049 // Load environment pointer into r11.
3050 // Offset of the environment pointer within the function descriptor.
3051 SDValue PtrOff = DAG.getIntPtrConstant(16);
3053 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3054 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3056 Chain = LoadEnvPtr.getValue(1);
3057 InFlag = LoadEnvPtr.getValue(2);
3059 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3061 Chain = EnvVal.getValue(0);
3062 InFlag = EnvVal.getValue(1);
3064 // Load TOC of the callee into r2. We are using a target-specific load
3065 // with r2 hard coded, because the result of a target-independent load
3066 // would never go directly into r2, since r2 is a reserved register (which
3067 // prevents the register allocator from allocating it), resulting in an
3068 // additional register being allocated and an unnecessary move instruction
3070 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3071 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3073 Chain = LoadTOCPtr.getValue(0);
3074 InFlag = LoadTOCPtr.getValue(1);
3076 MTCTROps[0] = Chain;
3077 MTCTROps[1] = LoadFuncPtr;
3078 MTCTROps[2] = InFlag;
3081 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3082 2 + (InFlag.getNode() != 0));
3083 InFlag = Chain.getValue(1);
3086 NodeTys.push_back(MVT::Other);
3087 NodeTys.push_back(MVT::Glue);
3088 Ops.push_back(Chain);
3089 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3091 // Add CTR register as callee so a bctr can be emitted later.
3093 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3096 // If this is a direct call, pass the chain and the callee.
3097 if (Callee.getNode()) {
3098 Ops.push_back(Chain);
3099 Ops.push_back(Callee);
3101 // If this is a tail call add stack pointer delta.
3103 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3105 // Add argument registers to the end of the list so that they are known live
3107 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3108 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3109 RegsToPass[i].second.getValueType()));
3115 bool isLocalCall(const SDValue &Callee)
3117 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3118 return !G->getGlobal()->isDeclaration() &&
3119 !G->getGlobal()->isWeakForLinker();
3124 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3125 CallingConv::ID CallConv, bool isVarArg,
3126 const SmallVectorImpl<ISD::InputArg> &Ins,
3127 DebugLoc dl, SelectionDAG &DAG,
3128 SmallVectorImpl<SDValue> &InVals) const {
3130 SmallVector<CCValAssign, 16> RVLocs;
3131 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3132 getTargetMachine(), RVLocs, *DAG.getContext());
3133 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3135 // Copy all of the result registers out of their specified physreg.
3136 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3137 CCValAssign &VA = RVLocs[i];
3138 EVT VT = VA.getValVT();
3139 assert(VA.isRegLoc() && "Can only return in registers!");
3140 Chain = DAG.getCopyFromReg(Chain, dl,
3141 VA.getLocReg(), VT, InFlag).getValue(1);
3142 InVals.push_back(Chain.getValue(0));
3143 InFlag = Chain.getValue(2);
3150 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3151 bool isTailCall, bool isVarArg,
3153 SmallVector<std::pair<unsigned, SDValue>, 8>
3155 SDValue InFlag, SDValue Chain,
3157 int SPDiff, unsigned NumBytes,
3158 const SmallVectorImpl<ISD::InputArg> &Ins,
3159 SmallVectorImpl<SDValue> &InVals) const {
3160 std::vector<EVT> NodeTys;
3161 SmallVector<SDValue, 8> Ops;
3162 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3163 isTailCall, RegsToPass, Ops, NodeTys,
3166 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3167 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3168 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3170 // When performing tail call optimization the callee pops its arguments off
3171 // the stack. Account for this here so these bytes can be pushed back on in
3172 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3173 int BytesCalleePops =
3174 (CallConv == CallingConv::Fast &&
3175 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3177 // Add a register mask operand representing the call-preserved registers.
3178 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3179 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3180 assert(Mask && "Missing call preserved mask for calling convention");
3181 Ops.push_back(DAG.getRegisterMask(Mask));
3183 if (InFlag.getNode())
3184 Ops.push_back(InFlag);
3188 // If this is the first return lowered for this function, add the regs
3189 // to the liveout set for the function.
3190 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3191 SmallVector<CCValAssign, 16> RVLocs;
3192 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3193 getTargetMachine(), RVLocs, *DAG.getContext());
3194 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3195 for (unsigned i = 0; i != RVLocs.size(); ++i)
3196 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3199 assert(((Callee.getOpcode() == ISD::Register &&
3200 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3201 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3202 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3203 isa<ConstantSDNode>(Callee)) &&
3204 "Expecting an global address, external symbol, absolute value or register");
3206 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3209 // Add a NOP immediately after the branch instruction when using the 64-bit
3210 // SVR4 ABI. At link time, if caller and callee are in a different module and
3211 // thus have a different TOC, the call will be replaced with a call to a stub
3212 // function which saves the current TOC, loads the TOC of the callee and
3213 // branches to the callee. The NOP will be replaced with a load instruction
3214 // which restores the TOC of the caller from the TOC save slot of the current
3215 // stack frame. If caller and callee belong to the same module (and have the
3216 // same TOC), the NOP will remain unchanged.
3218 bool needsTOCRestore = false;
3219 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3220 if (CallOpc == PPCISD::BCTRL_SVR4) {
3221 // This is a call through a function pointer.
3222 // Restore the caller TOC from the save area into R2.
3223 // See PrepareCall() for more information about calls through function
3224 // pointers in the 64-bit SVR4 ABI.
3225 // We are using a target-specific load with r2 hard coded, because the
3226 // result of a target-independent load would never go directly into r2,
3227 // since r2 is a reserved register (which prevents the register allocator
3228 // from allocating it), resulting in an additional register being
3229 // allocated and an unnecessary move instruction being generated.
3230 needsTOCRestore = true;
3231 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3232 // Otherwise insert NOP for non-local calls.
3233 CallOpc = PPCISD::CALL_NOP_SVR4;
3237 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3238 InFlag = Chain.getValue(1);
3240 if (needsTOCRestore) {
3241 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3242 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3243 InFlag = Chain.getValue(1);
3246 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3247 DAG.getIntPtrConstant(BytesCalleePops, true),
3250 InFlag = Chain.getValue(1);
3252 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3253 Ins, dl, DAG, InVals);
3257 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3258 SmallVectorImpl<SDValue> &InVals) const {
3259 SelectionDAG &DAG = CLI.DAG;
3260 DebugLoc &dl = CLI.DL;
3261 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3262 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3263 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3264 SDValue Chain = CLI.Chain;
3265 SDValue Callee = CLI.Callee;
3266 bool &isTailCall = CLI.IsTailCall;
3267 CallingConv::ID CallConv = CLI.CallConv;
3268 bool isVarArg = CLI.IsVarArg;
3271 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3274 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3275 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3276 isTailCall, Outs, OutVals, Ins,
3279 return LowerCall_Darwin_Or_64SVR4(Chain, Callee, CallConv, isVarArg,
3280 isTailCall, Outs, OutVals, Ins,
3285 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3286 CallingConv::ID CallConv, bool isVarArg,
3288 const SmallVectorImpl<ISD::OutputArg> &Outs,
3289 const SmallVectorImpl<SDValue> &OutVals,
3290 const SmallVectorImpl<ISD::InputArg> &Ins,
3291 DebugLoc dl, SelectionDAG &DAG,
3292 SmallVectorImpl<SDValue> &InVals) const {
3293 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3294 // of the 32-bit SVR4 ABI stack frame layout.
3296 assert((CallConv == CallingConv::C ||
3297 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3299 unsigned PtrByteSize = 4;
3301 MachineFunction &MF = DAG.getMachineFunction();
3303 // Mark this function as potentially containing a function that contains a
3304 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3305 // and restoring the callers stack pointer in this functions epilog. This is
3306 // done because by tail calling the called function might overwrite the value
3307 // in this function's (MF) stack pointer stack slot 0(SP).
3308 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3309 CallConv == CallingConv::Fast)
3310 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3312 // Count how many bytes are to be pushed on the stack, including the linkage
3313 // area, parameter list area and the part of the local variable space which
3314 // contains copies of aggregates which are passed by value.
3316 // Assign locations to all of the outgoing arguments.
3317 SmallVector<CCValAssign, 16> ArgLocs;
3318 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3319 getTargetMachine(), ArgLocs, *DAG.getContext());
3321 // Reserve space for the linkage area on the stack.
3322 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3325 // Handle fixed and variable vector arguments differently.
3326 // Fixed vector arguments go into registers as long as registers are
3327 // available. Variable vector arguments always go into memory.
3328 unsigned NumArgs = Outs.size();
3330 for (unsigned i = 0; i != NumArgs; ++i) {
3331 MVT ArgVT = Outs[i].VT;
3332 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3335 if (Outs[i].IsFixed) {
3336 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3339 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3345 errs() << "Call operand #" << i << " has unhandled type "
3346 << EVT(ArgVT).getEVTString() << "\n";
3348 llvm_unreachable(0);
3352 // All arguments are treated the same.
3353 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3356 // Assign locations to all of the outgoing aggregate by value arguments.
3357 SmallVector<CCValAssign, 16> ByValArgLocs;
3358 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3359 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3361 // Reserve stack space for the allocations in CCInfo.
3362 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3364 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3366 // Size of the linkage area, parameter list area and the part of the local
3367 // space variable where copies of aggregates which are passed by value are
3369 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3371 // Calculate by how many bytes the stack has to be adjusted in case of tail
3372 // call optimization.
3373 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3375 // Adjust the stack pointer for the new arguments...
3376 // These operations are automatically eliminated by the prolog/epilog pass
3377 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3378 SDValue CallSeqStart = Chain;
3380 // Load the return address and frame pointer so it can be moved somewhere else
3383 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3386 // Set up a copy of the stack pointer for use loading and storing any
3387 // arguments that may not fit in the registers available for argument
3389 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3391 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3392 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3393 SmallVector<SDValue, 8> MemOpChains;
3395 bool seenFloatArg = false;
3396 // Walk the register/memloc assignments, inserting copies/loads.
3397 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3400 CCValAssign &VA = ArgLocs[i];
3401 SDValue Arg = OutVals[i];
3402 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3404 if (Flags.isByVal()) {
3405 // Argument is an aggregate which is passed by value, thus we need to
3406 // create a copy of it in the local variable space of the current stack
3407 // frame (which is the stack frame of the caller) and pass the address of
3408 // this copy to the callee.
3409 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3410 CCValAssign &ByValVA = ByValArgLocs[j++];
3411 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3413 // Memory reserved in the local variable space of the callers stack frame.
3414 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3416 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3417 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3419 // Create a copy of the argument in the local area of the current
3421 SDValue MemcpyCall =
3422 CreateCopyOfByValArgument(Arg, PtrOff,
3423 CallSeqStart.getNode()->getOperand(0),
3426 // This must go outside the CALLSEQ_START..END.
3427 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3428 CallSeqStart.getNode()->getOperand(1));
3429 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3430 NewCallSeqStart.getNode());
3431 Chain = CallSeqStart = NewCallSeqStart;
3433 // Pass the address of the aggregate copy on the stack either in a
3434 // physical register or in the parameter list area of the current stack
3435 // frame to the callee.
3439 if (VA.isRegLoc()) {
3440 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3441 // Put argument in a physical register.
3442 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3444 // Put argument in the parameter list area of the current stack frame.
3445 assert(VA.isMemLoc());
3446 unsigned LocMemOffset = VA.getLocMemOffset();
3449 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3450 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3452 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3453 MachinePointerInfo(),
3456 // Calculate and remember argument location.
3457 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3463 if (!MemOpChains.empty())
3464 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3465 &MemOpChains[0], MemOpChains.size());
3467 // Build a sequence of copy-to-reg nodes chained together with token chain
3468 // and flag operands which copy the outgoing args into the appropriate regs.
3470 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3471 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3472 RegsToPass[i].second, InFlag);
3473 InFlag = Chain.getValue(1);
3476 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3479 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3480 SDValue Ops[] = { Chain, InFlag };
3482 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3483 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3485 InFlag = Chain.getValue(1);
3489 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3490 false, TailCallArguments);
3492 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3493 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3498 PPCTargetLowering::LowerCall_Darwin_Or_64SVR4(SDValue Chain, SDValue Callee,
3499 CallingConv::ID CallConv, bool isVarArg,
3501 const SmallVectorImpl<ISD::OutputArg> &Outs,
3502 const SmallVectorImpl<SDValue> &OutVals,
3503 const SmallVectorImpl<ISD::InputArg> &Ins,
3504 DebugLoc dl, SelectionDAG &DAG,
3505 SmallVectorImpl<SDValue> &InVals) const {
3507 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3509 unsigned NumOps = Outs.size();
3511 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3512 bool isPPC64 = PtrVT == MVT::i64;
3513 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3515 MachineFunction &MF = DAG.getMachineFunction();
3517 // Mark this function as potentially containing a function that contains a
3518 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3519 // and restoring the callers stack pointer in this functions epilog. This is
3520 // done because by tail calling the called function might overwrite the value
3521 // in this function's (MF) stack pointer stack slot 0(SP).
3522 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3523 CallConv == CallingConv::Fast)
3524 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3526 unsigned nAltivecParamsAtEnd = 0;
3528 // Count how many bytes are to be pushed on the stack, including the linkage
3529 // area, and parameter passing area. We start with 24/48 bytes, which is
3530 // prereserved space for [SP][CR][LR][3 x unused].
3532 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3534 nAltivecParamsAtEnd);
3536 // Calculate by how many bytes the stack has to be adjusted in case of tail
3537 // call optimization.
3538 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3540 // To protect arguments on the stack from being clobbered in a tail call,
3541 // force all the loads to happen before doing any other lowering.
3543 Chain = DAG.getStackArgumentTokenFactor(Chain);
3545 // Adjust the stack pointer for the new arguments...
3546 // These operations are automatically eliminated by the prolog/epilog pass
3547 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3548 SDValue CallSeqStart = Chain;
3550 // Load the return address and frame pointer so it can be move somewhere else
3553 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3556 // Set up a copy of the stack pointer for use loading and storing any
3557 // arguments that may not fit in the registers available for argument
3561 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3563 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3565 // Figure out which arguments are going to go in registers, and which in
3566 // memory. Also, if this is a vararg function, floating point operations
3567 // must be stored to our stack, and loaded into integer regs as well, if
3568 // any integer regs are available for argument passing.
3569 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3570 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3572 static const uint16_t GPR_32[] = { // 32-bit registers.
3573 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3574 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3576 static const uint16_t GPR_64[] = { // 64-bit registers.
3577 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3578 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3580 static const uint16_t *FPR = GetFPR();
3582 static const uint16_t VR[] = {
3583 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3584 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3586 const unsigned NumGPRs = array_lengthof(GPR_32);
3587 const unsigned NumFPRs = 13;
3588 const unsigned NumVRs = array_lengthof(VR);
3590 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3592 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3593 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3595 SmallVector<SDValue, 8> MemOpChains;
3596 for (unsigned i = 0; i != NumOps; ++i) {
3597 SDValue Arg = OutVals[i];
3598 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3600 // PtrOff will be used to store the current argument to the stack if a
3601 // register cannot be found for it.
3604 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3606 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3608 // On PPC64, promote integers to 64-bit values.
3609 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3610 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3611 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3612 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3615 // FIXME memcpy is used way more than necessary. Correctness first.
3616 // Note: "by value" is code for passing a structure by value, not
3618 if (Flags.isByVal()) {
3619 // Note: Size includes alignment padding, so
3620 // struct x { short a; char b; }
3621 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3622 // These are the proper values we need for right-justifying the
3623 // aggregate in a parameter register for 64-bit SVR4.
3624 unsigned Size = Flags.getByValSize();
3625 // FOR DARWIN ONLY: Very small objects are passed right-justified.
3626 // Everything else is passed left-justified.
3627 // FOR 64-BIT SVR4: All aggregates smaller than 8 bytes must
3628 // be passed right-justified.
3629 if (Size==1 || Size==2 ||
3630 (Size==4 && isSVR4ABI)) {
3631 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3632 if (GPR_idx != NumGPRs) {
3633 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3634 MachinePointerInfo(), VT,
3636 MemOpChains.push_back(Load.getValue(1));
3637 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3639 ArgOffset += PtrByteSize;
3641 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3642 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3643 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3644 CallSeqStart.getNode()->getOperand(0),
3646 // This must go outside the CALLSEQ_START..END.
3647 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3648 CallSeqStart.getNode()->getOperand(1));
3649 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3650 NewCallSeqStart.getNode());
3651 Chain = CallSeqStart = NewCallSeqStart;
3652 ArgOffset += PtrByteSize;
3656 // Copy entire object into memory. There are cases where gcc-generated
3657 // code assumes it is there, even if it could be put entirely into
3658 // registers. (This is not what the doc says.)
3660 // FIXME: The above statement is likely due to a misunderstanding of the
3661 // documents. At least for 64-bit SVR4, all arguments must be copied
3662 // into the parameter area BY THE CALLEE in the event that the callee
3663 // takes the address of any formal argument. That has not yet been
3664 // implemented. However, it is reasonable to use the stack area as a
3665 // staging area for the register load.
3667 // Skip this for small aggregates under 64-bit SVR4, as we will use
3668 // the same slot for a right-justified copy, below.
3669 if (Size >= 8 || !isSVR4ABI) {
3670 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3671 CallSeqStart.getNode()->getOperand(0),
3673 // This must go outside the CALLSEQ_START..END.
3674 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3675 CallSeqStart.getNode()->getOperand(1));
3676 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3677 NewCallSeqStart.getNode());
3678 Chain = CallSeqStart = NewCallSeqStart;
3681 // FOR 64-BIT SVR4: When a register is available, pass the
3682 // aggregate right-justified.
3683 if (isSVR4ABI && Size < 8 && GPR_idx != NumGPRs) {
3684 // The easiest way to get this right-justified in a register
3685 // is to copy the structure into the rightmost portion of a
3686 // local variable slot, then load the whole slot into the
3688 // FIXME: The memcpy seems to produce pretty awful code for
3689 // small aggregates, particularly for packed ones.
3690 // FIXME: It would be preferable to use the slot in the
3691 // parameter save area instead of a new local variable.
3692 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3693 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3694 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3695 CallSeqStart.getNode()->getOperand(0),
3698 // Place the memcpy outside the CALLSEQ_START..END.
3699 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3700 CallSeqStart.getNode()->getOperand(1));
3701 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3702 NewCallSeqStart.getNode());
3703 Chain = CallSeqStart = NewCallSeqStart;
3705 // Load the slot into the register.
3706 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3707 MachinePointerInfo(),
3708 false, false, false, 0);
3709 MemOpChains.push_back(Load.getValue(1));
3710 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3712 // Done with this argument.
3713 ArgOffset += PtrByteSize;
3717 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
3718 // copy the pieces of the object that fit into registers from the
3719 // parameter save area.
3720 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3721 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3722 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3723 if (GPR_idx != NumGPRs) {
3724 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3725 MachinePointerInfo(),
3726 false, false, false, 0);
3727 MemOpChains.push_back(Load.getValue(1));
3728 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3729 ArgOffset += PtrByteSize;
3731 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3738 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3739 default: llvm_unreachable("Unexpected ValueType for argument!");
3742 if (GPR_idx != NumGPRs) {
3743 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3745 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3746 isPPC64, isTailCall, false, MemOpChains,
3747 TailCallArguments, dl);
3749 ArgOffset += PtrByteSize;
3753 if (FPR_idx != NumFPRs) {
3754 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3757 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3758 MachinePointerInfo(), false, false, 0);
3759 MemOpChains.push_back(Store);
3761 // Float varargs are always shadowed in available integer registers
3762 if (GPR_idx != NumGPRs) {
3763 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3764 MachinePointerInfo(), false, false,
3766 MemOpChains.push_back(Load.getValue(1));
3767 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3769 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3770 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3771 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3772 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3773 MachinePointerInfo(),
3774 false, false, false, 0);
3775 MemOpChains.push_back(Load.getValue(1));
3776 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3779 // If we have any FPRs remaining, we may also have GPRs remaining.
3780 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3782 if (GPR_idx != NumGPRs)
3784 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3785 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3789 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3790 isPPC64, isTailCall, false, MemOpChains,
3791 TailCallArguments, dl);
3796 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3803 // These go aligned on the stack, or in the corresponding R registers
3804 // when within range. The Darwin PPC ABI doc claims they also go in
3805 // V registers; in fact gcc does this only for arguments that are
3806 // prototyped, not for those that match the ... We do it for all
3807 // arguments, seems to work.
3808 while (ArgOffset % 16 !=0) {
3809 ArgOffset += PtrByteSize;
3810 if (GPR_idx != NumGPRs)
3813 // We could elide this store in the case where the object fits
3814 // entirely in R registers. Maybe later.
3815 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3816 DAG.getConstant(ArgOffset, PtrVT));
3817 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3818 MachinePointerInfo(), false, false, 0);
3819 MemOpChains.push_back(Store);
3820 if (VR_idx != NumVRs) {
3821 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3822 MachinePointerInfo(),
3823 false, false, false, 0);
3824 MemOpChains.push_back(Load.getValue(1));
3825 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3828 for (unsigned i=0; i<16; i+=PtrByteSize) {
3829 if (GPR_idx == NumGPRs)
3831 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3832 DAG.getConstant(i, PtrVT));
3833 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3834 false, false, false, 0);
3835 MemOpChains.push_back(Load.getValue(1));
3836 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3841 // Non-varargs Altivec params generally go in registers, but have
3842 // stack space allocated at the end.
3843 if (VR_idx != NumVRs) {
3844 // Doesn't have GPR space allocated.
3845 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3846 } else if (nAltivecParamsAtEnd==0) {
3847 // We are emitting Altivec params in order.
3848 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3849 isPPC64, isTailCall, true, MemOpChains,
3850 TailCallArguments, dl);
3856 // If all Altivec parameters fit in registers, as they usually do,
3857 // they get stack space following the non-Altivec parameters. We
3858 // don't track this here because nobody below needs it.
3859 // If there are more Altivec parameters than fit in registers emit
3861 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3863 // Offset is aligned; skip 1st 12 params which go in V registers.
3864 ArgOffset = ((ArgOffset+15)/16)*16;
3866 for (unsigned i = 0; i != NumOps; ++i) {
3867 SDValue Arg = OutVals[i];
3868 EVT ArgType = Outs[i].VT;
3869 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3870 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3873 // We are emitting Altivec params in order.
3874 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3875 isPPC64, isTailCall, true, MemOpChains,
3876 TailCallArguments, dl);
3883 if (!MemOpChains.empty())
3884 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3885 &MemOpChains[0], MemOpChains.size());
3887 // Check if this is an indirect call (MTCTR/BCTRL).
3888 // See PrepareCall() for more information about calls through function
3889 // pointers in the 64-bit SVR4 ABI.
3890 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3891 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3892 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3893 !isBLACompatibleAddress(Callee, DAG)) {
3894 // Load r2 into a virtual register and store it to the TOC save area.
3895 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3896 // TOC save area offset.
3897 SDValue PtrOff = DAG.getIntPtrConstant(40);
3898 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3899 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3903 // On Darwin, R12 must contain the address of an indirect callee. This does
3904 // not mean the MTCTR instruction must use R12; it's easier to model this as
3905 // an extra parameter, so do that.
3907 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3908 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3909 !isBLACompatibleAddress(Callee, DAG))
3910 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3911 PPC::R12), Callee));
3913 // Build a sequence of copy-to-reg nodes chained together with token chain
3914 // and flag operands which copy the outgoing args into the appropriate regs.
3916 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3917 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3918 RegsToPass[i].second, InFlag);
3919 InFlag = Chain.getValue(1);
3923 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3924 FPOp, true, TailCallArguments);
3926 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3927 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3932 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3933 MachineFunction &MF, bool isVarArg,
3934 const SmallVectorImpl<ISD::OutputArg> &Outs,
3935 LLVMContext &Context) const {
3936 SmallVector<CCValAssign, 16> RVLocs;
3937 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
3939 return CCInfo.CheckReturn(Outs, RetCC_PPC);
3943 PPCTargetLowering::LowerReturn(SDValue Chain,
3944 CallingConv::ID CallConv, bool isVarArg,
3945 const SmallVectorImpl<ISD::OutputArg> &Outs,
3946 const SmallVectorImpl<SDValue> &OutVals,
3947 DebugLoc dl, SelectionDAG &DAG) const {
3949 SmallVector<CCValAssign, 16> RVLocs;
3950 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3951 getTargetMachine(), RVLocs, *DAG.getContext());
3952 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3954 // If this is the first return lowered for this function, add the regs to the
3955 // liveout set for the function.
3956 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3957 for (unsigned i = 0; i != RVLocs.size(); ++i)
3958 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3963 // Copy the result values into the output registers.
3964 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3965 CCValAssign &VA = RVLocs[i];
3966 assert(VA.isRegLoc() && "Can only return in registers!");
3967 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3969 Flag = Chain.getValue(1);
3973 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3975 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3978 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3979 const PPCSubtarget &Subtarget) const {
3980 // When we pop the dynamic allocation we need to restore the SP link.
3981 DebugLoc dl = Op.getDebugLoc();
3983 // Get the corect type for pointers.
3984 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3986 // Construct the stack pointer operand.
3987 bool isPPC64 = Subtarget.isPPC64();
3988 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3989 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3991 // Get the operands for the STACKRESTORE.
3992 SDValue Chain = Op.getOperand(0);
3993 SDValue SaveSP = Op.getOperand(1);
3995 // Load the old link SP.
3996 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3997 MachinePointerInfo(),
3998 false, false, false, 0);
4000 // Restore the stack pointer.
4001 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4003 // Store the old link SP.
4004 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4011 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4012 MachineFunction &MF = DAG.getMachineFunction();
4013 bool isPPC64 = PPCSubTarget.isPPC64();
4014 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4015 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4017 // Get current frame pointer save index. The users of this index will be
4018 // primarily DYNALLOC instructions.
4019 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4020 int RASI = FI->getReturnAddrSaveIndex();
4022 // If the frame pointer save index hasn't been defined yet.
4024 // Find out what the fix offset of the frame pointer save area.
4025 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4026 // Allocate the frame index for frame pointer save area.
4027 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4029 FI->setReturnAddrSaveIndex(RASI);
4031 return DAG.getFrameIndex(RASI, PtrVT);
4035 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4036 MachineFunction &MF = DAG.getMachineFunction();
4037 bool isPPC64 = PPCSubTarget.isPPC64();
4038 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4039 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4041 // Get current frame pointer save index. The users of this index will be
4042 // primarily DYNALLOC instructions.
4043 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4044 int FPSI = FI->getFramePointerSaveIndex();
4046 // If the frame pointer save index hasn't been defined yet.
4048 // Find out what the fix offset of the frame pointer save area.
4049 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4052 // Allocate the frame index for frame pointer save area.
4053 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4055 FI->setFramePointerSaveIndex(FPSI);
4057 return DAG.getFrameIndex(FPSI, PtrVT);
4060 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4062 const PPCSubtarget &Subtarget) const {
4064 SDValue Chain = Op.getOperand(0);
4065 SDValue Size = Op.getOperand(1);
4066 DebugLoc dl = Op.getDebugLoc();
4068 // Get the corect type for pointers.
4069 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4071 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4072 DAG.getConstant(0, PtrVT), Size);
4073 // Construct a node for the frame pointer save index.
4074 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4075 // Build a DYNALLOC node.
4076 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4077 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4078 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4081 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4083 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4084 // Not FP? Not a fsel.
4085 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4086 !Op.getOperand(2).getValueType().isFloatingPoint())
4089 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4091 // Cannot handle SETEQ/SETNE.
4092 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4094 EVT ResVT = Op.getValueType();
4095 EVT CmpVT = Op.getOperand(0).getValueType();
4096 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4097 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4098 DebugLoc dl = Op.getDebugLoc();
4100 // If the RHS of the comparison is a 0.0, we don't need to do the
4101 // subtraction at all.
4102 if (isFloatingPointZero(RHS))
4104 default: break; // SETUO etc aren't handled by fsel.
4107 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4110 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4111 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4112 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
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,
4121 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4126 default: break; // SETUO etc aren't handled by fsel.
4129 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4130 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4131 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4132 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4135 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4136 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4137 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4138 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4141 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4142 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4143 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4144 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4147 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4148 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4149 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4150 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4155 // FIXME: Split this code up when LegalizeDAGTypes lands.
4156 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4157 DebugLoc dl) const {
4158 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4159 SDValue Src = Op.getOperand(0);
4160 if (Src.getValueType() == MVT::f32)
4161 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4164 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4165 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4167 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4172 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4176 // Convert the FP value to an int value through memory.
4177 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4179 // Emit a store to the stack slot.
4180 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4181 MachinePointerInfo(), false, false, 0);
4183 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4185 if (Op.getValueType() == MVT::i32)
4186 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4187 DAG.getConstant(4, FIPtr.getValueType()));
4188 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4189 false, false, false, 0);
4192 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4193 SelectionDAG &DAG) const {
4194 DebugLoc dl = Op.getDebugLoc();
4195 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4196 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4199 if (Op.getOperand(0).getValueType() == MVT::i64) {
4200 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
4201 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4202 if (Op.getValueType() == MVT::f32)
4203 FP = DAG.getNode(ISD::FP_ROUND, dl,
4204 MVT::f32, FP, DAG.getIntPtrConstant(0));
4208 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4209 "Unhandled SINT_TO_FP type in custom expander!");
4210 // Since we only generate this in 64-bit mode, we can take advantage of
4211 // 64-bit registers. In particular, sign extend the input value into the
4212 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4213 // then lfd it and fcfid it.
4214 MachineFunction &MF = DAG.getMachineFunction();
4215 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4216 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4217 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4218 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4220 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4223 // STD the extended value into the stack slot.
4224 MachineMemOperand *MMO =
4225 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4226 MachineMemOperand::MOStore, 8, 8);
4227 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4229 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4230 Ops, 4, MVT::i64, MMO);
4231 // Load the value as a double.
4232 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4233 false, false, false, 0);
4235 // FCFID it and return it.
4236 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4237 if (Op.getValueType() == MVT::f32)
4238 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4242 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4243 SelectionDAG &DAG) const {
4244 DebugLoc dl = Op.getDebugLoc();
4246 The rounding mode is in bits 30:31 of FPSR, and has the following
4253 FLT_ROUNDS, on the other hand, expects the following:
4260 To perform the conversion, we do:
4261 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4264 MachineFunction &MF = DAG.getMachineFunction();
4265 EVT VT = Op.getValueType();
4266 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4267 std::vector<EVT> NodeTys;
4268 SDValue MFFSreg, InFlag;
4270 // Save FP Control Word to register
4271 NodeTys.push_back(MVT::f64); // return register
4272 NodeTys.push_back(MVT::Glue); // unused in this context
4273 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4275 // Save FP register to stack slot
4276 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4277 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4278 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4279 StackSlot, MachinePointerInfo(), false, false,0);
4281 // Load FP Control Word from low 32 bits of stack slot.
4282 SDValue Four = DAG.getConstant(4, PtrVT);
4283 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4284 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4285 false, false, false, 0);
4287 // Transform as necessary
4289 DAG.getNode(ISD::AND, dl, MVT::i32,
4290 CWD, DAG.getConstant(3, MVT::i32));
4292 DAG.getNode(ISD::SRL, dl, MVT::i32,
4293 DAG.getNode(ISD::AND, dl, MVT::i32,
4294 DAG.getNode(ISD::XOR, dl, MVT::i32,
4295 CWD, DAG.getConstant(3, MVT::i32)),
4296 DAG.getConstant(3, MVT::i32)),
4297 DAG.getConstant(1, MVT::i32));
4300 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4302 return DAG.getNode((VT.getSizeInBits() < 16 ?
4303 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4306 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4307 EVT VT = Op.getValueType();
4308 unsigned BitWidth = VT.getSizeInBits();
4309 DebugLoc dl = Op.getDebugLoc();
4310 assert(Op.getNumOperands() == 3 &&
4311 VT == Op.getOperand(1).getValueType() &&
4314 // Expand into a bunch of logical ops. Note that these ops
4315 // depend on the PPC behavior for oversized shift amounts.
4316 SDValue Lo = Op.getOperand(0);
4317 SDValue Hi = Op.getOperand(1);
4318 SDValue Amt = Op.getOperand(2);
4319 EVT AmtVT = Amt.getValueType();
4321 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4322 DAG.getConstant(BitWidth, AmtVT), Amt);
4323 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4324 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4325 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4326 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4327 DAG.getConstant(-BitWidth, AmtVT));
4328 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4329 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4330 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4331 SDValue OutOps[] = { OutLo, OutHi };
4332 return DAG.getMergeValues(OutOps, 2, dl);
4335 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4336 EVT VT = Op.getValueType();
4337 DebugLoc dl = Op.getDebugLoc();
4338 unsigned BitWidth = VT.getSizeInBits();
4339 assert(Op.getNumOperands() == 3 &&
4340 VT == Op.getOperand(1).getValueType() &&
4343 // Expand into a bunch of logical ops. Note that these ops
4344 // depend on the PPC behavior for oversized shift amounts.
4345 SDValue Lo = Op.getOperand(0);
4346 SDValue Hi = Op.getOperand(1);
4347 SDValue Amt = Op.getOperand(2);
4348 EVT AmtVT = Amt.getValueType();
4350 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4351 DAG.getConstant(BitWidth, AmtVT), Amt);
4352 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4353 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4354 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4355 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4356 DAG.getConstant(-BitWidth, AmtVT));
4357 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4358 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4359 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4360 SDValue OutOps[] = { OutLo, OutHi };
4361 return DAG.getMergeValues(OutOps, 2, dl);
4364 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4365 DebugLoc dl = Op.getDebugLoc();
4366 EVT VT = Op.getValueType();
4367 unsigned BitWidth = VT.getSizeInBits();
4368 assert(Op.getNumOperands() == 3 &&
4369 VT == Op.getOperand(1).getValueType() &&
4372 // Expand into a bunch of logical ops, followed by a select_cc.
4373 SDValue Lo = Op.getOperand(0);
4374 SDValue Hi = Op.getOperand(1);
4375 SDValue Amt = Op.getOperand(2);
4376 EVT AmtVT = Amt.getValueType();
4378 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4379 DAG.getConstant(BitWidth, AmtVT), Amt);
4380 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4381 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4382 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4383 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4384 DAG.getConstant(-BitWidth, AmtVT));
4385 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4386 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4387 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4388 Tmp4, Tmp6, ISD::SETLE);
4389 SDValue OutOps[] = { OutLo, OutHi };
4390 return DAG.getMergeValues(OutOps, 2, dl);
4393 //===----------------------------------------------------------------------===//
4394 // Vector related lowering.
4397 /// BuildSplatI - Build a canonical splati of Val with an element size of
4398 /// SplatSize. Cast the result to VT.
4399 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4400 SelectionDAG &DAG, DebugLoc dl) {
4401 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4403 static const EVT VTys[] = { // canonical VT to use for each size.
4404 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4407 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4409 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4413 EVT CanonicalVT = VTys[SplatSize-1];
4415 // Build a canonical splat for this value.
4416 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4417 SmallVector<SDValue, 8> Ops;
4418 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4419 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4420 &Ops[0], Ops.size());
4421 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4424 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4425 /// specified intrinsic ID.
4426 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4427 SelectionDAG &DAG, DebugLoc dl,
4428 EVT DestVT = MVT::Other) {
4429 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4430 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4431 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4434 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4435 /// specified intrinsic ID.
4436 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4437 SDValue Op2, SelectionDAG &DAG,
4438 DebugLoc dl, EVT DestVT = MVT::Other) {
4439 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4440 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4441 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4445 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4446 /// amount. The result has the specified value type.
4447 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4448 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4449 // Force LHS/RHS to be the right type.
4450 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4451 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4454 for (unsigned i = 0; i != 16; ++i)
4456 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4457 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4460 // If this is a case we can't handle, return null and let the default
4461 // expansion code take care of it. If we CAN select this case, and if it
4462 // selects to a single instruction, return Op. Otherwise, if we can codegen
4463 // this case more efficiently than a constant pool load, lower it to the
4464 // sequence of ops that should be used.
4465 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4466 SelectionDAG &DAG) const {
4467 DebugLoc dl = Op.getDebugLoc();
4468 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4469 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4471 // Check if this is a splat of a constant value.
4472 APInt APSplatBits, APSplatUndef;
4473 unsigned SplatBitSize;
4475 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4476 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4479 unsigned SplatBits = APSplatBits.getZExtValue();
4480 unsigned SplatUndef = APSplatUndef.getZExtValue();
4481 unsigned SplatSize = SplatBitSize / 8;
4483 // First, handle single instruction cases.
4486 if (SplatBits == 0) {
4487 // Canonicalize all zero vectors to be v4i32.
4488 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4489 SDValue Z = DAG.getConstant(0, MVT::i32);
4490 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4491 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4496 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4497 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4499 if (SextVal >= -16 && SextVal <= 15)
4500 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4503 // Two instruction sequences.
4505 // If this value is in the range [-32,30] and is even, use:
4506 // tmp = VSPLTI[bhw], result = add tmp, tmp
4507 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4508 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4509 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4510 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4513 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4514 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4516 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4517 // Make -1 and vspltisw -1:
4518 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4520 // Make the VSLW intrinsic, computing 0x8000_0000.
4521 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4524 // xor by OnesV to invert it.
4525 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4526 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4529 // Check to see if this is a wide variety of vsplti*, binop self cases.
4530 static const signed char SplatCsts[] = {
4531 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4532 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4535 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4536 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4537 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4538 int i = SplatCsts[idx];
4540 // Figure out what shift amount will be used by altivec if shifted by i in
4542 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4544 // vsplti + shl self.
4545 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4546 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4547 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4548 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4549 Intrinsic::ppc_altivec_vslw
4551 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4552 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4555 // vsplti + srl self.
4556 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4557 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4558 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4559 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4560 Intrinsic::ppc_altivec_vsrw
4562 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4563 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4566 // vsplti + sra self.
4567 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4568 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4569 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4570 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4571 Intrinsic::ppc_altivec_vsraw
4573 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4574 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4577 // vsplti + rol self.
4578 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4579 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4580 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4581 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4582 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4583 Intrinsic::ppc_altivec_vrlw
4585 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4586 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4589 // t = vsplti c, result = vsldoi t, t, 1
4590 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
4591 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4592 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4594 // t = vsplti c, result = vsldoi t, t, 2
4595 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
4596 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4597 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4599 // t = vsplti c, result = vsldoi t, t, 3
4600 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4601 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4602 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4606 // Three instruction sequences.
4608 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4609 if (SextVal >= 0 && SextVal <= 31) {
4610 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4611 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4612 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4613 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4615 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4616 if (SextVal >= -31 && SextVal <= 0) {
4617 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4618 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4619 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4620 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4626 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4627 /// the specified operations to build the shuffle.
4628 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4629 SDValue RHS, SelectionDAG &DAG,
4631 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4632 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4633 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4636 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4648 if (OpNum == OP_COPY) {
4649 if (LHSID == (1*9+2)*9+3) return LHS;
4650 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4654 SDValue OpLHS, OpRHS;
4655 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4656 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4660 default: llvm_unreachable("Unknown i32 permute!");
4662 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4663 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4664 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4665 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4668 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4669 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4670 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4671 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4674 for (unsigned i = 0; i != 16; ++i)
4675 ShufIdxs[i] = (i&3)+0;
4678 for (unsigned i = 0; i != 16; ++i)
4679 ShufIdxs[i] = (i&3)+4;
4682 for (unsigned i = 0; i != 16; ++i)
4683 ShufIdxs[i] = (i&3)+8;
4686 for (unsigned i = 0; i != 16; ++i)
4687 ShufIdxs[i] = (i&3)+12;
4690 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4692 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4694 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4696 EVT VT = OpLHS.getValueType();
4697 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4698 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4699 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4700 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4703 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4704 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4705 /// return the code it can be lowered into. Worst case, it can always be
4706 /// lowered into a vperm.
4707 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4708 SelectionDAG &DAG) const {
4709 DebugLoc dl = Op.getDebugLoc();
4710 SDValue V1 = Op.getOperand(0);
4711 SDValue V2 = Op.getOperand(1);
4712 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4713 EVT VT = Op.getValueType();
4715 // Cases that are handled by instructions that take permute immediates
4716 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4717 // selected by the instruction selector.
4718 if (V2.getOpcode() == ISD::UNDEF) {
4719 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4720 PPC::isSplatShuffleMask(SVOp, 2) ||
4721 PPC::isSplatShuffleMask(SVOp, 4) ||
4722 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4723 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4724 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4725 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4726 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4727 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4728 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4729 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4730 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4735 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4736 // and produce a fixed permutation. If any of these match, do not lower to
4738 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4739 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4740 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4741 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4742 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4743 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4744 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4745 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4746 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4749 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4750 // perfect shuffle table to emit an optimal matching sequence.
4751 ArrayRef<int> PermMask = SVOp->getMask();
4753 unsigned PFIndexes[4];
4754 bool isFourElementShuffle = true;
4755 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4756 unsigned EltNo = 8; // Start out undef.
4757 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4758 if (PermMask[i*4+j] < 0)
4759 continue; // Undef, ignore it.
4761 unsigned ByteSource = PermMask[i*4+j];
4762 if ((ByteSource & 3) != j) {
4763 isFourElementShuffle = false;
4768 EltNo = ByteSource/4;
4769 } else if (EltNo != ByteSource/4) {
4770 isFourElementShuffle = false;
4774 PFIndexes[i] = EltNo;
4777 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4778 // perfect shuffle vector to determine if it is cost effective to do this as
4779 // discrete instructions, or whether we should use a vperm.
4780 if (isFourElementShuffle) {
4781 // Compute the index in the perfect shuffle table.
4782 unsigned PFTableIndex =
4783 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4785 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4786 unsigned Cost = (PFEntry >> 30);
4788 // Determining when to avoid vperm is tricky. Many things affect the cost
4789 // of vperm, particularly how many times the perm mask needs to be computed.
4790 // For example, if the perm mask can be hoisted out of a loop or is already
4791 // used (perhaps because there are multiple permutes with the same shuffle
4792 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4793 // the loop requires an extra register.
4795 // As a compromise, we only emit discrete instructions if the shuffle can be
4796 // generated in 3 or fewer operations. When we have loop information
4797 // available, if this block is within a loop, we should avoid using vperm
4798 // for 3-operation perms and use a constant pool load instead.
4800 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4803 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4804 // vector that will get spilled to the constant pool.
4805 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4807 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4808 // that it is in input element units, not in bytes. Convert now.
4809 EVT EltVT = V1.getValueType().getVectorElementType();
4810 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4812 SmallVector<SDValue, 16> ResultMask;
4813 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4814 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4816 for (unsigned j = 0; j != BytesPerElement; ++j)
4817 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4821 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4822 &ResultMask[0], ResultMask.size());
4823 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4826 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4827 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4828 /// information about the intrinsic.
4829 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4831 unsigned IntrinsicID =
4832 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4835 switch (IntrinsicID) {
4836 default: return false;
4837 // Comparison predicates.
4838 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4839 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4840 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4841 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4842 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4843 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4844 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4845 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4846 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4847 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4848 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4849 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4850 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4852 // Normal Comparisons.
4853 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4854 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4855 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4856 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4857 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4858 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4859 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4860 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4861 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4862 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4863 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4864 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4865 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4870 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4871 /// lower, do it, otherwise return null.
4872 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4873 SelectionDAG &DAG) const {
4874 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4875 // opcode number of the comparison.
4876 DebugLoc dl = Op.getDebugLoc();
4879 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4880 return SDValue(); // Don't custom lower most intrinsics.
4882 // If this is a non-dot comparison, make the VCMP node and we are done.
4884 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4885 Op.getOperand(1), Op.getOperand(2),
4886 DAG.getConstant(CompareOpc, MVT::i32));
4887 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4890 // Create the PPCISD altivec 'dot' comparison node.
4892 Op.getOperand(2), // LHS
4893 Op.getOperand(3), // RHS
4894 DAG.getConstant(CompareOpc, MVT::i32)
4896 std::vector<EVT> VTs;
4897 VTs.push_back(Op.getOperand(2).getValueType());
4898 VTs.push_back(MVT::Glue);
4899 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4901 // Now that we have the comparison, emit a copy from the CR to a GPR.
4902 // This is flagged to the above dot comparison.
4903 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4904 DAG.getRegister(PPC::CR6, MVT::i32),
4905 CompNode.getValue(1));
4907 // Unpack the result based on how the target uses it.
4908 unsigned BitNo; // Bit # of CR6.
4909 bool InvertBit; // Invert result?
4910 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4911 default: // Can't happen, don't crash on invalid number though.
4912 case 0: // Return the value of the EQ bit of CR6.
4913 BitNo = 0; InvertBit = false;
4915 case 1: // Return the inverted value of the EQ bit of CR6.
4916 BitNo = 0; InvertBit = true;
4918 case 2: // Return the value of the LT bit of CR6.
4919 BitNo = 2; InvertBit = false;
4921 case 3: // Return the inverted value of the LT bit of CR6.
4922 BitNo = 2; InvertBit = true;
4926 // Shift the bit into the low position.
4927 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4928 DAG.getConstant(8-(3-BitNo), MVT::i32));
4930 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4931 DAG.getConstant(1, MVT::i32));
4933 // If we are supposed to, toggle the bit.
4935 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4936 DAG.getConstant(1, MVT::i32));
4940 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4941 SelectionDAG &DAG) const {
4942 DebugLoc dl = Op.getDebugLoc();
4943 // Create a stack slot that is 16-byte aligned.
4944 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4945 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4946 EVT PtrVT = getPointerTy();
4947 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4949 // Store the input value into Value#0 of the stack slot.
4950 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4951 Op.getOperand(0), FIdx, MachinePointerInfo(),
4954 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4955 false, false, false, 0);
4958 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4959 DebugLoc dl = Op.getDebugLoc();
4960 if (Op.getValueType() == MVT::v4i32) {
4961 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4963 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4964 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4966 SDValue RHSSwap = // = vrlw RHS, 16
4967 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4969 // Shrinkify inputs to v8i16.
4970 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4971 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4972 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4974 // Low parts multiplied together, generating 32-bit results (we ignore the
4976 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4977 LHS, RHS, DAG, dl, MVT::v4i32);
4979 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4980 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4981 // Shift the high parts up 16 bits.
4982 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4984 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4985 } else if (Op.getValueType() == MVT::v8i16) {
4986 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4988 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4990 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4991 LHS, RHS, Zero, DAG, dl);
4992 } else if (Op.getValueType() == MVT::v16i8) {
4993 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4995 // Multiply the even 8-bit parts, producing 16-bit sums.
4996 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4997 LHS, RHS, DAG, dl, MVT::v8i16);
4998 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5000 // Multiply the odd 8-bit parts, producing 16-bit sums.
5001 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5002 LHS, RHS, DAG, dl, MVT::v8i16);
5003 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5005 // Merge the results together.
5007 for (unsigned i = 0; i != 8; ++i) {
5009 Ops[i*2+1] = 2*i+1+16;
5011 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5013 llvm_unreachable("Unknown mul to lower!");
5017 /// LowerOperation - Provide custom lowering hooks for some operations.
5019 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5020 switch (Op.getOpcode()) {
5021 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5022 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5023 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5024 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5025 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5026 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5027 case ISD::SETCC: return LowerSETCC(Op, DAG);
5028 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5029 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5031 return LowerVASTART(Op, DAG, PPCSubTarget);
5034 return LowerVAARG(Op, DAG, PPCSubTarget);
5036 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5037 case ISD::DYNAMIC_STACKALLOC:
5038 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5040 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5041 case ISD::FP_TO_UINT:
5042 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5044 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5045 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5047 // Lower 64-bit shifts.
5048 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5049 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5050 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5052 // Vector-related lowering.
5053 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5054 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5055 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5056 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5057 case ISD::MUL: return LowerMUL(Op, DAG);
5059 // Frame & Return address.
5060 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5061 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5065 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5066 SmallVectorImpl<SDValue>&Results,
5067 SelectionDAG &DAG) const {
5068 const TargetMachine &TM = getTargetMachine();
5069 DebugLoc dl = N->getDebugLoc();
5070 switch (N->getOpcode()) {
5072 llvm_unreachable("Do not know how to custom type legalize this operation!");
5074 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5075 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5078 EVT VT = N->getValueType(0);
5080 if (VT == MVT::i64) {
5081 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5083 Results.push_back(NewNode);
5084 Results.push_back(NewNode.getValue(1));
5088 case ISD::FP_ROUND_INREG: {
5089 assert(N->getValueType(0) == MVT::ppcf128);
5090 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5091 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5092 MVT::f64, N->getOperand(0),
5093 DAG.getIntPtrConstant(0));
5094 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5095 MVT::f64, N->getOperand(0),
5096 DAG.getIntPtrConstant(1));
5098 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5099 // of the long double, and puts FPSCR back the way it was. We do not
5100 // actually model FPSCR.
5101 std::vector<EVT> NodeTys;
5102 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5104 NodeTys.push_back(MVT::f64); // Return register
5105 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5106 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5107 MFFSreg = Result.getValue(0);
5108 InFlag = Result.getValue(1);
5111 NodeTys.push_back(MVT::Glue); // Returns a flag
5112 Ops[0] = DAG.getConstant(31, MVT::i32);
5114 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5115 InFlag = Result.getValue(0);
5118 NodeTys.push_back(MVT::Glue); // Returns a flag
5119 Ops[0] = DAG.getConstant(30, MVT::i32);
5121 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5122 InFlag = Result.getValue(0);
5125 NodeTys.push_back(MVT::f64); // result of add
5126 NodeTys.push_back(MVT::Glue); // Returns a flag
5130 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5131 FPreg = Result.getValue(0);
5132 InFlag = Result.getValue(1);
5135 NodeTys.push_back(MVT::f64);
5136 Ops[0] = DAG.getConstant(1, MVT::i32);
5140 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5141 FPreg = Result.getValue(0);
5143 // We know the low half is about to be thrown away, so just use something
5145 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5149 case ISD::FP_TO_SINT:
5150 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5156 //===----------------------------------------------------------------------===//
5157 // Other Lowering Code
5158 //===----------------------------------------------------------------------===//
5161 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5162 bool is64bit, unsigned BinOpcode) const {
5163 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5164 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5166 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5167 MachineFunction *F = BB->getParent();
5168 MachineFunction::iterator It = BB;
5171 unsigned dest = MI->getOperand(0).getReg();
5172 unsigned ptrA = MI->getOperand(1).getReg();
5173 unsigned ptrB = MI->getOperand(2).getReg();
5174 unsigned incr = MI->getOperand(3).getReg();
5175 DebugLoc dl = MI->getDebugLoc();
5177 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5178 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5179 F->insert(It, loopMBB);
5180 F->insert(It, exitMBB);
5181 exitMBB->splice(exitMBB->begin(), BB,
5182 llvm::next(MachineBasicBlock::iterator(MI)),
5184 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5186 MachineRegisterInfo &RegInfo = F->getRegInfo();
5187 unsigned TmpReg = (!BinOpcode) ? incr :
5188 RegInfo.createVirtualRegister(
5189 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5190 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5194 // fallthrough --> loopMBB
5195 BB->addSuccessor(loopMBB);
5198 // l[wd]arx dest, ptr
5199 // add r0, dest, incr
5200 // st[wd]cx. r0, ptr
5202 // fallthrough --> exitMBB
5204 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5205 .addReg(ptrA).addReg(ptrB);
5207 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5208 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5209 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5210 BuildMI(BB, dl, TII->get(PPC::BCC))
5211 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5212 BB->addSuccessor(loopMBB);
5213 BB->addSuccessor(exitMBB);
5222 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5223 MachineBasicBlock *BB,
5224 bool is8bit, // operation
5225 unsigned BinOpcode) const {
5226 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5227 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5228 // In 64 bit mode we have to use 64 bits for addresses, even though the
5229 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5230 // registers without caring whether they're 32 or 64, but here we're
5231 // doing actual arithmetic on the addresses.
5232 bool is64bit = PPCSubTarget.isPPC64();
5233 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5235 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5236 MachineFunction *F = BB->getParent();
5237 MachineFunction::iterator It = BB;
5240 unsigned dest = MI->getOperand(0).getReg();
5241 unsigned ptrA = MI->getOperand(1).getReg();
5242 unsigned ptrB = MI->getOperand(2).getReg();
5243 unsigned incr = MI->getOperand(3).getReg();
5244 DebugLoc dl = MI->getDebugLoc();
5246 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5247 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5248 F->insert(It, loopMBB);
5249 F->insert(It, exitMBB);
5250 exitMBB->splice(exitMBB->begin(), BB,
5251 llvm::next(MachineBasicBlock::iterator(MI)),
5253 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5255 MachineRegisterInfo &RegInfo = F->getRegInfo();
5256 const TargetRegisterClass *RC =
5257 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5258 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5259 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5260 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5261 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5262 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5263 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5264 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5265 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5266 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5267 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5268 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5269 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5271 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5275 // fallthrough --> loopMBB
5276 BB->addSuccessor(loopMBB);
5278 // The 4-byte load must be aligned, while a char or short may be
5279 // anywhere in the word. Hence all this nasty bookkeeping code.
5280 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5281 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5282 // xori shift, shift1, 24 [16]
5283 // rlwinm ptr, ptr1, 0, 0, 29
5284 // slw incr2, incr, shift
5285 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5286 // slw mask, mask2, shift
5288 // lwarx tmpDest, ptr
5289 // add tmp, tmpDest, incr2
5290 // andc tmp2, tmpDest, mask
5291 // and tmp3, tmp, mask
5292 // or tmp4, tmp3, tmp2
5295 // fallthrough --> exitMBB
5296 // srw dest, tmpDest, shift
5297 if (ptrA != ZeroReg) {
5298 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5299 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5300 .addReg(ptrA).addReg(ptrB);
5304 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5305 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5306 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5307 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5309 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5310 .addReg(Ptr1Reg).addImm(0).addImm(61);
5312 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5313 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5314 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5315 .addReg(incr).addReg(ShiftReg);
5317 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5319 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5320 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5322 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5323 .addReg(Mask2Reg).addReg(ShiftReg);
5326 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5327 .addReg(ZeroReg).addReg(PtrReg);
5329 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5330 .addReg(Incr2Reg).addReg(TmpDestReg);
5331 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5332 .addReg(TmpDestReg).addReg(MaskReg);
5333 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5334 .addReg(TmpReg).addReg(MaskReg);
5335 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5336 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5337 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5338 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5339 BuildMI(BB, dl, TII->get(PPC::BCC))
5340 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5341 BB->addSuccessor(loopMBB);
5342 BB->addSuccessor(exitMBB);
5347 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5353 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5354 MachineBasicBlock *BB) const {
5355 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5357 // To "insert" these instructions we actually have to insert their
5358 // control-flow patterns.
5359 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5360 MachineFunction::iterator It = BB;
5363 MachineFunction *F = BB->getParent();
5365 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5366 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5367 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5368 PPC::ISEL8 : PPC::ISEL;
5369 unsigned SelectPred = MI->getOperand(4).getImm();
5370 DebugLoc dl = MI->getDebugLoc();
5372 // The SelectPred is ((BI << 5) | BO) for a BCC
5373 unsigned BO = SelectPred & 0xF;
5374 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5376 unsigned TrueOpNo, FalseOpNo;
5383 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5386 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5387 .addReg(MI->getOperand(TrueOpNo).getReg())
5388 .addReg(MI->getOperand(FalseOpNo).getReg())
5389 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5390 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5391 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5392 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5393 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5394 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5397 // The incoming instruction knows the destination vreg to set, the
5398 // condition code register to branch on, the true/false values to
5399 // select between, and a branch opcode to use.
5404 // cmpTY ccX, r1, r2
5406 // fallthrough --> copy0MBB
5407 MachineBasicBlock *thisMBB = BB;
5408 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5409 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5410 unsigned SelectPred = MI->getOperand(4).getImm();
5411 DebugLoc dl = MI->getDebugLoc();
5412 F->insert(It, copy0MBB);
5413 F->insert(It, sinkMBB);
5415 // Transfer the remainder of BB and its successor edges to sinkMBB.
5416 sinkMBB->splice(sinkMBB->begin(), BB,
5417 llvm::next(MachineBasicBlock::iterator(MI)),
5419 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5421 // Next, add the true and fallthrough blocks as its successors.
5422 BB->addSuccessor(copy0MBB);
5423 BB->addSuccessor(sinkMBB);
5425 BuildMI(BB, dl, TII->get(PPC::BCC))
5426 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5429 // %FalseValue = ...
5430 // # fallthrough to sinkMBB
5433 // Update machine-CFG edges
5434 BB->addSuccessor(sinkMBB);
5437 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5440 BuildMI(*BB, BB->begin(), dl,
5441 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5442 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5443 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5445 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5446 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5447 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5448 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5449 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5450 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5451 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5452 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5454 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5455 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5456 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5457 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5458 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5459 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5460 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5461 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5463 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5464 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5465 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5466 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5467 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5468 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5469 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5470 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5472 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5473 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5474 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5475 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5476 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5477 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5478 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5479 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5481 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5482 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5483 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5484 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5485 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5486 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5487 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5488 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5490 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5491 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5492 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5493 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5494 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5495 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5496 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5497 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5499 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5500 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5501 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5502 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5503 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5504 BB = EmitAtomicBinary(MI, BB, false, 0);
5505 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5506 BB = EmitAtomicBinary(MI, BB, true, 0);
5508 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5509 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5510 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5512 unsigned dest = MI->getOperand(0).getReg();
5513 unsigned ptrA = MI->getOperand(1).getReg();
5514 unsigned ptrB = MI->getOperand(2).getReg();
5515 unsigned oldval = MI->getOperand(3).getReg();
5516 unsigned newval = MI->getOperand(4).getReg();
5517 DebugLoc dl = MI->getDebugLoc();
5519 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5520 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5521 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5522 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5523 F->insert(It, loop1MBB);
5524 F->insert(It, loop2MBB);
5525 F->insert(It, midMBB);
5526 F->insert(It, exitMBB);
5527 exitMBB->splice(exitMBB->begin(), BB,
5528 llvm::next(MachineBasicBlock::iterator(MI)),
5530 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5534 // fallthrough --> loopMBB
5535 BB->addSuccessor(loop1MBB);
5538 // l[wd]arx dest, ptr
5539 // cmp[wd] dest, oldval
5542 // st[wd]cx. newval, ptr
5546 // st[wd]cx. dest, ptr
5549 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5550 .addReg(ptrA).addReg(ptrB);
5551 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5552 .addReg(oldval).addReg(dest);
5553 BuildMI(BB, dl, TII->get(PPC::BCC))
5554 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5555 BB->addSuccessor(loop2MBB);
5556 BB->addSuccessor(midMBB);
5559 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5560 .addReg(newval).addReg(ptrA).addReg(ptrB);
5561 BuildMI(BB, dl, TII->get(PPC::BCC))
5562 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5563 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5564 BB->addSuccessor(loop1MBB);
5565 BB->addSuccessor(exitMBB);
5568 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5569 .addReg(dest).addReg(ptrA).addReg(ptrB);
5570 BB->addSuccessor(exitMBB);
5575 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5576 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5577 // We must use 64-bit registers for addresses when targeting 64-bit,
5578 // since we're actually doing arithmetic on them. Other registers
5580 bool is64bit = PPCSubTarget.isPPC64();
5581 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5583 unsigned dest = MI->getOperand(0).getReg();
5584 unsigned ptrA = MI->getOperand(1).getReg();
5585 unsigned ptrB = MI->getOperand(2).getReg();
5586 unsigned oldval = MI->getOperand(3).getReg();
5587 unsigned newval = MI->getOperand(4).getReg();
5588 DebugLoc dl = MI->getDebugLoc();
5590 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5591 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5592 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5593 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5594 F->insert(It, loop1MBB);
5595 F->insert(It, loop2MBB);
5596 F->insert(It, midMBB);
5597 F->insert(It, exitMBB);
5598 exitMBB->splice(exitMBB->begin(), BB,
5599 llvm::next(MachineBasicBlock::iterator(MI)),
5601 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5603 MachineRegisterInfo &RegInfo = F->getRegInfo();
5604 const TargetRegisterClass *RC =
5605 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5606 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5607 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5608 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5609 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5610 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5611 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5612 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5613 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5614 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5615 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5616 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5617 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5618 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5619 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5621 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5622 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5625 // fallthrough --> loopMBB
5626 BB->addSuccessor(loop1MBB);
5628 // The 4-byte load must be aligned, while a char or short may be
5629 // anywhere in the word. Hence all this nasty bookkeeping code.
5630 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5631 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5632 // xori shift, shift1, 24 [16]
5633 // rlwinm ptr, ptr1, 0, 0, 29
5634 // slw newval2, newval, shift
5635 // slw oldval2, oldval,shift
5636 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5637 // slw mask, mask2, shift
5638 // and newval3, newval2, mask
5639 // and oldval3, oldval2, mask
5641 // lwarx tmpDest, ptr
5642 // and tmp, tmpDest, mask
5643 // cmpw tmp, oldval3
5646 // andc tmp2, tmpDest, mask
5647 // or tmp4, tmp2, newval3
5652 // stwcx. tmpDest, ptr
5654 // srw dest, tmpDest, shift
5655 if (ptrA != ZeroReg) {
5656 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5657 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5658 .addReg(ptrA).addReg(ptrB);
5662 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5663 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5664 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5665 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5667 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5668 .addReg(Ptr1Reg).addImm(0).addImm(61);
5670 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5671 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5672 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
5673 .addReg(newval).addReg(ShiftReg);
5674 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
5675 .addReg(oldval).addReg(ShiftReg);
5677 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5679 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5680 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5681 .addReg(Mask3Reg).addImm(65535);
5683 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5684 .addReg(Mask2Reg).addReg(ShiftReg);
5685 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5686 .addReg(NewVal2Reg).addReg(MaskReg);
5687 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5688 .addReg(OldVal2Reg).addReg(MaskReg);
5691 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5692 .addReg(ZeroReg).addReg(PtrReg);
5693 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5694 .addReg(TmpDestReg).addReg(MaskReg);
5695 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5696 .addReg(TmpReg).addReg(OldVal3Reg);
5697 BuildMI(BB, dl, TII->get(PPC::BCC))
5698 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5699 BB->addSuccessor(loop2MBB);
5700 BB->addSuccessor(midMBB);
5703 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5704 .addReg(TmpDestReg).addReg(MaskReg);
5705 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5706 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5707 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5708 .addReg(ZeroReg).addReg(PtrReg);
5709 BuildMI(BB, dl, TII->get(PPC::BCC))
5710 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5711 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5712 BB->addSuccessor(loop1MBB);
5713 BB->addSuccessor(exitMBB);
5716 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5717 .addReg(ZeroReg).addReg(PtrReg);
5718 BB->addSuccessor(exitMBB);
5723 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5726 llvm_unreachable("Unexpected instr type to insert");
5729 MI->eraseFromParent(); // The pseudo instruction is gone now.
5733 //===----------------------------------------------------------------------===//
5734 // Target Optimization Hooks
5735 //===----------------------------------------------------------------------===//
5737 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5738 DAGCombinerInfo &DCI) const {
5739 const TargetMachine &TM = getTargetMachine();
5740 SelectionDAG &DAG = DCI.DAG;
5741 DebugLoc dl = N->getDebugLoc();
5742 switch (N->getOpcode()) {
5745 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5746 if (C->isNullValue()) // 0 << V -> 0.
5747 return N->getOperand(0);
5751 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5752 if (C->isNullValue()) // 0 >>u V -> 0.
5753 return N->getOperand(0);
5757 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5758 if (C->isNullValue() || // 0 >>s V -> 0.
5759 C->isAllOnesValue()) // -1 >>s V -> -1.
5760 return N->getOperand(0);
5764 case ISD::SINT_TO_FP:
5765 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5766 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5767 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5768 // We allow the src/dst to be either f32/f64, but the intermediate
5769 // type must be i64.
5770 if (N->getOperand(0).getValueType() == MVT::i64 &&
5771 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5772 SDValue Val = N->getOperand(0).getOperand(0);
5773 if (Val.getValueType() == MVT::f32) {
5774 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5775 DCI.AddToWorklist(Val.getNode());
5778 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5779 DCI.AddToWorklist(Val.getNode());
5780 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5781 DCI.AddToWorklist(Val.getNode());
5782 if (N->getValueType(0) == MVT::f32) {
5783 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5784 DAG.getIntPtrConstant(0));
5785 DCI.AddToWorklist(Val.getNode());
5788 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5789 // If the intermediate type is i32, we can avoid the load/store here
5796 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5797 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5798 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5799 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5800 N->getOperand(1).getValueType() == MVT::i32 &&
5801 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5802 SDValue Val = N->getOperand(1).getOperand(0);
5803 if (Val.getValueType() == MVT::f32) {
5804 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5805 DCI.AddToWorklist(Val.getNode());
5807 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5808 DCI.AddToWorklist(Val.getNode());
5810 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5811 N->getOperand(2), N->getOperand(3));
5812 DCI.AddToWorklist(Val.getNode());
5816 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5817 if (cast<StoreSDNode>(N)->isUnindexed() &&
5818 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5819 N->getOperand(1).getNode()->hasOneUse() &&
5820 (N->getOperand(1).getValueType() == MVT::i32 ||
5821 N->getOperand(1).getValueType() == MVT::i16)) {
5822 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5823 // Do an any-extend to 32-bits if this is a half-word input.
5824 if (BSwapOp.getValueType() == MVT::i16)
5825 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5828 N->getOperand(0), BSwapOp, N->getOperand(2),
5829 DAG.getValueType(N->getOperand(1).getValueType())
5832 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5833 Ops, array_lengthof(Ops),
5834 cast<StoreSDNode>(N)->getMemoryVT(),
5835 cast<StoreSDNode>(N)->getMemOperand());
5839 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5840 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5841 N->getOperand(0).hasOneUse() &&
5842 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5843 SDValue Load = N->getOperand(0);
5844 LoadSDNode *LD = cast<LoadSDNode>(Load);
5845 // Create the byte-swapping load.
5847 LD->getChain(), // Chain
5848 LD->getBasePtr(), // Ptr
5849 DAG.getValueType(N->getValueType(0)) // VT
5852 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5853 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5854 LD->getMemoryVT(), LD->getMemOperand());
5856 // If this is an i16 load, insert the truncate.
5857 SDValue ResVal = BSLoad;
5858 if (N->getValueType(0) == MVT::i16)
5859 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5861 // First, combine the bswap away. This makes the value produced by the
5863 DCI.CombineTo(N, ResVal);
5865 // Next, combine the load away, we give it a bogus result value but a real
5866 // chain result. The result value is dead because the bswap is dead.
5867 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5869 // Return N so it doesn't get rechecked!
5870 return SDValue(N, 0);
5874 case PPCISD::VCMP: {
5875 // If a VCMPo node already exists with exactly the same operands as this
5876 // node, use its result instead of this node (VCMPo computes both a CR6 and
5877 // a normal output).
5879 if (!N->getOperand(0).hasOneUse() &&
5880 !N->getOperand(1).hasOneUse() &&
5881 !N->getOperand(2).hasOneUse()) {
5883 // Scan all of the users of the LHS, looking for VCMPo's that match.
5884 SDNode *VCMPoNode = 0;
5886 SDNode *LHSN = N->getOperand(0).getNode();
5887 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5889 if (UI->getOpcode() == PPCISD::VCMPo &&
5890 UI->getOperand(1) == N->getOperand(1) &&
5891 UI->getOperand(2) == N->getOperand(2) &&
5892 UI->getOperand(0) == N->getOperand(0)) {
5897 // If there is no VCMPo node, or if the flag value has a single use, don't
5899 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5902 // Look at the (necessarily single) use of the flag value. If it has a
5903 // chain, this transformation is more complex. Note that multiple things
5904 // could use the value result, which we should ignore.
5905 SDNode *FlagUser = 0;
5906 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5907 FlagUser == 0; ++UI) {
5908 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5910 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5911 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5918 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5919 // give up for right now.
5920 if (FlagUser->getOpcode() == PPCISD::MFCR)
5921 return SDValue(VCMPoNode, 0);
5926 // If this is a branch on an altivec predicate comparison, lower this so
5927 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5928 // lowering is done pre-legalize, because the legalizer lowers the predicate
5929 // compare down to code that is difficult to reassemble.
5930 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5931 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5935 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5936 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5937 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5938 assert(isDot && "Can't compare against a vector result!");
5940 // If this is a comparison against something other than 0/1, then we know
5941 // that the condition is never/always true.
5942 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5943 if (Val != 0 && Val != 1) {
5944 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5945 return N->getOperand(0);
5946 // Always !=, turn it into an unconditional branch.
5947 return DAG.getNode(ISD::BR, dl, MVT::Other,
5948 N->getOperand(0), N->getOperand(4));
5951 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5953 // Create the PPCISD altivec 'dot' comparison node.
5954 std::vector<EVT> VTs;
5956 LHS.getOperand(2), // LHS of compare
5957 LHS.getOperand(3), // RHS of compare
5958 DAG.getConstant(CompareOpc, MVT::i32)
5960 VTs.push_back(LHS.getOperand(2).getValueType());
5961 VTs.push_back(MVT::Glue);
5962 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5964 // Unpack the result based on how the target uses it.
5965 PPC::Predicate CompOpc;
5966 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5967 default: // Can't happen, don't crash on invalid number though.
5968 case 0: // Branch on the value of the EQ bit of CR6.
5969 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5971 case 1: // Branch on the inverted value of the EQ bit of CR6.
5972 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5974 case 2: // Branch on the value of the LT bit of CR6.
5975 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5977 case 3: // Branch on the inverted value of the LT bit of CR6.
5978 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5982 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5983 DAG.getConstant(CompOpc, MVT::i32),
5984 DAG.getRegister(PPC::CR6, MVT::i32),
5985 N->getOperand(4), CompNode.getValue(1));
5994 //===----------------------------------------------------------------------===//
5995 // Inline Assembly Support
5996 //===----------------------------------------------------------------------===//
5998 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6001 const SelectionDAG &DAG,
6002 unsigned Depth) const {
6003 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6004 switch (Op.getOpcode()) {
6006 case PPCISD::LBRX: {
6007 // lhbrx is known to have the top bits cleared out.
6008 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6009 KnownZero = 0xFFFF0000;
6012 case ISD::INTRINSIC_WO_CHAIN: {
6013 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6015 case Intrinsic::ppc_altivec_vcmpbfp_p:
6016 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6017 case Intrinsic::ppc_altivec_vcmpequb_p:
6018 case Intrinsic::ppc_altivec_vcmpequh_p:
6019 case Intrinsic::ppc_altivec_vcmpequw_p:
6020 case Intrinsic::ppc_altivec_vcmpgefp_p:
6021 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6022 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6023 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6024 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6025 case Intrinsic::ppc_altivec_vcmpgtub_p:
6026 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6027 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6028 KnownZero = ~1U; // All bits but the low one are known to be zero.
6036 /// getConstraintType - Given a constraint, return the type of
6037 /// constraint it is for this target.
6038 PPCTargetLowering::ConstraintType
6039 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6040 if (Constraint.size() == 1) {
6041 switch (Constraint[0]) {
6048 return C_RegisterClass;
6051 return TargetLowering::getConstraintType(Constraint);
6054 /// Examine constraint type and operand type and determine a weight value.
6055 /// This object must already have been set up with the operand type
6056 /// and the current alternative constraint selected.
6057 TargetLowering::ConstraintWeight
6058 PPCTargetLowering::getSingleConstraintMatchWeight(
6059 AsmOperandInfo &info, const char *constraint) const {
6060 ConstraintWeight weight = CW_Invalid;
6061 Value *CallOperandVal = info.CallOperandVal;
6062 // If we don't have a value, we can't do a match,
6063 // but allow it at the lowest weight.
6064 if (CallOperandVal == NULL)
6066 Type *type = CallOperandVal->getType();
6067 // Look at the constraint type.
6068 switch (*constraint) {
6070 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6073 if (type->isIntegerTy())
6074 weight = CW_Register;
6077 if (type->isFloatTy())
6078 weight = CW_Register;
6081 if (type->isDoubleTy())
6082 weight = CW_Register;
6085 if (type->isVectorTy())
6086 weight = CW_Register;
6089 weight = CW_Register;
6095 std::pair<unsigned, const TargetRegisterClass*>
6096 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6098 if (Constraint.size() == 1) {
6099 // GCC RS6000 Constraint Letters
6100 switch (Constraint[0]) {
6103 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6104 return std::make_pair(0U, &PPC::G8RCRegClass);
6105 return std::make_pair(0U, &PPC::GPRCRegClass);
6108 return std::make_pair(0U, &PPC::F4RCRegClass);
6110 return std::make_pair(0U, &PPC::F8RCRegClass);
6113 return std::make_pair(0U, &PPC::VRRCRegClass);
6115 return std::make_pair(0U, &PPC::CRRCRegClass);
6119 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6123 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6124 /// vector. If it is invalid, don't add anything to Ops.
6125 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6126 std::string &Constraint,
6127 std::vector<SDValue>&Ops,
6128 SelectionDAG &DAG) const {
6129 SDValue Result(0,0);
6131 // Only support length 1 constraints.
6132 if (Constraint.length() > 1) return;
6134 char Letter = Constraint[0];
6145 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6146 if (!CST) return; // Must be an immediate to match.
6147 unsigned Value = CST->getZExtValue();
6149 default: llvm_unreachable("Unknown constraint letter!");
6150 case 'I': // "I" is a signed 16-bit constant.
6151 if ((short)Value == (int)Value)
6152 Result = DAG.getTargetConstant(Value, Op.getValueType());
6154 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6155 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6156 if ((short)Value == 0)
6157 Result = DAG.getTargetConstant(Value, Op.getValueType());
6159 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6160 if ((Value >> 16) == 0)
6161 Result = DAG.getTargetConstant(Value, Op.getValueType());
6163 case 'M': // "M" is a constant that is greater than 31.
6165 Result = DAG.getTargetConstant(Value, Op.getValueType());
6167 case 'N': // "N" is a positive constant that is an exact power of two.
6168 if ((int)Value > 0 && isPowerOf2_32(Value))
6169 Result = DAG.getTargetConstant(Value, Op.getValueType());
6171 case 'O': // "O" is the constant zero.
6173 Result = DAG.getTargetConstant(Value, Op.getValueType());
6175 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6176 if ((short)-Value == (int)-Value)
6177 Result = DAG.getTargetConstant(Value, Op.getValueType());
6184 if (Result.getNode()) {
6185 Ops.push_back(Result);
6189 // Handle standard constraint letters.
6190 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6193 // isLegalAddressingMode - Return true if the addressing mode represented
6194 // by AM is legal for this target, for a load/store of the specified type.
6195 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6197 // FIXME: PPC does not allow r+i addressing modes for vectors!
6199 // PPC allows a sign-extended 16-bit immediate field.
6200 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6203 // No global is ever allowed as a base.
6207 // PPC only support r+r,
6209 case 0: // "r+i" or just "i", depending on HasBaseReg.
6212 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6214 // Otherwise we have r+r or r+i.
6217 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6219 // Allow 2*r as r+r.
6222 // No other scales are supported.
6229 /// isLegalAddressImmediate - Return true if the integer value can be used
6230 /// as the offset of the target addressing mode for load / store of the
6232 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6233 // PPC allows a sign-extended 16-bit immediate field.
6234 return (V > -(1 << 16) && V < (1 << 16)-1);
6237 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6241 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6242 SelectionDAG &DAG) const {
6243 MachineFunction &MF = DAG.getMachineFunction();
6244 MachineFrameInfo *MFI = MF.getFrameInfo();
6245 MFI->setReturnAddressIsTaken(true);
6247 DebugLoc dl = Op.getDebugLoc();
6248 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6250 // Make sure the function does not optimize away the store of the RA to
6252 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6253 FuncInfo->setLRStoreRequired();
6254 bool isPPC64 = PPCSubTarget.isPPC64();
6255 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6258 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6261 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6262 isPPC64? MVT::i64 : MVT::i32);
6263 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6264 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6266 MachinePointerInfo(), false, false, false, 0);
6269 // Just load the return address off the stack.
6270 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6271 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6272 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6275 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6276 SelectionDAG &DAG) const {
6277 DebugLoc dl = Op.getDebugLoc();
6278 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6280 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6281 bool isPPC64 = PtrVT == MVT::i64;
6283 MachineFunction &MF = DAG.getMachineFunction();
6284 MachineFrameInfo *MFI = MF.getFrameInfo();
6285 MFI->setFrameAddressIsTaken(true);
6286 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6287 MFI->hasVarSizedObjects()) &&
6288 MFI->getStackSize() &&
6289 !MF.getFunction()->getFnAttributes().hasNakedAttr();
6290 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6291 (is31 ? PPC::R31 : PPC::R1);
6292 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6295 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6296 FrameAddr, MachinePointerInfo(), false, false,
6302 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6303 // The PowerPC target isn't yet aware of offsets.
6307 /// getOptimalMemOpType - Returns the target specific optimal type for load
6308 /// and store operations as a result of memset, memcpy, and memmove
6309 /// lowering. If DstAlign is zero that means it's safe to destination
6310 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6311 /// means there isn't a need to check it against alignment requirement,
6312 /// probably because the source does not need to be loaded. If
6313 /// 'IsZeroVal' is true, that means it's safe to return a
6314 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6315 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6316 /// constant so it does not need to be loaded.
6317 /// It returns EVT::Other if the type should be determined using generic
6318 /// target-independent logic.
6319 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6320 unsigned DstAlign, unsigned SrcAlign,
6323 MachineFunction &MF) const {
6324 if (this->PPCSubTarget.isPPC64()) {
6331 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6332 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6333 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6334 /// is expanded to mul + add.
6335 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6339 switch (VT.getSimpleVT().SimpleTy) {
6351 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6353 return TargetLowering::getSchedulingPreference(N);