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);
1503 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1504 *DAG.getContext(), AS);
1506 TargetLowering::ArgListTy Args;
1507 TargetLowering::ArgListEntry Entry;
1509 Entry.Ty = IntPtrTy;
1510 Entry.Node = Trmp; Args.push_back(Entry);
1512 // TrampSize == (isPPC64 ? 48 : 40);
1513 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1514 isPPC64 ? MVT::i64 : MVT::i32);
1515 Args.push_back(Entry);
1517 Entry.Node = FPtr; Args.push_back(Entry);
1518 Entry.Node = Nest; Args.push_back(Entry);
1520 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1521 TargetLowering::CallLoweringInfo CLI(Chain,
1522 Type::getVoidTy(*DAG.getContext()),
1523 false, false, false, false, 0,
1525 /*isTailCall=*/false,
1526 /*doesNotRet=*/false,
1527 /*isReturnValueUsed=*/true,
1528 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1530 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1532 return CallResult.second;
1535 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1536 const PPCSubtarget &Subtarget) const {
1537 MachineFunction &MF = DAG.getMachineFunction();
1538 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1540 DebugLoc dl = Op.getDebugLoc();
1542 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1543 // vastart just stores the address of the VarArgsFrameIndex slot into the
1544 // memory location argument.
1545 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1546 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1547 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1548 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1549 MachinePointerInfo(SV),
1553 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1554 // We suppose the given va_list is already allocated.
1557 // char gpr; /* index into the array of 8 GPRs
1558 // * stored in the register save area
1559 // * gpr=0 corresponds to r3,
1560 // * gpr=1 to r4, etc.
1562 // char fpr; /* index into the array of 8 FPRs
1563 // * stored in the register save area
1564 // * fpr=0 corresponds to f1,
1565 // * fpr=1 to f2, etc.
1567 // char *overflow_arg_area;
1568 // /* location on stack that holds
1569 // * the next overflow argument
1571 // char *reg_save_area;
1572 // /* where r3:r10 and f1:f8 (if saved)
1578 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1579 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1582 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1584 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1586 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1589 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1590 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1592 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1593 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1595 uint64_t FPROffset = 1;
1596 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1598 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1600 // Store first byte : number of int regs
1601 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1603 MachinePointerInfo(SV),
1604 MVT::i8, false, false, 0);
1605 uint64_t nextOffset = FPROffset;
1606 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1609 // Store second byte : number of float regs
1610 SDValue secondStore =
1611 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1612 MachinePointerInfo(SV, nextOffset), MVT::i8,
1614 nextOffset += StackOffset;
1615 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1617 // Store second word : arguments given on stack
1618 SDValue thirdStore =
1619 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1620 MachinePointerInfo(SV, nextOffset),
1622 nextOffset += FrameOffset;
1623 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1625 // Store third word : arguments given in registers
1626 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1627 MachinePointerInfo(SV, nextOffset),
1632 #include "PPCGenCallingConv.inc"
1634 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1635 CCValAssign::LocInfo &LocInfo,
1636 ISD::ArgFlagsTy &ArgFlags,
1641 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1643 CCValAssign::LocInfo &LocInfo,
1644 ISD::ArgFlagsTy &ArgFlags,
1646 static const uint16_t ArgRegs[] = {
1647 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1648 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1650 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1652 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1654 // Skip one register if the first unallocated register has an even register
1655 // number and there are still argument registers available which have not been
1656 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1657 // need to skip a register if RegNum is odd.
1658 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1659 State.AllocateReg(ArgRegs[RegNum]);
1662 // Always return false here, as this function only makes sure that the first
1663 // unallocated register has an odd register number and does not actually
1664 // allocate a register for the current argument.
1668 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1670 CCValAssign::LocInfo &LocInfo,
1671 ISD::ArgFlagsTy &ArgFlags,
1673 static const uint16_t ArgRegs[] = {
1674 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1678 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1680 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1682 // If there is only one Floating-point register left we need to put both f64
1683 // values of a split ppc_fp128 value on the stack.
1684 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1685 State.AllocateReg(ArgRegs[RegNum]);
1688 // Always return false here, as this function only makes sure that the two f64
1689 // values a ppc_fp128 value is split into are both passed in registers or both
1690 // passed on the stack and does not actually allocate a register for the
1691 // current argument.
1695 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1697 static const uint16_t *GetFPR() {
1698 static const uint16_t FPR[] = {
1699 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1700 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1706 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1708 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1709 unsigned PtrByteSize) {
1710 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1711 if (Flags.isByVal())
1712 ArgSize = Flags.getByValSize();
1713 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1719 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1720 CallingConv::ID CallConv, bool isVarArg,
1721 const SmallVectorImpl<ISD::InputArg>
1723 DebugLoc dl, SelectionDAG &DAG,
1724 SmallVectorImpl<SDValue> &InVals)
1726 if (PPCSubTarget.isSVR4ABI()) {
1727 if (PPCSubTarget.isPPC64())
1728 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1731 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1734 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1740 PPCTargetLowering::LowerFormalArguments_32SVR4(
1742 CallingConv::ID CallConv, bool isVarArg,
1743 const SmallVectorImpl<ISD::InputArg>
1745 DebugLoc dl, SelectionDAG &DAG,
1746 SmallVectorImpl<SDValue> &InVals) const {
1748 // 32-bit SVR4 ABI Stack Frame Layout:
1749 // +-----------------------------------+
1750 // +--> | Back chain |
1751 // | +-----------------------------------+
1752 // | | Floating-point register save area |
1753 // | +-----------------------------------+
1754 // | | General register save area |
1755 // | +-----------------------------------+
1756 // | | CR save word |
1757 // | +-----------------------------------+
1758 // | | VRSAVE save word |
1759 // | +-----------------------------------+
1760 // | | Alignment padding |
1761 // | +-----------------------------------+
1762 // | | Vector register save area |
1763 // | +-----------------------------------+
1764 // | | Local variable space |
1765 // | +-----------------------------------+
1766 // | | Parameter list area |
1767 // | +-----------------------------------+
1768 // | | LR save word |
1769 // | +-----------------------------------+
1770 // SP--> +--- | Back chain |
1771 // +-----------------------------------+
1774 // System V Application Binary Interface PowerPC Processor Supplement
1775 // AltiVec Technology Programming Interface Manual
1777 MachineFunction &MF = DAG.getMachineFunction();
1778 MachineFrameInfo *MFI = MF.getFrameInfo();
1779 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1781 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1782 // Potential tail calls could cause overwriting of argument stack slots.
1783 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1784 (CallConv == CallingConv::Fast));
1785 unsigned PtrByteSize = 4;
1787 // Assign locations to all of the incoming arguments.
1788 SmallVector<CCValAssign, 16> ArgLocs;
1789 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1790 getTargetMachine(), ArgLocs, *DAG.getContext());
1792 // Reserve space for the linkage area on the stack.
1793 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1795 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1797 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1798 CCValAssign &VA = ArgLocs[i];
1800 // Arguments stored in registers.
1801 if (VA.isRegLoc()) {
1802 const TargetRegisterClass *RC;
1803 EVT ValVT = VA.getValVT();
1805 switch (ValVT.getSimpleVT().SimpleTy) {
1807 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1809 RC = &PPC::GPRCRegClass;
1812 RC = &PPC::F4RCRegClass;
1815 RC = &PPC::F8RCRegClass;
1821 RC = &PPC::VRRCRegClass;
1825 // Transform the arguments stored in physical registers into virtual ones.
1826 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1827 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1829 InVals.push_back(ArgValue);
1831 // Argument stored in memory.
1832 assert(VA.isMemLoc());
1834 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1835 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1838 // Create load nodes to retrieve arguments from the stack.
1839 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1840 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1841 MachinePointerInfo(),
1842 false, false, false, 0));
1846 // Assign locations to all of the incoming aggregate by value arguments.
1847 // Aggregates passed by value are stored in the local variable space of the
1848 // caller's stack frame, right above the parameter list area.
1849 SmallVector<CCValAssign, 16> ByValArgLocs;
1850 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1851 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1853 // Reserve stack space for the allocations in CCInfo.
1854 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1856 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1858 // Area that is at least reserved in the caller of this function.
1859 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1861 // Set the size that is at least reserved in caller of this function. Tail
1862 // call optimized function's reserved stack space needs to be aligned so that
1863 // taking the difference between two stack areas will result in an aligned
1865 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1868 std::max(MinReservedArea,
1869 PPCFrameLowering::getMinCallFrameSize(false, false));
1871 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1872 getStackAlignment();
1873 unsigned AlignMask = TargetAlign-1;
1874 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1876 FI->setMinReservedArea(MinReservedArea);
1878 SmallVector<SDValue, 8> MemOps;
1880 // If the function takes variable number of arguments, make a frame index for
1881 // the start of the first vararg value... for expansion of llvm.va_start.
1883 static const uint16_t GPArgRegs[] = {
1884 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1885 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1887 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1889 static const uint16_t FPArgRegs[] = {
1890 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1893 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1895 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1897 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1900 // Make room for NumGPArgRegs and NumFPArgRegs.
1901 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1902 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1904 FuncInfo->setVarArgsStackOffset(
1905 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1906 CCInfo.getNextStackOffset(), true));
1908 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1909 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1911 // The fixed integer arguments of a variadic function are stored to the
1912 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1913 // the result of va_next.
1914 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1915 // Get an existing live-in vreg, or add a new one.
1916 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1918 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1920 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1921 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1922 MachinePointerInfo(), false, false, 0);
1923 MemOps.push_back(Store);
1924 // Increment the address by four for the next argument to store
1925 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1926 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1929 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1931 // The double arguments are stored to the VarArgsFrameIndex
1933 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1934 // Get an existing live-in vreg, or add a new one.
1935 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1937 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1939 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1940 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1941 MachinePointerInfo(), false, false, 0);
1942 MemOps.push_back(Store);
1943 // Increment the address by eight for the next argument to store
1944 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1946 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1950 if (!MemOps.empty())
1951 Chain = DAG.getNode(ISD::TokenFactor, dl,
1952 MVT::Other, &MemOps[0], MemOps.size());
1957 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1958 // value to MVT::i64 and then truncate to the correct register size.
1960 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
1961 SelectionDAG &DAG, SDValue ArgVal,
1962 DebugLoc dl) const {
1964 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1965 DAG.getValueType(ObjectVT));
1966 else if (Flags.isZExt())
1967 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1968 DAG.getValueType(ObjectVT));
1970 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1973 // Set the size that is at least reserved in caller of this function. Tail
1974 // call optimized functions' reserved stack space needs to be aligned so that
1975 // taking the difference between two stack areas will result in an aligned
1978 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
1979 unsigned nAltivecParamsAtEnd,
1980 unsigned MinReservedArea,
1981 bool isPPC64) const {
1982 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1983 // Add the Altivec parameters at the end, if needed.
1984 if (nAltivecParamsAtEnd) {
1985 MinReservedArea = ((MinReservedArea+15)/16)*16;
1986 MinReservedArea += 16*nAltivecParamsAtEnd;
1989 std::max(MinReservedArea,
1990 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
1991 unsigned TargetAlign
1992 = DAG.getMachineFunction().getTarget().getFrameLowering()->
1993 getStackAlignment();
1994 unsigned AlignMask = TargetAlign-1;
1995 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1996 FI->setMinReservedArea(MinReservedArea);
2000 PPCTargetLowering::LowerFormalArguments_64SVR4(
2002 CallingConv::ID CallConv, bool isVarArg,
2003 const SmallVectorImpl<ISD::InputArg>
2005 DebugLoc dl, SelectionDAG &DAG,
2006 SmallVectorImpl<SDValue> &InVals) const {
2007 // TODO: add description of PPC stack frame format, or at least some docs.
2009 MachineFunction &MF = DAG.getMachineFunction();
2010 MachineFrameInfo *MFI = MF.getFrameInfo();
2011 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2013 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2014 // Potential tail calls could cause overwriting of argument stack slots.
2015 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2016 (CallConv == CallingConv::Fast));
2017 unsigned PtrByteSize = 8;
2019 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2020 // Area that is at least reserved in caller of this function.
2021 unsigned MinReservedArea = ArgOffset;
2023 static const uint16_t GPR[] = {
2024 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2025 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2028 static const uint16_t *FPR = GetFPR();
2030 static const uint16_t VR[] = {
2031 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2032 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2035 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2036 const unsigned Num_FPR_Regs = 13;
2037 const unsigned Num_VR_Regs = array_lengthof(VR);
2039 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2041 // Add DAG nodes to load the arguments or copy them out of registers. On
2042 // entry to a function on PPC, the arguments start after the linkage area,
2043 // although the first ones are often in registers.
2045 SmallVector<SDValue, 8> MemOps;
2046 unsigned nAltivecParamsAtEnd = 0;
2047 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2048 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2050 bool needsLoad = false;
2051 EVT ObjectVT = Ins[ArgNo].VT;
2052 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2053 unsigned ArgSize = ObjSize;
2054 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2056 unsigned CurArgOffset = ArgOffset;
2058 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2059 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2060 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2062 MinReservedArea = ((MinReservedArea+15)/16)*16;
2063 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2067 nAltivecParamsAtEnd++;
2069 // Calculate min reserved area.
2070 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2074 // FIXME the codegen can be much improved in some cases.
2075 // We do not have to keep everything in memory.
2076 if (Flags.isByVal()) {
2077 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2078 ObjSize = Flags.getByValSize();
2079 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2080 // All aggregates smaller than 8 bytes must be passed right-justified.
2081 if (ObjSize < PtrByteSize)
2082 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2083 // The value of the object is its address.
2084 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2085 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2086 InVals.push_back(FIN);
2089 if (GPR_idx != Num_GPR_Regs) {
2090 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2091 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2094 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2095 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2096 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2097 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2098 MachinePointerInfo(FuncArg, CurArgOffset),
2099 ObjType, false, false, 0);
2101 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2102 // store the whole register as-is to the parameter save area
2103 // slot. The address of the parameter was already calculated
2104 // above (InVals.push_back(FIN)) to be the right-justified
2105 // offset within the slot. For this store, we need a new
2106 // frame index that points at the beginning of the slot.
2107 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2108 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2109 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2110 MachinePointerInfo(FuncArg, ArgOffset),
2114 MemOps.push_back(Store);
2117 // Whether we copied from a register or not, advance the offset
2118 // into the parameter save area by a full doubleword.
2119 ArgOffset += PtrByteSize;
2123 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2124 // Store whatever pieces of the object are in registers
2125 // to memory. ArgOffset will be the address of the beginning
2127 if (GPR_idx != Num_GPR_Regs) {
2129 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2130 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2131 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2132 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2133 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2134 MachinePointerInfo(FuncArg, ArgOffset),
2136 MemOps.push_back(Store);
2138 ArgOffset += PtrByteSize;
2140 ArgOffset += ArgSize - j;
2147 switch (ObjectVT.getSimpleVT().SimpleTy) {
2148 default: llvm_unreachable("Unhandled argument type!");
2151 if (GPR_idx != Num_GPR_Regs) {
2152 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2153 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2155 if (ObjectVT == MVT::i32)
2156 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2157 // value to MVT::i64 and then truncate to the correct register size.
2158 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2163 ArgSize = PtrByteSize;
2170 // Every 8 bytes of argument space consumes one of the GPRs available for
2171 // argument passing.
2172 if (GPR_idx != Num_GPR_Regs) {
2175 if (FPR_idx != Num_FPR_Regs) {
2178 if (ObjectVT == MVT::f32)
2179 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2181 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2183 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2187 ArgSize = PtrByteSize;
2196 // Note that vector arguments in registers don't reserve stack space,
2197 // except in varargs functions.
2198 if (VR_idx != Num_VR_Regs) {
2199 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2200 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2202 while ((ArgOffset % 16) != 0) {
2203 ArgOffset += PtrByteSize;
2204 if (GPR_idx != Num_GPR_Regs)
2208 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2212 // Vectors are aligned.
2213 ArgOffset = ((ArgOffset+15)/16)*16;
2214 CurArgOffset = ArgOffset;
2221 // We need to load the argument to a virtual register if we determined
2222 // above that we ran out of physical registers of the appropriate type.
2224 int FI = MFI->CreateFixedObject(ObjSize,
2225 CurArgOffset + (ArgSize - ObjSize),
2227 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2228 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2229 false, false, false, 0);
2232 InVals.push_back(ArgVal);
2235 // Set the size that is at least reserved in caller of this function. Tail
2236 // call optimized functions' reserved stack space needs to be aligned so that
2237 // taking the difference between two stack areas will result in an aligned
2239 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2241 // If the function takes variable number of arguments, make a frame index for
2242 // the start of the first vararg value... for expansion of llvm.va_start.
2244 int Depth = ArgOffset;
2246 FuncInfo->setVarArgsFrameIndex(
2247 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2248 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2250 // If this function is vararg, store any remaining integer argument regs
2251 // to their spots on the stack so that they may be loaded by deferencing the
2252 // result of va_next.
2253 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2254 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2255 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2256 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2257 MachinePointerInfo(), false, false, 0);
2258 MemOps.push_back(Store);
2259 // Increment the address by four for the next argument to store
2260 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2261 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2265 if (!MemOps.empty())
2266 Chain = DAG.getNode(ISD::TokenFactor, dl,
2267 MVT::Other, &MemOps[0], MemOps.size());
2273 PPCTargetLowering::LowerFormalArguments_Darwin(
2275 CallingConv::ID CallConv, bool isVarArg,
2276 const SmallVectorImpl<ISD::InputArg>
2278 DebugLoc dl, SelectionDAG &DAG,
2279 SmallVectorImpl<SDValue> &InVals) const {
2280 // TODO: add description of PPC stack frame format, or at least some docs.
2282 MachineFunction &MF = DAG.getMachineFunction();
2283 MachineFrameInfo *MFI = MF.getFrameInfo();
2284 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2286 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2287 bool isPPC64 = PtrVT == MVT::i64;
2288 // Potential tail calls could cause overwriting of argument stack slots.
2289 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2290 (CallConv == CallingConv::Fast));
2291 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2293 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2294 // Area that is at least reserved in caller of this function.
2295 unsigned MinReservedArea = ArgOffset;
2297 static const uint16_t GPR_32[] = { // 32-bit registers.
2298 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2299 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2301 static const uint16_t GPR_64[] = { // 64-bit registers.
2302 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2303 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2306 static const uint16_t *FPR = GetFPR();
2308 static const uint16_t VR[] = {
2309 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2310 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2313 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2314 const unsigned Num_FPR_Regs = 13;
2315 const unsigned Num_VR_Regs = array_lengthof( VR);
2317 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2319 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2321 // In 32-bit non-varargs functions, the stack space for vectors is after the
2322 // stack space for non-vectors. We do not use this space unless we have
2323 // too many vectors to fit in registers, something that only occurs in
2324 // constructed examples:), but we have to walk the arglist to figure
2325 // that out...for the pathological case, compute VecArgOffset as the
2326 // start of the vector parameter area. Computing VecArgOffset is the
2327 // entire point of the following loop.
2328 unsigned VecArgOffset = ArgOffset;
2329 if (!isVarArg && !isPPC64) {
2330 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2332 EVT ObjectVT = Ins[ArgNo].VT;
2333 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2335 if (Flags.isByVal()) {
2336 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2337 unsigned ObjSize = Flags.getByValSize();
2339 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2340 VecArgOffset += ArgSize;
2344 switch(ObjectVT.getSimpleVT().SimpleTy) {
2345 default: llvm_unreachable("Unhandled argument type!");
2350 case MVT::i64: // PPC64
2352 // FIXME: We are guaranteed to be !isPPC64 at this point.
2353 // Does MVT::i64 apply?
2360 // Nothing to do, we're only looking at Nonvector args here.
2365 // We've found where the vector parameter area in memory is. Skip the
2366 // first 12 parameters; these don't use that memory.
2367 VecArgOffset = ((VecArgOffset+15)/16)*16;
2368 VecArgOffset += 12*16;
2370 // Add DAG nodes to load the arguments or copy them out of registers. On
2371 // entry to a function on PPC, the arguments start after the linkage area,
2372 // although the first ones are often in registers.
2374 SmallVector<SDValue, 8> MemOps;
2375 unsigned nAltivecParamsAtEnd = 0;
2376 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2377 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2379 bool needsLoad = false;
2380 EVT ObjectVT = Ins[ArgNo].VT;
2381 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2382 unsigned ArgSize = ObjSize;
2383 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2385 unsigned CurArgOffset = ArgOffset;
2387 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2388 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2389 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2390 if (isVarArg || isPPC64) {
2391 MinReservedArea = ((MinReservedArea+15)/16)*16;
2392 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2395 } else nAltivecParamsAtEnd++;
2397 // Calculate min reserved area.
2398 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2402 // FIXME the codegen can be much improved in some cases.
2403 // We do not have to keep everything in memory.
2404 if (Flags.isByVal()) {
2405 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2406 ObjSize = Flags.getByValSize();
2407 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2408 // Objects of size 1 and 2 are right justified, everything else is
2409 // left justified. This means the memory address is adjusted forwards.
2410 if (ObjSize==1 || ObjSize==2) {
2411 CurArgOffset = CurArgOffset + (4 - ObjSize);
2413 // The value of the object is its address.
2414 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2415 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2416 InVals.push_back(FIN);
2417 if (ObjSize==1 || ObjSize==2) {
2418 if (GPR_idx != Num_GPR_Regs) {
2421 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2423 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2424 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2425 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2426 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2427 MachinePointerInfo(FuncArg,
2429 ObjType, false, false, 0);
2430 MemOps.push_back(Store);
2434 ArgOffset += PtrByteSize;
2438 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2439 // Store whatever pieces of the object are in registers
2440 // to memory. ArgOffset will be the address of the beginning
2442 if (GPR_idx != Num_GPR_Regs) {
2445 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2447 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2448 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2449 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2450 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2451 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2452 MachinePointerInfo(FuncArg, ArgOffset),
2454 MemOps.push_back(Store);
2456 ArgOffset += PtrByteSize;
2458 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2465 switch (ObjectVT.getSimpleVT().SimpleTy) {
2466 default: llvm_unreachable("Unhandled argument type!");
2469 if (GPR_idx != Num_GPR_Regs) {
2470 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2471 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2475 ArgSize = PtrByteSize;
2477 // All int arguments reserve stack space in the Darwin ABI.
2478 ArgOffset += PtrByteSize;
2482 case MVT::i64: // PPC64
2483 if (GPR_idx != Num_GPR_Regs) {
2484 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2485 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2487 if (ObjectVT == MVT::i32)
2488 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2489 // value to MVT::i64 and then truncate to the correct register size.
2490 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2495 ArgSize = PtrByteSize;
2497 // All int arguments reserve stack space in the Darwin ABI.
2503 // Every 4 bytes of argument space consumes one of the GPRs available for
2504 // argument passing.
2505 if (GPR_idx != Num_GPR_Regs) {
2507 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2510 if (FPR_idx != Num_FPR_Regs) {
2513 if (ObjectVT == MVT::f32)
2514 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2516 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2518 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2524 // All FP arguments reserve stack space in the Darwin ABI.
2525 ArgOffset += isPPC64 ? 8 : ObjSize;
2531 // Note that vector arguments in registers don't reserve stack space,
2532 // except in varargs functions.
2533 if (VR_idx != Num_VR_Regs) {
2534 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2535 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2537 while ((ArgOffset % 16) != 0) {
2538 ArgOffset += PtrByteSize;
2539 if (GPR_idx != Num_GPR_Regs)
2543 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2547 if (!isVarArg && !isPPC64) {
2548 // Vectors go after all the nonvectors.
2549 CurArgOffset = VecArgOffset;
2552 // Vectors are aligned.
2553 ArgOffset = ((ArgOffset+15)/16)*16;
2554 CurArgOffset = ArgOffset;
2562 // We need to load the argument to a virtual register if we determined above
2563 // that we ran out of physical registers of the appropriate type.
2565 int FI = MFI->CreateFixedObject(ObjSize,
2566 CurArgOffset + (ArgSize - ObjSize),
2568 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2569 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2570 false, false, false, 0);
2573 InVals.push_back(ArgVal);
2576 // Set the size that is at least reserved in caller of this function. Tail
2577 // call optimized functions' reserved stack space needs to be aligned so that
2578 // taking the difference between two stack areas will result in an aligned
2580 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2582 // If the function takes variable number of arguments, make a frame index for
2583 // the start of the first vararg value... for expansion of llvm.va_start.
2585 int Depth = ArgOffset;
2587 FuncInfo->setVarArgsFrameIndex(
2588 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2590 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2592 // If this function is vararg, store any remaining integer argument regs
2593 // to their spots on the stack so that they may be loaded by deferencing the
2594 // result of va_next.
2595 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2599 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2601 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2603 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2604 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2605 MachinePointerInfo(), false, false, 0);
2606 MemOps.push_back(Store);
2607 // Increment the address by four for the next argument to store
2608 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2609 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2613 if (!MemOps.empty())
2614 Chain = DAG.getNode(ISD::TokenFactor, dl,
2615 MVT::Other, &MemOps[0], MemOps.size());
2620 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2621 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2623 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2627 const SmallVectorImpl<ISD::OutputArg>
2629 const SmallVectorImpl<SDValue> &OutVals,
2630 unsigned &nAltivecParamsAtEnd) {
2631 // Count how many bytes are to be pushed on the stack, including the linkage
2632 // area, and parameter passing area. We start with 24/48 bytes, which is
2633 // prereserved space for [SP][CR][LR][3 x unused].
2634 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2635 unsigned NumOps = Outs.size();
2636 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2638 // Add up all the space actually used.
2639 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2640 // they all go in registers, but we must reserve stack space for them for
2641 // possible use by the caller. In varargs or 64-bit calls, parameters are
2642 // assigned stack space in order, with padding so Altivec parameters are
2644 nAltivecParamsAtEnd = 0;
2645 for (unsigned i = 0; i != NumOps; ++i) {
2646 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2647 EVT ArgVT = Outs[i].VT;
2648 // Varargs Altivec parameters are padded to a 16 byte boundary.
2649 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2650 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2651 if (!isVarArg && !isPPC64) {
2652 // Non-varargs Altivec parameters go after all the non-Altivec
2653 // parameters; handle those later so we know how much padding we need.
2654 nAltivecParamsAtEnd++;
2657 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2658 NumBytes = ((NumBytes+15)/16)*16;
2660 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2663 // Allow for Altivec parameters at the end, if needed.
2664 if (nAltivecParamsAtEnd) {
2665 NumBytes = ((NumBytes+15)/16)*16;
2666 NumBytes += 16*nAltivecParamsAtEnd;
2669 // The prolog code of the callee may store up to 8 GPR argument registers to
2670 // the stack, allowing va_start to index over them in memory if its varargs.
2671 // Because we cannot tell if this is needed on the caller side, we have to
2672 // conservatively assume that it is needed. As such, make sure we have at
2673 // least enough stack space for the caller to store the 8 GPRs.
2674 NumBytes = std::max(NumBytes,
2675 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2677 // Tail call needs the stack to be aligned.
2678 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2679 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2680 getFrameLowering()->getStackAlignment();
2681 unsigned AlignMask = TargetAlign-1;
2682 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2688 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2689 /// adjusted to accommodate the arguments for the tailcall.
2690 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2691 unsigned ParamSize) {
2693 if (!isTailCall) return 0;
2695 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2696 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2697 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2698 // Remember only if the new adjustement is bigger.
2699 if (SPDiff < FI->getTailCallSPDelta())
2700 FI->setTailCallSPDelta(SPDiff);
2705 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2706 /// for tail call optimization. Targets which want to do tail call
2707 /// optimization should implement this function.
2709 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2710 CallingConv::ID CalleeCC,
2712 const SmallVectorImpl<ISD::InputArg> &Ins,
2713 SelectionDAG& DAG) const {
2714 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2717 // Variable argument functions are not supported.
2721 MachineFunction &MF = DAG.getMachineFunction();
2722 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2723 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2724 // Functions containing by val parameters are not supported.
2725 for (unsigned i = 0; i != Ins.size(); i++) {
2726 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2727 if (Flags.isByVal()) return false;
2730 // Non PIC/GOT tail calls are supported.
2731 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2734 // At the moment we can only do local tail calls (in same module, hidden
2735 // or protected) if we are generating PIC.
2736 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2737 return G->getGlobal()->hasHiddenVisibility()
2738 || G->getGlobal()->hasProtectedVisibility();
2744 /// isCallCompatibleAddress - Return the immediate to use if the specified
2745 /// 32-bit value is representable in the immediate field of a BxA instruction.
2746 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2747 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2750 int Addr = C->getZExtValue();
2751 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2752 SignExtend32<26>(Addr) != Addr)
2753 return 0; // Top 6 bits have to be sext of immediate.
2755 return DAG.getConstant((int)C->getZExtValue() >> 2,
2756 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2761 struct TailCallArgumentInfo {
2766 TailCallArgumentInfo() : FrameIdx(0) {}
2771 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2773 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2775 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2776 SmallVector<SDValue, 8> &MemOpChains,
2778 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2779 SDValue Arg = TailCallArgs[i].Arg;
2780 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2781 int FI = TailCallArgs[i].FrameIdx;
2782 // Store relative to framepointer.
2783 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2784 MachinePointerInfo::getFixedStack(FI),
2789 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2790 /// the appropriate stack slot for the tail call optimized function call.
2791 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2792 MachineFunction &MF,
2801 // Calculate the new stack slot for the return address.
2802 int SlotSize = isPPC64 ? 8 : 4;
2803 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2805 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2806 NewRetAddrLoc, true);
2807 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2808 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2809 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2810 MachinePointerInfo::getFixedStack(NewRetAddr),
2813 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2814 // slot as the FP is never overwritten.
2817 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2818 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2820 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2821 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2822 MachinePointerInfo::getFixedStack(NewFPIdx),
2829 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2830 /// the position of the argument.
2832 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2833 SDValue Arg, int SPDiff, unsigned ArgOffset,
2834 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2835 int Offset = ArgOffset + SPDiff;
2836 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2837 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2838 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2839 SDValue FIN = DAG.getFrameIndex(FI, VT);
2840 TailCallArgumentInfo Info;
2842 Info.FrameIdxOp = FIN;
2844 TailCallArguments.push_back(Info);
2847 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2848 /// stack slot. Returns the chain as result and the loaded frame pointers in
2849 /// LROpOut/FPOpout. Used when tail calling.
2850 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2856 DebugLoc dl) const {
2858 // Load the LR and FP stack slot for later adjusting.
2859 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2860 LROpOut = getReturnAddrFrameIndex(DAG);
2861 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2862 false, false, false, 0);
2863 Chain = SDValue(LROpOut.getNode(), 1);
2865 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2866 // slot as the FP is never overwritten.
2868 FPOpOut = getFramePointerFrameIndex(DAG);
2869 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2870 false, false, false, 0);
2871 Chain = SDValue(FPOpOut.getNode(), 1);
2877 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2878 /// by "Src" to address "Dst" of size "Size". Alignment information is
2879 /// specified by the specific parameter attribute. The copy will be passed as
2880 /// a byval function parameter.
2881 /// Sometimes what we are copying is the end of a larger object, the part that
2882 /// does not fit in registers.
2884 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2885 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2887 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2888 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2889 false, false, MachinePointerInfo(0),
2890 MachinePointerInfo(0));
2893 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2896 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2897 SDValue Arg, SDValue PtrOff, int SPDiff,
2898 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2899 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2900 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2902 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2907 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2909 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2910 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2911 DAG.getConstant(ArgOffset, PtrVT));
2913 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2914 MachinePointerInfo(), false, false, 0));
2915 // Calculate and remember argument location.
2916 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2921 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2922 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2923 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2924 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2925 MachineFunction &MF = DAG.getMachineFunction();
2927 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2928 // might overwrite each other in case of tail call optimization.
2929 SmallVector<SDValue, 8> MemOpChains2;
2930 // Do not flag preceding copytoreg stuff together with the following stuff.
2932 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2934 if (!MemOpChains2.empty())
2935 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2936 &MemOpChains2[0], MemOpChains2.size());
2938 // Store the return address to the appropriate stack slot.
2939 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2940 isPPC64, isDarwinABI, dl);
2942 // Emit callseq_end just before tailcall node.
2943 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2944 DAG.getIntPtrConstant(0, true), InFlag);
2945 InFlag = Chain.getValue(1);
2949 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2950 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2951 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2952 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2953 const PPCSubtarget &PPCSubTarget) {
2955 bool isPPC64 = PPCSubTarget.isPPC64();
2956 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2958 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2959 NodeTys.push_back(MVT::Other); // Returns a chain
2960 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2962 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2964 bool needIndirectCall = true;
2965 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2966 // If this is an absolute destination address, use the munged value.
2967 Callee = SDValue(Dest, 0);
2968 needIndirectCall = false;
2971 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2972 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2973 // Use indirect calls for ALL functions calls in JIT mode, since the
2974 // far-call stubs may be outside relocation limits for a BL instruction.
2975 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2976 unsigned OpFlags = 0;
2977 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2978 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2979 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2980 (G->getGlobal()->isDeclaration() ||
2981 G->getGlobal()->isWeakForLinker())) {
2982 // PC-relative references to external symbols should go through $stub,
2983 // unless we're building with the leopard linker or later, which
2984 // automatically synthesizes these stubs.
2985 OpFlags = PPCII::MO_DARWIN_STUB;
2988 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2989 // every direct call is) turn it into a TargetGlobalAddress /
2990 // TargetExternalSymbol node so that legalize doesn't hack it.
2991 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2992 Callee.getValueType(),
2994 needIndirectCall = false;
2998 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2999 unsigned char OpFlags = 0;
3001 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3002 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3003 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3004 // PC-relative references to external symbols should go through $stub,
3005 // unless we're building with the leopard linker or later, which
3006 // automatically synthesizes these stubs.
3007 OpFlags = PPCII::MO_DARWIN_STUB;
3010 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3012 needIndirectCall = false;
3015 if (needIndirectCall) {
3016 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3017 // to do the call, we can't use PPCISD::CALL.
3018 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3020 if (isSVR4ABI && isPPC64) {
3021 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3022 // entry point, but to the function descriptor (the function entry point
3023 // address is part of the function descriptor though).
3024 // The function descriptor is a three doubleword structure with the
3025 // following fields: function entry point, TOC base address and
3026 // environment pointer.
3027 // Thus for a call through a function pointer, the following actions need
3029 // 1. Save the TOC of the caller in the TOC save area of its stack
3030 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3031 // 2. Load the address of the function entry point from the function
3033 // 3. Load the TOC of the callee from the function descriptor into r2.
3034 // 4. Load the environment pointer from the function descriptor into
3036 // 5. Branch to the function entry point address.
3037 // 6. On return of the callee, the TOC of the caller needs to be
3038 // restored (this is done in FinishCall()).
3040 // All those operations are flagged together to ensure that no other
3041 // operations can be scheduled in between. E.g. without flagging the
3042 // operations together, a TOC access in the caller could be scheduled
3043 // between the load of the callee TOC and the branch to the callee, which
3044 // results in the TOC access going through the TOC of the callee instead
3045 // of going through the TOC of the caller, which leads to incorrect code.
3047 // Load the address of the function entry point from the function
3049 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3050 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3051 InFlag.getNode() ? 3 : 2);
3052 Chain = LoadFuncPtr.getValue(1);
3053 InFlag = LoadFuncPtr.getValue(2);
3055 // Load environment pointer into r11.
3056 // Offset of the environment pointer within the function descriptor.
3057 SDValue PtrOff = DAG.getIntPtrConstant(16);
3059 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3060 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3062 Chain = LoadEnvPtr.getValue(1);
3063 InFlag = LoadEnvPtr.getValue(2);
3065 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3067 Chain = EnvVal.getValue(0);
3068 InFlag = EnvVal.getValue(1);
3070 // Load TOC of the callee into r2. We are using a target-specific load
3071 // with r2 hard coded, because the result of a target-independent load
3072 // would never go directly into r2, since r2 is a reserved register (which
3073 // prevents the register allocator from allocating it), resulting in an
3074 // additional register being allocated and an unnecessary move instruction
3076 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3077 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3079 Chain = LoadTOCPtr.getValue(0);
3080 InFlag = LoadTOCPtr.getValue(1);
3082 MTCTROps[0] = Chain;
3083 MTCTROps[1] = LoadFuncPtr;
3084 MTCTROps[2] = InFlag;
3087 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3088 2 + (InFlag.getNode() != 0));
3089 InFlag = Chain.getValue(1);
3092 NodeTys.push_back(MVT::Other);
3093 NodeTys.push_back(MVT::Glue);
3094 Ops.push_back(Chain);
3095 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3097 // Add CTR register as callee so a bctr can be emitted later.
3099 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3102 // If this is a direct call, pass the chain and the callee.
3103 if (Callee.getNode()) {
3104 Ops.push_back(Chain);
3105 Ops.push_back(Callee);
3107 // If this is a tail call add stack pointer delta.
3109 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3111 // Add argument registers to the end of the list so that they are known live
3113 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3114 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3115 RegsToPass[i].second.getValueType()));
3121 bool isLocalCall(const SDValue &Callee)
3123 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3124 return !G->getGlobal()->isDeclaration() &&
3125 !G->getGlobal()->isWeakForLinker();
3130 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3131 CallingConv::ID CallConv, bool isVarArg,
3132 const SmallVectorImpl<ISD::InputArg> &Ins,
3133 DebugLoc dl, SelectionDAG &DAG,
3134 SmallVectorImpl<SDValue> &InVals) const {
3136 SmallVector<CCValAssign, 16> RVLocs;
3137 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3138 getTargetMachine(), RVLocs, *DAG.getContext());
3139 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3141 // Copy all of the result registers out of their specified physreg.
3142 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3143 CCValAssign &VA = RVLocs[i];
3144 EVT VT = VA.getValVT();
3145 assert(VA.isRegLoc() && "Can only return in registers!");
3146 Chain = DAG.getCopyFromReg(Chain, dl,
3147 VA.getLocReg(), VT, InFlag).getValue(1);
3148 InVals.push_back(Chain.getValue(0));
3149 InFlag = Chain.getValue(2);
3156 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3157 bool isTailCall, bool isVarArg,
3159 SmallVector<std::pair<unsigned, SDValue>, 8>
3161 SDValue InFlag, SDValue Chain,
3163 int SPDiff, unsigned NumBytes,
3164 const SmallVectorImpl<ISD::InputArg> &Ins,
3165 SmallVectorImpl<SDValue> &InVals) const {
3166 std::vector<EVT> NodeTys;
3167 SmallVector<SDValue, 8> Ops;
3168 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3169 isTailCall, RegsToPass, Ops, NodeTys,
3172 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3173 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3174 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3176 // When performing tail call optimization the callee pops its arguments off
3177 // the stack. Account for this here so these bytes can be pushed back on in
3178 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3179 int BytesCalleePops =
3180 (CallConv == CallingConv::Fast &&
3181 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3183 // Add a register mask operand representing the call-preserved registers.
3184 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3185 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3186 assert(Mask && "Missing call preserved mask for calling convention");
3187 Ops.push_back(DAG.getRegisterMask(Mask));
3189 if (InFlag.getNode())
3190 Ops.push_back(InFlag);
3194 // If this is the first return lowered for this function, add the regs
3195 // to the liveout set for the function.
3196 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3197 SmallVector<CCValAssign, 16> RVLocs;
3198 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3199 getTargetMachine(), RVLocs, *DAG.getContext());
3200 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3201 for (unsigned i = 0; i != RVLocs.size(); ++i)
3202 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3205 assert(((Callee.getOpcode() == ISD::Register &&
3206 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3207 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3208 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3209 isa<ConstantSDNode>(Callee)) &&
3210 "Expecting an global address, external symbol, absolute value or register");
3212 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3215 // Add a NOP immediately after the branch instruction when using the 64-bit
3216 // SVR4 ABI. At link time, if caller and callee are in a different module and
3217 // thus have a different TOC, the call will be replaced with a call to a stub
3218 // function which saves the current TOC, loads the TOC of the callee and
3219 // branches to the callee. The NOP will be replaced with a load instruction
3220 // which restores the TOC of the caller from the TOC save slot of the current
3221 // stack frame. If caller and callee belong to the same module (and have the
3222 // same TOC), the NOP will remain unchanged.
3224 bool needsTOCRestore = false;
3225 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3226 if (CallOpc == PPCISD::BCTRL_SVR4) {
3227 // This is a call through a function pointer.
3228 // Restore the caller TOC from the save area into R2.
3229 // See PrepareCall() for more information about calls through function
3230 // pointers in the 64-bit SVR4 ABI.
3231 // We are using a target-specific load with r2 hard coded, because the
3232 // result of a target-independent load would never go directly into r2,
3233 // since r2 is a reserved register (which prevents the register allocator
3234 // from allocating it), resulting in an additional register being
3235 // allocated and an unnecessary move instruction being generated.
3236 needsTOCRestore = true;
3237 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3238 // Otherwise insert NOP for non-local calls.
3239 CallOpc = PPCISD::CALL_NOP_SVR4;
3243 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3244 InFlag = Chain.getValue(1);
3246 if (needsTOCRestore) {
3247 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3248 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3249 InFlag = Chain.getValue(1);
3252 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3253 DAG.getIntPtrConstant(BytesCalleePops, true),
3256 InFlag = Chain.getValue(1);
3258 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3259 Ins, dl, DAG, InVals);
3263 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3264 SmallVectorImpl<SDValue> &InVals) const {
3265 SelectionDAG &DAG = CLI.DAG;
3266 DebugLoc &dl = CLI.DL;
3267 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3268 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3269 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3270 SDValue Chain = CLI.Chain;
3271 SDValue Callee = CLI.Callee;
3272 bool &isTailCall = CLI.IsTailCall;
3273 CallingConv::ID CallConv = CLI.CallConv;
3274 bool isVarArg = CLI.IsVarArg;
3277 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3280 if (PPCSubTarget.isSVR4ABI()) {
3281 if (PPCSubTarget.isPPC64())
3282 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3283 isTailCall, Outs, OutVals, Ins,
3286 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3287 isTailCall, Outs, OutVals, Ins,
3291 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3292 isTailCall, Outs, OutVals, Ins,
3297 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3298 CallingConv::ID CallConv, bool isVarArg,
3300 const SmallVectorImpl<ISD::OutputArg> &Outs,
3301 const SmallVectorImpl<SDValue> &OutVals,
3302 const SmallVectorImpl<ISD::InputArg> &Ins,
3303 DebugLoc dl, SelectionDAG &DAG,
3304 SmallVectorImpl<SDValue> &InVals) const {
3305 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3306 // of the 32-bit SVR4 ABI stack frame layout.
3308 assert((CallConv == CallingConv::C ||
3309 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3311 unsigned PtrByteSize = 4;
3313 MachineFunction &MF = DAG.getMachineFunction();
3315 // Mark this function as potentially containing a function that contains a
3316 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3317 // and restoring the callers stack pointer in this functions epilog. This is
3318 // done because by tail calling the called function might overwrite the value
3319 // in this function's (MF) stack pointer stack slot 0(SP).
3320 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3321 CallConv == CallingConv::Fast)
3322 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3324 // Count how many bytes are to be pushed on the stack, including the linkage
3325 // area, parameter list area and the part of the local variable space which
3326 // contains copies of aggregates which are passed by value.
3328 // Assign locations to all of the outgoing arguments.
3329 SmallVector<CCValAssign, 16> ArgLocs;
3330 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3331 getTargetMachine(), ArgLocs, *DAG.getContext());
3333 // Reserve space for the linkage area on the stack.
3334 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3337 // Handle fixed and variable vector arguments differently.
3338 // Fixed vector arguments go into registers as long as registers are
3339 // available. Variable vector arguments always go into memory.
3340 unsigned NumArgs = Outs.size();
3342 for (unsigned i = 0; i != NumArgs; ++i) {
3343 MVT ArgVT = Outs[i].VT;
3344 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3347 if (Outs[i].IsFixed) {
3348 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3351 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3357 errs() << "Call operand #" << i << " has unhandled type "
3358 << EVT(ArgVT).getEVTString() << "\n";
3360 llvm_unreachable(0);
3364 // All arguments are treated the same.
3365 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3368 // Assign locations to all of the outgoing aggregate by value arguments.
3369 SmallVector<CCValAssign, 16> ByValArgLocs;
3370 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3371 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3373 // Reserve stack space for the allocations in CCInfo.
3374 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3376 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3378 // Size of the linkage area, parameter list area and the part of the local
3379 // space variable where copies of aggregates which are passed by value are
3381 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3383 // Calculate by how many bytes the stack has to be adjusted in case of tail
3384 // call optimization.
3385 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3387 // Adjust the stack pointer for the new arguments...
3388 // These operations are automatically eliminated by the prolog/epilog pass
3389 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3390 SDValue CallSeqStart = Chain;
3392 // Load the return address and frame pointer so it can be moved somewhere else
3395 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3398 // Set up a copy of the stack pointer for use loading and storing any
3399 // arguments that may not fit in the registers available for argument
3401 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3403 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3404 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3405 SmallVector<SDValue, 8> MemOpChains;
3407 bool seenFloatArg = false;
3408 // Walk the register/memloc assignments, inserting copies/loads.
3409 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3412 CCValAssign &VA = ArgLocs[i];
3413 SDValue Arg = OutVals[i];
3414 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3416 if (Flags.isByVal()) {
3417 // Argument is an aggregate which is passed by value, thus we need to
3418 // create a copy of it in the local variable space of the current stack
3419 // frame (which is the stack frame of the caller) and pass the address of
3420 // this copy to the callee.
3421 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3422 CCValAssign &ByValVA = ByValArgLocs[j++];
3423 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3425 // Memory reserved in the local variable space of the callers stack frame.
3426 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3428 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3429 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3431 // Create a copy of the argument in the local area of the current
3433 SDValue MemcpyCall =
3434 CreateCopyOfByValArgument(Arg, PtrOff,
3435 CallSeqStart.getNode()->getOperand(0),
3438 // This must go outside the CALLSEQ_START..END.
3439 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3440 CallSeqStart.getNode()->getOperand(1));
3441 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3442 NewCallSeqStart.getNode());
3443 Chain = CallSeqStart = NewCallSeqStart;
3445 // Pass the address of the aggregate copy on the stack either in a
3446 // physical register or in the parameter list area of the current stack
3447 // frame to the callee.
3451 if (VA.isRegLoc()) {
3452 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3453 // Put argument in a physical register.
3454 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3456 // Put argument in the parameter list area of the current stack frame.
3457 assert(VA.isMemLoc());
3458 unsigned LocMemOffset = VA.getLocMemOffset();
3461 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3462 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3464 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3465 MachinePointerInfo(),
3468 // Calculate and remember argument location.
3469 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3475 if (!MemOpChains.empty())
3476 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3477 &MemOpChains[0], MemOpChains.size());
3479 // Build a sequence of copy-to-reg nodes chained together with token chain
3480 // and flag operands which copy the outgoing args into the appropriate regs.
3482 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3483 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3484 RegsToPass[i].second, InFlag);
3485 InFlag = Chain.getValue(1);
3488 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3491 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3492 SDValue Ops[] = { Chain, InFlag };
3494 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3495 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3497 InFlag = Chain.getValue(1);
3501 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3502 false, TailCallArguments);
3504 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3505 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3509 // Copy an argument into memory, being careful to do this outside the
3510 // call sequence for the call to which the argument belongs.
3512 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3513 SDValue CallSeqStart,
3514 ISD::ArgFlagsTy Flags,
3516 DebugLoc dl) const {
3517 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3518 CallSeqStart.getNode()->getOperand(0),
3520 // The MEMCPY must go outside the CALLSEQ_START..END.
3521 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3522 CallSeqStart.getNode()->getOperand(1));
3523 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3524 NewCallSeqStart.getNode());
3525 return NewCallSeqStart;
3529 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3530 CallingConv::ID CallConv, bool isVarArg,
3532 const SmallVectorImpl<ISD::OutputArg> &Outs,
3533 const SmallVectorImpl<SDValue> &OutVals,
3534 const SmallVectorImpl<ISD::InputArg> &Ins,
3535 DebugLoc dl, SelectionDAG &DAG,
3536 SmallVectorImpl<SDValue> &InVals) const {
3538 unsigned NumOps = Outs.size();
3540 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3541 unsigned PtrByteSize = 8;
3543 MachineFunction &MF = DAG.getMachineFunction();
3545 // Mark this function as potentially containing a function that contains a
3546 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3547 // and restoring the callers stack pointer in this functions epilog. This is
3548 // done because by tail calling the called function might overwrite the value
3549 // in this function's (MF) stack pointer stack slot 0(SP).
3550 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3551 CallConv == CallingConv::Fast)
3552 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3554 unsigned nAltivecParamsAtEnd = 0;
3556 // Count how many bytes are to be pushed on the stack, including the linkage
3557 // area, and parameter passing area. We start with at least 48 bytes, which
3558 // is reserved space for [SP][CR][LR][3 x unused].
3559 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3562 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3563 Outs, OutVals, nAltivecParamsAtEnd);
3565 // Calculate by how many bytes the stack has to be adjusted in case of tail
3566 // call optimization.
3567 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3569 // To protect arguments on the stack from being clobbered in a tail call,
3570 // force all the loads to happen before doing any other lowering.
3572 Chain = DAG.getStackArgumentTokenFactor(Chain);
3574 // Adjust the stack pointer for the new arguments...
3575 // These operations are automatically eliminated by the prolog/epilog pass
3576 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3577 SDValue CallSeqStart = Chain;
3579 // Load the return address and frame pointer so it can be move somewhere else
3582 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3585 // Set up a copy of the stack pointer for use loading and storing any
3586 // arguments that may not fit in the registers available for argument
3588 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3590 // Figure out which arguments are going to go in registers, and which in
3591 // memory. Also, if this is a vararg function, floating point operations
3592 // must be stored to our stack, and loaded into integer regs as well, if
3593 // any integer regs are available for argument passing.
3594 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3595 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3597 static const uint16_t GPR[] = {
3598 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3599 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3601 static const uint16_t *FPR = GetFPR();
3603 static const uint16_t VR[] = {
3604 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3605 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3607 const unsigned NumGPRs = array_lengthof(GPR);
3608 const unsigned NumFPRs = 13;
3609 const unsigned NumVRs = array_lengthof(VR);
3611 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3612 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3614 SmallVector<SDValue, 8> MemOpChains;
3615 for (unsigned i = 0; i != NumOps; ++i) {
3616 SDValue Arg = OutVals[i];
3617 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3619 // PtrOff will be used to store the current argument to the stack if a
3620 // register cannot be found for it.
3623 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3625 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3627 // Promote integers to 64-bit values.
3628 if (Arg.getValueType() == MVT::i32) {
3629 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3630 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3631 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3634 // FIXME memcpy is used way more than necessary. Correctness first.
3635 // Note: "by value" is code for passing a structure by value, not
3637 if (Flags.isByVal()) {
3638 // Note: Size includes alignment padding, so
3639 // struct x { short a; char b; }
3640 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3641 // These are the proper values we need for right-justifying the
3642 // aggregate in a parameter register.
3643 unsigned Size = Flags.getByValSize();
3644 // All aggregates smaller than 8 bytes must be passed right-justified.
3645 if (Size==1 || Size==2 || Size==4) {
3646 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3647 if (GPR_idx != NumGPRs) {
3648 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3649 MachinePointerInfo(), VT,
3651 MemOpChains.push_back(Load.getValue(1));
3652 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3654 ArgOffset += PtrByteSize;
3659 if (GPR_idx == NumGPRs && Size < 8) {
3660 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3661 PtrOff.getValueType());
3662 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3663 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3666 ArgOffset += PtrByteSize;
3669 // Copy entire object into memory. There are cases where gcc-generated
3670 // code assumes it is there, even if it could be put entirely into
3671 // registers. (This is not what the doc says.)
3673 // FIXME: The above statement is likely due to a misunderstanding of the
3674 // documents. All arguments must be copied into the parameter area BY
3675 // THE CALLEE in the event that the callee takes the address of any
3676 // formal argument. That has not yet been implemented. However, it is
3677 // reasonable to use the stack area as a staging area for the register
3680 // Skip this for small aggregates, as we will use the same slot for a
3681 // right-justified copy, below.
3683 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3687 // When a register is available, pass a small aggregate right-justified.
3688 if (Size < 8 && GPR_idx != NumGPRs) {
3689 // The easiest way to get this right-justified in a register
3690 // is to copy the structure into the rightmost portion of a
3691 // local variable slot, then load the whole slot into the
3693 // FIXME: The memcpy seems to produce pretty awful code for
3694 // small aggregates, particularly for packed ones.
3695 // FIXME: It would be preferable to use the slot in the
3696 // parameter save area instead of a new local variable.
3697 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3698 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3699 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3703 // Load the slot into the register.
3704 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3705 MachinePointerInfo(),
3706 false, false, false, 0);
3707 MemOpChains.push_back(Load.getValue(1));
3708 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3710 // Done with this argument.
3711 ArgOffset += PtrByteSize;
3715 // For aggregates larger than PtrByteSize, copy the pieces of the
3716 // object that fit into registers from the parameter save area.
3717 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3718 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3719 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3720 if (GPR_idx != NumGPRs) {
3721 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3722 MachinePointerInfo(),
3723 false, false, false, 0);
3724 MemOpChains.push_back(Load.getValue(1));
3725 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3726 ArgOffset += PtrByteSize;
3728 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3735 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3736 default: llvm_unreachable("Unexpected ValueType for argument!");
3739 if (GPR_idx != NumGPRs) {
3740 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3742 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3743 true, isTailCall, false, MemOpChains,
3744 TailCallArguments, dl);
3746 ArgOffset += PtrByteSize;
3750 if (FPR_idx != NumFPRs) {
3751 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3754 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3755 MachinePointerInfo(), false, false, 0);
3756 MemOpChains.push_back(Store);
3758 // Float varargs are always shadowed in available integer registers
3759 if (GPR_idx != NumGPRs) {
3760 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3761 MachinePointerInfo(), false, false,
3763 MemOpChains.push_back(Load.getValue(1));
3764 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3766 } else if (GPR_idx != NumGPRs)
3767 // If we have any FPRs remaining, we may also have GPRs remaining.
3770 // Single-precision floating-point values are mapped to the
3771 // second (rightmost) word of the stack doubleword.
3772 if (Arg.getValueType() == MVT::f32) {
3773 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3774 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3777 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3778 true, isTailCall, false, MemOpChains,
3779 TailCallArguments, dl);
3788 // These go aligned on the stack, or in the corresponding R registers
3789 // when within range. The Darwin PPC ABI doc claims they also go in
3790 // V registers; in fact gcc does this only for arguments that are
3791 // prototyped, not for those that match the ... We do it for all
3792 // arguments, seems to work.
3793 while (ArgOffset % 16 !=0) {
3794 ArgOffset += PtrByteSize;
3795 if (GPR_idx != NumGPRs)
3798 // We could elide this store in the case where the object fits
3799 // entirely in R registers. Maybe later.
3800 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3801 DAG.getConstant(ArgOffset, PtrVT));
3802 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3803 MachinePointerInfo(), false, false, 0);
3804 MemOpChains.push_back(Store);
3805 if (VR_idx != NumVRs) {
3806 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3807 MachinePointerInfo(),
3808 false, false, false, 0);
3809 MemOpChains.push_back(Load.getValue(1));
3810 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3813 for (unsigned i=0; i<16; i+=PtrByteSize) {
3814 if (GPR_idx == NumGPRs)
3816 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3817 DAG.getConstant(i, PtrVT));
3818 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3819 false, false, false, 0);
3820 MemOpChains.push_back(Load.getValue(1));
3821 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3826 // Non-varargs Altivec params generally go in registers, but have
3827 // stack space allocated at the end.
3828 if (VR_idx != NumVRs) {
3829 // Doesn't have GPR space allocated.
3830 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3832 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3833 true, isTailCall, true, MemOpChains,
3834 TailCallArguments, dl);
3841 if (!MemOpChains.empty())
3842 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3843 &MemOpChains[0], MemOpChains.size());
3845 // Check if this is an indirect call (MTCTR/BCTRL).
3846 // See PrepareCall() for more information about calls through function
3847 // pointers in the 64-bit SVR4 ABI.
3849 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3850 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3851 !isBLACompatibleAddress(Callee, DAG)) {
3852 // Load r2 into a virtual register and store it to the TOC save area.
3853 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3854 // TOC save area offset.
3855 SDValue PtrOff = DAG.getIntPtrConstant(40);
3856 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3857 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3859 // R12 must contain the address of an indirect callee. This does not
3860 // mean the MTCTR instruction must use R12; it's easier to model this
3861 // as an extra parameter, so do that.
3862 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
3865 // Build a sequence of copy-to-reg nodes chained together with token chain
3866 // and flag operands which copy the outgoing args into the appropriate regs.
3868 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3869 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3870 RegsToPass[i].second, InFlag);
3871 InFlag = Chain.getValue(1);
3875 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
3876 FPOp, true, TailCallArguments);
3878 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3879 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3884 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3885 CallingConv::ID CallConv, bool isVarArg,
3887 const SmallVectorImpl<ISD::OutputArg> &Outs,
3888 const SmallVectorImpl<SDValue> &OutVals,
3889 const SmallVectorImpl<ISD::InputArg> &Ins,
3890 DebugLoc dl, SelectionDAG &DAG,
3891 SmallVectorImpl<SDValue> &InVals) const {
3893 unsigned NumOps = Outs.size();
3895 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3896 bool isPPC64 = PtrVT == MVT::i64;
3897 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3899 MachineFunction &MF = DAG.getMachineFunction();
3901 // Mark this function as potentially containing a function that contains a
3902 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3903 // and restoring the callers stack pointer in this functions epilog. This is
3904 // done because by tail calling the called function might overwrite the value
3905 // in this function's (MF) stack pointer stack slot 0(SP).
3906 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3907 CallConv == CallingConv::Fast)
3908 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3910 unsigned nAltivecParamsAtEnd = 0;
3912 // Count how many bytes are to be pushed on the stack, including the linkage
3913 // area, and parameter passing area. We start with 24/48 bytes, which is
3914 // prereserved space for [SP][CR][LR][3 x unused].
3916 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3918 nAltivecParamsAtEnd);
3920 // Calculate by how many bytes the stack has to be adjusted in case of tail
3921 // call optimization.
3922 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3924 // To protect arguments on the stack from being clobbered in a tail call,
3925 // force all the loads to happen before doing any other lowering.
3927 Chain = DAG.getStackArgumentTokenFactor(Chain);
3929 // Adjust the stack pointer for the new arguments...
3930 // These operations are automatically eliminated by the prolog/epilog pass
3931 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3932 SDValue CallSeqStart = Chain;
3934 // Load the return address and frame pointer so it can be move somewhere else
3937 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3940 // Set up a copy of the stack pointer for use loading and storing any
3941 // arguments that may not fit in the registers available for argument
3945 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3947 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3949 // Figure out which arguments are going to go in registers, and which in
3950 // memory. Also, if this is a vararg function, floating point operations
3951 // must be stored to our stack, and loaded into integer regs as well, if
3952 // any integer regs are available for argument passing.
3953 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3954 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3956 static const uint16_t GPR_32[] = { // 32-bit registers.
3957 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3958 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3960 static const uint16_t GPR_64[] = { // 64-bit registers.
3961 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3962 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3964 static const uint16_t *FPR = GetFPR();
3966 static const uint16_t VR[] = {
3967 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3968 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3970 const unsigned NumGPRs = array_lengthof(GPR_32);
3971 const unsigned NumFPRs = 13;
3972 const unsigned NumVRs = array_lengthof(VR);
3974 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3976 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3977 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3979 SmallVector<SDValue, 8> MemOpChains;
3980 for (unsigned i = 0; i != NumOps; ++i) {
3981 SDValue Arg = OutVals[i];
3982 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3984 // PtrOff will be used to store the current argument to the stack if a
3985 // register cannot be found for it.
3988 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3990 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3992 // On PPC64, promote integers to 64-bit values.
3993 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3994 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3995 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3996 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3999 // FIXME memcpy is used way more than necessary. Correctness first.
4000 // Note: "by value" is code for passing a structure by value, not
4002 if (Flags.isByVal()) {
4003 unsigned Size = Flags.getByValSize();
4004 // Very small objects are passed right-justified. Everything else is
4005 // passed left-justified.
4006 if (Size==1 || Size==2) {
4007 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4008 if (GPR_idx != NumGPRs) {
4009 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4010 MachinePointerInfo(), VT,
4012 MemOpChains.push_back(Load.getValue(1));
4013 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4015 ArgOffset += PtrByteSize;
4017 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4018 PtrOff.getValueType());
4019 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4020 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4023 ArgOffset += PtrByteSize;
4027 // Copy entire object into memory. There are cases where gcc-generated
4028 // code assumes it is there, even if it could be put entirely into
4029 // registers. (This is not what the doc says.)
4030 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4034 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4035 // copy the pieces of the object that fit into registers from the
4036 // parameter save area.
4037 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4038 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4039 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4040 if (GPR_idx != NumGPRs) {
4041 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4042 MachinePointerInfo(),
4043 false, false, false, 0);
4044 MemOpChains.push_back(Load.getValue(1));
4045 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4046 ArgOffset += PtrByteSize;
4048 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4055 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4056 default: llvm_unreachable("Unexpected ValueType for argument!");
4059 if (GPR_idx != NumGPRs) {
4060 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4062 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4063 isPPC64, isTailCall, false, MemOpChains,
4064 TailCallArguments, dl);
4066 ArgOffset += PtrByteSize;
4070 if (FPR_idx != NumFPRs) {
4071 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4074 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4075 MachinePointerInfo(), false, false, 0);
4076 MemOpChains.push_back(Store);
4078 // Float varargs are always shadowed in available integer registers
4079 if (GPR_idx != NumGPRs) {
4080 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4081 MachinePointerInfo(), false, false,
4083 MemOpChains.push_back(Load.getValue(1));
4084 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4086 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4087 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4088 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4089 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4090 MachinePointerInfo(),
4091 false, false, false, 0);
4092 MemOpChains.push_back(Load.getValue(1));
4093 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4096 // If we have any FPRs remaining, we may also have GPRs remaining.
4097 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4099 if (GPR_idx != NumGPRs)
4101 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4102 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4106 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4107 isPPC64, isTailCall, false, MemOpChains,
4108 TailCallArguments, dl);
4112 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4119 // These go aligned on the stack, or in the corresponding R registers
4120 // when within range. The Darwin PPC ABI doc claims they also go in
4121 // V registers; in fact gcc does this only for arguments that are
4122 // prototyped, not for those that match the ... We do it for all
4123 // arguments, seems to work.
4124 while (ArgOffset % 16 !=0) {
4125 ArgOffset += PtrByteSize;
4126 if (GPR_idx != NumGPRs)
4129 // We could elide this store in the case where the object fits
4130 // entirely in R registers. Maybe later.
4131 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4132 DAG.getConstant(ArgOffset, PtrVT));
4133 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4134 MachinePointerInfo(), false, false, 0);
4135 MemOpChains.push_back(Store);
4136 if (VR_idx != NumVRs) {
4137 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4138 MachinePointerInfo(),
4139 false, false, false, 0);
4140 MemOpChains.push_back(Load.getValue(1));
4141 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4144 for (unsigned i=0; i<16; i+=PtrByteSize) {
4145 if (GPR_idx == NumGPRs)
4147 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4148 DAG.getConstant(i, PtrVT));
4149 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4150 false, false, false, 0);
4151 MemOpChains.push_back(Load.getValue(1));
4152 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4157 // Non-varargs Altivec params generally go in registers, but have
4158 // stack space allocated at the end.
4159 if (VR_idx != NumVRs) {
4160 // Doesn't have GPR space allocated.
4161 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4162 } else if (nAltivecParamsAtEnd==0) {
4163 // We are emitting Altivec params in order.
4164 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4165 isPPC64, isTailCall, true, MemOpChains,
4166 TailCallArguments, dl);
4172 // If all Altivec parameters fit in registers, as they usually do,
4173 // they get stack space following the non-Altivec parameters. We
4174 // don't track this here because nobody below needs it.
4175 // If there are more Altivec parameters than fit in registers emit
4177 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4179 // Offset is aligned; skip 1st 12 params which go in V registers.
4180 ArgOffset = ((ArgOffset+15)/16)*16;
4182 for (unsigned i = 0; i != NumOps; ++i) {
4183 SDValue Arg = OutVals[i];
4184 EVT ArgType = Outs[i].VT;
4185 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4186 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4189 // We are emitting Altivec params in order.
4190 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4191 isPPC64, isTailCall, true, MemOpChains,
4192 TailCallArguments, dl);
4199 if (!MemOpChains.empty())
4200 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4201 &MemOpChains[0], MemOpChains.size());
4203 // On Darwin, R12 must contain the address of an indirect callee. This does
4204 // not mean the MTCTR instruction must use R12; it's easier to model this as
4205 // an extra parameter, so do that.
4207 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4208 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4209 !isBLACompatibleAddress(Callee, DAG))
4210 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4211 PPC::R12), Callee));
4213 // Build a sequence of copy-to-reg nodes chained together with token chain
4214 // and flag operands which copy the outgoing args into the appropriate regs.
4216 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4217 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4218 RegsToPass[i].second, InFlag);
4219 InFlag = Chain.getValue(1);
4223 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4224 FPOp, true, TailCallArguments);
4226 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4227 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4232 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4233 MachineFunction &MF, bool isVarArg,
4234 const SmallVectorImpl<ISD::OutputArg> &Outs,
4235 LLVMContext &Context) const {
4236 SmallVector<CCValAssign, 16> RVLocs;
4237 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4239 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4243 PPCTargetLowering::LowerReturn(SDValue Chain,
4244 CallingConv::ID CallConv, bool isVarArg,
4245 const SmallVectorImpl<ISD::OutputArg> &Outs,
4246 const SmallVectorImpl<SDValue> &OutVals,
4247 DebugLoc dl, SelectionDAG &DAG) const {
4249 SmallVector<CCValAssign, 16> RVLocs;
4250 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4251 getTargetMachine(), RVLocs, *DAG.getContext());
4252 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4254 // If this is the first return lowered for this function, add the regs to the
4255 // liveout set for the function.
4256 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
4257 for (unsigned i = 0; i != RVLocs.size(); ++i)
4258 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
4263 // Copy the result values into the output registers.
4264 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4265 CCValAssign &VA = RVLocs[i];
4266 assert(VA.isRegLoc() && "Can only return in registers!");
4267 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
4269 Flag = Chain.getValue(1);
4273 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
4275 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
4278 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4279 const PPCSubtarget &Subtarget) const {
4280 // When we pop the dynamic allocation we need to restore the SP link.
4281 DebugLoc dl = Op.getDebugLoc();
4283 // Get the corect type for pointers.
4284 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4286 // Construct the stack pointer operand.
4287 bool isPPC64 = Subtarget.isPPC64();
4288 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4289 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4291 // Get the operands for the STACKRESTORE.
4292 SDValue Chain = Op.getOperand(0);
4293 SDValue SaveSP = Op.getOperand(1);
4295 // Load the old link SP.
4296 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4297 MachinePointerInfo(),
4298 false, false, false, 0);
4300 // Restore the stack pointer.
4301 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4303 // Store the old link SP.
4304 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4311 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4312 MachineFunction &MF = DAG.getMachineFunction();
4313 bool isPPC64 = PPCSubTarget.isPPC64();
4314 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4315 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4317 // Get current frame pointer save index. The users of this index will be
4318 // primarily DYNALLOC instructions.
4319 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4320 int RASI = FI->getReturnAddrSaveIndex();
4322 // If the frame pointer save index hasn't been defined yet.
4324 // Find out what the fix offset of the frame pointer save area.
4325 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4326 // Allocate the frame index for frame pointer save area.
4327 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4329 FI->setReturnAddrSaveIndex(RASI);
4331 return DAG.getFrameIndex(RASI, PtrVT);
4335 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4336 MachineFunction &MF = DAG.getMachineFunction();
4337 bool isPPC64 = PPCSubTarget.isPPC64();
4338 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4339 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4341 // Get current frame pointer save index. The users of this index will be
4342 // primarily DYNALLOC instructions.
4343 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4344 int FPSI = FI->getFramePointerSaveIndex();
4346 // If the frame pointer save index hasn't been defined yet.
4348 // Find out what the fix offset of the frame pointer save area.
4349 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4352 // Allocate the frame index for frame pointer save area.
4353 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4355 FI->setFramePointerSaveIndex(FPSI);
4357 return DAG.getFrameIndex(FPSI, PtrVT);
4360 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4362 const PPCSubtarget &Subtarget) const {
4364 SDValue Chain = Op.getOperand(0);
4365 SDValue Size = Op.getOperand(1);
4366 DebugLoc dl = Op.getDebugLoc();
4368 // Get the corect type for pointers.
4369 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4371 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4372 DAG.getConstant(0, PtrVT), Size);
4373 // Construct a node for the frame pointer save index.
4374 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4375 // Build a DYNALLOC node.
4376 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4377 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4378 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4381 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4383 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4384 // Not FP? Not a fsel.
4385 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4386 !Op.getOperand(2).getValueType().isFloatingPoint())
4389 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4391 // Cannot handle SETEQ/SETNE.
4392 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4394 EVT ResVT = Op.getValueType();
4395 EVT CmpVT = Op.getOperand(0).getValueType();
4396 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4397 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4398 DebugLoc dl = Op.getDebugLoc();
4400 // If the RHS of the comparison is a 0.0, we don't need to do the
4401 // subtraction at all.
4402 if (isFloatingPointZero(RHS))
4404 default: break; // SETUO etc aren't handled by fsel.
4407 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4410 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4411 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4412 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4415 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4418 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4419 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4420 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4421 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4426 default: break; // SETUO etc aren't handled by fsel.
4429 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4430 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4431 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4432 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4435 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4436 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4437 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4438 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4441 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4442 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4443 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4444 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4447 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4448 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4449 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4450 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4455 // FIXME: Split this code up when LegalizeDAGTypes lands.
4456 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4457 DebugLoc dl) const {
4458 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4459 SDValue Src = Op.getOperand(0);
4460 if (Src.getValueType() == MVT::f32)
4461 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4464 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4465 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4467 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4472 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4476 // Convert the FP value to an int value through memory.
4477 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4479 // Emit a store to the stack slot.
4480 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4481 MachinePointerInfo(), false, false, 0);
4483 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4485 if (Op.getValueType() == MVT::i32)
4486 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4487 DAG.getConstant(4, FIPtr.getValueType()));
4488 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4489 false, false, false, 0);
4492 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4493 SelectionDAG &DAG) const {
4494 DebugLoc dl = Op.getDebugLoc();
4495 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4496 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4499 if (Op.getOperand(0).getValueType() == MVT::i64) {
4500 SDValue SINT = Op.getOperand(0);
4501 // When converting to single-precision, we actually need to convert
4502 // to double-precision first and then round to single-precision.
4503 // To avoid double-rounding effects during that operation, we have
4504 // to prepare the input operand. Bits that might be truncated when
4505 // converting to double-precision are replaced by a bit that won't
4506 // be lost at this stage, but is below the single-precision rounding
4509 // However, if -enable-unsafe-fp-math is in effect, accept double
4510 // rounding to avoid the extra overhead.
4511 if (Op.getValueType() == MVT::f32 &&
4512 !DAG.getTarget().Options.UnsafeFPMath) {
4514 // Twiddle input to make sure the low 11 bits are zero. (If this
4515 // is the case, we are guaranteed the value will fit into the 53 bit
4516 // mantissa of an IEEE double-precision value without rounding.)
4517 // If any of those low 11 bits were not zero originally, make sure
4518 // bit 12 (value 2048) is set instead, so that the final rounding
4519 // to single-precision gets the correct result.
4520 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4521 SINT, DAG.getConstant(2047, MVT::i64));
4522 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4523 Round, DAG.getConstant(2047, MVT::i64));
4524 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4525 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4526 Round, DAG.getConstant(-2048, MVT::i64));
4528 // However, we cannot use that value unconditionally: if the magnitude
4529 // of the input value is small, the bit-twiddling we did above might
4530 // end up visibly changing the output. Fortunately, in that case, we
4531 // don't need to twiddle bits since the original input will convert
4532 // exactly to double-precision floating-point already. Therefore,
4533 // construct a conditional to use the original value if the top 11
4534 // bits are all sign-bit copies, and use the rounded value computed
4536 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4537 SINT, DAG.getConstant(53, MVT::i32));
4538 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4539 Cond, DAG.getConstant(1, MVT::i64));
4540 Cond = DAG.getSetCC(dl, MVT::i32,
4541 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4543 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4545 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4546 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4547 if (Op.getValueType() == MVT::f32)
4548 FP = DAG.getNode(ISD::FP_ROUND, dl,
4549 MVT::f32, FP, DAG.getIntPtrConstant(0));
4553 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4554 "Unhandled SINT_TO_FP type in custom expander!");
4555 // Since we only generate this in 64-bit mode, we can take advantage of
4556 // 64-bit registers. In particular, sign extend the input value into the
4557 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4558 // then lfd it and fcfid it.
4559 MachineFunction &MF = DAG.getMachineFunction();
4560 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4561 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4562 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4563 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4565 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4568 // STD the extended value into the stack slot.
4569 MachineMemOperand *MMO =
4570 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4571 MachineMemOperand::MOStore, 8, 8);
4572 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4574 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4575 Ops, 4, MVT::i64, MMO);
4576 // Load the value as a double.
4577 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4578 false, false, false, 0);
4580 // FCFID it and return it.
4581 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4582 if (Op.getValueType() == MVT::f32)
4583 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4587 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4588 SelectionDAG &DAG) const {
4589 DebugLoc dl = Op.getDebugLoc();
4591 The rounding mode is in bits 30:31 of FPSR, and has the following
4598 FLT_ROUNDS, on the other hand, expects the following:
4605 To perform the conversion, we do:
4606 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4609 MachineFunction &MF = DAG.getMachineFunction();
4610 EVT VT = Op.getValueType();
4611 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4612 std::vector<EVT> NodeTys;
4613 SDValue MFFSreg, InFlag;
4615 // Save FP Control Word to register
4616 NodeTys.push_back(MVT::f64); // return register
4617 NodeTys.push_back(MVT::Glue); // unused in this context
4618 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4620 // Save FP register to stack slot
4621 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4622 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4623 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4624 StackSlot, MachinePointerInfo(), false, false,0);
4626 // Load FP Control Word from low 32 bits of stack slot.
4627 SDValue Four = DAG.getConstant(4, PtrVT);
4628 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4629 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4630 false, false, false, 0);
4632 // Transform as necessary
4634 DAG.getNode(ISD::AND, dl, MVT::i32,
4635 CWD, DAG.getConstant(3, MVT::i32));
4637 DAG.getNode(ISD::SRL, dl, MVT::i32,
4638 DAG.getNode(ISD::AND, dl, MVT::i32,
4639 DAG.getNode(ISD::XOR, dl, MVT::i32,
4640 CWD, DAG.getConstant(3, MVT::i32)),
4641 DAG.getConstant(3, MVT::i32)),
4642 DAG.getConstant(1, MVT::i32));
4645 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4647 return DAG.getNode((VT.getSizeInBits() < 16 ?
4648 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4651 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4652 EVT VT = Op.getValueType();
4653 unsigned BitWidth = VT.getSizeInBits();
4654 DebugLoc dl = Op.getDebugLoc();
4655 assert(Op.getNumOperands() == 3 &&
4656 VT == Op.getOperand(1).getValueType() &&
4659 // Expand into a bunch of logical ops. Note that these ops
4660 // depend on the PPC behavior for oversized shift amounts.
4661 SDValue Lo = Op.getOperand(0);
4662 SDValue Hi = Op.getOperand(1);
4663 SDValue Amt = Op.getOperand(2);
4664 EVT AmtVT = Amt.getValueType();
4666 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4667 DAG.getConstant(BitWidth, AmtVT), Amt);
4668 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4669 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4670 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4671 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4672 DAG.getConstant(-BitWidth, AmtVT));
4673 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4674 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4675 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4676 SDValue OutOps[] = { OutLo, OutHi };
4677 return DAG.getMergeValues(OutOps, 2, dl);
4680 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4681 EVT VT = Op.getValueType();
4682 DebugLoc dl = Op.getDebugLoc();
4683 unsigned BitWidth = VT.getSizeInBits();
4684 assert(Op.getNumOperands() == 3 &&
4685 VT == Op.getOperand(1).getValueType() &&
4688 // Expand into a bunch of logical ops. Note that these ops
4689 // depend on the PPC behavior for oversized shift amounts.
4690 SDValue Lo = Op.getOperand(0);
4691 SDValue Hi = Op.getOperand(1);
4692 SDValue Amt = Op.getOperand(2);
4693 EVT AmtVT = Amt.getValueType();
4695 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4696 DAG.getConstant(BitWidth, AmtVT), Amt);
4697 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4698 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4699 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4700 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4701 DAG.getConstant(-BitWidth, AmtVT));
4702 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4703 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4704 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4705 SDValue OutOps[] = { OutLo, OutHi };
4706 return DAG.getMergeValues(OutOps, 2, dl);
4709 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4710 DebugLoc dl = Op.getDebugLoc();
4711 EVT VT = Op.getValueType();
4712 unsigned BitWidth = VT.getSizeInBits();
4713 assert(Op.getNumOperands() == 3 &&
4714 VT == Op.getOperand(1).getValueType() &&
4717 // Expand into a bunch of logical ops, followed by a select_cc.
4718 SDValue Lo = Op.getOperand(0);
4719 SDValue Hi = Op.getOperand(1);
4720 SDValue Amt = Op.getOperand(2);
4721 EVT AmtVT = Amt.getValueType();
4723 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4724 DAG.getConstant(BitWidth, AmtVT), Amt);
4725 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4726 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4727 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4728 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4729 DAG.getConstant(-BitWidth, AmtVT));
4730 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4731 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4732 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4733 Tmp4, Tmp6, ISD::SETLE);
4734 SDValue OutOps[] = { OutLo, OutHi };
4735 return DAG.getMergeValues(OutOps, 2, dl);
4738 //===----------------------------------------------------------------------===//
4739 // Vector related lowering.
4742 /// BuildSplatI - Build a canonical splati of Val with an element size of
4743 /// SplatSize. Cast the result to VT.
4744 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4745 SelectionDAG &DAG, DebugLoc dl) {
4746 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4748 static const EVT VTys[] = { // canonical VT to use for each size.
4749 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4752 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4754 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4758 EVT CanonicalVT = VTys[SplatSize-1];
4760 // Build a canonical splat for this value.
4761 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4762 SmallVector<SDValue, 8> Ops;
4763 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4764 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4765 &Ops[0], Ops.size());
4766 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4769 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4770 /// specified intrinsic ID.
4771 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4772 SelectionDAG &DAG, DebugLoc dl,
4773 EVT DestVT = MVT::Other) {
4774 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4775 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4776 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4779 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4780 /// specified intrinsic ID.
4781 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4782 SDValue Op2, SelectionDAG &DAG,
4783 DebugLoc dl, EVT DestVT = MVT::Other) {
4784 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4785 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4786 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4790 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4791 /// amount. The result has the specified value type.
4792 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4793 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4794 // Force LHS/RHS to be the right type.
4795 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4796 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4799 for (unsigned i = 0; i != 16; ++i)
4801 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4802 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4805 // If this is a case we can't handle, return null and let the default
4806 // expansion code take care of it. If we CAN select this case, and if it
4807 // selects to a single instruction, return Op. Otherwise, if we can codegen
4808 // this case more efficiently than a constant pool load, lower it to the
4809 // sequence of ops that should be used.
4810 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4811 SelectionDAG &DAG) const {
4812 DebugLoc dl = Op.getDebugLoc();
4813 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4814 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4816 // Check if this is a splat of a constant value.
4817 APInt APSplatBits, APSplatUndef;
4818 unsigned SplatBitSize;
4820 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4821 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4824 unsigned SplatBits = APSplatBits.getZExtValue();
4825 unsigned SplatUndef = APSplatUndef.getZExtValue();
4826 unsigned SplatSize = SplatBitSize / 8;
4828 // First, handle single instruction cases.
4831 if (SplatBits == 0) {
4832 // Canonicalize all zero vectors to be v4i32.
4833 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4834 SDValue Z = DAG.getConstant(0, MVT::i32);
4835 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4836 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4841 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4842 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4844 if (SextVal >= -16 && SextVal <= 15)
4845 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4848 // Two instruction sequences.
4850 // If this value is in the range [-32,30] and is even, use:
4851 // tmp = VSPLTI[bhw], result = add tmp, tmp
4852 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4853 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4854 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4855 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4858 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4859 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4861 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4862 // Make -1 and vspltisw -1:
4863 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4865 // Make the VSLW intrinsic, computing 0x8000_0000.
4866 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4869 // xor by OnesV to invert it.
4870 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4871 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4874 // Check to see if this is a wide variety of vsplti*, binop self cases.
4875 static const signed char SplatCsts[] = {
4876 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4877 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4880 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4881 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4882 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4883 int i = SplatCsts[idx];
4885 // Figure out what shift amount will be used by altivec if shifted by i in
4887 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4889 // vsplti + shl self.
4890 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4891 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4892 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4893 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4894 Intrinsic::ppc_altivec_vslw
4896 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4897 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4900 // vsplti + srl self.
4901 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4902 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4903 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4904 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4905 Intrinsic::ppc_altivec_vsrw
4907 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4908 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4911 // vsplti + sra self.
4912 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4913 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4914 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4915 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4916 Intrinsic::ppc_altivec_vsraw
4918 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4919 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4922 // vsplti + rol self.
4923 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4924 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4925 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4926 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4927 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4928 Intrinsic::ppc_altivec_vrlw
4930 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4931 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4934 // t = vsplti c, result = vsldoi t, t, 1
4935 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
4936 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4937 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4939 // t = vsplti c, result = vsldoi t, t, 2
4940 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
4941 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4942 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4944 // t = vsplti c, result = vsldoi t, t, 3
4945 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4946 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4947 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4951 // Three instruction sequences.
4953 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4954 if (SextVal >= 0 && SextVal <= 31) {
4955 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4956 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4957 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4958 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4960 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4961 if (SextVal >= -31 && SextVal <= 0) {
4962 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4963 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4964 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4965 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4971 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4972 /// the specified operations to build the shuffle.
4973 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4974 SDValue RHS, SelectionDAG &DAG,
4976 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4977 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4978 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4981 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4993 if (OpNum == OP_COPY) {
4994 if (LHSID == (1*9+2)*9+3) return LHS;
4995 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4999 SDValue OpLHS, OpRHS;
5000 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5001 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5005 default: llvm_unreachable("Unknown i32 permute!");
5007 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5008 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5009 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5010 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5013 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5014 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5015 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5016 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5019 for (unsigned i = 0; i != 16; ++i)
5020 ShufIdxs[i] = (i&3)+0;
5023 for (unsigned i = 0; i != 16; ++i)
5024 ShufIdxs[i] = (i&3)+4;
5027 for (unsigned i = 0; i != 16; ++i)
5028 ShufIdxs[i] = (i&3)+8;
5031 for (unsigned i = 0; i != 16; ++i)
5032 ShufIdxs[i] = (i&3)+12;
5035 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5037 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5039 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5041 EVT VT = OpLHS.getValueType();
5042 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5043 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5044 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5045 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5048 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5049 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5050 /// return the code it can be lowered into. Worst case, it can always be
5051 /// lowered into a vperm.
5052 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5053 SelectionDAG &DAG) const {
5054 DebugLoc dl = Op.getDebugLoc();
5055 SDValue V1 = Op.getOperand(0);
5056 SDValue V2 = Op.getOperand(1);
5057 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5058 EVT VT = Op.getValueType();
5060 // Cases that are handled by instructions that take permute immediates
5061 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5062 // selected by the instruction selector.
5063 if (V2.getOpcode() == ISD::UNDEF) {
5064 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5065 PPC::isSplatShuffleMask(SVOp, 2) ||
5066 PPC::isSplatShuffleMask(SVOp, 4) ||
5067 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5068 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5069 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5070 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5071 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5072 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5073 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5074 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5075 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5080 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5081 // and produce a fixed permutation. If any of these match, do not lower to
5083 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5084 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5085 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5086 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5087 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5088 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5089 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5090 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5091 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5094 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5095 // perfect shuffle table to emit an optimal matching sequence.
5096 ArrayRef<int> PermMask = SVOp->getMask();
5098 unsigned PFIndexes[4];
5099 bool isFourElementShuffle = true;
5100 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5101 unsigned EltNo = 8; // Start out undef.
5102 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5103 if (PermMask[i*4+j] < 0)
5104 continue; // Undef, ignore it.
5106 unsigned ByteSource = PermMask[i*4+j];
5107 if ((ByteSource & 3) != j) {
5108 isFourElementShuffle = false;
5113 EltNo = ByteSource/4;
5114 } else if (EltNo != ByteSource/4) {
5115 isFourElementShuffle = false;
5119 PFIndexes[i] = EltNo;
5122 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5123 // perfect shuffle vector to determine if it is cost effective to do this as
5124 // discrete instructions, or whether we should use a vperm.
5125 if (isFourElementShuffle) {
5126 // Compute the index in the perfect shuffle table.
5127 unsigned PFTableIndex =
5128 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5130 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5131 unsigned Cost = (PFEntry >> 30);
5133 // Determining when to avoid vperm is tricky. Many things affect the cost
5134 // of vperm, particularly how many times the perm mask needs to be computed.
5135 // For example, if the perm mask can be hoisted out of a loop or is already
5136 // used (perhaps because there are multiple permutes with the same shuffle
5137 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5138 // the loop requires an extra register.
5140 // As a compromise, we only emit discrete instructions if the shuffle can be
5141 // generated in 3 or fewer operations. When we have loop information
5142 // available, if this block is within a loop, we should avoid using vperm
5143 // for 3-operation perms and use a constant pool load instead.
5145 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5148 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5149 // vector that will get spilled to the constant pool.
5150 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5152 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5153 // that it is in input element units, not in bytes. Convert now.
5154 EVT EltVT = V1.getValueType().getVectorElementType();
5155 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5157 SmallVector<SDValue, 16> ResultMask;
5158 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5159 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5161 for (unsigned j = 0; j != BytesPerElement; ++j)
5162 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5166 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5167 &ResultMask[0], ResultMask.size());
5168 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5171 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5172 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5173 /// information about the intrinsic.
5174 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5176 unsigned IntrinsicID =
5177 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5180 switch (IntrinsicID) {
5181 default: return false;
5182 // Comparison predicates.
5183 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5184 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5185 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5186 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5187 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5188 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5189 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5190 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5191 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5192 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5193 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5194 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5195 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5197 // Normal Comparisons.
5198 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5199 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5200 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5201 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5202 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5203 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5204 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5205 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5206 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5207 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5208 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5209 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5210 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5215 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5216 /// lower, do it, otherwise return null.
5217 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5218 SelectionDAG &DAG) const {
5219 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5220 // opcode number of the comparison.
5221 DebugLoc dl = Op.getDebugLoc();
5224 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5225 return SDValue(); // Don't custom lower most intrinsics.
5227 // If this is a non-dot comparison, make the VCMP node and we are done.
5229 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5230 Op.getOperand(1), Op.getOperand(2),
5231 DAG.getConstant(CompareOpc, MVT::i32));
5232 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5235 // Create the PPCISD altivec 'dot' comparison node.
5237 Op.getOperand(2), // LHS
5238 Op.getOperand(3), // RHS
5239 DAG.getConstant(CompareOpc, MVT::i32)
5241 std::vector<EVT> VTs;
5242 VTs.push_back(Op.getOperand(2).getValueType());
5243 VTs.push_back(MVT::Glue);
5244 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5246 // Now that we have the comparison, emit a copy from the CR to a GPR.
5247 // This is flagged to the above dot comparison.
5248 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5249 DAG.getRegister(PPC::CR6, MVT::i32),
5250 CompNode.getValue(1));
5252 // Unpack the result based on how the target uses it.
5253 unsigned BitNo; // Bit # of CR6.
5254 bool InvertBit; // Invert result?
5255 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5256 default: // Can't happen, don't crash on invalid number though.
5257 case 0: // Return the value of the EQ bit of CR6.
5258 BitNo = 0; InvertBit = false;
5260 case 1: // Return the inverted value of the EQ bit of CR6.
5261 BitNo = 0; InvertBit = true;
5263 case 2: // Return the value of the LT bit of CR6.
5264 BitNo = 2; InvertBit = false;
5266 case 3: // Return the inverted value of the LT bit of CR6.
5267 BitNo = 2; InvertBit = true;
5271 // Shift the bit into the low position.
5272 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5273 DAG.getConstant(8-(3-BitNo), MVT::i32));
5275 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5276 DAG.getConstant(1, MVT::i32));
5278 // If we are supposed to, toggle the bit.
5280 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5281 DAG.getConstant(1, MVT::i32));
5285 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5286 SelectionDAG &DAG) const {
5287 DebugLoc dl = Op.getDebugLoc();
5288 // Create a stack slot that is 16-byte aligned.
5289 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5290 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5291 EVT PtrVT = getPointerTy();
5292 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5294 // Store the input value into Value#0 of the stack slot.
5295 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5296 Op.getOperand(0), FIdx, MachinePointerInfo(),
5299 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5300 false, false, false, 0);
5303 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5304 DebugLoc dl = Op.getDebugLoc();
5305 if (Op.getValueType() == MVT::v4i32) {
5306 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5308 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5309 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5311 SDValue RHSSwap = // = vrlw RHS, 16
5312 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5314 // Shrinkify inputs to v8i16.
5315 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5316 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5317 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5319 // Low parts multiplied together, generating 32-bit results (we ignore the
5321 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5322 LHS, RHS, DAG, dl, MVT::v4i32);
5324 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5325 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5326 // Shift the high parts up 16 bits.
5327 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5329 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5330 } else if (Op.getValueType() == MVT::v8i16) {
5331 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5333 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5335 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5336 LHS, RHS, Zero, DAG, dl);
5337 } else if (Op.getValueType() == MVT::v16i8) {
5338 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5340 // Multiply the even 8-bit parts, producing 16-bit sums.
5341 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5342 LHS, RHS, DAG, dl, MVT::v8i16);
5343 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5345 // Multiply the odd 8-bit parts, producing 16-bit sums.
5346 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5347 LHS, RHS, DAG, dl, MVT::v8i16);
5348 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5350 // Merge the results together.
5352 for (unsigned i = 0; i != 8; ++i) {
5354 Ops[i*2+1] = 2*i+1+16;
5356 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5358 llvm_unreachable("Unknown mul to lower!");
5362 /// LowerOperation - Provide custom lowering hooks for some operations.
5364 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5365 switch (Op.getOpcode()) {
5366 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5367 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5368 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5369 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5370 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5371 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5372 case ISD::SETCC: return LowerSETCC(Op, DAG);
5373 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5374 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5376 return LowerVASTART(Op, DAG, PPCSubTarget);
5379 return LowerVAARG(Op, DAG, PPCSubTarget);
5381 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5382 case ISD::DYNAMIC_STACKALLOC:
5383 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5385 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5386 case ISD::FP_TO_UINT:
5387 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5389 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5390 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5392 // Lower 64-bit shifts.
5393 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5394 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5395 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5397 // Vector-related lowering.
5398 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5399 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5400 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5401 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5402 case ISD::MUL: return LowerMUL(Op, DAG);
5404 // Frame & Return address.
5405 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5406 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5410 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5411 SmallVectorImpl<SDValue>&Results,
5412 SelectionDAG &DAG) const {
5413 const TargetMachine &TM = getTargetMachine();
5414 DebugLoc dl = N->getDebugLoc();
5415 switch (N->getOpcode()) {
5417 llvm_unreachable("Do not know how to custom type legalize this operation!");
5419 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5420 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5423 EVT VT = N->getValueType(0);
5425 if (VT == MVT::i64) {
5426 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5428 Results.push_back(NewNode);
5429 Results.push_back(NewNode.getValue(1));
5433 case ISD::FP_ROUND_INREG: {
5434 assert(N->getValueType(0) == MVT::ppcf128);
5435 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5436 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5437 MVT::f64, N->getOperand(0),
5438 DAG.getIntPtrConstant(0));
5439 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5440 MVT::f64, N->getOperand(0),
5441 DAG.getIntPtrConstant(1));
5443 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5444 // of the long double, and puts FPSCR back the way it was. We do not
5445 // actually model FPSCR.
5446 std::vector<EVT> NodeTys;
5447 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5449 NodeTys.push_back(MVT::f64); // Return register
5450 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5451 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5452 MFFSreg = Result.getValue(0);
5453 InFlag = Result.getValue(1);
5456 NodeTys.push_back(MVT::Glue); // Returns a flag
5457 Ops[0] = DAG.getConstant(31, MVT::i32);
5459 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5460 InFlag = Result.getValue(0);
5463 NodeTys.push_back(MVT::Glue); // Returns a flag
5464 Ops[0] = DAG.getConstant(30, MVT::i32);
5466 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5467 InFlag = Result.getValue(0);
5470 NodeTys.push_back(MVT::f64); // result of add
5471 NodeTys.push_back(MVT::Glue); // Returns a flag
5475 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5476 FPreg = Result.getValue(0);
5477 InFlag = Result.getValue(1);
5480 NodeTys.push_back(MVT::f64);
5481 Ops[0] = DAG.getConstant(1, MVT::i32);
5485 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5486 FPreg = Result.getValue(0);
5488 // We know the low half is about to be thrown away, so just use something
5490 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5494 case ISD::FP_TO_SINT:
5495 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5501 //===----------------------------------------------------------------------===//
5502 // Other Lowering Code
5503 //===----------------------------------------------------------------------===//
5506 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5507 bool is64bit, unsigned BinOpcode) const {
5508 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5509 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5511 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5512 MachineFunction *F = BB->getParent();
5513 MachineFunction::iterator It = BB;
5516 unsigned dest = MI->getOperand(0).getReg();
5517 unsigned ptrA = MI->getOperand(1).getReg();
5518 unsigned ptrB = MI->getOperand(2).getReg();
5519 unsigned incr = MI->getOperand(3).getReg();
5520 DebugLoc dl = MI->getDebugLoc();
5522 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5523 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5524 F->insert(It, loopMBB);
5525 F->insert(It, exitMBB);
5526 exitMBB->splice(exitMBB->begin(), BB,
5527 llvm::next(MachineBasicBlock::iterator(MI)),
5529 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5531 MachineRegisterInfo &RegInfo = F->getRegInfo();
5532 unsigned TmpReg = (!BinOpcode) ? incr :
5533 RegInfo.createVirtualRegister(
5534 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5535 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5539 // fallthrough --> loopMBB
5540 BB->addSuccessor(loopMBB);
5543 // l[wd]arx dest, ptr
5544 // add r0, dest, incr
5545 // st[wd]cx. r0, ptr
5547 // fallthrough --> exitMBB
5549 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5550 .addReg(ptrA).addReg(ptrB);
5552 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5553 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5554 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5555 BuildMI(BB, dl, TII->get(PPC::BCC))
5556 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5557 BB->addSuccessor(loopMBB);
5558 BB->addSuccessor(exitMBB);
5567 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5568 MachineBasicBlock *BB,
5569 bool is8bit, // operation
5570 unsigned BinOpcode) const {
5571 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5572 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5573 // In 64 bit mode we have to use 64 bits for addresses, even though the
5574 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5575 // registers without caring whether they're 32 or 64, but here we're
5576 // doing actual arithmetic on the addresses.
5577 bool is64bit = PPCSubTarget.isPPC64();
5578 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5580 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5581 MachineFunction *F = BB->getParent();
5582 MachineFunction::iterator It = BB;
5585 unsigned dest = MI->getOperand(0).getReg();
5586 unsigned ptrA = MI->getOperand(1).getReg();
5587 unsigned ptrB = MI->getOperand(2).getReg();
5588 unsigned incr = MI->getOperand(3).getReg();
5589 DebugLoc dl = MI->getDebugLoc();
5591 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5592 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5593 F->insert(It, loopMBB);
5594 F->insert(It, exitMBB);
5595 exitMBB->splice(exitMBB->begin(), BB,
5596 llvm::next(MachineBasicBlock::iterator(MI)),
5598 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5600 MachineRegisterInfo &RegInfo = F->getRegInfo();
5601 const TargetRegisterClass *RC =
5602 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5603 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5604 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5605 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5606 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5607 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5608 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5609 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5610 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5611 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5612 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5613 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5614 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5616 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5620 // fallthrough --> loopMBB
5621 BB->addSuccessor(loopMBB);
5623 // The 4-byte load must be aligned, while a char or short may be
5624 // anywhere in the word. Hence all this nasty bookkeeping code.
5625 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5626 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5627 // xori shift, shift1, 24 [16]
5628 // rlwinm ptr, ptr1, 0, 0, 29
5629 // slw incr2, incr, shift
5630 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5631 // slw mask, mask2, shift
5633 // lwarx tmpDest, ptr
5634 // add tmp, tmpDest, incr2
5635 // andc tmp2, tmpDest, mask
5636 // and tmp3, tmp, mask
5637 // or tmp4, tmp3, tmp2
5640 // fallthrough --> exitMBB
5641 // srw dest, tmpDest, shift
5642 if (ptrA != ZeroReg) {
5643 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5644 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5645 .addReg(ptrA).addReg(ptrB);
5649 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5650 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5651 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5652 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5654 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5655 .addReg(Ptr1Reg).addImm(0).addImm(61);
5657 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5658 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5659 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5660 .addReg(incr).addReg(ShiftReg);
5662 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5664 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5665 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5667 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5668 .addReg(Mask2Reg).addReg(ShiftReg);
5671 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5672 .addReg(ZeroReg).addReg(PtrReg);
5674 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5675 .addReg(Incr2Reg).addReg(TmpDestReg);
5676 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5677 .addReg(TmpDestReg).addReg(MaskReg);
5678 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5679 .addReg(TmpReg).addReg(MaskReg);
5680 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5681 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5682 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5683 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5684 BuildMI(BB, dl, TII->get(PPC::BCC))
5685 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5686 BB->addSuccessor(loopMBB);
5687 BB->addSuccessor(exitMBB);
5692 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5698 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5699 MachineBasicBlock *BB) const {
5700 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5702 // To "insert" these instructions we actually have to insert their
5703 // control-flow patterns.
5704 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5705 MachineFunction::iterator It = BB;
5708 MachineFunction *F = BB->getParent();
5710 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5711 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5712 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5713 PPC::ISEL8 : PPC::ISEL;
5714 unsigned SelectPred = MI->getOperand(4).getImm();
5715 DebugLoc dl = MI->getDebugLoc();
5717 // The SelectPred is ((BI << 5) | BO) for a BCC
5718 unsigned BO = SelectPred & 0xF;
5719 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5721 unsigned TrueOpNo, FalseOpNo;
5728 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5731 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5732 .addReg(MI->getOperand(TrueOpNo).getReg())
5733 .addReg(MI->getOperand(FalseOpNo).getReg())
5734 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5735 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5736 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5737 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5738 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5739 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5742 // The incoming instruction knows the destination vreg to set, the
5743 // condition code register to branch on, the true/false values to
5744 // select between, and a branch opcode to use.
5749 // cmpTY ccX, r1, r2
5751 // fallthrough --> copy0MBB
5752 MachineBasicBlock *thisMBB = BB;
5753 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5754 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5755 unsigned SelectPred = MI->getOperand(4).getImm();
5756 DebugLoc dl = MI->getDebugLoc();
5757 F->insert(It, copy0MBB);
5758 F->insert(It, sinkMBB);
5760 // Transfer the remainder of BB and its successor edges to sinkMBB.
5761 sinkMBB->splice(sinkMBB->begin(), BB,
5762 llvm::next(MachineBasicBlock::iterator(MI)),
5764 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5766 // Next, add the true and fallthrough blocks as its successors.
5767 BB->addSuccessor(copy0MBB);
5768 BB->addSuccessor(sinkMBB);
5770 BuildMI(BB, dl, TII->get(PPC::BCC))
5771 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5774 // %FalseValue = ...
5775 // # fallthrough to sinkMBB
5778 // Update machine-CFG edges
5779 BB->addSuccessor(sinkMBB);
5782 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5785 BuildMI(*BB, BB->begin(), dl,
5786 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5787 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5788 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5790 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5791 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5792 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5793 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5794 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5795 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5796 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5797 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5799 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5800 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5801 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5802 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5803 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5804 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5805 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5806 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5808 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5809 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5810 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5811 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5812 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5813 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5814 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5815 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5817 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5818 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5819 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5820 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5821 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5822 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5823 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5824 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5826 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5827 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5828 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5829 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5830 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5831 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5832 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5833 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5835 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5836 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5837 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5838 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5839 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5840 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5841 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5842 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5844 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5845 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5846 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5847 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5848 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5849 BB = EmitAtomicBinary(MI, BB, false, 0);
5850 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5851 BB = EmitAtomicBinary(MI, BB, true, 0);
5853 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5854 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5855 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5857 unsigned dest = MI->getOperand(0).getReg();
5858 unsigned ptrA = MI->getOperand(1).getReg();
5859 unsigned ptrB = MI->getOperand(2).getReg();
5860 unsigned oldval = MI->getOperand(3).getReg();
5861 unsigned newval = MI->getOperand(4).getReg();
5862 DebugLoc dl = MI->getDebugLoc();
5864 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5865 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5866 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5867 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5868 F->insert(It, loop1MBB);
5869 F->insert(It, loop2MBB);
5870 F->insert(It, midMBB);
5871 F->insert(It, exitMBB);
5872 exitMBB->splice(exitMBB->begin(), BB,
5873 llvm::next(MachineBasicBlock::iterator(MI)),
5875 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5879 // fallthrough --> loopMBB
5880 BB->addSuccessor(loop1MBB);
5883 // l[wd]arx dest, ptr
5884 // cmp[wd] dest, oldval
5887 // st[wd]cx. newval, ptr
5891 // st[wd]cx. dest, ptr
5894 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5895 .addReg(ptrA).addReg(ptrB);
5896 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5897 .addReg(oldval).addReg(dest);
5898 BuildMI(BB, dl, TII->get(PPC::BCC))
5899 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5900 BB->addSuccessor(loop2MBB);
5901 BB->addSuccessor(midMBB);
5904 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5905 .addReg(newval).addReg(ptrA).addReg(ptrB);
5906 BuildMI(BB, dl, TII->get(PPC::BCC))
5907 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5908 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5909 BB->addSuccessor(loop1MBB);
5910 BB->addSuccessor(exitMBB);
5913 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5914 .addReg(dest).addReg(ptrA).addReg(ptrB);
5915 BB->addSuccessor(exitMBB);
5920 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5921 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5922 // We must use 64-bit registers for addresses when targeting 64-bit,
5923 // since we're actually doing arithmetic on them. Other registers
5925 bool is64bit = PPCSubTarget.isPPC64();
5926 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5928 unsigned dest = MI->getOperand(0).getReg();
5929 unsigned ptrA = MI->getOperand(1).getReg();
5930 unsigned ptrB = MI->getOperand(2).getReg();
5931 unsigned oldval = MI->getOperand(3).getReg();
5932 unsigned newval = MI->getOperand(4).getReg();
5933 DebugLoc dl = MI->getDebugLoc();
5935 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5936 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5937 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5938 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5939 F->insert(It, loop1MBB);
5940 F->insert(It, loop2MBB);
5941 F->insert(It, midMBB);
5942 F->insert(It, exitMBB);
5943 exitMBB->splice(exitMBB->begin(), BB,
5944 llvm::next(MachineBasicBlock::iterator(MI)),
5946 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5948 MachineRegisterInfo &RegInfo = F->getRegInfo();
5949 const TargetRegisterClass *RC =
5950 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5951 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5952 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5953 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5954 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5955 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5956 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5957 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5958 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5959 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5960 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5961 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5962 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5963 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5964 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5966 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5967 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5970 // fallthrough --> loopMBB
5971 BB->addSuccessor(loop1MBB);
5973 // The 4-byte load must be aligned, while a char or short may be
5974 // anywhere in the word. Hence all this nasty bookkeeping code.
5975 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5976 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5977 // xori shift, shift1, 24 [16]
5978 // rlwinm ptr, ptr1, 0, 0, 29
5979 // slw newval2, newval, shift
5980 // slw oldval2, oldval,shift
5981 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5982 // slw mask, mask2, shift
5983 // and newval3, newval2, mask
5984 // and oldval3, oldval2, mask
5986 // lwarx tmpDest, ptr
5987 // and tmp, tmpDest, mask
5988 // cmpw tmp, oldval3
5991 // andc tmp2, tmpDest, mask
5992 // or tmp4, tmp2, newval3
5997 // stwcx. tmpDest, ptr
5999 // srw dest, tmpDest, shift
6000 if (ptrA != ZeroReg) {
6001 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6002 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6003 .addReg(ptrA).addReg(ptrB);
6007 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6008 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6009 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6010 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6012 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6013 .addReg(Ptr1Reg).addImm(0).addImm(61);
6015 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6016 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6017 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6018 .addReg(newval).addReg(ShiftReg);
6019 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6020 .addReg(oldval).addReg(ShiftReg);
6022 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6024 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6025 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6026 .addReg(Mask3Reg).addImm(65535);
6028 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6029 .addReg(Mask2Reg).addReg(ShiftReg);
6030 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6031 .addReg(NewVal2Reg).addReg(MaskReg);
6032 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6033 .addReg(OldVal2Reg).addReg(MaskReg);
6036 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6037 .addReg(ZeroReg).addReg(PtrReg);
6038 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6039 .addReg(TmpDestReg).addReg(MaskReg);
6040 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6041 .addReg(TmpReg).addReg(OldVal3Reg);
6042 BuildMI(BB, dl, TII->get(PPC::BCC))
6043 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6044 BB->addSuccessor(loop2MBB);
6045 BB->addSuccessor(midMBB);
6048 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6049 .addReg(TmpDestReg).addReg(MaskReg);
6050 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6051 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6052 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6053 .addReg(ZeroReg).addReg(PtrReg);
6054 BuildMI(BB, dl, TII->get(PPC::BCC))
6055 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6056 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6057 BB->addSuccessor(loop1MBB);
6058 BB->addSuccessor(exitMBB);
6061 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6062 .addReg(ZeroReg).addReg(PtrReg);
6063 BB->addSuccessor(exitMBB);
6068 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6071 llvm_unreachable("Unexpected instr type to insert");
6074 MI->eraseFromParent(); // The pseudo instruction is gone now.
6078 //===----------------------------------------------------------------------===//
6079 // Target Optimization Hooks
6080 //===----------------------------------------------------------------------===//
6082 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6083 DAGCombinerInfo &DCI) const {
6084 const TargetMachine &TM = getTargetMachine();
6085 SelectionDAG &DAG = DCI.DAG;
6086 DebugLoc dl = N->getDebugLoc();
6087 switch (N->getOpcode()) {
6090 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6091 if (C->isNullValue()) // 0 << V -> 0.
6092 return N->getOperand(0);
6096 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6097 if (C->isNullValue()) // 0 >>u V -> 0.
6098 return N->getOperand(0);
6102 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6103 if (C->isNullValue() || // 0 >>s V -> 0.
6104 C->isAllOnesValue()) // -1 >>s V -> -1.
6105 return N->getOperand(0);
6109 case ISD::SINT_TO_FP:
6110 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6111 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6112 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6113 // We allow the src/dst to be either f32/f64, but the intermediate
6114 // type must be i64.
6115 if (N->getOperand(0).getValueType() == MVT::i64 &&
6116 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6117 SDValue Val = N->getOperand(0).getOperand(0);
6118 if (Val.getValueType() == MVT::f32) {
6119 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6120 DCI.AddToWorklist(Val.getNode());
6123 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6124 DCI.AddToWorklist(Val.getNode());
6125 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6126 DCI.AddToWorklist(Val.getNode());
6127 if (N->getValueType(0) == MVT::f32) {
6128 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6129 DAG.getIntPtrConstant(0));
6130 DCI.AddToWorklist(Val.getNode());
6133 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6134 // If the intermediate type is i32, we can avoid the load/store here
6141 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6142 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6143 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6144 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6145 N->getOperand(1).getValueType() == MVT::i32 &&
6146 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6147 SDValue Val = N->getOperand(1).getOperand(0);
6148 if (Val.getValueType() == MVT::f32) {
6149 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6150 DCI.AddToWorklist(Val.getNode());
6152 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6153 DCI.AddToWorklist(Val.getNode());
6155 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6156 N->getOperand(2), N->getOperand(3));
6157 DCI.AddToWorklist(Val.getNode());
6161 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6162 if (cast<StoreSDNode>(N)->isUnindexed() &&
6163 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6164 N->getOperand(1).getNode()->hasOneUse() &&
6165 (N->getOperand(1).getValueType() == MVT::i32 ||
6166 N->getOperand(1).getValueType() == MVT::i16)) {
6167 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6168 // Do an any-extend to 32-bits if this is a half-word input.
6169 if (BSwapOp.getValueType() == MVT::i16)
6170 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6173 N->getOperand(0), BSwapOp, N->getOperand(2),
6174 DAG.getValueType(N->getOperand(1).getValueType())
6177 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6178 Ops, array_lengthof(Ops),
6179 cast<StoreSDNode>(N)->getMemoryVT(),
6180 cast<StoreSDNode>(N)->getMemOperand());
6184 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6185 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6186 N->getOperand(0).hasOneUse() &&
6187 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6188 SDValue Load = N->getOperand(0);
6189 LoadSDNode *LD = cast<LoadSDNode>(Load);
6190 // Create the byte-swapping load.
6192 LD->getChain(), // Chain
6193 LD->getBasePtr(), // Ptr
6194 DAG.getValueType(N->getValueType(0)) // VT
6197 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6198 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6199 LD->getMemoryVT(), LD->getMemOperand());
6201 // If this is an i16 load, insert the truncate.
6202 SDValue ResVal = BSLoad;
6203 if (N->getValueType(0) == MVT::i16)
6204 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6206 // First, combine the bswap away. This makes the value produced by the
6208 DCI.CombineTo(N, ResVal);
6210 // Next, combine the load away, we give it a bogus result value but a real
6211 // chain result. The result value is dead because the bswap is dead.
6212 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6214 // Return N so it doesn't get rechecked!
6215 return SDValue(N, 0);
6219 case PPCISD::VCMP: {
6220 // If a VCMPo node already exists with exactly the same operands as this
6221 // node, use its result instead of this node (VCMPo computes both a CR6 and
6222 // a normal output).
6224 if (!N->getOperand(0).hasOneUse() &&
6225 !N->getOperand(1).hasOneUse() &&
6226 !N->getOperand(2).hasOneUse()) {
6228 // Scan all of the users of the LHS, looking for VCMPo's that match.
6229 SDNode *VCMPoNode = 0;
6231 SDNode *LHSN = N->getOperand(0).getNode();
6232 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6234 if (UI->getOpcode() == PPCISD::VCMPo &&
6235 UI->getOperand(1) == N->getOperand(1) &&
6236 UI->getOperand(2) == N->getOperand(2) &&
6237 UI->getOperand(0) == N->getOperand(0)) {
6242 // If there is no VCMPo node, or if the flag value has a single use, don't
6244 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6247 // Look at the (necessarily single) use of the flag value. If it has a
6248 // chain, this transformation is more complex. Note that multiple things
6249 // could use the value result, which we should ignore.
6250 SDNode *FlagUser = 0;
6251 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6252 FlagUser == 0; ++UI) {
6253 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6255 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6256 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6263 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6264 // give up for right now.
6265 if (FlagUser->getOpcode() == PPCISD::MFCR)
6266 return SDValue(VCMPoNode, 0);
6271 // If this is a branch on an altivec predicate comparison, lower this so
6272 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6273 // lowering is done pre-legalize, because the legalizer lowers the predicate
6274 // compare down to code that is difficult to reassemble.
6275 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6276 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6280 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6281 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6282 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6283 assert(isDot && "Can't compare against a vector result!");
6285 // If this is a comparison against something other than 0/1, then we know
6286 // that the condition is never/always true.
6287 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6288 if (Val != 0 && Val != 1) {
6289 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6290 return N->getOperand(0);
6291 // Always !=, turn it into an unconditional branch.
6292 return DAG.getNode(ISD::BR, dl, MVT::Other,
6293 N->getOperand(0), N->getOperand(4));
6296 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6298 // Create the PPCISD altivec 'dot' comparison node.
6299 std::vector<EVT> VTs;
6301 LHS.getOperand(2), // LHS of compare
6302 LHS.getOperand(3), // RHS of compare
6303 DAG.getConstant(CompareOpc, MVT::i32)
6305 VTs.push_back(LHS.getOperand(2).getValueType());
6306 VTs.push_back(MVT::Glue);
6307 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6309 // Unpack the result based on how the target uses it.
6310 PPC::Predicate CompOpc;
6311 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6312 default: // Can't happen, don't crash on invalid number though.
6313 case 0: // Branch on the value of the EQ bit of CR6.
6314 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6316 case 1: // Branch on the inverted value of the EQ bit of CR6.
6317 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6319 case 2: // Branch on the value of the LT bit of CR6.
6320 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6322 case 3: // Branch on the inverted value of the LT bit of CR6.
6323 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6327 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6328 DAG.getConstant(CompOpc, MVT::i32),
6329 DAG.getRegister(PPC::CR6, MVT::i32),
6330 N->getOperand(4), CompNode.getValue(1));
6339 //===----------------------------------------------------------------------===//
6340 // Inline Assembly Support
6341 //===----------------------------------------------------------------------===//
6343 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6346 const SelectionDAG &DAG,
6347 unsigned Depth) const {
6348 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6349 switch (Op.getOpcode()) {
6351 case PPCISD::LBRX: {
6352 // lhbrx is known to have the top bits cleared out.
6353 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6354 KnownZero = 0xFFFF0000;
6357 case ISD::INTRINSIC_WO_CHAIN: {
6358 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6360 case Intrinsic::ppc_altivec_vcmpbfp_p:
6361 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6362 case Intrinsic::ppc_altivec_vcmpequb_p:
6363 case Intrinsic::ppc_altivec_vcmpequh_p:
6364 case Intrinsic::ppc_altivec_vcmpequw_p:
6365 case Intrinsic::ppc_altivec_vcmpgefp_p:
6366 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6367 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6368 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6369 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6370 case Intrinsic::ppc_altivec_vcmpgtub_p:
6371 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6372 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6373 KnownZero = ~1U; // All bits but the low one are known to be zero.
6381 /// getConstraintType - Given a constraint, return the type of
6382 /// constraint it is for this target.
6383 PPCTargetLowering::ConstraintType
6384 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6385 if (Constraint.size() == 1) {
6386 switch (Constraint[0]) {
6393 return C_RegisterClass;
6396 return TargetLowering::getConstraintType(Constraint);
6399 /// Examine constraint type and operand type and determine a weight value.
6400 /// This object must already have been set up with the operand type
6401 /// and the current alternative constraint selected.
6402 TargetLowering::ConstraintWeight
6403 PPCTargetLowering::getSingleConstraintMatchWeight(
6404 AsmOperandInfo &info, const char *constraint) const {
6405 ConstraintWeight weight = CW_Invalid;
6406 Value *CallOperandVal = info.CallOperandVal;
6407 // If we don't have a value, we can't do a match,
6408 // but allow it at the lowest weight.
6409 if (CallOperandVal == NULL)
6411 Type *type = CallOperandVal->getType();
6412 // Look at the constraint type.
6413 switch (*constraint) {
6415 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6418 if (type->isIntegerTy())
6419 weight = CW_Register;
6422 if (type->isFloatTy())
6423 weight = CW_Register;
6426 if (type->isDoubleTy())
6427 weight = CW_Register;
6430 if (type->isVectorTy())
6431 weight = CW_Register;
6434 weight = CW_Register;
6440 std::pair<unsigned, const TargetRegisterClass*>
6441 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6443 if (Constraint.size() == 1) {
6444 // GCC RS6000 Constraint Letters
6445 switch (Constraint[0]) {
6448 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6449 return std::make_pair(0U, &PPC::G8RCRegClass);
6450 return std::make_pair(0U, &PPC::GPRCRegClass);
6453 return std::make_pair(0U, &PPC::F4RCRegClass);
6455 return std::make_pair(0U, &PPC::F8RCRegClass);
6458 return std::make_pair(0U, &PPC::VRRCRegClass);
6460 return std::make_pair(0U, &PPC::CRRCRegClass);
6464 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6468 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6469 /// vector. If it is invalid, don't add anything to Ops.
6470 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6471 std::string &Constraint,
6472 std::vector<SDValue>&Ops,
6473 SelectionDAG &DAG) const {
6474 SDValue Result(0,0);
6476 // Only support length 1 constraints.
6477 if (Constraint.length() > 1) return;
6479 char Letter = Constraint[0];
6490 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6491 if (!CST) return; // Must be an immediate to match.
6492 unsigned Value = CST->getZExtValue();
6494 default: llvm_unreachable("Unknown constraint letter!");
6495 case 'I': // "I" is a signed 16-bit constant.
6496 if ((short)Value == (int)Value)
6497 Result = DAG.getTargetConstant(Value, Op.getValueType());
6499 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6500 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6501 if ((short)Value == 0)
6502 Result = DAG.getTargetConstant(Value, Op.getValueType());
6504 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6505 if ((Value >> 16) == 0)
6506 Result = DAG.getTargetConstant(Value, Op.getValueType());
6508 case 'M': // "M" is a constant that is greater than 31.
6510 Result = DAG.getTargetConstant(Value, Op.getValueType());
6512 case 'N': // "N" is a positive constant that is an exact power of two.
6513 if ((int)Value > 0 && isPowerOf2_32(Value))
6514 Result = DAG.getTargetConstant(Value, Op.getValueType());
6516 case 'O': // "O" is the constant zero.
6518 Result = DAG.getTargetConstant(Value, Op.getValueType());
6520 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6521 if ((short)-Value == (int)-Value)
6522 Result = DAG.getTargetConstant(Value, Op.getValueType());
6529 if (Result.getNode()) {
6530 Ops.push_back(Result);
6534 // Handle standard constraint letters.
6535 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6538 // isLegalAddressingMode - Return true if the addressing mode represented
6539 // by AM is legal for this target, for a load/store of the specified type.
6540 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6542 // FIXME: PPC does not allow r+i addressing modes for vectors!
6544 // PPC allows a sign-extended 16-bit immediate field.
6545 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6548 // No global is ever allowed as a base.
6552 // PPC only support r+r,
6554 case 0: // "r+i" or just "i", depending on HasBaseReg.
6557 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6559 // Otherwise we have r+r or r+i.
6562 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6564 // Allow 2*r as r+r.
6567 // No other scales are supported.
6574 /// isLegalAddressImmediate - Return true if the integer value can be used
6575 /// as the offset of the target addressing mode for load / store of the
6577 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6578 // PPC allows a sign-extended 16-bit immediate field.
6579 return (V > -(1 << 16) && V < (1 << 16)-1);
6582 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6586 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6587 SelectionDAG &DAG) const {
6588 MachineFunction &MF = DAG.getMachineFunction();
6589 MachineFrameInfo *MFI = MF.getFrameInfo();
6590 MFI->setReturnAddressIsTaken(true);
6592 DebugLoc dl = Op.getDebugLoc();
6593 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6595 // Make sure the function does not optimize away the store of the RA to
6597 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6598 FuncInfo->setLRStoreRequired();
6599 bool isPPC64 = PPCSubTarget.isPPC64();
6600 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6603 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6606 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6607 isPPC64? MVT::i64 : MVT::i32);
6608 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6609 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6611 MachinePointerInfo(), false, false, false, 0);
6614 // Just load the return address off the stack.
6615 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6616 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6617 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6620 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6621 SelectionDAG &DAG) const {
6622 DebugLoc dl = Op.getDebugLoc();
6623 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6625 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6626 bool isPPC64 = PtrVT == MVT::i64;
6628 MachineFunction &MF = DAG.getMachineFunction();
6629 MachineFrameInfo *MFI = MF.getFrameInfo();
6630 MFI->setFrameAddressIsTaken(true);
6631 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6632 MFI->hasVarSizedObjects()) &&
6633 MFI->getStackSize() &&
6634 !MF.getFunction()->getFnAttributes().
6635 hasAttribute(Attributes::Naked);
6636 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6637 (is31 ? PPC::R31 : PPC::R1);
6638 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6641 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6642 FrameAddr, MachinePointerInfo(), false, false,
6648 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6649 // The PowerPC target isn't yet aware of offsets.
6653 /// getOptimalMemOpType - Returns the target specific optimal type for load
6654 /// and store operations as a result of memset, memcpy, and memmove
6655 /// lowering. If DstAlign is zero that means it's safe to destination
6656 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6657 /// means there isn't a need to check it against alignment requirement,
6658 /// probably because the source does not need to be loaded. If
6659 /// 'IsZeroVal' is true, that means it's safe to return a
6660 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6661 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6662 /// constant so it does not need to be loaded.
6663 /// It returns EVT::Other if the type should be determined using generic
6664 /// target-independent logic.
6665 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6666 unsigned DstAlign, unsigned SrcAlign,
6669 MachineFunction &MF) const {
6670 if (this->PPCSubTarget.isPPC64()) {
6677 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6678 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6679 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6680 /// is expanded to mul + add.
6681 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6685 switch (VT.getSimpleVT().SimpleTy) {
6697 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6699 return TargetLowering::getSchedulingPreference(N);