1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCMachineFunctionInfo.h"
16 #include "PPCPerfectShuffle.h"
17 #include "PPCTargetMachine.h"
18 #include "MCTargetDesc/PPCPredicates.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstrBuilder.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/SelectionDAG.h"
31 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetOptions.h"
39 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
40 CCValAssign::LocInfo &LocInfo,
41 ISD::ArgFlagsTy &ArgFlags,
43 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
50 CCValAssign::LocInfo &LocInfo,
51 ISD::ArgFlagsTy &ArgFlags,
54 static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc",
55 cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden);
57 static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref",
58 cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden);
60 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
61 if (TM.getSubtargetImpl()->isDarwin())
62 return new TargetLoweringObjectFileMachO();
64 return new TargetLoweringObjectFileELF();
67 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
68 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
69 const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
73 // Use _setjmp/_longjmp instead of setjmp/longjmp.
74 setUseUnderscoreSetJmp(true);
75 setUseUnderscoreLongJmp(true);
77 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
78 // arguments are at least 4/8 bytes aligned.
79 bool isPPC64 = Subtarget->isPPC64();
80 setMinStackArgumentAlignment(isPPC64 ? 8:4);
82 // Set up the register classes.
83 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
84 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
85 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
87 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
88 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
89 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
91 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
93 // PowerPC has pre-inc load and store's.
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
96 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
97 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
98 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
101 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
102 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
103 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
105 // This is used in the ppcf128->int sequence. Note it has different semantics
106 // from FP_ROUND: that rounds to nearest, this rounds to zero.
107 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
109 // We do not currently implement these libm ops for PowerPC.
110 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
111 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
112 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
113 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
114 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
116 // PowerPC has no SREM/UREM instructions
117 setOperationAction(ISD::SREM, MVT::i32, Expand);
118 setOperationAction(ISD::UREM, MVT::i32, Expand);
119 setOperationAction(ISD::SREM, MVT::i64, Expand);
120 setOperationAction(ISD::UREM, MVT::i64, Expand);
122 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
123 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
124 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
125 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
126 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
127 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
128 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
129 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
130 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
132 // We don't support sin/cos/sqrt/fmod/pow
133 setOperationAction(ISD::FSIN , MVT::f64, Expand);
134 setOperationAction(ISD::FCOS , MVT::f64, Expand);
135 setOperationAction(ISD::FREM , MVT::f64, Expand);
136 setOperationAction(ISD::FPOW , MVT::f64, Expand);
137 setOperationAction(ISD::FMA , MVT::f64, Legal);
138 setOperationAction(ISD::FSIN , MVT::f32, Expand);
139 setOperationAction(ISD::FCOS , MVT::f32, Expand);
140 setOperationAction(ISD::FREM , MVT::f32, Expand);
141 setOperationAction(ISD::FPOW , MVT::f32, Expand);
142 setOperationAction(ISD::FMA , MVT::f32, Legal);
144 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
146 // If we're enabling GP optimizations, use hardware square root
147 if (!Subtarget->hasFSQRT()) {
148 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
149 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
152 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
153 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
155 // PowerPC does not have BSWAP, CTPOP or CTTZ
156 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
157 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
158 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
160 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
161 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
162 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
163 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
164 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
165 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
167 // PowerPC does not have ROTR
168 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
169 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
171 // PowerPC does not have Select
172 setOperationAction(ISD::SELECT, MVT::i32, Expand);
173 setOperationAction(ISD::SELECT, MVT::i64, Expand);
174 setOperationAction(ISD::SELECT, MVT::f32, Expand);
175 setOperationAction(ISD::SELECT, MVT::f64, Expand);
177 // PowerPC wants to turn select_cc of FP into fsel when possible.
178 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
179 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
181 // PowerPC wants to optimize integer setcc a bit
182 setOperationAction(ISD::SETCC, MVT::i32, Custom);
184 // PowerPC does not have BRCOND which requires SetCC
185 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
187 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
189 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
190 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
192 // PowerPC does not have [U|S]INT_TO_FP
193 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
194 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
196 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
197 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
198 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
199 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
201 // We cannot sextinreg(i1). Expand to shifts.
202 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
204 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
205 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
206 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
207 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
210 // We want to legalize GlobalAddress and ConstantPool nodes into the
211 // appropriate instructions to materialize the address.
212 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
213 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
214 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
215 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
216 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
217 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
218 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
219 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
220 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
221 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
224 setOperationAction(ISD::TRAP, MVT::Other, Legal);
226 // TRAMPOLINE is custom lowered.
227 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
228 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
230 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
231 setOperationAction(ISD::VASTART , MVT::Other, Custom);
233 if (Subtarget->isSVR4ABI()) {
235 // VAARG always uses double-word chunks, so promote anything smaller.
236 setOperationAction(ISD::VAARG, MVT::i1, Promote);
237 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
238 setOperationAction(ISD::VAARG, MVT::i8, Promote);
239 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
240 setOperationAction(ISD::VAARG, MVT::i16, Promote);
241 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
242 setOperationAction(ISD::VAARG, MVT::i32, Promote);
243 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
244 setOperationAction(ISD::VAARG, MVT::Other, Expand);
246 // VAARG is custom lowered with the 32-bit SVR4 ABI.
247 setOperationAction(ISD::VAARG, MVT::Other, Custom);
248 setOperationAction(ISD::VAARG, MVT::i64, Custom);
251 setOperationAction(ISD::VAARG, MVT::Other, Expand);
253 // Use the default implementation.
254 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
255 setOperationAction(ISD::VAEND , MVT::Other, Expand);
256 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
257 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
258 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
259 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
261 // We want to custom lower some of our intrinsics.
262 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
264 // Comparisons that require checking two conditions.
265 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
266 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
267 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
270 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
271 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
272 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
273 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
274 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
275 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
276 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
278 if (Subtarget->has64BitSupport()) {
279 // They also have instructions for converting between i64 and fp.
280 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
281 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
282 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
283 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
284 // This is just the low 32 bits of a (signed) fp->i64 conversion.
285 // We cannot do this with Promote because i64 is not a legal type.
286 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
288 // FIXME: disable this lowered code. This generates 64-bit register values,
289 // and we don't model the fact that the top part is clobbered by calls. We
290 // need to flag these together so that the value isn't live across a call.
291 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
293 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
294 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
297 if (Subtarget->use64BitRegs()) {
298 // 64-bit PowerPC implementations can support i64 types directly
299 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
300 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
301 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
302 // 64-bit PowerPC wants to expand i128 shifts itself.
303 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
304 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
305 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
307 // 32-bit PowerPC wants to expand i64 shifts itself.
308 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
309 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
310 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
313 if (Subtarget->hasAltivec()) {
314 // First set operation action for all vector types to expand. Then we
315 // will selectively turn on ones that can be effectively codegen'd.
316 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
317 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
318 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
320 // add/sub are legal for all supported vector VT's.
321 setOperationAction(ISD::ADD , VT, Legal);
322 setOperationAction(ISD::SUB , VT, Legal);
324 // We promote all shuffles to v16i8.
325 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
326 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
328 // We promote all non-typed operations to v4i32.
329 setOperationAction(ISD::AND , VT, Promote);
330 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
331 setOperationAction(ISD::OR , VT, Promote);
332 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
333 setOperationAction(ISD::XOR , VT, Promote);
334 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
335 setOperationAction(ISD::LOAD , VT, Promote);
336 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
337 setOperationAction(ISD::SELECT, VT, Promote);
338 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
339 setOperationAction(ISD::STORE, VT, Promote);
340 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
342 // No other operations are legal.
343 setOperationAction(ISD::MUL , VT, Expand);
344 setOperationAction(ISD::SDIV, VT, Expand);
345 setOperationAction(ISD::SREM, VT, Expand);
346 setOperationAction(ISD::UDIV, VT, Expand);
347 setOperationAction(ISD::UREM, VT, Expand);
348 setOperationAction(ISD::FDIV, VT, Expand);
349 setOperationAction(ISD::FNEG, VT, Expand);
350 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
351 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
352 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
353 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
354 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
355 setOperationAction(ISD::UDIVREM, VT, Expand);
356 setOperationAction(ISD::SDIVREM, VT, Expand);
357 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
358 setOperationAction(ISD::FPOW, VT, Expand);
359 setOperationAction(ISD::CTPOP, VT, Expand);
360 setOperationAction(ISD::CTLZ, VT, Expand);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
362 setOperationAction(ISD::CTTZ, VT, Expand);
363 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
366 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
367 // with merges, splats, etc.
368 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
370 setOperationAction(ISD::AND , MVT::v4i32, Legal);
371 setOperationAction(ISD::OR , MVT::v4i32, Legal);
372 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
373 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
374 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
375 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
376 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
377 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
378 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
379 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
381 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
382 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
383 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
384 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
386 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
387 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
388 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
389 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
390 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
392 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
393 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
395 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
396 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
397 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
398 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
401 if (Subtarget->has64BitSupport()) {
402 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
403 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
406 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
407 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
409 setBooleanContents(ZeroOrOneBooleanContent);
410 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
413 setStackPointerRegisterToSaveRestore(PPC::X1);
414 setExceptionPointerRegister(PPC::X3);
415 setExceptionSelectorRegister(PPC::X4);
417 setStackPointerRegisterToSaveRestore(PPC::R1);
418 setExceptionPointerRegister(PPC::R3);
419 setExceptionSelectorRegister(PPC::R4);
422 // We have target-specific dag combine patterns for the following nodes:
423 setTargetDAGCombine(ISD::SINT_TO_FP);
424 setTargetDAGCombine(ISD::STORE);
425 setTargetDAGCombine(ISD::BR_CC);
426 setTargetDAGCombine(ISD::BSWAP);
428 // Darwin long double math library functions have $LDBL128 appended.
429 if (Subtarget->isDarwin()) {
430 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
431 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
432 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
433 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
434 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
435 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
436 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
437 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
438 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
439 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
442 setMinFunctionAlignment(2);
443 if (PPCSubTarget.isDarwin())
444 setPrefFunctionAlignment(4);
446 if (isPPC64 && Subtarget->isJITCodeModel())
447 // Temporary workaround for the inability of PPC64 JIT to handle jump
449 setSupportJumpTables(false);
451 setInsertFencesForAtomic(true);
453 setSchedulingPreference(Sched::Hybrid);
455 computeRegisterProperties();
457 // The Freescale cores does better with aggressive inlining of memcpy and
458 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
459 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
460 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
461 maxStoresPerMemset = 32;
462 maxStoresPerMemsetOptSize = 16;
463 maxStoresPerMemcpy = 32;
464 maxStoresPerMemcpyOptSize = 8;
465 maxStoresPerMemmove = 32;
466 maxStoresPerMemmoveOptSize = 8;
468 setPrefFunctionAlignment(4);
469 benefitFromCodePlacementOpt = true;
473 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
474 /// function arguments in the caller parameter area.
475 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
476 const TargetMachine &TM = getTargetMachine();
477 // Darwin passes everything on 4 byte boundary.
478 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
481 // 16byte and wider vectors are passed on 16byte boundary.
482 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
483 if (VTy->getBitWidth() >= 128)
486 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
487 if (PPCSubTarget.isPPC64())
493 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
496 case PPCISD::FSEL: return "PPCISD::FSEL";
497 case PPCISD::FCFID: return "PPCISD::FCFID";
498 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
499 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
500 case PPCISD::STFIWX: return "PPCISD::STFIWX";
501 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
502 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
503 case PPCISD::VPERM: return "PPCISD::VPERM";
504 case PPCISD::Hi: return "PPCISD::Hi";
505 case PPCISD::Lo: return "PPCISD::Lo";
506 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
507 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
508 case PPCISD::LOAD: return "PPCISD::LOAD";
509 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
510 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
511 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
512 case PPCISD::SRL: return "PPCISD::SRL";
513 case PPCISD::SRA: return "PPCISD::SRA";
514 case PPCISD::SHL: return "PPCISD::SHL";
515 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
516 case PPCISD::STD_32: return "PPCISD::STD_32";
517 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
518 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
519 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
520 case PPCISD::NOP: return "PPCISD::NOP";
521 case PPCISD::MTCTR: return "PPCISD::MTCTR";
522 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
523 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
524 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
525 case PPCISD::MFCR: return "PPCISD::MFCR";
526 case PPCISD::VCMP: return "PPCISD::VCMP";
527 case PPCISD::VCMPo: return "PPCISD::VCMPo";
528 case PPCISD::LBRX: return "PPCISD::LBRX";
529 case PPCISD::STBRX: return "PPCISD::STBRX";
530 case PPCISD::LARX: return "PPCISD::LARX";
531 case PPCISD::STCX: return "PPCISD::STCX";
532 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
533 case PPCISD::MFFS: return "PPCISD::MFFS";
534 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
535 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
536 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
537 case PPCISD::MTFSF: return "PPCISD::MTFSF";
538 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
539 case PPCISD::CR6SET: return "PPCISD::CR6SET";
540 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
544 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
547 return VT.changeVectorElementTypeToInteger();
550 //===----------------------------------------------------------------------===//
551 // Node matching predicates, for use by the tblgen matching code.
552 //===----------------------------------------------------------------------===//
554 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
555 static bool isFloatingPointZero(SDValue Op) {
556 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
557 return CFP->getValueAPF().isZero();
558 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
559 // Maybe this has already been legalized into the constant pool?
560 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
561 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
562 return CFP->getValueAPF().isZero();
567 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
568 /// true if Op is undef or if it matches the specified value.
569 static bool isConstantOrUndef(int Op, int Val) {
570 return Op < 0 || Op == Val;
573 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
574 /// VPKUHUM instruction.
575 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
577 for (unsigned i = 0; i != 16; ++i)
578 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
581 for (unsigned i = 0; i != 8; ++i)
582 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
583 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
589 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
590 /// VPKUWUM instruction.
591 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
593 for (unsigned i = 0; i != 16; i += 2)
594 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
595 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
598 for (unsigned i = 0; i != 8; i += 2)
599 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
600 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
601 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
602 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
608 /// isVMerge - Common function, used to match vmrg* shuffles.
610 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
611 unsigned LHSStart, unsigned RHSStart) {
612 assert(N->getValueType(0) == MVT::v16i8 &&
613 "PPC only supports shuffles by bytes!");
614 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
615 "Unsupported merge size!");
617 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
618 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
619 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
620 LHSStart+j+i*UnitSize) ||
621 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
622 RHSStart+j+i*UnitSize))
628 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
629 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
630 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
633 return isVMerge(N, UnitSize, 8, 24);
634 return isVMerge(N, UnitSize, 8, 8);
637 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
638 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
639 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
642 return isVMerge(N, UnitSize, 0, 16);
643 return isVMerge(N, UnitSize, 0, 0);
647 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
648 /// amount, otherwise return -1.
649 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
650 assert(N->getValueType(0) == MVT::v16i8 &&
651 "PPC only supports shuffles by bytes!");
653 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
655 // Find the first non-undef value in the shuffle mask.
657 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
660 if (i == 16) return -1; // all undef.
662 // Otherwise, check to see if the rest of the elements are consecutively
663 // numbered from this value.
664 unsigned ShiftAmt = SVOp->getMaskElt(i);
665 if (ShiftAmt < i) return -1;
669 // Check the rest of the elements to see if they are consecutive.
670 for (++i; i != 16; ++i)
671 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
674 // Check the rest of the elements to see if they are consecutive.
675 for (++i; i != 16; ++i)
676 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
682 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
683 /// specifies a splat of a single element that is suitable for input to
684 /// VSPLTB/VSPLTH/VSPLTW.
685 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
686 assert(N->getValueType(0) == MVT::v16i8 &&
687 (EltSize == 1 || EltSize == 2 || EltSize == 4));
689 // This is a splat operation if each element of the permute is the same, and
690 // if the value doesn't reference the second vector.
691 unsigned ElementBase = N->getMaskElt(0);
693 // FIXME: Handle UNDEF elements too!
694 if (ElementBase >= 16)
697 // Check that the indices are consecutive, in the case of a multi-byte element
698 // splatted with a v16i8 mask.
699 for (unsigned i = 1; i != EltSize; ++i)
700 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
703 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
704 if (N->getMaskElt(i) < 0) continue;
705 for (unsigned j = 0; j != EltSize; ++j)
706 if (N->getMaskElt(i+j) != N->getMaskElt(j))
712 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
714 bool PPC::isAllNegativeZeroVector(SDNode *N) {
715 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
717 APInt APVal, APUndef;
721 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
722 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
723 return CFP->getValueAPF().isNegZero();
728 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
729 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
730 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
731 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
732 assert(isSplatShuffleMask(SVOp, EltSize));
733 return SVOp->getMaskElt(0) / EltSize;
736 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
737 /// by using a vspltis[bhw] instruction of the specified element size, return
738 /// the constant being splatted. The ByteSize field indicates the number of
739 /// bytes of each element [124] -> [bhw].
740 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
743 // If ByteSize of the splat is bigger than the element size of the
744 // build_vector, then we have a case where we are checking for a splat where
745 // multiple elements of the buildvector are folded together into a single
746 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
747 unsigned EltSize = 16/N->getNumOperands();
748 if (EltSize < ByteSize) {
749 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
750 SDValue UniquedVals[4];
751 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
753 // See if all of the elements in the buildvector agree across.
754 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
755 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
756 // If the element isn't a constant, bail fully out.
757 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
760 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
761 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
762 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
763 return SDValue(); // no match.
766 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
767 // either constant or undef values that are identical for each chunk. See
768 // if these chunks can form into a larger vspltis*.
770 // Check to see if all of the leading entries are either 0 or -1. If
771 // neither, then this won't fit into the immediate field.
772 bool LeadingZero = true;
773 bool LeadingOnes = true;
774 for (unsigned i = 0; i != Multiple-1; ++i) {
775 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
777 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
778 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
780 // Finally, check the least significant entry.
782 if (UniquedVals[Multiple-1].getNode() == 0)
783 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
784 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
786 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
789 if (UniquedVals[Multiple-1].getNode() == 0)
790 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
791 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
792 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
793 return DAG.getTargetConstant(Val, MVT::i32);
799 // Check to see if this buildvec has a single non-undef value in its elements.
800 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
801 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
802 if (OpVal.getNode() == 0)
803 OpVal = N->getOperand(i);
804 else if (OpVal != N->getOperand(i))
808 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
810 unsigned ValSizeInBytes = EltSize;
812 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
813 Value = CN->getZExtValue();
814 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
815 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
816 Value = FloatToBits(CN->getValueAPF().convertToFloat());
819 // If the splat value is larger than the element value, then we can never do
820 // this splat. The only case that we could fit the replicated bits into our
821 // immediate field for would be zero, and we prefer to use vxor for it.
822 if (ValSizeInBytes < ByteSize) return SDValue();
824 // If the element value is larger than the splat value, cut it in half and
825 // check to see if the two halves are equal. Continue doing this until we
826 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
827 while (ValSizeInBytes > ByteSize) {
828 ValSizeInBytes >>= 1;
830 // If the top half equals the bottom half, we're still ok.
831 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
832 (Value & ((1 << (8*ValSizeInBytes))-1)))
836 // Properly sign extend the value.
837 int MaskVal = SignExtend32(Value, ByteSize * 8);
839 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
840 if (MaskVal == 0) return SDValue();
842 // Finally, if this value fits in a 5 bit sext field, return it
843 if (SignExtend32<5>(MaskVal) == MaskVal)
844 return DAG.getTargetConstant(MaskVal, MVT::i32);
848 //===----------------------------------------------------------------------===//
849 // Addressing Mode Selection
850 //===----------------------------------------------------------------------===//
852 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
853 /// or 64-bit immediate, and if the value can be accurately represented as a
854 /// sign extension from a 16-bit value. If so, this returns true and the
856 static bool isIntS16Immediate(SDNode *N, short &Imm) {
857 if (N->getOpcode() != ISD::Constant)
860 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
861 if (N->getValueType(0) == MVT::i32)
862 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
864 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
866 static bool isIntS16Immediate(SDValue Op, short &Imm) {
867 return isIntS16Immediate(Op.getNode(), Imm);
871 /// SelectAddressRegReg - Given the specified addressed, check to see if it
872 /// can be represented as an indexed [r+r] operation. Returns false if it
873 /// can be more efficiently represented with [r+imm].
874 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
876 SelectionDAG &DAG) const {
878 if (N.getOpcode() == ISD::ADD) {
879 if (isIntS16Immediate(N.getOperand(1), imm))
881 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
884 Base = N.getOperand(0);
885 Index = N.getOperand(1);
887 } else if (N.getOpcode() == ISD::OR) {
888 if (isIntS16Immediate(N.getOperand(1), imm))
889 return false; // r+i can fold it if we can.
891 // If this is an or of disjoint bitfields, we can codegen this as an add
892 // (for better address arithmetic) if the LHS and RHS of the OR are provably
894 APInt LHSKnownZero, LHSKnownOne;
895 APInt RHSKnownZero, RHSKnownOne;
896 DAG.ComputeMaskedBits(N.getOperand(0),
897 LHSKnownZero, LHSKnownOne);
899 if (LHSKnownZero.getBoolValue()) {
900 DAG.ComputeMaskedBits(N.getOperand(1),
901 RHSKnownZero, RHSKnownOne);
902 // If all of the bits are known zero on the LHS or RHS, the add won't
904 if (~(LHSKnownZero | RHSKnownZero) == 0) {
905 Base = N.getOperand(0);
906 Index = N.getOperand(1);
915 /// Returns true if the address N can be represented by a base register plus
916 /// a signed 16-bit displacement [r+imm], and if it is not better
917 /// represented as reg+reg.
918 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
920 SelectionDAG &DAG) const {
921 // FIXME dl should come from parent load or store, not from address
922 DebugLoc dl = N.getDebugLoc();
923 // If this can be more profitably realized as r+r, fail.
924 if (SelectAddressRegReg(N, Disp, Base, DAG))
927 if (N.getOpcode() == ISD::ADD) {
929 if (isIntS16Immediate(N.getOperand(1), imm)) {
930 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
931 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
932 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
934 Base = N.getOperand(0);
936 return true; // [r+i]
937 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
938 // Match LOAD (ADD (X, Lo(G))).
939 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
940 && "Cannot handle constant offsets yet!");
941 Disp = N.getOperand(1).getOperand(0); // The global address.
942 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
943 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
944 Disp.getOpcode() == ISD::TargetConstantPool ||
945 Disp.getOpcode() == ISD::TargetJumpTable);
946 Base = N.getOperand(0);
947 return true; // [&g+r]
949 } else if (N.getOpcode() == ISD::OR) {
951 if (isIntS16Immediate(N.getOperand(1), imm)) {
952 // If this is an or of disjoint bitfields, we can codegen this as an add
953 // (for better address arithmetic) if the LHS and RHS of the OR are
954 // provably disjoint.
955 APInt LHSKnownZero, LHSKnownOne;
956 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
958 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
959 // If all of the bits are known zero on the LHS or RHS, the add won't
961 Base = N.getOperand(0);
962 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
966 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
967 // Loading from a constant address.
969 // If this address fits entirely in a 16-bit sext immediate field, codegen
972 if (isIntS16Immediate(CN, Imm)) {
973 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
974 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
975 CN->getValueType(0));
979 // Handle 32-bit sext immediates with LIS + addr mode.
980 if (CN->getValueType(0) == MVT::i32 ||
981 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
982 int Addr = (int)CN->getZExtValue();
984 // Otherwise, break this down into an LIS + disp.
985 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
987 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
988 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
989 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
994 Disp = DAG.getTargetConstant(0, getPointerTy());
995 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
996 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
999 return true; // [r+0]
1002 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1003 /// represented as an indexed [r+r] operation.
1004 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1006 SelectionDAG &DAG) const {
1007 // Check to see if we can easily represent this as an [r+r] address. This
1008 // will fail if it thinks that the address is more profitably represented as
1009 // reg+imm, e.g. where imm = 0.
1010 if (SelectAddressRegReg(N, Base, Index, DAG))
1013 // If the operand is an addition, always emit this as [r+r], since this is
1014 // better (for code size, and execution, as the memop does the add for free)
1015 // than emitting an explicit add.
1016 if (N.getOpcode() == ISD::ADD) {
1017 Base = N.getOperand(0);
1018 Index = N.getOperand(1);
1022 // Otherwise, do it the hard way, using R0 as the base register.
1023 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1029 /// SelectAddressRegImmShift - Returns true if the address N can be
1030 /// represented by a base register plus a signed 14-bit displacement
1031 /// [r+imm*4]. Suitable for use by STD and friends.
1032 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1034 SelectionDAG &DAG) const {
1035 // FIXME dl should come from the parent load or store, not the address
1036 DebugLoc dl = N.getDebugLoc();
1037 // If this can be more profitably realized as r+r, fail.
1038 if (SelectAddressRegReg(N, Disp, Base, DAG))
1041 if (N.getOpcode() == ISD::ADD) {
1043 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1044 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1045 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1046 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1048 Base = N.getOperand(0);
1050 return true; // [r+i]
1051 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1052 // Match LOAD (ADD (X, Lo(G))).
1053 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1054 && "Cannot handle constant offsets yet!");
1055 Disp = N.getOperand(1).getOperand(0); // The global address.
1056 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1057 Disp.getOpcode() == ISD::TargetConstantPool ||
1058 Disp.getOpcode() == ISD::TargetJumpTable);
1059 Base = N.getOperand(0);
1060 return true; // [&g+r]
1062 } else if (N.getOpcode() == ISD::OR) {
1064 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1065 // If this is an or of disjoint bitfields, we can codegen this as an add
1066 // (for better address arithmetic) if the LHS and RHS of the OR are
1067 // provably disjoint.
1068 APInt LHSKnownZero, LHSKnownOne;
1069 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1070 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1071 // If all of the bits are known zero on the LHS or RHS, the add won't
1073 Base = N.getOperand(0);
1074 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1078 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1079 // Loading from a constant address. Verify low two bits are clear.
1080 if ((CN->getZExtValue() & 3) == 0) {
1081 // If this address fits entirely in a 14-bit sext immediate field, codegen
1084 if (isIntS16Immediate(CN, Imm)) {
1085 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1086 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1087 CN->getValueType(0));
1091 // Fold the low-part of 32-bit absolute addresses into addr mode.
1092 if (CN->getValueType(0) == MVT::i32 ||
1093 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1094 int Addr = (int)CN->getZExtValue();
1096 // Otherwise, break this down into an LIS + disp.
1097 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1098 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1099 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1100 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1106 Disp = DAG.getTargetConstant(0, getPointerTy());
1107 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1108 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1111 return true; // [r+0]
1115 /// getPreIndexedAddressParts - returns true by value, base pointer and
1116 /// offset pointer and addressing mode by reference if the node's address
1117 /// can be legally represented as pre-indexed load / store address.
1118 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1120 ISD::MemIndexedMode &AM,
1121 SelectionDAG &DAG) const {
1122 if (DisablePPCPreinc) return false;
1126 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1127 Ptr = LD->getBasePtr();
1128 VT = LD->getMemoryVT();
1130 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1131 Ptr = ST->getBasePtr();
1132 VT = ST->getMemoryVT();
1136 // PowerPC doesn't have preinc load/store instructions for vectors.
1140 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1145 // LDU/STU use reg+imm*4, others use reg+imm.
1146 if (VT != MVT::i64) {
1148 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1152 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1156 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1157 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1158 // sext i32 to i64 when addr mode is r+i.
1159 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1160 LD->getExtensionType() == ISD::SEXTLOAD &&
1161 isa<ConstantSDNode>(Offset))
1169 //===----------------------------------------------------------------------===//
1170 // LowerOperation implementation
1171 //===----------------------------------------------------------------------===//
1173 /// GetLabelAccessInfo - Return true if we should reference labels using a
1174 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1175 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1176 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1177 HiOpFlags = PPCII::MO_HA16;
1178 LoOpFlags = PPCII::MO_LO16;
1180 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1181 // non-darwin platform. We don't support PIC on other platforms yet.
1182 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1183 TM.getSubtarget<PPCSubtarget>().isDarwin();
1185 HiOpFlags |= PPCII::MO_PIC_FLAG;
1186 LoOpFlags |= PPCII::MO_PIC_FLAG;
1189 // If this is a reference to a global value that requires a non-lazy-ptr, make
1190 // sure that instruction lowering adds it.
1191 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1192 HiOpFlags |= PPCII::MO_NLP_FLAG;
1193 LoOpFlags |= PPCII::MO_NLP_FLAG;
1195 if (GV->hasHiddenVisibility()) {
1196 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1197 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1204 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1205 SelectionDAG &DAG) {
1206 EVT PtrVT = HiPart.getValueType();
1207 SDValue Zero = DAG.getConstant(0, PtrVT);
1208 DebugLoc DL = HiPart.getDebugLoc();
1210 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1211 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1213 // With PIC, the first instruction is actually "GR+hi(&G)".
1215 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1216 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1218 // Generate non-pic code that has direct accesses to the constant pool.
1219 // The address of the global is just (hi(&g)+lo(&g)).
1220 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1223 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1224 SelectionDAG &DAG) const {
1225 EVT PtrVT = Op.getValueType();
1226 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1227 const Constant *C = CP->getConstVal();
1229 // 64-bit SVR4 ABI code is always position-independent.
1230 // The actual address of the GlobalValue is stored in the TOC.
1231 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1232 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1233 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1234 DAG.getRegister(PPC::X2, MVT::i64));
1237 unsigned MOHiFlag, MOLoFlag;
1238 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1240 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1242 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1243 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1246 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1247 EVT PtrVT = Op.getValueType();
1248 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1250 // 64-bit SVR4 ABI code is always position-independent.
1251 // The actual address of the GlobalValue is stored in the TOC.
1252 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1253 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1254 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1255 DAG.getRegister(PPC::X2, MVT::i64));
1258 unsigned MOHiFlag, MOLoFlag;
1259 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1260 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1261 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1262 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1265 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1266 SelectionDAG &DAG) const {
1267 EVT PtrVT = Op.getValueType();
1269 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1271 unsigned MOHiFlag, MOLoFlag;
1272 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1273 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1274 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1275 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1278 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1279 SelectionDAG &DAG) const {
1281 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1282 DebugLoc dl = GA->getDebugLoc();
1283 const GlobalValue *GV = GA->getGlobal();
1284 EVT PtrVT = getPointerTy();
1285 bool is64bit = PPCSubTarget.isPPC64();
1287 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1289 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1290 PPCII::MO_TPREL16_HA);
1291 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1292 PPCII::MO_TPREL16_LO);
1294 if (model != TLSModel::LocalExec)
1295 llvm_unreachable("only local-exec TLS mode supported");
1296 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1297 is64bit ? MVT::i64 : MVT::i32);
1298 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1299 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1302 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1303 SelectionDAG &DAG) const {
1304 EVT PtrVT = Op.getValueType();
1305 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1306 DebugLoc DL = GSDN->getDebugLoc();
1307 const GlobalValue *GV = GSDN->getGlobal();
1309 // 64-bit SVR4 ABI code is always position-independent.
1310 // The actual address of the GlobalValue is stored in the TOC.
1311 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1312 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1313 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1314 DAG.getRegister(PPC::X2, MVT::i64));
1317 unsigned MOHiFlag, MOLoFlag;
1318 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1321 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1323 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1325 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1327 // If the global reference is actually to a non-lazy-pointer, we have to do an
1328 // extra load to get the address of the global.
1329 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1330 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1331 false, false, false, 0);
1335 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1336 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1337 DebugLoc dl = Op.getDebugLoc();
1339 // If we're comparing for equality to zero, expose the fact that this is
1340 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1341 // fold the new nodes.
1342 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1343 if (C->isNullValue() && CC == ISD::SETEQ) {
1344 EVT VT = Op.getOperand(0).getValueType();
1345 SDValue Zext = Op.getOperand(0);
1346 if (VT.bitsLT(MVT::i32)) {
1348 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1350 unsigned Log2b = Log2_32(VT.getSizeInBits());
1351 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1352 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1353 DAG.getConstant(Log2b, MVT::i32));
1354 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1356 // Leave comparisons against 0 and -1 alone for now, since they're usually
1357 // optimized. FIXME: revisit this when we can custom lower all setcc
1359 if (C->isAllOnesValue() || C->isNullValue())
1363 // If we have an integer seteq/setne, turn it into a compare against zero
1364 // by xor'ing the rhs with the lhs, which is faster than setting a
1365 // condition register, reading it back out, and masking the correct bit. The
1366 // normal approach here uses sub to do this instead of xor. Using xor exposes
1367 // the result to other bit-twiddling opportunities.
1368 EVT LHSVT = Op.getOperand(0).getValueType();
1369 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1370 EVT VT = Op.getValueType();
1371 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1373 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1378 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1379 const PPCSubtarget &Subtarget) const {
1380 SDNode *Node = Op.getNode();
1381 EVT VT = Node->getValueType(0);
1382 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1383 SDValue InChain = Node->getOperand(0);
1384 SDValue VAListPtr = Node->getOperand(1);
1385 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1386 DebugLoc dl = Node->getDebugLoc();
1388 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1391 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1392 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1394 InChain = GprIndex.getValue(1);
1396 if (VT == MVT::i64) {
1397 // Check if GprIndex is even
1398 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1399 DAG.getConstant(1, MVT::i32));
1400 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1401 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1402 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1403 DAG.getConstant(1, MVT::i32));
1404 // Align GprIndex to be even if it isn't
1405 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1409 // fpr index is 1 byte after gpr
1410 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1411 DAG.getConstant(1, MVT::i32));
1414 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1415 FprPtr, MachinePointerInfo(SV), MVT::i8,
1417 InChain = FprIndex.getValue(1);
1419 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1420 DAG.getConstant(8, MVT::i32));
1422 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1423 DAG.getConstant(4, MVT::i32));
1426 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1427 MachinePointerInfo(), false, false,
1429 InChain = OverflowArea.getValue(1);
1431 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1432 MachinePointerInfo(), false, false,
1434 InChain = RegSaveArea.getValue(1);
1436 // select overflow_area if index > 8
1437 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1438 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1440 // adjustment constant gpr_index * 4/8
1441 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1442 VT.isInteger() ? GprIndex : FprIndex,
1443 DAG.getConstant(VT.isInteger() ? 4 : 8,
1446 // OurReg = RegSaveArea + RegConstant
1447 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1450 // Floating types are 32 bytes into RegSaveArea
1451 if (VT.isFloatingPoint())
1452 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1453 DAG.getConstant(32, MVT::i32));
1455 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1456 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1457 VT.isInteger() ? GprIndex : FprIndex,
1458 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1461 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1462 VT.isInteger() ? VAListPtr : FprPtr,
1463 MachinePointerInfo(SV),
1464 MVT::i8, false, false, 0);
1466 // determine if we should load from reg_save_area or overflow_area
1467 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1469 // increase overflow_area by 4/8 if gpr/fpr > 8
1470 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1471 DAG.getConstant(VT.isInteger() ? 4 : 8,
1474 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1477 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1479 MachinePointerInfo(),
1480 MVT::i32, false, false, 0);
1482 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1483 false, false, false, 0);
1486 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1487 SelectionDAG &DAG) const {
1488 return Op.getOperand(0);
1491 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1492 SelectionDAG &DAG) const {
1493 SDValue Chain = Op.getOperand(0);
1494 SDValue Trmp = Op.getOperand(1); // trampoline
1495 SDValue FPtr = Op.getOperand(2); // nested function
1496 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1497 DebugLoc dl = Op.getDebugLoc();
1499 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1500 bool isPPC64 = (PtrVT == MVT::i64);
1502 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1505 TargetLowering::ArgListTy Args;
1506 TargetLowering::ArgListEntry Entry;
1508 Entry.Ty = IntPtrTy;
1509 Entry.Node = Trmp; Args.push_back(Entry);
1511 // TrampSize == (isPPC64 ? 48 : 40);
1512 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1513 isPPC64 ? MVT::i64 : MVT::i32);
1514 Args.push_back(Entry);
1516 Entry.Node = FPtr; Args.push_back(Entry);
1517 Entry.Node = Nest; Args.push_back(Entry);
1519 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1520 TargetLowering::CallLoweringInfo CLI(Chain,
1521 Type::getVoidTy(*DAG.getContext()),
1522 false, false, false, false, 0,
1524 /*isTailCall=*/false,
1525 /*doesNotRet=*/false,
1526 /*isReturnValueUsed=*/true,
1527 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1529 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1531 return CallResult.second;
1534 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1535 const PPCSubtarget &Subtarget) const {
1536 MachineFunction &MF = DAG.getMachineFunction();
1537 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1539 DebugLoc dl = Op.getDebugLoc();
1541 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1542 // vastart just stores the address of the VarArgsFrameIndex slot into the
1543 // memory location argument.
1544 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1545 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1546 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1547 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1548 MachinePointerInfo(SV),
1552 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1553 // We suppose the given va_list is already allocated.
1556 // char gpr; /* index into the array of 8 GPRs
1557 // * stored in the register save area
1558 // * gpr=0 corresponds to r3,
1559 // * gpr=1 to r4, etc.
1561 // char fpr; /* index into the array of 8 FPRs
1562 // * stored in the register save area
1563 // * fpr=0 corresponds to f1,
1564 // * fpr=1 to f2, etc.
1566 // char *overflow_arg_area;
1567 // /* location on stack that holds
1568 // * the next overflow argument
1570 // char *reg_save_area;
1571 // /* where r3:r10 and f1:f8 (if saved)
1577 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1578 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1581 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1583 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1585 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1588 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1589 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1591 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1592 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1594 uint64_t FPROffset = 1;
1595 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1597 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1599 // Store first byte : number of int regs
1600 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1602 MachinePointerInfo(SV),
1603 MVT::i8, false, false, 0);
1604 uint64_t nextOffset = FPROffset;
1605 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1608 // Store second byte : number of float regs
1609 SDValue secondStore =
1610 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1611 MachinePointerInfo(SV, nextOffset), MVT::i8,
1613 nextOffset += StackOffset;
1614 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1616 // Store second word : arguments given on stack
1617 SDValue thirdStore =
1618 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1619 MachinePointerInfo(SV, nextOffset),
1621 nextOffset += FrameOffset;
1622 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1624 // Store third word : arguments given in registers
1625 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1626 MachinePointerInfo(SV, nextOffset),
1631 #include "PPCGenCallingConv.inc"
1633 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1634 CCValAssign::LocInfo &LocInfo,
1635 ISD::ArgFlagsTy &ArgFlags,
1640 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1642 CCValAssign::LocInfo &LocInfo,
1643 ISD::ArgFlagsTy &ArgFlags,
1645 static const uint16_t ArgRegs[] = {
1646 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1647 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1649 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1651 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1653 // Skip one register if the first unallocated register has an even register
1654 // number and there are still argument registers available which have not been
1655 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1656 // need to skip a register if RegNum is odd.
1657 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1658 State.AllocateReg(ArgRegs[RegNum]);
1661 // Always return false here, as this function only makes sure that the first
1662 // unallocated register has an odd register number and does not actually
1663 // allocate a register for the current argument.
1667 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1669 CCValAssign::LocInfo &LocInfo,
1670 ISD::ArgFlagsTy &ArgFlags,
1672 static const uint16_t ArgRegs[] = {
1673 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1677 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1679 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1681 // If there is only one Floating-point register left we need to put both f64
1682 // values of a split ppc_fp128 value on the stack.
1683 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1684 State.AllocateReg(ArgRegs[RegNum]);
1687 // Always return false here, as this function only makes sure that the two f64
1688 // values a ppc_fp128 value is split into are both passed in registers or both
1689 // passed on the stack and does not actually allocate a register for the
1690 // current argument.
1694 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1696 static const uint16_t *GetFPR() {
1697 static const uint16_t FPR[] = {
1698 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1699 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1705 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1707 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1708 unsigned PtrByteSize) {
1709 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1710 if (Flags.isByVal())
1711 ArgSize = Flags.getByValSize();
1712 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1718 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1719 CallingConv::ID CallConv, bool isVarArg,
1720 const SmallVectorImpl<ISD::InputArg>
1722 DebugLoc dl, SelectionDAG &DAG,
1723 SmallVectorImpl<SDValue> &InVals)
1725 if (PPCSubTarget.isSVR4ABI()) {
1726 if (PPCSubTarget.isPPC64())
1727 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1730 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1733 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1739 PPCTargetLowering::LowerFormalArguments_32SVR4(
1741 CallingConv::ID CallConv, bool isVarArg,
1742 const SmallVectorImpl<ISD::InputArg>
1744 DebugLoc dl, SelectionDAG &DAG,
1745 SmallVectorImpl<SDValue> &InVals) const {
1747 // 32-bit SVR4 ABI Stack Frame Layout:
1748 // +-----------------------------------+
1749 // +--> | Back chain |
1750 // | +-----------------------------------+
1751 // | | Floating-point register save area |
1752 // | +-----------------------------------+
1753 // | | General register save area |
1754 // | +-----------------------------------+
1755 // | | CR save word |
1756 // | +-----------------------------------+
1757 // | | VRSAVE save word |
1758 // | +-----------------------------------+
1759 // | | Alignment padding |
1760 // | +-----------------------------------+
1761 // | | Vector register save area |
1762 // | +-----------------------------------+
1763 // | | Local variable space |
1764 // | +-----------------------------------+
1765 // | | Parameter list area |
1766 // | +-----------------------------------+
1767 // | | LR save word |
1768 // | +-----------------------------------+
1769 // SP--> +--- | Back chain |
1770 // +-----------------------------------+
1773 // System V Application Binary Interface PowerPC Processor Supplement
1774 // AltiVec Technology Programming Interface Manual
1776 MachineFunction &MF = DAG.getMachineFunction();
1777 MachineFrameInfo *MFI = MF.getFrameInfo();
1778 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1780 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1781 // Potential tail calls could cause overwriting of argument stack slots.
1782 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1783 (CallConv == CallingConv::Fast));
1784 unsigned PtrByteSize = 4;
1786 // Assign locations to all of the incoming arguments.
1787 SmallVector<CCValAssign, 16> ArgLocs;
1788 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1789 getTargetMachine(), ArgLocs, *DAG.getContext());
1791 // Reserve space for the linkage area on the stack.
1792 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1794 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1796 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1797 CCValAssign &VA = ArgLocs[i];
1799 // Arguments stored in registers.
1800 if (VA.isRegLoc()) {
1801 const TargetRegisterClass *RC;
1802 EVT ValVT = VA.getValVT();
1804 switch (ValVT.getSimpleVT().SimpleTy) {
1806 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1808 RC = &PPC::GPRCRegClass;
1811 RC = &PPC::F4RCRegClass;
1814 RC = &PPC::F8RCRegClass;
1820 RC = &PPC::VRRCRegClass;
1824 // Transform the arguments stored in physical registers into virtual ones.
1825 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1826 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1828 InVals.push_back(ArgValue);
1830 // Argument stored in memory.
1831 assert(VA.isMemLoc());
1833 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1834 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1837 // Create load nodes to retrieve arguments from the stack.
1838 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1839 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1840 MachinePointerInfo(),
1841 false, false, false, 0));
1845 // Assign locations to all of the incoming aggregate by value arguments.
1846 // Aggregates passed by value are stored in the local variable space of the
1847 // caller's stack frame, right above the parameter list area.
1848 SmallVector<CCValAssign, 16> ByValArgLocs;
1849 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1850 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1852 // Reserve stack space for the allocations in CCInfo.
1853 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1855 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1857 // Area that is at least reserved in the caller of this function.
1858 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1860 // Set the size that is at least reserved in caller of this function. Tail
1861 // call optimized function's reserved stack space needs to be aligned so that
1862 // taking the difference between two stack areas will result in an aligned
1864 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1867 std::max(MinReservedArea,
1868 PPCFrameLowering::getMinCallFrameSize(false, false));
1870 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1871 getStackAlignment();
1872 unsigned AlignMask = TargetAlign-1;
1873 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1875 FI->setMinReservedArea(MinReservedArea);
1877 SmallVector<SDValue, 8> MemOps;
1879 // If the function takes variable number of arguments, make a frame index for
1880 // the start of the first vararg value... for expansion of llvm.va_start.
1882 static const uint16_t GPArgRegs[] = {
1883 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1884 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1886 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1888 static const uint16_t FPArgRegs[] = {
1889 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1892 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1894 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1896 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1899 // Make room for NumGPArgRegs and NumFPArgRegs.
1900 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1901 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1903 FuncInfo->setVarArgsStackOffset(
1904 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1905 CCInfo.getNextStackOffset(), true));
1907 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1908 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1910 // The fixed integer arguments of a variadic function are stored to the
1911 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1912 // the result of va_next.
1913 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1914 // Get an existing live-in vreg, or add a new one.
1915 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1917 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1919 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1920 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1921 MachinePointerInfo(), false, false, 0);
1922 MemOps.push_back(Store);
1923 // Increment the address by four for the next argument to store
1924 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1925 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1928 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1930 // The double arguments are stored to the VarArgsFrameIndex
1932 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1933 // Get an existing live-in vreg, or add a new one.
1934 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1936 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1938 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1939 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1940 MachinePointerInfo(), false, false, 0);
1941 MemOps.push_back(Store);
1942 // Increment the address by eight for the next argument to store
1943 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1945 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1949 if (!MemOps.empty())
1950 Chain = DAG.getNode(ISD::TokenFactor, dl,
1951 MVT::Other, &MemOps[0], MemOps.size());
1956 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1957 // value to MVT::i64 and then truncate to the correct register size.
1959 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
1960 SelectionDAG &DAG, SDValue ArgVal,
1961 DebugLoc dl) const {
1963 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1964 DAG.getValueType(ObjectVT));
1965 else if (Flags.isZExt())
1966 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1967 DAG.getValueType(ObjectVT));
1969 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1972 // Set the size that is at least reserved in caller of this function. Tail
1973 // call optimized functions' reserved stack space needs to be aligned so that
1974 // taking the difference between two stack areas will result in an aligned
1977 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
1978 unsigned nAltivecParamsAtEnd,
1979 unsigned MinReservedArea,
1980 bool isPPC64) const {
1981 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1982 // Add the Altivec parameters at the end, if needed.
1983 if (nAltivecParamsAtEnd) {
1984 MinReservedArea = ((MinReservedArea+15)/16)*16;
1985 MinReservedArea += 16*nAltivecParamsAtEnd;
1988 std::max(MinReservedArea,
1989 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
1990 unsigned TargetAlign
1991 = DAG.getMachineFunction().getTarget().getFrameLowering()->
1992 getStackAlignment();
1993 unsigned AlignMask = TargetAlign-1;
1994 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1995 FI->setMinReservedArea(MinReservedArea);
1999 PPCTargetLowering::LowerFormalArguments_64SVR4(
2001 CallingConv::ID CallConv, bool isVarArg,
2002 const SmallVectorImpl<ISD::InputArg>
2004 DebugLoc dl, SelectionDAG &DAG,
2005 SmallVectorImpl<SDValue> &InVals) const {
2006 // TODO: add description of PPC stack frame format, or at least some docs.
2008 MachineFunction &MF = DAG.getMachineFunction();
2009 MachineFrameInfo *MFI = MF.getFrameInfo();
2010 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2012 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2013 // Potential tail calls could cause overwriting of argument stack slots.
2014 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2015 (CallConv == CallingConv::Fast));
2016 unsigned PtrByteSize = 8;
2018 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2019 // Area that is at least reserved in caller of this function.
2020 unsigned MinReservedArea = ArgOffset;
2022 static const uint16_t GPR[] = {
2023 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2024 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2027 static const uint16_t *FPR = GetFPR();
2029 static const uint16_t VR[] = {
2030 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2031 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2034 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2035 const unsigned Num_FPR_Regs = 13;
2036 const unsigned Num_VR_Regs = array_lengthof(VR);
2038 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2040 // Add DAG nodes to load the arguments or copy them out of registers. On
2041 // entry to a function on PPC, the arguments start after the linkage area,
2042 // although the first ones are often in registers.
2044 SmallVector<SDValue, 8> MemOps;
2045 unsigned nAltivecParamsAtEnd = 0;
2046 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2047 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2049 bool needsLoad = false;
2050 EVT ObjectVT = Ins[ArgNo].VT;
2051 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2052 unsigned ArgSize = ObjSize;
2053 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2055 unsigned CurArgOffset = ArgOffset;
2057 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2058 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2059 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2061 MinReservedArea = ((MinReservedArea+15)/16)*16;
2062 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2066 nAltivecParamsAtEnd++;
2068 // Calculate min reserved area.
2069 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2073 // FIXME the codegen can be much improved in some cases.
2074 // We do not have to keep everything in memory.
2075 if (Flags.isByVal()) {
2076 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2077 ObjSize = Flags.getByValSize();
2078 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2079 // All aggregates smaller than 8 bytes must be passed right-justified.
2080 if (ObjSize < PtrByteSize)
2081 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2082 // The value of the object is its address.
2083 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2084 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2085 InVals.push_back(FIN);
2086 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2087 if (GPR_idx != Num_GPR_Regs) {
2089 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2090 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2091 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2092 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2093 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2094 MachinePointerInfo(FuncArg,
2096 ObjType, false, false, 0);
2097 MemOps.push_back(Store);
2101 ArgOffset += PtrByteSize;
2105 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2106 // Store whatever pieces of the object are in registers
2107 // to memory. ArgOffset will be the address of the beginning
2109 if (GPR_idx != Num_GPR_Regs) {
2111 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2112 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2113 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2114 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2115 SDValue Shifted = Val;
2117 // For 64-bit SVR4, small structs come in right-adjusted.
2118 // Shift them left so the following logic works as expected.
2120 SDValue ShiftAmt = DAG.getConstant(64 - 8 * ObjSize, PtrVT);
2121 Shifted = DAG.getNode(ISD::SHL, dl, PtrVT, Val, ShiftAmt);
2124 SDValue Store = DAG.getStore(Val.getValue(1), dl, Shifted, FIN,
2125 MachinePointerInfo(FuncArg, ArgOffset),
2127 MemOps.push_back(Store);
2129 ArgOffset += PtrByteSize;
2131 ArgOffset += ArgSize - j;
2138 switch (ObjectVT.getSimpleVT().SimpleTy) {
2139 default: llvm_unreachable("Unhandled argument type!");
2142 if (GPR_idx != Num_GPR_Regs) {
2143 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2144 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2146 if (ObjectVT == MVT::i32)
2147 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2148 // value to MVT::i64 and then truncate to the correct register size.
2149 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2154 ArgSize = PtrByteSize;
2161 // Every 8 bytes of argument space consumes one of the GPRs available for
2162 // argument passing.
2163 if (GPR_idx != Num_GPR_Regs) {
2166 if (FPR_idx != Num_FPR_Regs) {
2169 if (ObjectVT == MVT::f32)
2170 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2172 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2174 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2178 ArgSize = PtrByteSize;
2187 // Note that vector arguments in registers don't reserve stack space,
2188 // except in varargs functions.
2189 if (VR_idx != Num_VR_Regs) {
2190 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2191 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2193 while ((ArgOffset % 16) != 0) {
2194 ArgOffset += PtrByteSize;
2195 if (GPR_idx != Num_GPR_Regs)
2199 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2203 // Vectors are aligned.
2204 ArgOffset = ((ArgOffset+15)/16)*16;
2205 CurArgOffset = ArgOffset;
2212 // We need to load the argument to a virtual register if we determined
2213 // above that we ran out of physical registers of the appropriate type.
2215 int FI = MFI->CreateFixedObject(ObjSize,
2216 CurArgOffset + (ArgSize - ObjSize),
2218 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2219 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2220 false, false, false, 0);
2223 InVals.push_back(ArgVal);
2226 // Set the size that is at least reserved in caller of this function. Tail
2227 // call optimized functions' reserved stack space needs to be aligned so that
2228 // taking the difference between two stack areas will result in an aligned
2230 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2232 // If the function takes variable number of arguments, make a frame index for
2233 // the start of the first vararg value... for expansion of llvm.va_start.
2235 int Depth = ArgOffset;
2237 FuncInfo->setVarArgsFrameIndex(
2238 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2239 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2241 // If this function is vararg, store any remaining integer argument regs
2242 // to their spots on the stack so that they may be loaded by deferencing the
2243 // result of va_next.
2244 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2245 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2246 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2247 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2248 MachinePointerInfo(), false, false, 0);
2249 MemOps.push_back(Store);
2250 // Increment the address by four for the next argument to store
2251 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2252 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2256 if (!MemOps.empty())
2257 Chain = DAG.getNode(ISD::TokenFactor, dl,
2258 MVT::Other, &MemOps[0], MemOps.size());
2264 PPCTargetLowering::LowerFormalArguments_Darwin(
2266 CallingConv::ID CallConv, bool isVarArg,
2267 const SmallVectorImpl<ISD::InputArg>
2269 DebugLoc dl, SelectionDAG &DAG,
2270 SmallVectorImpl<SDValue> &InVals) const {
2271 // TODO: add description of PPC stack frame format, or at least some docs.
2273 MachineFunction &MF = DAG.getMachineFunction();
2274 MachineFrameInfo *MFI = MF.getFrameInfo();
2275 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2277 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2278 bool isPPC64 = PtrVT == MVT::i64;
2279 // Potential tail calls could cause overwriting of argument stack slots.
2280 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2281 (CallConv == CallingConv::Fast));
2282 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2284 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2285 // Area that is at least reserved in caller of this function.
2286 unsigned MinReservedArea = ArgOffset;
2288 static const uint16_t GPR_32[] = { // 32-bit registers.
2289 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2290 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2292 static const uint16_t GPR_64[] = { // 64-bit registers.
2293 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2294 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2297 static const uint16_t *FPR = GetFPR();
2299 static const uint16_t VR[] = {
2300 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2301 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2304 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2305 const unsigned Num_FPR_Regs = 13;
2306 const unsigned Num_VR_Regs = array_lengthof( VR);
2308 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2310 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2312 // In 32-bit non-varargs functions, the stack space for vectors is after the
2313 // stack space for non-vectors. We do not use this space unless we have
2314 // too many vectors to fit in registers, something that only occurs in
2315 // constructed examples:), but we have to walk the arglist to figure
2316 // that out...for the pathological case, compute VecArgOffset as the
2317 // start of the vector parameter area. Computing VecArgOffset is the
2318 // entire point of the following loop.
2319 unsigned VecArgOffset = ArgOffset;
2320 if (!isVarArg && !isPPC64) {
2321 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2323 EVT ObjectVT = Ins[ArgNo].VT;
2324 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2326 if (Flags.isByVal()) {
2327 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2328 unsigned ObjSize = Flags.getByValSize();
2330 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2331 VecArgOffset += ArgSize;
2335 switch(ObjectVT.getSimpleVT().SimpleTy) {
2336 default: llvm_unreachable("Unhandled argument type!");
2341 case MVT::i64: // PPC64
2343 // FIXME: We are guaranteed to be !isPPC64 at this point.
2344 // Does MVT::i64 apply?
2351 // Nothing to do, we're only looking at Nonvector args here.
2356 // We've found where the vector parameter area in memory is. Skip the
2357 // first 12 parameters; these don't use that memory.
2358 VecArgOffset = ((VecArgOffset+15)/16)*16;
2359 VecArgOffset += 12*16;
2361 // Add DAG nodes to load the arguments or copy them out of registers. On
2362 // entry to a function on PPC, the arguments start after the linkage area,
2363 // although the first ones are often in registers.
2365 SmallVector<SDValue, 8> MemOps;
2366 unsigned nAltivecParamsAtEnd = 0;
2367 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2368 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2370 bool needsLoad = false;
2371 EVT ObjectVT = Ins[ArgNo].VT;
2372 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2373 unsigned ArgSize = ObjSize;
2374 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2376 unsigned CurArgOffset = ArgOffset;
2378 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2379 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2380 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2381 if (isVarArg || isPPC64) {
2382 MinReservedArea = ((MinReservedArea+15)/16)*16;
2383 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2386 } else nAltivecParamsAtEnd++;
2388 // Calculate min reserved area.
2389 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2393 // FIXME the codegen can be much improved in some cases.
2394 // We do not have to keep everything in memory.
2395 if (Flags.isByVal()) {
2396 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2397 ObjSize = Flags.getByValSize();
2398 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2399 // Objects of size 1 and 2 are right justified, everything else is
2400 // left justified. This means the memory address is adjusted forwards.
2401 if (ObjSize==1 || ObjSize==2) {
2402 CurArgOffset = CurArgOffset + (4 - ObjSize);
2404 // The value of the object is its address.
2405 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2406 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2407 InVals.push_back(FIN);
2408 if (ObjSize==1 || ObjSize==2) {
2409 if (GPR_idx != Num_GPR_Regs) {
2412 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2414 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2415 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2416 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2417 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2418 MachinePointerInfo(FuncArg,
2420 ObjType, false, false, 0);
2421 MemOps.push_back(Store);
2425 ArgOffset += PtrByteSize;
2429 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2430 // Store whatever pieces of the object are in registers
2431 // to memory. ArgOffset will be the address of the beginning
2433 if (GPR_idx != Num_GPR_Regs) {
2436 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2438 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2439 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2440 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2441 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2442 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2443 MachinePointerInfo(FuncArg, ArgOffset),
2445 MemOps.push_back(Store);
2447 ArgOffset += PtrByteSize;
2449 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2456 switch (ObjectVT.getSimpleVT().SimpleTy) {
2457 default: llvm_unreachable("Unhandled argument type!");
2460 if (GPR_idx != Num_GPR_Regs) {
2461 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2462 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2466 ArgSize = PtrByteSize;
2468 // All int arguments reserve stack space in the Darwin ABI.
2469 ArgOffset += PtrByteSize;
2473 case MVT::i64: // PPC64
2474 if (GPR_idx != Num_GPR_Regs) {
2475 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2476 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2478 if (ObjectVT == MVT::i32)
2479 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2480 // value to MVT::i64 and then truncate to the correct register size.
2481 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2486 ArgSize = PtrByteSize;
2488 // All int arguments reserve stack space in the Darwin ABI.
2494 // Every 4 bytes of argument space consumes one of the GPRs available for
2495 // argument passing.
2496 if (GPR_idx != Num_GPR_Regs) {
2498 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2501 if (FPR_idx != Num_FPR_Regs) {
2504 if (ObjectVT == MVT::f32)
2505 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2507 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2509 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2515 // All FP arguments reserve stack space in the Darwin ABI.
2516 ArgOffset += isPPC64 ? 8 : ObjSize;
2522 // Note that vector arguments in registers don't reserve stack space,
2523 // except in varargs functions.
2524 if (VR_idx != Num_VR_Regs) {
2525 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2526 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2528 while ((ArgOffset % 16) != 0) {
2529 ArgOffset += PtrByteSize;
2530 if (GPR_idx != Num_GPR_Regs)
2534 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2538 if (!isVarArg && !isPPC64) {
2539 // Vectors go after all the nonvectors.
2540 CurArgOffset = VecArgOffset;
2543 // Vectors are aligned.
2544 ArgOffset = ((ArgOffset+15)/16)*16;
2545 CurArgOffset = ArgOffset;
2553 // We need to load the argument to a virtual register if we determined above
2554 // that we ran out of physical registers of the appropriate type.
2556 int FI = MFI->CreateFixedObject(ObjSize,
2557 CurArgOffset + (ArgSize - ObjSize),
2559 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2560 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2561 false, false, false, 0);
2564 InVals.push_back(ArgVal);
2567 // Set the size that is at least reserved in caller of this function. Tail
2568 // call optimized functions' reserved stack space needs to be aligned so that
2569 // taking the difference between two stack areas will result in an aligned
2571 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2573 // If the function takes variable number of arguments, make a frame index for
2574 // the start of the first vararg value... for expansion of llvm.va_start.
2576 int Depth = ArgOffset;
2578 FuncInfo->setVarArgsFrameIndex(
2579 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2581 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2583 // If this function is vararg, store any remaining integer argument regs
2584 // to their spots on the stack so that they may be loaded by deferencing the
2585 // result of va_next.
2586 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2590 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2592 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2594 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2595 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2596 MachinePointerInfo(), false, false, 0);
2597 MemOps.push_back(Store);
2598 // Increment the address by four for the next argument to store
2599 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2600 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2604 if (!MemOps.empty())
2605 Chain = DAG.getNode(ISD::TokenFactor, dl,
2606 MVT::Other, &MemOps[0], MemOps.size());
2611 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2612 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2614 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2618 const SmallVectorImpl<ISD::OutputArg>
2620 const SmallVectorImpl<SDValue> &OutVals,
2621 unsigned &nAltivecParamsAtEnd) {
2622 // Count how many bytes are to be pushed on the stack, including the linkage
2623 // area, and parameter passing area. We start with 24/48 bytes, which is
2624 // prereserved space for [SP][CR][LR][3 x unused].
2625 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2626 unsigned NumOps = Outs.size();
2627 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2629 // Add up all the space actually used.
2630 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2631 // they all go in registers, but we must reserve stack space for them for
2632 // possible use by the caller. In varargs or 64-bit calls, parameters are
2633 // assigned stack space in order, with padding so Altivec parameters are
2635 nAltivecParamsAtEnd = 0;
2636 for (unsigned i = 0; i != NumOps; ++i) {
2637 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2638 EVT ArgVT = Outs[i].VT;
2639 // Varargs Altivec parameters are padded to a 16 byte boundary.
2640 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2641 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2642 if (!isVarArg && !isPPC64) {
2643 // Non-varargs Altivec parameters go after all the non-Altivec
2644 // parameters; handle those later so we know how much padding we need.
2645 nAltivecParamsAtEnd++;
2648 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2649 NumBytes = ((NumBytes+15)/16)*16;
2651 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2654 // Allow for Altivec parameters at the end, if needed.
2655 if (nAltivecParamsAtEnd) {
2656 NumBytes = ((NumBytes+15)/16)*16;
2657 NumBytes += 16*nAltivecParamsAtEnd;
2660 // The prolog code of the callee may store up to 8 GPR argument registers to
2661 // the stack, allowing va_start to index over them in memory if its varargs.
2662 // Because we cannot tell if this is needed on the caller side, we have to
2663 // conservatively assume that it is needed. As such, make sure we have at
2664 // least enough stack space for the caller to store the 8 GPRs.
2665 NumBytes = std::max(NumBytes,
2666 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2668 // Tail call needs the stack to be aligned.
2669 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2670 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2671 getFrameLowering()->getStackAlignment();
2672 unsigned AlignMask = TargetAlign-1;
2673 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2679 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2680 /// adjusted to accommodate the arguments for the tailcall.
2681 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2682 unsigned ParamSize) {
2684 if (!isTailCall) return 0;
2686 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2687 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2688 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2689 // Remember only if the new adjustement is bigger.
2690 if (SPDiff < FI->getTailCallSPDelta())
2691 FI->setTailCallSPDelta(SPDiff);
2696 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2697 /// for tail call optimization. Targets which want to do tail call
2698 /// optimization should implement this function.
2700 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2701 CallingConv::ID CalleeCC,
2703 const SmallVectorImpl<ISD::InputArg> &Ins,
2704 SelectionDAG& DAG) const {
2705 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2708 // Variable argument functions are not supported.
2712 MachineFunction &MF = DAG.getMachineFunction();
2713 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2714 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2715 // Functions containing by val parameters are not supported.
2716 for (unsigned i = 0; i != Ins.size(); i++) {
2717 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2718 if (Flags.isByVal()) return false;
2721 // Non PIC/GOT tail calls are supported.
2722 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2725 // At the moment we can only do local tail calls (in same module, hidden
2726 // or protected) if we are generating PIC.
2727 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2728 return G->getGlobal()->hasHiddenVisibility()
2729 || G->getGlobal()->hasProtectedVisibility();
2735 /// isCallCompatibleAddress - Return the immediate to use if the specified
2736 /// 32-bit value is representable in the immediate field of a BxA instruction.
2737 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2738 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2741 int Addr = C->getZExtValue();
2742 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2743 SignExtend32<26>(Addr) != Addr)
2744 return 0; // Top 6 bits have to be sext of immediate.
2746 return DAG.getConstant((int)C->getZExtValue() >> 2,
2747 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2752 struct TailCallArgumentInfo {
2757 TailCallArgumentInfo() : FrameIdx(0) {}
2762 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2764 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2766 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2767 SmallVector<SDValue, 8> &MemOpChains,
2769 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2770 SDValue Arg = TailCallArgs[i].Arg;
2771 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2772 int FI = TailCallArgs[i].FrameIdx;
2773 // Store relative to framepointer.
2774 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2775 MachinePointerInfo::getFixedStack(FI),
2780 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2781 /// the appropriate stack slot for the tail call optimized function call.
2782 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2783 MachineFunction &MF,
2792 // Calculate the new stack slot for the return address.
2793 int SlotSize = isPPC64 ? 8 : 4;
2794 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2796 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2797 NewRetAddrLoc, true);
2798 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2799 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2800 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2801 MachinePointerInfo::getFixedStack(NewRetAddr),
2804 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2805 // slot as the FP is never overwritten.
2808 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2809 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2811 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2812 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2813 MachinePointerInfo::getFixedStack(NewFPIdx),
2820 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2821 /// the position of the argument.
2823 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2824 SDValue Arg, int SPDiff, unsigned ArgOffset,
2825 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2826 int Offset = ArgOffset + SPDiff;
2827 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2828 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2829 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2830 SDValue FIN = DAG.getFrameIndex(FI, VT);
2831 TailCallArgumentInfo Info;
2833 Info.FrameIdxOp = FIN;
2835 TailCallArguments.push_back(Info);
2838 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2839 /// stack slot. Returns the chain as result and the loaded frame pointers in
2840 /// LROpOut/FPOpout. Used when tail calling.
2841 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2847 DebugLoc dl) const {
2849 // Load the LR and FP stack slot for later adjusting.
2850 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2851 LROpOut = getReturnAddrFrameIndex(DAG);
2852 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2853 false, false, false, 0);
2854 Chain = SDValue(LROpOut.getNode(), 1);
2856 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2857 // slot as the FP is never overwritten.
2859 FPOpOut = getFramePointerFrameIndex(DAG);
2860 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2861 false, false, false, 0);
2862 Chain = SDValue(FPOpOut.getNode(), 1);
2868 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2869 /// by "Src" to address "Dst" of size "Size". Alignment information is
2870 /// specified by the specific parameter attribute. The copy will be passed as
2871 /// a byval function parameter.
2872 /// Sometimes what we are copying is the end of a larger object, the part that
2873 /// does not fit in registers.
2875 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2876 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2878 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2879 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2880 false, false, MachinePointerInfo(0),
2881 MachinePointerInfo(0));
2884 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2887 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2888 SDValue Arg, SDValue PtrOff, int SPDiff,
2889 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2890 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2891 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2893 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2898 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2900 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2901 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2902 DAG.getConstant(ArgOffset, PtrVT));
2904 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2905 MachinePointerInfo(), false, false, 0));
2906 // Calculate and remember argument location.
2907 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2912 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2913 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2914 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2915 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2916 MachineFunction &MF = DAG.getMachineFunction();
2918 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2919 // might overwrite each other in case of tail call optimization.
2920 SmallVector<SDValue, 8> MemOpChains2;
2921 // Do not flag preceding copytoreg stuff together with the following stuff.
2923 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2925 if (!MemOpChains2.empty())
2926 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2927 &MemOpChains2[0], MemOpChains2.size());
2929 // Store the return address to the appropriate stack slot.
2930 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2931 isPPC64, isDarwinABI, dl);
2933 // Emit callseq_end just before tailcall node.
2934 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2935 DAG.getIntPtrConstant(0, true), InFlag);
2936 InFlag = Chain.getValue(1);
2940 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2941 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2942 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2943 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2944 const PPCSubtarget &PPCSubTarget) {
2946 bool isPPC64 = PPCSubTarget.isPPC64();
2947 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2949 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2950 NodeTys.push_back(MVT::Other); // Returns a chain
2951 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2953 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2955 bool needIndirectCall = true;
2956 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2957 // If this is an absolute destination address, use the munged value.
2958 Callee = SDValue(Dest, 0);
2959 needIndirectCall = false;
2962 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2963 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2964 // Use indirect calls for ALL functions calls in JIT mode, since the
2965 // far-call stubs may be outside relocation limits for a BL instruction.
2966 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2967 unsigned OpFlags = 0;
2968 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2969 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2970 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2971 (G->getGlobal()->isDeclaration() ||
2972 G->getGlobal()->isWeakForLinker())) {
2973 // PC-relative references to external symbols should go through $stub,
2974 // unless we're building with the leopard linker or later, which
2975 // automatically synthesizes these stubs.
2976 OpFlags = PPCII::MO_DARWIN_STUB;
2979 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2980 // every direct call is) turn it into a TargetGlobalAddress /
2981 // TargetExternalSymbol node so that legalize doesn't hack it.
2982 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2983 Callee.getValueType(),
2985 needIndirectCall = false;
2989 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2990 unsigned char OpFlags = 0;
2992 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2993 (PPCSubTarget.getTargetTriple().isMacOSX() &&
2994 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2995 // PC-relative references to external symbols should go through $stub,
2996 // unless we're building with the leopard linker or later, which
2997 // automatically synthesizes these stubs.
2998 OpFlags = PPCII::MO_DARWIN_STUB;
3001 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3003 needIndirectCall = false;
3006 if (needIndirectCall) {
3007 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3008 // to do the call, we can't use PPCISD::CALL.
3009 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3011 if (isSVR4ABI && isPPC64) {
3012 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3013 // entry point, but to the function descriptor (the function entry point
3014 // address is part of the function descriptor though).
3015 // The function descriptor is a three doubleword structure with the
3016 // following fields: function entry point, TOC base address and
3017 // environment pointer.
3018 // Thus for a call through a function pointer, the following actions need
3020 // 1. Save the TOC of the caller in the TOC save area of its stack
3021 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3022 // 2. Load the address of the function entry point from the function
3024 // 3. Load the TOC of the callee from the function descriptor into r2.
3025 // 4. Load the environment pointer from the function descriptor into
3027 // 5. Branch to the function entry point address.
3028 // 6. On return of the callee, the TOC of the caller needs to be
3029 // restored (this is done in FinishCall()).
3031 // All those operations are flagged together to ensure that no other
3032 // operations can be scheduled in between. E.g. without flagging the
3033 // operations together, a TOC access in the caller could be scheduled
3034 // between the load of the callee TOC and the branch to the callee, which
3035 // results in the TOC access going through the TOC of the callee instead
3036 // of going through the TOC of the caller, which leads to incorrect code.
3038 // Load the address of the function entry point from the function
3040 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3041 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3042 InFlag.getNode() ? 3 : 2);
3043 Chain = LoadFuncPtr.getValue(1);
3044 InFlag = LoadFuncPtr.getValue(2);
3046 // Load environment pointer into r11.
3047 // Offset of the environment pointer within the function descriptor.
3048 SDValue PtrOff = DAG.getIntPtrConstant(16);
3050 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3051 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3053 Chain = LoadEnvPtr.getValue(1);
3054 InFlag = LoadEnvPtr.getValue(2);
3056 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3058 Chain = EnvVal.getValue(0);
3059 InFlag = EnvVal.getValue(1);
3061 // Load TOC of the callee into r2. We are using a target-specific load
3062 // with r2 hard coded, because the result of a target-independent load
3063 // would never go directly into r2, since r2 is a reserved register (which
3064 // prevents the register allocator from allocating it), resulting in an
3065 // additional register being allocated and an unnecessary move instruction
3067 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3068 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3070 Chain = LoadTOCPtr.getValue(0);
3071 InFlag = LoadTOCPtr.getValue(1);
3073 MTCTROps[0] = Chain;
3074 MTCTROps[1] = LoadFuncPtr;
3075 MTCTROps[2] = InFlag;
3078 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3079 2 + (InFlag.getNode() != 0));
3080 InFlag = Chain.getValue(1);
3083 NodeTys.push_back(MVT::Other);
3084 NodeTys.push_back(MVT::Glue);
3085 Ops.push_back(Chain);
3086 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3088 // Add CTR register as callee so a bctr can be emitted later.
3090 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3093 // If this is a direct call, pass the chain and the callee.
3094 if (Callee.getNode()) {
3095 Ops.push_back(Chain);
3096 Ops.push_back(Callee);
3098 // If this is a tail call add stack pointer delta.
3100 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3102 // Add argument registers to the end of the list so that they are known live
3104 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3105 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3106 RegsToPass[i].second.getValueType()));
3112 bool isLocalCall(const SDValue &Callee)
3114 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3115 return !G->getGlobal()->isDeclaration() &&
3116 !G->getGlobal()->isWeakForLinker();
3121 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3122 CallingConv::ID CallConv, bool isVarArg,
3123 const SmallVectorImpl<ISD::InputArg> &Ins,
3124 DebugLoc dl, SelectionDAG &DAG,
3125 SmallVectorImpl<SDValue> &InVals) const {
3127 SmallVector<CCValAssign, 16> RVLocs;
3128 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3129 getTargetMachine(), RVLocs, *DAG.getContext());
3130 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3132 // Copy all of the result registers out of their specified physreg.
3133 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3134 CCValAssign &VA = RVLocs[i];
3135 EVT VT = VA.getValVT();
3136 assert(VA.isRegLoc() && "Can only return in registers!");
3137 Chain = DAG.getCopyFromReg(Chain, dl,
3138 VA.getLocReg(), VT, InFlag).getValue(1);
3139 InVals.push_back(Chain.getValue(0));
3140 InFlag = Chain.getValue(2);
3147 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3148 bool isTailCall, bool isVarArg,
3150 SmallVector<std::pair<unsigned, SDValue>, 8>
3152 SDValue InFlag, SDValue Chain,
3154 int SPDiff, unsigned NumBytes,
3155 const SmallVectorImpl<ISD::InputArg> &Ins,
3156 SmallVectorImpl<SDValue> &InVals) const {
3157 std::vector<EVT> NodeTys;
3158 SmallVector<SDValue, 8> Ops;
3159 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3160 isTailCall, RegsToPass, Ops, NodeTys,
3163 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3164 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3165 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3167 // When performing tail call optimization the callee pops its arguments off
3168 // the stack. Account for this here so these bytes can be pushed back on in
3169 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3170 int BytesCalleePops =
3171 (CallConv == CallingConv::Fast &&
3172 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3174 // Add a register mask operand representing the call-preserved registers.
3175 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3176 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3177 assert(Mask && "Missing call preserved mask for calling convention");
3178 Ops.push_back(DAG.getRegisterMask(Mask));
3180 if (InFlag.getNode())
3181 Ops.push_back(InFlag);
3185 // If this is the first return lowered for this function, add the regs
3186 // to the liveout set for the function.
3187 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3188 SmallVector<CCValAssign, 16> RVLocs;
3189 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3190 getTargetMachine(), RVLocs, *DAG.getContext());
3191 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3192 for (unsigned i = 0; i != RVLocs.size(); ++i)
3193 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3196 assert(((Callee.getOpcode() == ISD::Register &&
3197 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3198 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3199 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3200 isa<ConstantSDNode>(Callee)) &&
3201 "Expecting an global address, external symbol, absolute value or register");
3203 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3206 // Add a NOP immediately after the branch instruction when using the 64-bit
3207 // SVR4 ABI. At link time, if caller and callee are in a different module and
3208 // thus have a different TOC, the call will be replaced with a call to a stub
3209 // function which saves the current TOC, loads the TOC of the callee and
3210 // branches to the callee. The NOP will be replaced with a load instruction
3211 // which restores the TOC of the caller from the TOC save slot of the current
3212 // stack frame. If caller and callee belong to the same module (and have the
3213 // same TOC), the NOP will remain unchanged.
3215 bool needsTOCRestore = false;
3216 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3217 if (CallOpc == PPCISD::BCTRL_SVR4) {
3218 // This is a call through a function pointer.
3219 // Restore the caller TOC from the save area into R2.
3220 // See PrepareCall() for more information about calls through function
3221 // pointers in the 64-bit SVR4 ABI.
3222 // We are using a target-specific load with r2 hard coded, because the
3223 // result of a target-independent load would never go directly into r2,
3224 // since r2 is a reserved register (which prevents the register allocator
3225 // from allocating it), resulting in an additional register being
3226 // allocated and an unnecessary move instruction being generated.
3227 needsTOCRestore = true;
3228 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3229 // Otherwise insert NOP for non-local calls.
3230 CallOpc = PPCISD::CALL_NOP_SVR4;
3234 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3235 InFlag = Chain.getValue(1);
3237 if (needsTOCRestore) {
3238 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3239 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3240 InFlag = Chain.getValue(1);
3243 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3244 DAG.getIntPtrConstant(BytesCalleePops, true),
3247 InFlag = Chain.getValue(1);
3249 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3250 Ins, dl, DAG, InVals);
3254 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3255 SmallVectorImpl<SDValue> &InVals) const {
3256 SelectionDAG &DAG = CLI.DAG;
3257 DebugLoc &dl = CLI.DL;
3258 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3259 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3260 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3261 SDValue Chain = CLI.Chain;
3262 SDValue Callee = CLI.Callee;
3263 bool &isTailCall = CLI.IsTailCall;
3264 CallingConv::ID CallConv = CLI.CallConv;
3265 bool isVarArg = CLI.IsVarArg;
3268 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3271 if (PPCSubTarget.isSVR4ABI()) {
3272 if (PPCSubTarget.isPPC64())
3273 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3274 isTailCall, Outs, OutVals, Ins,
3277 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3278 isTailCall, Outs, OutVals, Ins,
3282 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3283 isTailCall, Outs, OutVals, Ins,
3288 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3289 CallingConv::ID CallConv, bool isVarArg,
3291 const SmallVectorImpl<ISD::OutputArg> &Outs,
3292 const SmallVectorImpl<SDValue> &OutVals,
3293 const SmallVectorImpl<ISD::InputArg> &Ins,
3294 DebugLoc dl, SelectionDAG &DAG,
3295 SmallVectorImpl<SDValue> &InVals) const {
3296 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3297 // of the 32-bit SVR4 ABI stack frame layout.
3299 assert((CallConv == CallingConv::C ||
3300 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3302 unsigned PtrByteSize = 4;
3304 MachineFunction &MF = DAG.getMachineFunction();
3306 // Mark this function as potentially containing a function that contains a
3307 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3308 // and restoring the callers stack pointer in this functions epilog. This is
3309 // done because by tail calling the called function might overwrite the value
3310 // in this function's (MF) stack pointer stack slot 0(SP).
3311 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3312 CallConv == CallingConv::Fast)
3313 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3315 // Count how many bytes are to be pushed on the stack, including the linkage
3316 // area, parameter list area and the part of the local variable space which
3317 // contains copies of aggregates which are passed by value.
3319 // Assign locations to all of the outgoing arguments.
3320 SmallVector<CCValAssign, 16> ArgLocs;
3321 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3322 getTargetMachine(), ArgLocs, *DAG.getContext());
3324 // Reserve space for the linkage area on the stack.
3325 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3328 // Handle fixed and variable vector arguments differently.
3329 // Fixed vector arguments go into registers as long as registers are
3330 // available. Variable vector arguments always go into memory.
3331 unsigned NumArgs = Outs.size();
3333 for (unsigned i = 0; i != NumArgs; ++i) {
3334 MVT ArgVT = Outs[i].VT;
3335 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3338 if (Outs[i].IsFixed) {
3339 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3342 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3348 errs() << "Call operand #" << i << " has unhandled type "
3349 << EVT(ArgVT).getEVTString() << "\n";
3351 llvm_unreachable(0);
3355 // All arguments are treated the same.
3356 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3359 // Assign locations to all of the outgoing aggregate by value arguments.
3360 SmallVector<CCValAssign, 16> ByValArgLocs;
3361 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3362 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3364 // Reserve stack space for the allocations in CCInfo.
3365 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3367 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3369 // Size of the linkage area, parameter list area and the part of the local
3370 // space variable where copies of aggregates which are passed by value are
3372 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3374 // Calculate by how many bytes the stack has to be adjusted in case of tail
3375 // call optimization.
3376 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3378 // Adjust the stack pointer for the new arguments...
3379 // These operations are automatically eliminated by the prolog/epilog pass
3380 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3381 SDValue CallSeqStart = Chain;
3383 // Load the return address and frame pointer so it can be moved somewhere else
3386 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3389 // Set up a copy of the stack pointer for use loading and storing any
3390 // arguments that may not fit in the registers available for argument
3392 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3394 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3395 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3396 SmallVector<SDValue, 8> MemOpChains;
3398 bool seenFloatArg = false;
3399 // Walk the register/memloc assignments, inserting copies/loads.
3400 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3403 CCValAssign &VA = ArgLocs[i];
3404 SDValue Arg = OutVals[i];
3405 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3407 if (Flags.isByVal()) {
3408 // Argument is an aggregate which is passed by value, thus we need to
3409 // create a copy of it in the local variable space of the current stack
3410 // frame (which is the stack frame of the caller) and pass the address of
3411 // this copy to the callee.
3412 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3413 CCValAssign &ByValVA = ByValArgLocs[j++];
3414 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3416 // Memory reserved in the local variable space of the callers stack frame.
3417 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3419 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3420 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3422 // Create a copy of the argument in the local area of the current
3424 SDValue MemcpyCall =
3425 CreateCopyOfByValArgument(Arg, PtrOff,
3426 CallSeqStart.getNode()->getOperand(0),
3429 // This must go outside the CALLSEQ_START..END.
3430 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3431 CallSeqStart.getNode()->getOperand(1));
3432 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3433 NewCallSeqStart.getNode());
3434 Chain = CallSeqStart = NewCallSeqStart;
3436 // Pass the address of the aggregate copy on the stack either in a
3437 // physical register or in the parameter list area of the current stack
3438 // frame to the callee.
3442 if (VA.isRegLoc()) {
3443 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3444 // Put argument in a physical register.
3445 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3447 // Put argument in the parameter list area of the current stack frame.
3448 assert(VA.isMemLoc());
3449 unsigned LocMemOffset = VA.getLocMemOffset();
3452 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3453 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3455 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3456 MachinePointerInfo(),
3459 // Calculate and remember argument location.
3460 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3466 if (!MemOpChains.empty())
3467 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3468 &MemOpChains[0], MemOpChains.size());
3470 // Build a sequence of copy-to-reg nodes chained together with token chain
3471 // and flag operands which copy the outgoing args into the appropriate regs.
3473 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3474 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3475 RegsToPass[i].second, InFlag);
3476 InFlag = Chain.getValue(1);
3479 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3482 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3483 SDValue Ops[] = { Chain, InFlag };
3485 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3486 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3488 InFlag = Chain.getValue(1);
3492 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3493 false, TailCallArguments);
3495 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3496 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3500 // Copy an argument into memory, being careful to do this outside the
3501 // call sequence for the call to which the argument belongs.
3503 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3504 SDValue CallSeqStart,
3505 ISD::ArgFlagsTy Flags,
3507 DebugLoc dl) const {
3508 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3509 CallSeqStart.getNode()->getOperand(0),
3511 // The MEMCPY must go outside the CALLSEQ_START..END.
3512 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3513 CallSeqStart.getNode()->getOperand(1));
3514 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3515 NewCallSeqStart.getNode());
3516 return NewCallSeqStart;
3520 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3521 CallingConv::ID CallConv, bool isVarArg,
3523 const SmallVectorImpl<ISD::OutputArg> &Outs,
3524 const SmallVectorImpl<SDValue> &OutVals,
3525 const SmallVectorImpl<ISD::InputArg> &Ins,
3526 DebugLoc dl, SelectionDAG &DAG,
3527 SmallVectorImpl<SDValue> &InVals) const {
3529 unsigned NumOps = Outs.size();
3531 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3532 unsigned PtrByteSize = 8;
3534 MachineFunction &MF = DAG.getMachineFunction();
3536 // Mark this function as potentially containing a function that contains a
3537 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3538 // and restoring the callers stack pointer in this functions epilog. This is
3539 // done because by tail calling the called function might overwrite the value
3540 // in this function's (MF) stack pointer stack slot 0(SP).
3541 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3542 CallConv == CallingConv::Fast)
3543 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3545 unsigned nAltivecParamsAtEnd = 0;
3547 // Count how many bytes are to be pushed on the stack, including the linkage
3548 // area, and parameter passing area. We start with at least 48 bytes, which
3549 // is reserved space for [SP][CR][LR][3 x unused].
3550 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3553 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3554 Outs, OutVals, nAltivecParamsAtEnd);
3556 // Calculate by how many bytes the stack has to be adjusted in case of tail
3557 // call optimization.
3558 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3560 // To protect arguments on the stack from being clobbered in a tail call,
3561 // force all the loads to happen before doing any other lowering.
3563 Chain = DAG.getStackArgumentTokenFactor(Chain);
3565 // Adjust the stack pointer for the new arguments...
3566 // These operations are automatically eliminated by the prolog/epilog pass
3567 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3568 SDValue CallSeqStart = Chain;
3570 // Load the return address and frame pointer so it can be move somewhere else
3573 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3576 // Set up a copy of the stack pointer for use loading and storing any
3577 // arguments that may not fit in the registers available for argument
3579 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3581 // Figure out which arguments are going to go in registers, and which in
3582 // memory. Also, if this is a vararg function, floating point operations
3583 // must be stored to our stack, and loaded into integer regs as well, if
3584 // any integer regs are available for argument passing.
3585 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3586 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3588 static const uint16_t GPR[] = {
3589 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3590 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3592 static const uint16_t *FPR = GetFPR();
3594 static const uint16_t VR[] = {
3595 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3596 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3598 const unsigned NumGPRs = array_lengthof(GPR);
3599 const unsigned NumFPRs = 13;
3600 const unsigned NumVRs = array_lengthof(VR);
3602 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3603 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3605 SmallVector<SDValue, 8> MemOpChains;
3606 for (unsigned i = 0; i != NumOps; ++i) {
3607 SDValue Arg = OutVals[i];
3608 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3610 // PtrOff will be used to store the current argument to the stack if a
3611 // register cannot be found for it.
3614 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3616 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3618 // Promote integers to 64-bit values.
3619 if (Arg.getValueType() == MVT::i32) {
3620 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3621 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3622 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3625 // FIXME memcpy is used way more than necessary. Correctness first.
3626 // Note: "by value" is code for passing a structure by value, not
3628 if (Flags.isByVal()) {
3629 // Note: Size includes alignment padding, so
3630 // struct x { short a; char b; }
3631 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3632 // These are the proper values we need for right-justifying the
3633 // aggregate in a parameter register.
3634 unsigned Size = Flags.getByValSize();
3635 // All aggregates smaller than 8 bytes must be passed right-justified.
3636 if (Size==1 || Size==2 || Size==4) {
3637 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3638 if (GPR_idx != NumGPRs) {
3639 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3640 MachinePointerInfo(), VT,
3642 MemOpChains.push_back(Load.getValue(1));
3643 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3645 ArgOffset += PtrByteSize;
3650 if (GPR_idx == NumGPRs && Size < 8) {
3651 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3652 PtrOff.getValueType());
3653 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3654 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3657 ArgOffset += PtrByteSize;
3660 // Copy entire object into memory. There are cases where gcc-generated
3661 // code assumes it is there, even if it could be put entirely into
3662 // registers. (This is not what the doc says.)
3664 // FIXME: The above statement is likely due to a misunderstanding of the
3665 // documents. All arguments must be copied into the parameter area BY
3666 // THE CALLEE in the event that the callee takes the address of any
3667 // formal argument. That has not yet been implemented. However, it is
3668 // reasonable to use the stack area as a staging area for the register
3671 // Skip this for small aggregates, as we will use the same slot for a
3672 // right-justified copy, below.
3674 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3678 // When a register is available, pass a small aggregate right-justified.
3679 if (Size < 8 && GPR_idx != NumGPRs) {
3680 // The easiest way to get this right-justified in a register
3681 // is to copy the structure into the rightmost portion of a
3682 // local variable slot, then load the whole slot into the
3684 // FIXME: The memcpy seems to produce pretty awful code for
3685 // small aggregates, particularly for packed ones.
3686 // FIXME: It would be preferable to use the slot in the
3687 // parameter save area instead of a new local variable.
3688 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3689 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3690 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3694 // Load the slot into the register.
3695 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3696 MachinePointerInfo(),
3697 false, false, false, 0);
3698 MemOpChains.push_back(Load.getValue(1));
3699 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3701 // Done with this argument.
3702 ArgOffset += PtrByteSize;
3706 // For aggregates larger than PtrByteSize, copy the pieces of the
3707 // object that fit into registers from the parameter save area.
3708 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3709 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3710 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3711 if (GPR_idx != NumGPRs) {
3712 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3713 MachinePointerInfo(),
3714 false, false, false, 0);
3715 MemOpChains.push_back(Load.getValue(1));
3716 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3717 ArgOffset += PtrByteSize;
3719 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3726 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3727 default: llvm_unreachable("Unexpected ValueType for argument!");
3730 if (GPR_idx != NumGPRs) {
3731 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3733 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3734 true, isTailCall, false, MemOpChains,
3735 TailCallArguments, dl);
3737 ArgOffset += PtrByteSize;
3741 if (FPR_idx != NumFPRs) {
3742 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3745 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3746 MachinePointerInfo(), false, false, 0);
3747 MemOpChains.push_back(Store);
3749 // Float varargs are always shadowed in available integer registers
3750 if (GPR_idx != NumGPRs) {
3751 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3752 MachinePointerInfo(), false, false,
3754 MemOpChains.push_back(Load.getValue(1));
3755 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3757 } else if (GPR_idx != NumGPRs)
3758 // If we have any FPRs remaining, we may also have GPRs remaining.
3761 // Single-precision floating-point values are mapped to the
3762 // second (rightmost) word of the stack doubleword.
3763 if (Arg.getValueType() == MVT::f32) {
3764 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3765 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3768 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3769 true, isTailCall, false, MemOpChains,
3770 TailCallArguments, dl);
3779 // These go aligned on the stack, or in the corresponding R registers
3780 // when within range. The Darwin PPC ABI doc claims they also go in
3781 // V registers; in fact gcc does this only for arguments that are
3782 // prototyped, not for those that match the ... We do it for all
3783 // arguments, seems to work.
3784 while (ArgOffset % 16 !=0) {
3785 ArgOffset += PtrByteSize;
3786 if (GPR_idx != NumGPRs)
3789 // We could elide this store in the case where the object fits
3790 // entirely in R registers. Maybe later.
3791 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3792 DAG.getConstant(ArgOffset, PtrVT));
3793 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3794 MachinePointerInfo(), false, false, 0);
3795 MemOpChains.push_back(Store);
3796 if (VR_idx != NumVRs) {
3797 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3798 MachinePointerInfo(),
3799 false, false, false, 0);
3800 MemOpChains.push_back(Load.getValue(1));
3801 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3804 for (unsigned i=0; i<16; i+=PtrByteSize) {
3805 if (GPR_idx == NumGPRs)
3807 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3808 DAG.getConstant(i, PtrVT));
3809 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3810 false, false, false, 0);
3811 MemOpChains.push_back(Load.getValue(1));
3812 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3817 // Non-varargs Altivec params generally go in registers, but have
3818 // stack space allocated at the end.
3819 if (VR_idx != NumVRs) {
3820 // Doesn't have GPR space allocated.
3821 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3823 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3824 true, isTailCall, true, MemOpChains,
3825 TailCallArguments, dl);
3832 if (!MemOpChains.empty())
3833 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3834 &MemOpChains[0], MemOpChains.size());
3836 // Check if this is an indirect call (MTCTR/BCTRL).
3837 // See PrepareCall() for more information about calls through function
3838 // pointers in the 64-bit SVR4 ABI.
3840 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3841 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3842 !isBLACompatibleAddress(Callee, DAG)) {
3843 // Load r2 into a virtual register and store it to the TOC save area.
3844 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3845 // TOC save area offset.
3846 SDValue PtrOff = DAG.getIntPtrConstant(40);
3847 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3848 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3850 // R12 must contain the address of an indirect callee. This does not
3851 // mean the MTCTR instruction must use R12; it's easier to model this
3852 // as an extra parameter, so do that.
3853 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
3856 // Build a sequence of copy-to-reg nodes chained together with token chain
3857 // and flag operands which copy the outgoing args into the appropriate regs.
3859 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3860 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3861 RegsToPass[i].second, InFlag);
3862 InFlag = Chain.getValue(1);
3866 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
3867 FPOp, true, TailCallArguments);
3869 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3870 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3875 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3876 CallingConv::ID CallConv, bool isVarArg,
3878 const SmallVectorImpl<ISD::OutputArg> &Outs,
3879 const SmallVectorImpl<SDValue> &OutVals,
3880 const SmallVectorImpl<ISD::InputArg> &Ins,
3881 DebugLoc dl, SelectionDAG &DAG,
3882 SmallVectorImpl<SDValue> &InVals) const {
3884 unsigned NumOps = Outs.size();
3886 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3887 bool isPPC64 = PtrVT == MVT::i64;
3888 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3890 MachineFunction &MF = DAG.getMachineFunction();
3892 // Mark this function as potentially containing a function that contains a
3893 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3894 // and restoring the callers stack pointer in this functions epilog. This is
3895 // done because by tail calling the called function might overwrite the value
3896 // in this function's (MF) stack pointer stack slot 0(SP).
3897 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3898 CallConv == CallingConv::Fast)
3899 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3901 unsigned nAltivecParamsAtEnd = 0;
3903 // Count how many bytes are to be pushed on the stack, including the linkage
3904 // area, and parameter passing area. We start with 24/48 bytes, which is
3905 // prereserved space for [SP][CR][LR][3 x unused].
3907 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3909 nAltivecParamsAtEnd);
3911 // Calculate by how many bytes the stack has to be adjusted in case of tail
3912 // call optimization.
3913 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3915 // To protect arguments on the stack from being clobbered in a tail call,
3916 // force all the loads to happen before doing any other lowering.
3918 Chain = DAG.getStackArgumentTokenFactor(Chain);
3920 // Adjust the stack pointer for the new arguments...
3921 // These operations are automatically eliminated by the prolog/epilog pass
3922 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3923 SDValue CallSeqStart = Chain;
3925 // Load the return address and frame pointer so it can be move somewhere else
3928 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3931 // Set up a copy of the stack pointer for use loading and storing any
3932 // arguments that may not fit in the registers available for argument
3936 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3938 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3940 // Figure out which arguments are going to go in registers, and which in
3941 // memory. Also, if this is a vararg function, floating point operations
3942 // must be stored to our stack, and loaded into integer regs as well, if
3943 // any integer regs are available for argument passing.
3944 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3945 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3947 static const uint16_t GPR_32[] = { // 32-bit registers.
3948 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3949 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3951 static const uint16_t GPR_64[] = { // 64-bit registers.
3952 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3953 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3955 static const uint16_t *FPR = GetFPR();
3957 static const uint16_t VR[] = {
3958 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3959 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3961 const unsigned NumGPRs = array_lengthof(GPR_32);
3962 const unsigned NumFPRs = 13;
3963 const unsigned NumVRs = array_lengthof(VR);
3965 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
3967 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3968 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3970 SmallVector<SDValue, 8> MemOpChains;
3971 for (unsigned i = 0; i != NumOps; ++i) {
3972 SDValue Arg = OutVals[i];
3973 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3975 // PtrOff will be used to store the current argument to the stack if a
3976 // register cannot be found for it.
3979 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3981 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3983 // On PPC64, promote integers to 64-bit values.
3984 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3985 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3986 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3987 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3990 // FIXME memcpy is used way more than necessary. Correctness first.
3991 // Note: "by value" is code for passing a structure by value, not
3993 if (Flags.isByVal()) {
3994 unsigned Size = Flags.getByValSize();
3995 // Very small objects are passed right-justified. Everything else is
3996 // passed left-justified.
3997 if (Size==1 || Size==2) {
3998 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3999 if (GPR_idx != NumGPRs) {
4000 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4001 MachinePointerInfo(), VT,
4003 MemOpChains.push_back(Load.getValue(1));
4004 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4006 ArgOffset += PtrByteSize;
4008 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4009 PtrOff.getValueType());
4010 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4011 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4014 ArgOffset += PtrByteSize;
4018 // Copy entire object into memory. There are cases where gcc-generated
4019 // code assumes it is there, even if it could be put entirely into
4020 // registers. (This is not what the doc says.)
4021 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4025 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4026 // copy the pieces of the object that fit into registers from the
4027 // parameter save area.
4028 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4029 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4030 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4031 if (GPR_idx != NumGPRs) {
4032 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4033 MachinePointerInfo(),
4034 false, false, false, 0);
4035 MemOpChains.push_back(Load.getValue(1));
4036 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4037 ArgOffset += PtrByteSize;
4039 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4046 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4047 default: llvm_unreachable("Unexpected ValueType for argument!");
4050 if (GPR_idx != NumGPRs) {
4051 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4053 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4054 isPPC64, isTailCall, false, MemOpChains,
4055 TailCallArguments, dl);
4057 ArgOffset += PtrByteSize;
4061 if (FPR_idx != NumFPRs) {
4062 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4065 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4066 MachinePointerInfo(), false, false, 0);
4067 MemOpChains.push_back(Store);
4069 // Float varargs are always shadowed in available integer registers
4070 if (GPR_idx != NumGPRs) {
4071 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4072 MachinePointerInfo(), false, false,
4074 MemOpChains.push_back(Load.getValue(1));
4075 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4077 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4078 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4079 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4080 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4081 MachinePointerInfo(),
4082 false, false, false, 0);
4083 MemOpChains.push_back(Load.getValue(1));
4084 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4087 // If we have any FPRs remaining, we may also have GPRs remaining.
4088 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4090 if (GPR_idx != NumGPRs)
4092 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4093 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4097 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4098 isPPC64, isTailCall, false, MemOpChains,
4099 TailCallArguments, dl);
4103 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4110 // These go aligned on the stack, or in the corresponding R registers
4111 // when within range. The Darwin PPC ABI doc claims they also go in
4112 // V registers; in fact gcc does this only for arguments that are
4113 // prototyped, not for those that match the ... We do it for all
4114 // arguments, seems to work.
4115 while (ArgOffset % 16 !=0) {
4116 ArgOffset += PtrByteSize;
4117 if (GPR_idx != NumGPRs)
4120 // We could elide this store in the case where the object fits
4121 // entirely in R registers. Maybe later.
4122 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4123 DAG.getConstant(ArgOffset, PtrVT));
4124 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4125 MachinePointerInfo(), false, false, 0);
4126 MemOpChains.push_back(Store);
4127 if (VR_idx != NumVRs) {
4128 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4129 MachinePointerInfo(),
4130 false, false, false, 0);
4131 MemOpChains.push_back(Load.getValue(1));
4132 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4135 for (unsigned i=0; i<16; i+=PtrByteSize) {
4136 if (GPR_idx == NumGPRs)
4138 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4139 DAG.getConstant(i, PtrVT));
4140 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4141 false, false, false, 0);
4142 MemOpChains.push_back(Load.getValue(1));
4143 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4148 // Non-varargs Altivec params generally go in registers, but have
4149 // stack space allocated at the end.
4150 if (VR_idx != NumVRs) {
4151 // Doesn't have GPR space allocated.
4152 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4153 } else if (nAltivecParamsAtEnd==0) {
4154 // We are emitting Altivec params in order.
4155 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4156 isPPC64, isTailCall, true, MemOpChains,
4157 TailCallArguments, dl);
4163 // If all Altivec parameters fit in registers, as they usually do,
4164 // they get stack space following the non-Altivec parameters. We
4165 // don't track this here because nobody below needs it.
4166 // If there are more Altivec parameters than fit in registers emit
4168 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4170 // Offset is aligned; skip 1st 12 params which go in V registers.
4171 ArgOffset = ((ArgOffset+15)/16)*16;
4173 for (unsigned i = 0; i != NumOps; ++i) {
4174 SDValue Arg = OutVals[i];
4175 EVT ArgType = Outs[i].VT;
4176 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4177 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4180 // We are emitting Altivec params in order.
4181 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4182 isPPC64, isTailCall, true, MemOpChains,
4183 TailCallArguments, dl);
4190 if (!MemOpChains.empty())
4191 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4192 &MemOpChains[0], MemOpChains.size());
4194 // On Darwin, R12 must contain the address of an indirect callee. This does
4195 // not mean the MTCTR instruction must use R12; it's easier to model this as
4196 // an extra parameter, so do that.
4198 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4199 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4200 !isBLACompatibleAddress(Callee, DAG))
4201 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4202 PPC::R12), Callee));
4204 // Build a sequence of copy-to-reg nodes chained together with token chain
4205 // and flag operands which copy the outgoing args into the appropriate regs.
4207 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4208 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4209 RegsToPass[i].second, InFlag);
4210 InFlag = Chain.getValue(1);
4214 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4215 FPOp, true, TailCallArguments);
4217 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4218 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4223 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4224 MachineFunction &MF, bool isVarArg,
4225 const SmallVectorImpl<ISD::OutputArg> &Outs,
4226 LLVMContext &Context) const {
4227 SmallVector<CCValAssign, 16> RVLocs;
4228 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4230 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4234 PPCTargetLowering::LowerReturn(SDValue Chain,
4235 CallingConv::ID CallConv, bool isVarArg,
4236 const SmallVectorImpl<ISD::OutputArg> &Outs,
4237 const SmallVectorImpl<SDValue> &OutVals,
4238 DebugLoc dl, SelectionDAG &DAG) const {
4240 SmallVector<CCValAssign, 16> RVLocs;
4241 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4242 getTargetMachine(), RVLocs, *DAG.getContext());
4243 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4245 // If this is the first return lowered for this function, add the regs to the
4246 // liveout set for the function.
4247 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
4248 for (unsigned i = 0; i != RVLocs.size(); ++i)
4249 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
4254 // Copy the result values into the output registers.
4255 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4256 CCValAssign &VA = RVLocs[i];
4257 assert(VA.isRegLoc() && "Can only return in registers!");
4258 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
4260 Flag = Chain.getValue(1);
4264 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
4266 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
4269 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4270 const PPCSubtarget &Subtarget) const {
4271 // When we pop the dynamic allocation we need to restore the SP link.
4272 DebugLoc dl = Op.getDebugLoc();
4274 // Get the corect type for pointers.
4275 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4277 // Construct the stack pointer operand.
4278 bool isPPC64 = Subtarget.isPPC64();
4279 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4280 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4282 // Get the operands for the STACKRESTORE.
4283 SDValue Chain = Op.getOperand(0);
4284 SDValue SaveSP = Op.getOperand(1);
4286 // Load the old link SP.
4287 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4288 MachinePointerInfo(),
4289 false, false, false, 0);
4291 // Restore the stack pointer.
4292 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4294 // Store the old link SP.
4295 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4302 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4303 MachineFunction &MF = DAG.getMachineFunction();
4304 bool isPPC64 = PPCSubTarget.isPPC64();
4305 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4306 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4308 // Get current frame pointer save index. The users of this index will be
4309 // primarily DYNALLOC instructions.
4310 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4311 int RASI = FI->getReturnAddrSaveIndex();
4313 // If the frame pointer save index hasn't been defined yet.
4315 // Find out what the fix offset of the frame pointer save area.
4316 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4317 // Allocate the frame index for frame pointer save area.
4318 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4320 FI->setReturnAddrSaveIndex(RASI);
4322 return DAG.getFrameIndex(RASI, PtrVT);
4326 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4327 MachineFunction &MF = DAG.getMachineFunction();
4328 bool isPPC64 = PPCSubTarget.isPPC64();
4329 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4330 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4332 // Get current frame pointer save index. The users of this index will be
4333 // primarily DYNALLOC instructions.
4334 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4335 int FPSI = FI->getFramePointerSaveIndex();
4337 // If the frame pointer save index hasn't been defined yet.
4339 // Find out what the fix offset of the frame pointer save area.
4340 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4343 // Allocate the frame index for frame pointer save area.
4344 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4346 FI->setFramePointerSaveIndex(FPSI);
4348 return DAG.getFrameIndex(FPSI, PtrVT);
4351 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4353 const PPCSubtarget &Subtarget) const {
4355 SDValue Chain = Op.getOperand(0);
4356 SDValue Size = Op.getOperand(1);
4357 DebugLoc dl = Op.getDebugLoc();
4359 // Get the corect type for pointers.
4360 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4362 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4363 DAG.getConstant(0, PtrVT), Size);
4364 // Construct a node for the frame pointer save index.
4365 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4366 // Build a DYNALLOC node.
4367 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4368 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4369 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4372 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4374 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4375 // Not FP? Not a fsel.
4376 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4377 !Op.getOperand(2).getValueType().isFloatingPoint())
4380 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4382 // Cannot handle SETEQ/SETNE.
4383 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4385 EVT ResVT = Op.getValueType();
4386 EVT CmpVT = Op.getOperand(0).getValueType();
4387 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4388 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4389 DebugLoc dl = Op.getDebugLoc();
4391 // If the RHS of the comparison is a 0.0, we don't need to do the
4392 // subtraction at all.
4393 if (isFloatingPointZero(RHS))
4395 default: break; // SETUO etc aren't handled by fsel.
4398 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4401 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4402 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4403 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4406 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4409 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4410 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4411 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4412 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4417 default: break; // SETUO etc aren't handled by fsel.
4420 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4421 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4422 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4423 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4426 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4427 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4428 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4429 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4432 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4433 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4434 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4435 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4438 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4439 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4440 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4441 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4446 // FIXME: Split this code up when LegalizeDAGTypes lands.
4447 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4448 DebugLoc dl) const {
4449 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4450 SDValue Src = Op.getOperand(0);
4451 if (Src.getValueType() == MVT::f32)
4452 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4455 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4456 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4458 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4463 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4467 // Convert the FP value to an int value through memory.
4468 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4470 // Emit a store to the stack slot.
4471 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4472 MachinePointerInfo(), false, false, 0);
4474 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4476 if (Op.getValueType() == MVT::i32)
4477 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4478 DAG.getConstant(4, FIPtr.getValueType()));
4479 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4480 false, false, false, 0);
4483 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4484 SelectionDAG &DAG) const {
4485 DebugLoc dl = Op.getDebugLoc();
4486 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4487 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4490 if (Op.getOperand(0).getValueType() == MVT::i64) {
4491 SDValue SINT = Op.getOperand(0);
4492 // When converting to single-precision, we actually need to convert
4493 // to double-precision first and then round to single-precision.
4494 // To avoid double-rounding effects during that operation, we have
4495 // to prepare the input operand. Bits that might be truncated when
4496 // converting to double-precision are replaced by a bit that won't
4497 // be lost at this stage, but is below the single-precision rounding
4500 // However, if -enable-unsafe-fp-math is in effect, accept double
4501 // rounding to avoid the extra overhead.
4502 if (Op.getValueType() == MVT::f32 &&
4503 !DAG.getTarget().Options.UnsafeFPMath) {
4505 // Twiddle input to make sure the low 11 bits are zero. (If this
4506 // is the case, we are guaranteed the value will fit into the 53 bit
4507 // mantissa of an IEEE double-precision value without rounding.)
4508 // If any of those low 11 bits were not zero originally, make sure
4509 // bit 12 (value 2048) is set instead, so that the final rounding
4510 // to single-precision gets the correct result.
4511 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4512 SINT, DAG.getConstant(2047, MVT::i64));
4513 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4514 Round, DAG.getConstant(2047, MVT::i64));
4515 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4516 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4517 Round, DAG.getConstant(-2048, MVT::i64));
4519 // However, we cannot use that value unconditionally: if the magnitude
4520 // of the input value is small, the bit-twiddling we did above might
4521 // end up visibly changing the output. Fortunately, in that case, we
4522 // don't need to twiddle bits since the original input will convert
4523 // exactly to double-precision floating-point already. Therefore,
4524 // construct a conditional to use the original value if the top 11
4525 // bits are all sign-bit copies, and use the rounded value computed
4527 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4528 SINT, DAG.getConstant(53, MVT::i32));
4529 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4530 Cond, DAG.getConstant(1, MVT::i64));
4531 Cond = DAG.getSetCC(dl, MVT::i32,
4532 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4534 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4536 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4537 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4538 if (Op.getValueType() == MVT::f32)
4539 FP = DAG.getNode(ISD::FP_ROUND, dl,
4540 MVT::f32, FP, DAG.getIntPtrConstant(0));
4544 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4545 "Unhandled SINT_TO_FP type in custom expander!");
4546 // Since we only generate this in 64-bit mode, we can take advantage of
4547 // 64-bit registers. In particular, sign extend the input value into the
4548 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4549 // then lfd it and fcfid it.
4550 MachineFunction &MF = DAG.getMachineFunction();
4551 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4552 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4553 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4554 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4556 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4559 // STD the extended value into the stack slot.
4560 MachineMemOperand *MMO =
4561 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4562 MachineMemOperand::MOStore, 8, 8);
4563 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4565 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4566 Ops, 4, MVT::i64, MMO);
4567 // Load the value as a double.
4568 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4569 false, false, false, 0);
4571 // FCFID it and return it.
4572 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4573 if (Op.getValueType() == MVT::f32)
4574 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4578 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4579 SelectionDAG &DAG) const {
4580 DebugLoc dl = Op.getDebugLoc();
4582 The rounding mode is in bits 30:31 of FPSR, and has the following
4589 FLT_ROUNDS, on the other hand, expects the following:
4596 To perform the conversion, we do:
4597 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4600 MachineFunction &MF = DAG.getMachineFunction();
4601 EVT VT = Op.getValueType();
4602 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4603 std::vector<EVT> NodeTys;
4604 SDValue MFFSreg, InFlag;
4606 // Save FP Control Word to register
4607 NodeTys.push_back(MVT::f64); // return register
4608 NodeTys.push_back(MVT::Glue); // unused in this context
4609 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4611 // Save FP register to stack slot
4612 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4613 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4614 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4615 StackSlot, MachinePointerInfo(), false, false,0);
4617 // Load FP Control Word from low 32 bits of stack slot.
4618 SDValue Four = DAG.getConstant(4, PtrVT);
4619 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4620 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4621 false, false, false, 0);
4623 // Transform as necessary
4625 DAG.getNode(ISD::AND, dl, MVT::i32,
4626 CWD, DAG.getConstant(3, MVT::i32));
4628 DAG.getNode(ISD::SRL, dl, MVT::i32,
4629 DAG.getNode(ISD::AND, dl, MVT::i32,
4630 DAG.getNode(ISD::XOR, dl, MVT::i32,
4631 CWD, DAG.getConstant(3, MVT::i32)),
4632 DAG.getConstant(3, MVT::i32)),
4633 DAG.getConstant(1, MVT::i32));
4636 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4638 return DAG.getNode((VT.getSizeInBits() < 16 ?
4639 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4642 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4643 EVT VT = Op.getValueType();
4644 unsigned BitWidth = VT.getSizeInBits();
4645 DebugLoc dl = Op.getDebugLoc();
4646 assert(Op.getNumOperands() == 3 &&
4647 VT == Op.getOperand(1).getValueType() &&
4650 // Expand into a bunch of logical ops. Note that these ops
4651 // depend on the PPC behavior for oversized shift amounts.
4652 SDValue Lo = Op.getOperand(0);
4653 SDValue Hi = Op.getOperand(1);
4654 SDValue Amt = Op.getOperand(2);
4655 EVT AmtVT = Amt.getValueType();
4657 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4658 DAG.getConstant(BitWidth, AmtVT), Amt);
4659 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4660 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4661 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4662 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4663 DAG.getConstant(-BitWidth, AmtVT));
4664 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4665 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4666 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4667 SDValue OutOps[] = { OutLo, OutHi };
4668 return DAG.getMergeValues(OutOps, 2, dl);
4671 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4672 EVT VT = Op.getValueType();
4673 DebugLoc dl = Op.getDebugLoc();
4674 unsigned BitWidth = VT.getSizeInBits();
4675 assert(Op.getNumOperands() == 3 &&
4676 VT == Op.getOperand(1).getValueType() &&
4679 // Expand into a bunch of logical ops. Note that these ops
4680 // depend on the PPC behavior for oversized shift amounts.
4681 SDValue Lo = Op.getOperand(0);
4682 SDValue Hi = Op.getOperand(1);
4683 SDValue Amt = Op.getOperand(2);
4684 EVT AmtVT = Amt.getValueType();
4686 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4687 DAG.getConstant(BitWidth, AmtVT), Amt);
4688 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4689 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4690 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4691 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4692 DAG.getConstant(-BitWidth, AmtVT));
4693 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4694 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4695 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4696 SDValue OutOps[] = { OutLo, OutHi };
4697 return DAG.getMergeValues(OutOps, 2, dl);
4700 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4701 DebugLoc dl = Op.getDebugLoc();
4702 EVT VT = Op.getValueType();
4703 unsigned BitWidth = VT.getSizeInBits();
4704 assert(Op.getNumOperands() == 3 &&
4705 VT == Op.getOperand(1).getValueType() &&
4708 // Expand into a bunch of logical ops, followed by a select_cc.
4709 SDValue Lo = Op.getOperand(0);
4710 SDValue Hi = Op.getOperand(1);
4711 SDValue Amt = Op.getOperand(2);
4712 EVT AmtVT = Amt.getValueType();
4714 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4715 DAG.getConstant(BitWidth, AmtVT), Amt);
4716 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4717 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4718 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4719 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4720 DAG.getConstant(-BitWidth, AmtVT));
4721 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4722 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4723 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4724 Tmp4, Tmp6, ISD::SETLE);
4725 SDValue OutOps[] = { OutLo, OutHi };
4726 return DAG.getMergeValues(OutOps, 2, dl);
4729 //===----------------------------------------------------------------------===//
4730 // Vector related lowering.
4733 /// BuildSplatI - Build a canonical splati of Val with an element size of
4734 /// SplatSize. Cast the result to VT.
4735 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4736 SelectionDAG &DAG, DebugLoc dl) {
4737 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4739 static const EVT VTys[] = { // canonical VT to use for each size.
4740 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4743 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4745 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4749 EVT CanonicalVT = VTys[SplatSize-1];
4751 // Build a canonical splat for this value.
4752 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4753 SmallVector<SDValue, 8> Ops;
4754 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4755 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4756 &Ops[0], Ops.size());
4757 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4760 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4761 /// specified intrinsic ID.
4762 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4763 SelectionDAG &DAG, DebugLoc dl,
4764 EVT DestVT = MVT::Other) {
4765 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4766 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4767 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4770 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4771 /// specified intrinsic ID.
4772 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4773 SDValue Op2, SelectionDAG &DAG,
4774 DebugLoc dl, EVT DestVT = MVT::Other) {
4775 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4776 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4777 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4781 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4782 /// amount. The result has the specified value type.
4783 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4784 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4785 // Force LHS/RHS to be the right type.
4786 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4787 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4790 for (unsigned i = 0; i != 16; ++i)
4792 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4793 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4796 // If this is a case we can't handle, return null and let the default
4797 // expansion code take care of it. If we CAN select this case, and if it
4798 // selects to a single instruction, return Op. Otherwise, if we can codegen
4799 // this case more efficiently than a constant pool load, lower it to the
4800 // sequence of ops that should be used.
4801 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4802 SelectionDAG &DAG) const {
4803 DebugLoc dl = Op.getDebugLoc();
4804 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4805 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4807 // Check if this is a splat of a constant value.
4808 APInt APSplatBits, APSplatUndef;
4809 unsigned SplatBitSize;
4811 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4812 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4815 unsigned SplatBits = APSplatBits.getZExtValue();
4816 unsigned SplatUndef = APSplatUndef.getZExtValue();
4817 unsigned SplatSize = SplatBitSize / 8;
4819 // First, handle single instruction cases.
4822 if (SplatBits == 0) {
4823 // Canonicalize all zero vectors to be v4i32.
4824 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4825 SDValue Z = DAG.getConstant(0, MVT::i32);
4826 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4827 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4832 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4833 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4835 if (SextVal >= -16 && SextVal <= 15)
4836 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4839 // Two instruction sequences.
4841 // If this value is in the range [-32,30] and is even, use:
4842 // tmp = VSPLTI[bhw], result = add tmp, tmp
4843 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4844 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4845 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4846 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4849 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4850 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4852 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4853 // Make -1 and vspltisw -1:
4854 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4856 // Make the VSLW intrinsic, computing 0x8000_0000.
4857 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4860 // xor by OnesV to invert it.
4861 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4862 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4865 // Check to see if this is a wide variety of vsplti*, binop self cases.
4866 static const signed char SplatCsts[] = {
4867 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4868 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4871 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4872 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4873 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4874 int i = SplatCsts[idx];
4876 // Figure out what shift amount will be used by altivec if shifted by i in
4878 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4880 // vsplti + shl self.
4881 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4882 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4883 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4884 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4885 Intrinsic::ppc_altivec_vslw
4887 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4888 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4891 // vsplti + srl self.
4892 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4893 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4894 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4895 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4896 Intrinsic::ppc_altivec_vsrw
4898 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4899 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4902 // vsplti + sra self.
4903 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4904 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4905 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4906 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
4907 Intrinsic::ppc_altivec_vsraw
4909 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4910 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4913 // vsplti + rol self.
4914 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
4915 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
4916 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4917 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4918 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
4919 Intrinsic::ppc_altivec_vrlw
4921 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4922 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4925 // t = vsplti c, result = vsldoi t, t, 1
4926 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
4927 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4928 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
4930 // t = vsplti c, result = vsldoi t, t, 2
4931 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
4932 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4933 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
4935 // t = vsplti c, result = vsldoi t, t, 3
4936 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
4937 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
4938 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
4942 // Three instruction sequences.
4944 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
4945 if (SextVal >= 0 && SextVal <= 31) {
4946 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
4947 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4948 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
4949 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4951 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
4952 if (SextVal >= -31 && SextVal <= 0) {
4953 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
4954 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
4955 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
4956 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
4962 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4963 /// the specified operations to build the shuffle.
4964 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4965 SDValue RHS, SelectionDAG &DAG,
4967 unsigned OpNum = (PFEntry >> 26) & 0x0F;
4968 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4969 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4972 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4984 if (OpNum == OP_COPY) {
4985 if (LHSID == (1*9+2)*9+3) return LHS;
4986 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4990 SDValue OpLHS, OpRHS;
4991 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4992 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4996 default: llvm_unreachable("Unknown i32 permute!");
4998 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4999 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5000 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5001 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5004 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5005 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5006 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5007 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5010 for (unsigned i = 0; i != 16; ++i)
5011 ShufIdxs[i] = (i&3)+0;
5014 for (unsigned i = 0; i != 16; ++i)
5015 ShufIdxs[i] = (i&3)+4;
5018 for (unsigned i = 0; i != 16; ++i)
5019 ShufIdxs[i] = (i&3)+8;
5022 for (unsigned i = 0; i != 16; ++i)
5023 ShufIdxs[i] = (i&3)+12;
5026 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5028 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5030 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5032 EVT VT = OpLHS.getValueType();
5033 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5034 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5035 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5036 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5039 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5040 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5041 /// return the code it can be lowered into. Worst case, it can always be
5042 /// lowered into a vperm.
5043 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5044 SelectionDAG &DAG) const {
5045 DebugLoc dl = Op.getDebugLoc();
5046 SDValue V1 = Op.getOperand(0);
5047 SDValue V2 = Op.getOperand(1);
5048 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5049 EVT VT = Op.getValueType();
5051 // Cases that are handled by instructions that take permute immediates
5052 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5053 // selected by the instruction selector.
5054 if (V2.getOpcode() == ISD::UNDEF) {
5055 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5056 PPC::isSplatShuffleMask(SVOp, 2) ||
5057 PPC::isSplatShuffleMask(SVOp, 4) ||
5058 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5059 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5060 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5061 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5062 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5063 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5064 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5065 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5066 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5071 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5072 // and produce a fixed permutation. If any of these match, do not lower to
5074 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5075 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5076 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5077 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5078 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5079 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5080 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5081 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5082 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5085 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5086 // perfect shuffle table to emit an optimal matching sequence.
5087 ArrayRef<int> PermMask = SVOp->getMask();
5089 unsigned PFIndexes[4];
5090 bool isFourElementShuffle = true;
5091 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5092 unsigned EltNo = 8; // Start out undef.
5093 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5094 if (PermMask[i*4+j] < 0)
5095 continue; // Undef, ignore it.
5097 unsigned ByteSource = PermMask[i*4+j];
5098 if ((ByteSource & 3) != j) {
5099 isFourElementShuffle = false;
5104 EltNo = ByteSource/4;
5105 } else if (EltNo != ByteSource/4) {
5106 isFourElementShuffle = false;
5110 PFIndexes[i] = EltNo;
5113 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5114 // perfect shuffle vector to determine if it is cost effective to do this as
5115 // discrete instructions, or whether we should use a vperm.
5116 if (isFourElementShuffle) {
5117 // Compute the index in the perfect shuffle table.
5118 unsigned PFTableIndex =
5119 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5121 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5122 unsigned Cost = (PFEntry >> 30);
5124 // Determining when to avoid vperm is tricky. Many things affect the cost
5125 // of vperm, particularly how many times the perm mask needs to be computed.
5126 // For example, if the perm mask can be hoisted out of a loop or is already
5127 // used (perhaps because there are multiple permutes with the same shuffle
5128 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5129 // the loop requires an extra register.
5131 // As a compromise, we only emit discrete instructions if the shuffle can be
5132 // generated in 3 or fewer operations. When we have loop information
5133 // available, if this block is within a loop, we should avoid using vperm
5134 // for 3-operation perms and use a constant pool load instead.
5136 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5139 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5140 // vector that will get spilled to the constant pool.
5141 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5143 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5144 // that it is in input element units, not in bytes. Convert now.
5145 EVT EltVT = V1.getValueType().getVectorElementType();
5146 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5148 SmallVector<SDValue, 16> ResultMask;
5149 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5150 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5152 for (unsigned j = 0; j != BytesPerElement; ++j)
5153 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5157 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5158 &ResultMask[0], ResultMask.size());
5159 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5162 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5163 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5164 /// information about the intrinsic.
5165 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5167 unsigned IntrinsicID =
5168 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5171 switch (IntrinsicID) {
5172 default: return false;
5173 // Comparison predicates.
5174 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5175 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5176 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5177 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5178 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5179 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5180 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5181 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5182 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5183 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5184 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5185 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5186 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5188 // Normal Comparisons.
5189 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5190 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5191 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5192 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5193 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5194 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5195 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5196 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5197 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5198 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5199 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5200 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5201 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5206 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5207 /// lower, do it, otherwise return null.
5208 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5209 SelectionDAG &DAG) const {
5210 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5211 // opcode number of the comparison.
5212 DebugLoc dl = Op.getDebugLoc();
5215 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5216 return SDValue(); // Don't custom lower most intrinsics.
5218 // If this is a non-dot comparison, make the VCMP node and we are done.
5220 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5221 Op.getOperand(1), Op.getOperand(2),
5222 DAG.getConstant(CompareOpc, MVT::i32));
5223 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5226 // Create the PPCISD altivec 'dot' comparison node.
5228 Op.getOperand(2), // LHS
5229 Op.getOperand(3), // RHS
5230 DAG.getConstant(CompareOpc, MVT::i32)
5232 std::vector<EVT> VTs;
5233 VTs.push_back(Op.getOperand(2).getValueType());
5234 VTs.push_back(MVT::Glue);
5235 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5237 // Now that we have the comparison, emit a copy from the CR to a GPR.
5238 // This is flagged to the above dot comparison.
5239 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5240 DAG.getRegister(PPC::CR6, MVT::i32),
5241 CompNode.getValue(1));
5243 // Unpack the result based on how the target uses it.
5244 unsigned BitNo; // Bit # of CR6.
5245 bool InvertBit; // Invert result?
5246 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5247 default: // Can't happen, don't crash on invalid number though.
5248 case 0: // Return the value of the EQ bit of CR6.
5249 BitNo = 0; InvertBit = false;
5251 case 1: // Return the inverted value of the EQ bit of CR6.
5252 BitNo = 0; InvertBit = true;
5254 case 2: // Return the value of the LT bit of CR6.
5255 BitNo = 2; InvertBit = false;
5257 case 3: // Return the inverted value of the LT bit of CR6.
5258 BitNo = 2; InvertBit = true;
5262 // Shift the bit into the low position.
5263 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5264 DAG.getConstant(8-(3-BitNo), MVT::i32));
5266 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5267 DAG.getConstant(1, MVT::i32));
5269 // If we are supposed to, toggle the bit.
5271 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5272 DAG.getConstant(1, MVT::i32));
5276 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5277 SelectionDAG &DAG) const {
5278 DebugLoc dl = Op.getDebugLoc();
5279 // Create a stack slot that is 16-byte aligned.
5280 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5281 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5282 EVT PtrVT = getPointerTy();
5283 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5285 // Store the input value into Value#0 of the stack slot.
5286 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5287 Op.getOperand(0), FIdx, MachinePointerInfo(),
5290 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5291 false, false, false, 0);
5294 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5295 DebugLoc dl = Op.getDebugLoc();
5296 if (Op.getValueType() == MVT::v4i32) {
5297 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5299 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5300 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5302 SDValue RHSSwap = // = vrlw RHS, 16
5303 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5305 // Shrinkify inputs to v8i16.
5306 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5307 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5308 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5310 // Low parts multiplied together, generating 32-bit results (we ignore the
5312 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5313 LHS, RHS, DAG, dl, MVT::v4i32);
5315 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5316 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5317 // Shift the high parts up 16 bits.
5318 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5320 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5321 } else if (Op.getValueType() == MVT::v8i16) {
5322 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5324 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5326 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5327 LHS, RHS, Zero, DAG, dl);
5328 } else if (Op.getValueType() == MVT::v16i8) {
5329 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5331 // Multiply the even 8-bit parts, producing 16-bit sums.
5332 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5333 LHS, RHS, DAG, dl, MVT::v8i16);
5334 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5336 // Multiply the odd 8-bit parts, producing 16-bit sums.
5337 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5338 LHS, RHS, DAG, dl, MVT::v8i16);
5339 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5341 // Merge the results together.
5343 for (unsigned i = 0; i != 8; ++i) {
5345 Ops[i*2+1] = 2*i+1+16;
5347 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5349 llvm_unreachable("Unknown mul to lower!");
5353 /// LowerOperation - Provide custom lowering hooks for some operations.
5355 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5356 switch (Op.getOpcode()) {
5357 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5358 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5359 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5360 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5361 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5362 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5363 case ISD::SETCC: return LowerSETCC(Op, DAG);
5364 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5365 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5367 return LowerVASTART(Op, DAG, PPCSubTarget);
5370 return LowerVAARG(Op, DAG, PPCSubTarget);
5372 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5373 case ISD::DYNAMIC_STACKALLOC:
5374 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5376 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5377 case ISD::FP_TO_UINT:
5378 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5380 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5381 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5383 // Lower 64-bit shifts.
5384 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5385 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5386 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5388 // Vector-related lowering.
5389 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5390 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5391 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5392 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5393 case ISD::MUL: return LowerMUL(Op, DAG);
5395 // Frame & Return address.
5396 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5397 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5401 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5402 SmallVectorImpl<SDValue>&Results,
5403 SelectionDAG &DAG) const {
5404 const TargetMachine &TM = getTargetMachine();
5405 DebugLoc dl = N->getDebugLoc();
5406 switch (N->getOpcode()) {
5408 llvm_unreachable("Do not know how to custom type legalize this operation!");
5410 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5411 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5414 EVT VT = N->getValueType(0);
5416 if (VT == MVT::i64) {
5417 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5419 Results.push_back(NewNode);
5420 Results.push_back(NewNode.getValue(1));
5424 case ISD::FP_ROUND_INREG: {
5425 assert(N->getValueType(0) == MVT::ppcf128);
5426 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5427 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5428 MVT::f64, N->getOperand(0),
5429 DAG.getIntPtrConstant(0));
5430 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5431 MVT::f64, N->getOperand(0),
5432 DAG.getIntPtrConstant(1));
5434 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5435 // of the long double, and puts FPSCR back the way it was. We do not
5436 // actually model FPSCR.
5437 std::vector<EVT> NodeTys;
5438 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5440 NodeTys.push_back(MVT::f64); // Return register
5441 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5442 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5443 MFFSreg = Result.getValue(0);
5444 InFlag = Result.getValue(1);
5447 NodeTys.push_back(MVT::Glue); // Returns a flag
5448 Ops[0] = DAG.getConstant(31, MVT::i32);
5450 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5451 InFlag = Result.getValue(0);
5454 NodeTys.push_back(MVT::Glue); // Returns a flag
5455 Ops[0] = DAG.getConstant(30, MVT::i32);
5457 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5458 InFlag = Result.getValue(0);
5461 NodeTys.push_back(MVT::f64); // result of add
5462 NodeTys.push_back(MVT::Glue); // Returns a flag
5466 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5467 FPreg = Result.getValue(0);
5468 InFlag = Result.getValue(1);
5471 NodeTys.push_back(MVT::f64);
5472 Ops[0] = DAG.getConstant(1, MVT::i32);
5476 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5477 FPreg = Result.getValue(0);
5479 // We know the low half is about to be thrown away, so just use something
5481 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5485 case ISD::FP_TO_SINT:
5486 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5492 //===----------------------------------------------------------------------===//
5493 // Other Lowering Code
5494 //===----------------------------------------------------------------------===//
5497 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5498 bool is64bit, unsigned BinOpcode) const {
5499 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5500 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5502 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5503 MachineFunction *F = BB->getParent();
5504 MachineFunction::iterator It = BB;
5507 unsigned dest = MI->getOperand(0).getReg();
5508 unsigned ptrA = MI->getOperand(1).getReg();
5509 unsigned ptrB = MI->getOperand(2).getReg();
5510 unsigned incr = MI->getOperand(3).getReg();
5511 DebugLoc dl = MI->getDebugLoc();
5513 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5514 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5515 F->insert(It, loopMBB);
5516 F->insert(It, exitMBB);
5517 exitMBB->splice(exitMBB->begin(), BB,
5518 llvm::next(MachineBasicBlock::iterator(MI)),
5520 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5522 MachineRegisterInfo &RegInfo = F->getRegInfo();
5523 unsigned TmpReg = (!BinOpcode) ? incr :
5524 RegInfo.createVirtualRegister(
5525 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5526 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5530 // fallthrough --> loopMBB
5531 BB->addSuccessor(loopMBB);
5534 // l[wd]arx dest, ptr
5535 // add r0, dest, incr
5536 // st[wd]cx. r0, ptr
5538 // fallthrough --> exitMBB
5540 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5541 .addReg(ptrA).addReg(ptrB);
5543 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5544 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5545 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5546 BuildMI(BB, dl, TII->get(PPC::BCC))
5547 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5548 BB->addSuccessor(loopMBB);
5549 BB->addSuccessor(exitMBB);
5558 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5559 MachineBasicBlock *BB,
5560 bool is8bit, // operation
5561 unsigned BinOpcode) const {
5562 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5563 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5564 // In 64 bit mode we have to use 64 bits for addresses, even though the
5565 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5566 // registers without caring whether they're 32 or 64, but here we're
5567 // doing actual arithmetic on the addresses.
5568 bool is64bit = PPCSubTarget.isPPC64();
5569 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5571 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5572 MachineFunction *F = BB->getParent();
5573 MachineFunction::iterator It = BB;
5576 unsigned dest = MI->getOperand(0).getReg();
5577 unsigned ptrA = MI->getOperand(1).getReg();
5578 unsigned ptrB = MI->getOperand(2).getReg();
5579 unsigned incr = MI->getOperand(3).getReg();
5580 DebugLoc dl = MI->getDebugLoc();
5582 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5583 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5584 F->insert(It, loopMBB);
5585 F->insert(It, exitMBB);
5586 exitMBB->splice(exitMBB->begin(), BB,
5587 llvm::next(MachineBasicBlock::iterator(MI)),
5589 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5591 MachineRegisterInfo &RegInfo = F->getRegInfo();
5592 const TargetRegisterClass *RC =
5593 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5594 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5595 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5596 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5597 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5598 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5599 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5600 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5601 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5602 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5603 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5604 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5605 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5607 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5611 // fallthrough --> loopMBB
5612 BB->addSuccessor(loopMBB);
5614 // The 4-byte load must be aligned, while a char or short may be
5615 // anywhere in the word. Hence all this nasty bookkeeping code.
5616 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5617 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5618 // xori shift, shift1, 24 [16]
5619 // rlwinm ptr, ptr1, 0, 0, 29
5620 // slw incr2, incr, shift
5621 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5622 // slw mask, mask2, shift
5624 // lwarx tmpDest, ptr
5625 // add tmp, tmpDest, incr2
5626 // andc tmp2, tmpDest, mask
5627 // and tmp3, tmp, mask
5628 // or tmp4, tmp3, tmp2
5631 // fallthrough --> exitMBB
5632 // srw dest, tmpDest, shift
5633 if (ptrA != ZeroReg) {
5634 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5635 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5636 .addReg(ptrA).addReg(ptrB);
5640 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5641 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5642 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5643 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5645 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5646 .addReg(Ptr1Reg).addImm(0).addImm(61);
5648 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5649 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5650 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5651 .addReg(incr).addReg(ShiftReg);
5653 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5655 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5656 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5658 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5659 .addReg(Mask2Reg).addReg(ShiftReg);
5662 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5663 .addReg(ZeroReg).addReg(PtrReg);
5665 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5666 .addReg(Incr2Reg).addReg(TmpDestReg);
5667 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5668 .addReg(TmpDestReg).addReg(MaskReg);
5669 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5670 .addReg(TmpReg).addReg(MaskReg);
5671 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5672 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5673 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5674 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5675 BuildMI(BB, dl, TII->get(PPC::BCC))
5676 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5677 BB->addSuccessor(loopMBB);
5678 BB->addSuccessor(exitMBB);
5683 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5689 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5690 MachineBasicBlock *BB) const {
5691 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5693 // To "insert" these instructions we actually have to insert their
5694 // control-flow patterns.
5695 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5696 MachineFunction::iterator It = BB;
5699 MachineFunction *F = BB->getParent();
5701 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5702 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5703 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5704 PPC::ISEL8 : PPC::ISEL;
5705 unsigned SelectPred = MI->getOperand(4).getImm();
5706 DebugLoc dl = MI->getDebugLoc();
5708 // The SelectPred is ((BI << 5) | BO) for a BCC
5709 unsigned BO = SelectPred & 0xF;
5710 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5712 unsigned TrueOpNo, FalseOpNo;
5719 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5722 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5723 .addReg(MI->getOperand(TrueOpNo).getReg())
5724 .addReg(MI->getOperand(FalseOpNo).getReg())
5725 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5726 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5727 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5728 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5729 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5730 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5733 // The incoming instruction knows the destination vreg to set, the
5734 // condition code register to branch on, the true/false values to
5735 // select between, and a branch opcode to use.
5740 // cmpTY ccX, r1, r2
5742 // fallthrough --> copy0MBB
5743 MachineBasicBlock *thisMBB = BB;
5744 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5745 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5746 unsigned SelectPred = MI->getOperand(4).getImm();
5747 DebugLoc dl = MI->getDebugLoc();
5748 F->insert(It, copy0MBB);
5749 F->insert(It, sinkMBB);
5751 // Transfer the remainder of BB and its successor edges to sinkMBB.
5752 sinkMBB->splice(sinkMBB->begin(), BB,
5753 llvm::next(MachineBasicBlock::iterator(MI)),
5755 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5757 // Next, add the true and fallthrough blocks as its successors.
5758 BB->addSuccessor(copy0MBB);
5759 BB->addSuccessor(sinkMBB);
5761 BuildMI(BB, dl, TII->get(PPC::BCC))
5762 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5765 // %FalseValue = ...
5766 // # fallthrough to sinkMBB
5769 // Update machine-CFG edges
5770 BB->addSuccessor(sinkMBB);
5773 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5776 BuildMI(*BB, BB->begin(), dl,
5777 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5778 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5779 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5781 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5782 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5783 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5784 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5785 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5786 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5787 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5788 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5790 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5791 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5792 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5793 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5794 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5795 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5796 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5797 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5799 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5800 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5801 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5802 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5803 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5804 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5805 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5806 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5808 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5809 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5810 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5811 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5812 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5813 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5814 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5815 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5817 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5818 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5819 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5820 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5821 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5822 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5823 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5824 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5826 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5827 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5828 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5829 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5830 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5831 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5832 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5833 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5835 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5836 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5837 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5838 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5839 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5840 BB = EmitAtomicBinary(MI, BB, false, 0);
5841 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5842 BB = EmitAtomicBinary(MI, BB, true, 0);
5844 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5845 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5846 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5848 unsigned dest = MI->getOperand(0).getReg();
5849 unsigned ptrA = MI->getOperand(1).getReg();
5850 unsigned ptrB = MI->getOperand(2).getReg();
5851 unsigned oldval = MI->getOperand(3).getReg();
5852 unsigned newval = MI->getOperand(4).getReg();
5853 DebugLoc dl = MI->getDebugLoc();
5855 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5856 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5857 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5858 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5859 F->insert(It, loop1MBB);
5860 F->insert(It, loop2MBB);
5861 F->insert(It, midMBB);
5862 F->insert(It, exitMBB);
5863 exitMBB->splice(exitMBB->begin(), BB,
5864 llvm::next(MachineBasicBlock::iterator(MI)),
5866 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5870 // fallthrough --> loopMBB
5871 BB->addSuccessor(loop1MBB);
5874 // l[wd]arx dest, ptr
5875 // cmp[wd] dest, oldval
5878 // st[wd]cx. newval, ptr
5882 // st[wd]cx. dest, ptr
5885 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5886 .addReg(ptrA).addReg(ptrB);
5887 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5888 .addReg(oldval).addReg(dest);
5889 BuildMI(BB, dl, TII->get(PPC::BCC))
5890 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5891 BB->addSuccessor(loop2MBB);
5892 BB->addSuccessor(midMBB);
5895 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5896 .addReg(newval).addReg(ptrA).addReg(ptrB);
5897 BuildMI(BB, dl, TII->get(PPC::BCC))
5898 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5899 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5900 BB->addSuccessor(loop1MBB);
5901 BB->addSuccessor(exitMBB);
5904 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5905 .addReg(dest).addReg(ptrA).addReg(ptrB);
5906 BB->addSuccessor(exitMBB);
5911 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
5912 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
5913 // We must use 64-bit registers for addresses when targeting 64-bit,
5914 // since we're actually doing arithmetic on them. Other registers
5916 bool is64bit = PPCSubTarget.isPPC64();
5917 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
5919 unsigned dest = MI->getOperand(0).getReg();
5920 unsigned ptrA = MI->getOperand(1).getReg();
5921 unsigned ptrB = MI->getOperand(2).getReg();
5922 unsigned oldval = MI->getOperand(3).getReg();
5923 unsigned newval = MI->getOperand(4).getReg();
5924 DebugLoc dl = MI->getDebugLoc();
5926 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5927 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5928 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5929 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5930 F->insert(It, loop1MBB);
5931 F->insert(It, loop2MBB);
5932 F->insert(It, midMBB);
5933 F->insert(It, exitMBB);
5934 exitMBB->splice(exitMBB->begin(), BB,
5935 llvm::next(MachineBasicBlock::iterator(MI)),
5937 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5939 MachineRegisterInfo &RegInfo = F->getRegInfo();
5940 const TargetRegisterClass *RC =
5941 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5942 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5943 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5944 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5945 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5946 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
5947 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
5948 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
5949 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
5950 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5951 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5952 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5953 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5954 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5955 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5957 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
5958 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5961 // fallthrough --> loopMBB
5962 BB->addSuccessor(loop1MBB);
5964 // The 4-byte load must be aligned, while a char or short may be
5965 // anywhere in the word. Hence all this nasty bookkeeping code.
5966 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5967 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5968 // xori shift, shift1, 24 [16]
5969 // rlwinm ptr, ptr1, 0, 0, 29
5970 // slw newval2, newval, shift
5971 // slw oldval2, oldval,shift
5972 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5973 // slw mask, mask2, shift
5974 // and newval3, newval2, mask
5975 // and oldval3, oldval2, mask
5977 // lwarx tmpDest, ptr
5978 // and tmp, tmpDest, mask
5979 // cmpw tmp, oldval3
5982 // andc tmp2, tmpDest, mask
5983 // or tmp4, tmp2, newval3
5988 // stwcx. tmpDest, ptr
5990 // srw dest, tmpDest, shift
5991 if (ptrA != ZeroReg) {
5992 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5993 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5994 .addReg(ptrA).addReg(ptrB);
5998 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5999 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6000 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6001 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6003 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6004 .addReg(Ptr1Reg).addImm(0).addImm(61);
6006 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6007 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6008 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6009 .addReg(newval).addReg(ShiftReg);
6010 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6011 .addReg(oldval).addReg(ShiftReg);
6013 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6015 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6016 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6017 .addReg(Mask3Reg).addImm(65535);
6019 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6020 .addReg(Mask2Reg).addReg(ShiftReg);
6021 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6022 .addReg(NewVal2Reg).addReg(MaskReg);
6023 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6024 .addReg(OldVal2Reg).addReg(MaskReg);
6027 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6028 .addReg(ZeroReg).addReg(PtrReg);
6029 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6030 .addReg(TmpDestReg).addReg(MaskReg);
6031 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6032 .addReg(TmpReg).addReg(OldVal3Reg);
6033 BuildMI(BB, dl, TII->get(PPC::BCC))
6034 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6035 BB->addSuccessor(loop2MBB);
6036 BB->addSuccessor(midMBB);
6039 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6040 .addReg(TmpDestReg).addReg(MaskReg);
6041 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6042 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6043 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6044 .addReg(ZeroReg).addReg(PtrReg);
6045 BuildMI(BB, dl, TII->get(PPC::BCC))
6046 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6047 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6048 BB->addSuccessor(loop1MBB);
6049 BB->addSuccessor(exitMBB);
6052 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6053 .addReg(ZeroReg).addReg(PtrReg);
6054 BB->addSuccessor(exitMBB);
6059 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6062 llvm_unreachable("Unexpected instr type to insert");
6065 MI->eraseFromParent(); // The pseudo instruction is gone now.
6069 //===----------------------------------------------------------------------===//
6070 // Target Optimization Hooks
6071 //===----------------------------------------------------------------------===//
6073 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6074 DAGCombinerInfo &DCI) const {
6075 const TargetMachine &TM = getTargetMachine();
6076 SelectionDAG &DAG = DCI.DAG;
6077 DebugLoc dl = N->getDebugLoc();
6078 switch (N->getOpcode()) {
6081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6082 if (C->isNullValue()) // 0 << V -> 0.
6083 return N->getOperand(0);
6087 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6088 if (C->isNullValue()) // 0 >>u V -> 0.
6089 return N->getOperand(0);
6093 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6094 if (C->isNullValue() || // 0 >>s V -> 0.
6095 C->isAllOnesValue()) // -1 >>s V -> -1.
6096 return N->getOperand(0);
6100 case ISD::SINT_TO_FP:
6101 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6102 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6103 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6104 // We allow the src/dst to be either f32/f64, but the intermediate
6105 // type must be i64.
6106 if (N->getOperand(0).getValueType() == MVT::i64 &&
6107 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6108 SDValue Val = N->getOperand(0).getOperand(0);
6109 if (Val.getValueType() == MVT::f32) {
6110 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6111 DCI.AddToWorklist(Val.getNode());
6114 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6115 DCI.AddToWorklist(Val.getNode());
6116 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6117 DCI.AddToWorklist(Val.getNode());
6118 if (N->getValueType(0) == MVT::f32) {
6119 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6120 DAG.getIntPtrConstant(0));
6121 DCI.AddToWorklist(Val.getNode());
6124 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6125 // If the intermediate type is i32, we can avoid the load/store here
6132 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6133 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6134 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6135 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6136 N->getOperand(1).getValueType() == MVT::i32 &&
6137 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6138 SDValue Val = N->getOperand(1).getOperand(0);
6139 if (Val.getValueType() == MVT::f32) {
6140 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6141 DCI.AddToWorklist(Val.getNode());
6143 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6144 DCI.AddToWorklist(Val.getNode());
6146 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6147 N->getOperand(2), N->getOperand(3));
6148 DCI.AddToWorklist(Val.getNode());
6152 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6153 if (cast<StoreSDNode>(N)->isUnindexed() &&
6154 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6155 N->getOperand(1).getNode()->hasOneUse() &&
6156 (N->getOperand(1).getValueType() == MVT::i32 ||
6157 N->getOperand(1).getValueType() == MVT::i16)) {
6158 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6159 // Do an any-extend to 32-bits if this is a half-word input.
6160 if (BSwapOp.getValueType() == MVT::i16)
6161 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6164 N->getOperand(0), BSwapOp, N->getOperand(2),
6165 DAG.getValueType(N->getOperand(1).getValueType())
6168 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6169 Ops, array_lengthof(Ops),
6170 cast<StoreSDNode>(N)->getMemoryVT(),
6171 cast<StoreSDNode>(N)->getMemOperand());
6175 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6176 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6177 N->getOperand(0).hasOneUse() &&
6178 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6179 SDValue Load = N->getOperand(0);
6180 LoadSDNode *LD = cast<LoadSDNode>(Load);
6181 // Create the byte-swapping load.
6183 LD->getChain(), // Chain
6184 LD->getBasePtr(), // Ptr
6185 DAG.getValueType(N->getValueType(0)) // VT
6188 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6189 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6190 LD->getMemoryVT(), LD->getMemOperand());
6192 // If this is an i16 load, insert the truncate.
6193 SDValue ResVal = BSLoad;
6194 if (N->getValueType(0) == MVT::i16)
6195 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6197 // First, combine the bswap away. This makes the value produced by the
6199 DCI.CombineTo(N, ResVal);
6201 // Next, combine the load away, we give it a bogus result value but a real
6202 // chain result. The result value is dead because the bswap is dead.
6203 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6205 // Return N so it doesn't get rechecked!
6206 return SDValue(N, 0);
6210 case PPCISD::VCMP: {
6211 // If a VCMPo node already exists with exactly the same operands as this
6212 // node, use its result instead of this node (VCMPo computes both a CR6 and
6213 // a normal output).
6215 if (!N->getOperand(0).hasOneUse() &&
6216 !N->getOperand(1).hasOneUse() &&
6217 !N->getOperand(2).hasOneUse()) {
6219 // Scan all of the users of the LHS, looking for VCMPo's that match.
6220 SDNode *VCMPoNode = 0;
6222 SDNode *LHSN = N->getOperand(0).getNode();
6223 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6225 if (UI->getOpcode() == PPCISD::VCMPo &&
6226 UI->getOperand(1) == N->getOperand(1) &&
6227 UI->getOperand(2) == N->getOperand(2) &&
6228 UI->getOperand(0) == N->getOperand(0)) {
6233 // If there is no VCMPo node, or if the flag value has a single use, don't
6235 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6238 // Look at the (necessarily single) use of the flag value. If it has a
6239 // chain, this transformation is more complex. Note that multiple things
6240 // could use the value result, which we should ignore.
6241 SDNode *FlagUser = 0;
6242 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6243 FlagUser == 0; ++UI) {
6244 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6246 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6247 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6254 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6255 // give up for right now.
6256 if (FlagUser->getOpcode() == PPCISD::MFCR)
6257 return SDValue(VCMPoNode, 0);
6262 // If this is a branch on an altivec predicate comparison, lower this so
6263 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6264 // lowering is done pre-legalize, because the legalizer lowers the predicate
6265 // compare down to code that is difficult to reassemble.
6266 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6267 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6271 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6272 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6273 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6274 assert(isDot && "Can't compare against a vector result!");
6276 // If this is a comparison against something other than 0/1, then we know
6277 // that the condition is never/always true.
6278 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6279 if (Val != 0 && Val != 1) {
6280 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6281 return N->getOperand(0);
6282 // Always !=, turn it into an unconditional branch.
6283 return DAG.getNode(ISD::BR, dl, MVT::Other,
6284 N->getOperand(0), N->getOperand(4));
6287 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6289 // Create the PPCISD altivec 'dot' comparison node.
6290 std::vector<EVT> VTs;
6292 LHS.getOperand(2), // LHS of compare
6293 LHS.getOperand(3), // RHS of compare
6294 DAG.getConstant(CompareOpc, MVT::i32)
6296 VTs.push_back(LHS.getOperand(2).getValueType());
6297 VTs.push_back(MVT::Glue);
6298 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6300 // Unpack the result based on how the target uses it.
6301 PPC::Predicate CompOpc;
6302 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6303 default: // Can't happen, don't crash on invalid number though.
6304 case 0: // Branch on the value of the EQ bit of CR6.
6305 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6307 case 1: // Branch on the inverted value of the EQ bit of CR6.
6308 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6310 case 2: // Branch on the value of the LT bit of CR6.
6311 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6313 case 3: // Branch on the inverted value of the LT bit of CR6.
6314 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6318 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6319 DAG.getConstant(CompOpc, MVT::i32),
6320 DAG.getRegister(PPC::CR6, MVT::i32),
6321 N->getOperand(4), CompNode.getValue(1));
6330 //===----------------------------------------------------------------------===//
6331 // Inline Assembly Support
6332 //===----------------------------------------------------------------------===//
6334 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6337 const SelectionDAG &DAG,
6338 unsigned Depth) const {
6339 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6340 switch (Op.getOpcode()) {
6342 case PPCISD::LBRX: {
6343 // lhbrx is known to have the top bits cleared out.
6344 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6345 KnownZero = 0xFFFF0000;
6348 case ISD::INTRINSIC_WO_CHAIN: {
6349 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6351 case Intrinsic::ppc_altivec_vcmpbfp_p:
6352 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6353 case Intrinsic::ppc_altivec_vcmpequb_p:
6354 case Intrinsic::ppc_altivec_vcmpequh_p:
6355 case Intrinsic::ppc_altivec_vcmpequw_p:
6356 case Intrinsic::ppc_altivec_vcmpgefp_p:
6357 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6358 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6359 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6360 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6361 case Intrinsic::ppc_altivec_vcmpgtub_p:
6362 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6363 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6364 KnownZero = ~1U; // All bits but the low one are known to be zero.
6372 /// getConstraintType - Given a constraint, return the type of
6373 /// constraint it is for this target.
6374 PPCTargetLowering::ConstraintType
6375 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6376 if (Constraint.size() == 1) {
6377 switch (Constraint[0]) {
6384 return C_RegisterClass;
6387 return TargetLowering::getConstraintType(Constraint);
6390 /// Examine constraint type and operand type and determine a weight value.
6391 /// This object must already have been set up with the operand type
6392 /// and the current alternative constraint selected.
6393 TargetLowering::ConstraintWeight
6394 PPCTargetLowering::getSingleConstraintMatchWeight(
6395 AsmOperandInfo &info, const char *constraint) const {
6396 ConstraintWeight weight = CW_Invalid;
6397 Value *CallOperandVal = info.CallOperandVal;
6398 // If we don't have a value, we can't do a match,
6399 // but allow it at the lowest weight.
6400 if (CallOperandVal == NULL)
6402 Type *type = CallOperandVal->getType();
6403 // Look at the constraint type.
6404 switch (*constraint) {
6406 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6409 if (type->isIntegerTy())
6410 weight = CW_Register;
6413 if (type->isFloatTy())
6414 weight = CW_Register;
6417 if (type->isDoubleTy())
6418 weight = CW_Register;
6421 if (type->isVectorTy())
6422 weight = CW_Register;
6425 weight = CW_Register;
6431 std::pair<unsigned, const TargetRegisterClass*>
6432 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6434 if (Constraint.size() == 1) {
6435 // GCC RS6000 Constraint Letters
6436 switch (Constraint[0]) {
6439 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6440 return std::make_pair(0U, &PPC::G8RCRegClass);
6441 return std::make_pair(0U, &PPC::GPRCRegClass);
6444 return std::make_pair(0U, &PPC::F4RCRegClass);
6446 return std::make_pair(0U, &PPC::F8RCRegClass);
6449 return std::make_pair(0U, &PPC::VRRCRegClass);
6451 return std::make_pair(0U, &PPC::CRRCRegClass);
6455 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6459 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6460 /// vector. If it is invalid, don't add anything to Ops.
6461 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6462 std::string &Constraint,
6463 std::vector<SDValue>&Ops,
6464 SelectionDAG &DAG) const {
6465 SDValue Result(0,0);
6467 // Only support length 1 constraints.
6468 if (Constraint.length() > 1) return;
6470 char Letter = Constraint[0];
6481 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6482 if (!CST) return; // Must be an immediate to match.
6483 unsigned Value = CST->getZExtValue();
6485 default: llvm_unreachable("Unknown constraint letter!");
6486 case 'I': // "I" is a signed 16-bit constant.
6487 if ((short)Value == (int)Value)
6488 Result = DAG.getTargetConstant(Value, Op.getValueType());
6490 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6491 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6492 if ((short)Value == 0)
6493 Result = DAG.getTargetConstant(Value, Op.getValueType());
6495 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6496 if ((Value >> 16) == 0)
6497 Result = DAG.getTargetConstant(Value, Op.getValueType());
6499 case 'M': // "M" is a constant that is greater than 31.
6501 Result = DAG.getTargetConstant(Value, Op.getValueType());
6503 case 'N': // "N" is a positive constant that is an exact power of two.
6504 if ((int)Value > 0 && isPowerOf2_32(Value))
6505 Result = DAG.getTargetConstant(Value, Op.getValueType());
6507 case 'O': // "O" is the constant zero.
6509 Result = DAG.getTargetConstant(Value, Op.getValueType());
6511 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6512 if ((short)-Value == (int)-Value)
6513 Result = DAG.getTargetConstant(Value, Op.getValueType());
6520 if (Result.getNode()) {
6521 Ops.push_back(Result);
6525 // Handle standard constraint letters.
6526 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6529 // isLegalAddressingMode - Return true if the addressing mode represented
6530 // by AM is legal for this target, for a load/store of the specified type.
6531 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6533 // FIXME: PPC does not allow r+i addressing modes for vectors!
6535 // PPC allows a sign-extended 16-bit immediate field.
6536 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6539 // No global is ever allowed as a base.
6543 // PPC only support r+r,
6545 case 0: // "r+i" or just "i", depending on HasBaseReg.
6548 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6550 // Otherwise we have r+r or r+i.
6553 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6555 // Allow 2*r as r+r.
6558 // No other scales are supported.
6565 /// isLegalAddressImmediate - Return true if the integer value can be used
6566 /// as the offset of the target addressing mode for load / store of the
6568 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6569 // PPC allows a sign-extended 16-bit immediate field.
6570 return (V > -(1 << 16) && V < (1 << 16)-1);
6573 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6577 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6578 SelectionDAG &DAG) const {
6579 MachineFunction &MF = DAG.getMachineFunction();
6580 MachineFrameInfo *MFI = MF.getFrameInfo();
6581 MFI->setReturnAddressIsTaken(true);
6583 DebugLoc dl = Op.getDebugLoc();
6584 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6586 // Make sure the function does not optimize away the store of the RA to
6588 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6589 FuncInfo->setLRStoreRequired();
6590 bool isPPC64 = PPCSubTarget.isPPC64();
6591 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6594 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6597 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6598 isPPC64? MVT::i64 : MVT::i32);
6599 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6600 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6602 MachinePointerInfo(), false, false, false, 0);
6605 // Just load the return address off the stack.
6606 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6607 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6608 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6611 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6612 SelectionDAG &DAG) const {
6613 DebugLoc dl = Op.getDebugLoc();
6614 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6616 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6617 bool isPPC64 = PtrVT == MVT::i64;
6619 MachineFunction &MF = DAG.getMachineFunction();
6620 MachineFrameInfo *MFI = MF.getFrameInfo();
6621 MFI->setFrameAddressIsTaken(true);
6622 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6623 MFI->hasVarSizedObjects()) &&
6624 MFI->getStackSize() &&
6625 !MF.getFunction()->getFnAttributes().
6626 hasAttribute(Attributes::Naked);
6627 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6628 (is31 ? PPC::R31 : PPC::R1);
6629 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6632 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6633 FrameAddr, MachinePointerInfo(), false, false,
6639 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6640 // The PowerPC target isn't yet aware of offsets.
6644 /// getOptimalMemOpType - Returns the target specific optimal type for load
6645 /// and store operations as a result of memset, memcpy, and memmove
6646 /// lowering. If DstAlign is zero that means it's safe to destination
6647 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6648 /// means there isn't a need to check it against alignment requirement,
6649 /// probably because the source does not need to be loaded. If
6650 /// 'IsZeroVal' is true, that means it's safe to return a
6651 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6652 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6653 /// constant so it does not need to be loaded.
6654 /// It returns EVT::Other if the type should be determined using generic
6655 /// target-independent logic.
6656 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6657 unsigned DstAlign, unsigned SrcAlign,
6660 MachineFunction &MF) const {
6661 if (this->PPCSubTarget.isPPC64()) {
6668 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6669 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6670 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6671 /// is expanded to mul + add.
6672 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6676 switch (VT.getSimpleVT().SimpleTy) {
6688 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6690 return TargetLowering::getSchedulingPreference(N);