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);
364 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
366 for (unsigned j = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
367 j <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++j) {
368 MVT::SimpleValueType InnerVT = (MVT::SimpleValueType)j;
369 setTruncStoreAction(VT, InnerVT, Expand);
371 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
372 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
373 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
376 for (unsigned i = (unsigned)MVT::FIRST_FP_VECTOR_VALUETYPE;
377 i <= (unsigned)MVT::LAST_FP_VECTOR_VALUETYPE; ++i) {
378 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
379 setOperationAction(ISD::FSQRT, VT, Expand);
382 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
383 // with merges, splats, etc.
384 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
386 setOperationAction(ISD::AND , MVT::v4i32, Legal);
387 setOperationAction(ISD::OR , MVT::v4i32, Legal);
388 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
389 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
390 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
391 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
392 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
393 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
394 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
395 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
397 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
398 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
399 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
400 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
402 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
403 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
404 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
405 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
406 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
408 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
409 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
411 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
412 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
413 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
414 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
416 // Altivec does not contain unordered floating-point compare instructions
417 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
418 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
419 setCondCodeAction(ISD::SETUGT, MVT::v4f32, Expand);
420 setCondCodeAction(ISD::SETUGE, MVT::v4f32, Expand);
421 setCondCodeAction(ISD::SETULT, MVT::v4f32, Expand);
422 setCondCodeAction(ISD::SETULE, MVT::v4f32, Expand);
425 if (Subtarget->has64BitSupport()) {
426 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
427 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
430 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
431 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
433 setBooleanContents(ZeroOrOneBooleanContent);
434 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
437 setStackPointerRegisterToSaveRestore(PPC::X1);
438 setExceptionPointerRegister(PPC::X3);
439 setExceptionSelectorRegister(PPC::X4);
441 setStackPointerRegisterToSaveRestore(PPC::R1);
442 setExceptionPointerRegister(PPC::R3);
443 setExceptionSelectorRegister(PPC::R4);
446 // We have target-specific dag combine patterns for the following nodes:
447 setTargetDAGCombine(ISD::SINT_TO_FP);
448 setTargetDAGCombine(ISD::STORE);
449 setTargetDAGCombine(ISD::BR_CC);
450 setTargetDAGCombine(ISD::BSWAP);
452 // Darwin long double math library functions have $LDBL128 appended.
453 if (Subtarget->isDarwin()) {
454 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
455 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
456 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
457 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
458 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
459 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
460 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
461 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
462 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
463 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
466 setMinFunctionAlignment(2);
467 if (PPCSubTarget.isDarwin())
468 setPrefFunctionAlignment(4);
470 if (isPPC64 && Subtarget->isJITCodeModel())
471 // Temporary workaround for the inability of PPC64 JIT to handle jump
473 setSupportJumpTables(false);
475 setInsertFencesForAtomic(true);
477 setSchedulingPreference(Sched::Hybrid);
479 computeRegisterProperties();
481 // The Freescale cores does better with aggressive inlining of memcpy and
482 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
483 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
484 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
485 maxStoresPerMemset = 32;
486 maxStoresPerMemsetOptSize = 16;
487 maxStoresPerMemcpy = 32;
488 maxStoresPerMemcpyOptSize = 8;
489 maxStoresPerMemmove = 32;
490 maxStoresPerMemmoveOptSize = 8;
492 setPrefFunctionAlignment(4);
493 benefitFromCodePlacementOpt = true;
497 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
498 /// function arguments in the caller parameter area.
499 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
500 const TargetMachine &TM = getTargetMachine();
501 // Darwin passes everything on 4 byte boundary.
502 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
505 // 16byte and wider vectors are passed on 16byte boundary.
506 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
507 if (VTy->getBitWidth() >= 128)
510 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
511 if (PPCSubTarget.isPPC64())
517 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
520 case PPCISD::FSEL: return "PPCISD::FSEL";
521 case PPCISD::FCFID: return "PPCISD::FCFID";
522 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
523 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
524 case PPCISD::STFIWX: return "PPCISD::STFIWX";
525 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
526 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
527 case PPCISD::VPERM: return "PPCISD::VPERM";
528 case PPCISD::Hi: return "PPCISD::Hi";
529 case PPCISD::Lo: return "PPCISD::Lo";
530 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
531 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
532 case PPCISD::LOAD: return "PPCISD::LOAD";
533 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
534 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
535 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
536 case PPCISD::SRL: return "PPCISD::SRL";
537 case PPCISD::SRA: return "PPCISD::SRA";
538 case PPCISD::SHL: return "PPCISD::SHL";
539 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
540 case PPCISD::STD_32: return "PPCISD::STD_32";
541 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
542 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
543 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
544 case PPCISD::NOP: return "PPCISD::NOP";
545 case PPCISD::MTCTR: return "PPCISD::MTCTR";
546 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
547 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
548 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
549 case PPCISD::MFCR: return "PPCISD::MFCR";
550 case PPCISD::VCMP: return "PPCISD::VCMP";
551 case PPCISD::VCMPo: return "PPCISD::VCMPo";
552 case PPCISD::LBRX: return "PPCISD::LBRX";
553 case PPCISD::STBRX: return "PPCISD::STBRX";
554 case PPCISD::LARX: return "PPCISD::LARX";
555 case PPCISD::STCX: return "PPCISD::STCX";
556 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
557 case PPCISD::MFFS: return "PPCISD::MFFS";
558 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
559 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
560 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
561 case PPCISD::MTFSF: return "PPCISD::MTFSF";
562 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
563 case PPCISD::CR6SET: return "PPCISD::CR6SET";
564 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
568 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
571 return VT.changeVectorElementTypeToInteger();
574 //===----------------------------------------------------------------------===//
575 // Node matching predicates, for use by the tblgen matching code.
576 //===----------------------------------------------------------------------===//
578 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
579 static bool isFloatingPointZero(SDValue Op) {
580 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
581 return CFP->getValueAPF().isZero();
582 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
583 // Maybe this has already been legalized into the constant pool?
584 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
585 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
586 return CFP->getValueAPF().isZero();
591 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
592 /// true if Op is undef or if it matches the specified value.
593 static bool isConstantOrUndef(int Op, int Val) {
594 return Op < 0 || Op == Val;
597 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
598 /// VPKUHUM instruction.
599 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
601 for (unsigned i = 0; i != 16; ++i)
602 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
605 for (unsigned i = 0; i != 8; ++i)
606 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
607 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
613 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
614 /// VPKUWUM instruction.
615 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
617 for (unsigned i = 0; i != 16; i += 2)
618 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
619 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
622 for (unsigned i = 0; i != 8; i += 2)
623 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
624 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
625 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
626 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
632 /// isVMerge - Common function, used to match vmrg* shuffles.
634 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
635 unsigned LHSStart, unsigned RHSStart) {
636 assert(N->getValueType(0) == MVT::v16i8 &&
637 "PPC only supports shuffles by bytes!");
638 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
639 "Unsupported merge size!");
641 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
642 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
643 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
644 LHSStart+j+i*UnitSize) ||
645 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
646 RHSStart+j+i*UnitSize))
652 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
653 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
654 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
657 return isVMerge(N, UnitSize, 8, 24);
658 return isVMerge(N, UnitSize, 8, 8);
661 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
662 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
663 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
666 return isVMerge(N, UnitSize, 0, 16);
667 return isVMerge(N, UnitSize, 0, 0);
671 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
672 /// amount, otherwise return -1.
673 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
674 assert(N->getValueType(0) == MVT::v16i8 &&
675 "PPC only supports shuffles by bytes!");
677 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
679 // Find the first non-undef value in the shuffle mask.
681 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
684 if (i == 16) return -1; // all undef.
686 // Otherwise, check to see if the rest of the elements are consecutively
687 // numbered from this value.
688 unsigned ShiftAmt = SVOp->getMaskElt(i);
689 if (ShiftAmt < i) return -1;
693 // Check the rest of the elements to see if they are consecutive.
694 for (++i; i != 16; ++i)
695 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
698 // Check the rest of the elements to see if they are consecutive.
699 for (++i; i != 16; ++i)
700 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
706 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
707 /// specifies a splat of a single element that is suitable for input to
708 /// VSPLTB/VSPLTH/VSPLTW.
709 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
710 assert(N->getValueType(0) == MVT::v16i8 &&
711 (EltSize == 1 || EltSize == 2 || EltSize == 4));
713 // This is a splat operation if each element of the permute is the same, and
714 // if the value doesn't reference the second vector.
715 unsigned ElementBase = N->getMaskElt(0);
717 // FIXME: Handle UNDEF elements too!
718 if (ElementBase >= 16)
721 // Check that the indices are consecutive, in the case of a multi-byte element
722 // splatted with a v16i8 mask.
723 for (unsigned i = 1; i != EltSize; ++i)
724 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
727 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
728 if (N->getMaskElt(i) < 0) continue;
729 for (unsigned j = 0; j != EltSize; ++j)
730 if (N->getMaskElt(i+j) != N->getMaskElt(j))
736 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
738 bool PPC::isAllNegativeZeroVector(SDNode *N) {
739 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
741 APInt APVal, APUndef;
745 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
746 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
747 return CFP->getValueAPF().isNegZero();
752 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
753 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
754 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
755 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
756 assert(isSplatShuffleMask(SVOp, EltSize));
757 return SVOp->getMaskElt(0) / EltSize;
760 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
761 /// by using a vspltis[bhw] instruction of the specified element size, return
762 /// the constant being splatted. The ByteSize field indicates the number of
763 /// bytes of each element [124] -> [bhw].
764 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
767 // If ByteSize of the splat is bigger than the element size of the
768 // build_vector, then we have a case where we are checking for a splat where
769 // multiple elements of the buildvector are folded together into a single
770 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
771 unsigned EltSize = 16/N->getNumOperands();
772 if (EltSize < ByteSize) {
773 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
774 SDValue UniquedVals[4];
775 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
777 // See if all of the elements in the buildvector agree across.
778 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
779 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
780 // If the element isn't a constant, bail fully out.
781 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
784 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
785 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
786 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
787 return SDValue(); // no match.
790 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
791 // either constant or undef values that are identical for each chunk. See
792 // if these chunks can form into a larger vspltis*.
794 // Check to see if all of the leading entries are either 0 or -1. If
795 // neither, then this won't fit into the immediate field.
796 bool LeadingZero = true;
797 bool LeadingOnes = true;
798 for (unsigned i = 0; i != Multiple-1; ++i) {
799 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
801 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
802 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
804 // Finally, check the least significant entry.
806 if (UniquedVals[Multiple-1].getNode() == 0)
807 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
808 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
810 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
813 if (UniquedVals[Multiple-1].getNode() == 0)
814 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
815 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
816 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
817 return DAG.getTargetConstant(Val, MVT::i32);
823 // Check to see if this buildvec has a single non-undef value in its elements.
824 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
825 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
826 if (OpVal.getNode() == 0)
827 OpVal = N->getOperand(i);
828 else if (OpVal != N->getOperand(i))
832 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
834 unsigned ValSizeInBytes = EltSize;
836 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
837 Value = CN->getZExtValue();
838 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
839 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
840 Value = FloatToBits(CN->getValueAPF().convertToFloat());
843 // If the splat value is larger than the element value, then we can never do
844 // this splat. The only case that we could fit the replicated bits into our
845 // immediate field for would be zero, and we prefer to use vxor for it.
846 if (ValSizeInBytes < ByteSize) return SDValue();
848 // If the element value is larger than the splat value, cut it in half and
849 // check to see if the two halves are equal. Continue doing this until we
850 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
851 while (ValSizeInBytes > ByteSize) {
852 ValSizeInBytes >>= 1;
854 // If the top half equals the bottom half, we're still ok.
855 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
856 (Value & ((1 << (8*ValSizeInBytes))-1)))
860 // Properly sign extend the value.
861 int MaskVal = SignExtend32(Value, ByteSize * 8);
863 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
864 if (MaskVal == 0) return SDValue();
866 // Finally, if this value fits in a 5 bit sext field, return it
867 if (SignExtend32<5>(MaskVal) == MaskVal)
868 return DAG.getTargetConstant(MaskVal, MVT::i32);
872 //===----------------------------------------------------------------------===//
873 // Addressing Mode Selection
874 //===----------------------------------------------------------------------===//
876 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
877 /// or 64-bit immediate, and if the value can be accurately represented as a
878 /// sign extension from a 16-bit value. If so, this returns true and the
880 static bool isIntS16Immediate(SDNode *N, short &Imm) {
881 if (N->getOpcode() != ISD::Constant)
884 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
885 if (N->getValueType(0) == MVT::i32)
886 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
888 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
890 static bool isIntS16Immediate(SDValue Op, short &Imm) {
891 return isIntS16Immediate(Op.getNode(), Imm);
895 /// SelectAddressRegReg - Given the specified addressed, check to see if it
896 /// can be represented as an indexed [r+r] operation. Returns false if it
897 /// can be more efficiently represented with [r+imm].
898 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
900 SelectionDAG &DAG) const {
902 if (N.getOpcode() == ISD::ADD) {
903 if (isIntS16Immediate(N.getOperand(1), imm))
905 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
908 Base = N.getOperand(0);
909 Index = N.getOperand(1);
911 } else if (N.getOpcode() == ISD::OR) {
912 if (isIntS16Immediate(N.getOperand(1), imm))
913 return false; // r+i can fold it if we can.
915 // If this is an or of disjoint bitfields, we can codegen this as an add
916 // (for better address arithmetic) if the LHS and RHS of the OR are provably
918 APInt LHSKnownZero, LHSKnownOne;
919 APInt RHSKnownZero, RHSKnownOne;
920 DAG.ComputeMaskedBits(N.getOperand(0),
921 LHSKnownZero, LHSKnownOne);
923 if (LHSKnownZero.getBoolValue()) {
924 DAG.ComputeMaskedBits(N.getOperand(1),
925 RHSKnownZero, RHSKnownOne);
926 // If all of the bits are known zero on the LHS or RHS, the add won't
928 if (~(LHSKnownZero | RHSKnownZero) == 0) {
929 Base = N.getOperand(0);
930 Index = N.getOperand(1);
939 /// Returns true if the address N can be represented by a base register plus
940 /// a signed 16-bit displacement [r+imm], and if it is not better
941 /// represented as reg+reg.
942 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
944 SelectionDAG &DAG) const {
945 // FIXME dl should come from parent load or store, not from address
946 DebugLoc dl = N.getDebugLoc();
947 // If this can be more profitably realized as r+r, fail.
948 if (SelectAddressRegReg(N, Disp, Base, DAG))
951 if (N.getOpcode() == ISD::ADD) {
953 if (isIntS16Immediate(N.getOperand(1), imm)) {
954 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
955 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
956 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
958 Base = N.getOperand(0);
960 return true; // [r+i]
961 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
962 // Match LOAD (ADD (X, Lo(G))).
963 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
964 && "Cannot handle constant offsets yet!");
965 Disp = N.getOperand(1).getOperand(0); // The global address.
966 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
967 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
968 Disp.getOpcode() == ISD::TargetConstantPool ||
969 Disp.getOpcode() == ISD::TargetJumpTable);
970 Base = N.getOperand(0);
971 return true; // [&g+r]
973 } else if (N.getOpcode() == ISD::OR) {
975 if (isIntS16Immediate(N.getOperand(1), imm)) {
976 // If this is an or of disjoint bitfields, we can codegen this as an add
977 // (for better address arithmetic) if the LHS and RHS of the OR are
978 // provably disjoint.
979 APInt LHSKnownZero, LHSKnownOne;
980 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
982 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
983 // If all of the bits are known zero on the LHS or RHS, the add won't
985 Base = N.getOperand(0);
986 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
990 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
991 // Loading from a constant address.
993 // If this address fits entirely in a 16-bit sext immediate field, codegen
996 if (isIntS16Immediate(CN, Imm)) {
997 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
998 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
999 CN->getValueType(0));
1003 // Handle 32-bit sext immediates with LIS + addr mode.
1004 if (CN->getValueType(0) == MVT::i32 ||
1005 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1006 int Addr = (int)CN->getZExtValue();
1008 // Otherwise, break this down into an LIS + disp.
1009 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1011 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1012 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1013 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1018 Disp = DAG.getTargetConstant(0, getPointerTy());
1019 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1020 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1023 return true; // [r+0]
1026 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1027 /// represented as an indexed [r+r] operation.
1028 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1030 SelectionDAG &DAG) const {
1031 // Check to see if we can easily represent this as an [r+r] address. This
1032 // will fail if it thinks that the address is more profitably represented as
1033 // reg+imm, e.g. where imm = 0.
1034 if (SelectAddressRegReg(N, Base, Index, DAG))
1037 // If the operand is an addition, always emit this as [r+r], since this is
1038 // better (for code size, and execution, as the memop does the add for free)
1039 // than emitting an explicit add.
1040 if (N.getOpcode() == ISD::ADD) {
1041 Base = N.getOperand(0);
1042 Index = N.getOperand(1);
1046 // Otherwise, do it the hard way, using R0 as the base register.
1047 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1053 /// SelectAddressRegImmShift - Returns true if the address N can be
1054 /// represented by a base register plus a signed 14-bit displacement
1055 /// [r+imm*4]. Suitable for use by STD and friends.
1056 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1058 SelectionDAG &DAG) const {
1059 // FIXME dl should come from the parent load or store, not the address
1060 DebugLoc dl = N.getDebugLoc();
1061 // If this can be more profitably realized as r+r, fail.
1062 if (SelectAddressRegReg(N, Disp, Base, DAG))
1065 if (N.getOpcode() == ISD::ADD) {
1067 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1068 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1069 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1070 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1072 Base = N.getOperand(0);
1074 return true; // [r+i]
1075 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1076 // Match LOAD (ADD (X, Lo(G))).
1077 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1078 && "Cannot handle constant offsets yet!");
1079 Disp = N.getOperand(1).getOperand(0); // The global address.
1080 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1081 Disp.getOpcode() == ISD::TargetConstantPool ||
1082 Disp.getOpcode() == ISD::TargetJumpTable);
1083 Base = N.getOperand(0);
1084 return true; // [&g+r]
1086 } else if (N.getOpcode() == ISD::OR) {
1088 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1089 // If this is an or of disjoint bitfields, we can codegen this as an add
1090 // (for better address arithmetic) if the LHS and RHS of the OR are
1091 // provably disjoint.
1092 APInt LHSKnownZero, LHSKnownOne;
1093 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1094 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1095 // If all of the bits are known zero on the LHS or RHS, the add won't
1097 Base = N.getOperand(0);
1098 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1102 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1103 // Loading from a constant address. Verify low two bits are clear.
1104 if ((CN->getZExtValue() & 3) == 0) {
1105 // If this address fits entirely in a 14-bit sext immediate field, codegen
1108 if (isIntS16Immediate(CN, Imm)) {
1109 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1110 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1111 CN->getValueType(0));
1115 // Fold the low-part of 32-bit absolute addresses into addr mode.
1116 if (CN->getValueType(0) == MVT::i32 ||
1117 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1118 int Addr = (int)CN->getZExtValue();
1120 // Otherwise, break this down into an LIS + disp.
1121 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1122 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1123 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1124 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1130 Disp = DAG.getTargetConstant(0, getPointerTy());
1131 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1132 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1135 return true; // [r+0]
1139 /// getPreIndexedAddressParts - returns true by value, base pointer and
1140 /// offset pointer and addressing mode by reference if the node's address
1141 /// can be legally represented as pre-indexed load / store address.
1142 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1144 ISD::MemIndexedMode &AM,
1145 SelectionDAG &DAG) const {
1146 if (DisablePPCPreinc) return false;
1150 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1151 Ptr = LD->getBasePtr();
1152 VT = LD->getMemoryVT();
1154 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1155 Ptr = ST->getBasePtr();
1156 VT = ST->getMemoryVT();
1160 // PowerPC doesn't have preinc load/store instructions for vectors.
1164 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1169 // LDU/STU use reg+imm*4, others use reg+imm.
1170 if (VT != MVT::i64) {
1172 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1176 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1180 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1181 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1182 // sext i32 to i64 when addr mode is r+i.
1183 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1184 LD->getExtensionType() == ISD::SEXTLOAD &&
1185 isa<ConstantSDNode>(Offset))
1193 //===----------------------------------------------------------------------===//
1194 // LowerOperation implementation
1195 //===----------------------------------------------------------------------===//
1197 /// GetLabelAccessInfo - Return true if we should reference labels using a
1198 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1199 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1200 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1201 HiOpFlags = PPCII::MO_HA16;
1202 LoOpFlags = PPCII::MO_LO16;
1204 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1205 // non-darwin platform. We don't support PIC on other platforms yet.
1206 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1207 TM.getSubtarget<PPCSubtarget>().isDarwin();
1209 HiOpFlags |= PPCII::MO_PIC_FLAG;
1210 LoOpFlags |= PPCII::MO_PIC_FLAG;
1213 // If this is a reference to a global value that requires a non-lazy-ptr, make
1214 // sure that instruction lowering adds it.
1215 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1216 HiOpFlags |= PPCII::MO_NLP_FLAG;
1217 LoOpFlags |= PPCII::MO_NLP_FLAG;
1219 if (GV->hasHiddenVisibility()) {
1220 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1221 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1228 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1229 SelectionDAG &DAG) {
1230 EVT PtrVT = HiPart.getValueType();
1231 SDValue Zero = DAG.getConstant(0, PtrVT);
1232 DebugLoc DL = HiPart.getDebugLoc();
1234 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1235 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1237 // With PIC, the first instruction is actually "GR+hi(&G)".
1239 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1240 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1242 // Generate non-pic code that has direct accesses to the constant pool.
1243 // The address of the global is just (hi(&g)+lo(&g)).
1244 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1247 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1248 SelectionDAG &DAG) const {
1249 EVT PtrVT = Op.getValueType();
1250 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1251 const Constant *C = CP->getConstVal();
1253 // 64-bit SVR4 ABI code is always position-independent.
1254 // The actual address of the GlobalValue is stored in the TOC.
1255 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1256 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1257 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1258 DAG.getRegister(PPC::X2, MVT::i64));
1261 unsigned MOHiFlag, MOLoFlag;
1262 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1264 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1266 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1267 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1270 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1271 EVT PtrVT = Op.getValueType();
1272 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1274 // 64-bit SVR4 ABI code is always position-independent.
1275 // The actual address of the GlobalValue is stored in the TOC.
1276 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1277 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1278 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1279 DAG.getRegister(PPC::X2, MVT::i64));
1282 unsigned MOHiFlag, MOLoFlag;
1283 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1284 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1285 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1286 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1289 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1290 SelectionDAG &DAG) const {
1291 EVT PtrVT = Op.getValueType();
1293 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1295 unsigned MOHiFlag, MOLoFlag;
1296 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1297 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1298 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1299 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1302 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1303 SelectionDAG &DAG) const {
1305 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1306 DebugLoc dl = GA->getDebugLoc();
1307 const GlobalValue *GV = GA->getGlobal();
1308 EVT PtrVT = getPointerTy();
1309 bool is64bit = PPCSubTarget.isPPC64();
1311 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1313 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1314 PPCII::MO_TPREL16_HA);
1315 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1316 PPCII::MO_TPREL16_LO);
1318 if (model != TLSModel::LocalExec)
1319 llvm_unreachable("only local-exec TLS mode supported");
1320 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1321 is64bit ? MVT::i64 : MVT::i32);
1322 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1323 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1326 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1327 SelectionDAG &DAG) const {
1328 EVT PtrVT = Op.getValueType();
1329 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1330 DebugLoc DL = GSDN->getDebugLoc();
1331 const GlobalValue *GV = GSDN->getGlobal();
1333 // 64-bit SVR4 ABI code is always position-independent.
1334 // The actual address of the GlobalValue is stored in the TOC.
1335 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1336 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1337 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1338 DAG.getRegister(PPC::X2, MVT::i64));
1341 unsigned MOHiFlag, MOLoFlag;
1342 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1345 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1347 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1349 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1351 // If the global reference is actually to a non-lazy-pointer, we have to do an
1352 // extra load to get the address of the global.
1353 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1354 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1355 false, false, false, 0);
1359 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1360 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1361 DebugLoc dl = Op.getDebugLoc();
1363 // If we're comparing for equality to zero, expose the fact that this is
1364 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1365 // fold the new nodes.
1366 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1367 if (C->isNullValue() && CC == ISD::SETEQ) {
1368 EVT VT = Op.getOperand(0).getValueType();
1369 SDValue Zext = Op.getOperand(0);
1370 if (VT.bitsLT(MVT::i32)) {
1372 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1374 unsigned Log2b = Log2_32(VT.getSizeInBits());
1375 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1376 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1377 DAG.getConstant(Log2b, MVT::i32));
1378 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1380 // Leave comparisons against 0 and -1 alone for now, since they're usually
1381 // optimized. FIXME: revisit this when we can custom lower all setcc
1383 if (C->isAllOnesValue() || C->isNullValue())
1387 // If we have an integer seteq/setne, turn it into a compare against zero
1388 // by xor'ing the rhs with the lhs, which is faster than setting a
1389 // condition register, reading it back out, and masking the correct bit. The
1390 // normal approach here uses sub to do this instead of xor. Using xor exposes
1391 // the result to other bit-twiddling opportunities.
1392 EVT LHSVT = Op.getOperand(0).getValueType();
1393 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1394 EVT VT = Op.getValueType();
1395 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1397 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1402 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1403 const PPCSubtarget &Subtarget) const {
1404 SDNode *Node = Op.getNode();
1405 EVT VT = Node->getValueType(0);
1406 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1407 SDValue InChain = Node->getOperand(0);
1408 SDValue VAListPtr = Node->getOperand(1);
1409 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1410 DebugLoc dl = Node->getDebugLoc();
1412 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1415 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1416 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1418 InChain = GprIndex.getValue(1);
1420 if (VT == MVT::i64) {
1421 // Check if GprIndex is even
1422 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1423 DAG.getConstant(1, MVT::i32));
1424 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1425 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1426 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1427 DAG.getConstant(1, MVT::i32));
1428 // Align GprIndex to be even if it isn't
1429 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1433 // fpr index is 1 byte after gpr
1434 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1435 DAG.getConstant(1, MVT::i32));
1438 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1439 FprPtr, MachinePointerInfo(SV), MVT::i8,
1441 InChain = FprIndex.getValue(1);
1443 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1444 DAG.getConstant(8, MVT::i32));
1446 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1447 DAG.getConstant(4, MVT::i32));
1450 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1451 MachinePointerInfo(), false, false,
1453 InChain = OverflowArea.getValue(1);
1455 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1456 MachinePointerInfo(), false, false,
1458 InChain = RegSaveArea.getValue(1);
1460 // select overflow_area if index > 8
1461 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1462 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1464 // adjustment constant gpr_index * 4/8
1465 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1466 VT.isInteger() ? GprIndex : FprIndex,
1467 DAG.getConstant(VT.isInteger() ? 4 : 8,
1470 // OurReg = RegSaveArea + RegConstant
1471 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1474 // Floating types are 32 bytes into RegSaveArea
1475 if (VT.isFloatingPoint())
1476 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1477 DAG.getConstant(32, MVT::i32));
1479 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1480 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1481 VT.isInteger() ? GprIndex : FprIndex,
1482 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1485 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1486 VT.isInteger() ? VAListPtr : FprPtr,
1487 MachinePointerInfo(SV),
1488 MVT::i8, false, false, 0);
1490 // determine if we should load from reg_save_area or overflow_area
1491 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1493 // increase overflow_area by 4/8 if gpr/fpr > 8
1494 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1495 DAG.getConstant(VT.isInteger() ? 4 : 8,
1498 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1501 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1503 MachinePointerInfo(),
1504 MVT::i32, false, false, 0);
1506 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1507 false, false, false, 0);
1510 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1511 SelectionDAG &DAG) const {
1512 return Op.getOperand(0);
1515 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1516 SelectionDAG &DAG) const {
1517 SDValue Chain = Op.getOperand(0);
1518 SDValue Trmp = Op.getOperand(1); // trampoline
1519 SDValue FPtr = Op.getOperand(2); // nested function
1520 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1521 DebugLoc dl = Op.getDebugLoc();
1523 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1524 bool isPPC64 = (PtrVT == MVT::i64);
1526 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1529 TargetLowering::ArgListTy Args;
1530 TargetLowering::ArgListEntry Entry;
1532 Entry.Ty = IntPtrTy;
1533 Entry.Node = Trmp; Args.push_back(Entry);
1535 // TrampSize == (isPPC64 ? 48 : 40);
1536 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1537 isPPC64 ? MVT::i64 : MVT::i32);
1538 Args.push_back(Entry);
1540 Entry.Node = FPtr; Args.push_back(Entry);
1541 Entry.Node = Nest; Args.push_back(Entry);
1543 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1544 TargetLowering::CallLoweringInfo CLI(Chain,
1545 Type::getVoidTy(*DAG.getContext()),
1546 false, false, false, false, 0,
1548 /*isTailCall=*/false,
1549 /*doesNotRet=*/false,
1550 /*isReturnValueUsed=*/true,
1551 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1553 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1555 return CallResult.second;
1558 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1559 const PPCSubtarget &Subtarget) const {
1560 MachineFunction &MF = DAG.getMachineFunction();
1561 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1563 DebugLoc dl = Op.getDebugLoc();
1565 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1566 // vastart just stores the address of the VarArgsFrameIndex slot into the
1567 // memory location argument.
1568 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1569 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1570 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1571 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1572 MachinePointerInfo(SV),
1576 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1577 // We suppose the given va_list is already allocated.
1580 // char gpr; /* index into the array of 8 GPRs
1581 // * stored in the register save area
1582 // * gpr=0 corresponds to r3,
1583 // * gpr=1 to r4, etc.
1585 // char fpr; /* index into the array of 8 FPRs
1586 // * stored in the register save area
1587 // * fpr=0 corresponds to f1,
1588 // * fpr=1 to f2, etc.
1590 // char *overflow_arg_area;
1591 // /* location on stack that holds
1592 // * the next overflow argument
1594 // char *reg_save_area;
1595 // /* where r3:r10 and f1:f8 (if saved)
1601 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1602 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1605 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1607 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1609 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1612 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1613 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1615 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1616 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1618 uint64_t FPROffset = 1;
1619 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1621 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1623 // Store first byte : number of int regs
1624 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1626 MachinePointerInfo(SV),
1627 MVT::i8, false, false, 0);
1628 uint64_t nextOffset = FPROffset;
1629 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1632 // Store second byte : number of float regs
1633 SDValue secondStore =
1634 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1635 MachinePointerInfo(SV, nextOffset), MVT::i8,
1637 nextOffset += StackOffset;
1638 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1640 // Store second word : arguments given on stack
1641 SDValue thirdStore =
1642 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1643 MachinePointerInfo(SV, nextOffset),
1645 nextOffset += FrameOffset;
1646 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1648 // Store third word : arguments given in registers
1649 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1650 MachinePointerInfo(SV, nextOffset),
1655 #include "PPCGenCallingConv.inc"
1657 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1658 CCValAssign::LocInfo &LocInfo,
1659 ISD::ArgFlagsTy &ArgFlags,
1664 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1666 CCValAssign::LocInfo &LocInfo,
1667 ISD::ArgFlagsTy &ArgFlags,
1669 static const uint16_t ArgRegs[] = {
1670 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1671 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1673 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1675 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1677 // Skip one register if the first unallocated register has an even register
1678 // number and there are still argument registers available which have not been
1679 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1680 // need to skip a register if RegNum is odd.
1681 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1682 State.AllocateReg(ArgRegs[RegNum]);
1685 // Always return false here, as this function only makes sure that the first
1686 // unallocated register has an odd register number and does not actually
1687 // allocate a register for the current argument.
1691 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1693 CCValAssign::LocInfo &LocInfo,
1694 ISD::ArgFlagsTy &ArgFlags,
1696 static const uint16_t ArgRegs[] = {
1697 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1701 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1703 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1705 // If there is only one Floating-point register left we need to put both f64
1706 // values of a split ppc_fp128 value on the stack.
1707 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1708 State.AllocateReg(ArgRegs[RegNum]);
1711 // Always return false here, as this function only makes sure that the two f64
1712 // values a ppc_fp128 value is split into are both passed in registers or both
1713 // passed on the stack and does not actually allocate a register for the
1714 // current argument.
1718 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1720 static const uint16_t *GetFPR() {
1721 static const uint16_t FPR[] = {
1722 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1723 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1729 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1731 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1732 unsigned PtrByteSize) {
1733 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1734 if (Flags.isByVal())
1735 ArgSize = Flags.getByValSize();
1736 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1742 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1743 CallingConv::ID CallConv, bool isVarArg,
1744 const SmallVectorImpl<ISD::InputArg>
1746 DebugLoc dl, SelectionDAG &DAG,
1747 SmallVectorImpl<SDValue> &InVals)
1749 if (PPCSubTarget.isSVR4ABI()) {
1750 if (PPCSubTarget.isPPC64())
1751 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1754 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1757 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1763 PPCTargetLowering::LowerFormalArguments_32SVR4(
1765 CallingConv::ID CallConv, bool isVarArg,
1766 const SmallVectorImpl<ISD::InputArg>
1768 DebugLoc dl, SelectionDAG &DAG,
1769 SmallVectorImpl<SDValue> &InVals) const {
1771 // 32-bit SVR4 ABI Stack Frame Layout:
1772 // +-----------------------------------+
1773 // +--> | Back chain |
1774 // | +-----------------------------------+
1775 // | | Floating-point register save area |
1776 // | +-----------------------------------+
1777 // | | General register save area |
1778 // | +-----------------------------------+
1779 // | | CR save word |
1780 // | +-----------------------------------+
1781 // | | VRSAVE save word |
1782 // | +-----------------------------------+
1783 // | | Alignment padding |
1784 // | +-----------------------------------+
1785 // | | Vector register save area |
1786 // | +-----------------------------------+
1787 // | | Local variable space |
1788 // | +-----------------------------------+
1789 // | | Parameter list area |
1790 // | +-----------------------------------+
1791 // | | LR save word |
1792 // | +-----------------------------------+
1793 // SP--> +--- | Back chain |
1794 // +-----------------------------------+
1797 // System V Application Binary Interface PowerPC Processor Supplement
1798 // AltiVec Technology Programming Interface Manual
1800 MachineFunction &MF = DAG.getMachineFunction();
1801 MachineFrameInfo *MFI = MF.getFrameInfo();
1802 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1804 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1805 // Potential tail calls could cause overwriting of argument stack slots.
1806 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1807 (CallConv == CallingConv::Fast));
1808 unsigned PtrByteSize = 4;
1810 // Assign locations to all of the incoming arguments.
1811 SmallVector<CCValAssign, 16> ArgLocs;
1812 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1813 getTargetMachine(), ArgLocs, *DAG.getContext());
1815 // Reserve space for the linkage area on the stack.
1816 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1818 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1820 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1821 CCValAssign &VA = ArgLocs[i];
1823 // Arguments stored in registers.
1824 if (VA.isRegLoc()) {
1825 const TargetRegisterClass *RC;
1826 EVT ValVT = VA.getValVT();
1828 switch (ValVT.getSimpleVT().SimpleTy) {
1830 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1832 RC = &PPC::GPRCRegClass;
1835 RC = &PPC::F4RCRegClass;
1838 RC = &PPC::F8RCRegClass;
1844 RC = &PPC::VRRCRegClass;
1848 // Transform the arguments stored in physical registers into virtual ones.
1849 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1850 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1852 InVals.push_back(ArgValue);
1854 // Argument stored in memory.
1855 assert(VA.isMemLoc());
1857 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1858 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1861 // Create load nodes to retrieve arguments from the stack.
1862 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1863 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1864 MachinePointerInfo(),
1865 false, false, false, 0));
1869 // Assign locations to all of the incoming aggregate by value arguments.
1870 // Aggregates passed by value are stored in the local variable space of the
1871 // caller's stack frame, right above the parameter list area.
1872 SmallVector<CCValAssign, 16> ByValArgLocs;
1873 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1874 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1876 // Reserve stack space for the allocations in CCInfo.
1877 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1879 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1881 // Area that is at least reserved in the caller of this function.
1882 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1884 // Set the size that is at least reserved in caller of this function. Tail
1885 // call optimized function's reserved stack space needs to be aligned so that
1886 // taking the difference between two stack areas will result in an aligned
1888 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1891 std::max(MinReservedArea,
1892 PPCFrameLowering::getMinCallFrameSize(false, false));
1894 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1895 getStackAlignment();
1896 unsigned AlignMask = TargetAlign-1;
1897 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1899 FI->setMinReservedArea(MinReservedArea);
1901 SmallVector<SDValue, 8> MemOps;
1903 // If the function takes variable number of arguments, make a frame index for
1904 // the start of the first vararg value... for expansion of llvm.va_start.
1906 static const uint16_t GPArgRegs[] = {
1907 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1908 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1910 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1912 static const uint16_t FPArgRegs[] = {
1913 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1916 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1918 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1920 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1923 // Make room for NumGPArgRegs and NumFPArgRegs.
1924 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1925 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1927 FuncInfo->setVarArgsStackOffset(
1928 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1929 CCInfo.getNextStackOffset(), true));
1931 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1932 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1934 // The fixed integer arguments of a variadic function are stored to the
1935 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1936 // the result of va_next.
1937 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1938 // Get an existing live-in vreg, or add a new one.
1939 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1941 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1943 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1944 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1945 MachinePointerInfo(), false, false, 0);
1946 MemOps.push_back(Store);
1947 // Increment the address by four for the next argument to store
1948 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1949 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1952 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1954 // The double arguments are stored to the VarArgsFrameIndex
1956 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1957 // Get an existing live-in vreg, or add a new one.
1958 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1960 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1962 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1963 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1964 MachinePointerInfo(), false, false, 0);
1965 MemOps.push_back(Store);
1966 // Increment the address by eight for the next argument to store
1967 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1969 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1973 if (!MemOps.empty())
1974 Chain = DAG.getNode(ISD::TokenFactor, dl,
1975 MVT::Other, &MemOps[0], MemOps.size());
1980 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1981 // value to MVT::i64 and then truncate to the correct register size.
1983 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
1984 SelectionDAG &DAG, SDValue ArgVal,
1985 DebugLoc dl) const {
1987 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1988 DAG.getValueType(ObjectVT));
1989 else if (Flags.isZExt())
1990 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1991 DAG.getValueType(ObjectVT));
1993 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1996 // Set the size that is at least reserved in caller of this function. Tail
1997 // call optimized functions' reserved stack space needs to be aligned so that
1998 // taking the difference between two stack areas will result in an aligned
2001 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
2002 unsigned nAltivecParamsAtEnd,
2003 unsigned MinReservedArea,
2004 bool isPPC64) const {
2005 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2006 // Add the Altivec parameters at the end, if needed.
2007 if (nAltivecParamsAtEnd) {
2008 MinReservedArea = ((MinReservedArea+15)/16)*16;
2009 MinReservedArea += 16*nAltivecParamsAtEnd;
2012 std::max(MinReservedArea,
2013 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2014 unsigned TargetAlign
2015 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2016 getStackAlignment();
2017 unsigned AlignMask = TargetAlign-1;
2018 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2019 FI->setMinReservedArea(MinReservedArea);
2023 PPCTargetLowering::LowerFormalArguments_64SVR4(
2025 CallingConv::ID CallConv, bool isVarArg,
2026 const SmallVectorImpl<ISD::InputArg>
2028 DebugLoc dl, SelectionDAG &DAG,
2029 SmallVectorImpl<SDValue> &InVals) const {
2030 // TODO: add description of PPC stack frame format, or at least some docs.
2032 MachineFunction &MF = DAG.getMachineFunction();
2033 MachineFrameInfo *MFI = MF.getFrameInfo();
2034 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2036 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2037 // Potential tail calls could cause overwriting of argument stack slots.
2038 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2039 (CallConv == CallingConv::Fast));
2040 unsigned PtrByteSize = 8;
2042 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2043 // Area that is at least reserved in caller of this function.
2044 unsigned MinReservedArea = ArgOffset;
2046 static const uint16_t GPR[] = {
2047 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2048 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2051 static const uint16_t *FPR = GetFPR();
2053 static const uint16_t VR[] = {
2054 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2055 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2058 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2059 const unsigned Num_FPR_Regs = 13;
2060 const unsigned Num_VR_Regs = array_lengthof(VR);
2062 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2064 // Add DAG nodes to load the arguments or copy them out of registers. On
2065 // entry to a function on PPC, the arguments start after the linkage area,
2066 // although the first ones are often in registers.
2068 SmallVector<SDValue, 8> MemOps;
2069 unsigned nAltivecParamsAtEnd = 0;
2070 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2071 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2073 bool needsLoad = false;
2074 EVT ObjectVT = Ins[ArgNo].VT;
2075 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2076 unsigned ArgSize = ObjSize;
2077 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2079 unsigned CurArgOffset = ArgOffset;
2081 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2082 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2083 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2085 MinReservedArea = ((MinReservedArea+15)/16)*16;
2086 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2090 nAltivecParamsAtEnd++;
2092 // Calculate min reserved area.
2093 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2097 // FIXME the codegen can be much improved in some cases.
2098 // We do not have to keep everything in memory.
2099 if (Flags.isByVal()) {
2100 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2101 ObjSize = Flags.getByValSize();
2102 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2103 // Empty aggregate parameters do not take up registers. Examples:
2107 // etc. However, we have to provide a place-holder in InVals, so
2108 // pretend we have an 8-byte item at the current address for that
2111 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2112 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2113 InVals.push_back(FIN);
2116 // All aggregates smaller than 8 bytes must be passed right-justified.
2117 if (ObjSize < PtrByteSize)
2118 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2119 // The value of the object is its address.
2120 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2121 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2122 InVals.push_back(FIN);
2125 if (GPR_idx != Num_GPR_Regs) {
2126 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2127 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2130 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2131 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2132 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2133 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2134 MachinePointerInfo(FuncArg, CurArgOffset),
2135 ObjType, false, false, 0);
2137 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2138 // store the whole register as-is to the parameter save area
2139 // slot. The address of the parameter was already calculated
2140 // above (InVals.push_back(FIN)) to be the right-justified
2141 // offset within the slot. For this store, we need a new
2142 // frame index that points at the beginning of the slot.
2143 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2144 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2145 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2146 MachinePointerInfo(FuncArg, ArgOffset),
2150 MemOps.push_back(Store);
2153 // Whether we copied from a register or not, advance the offset
2154 // into the parameter save area by a full doubleword.
2155 ArgOffset += PtrByteSize;
2159 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2160 // Store whatever pieces of the object are in registers
2161 // to memory. ArgOffset will be the address of the beginning
2163 if (GPR_idx != Num_GPR_Regs) {
2165 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2166 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2167 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2168 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2169 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2170 MachinePointerInfo(FuncArg, ArgOffset),
2172 MemOps.push_back(Store);
2174 ArgOffset += PtrByteSize;
2176 ArgOffset += ArgSize - j;
2183 switch (ObjectVT.getSimpleVT().SimpleTy) {
2184 default: llvm_unreachable("Unhandled argument type!");
2187 if (GPR_idx != Num_GPR_Regs) {
2188 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2189 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2191 if (ObjectVT == MVT::i32)
2192 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2193 // value to MVT::i64 and then truncate to the correct register size.
2194 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2199 ArgSize = PtrByteSize;
2206 // Every 8 bytes of argument space consumes one of the GPRs available for
2207 // argument passing.
2208 if (GPR_idx != Num_GPR_Regs) {
2211 if (FPR_idx != Num_FPR_Regs) {
2214 if (ObjectVT == MVT::f32)
2215 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2217 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2219 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2223 ArgSize = PtrByteSize;
2232 // Note that vector arguments in registers don't reserve stack space,
2233 // except in varargs functions.
2234 if (VR_idx != Num_VR_Regs) {
2235 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2236 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2238 while ((ArgOffset % 16) != 0) {
2239 ArgOffset += PtrByteSize;
2240 if (GPR_idx != Num_GPR_Regs)
2244 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2248 // Vectors are aligned.
2249 ArgOffset = ((ArgOffset+15)/16)*16;
2250 CurArgOffset = ArgOffset;
2257 // We need to load the argument to a virtual register if we determined
2258 // above that we ran out of physical registers of the appropriate type.
2260 int FI = MFI->CreateFixedObject(ObjSize,
2261 CurArgOffset + (ArgSize - ObjSize),
2263 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2264 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2265 false, false, false, 0);
2268 InVals.push_back(ArgVal);
2271 // Set the size that is at least reserved in caller of this function. Tail
2272 // call optimized functions' reserved stack space needs to be aligned so that
2273 // taking the difference between two stack areas will result in an aligned
2275 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2277 // If the function takes variable number of arguments, make a frame index for
2278 // the start of the first vararg value... for expansion of llvm.va_start.
2280 int Depth = ArgOffset;
2282 FuncInfo->setVarArgsFrameIndex(
2283 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2284 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2286 // If this function is vararg, store any remaining integer argument regs
2287 // to their spots on the stack so that they may be loaded by deferencing the
2288 // result of va_next.
2289 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2290 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2291 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2292 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2293 MachinePointerInfo(), false, false, 0);
2294 MemOps.push_back(Store);
2295 // Increment the address by four for the next argument to store
2296 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2297 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2301 if (!MemOps.empty())
2302 Chain = DAG.getNode(ISD::TokenFactor, dl,
2303 MVT::Other, &MemOps[0], MemOps.size());
2309 PPCTargetLowering::LowerFormalArguments_Darwin(
2311 CallingConv::ID CallConv, bool isVarArg,
2312 const SmallVectorImpl<ISD::InputArg>
2314 DebugLoc dl, SelectionDAG &DAG,
2315 SmallVectorImpl<SDValue> &InVals) const {
2316 // TODO: add description of PPC stack frame format, or at least some docs.
2318 MachineFunction &MF = DAG.getMachineFunction();
2319 MachineFrameInfo *MFI = MF.getFrameInfo();
2320 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2322 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2323 bool isPPC64 = PtrVT == MVT::i64;
2324 // Potential tail calls could cause overwriting of argument stack slots.
2325 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2326 (CallConv == CallingConv::Fast));
2327 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2329 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2330 // Area that is at least reserved in caller of this function.
2331 unsigned MinReservedArea = ArgOffset;
2333 static const uint16_t GPR_32[] = { // 32-bit registers.
2334 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2335 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2337 static const uint16_t GPR_64[] = { // 64-bit registers.
2338 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2339 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2342 static const uint16_t *FPR = GetFPR();
2344 static const uint16_t VR[] = {
2345 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2346 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2349 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2350 const unsigned Num_FPR_Regs = 13;
2351 const unsigned Num_VR_Regs = array_lengthof( VR);
2353 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2355 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2357 // In 32-bit non-varargs functions, the stack space for vectors is after the
2358 // stack space for non-vectors. We do not use this space unless we have
2359 // too many vectors to fit in registers, something that only occurs in
2360 // constructed examples:), but we have to walk the arglist to figure
2361 // that out...for the pathological case, compute VecArgOffset as the
2362 // start of the vector parameter area. Computing VecArgOffset is the
2363 // entire point of the following loop.
2364 unsigned VecArgOffset = ArgOffset;
2365 if (!isVarArg && !isPPC64) {
2366 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2368 EVT ObjectVT = Ins[ArgNo].VT;
2369 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2371 if (Flags.isByVal()) {
2372 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2373 unsigned ObjSize = Flags.getByValSize();
2375 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2376 VecArgOffset += ArgSize;
2380 switch(ObjectVT.getSimpleVT().SimpleTy) {
2381 default: llvm_unreachable("Unhandled argument type!");
2386 case MVT::i64: // PPC64
2388 // FIXME: We are guaranteed to be !isPPC64 at this point.
2389 // Does MVT::i64 apply?
2396 // Nothing to do, we're only looking at Nonvector args here.
2401 // We've found where the vector parameter area in memory is. Skip the
2402 // first 12 parameters; these don't use that memory.
2403 VecArgOffset = ((VecArgOffset+15)/16)*16;
2404 VecArgOffset += 12*16;
2406 // Add DAG nodes to load the arguments or copy them out of registers. On
2407 // entry to a function on PPC, the arguments start after the linkage area,
2408 // although the first ones are often in registers.
2410 SmallVector<SDValue, 8> MemOps;
2411 unsigned nAltivecParamsAtEnd = 0;
2412 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2413 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2415 bool needsLoad = false;
2416 EVT ObjectVT = Ins[ArgNo].VT;
2417 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2418 unsigned ArgSize = ObjSize;
2419 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2421 unsigned CurArgOffset = ArgOffset;
2423 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2424 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2425 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2426 if (isVarArg || isPPC64) {
2427 MinReservedArea = ((MinReservedArea+15)/16)*16;
2428 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2431 } else nAltivecParamsAtEnd++;
2433 // Calculate min reserved area.
2434 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2438 // FIXME the codegen can be much improved in some cases.
2439 // We do not have to keep everything in memory.
2440 if (Flags.isByVal()) {
2441 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2442 ObjSize = Flags.getByValSize();
2443 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2444 // Objects of size 1 and 2 are right justified, everything else is
2445 // left justified. This means the memory address is adjusted forwards.
2446 if (ObjSize==1 || ObjSize==2) {
2447 CurArgOffset = CurArgOffset + (4 - ObjSize);
2449 // The value of the object is its address.
2450 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2451 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2452 InVals.push_back(FIN);
2453 if (ObjSize==1 || ObjSize==2) {
2454 if (GPR_idx != Num_GPR_Regs) {
2457 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2459 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2460 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2461 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2462 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2463 MachinePointerInfo(FuncArg,
2465 ObjType, false, false, 0);
2466 MemOps.push_back(Store);
2470 ArgOffset += PtrByteSize;
2474 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2475 // Store whatever pieces of the object are in registers
2476 // to memory. ArgOffset will be the address of the beginning
2478 if (GPR_idx != Num_GPR_Regs) {
2481 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2483 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2484 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2485 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2486 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2487 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2488 MachinePointerInfo(FuncArg, ArgOffset),
2490 MemOps.push_back(Store);
2492 ArgOffset += PtrByteSize;
2494 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2501 switch (ObjectVT.getSimpleVT().SimpleTy) {
2502 default: llvm_unreachable("Unhandled argument type!");
2505 if (GPR_idx != Num_GPR_Regs) {
2506 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2507 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2511 ArgSize = PtrByteSize;
2513 // All int arguments reserve stack space in the Darwin ABI.
2514 ArgOffset += PtrByteSize;
2518 case MVT::i64: // PPC64
2519 if (GPR_idx != Num_GPR_Regs) {
2520 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2521 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2523 if (ObjectVT == MVT::i32)
2524 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2525 // value to MVT::i64 and then truncate to the correct register size.
2526 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2531 ArgSize = PtrByteSize;
2533 // All int arguments reserve stack space in the Darwin ABI.
2539 // Every 4 bytes of argument space consumes one of the GPRs available for
2540 // argument passing.
2541 if (GPR_idx != Num_GPR_Regs) {
2543 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2546 if (FPR_idx != Num_FPR_Regs) {
2549 if (ObjectVT == MVT::f32)
2550 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2552 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2554 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2560 // All FP arguments reserve stack space in the Darwin ABI.
2561 ArgOffset += isPPC64 ? 8 : ObjSize;
2567 // Note that vector arguments in registers don't reserve stack space,
2568 // except in varargs functions.
2569 if (VR_idx != Num_VR_Regs) {
2570 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2571 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2573 while ((ArgOffset % 16) != 0) {
2574 ArgOffset += PtrByteSize;
2575 if (GPR_idx != Num_GPR_Regs)
2579 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2583 if (!isVarArg && !isPPC64) {
2584 // Vectors go after all the nonvectors.
2585 CurArgOffset = VecArgOffset;
2588 // Vectors are aligned.
2589 ArgOffset = ((ArgOffset+15)/16)*16;
2590 CurArgOffset = ArgOffset;
2598 // We need to load the argument to a virtual register if we determined above
2599 // that we ran out of physical registers of the appropriate type.
2601 int FI = MFI->CreateFixedObject(ObjSize,
2602 CurArgOffset + (ArgSize - ObjSize),
2604 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2605 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2606 false, false, false, 0);
2609 InVals.push_back(ArgVal);
2612 // Set the size that is at least reserved in caller of this function. Tail
2613 // call optimized functions' reserved stack space needs to be aligned so that
2614 // taking the difference between two stack areas will result in an aligned
2616 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2618 // If the function takes variable number of arguments, make a frame index for
2619 // the start of the first vararg value... for expansion of llvm.va_start.
2621 int Depth = ArgOffset;
2623 FuncInfo->setVarArgsFrameIndex(
2624 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2626 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2628 // If this function is vararg, store any remaining integer argument regs
2629 // to their spots on the stack so that they may be loaded by deferencing the
2630 // result of va_next.
2631 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2635 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2637 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2639 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2640 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2641 MachinePointerInfo(), false, false, 0);
2642 MemOps.push_back(Store);
2643 // Increment the address by four for the next argument to store
2644 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2645 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2649 if (!MemOps.empty())
2650 Chain = DAG.getNode(ISD::TokenFactor, dl,
2651 MVT::Other, &MemOps[0], MemOps.size());
2656 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2657 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2659 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2663 const SmallVectorImpl<ISD::OutputArg>
2665 const SmallVectorImpl<SDValue> &OutVals,
2666 unsigned &nAltivecParamsAtEnd) {
2667 // Count how many bytes are to be pushed on the stack, including the linkage
2668 // area, and parameter passing area. We start with 24/48 bytes, which is
2669 // prereserved space for [SP][CR][LR][3 x unused].
2670 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2671 unsigned NumOps = Outs.size();
2672 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2674 // Add up all the space actually used.
2675 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2676 // they all go in registers, but we must reserve stack space for them for
2677 // possible use by the caller. In varargs or 64-bit calls, parameters are
2678 // assigned stack space in order, with padding so Altivec parameters are
2680 nAltivecParamsAtEnd = 0;
2681 for (unsigned i = 0; i != NumOps; ++i) {
2682 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2683 EVT ArgVT = Outs[i].VT;
2684 // Varargs Altivec parameters are padded to a 16 byte boundary.
2685 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2686 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2687 if (!isVarArg && !isPPC64) {
2688 // Non-varargs Altivec parameters go after all the non-Altivec
2689 // parameters; handle those later so we know how much padding we need.
2690 nAltivecParamsAtEnd++;
2693 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2694 NumBytes = ((NumBytes+15)/16)*16;
2696 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2699 // Allow for Altivec parameters at the end, if needed.
2700 if (nAltivecParamsAtEnd) {
2701 NumBytes = ((NumBytes+15)/16)*16;
2702 NumBytes += 16*nAltivecParamsAtEnd;
2705 // The prolog code of the callee may store up to 8 GPR argument registers to
2706 // the stack, allowing va_start to index over them in memory if its varargs.
2707 // Because we cannot tell if this is needed on the caller side, we have to
2708 // conservatively assume that it is needed. As such, make sure we have at
2709 // least enough stack space for the caller to store the 8 GPRs.
2710 NumBytes = std::max(NumBytes,
2711 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2713 // Tail call needs the stack to be aligned.
2714 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2715 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2716 getFrameLowering()->getStackAlignment();
2717 unsigned AlignMask = TargetAlign-1;
2718 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2724 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2725 /// adjusted to accommodate the arguments for the tailcall.
2726 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2727 unsigned ParamSize) {
2729 if (!isTailCall) return 0;
2731 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2732 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2733 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2734 // Remember only if the new adjustement is bigger.
2735 if (SPDiff < FI->getTailCallSPDelta())
2736 FI->setTailCallSPDelta(SPDiff);
2741 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2742 /// for tail call optimization. Targets which want to do tail call
2743 /// optimization should implement this function.
2745 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2746 CallingConv::ID CalleeCC,
2748 const SmallVectorImpl<ISD::InputArg> &Ins,
2749 SelectionDAG& DAG) const {
2750 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2753 // Variable argument functions are not supported.
2757 MachineFunction &MF = DAG.getMachineFunction();
2758 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2759 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2760 // Functions containing by val parameters are not supported.
2761 for (unsigned i = 0; i != Ins.size(); i++) {
2762 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2763 if (Flags.isByVal()) return false;
2766 // Non PIC/GOT tail calls are supported.
2767 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2770 // At the moment we can only do local tail calls (in same module, hidden
2771 // or protected) if we are generating PIC.
2772 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2773 return G->getGlobal()->hasHiddenVisibility()
2774 || G->getGlobal()->hasProtectedVisibility();
2780 /// isCallCompatibleAddress - Return the immediate to use if the specified
2781 /// 32-bit value is representable in the immediate field of a BxA instruction.
2782 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2783 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2786 int Addr = C->getZExtValue();
2787 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2788 SignExtend32<26>(Addr) != Addr)
2789 return 0; // Top 6 bits have to be sext of immediate.
2791 return DAG.getConstant((int)C->getZExtValue() >> 2,
2792 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2797 struct TailCallArgumentInfo {
2802 TailCallArgumentInfo() : FrameIdx(0) {}
2807 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2809 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2811 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2812 SmallVector<SDValue, 8> &MemOpChains,
2814 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2815 SDValue Arg = TailCallArgs[i].Arg;
2816 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2817 int FI = TailCallArgs[i].FrameIdx;
2818 // Store relative to framepointer.
2819 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2820 MachinePointerInfo::getFixedStack(FI),
2825 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2826 /// the appropriate stack slot for the tail call optimized function call.
2827 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2828 MachineFunction &MF,
2837 // Calculate the new stack slot for the return address.
2838 int SlotSize = isPPC64 ? 8 : 4;
2839 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2841 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2842 NewRetAddrLoc, true);
2843 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2844 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2845 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2846 MachinePointerInfo::getFixedStack(NewRetAddr),
2849 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2850 // slot as the FP is never overwritten.
2853 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2854 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2856 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2857 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2858 MachinePointerInfo::getFixedStack(NewFPIdx),
2865 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2866 /// the position of the argument.
2868 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2869 SDValue Arg, int SPDiff, unsigned ArgOffset,
2870 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2871 int Offset = ArgOffset + SPDiff;
2872 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2873 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2874 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2875 SDValue FIN = DAG.getFrameIndex(FI, VT);
2876 TailCallArgumentInfo Info;
2878 Info.FrameIdxOp = FIN;
2880 TailCallArguments.push_back(Info);
2883 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2884 /// stack slot. Returns the chain as result and the loaded frame pointers in
2885 /// LROpOut/FPOpout. Used when tail calling.
2886 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2892 DebugLoc dl) const {
2894 // Load the LR and FP stack slot for later adjusting.
2895 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2896 LROpOut = getReturnAddrFrameIndex(DAG);
2897 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2898 false, false, false, 0);
2899 Chain = SDValue(LROpOut.getNode(), 1);
2901 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2902 // slot as the FP is never overwritten.
2904 FPOpOut = getFramePointerFrameIndex(DAG);
2905 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2906 false, false, false, 0);
2907 Chain = SDValue(FPOpOut.getNode(), 1);
2913 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2914 /// by "Src" to address "Dst" of size "Size". Alignment information is
2915 /// specified by the specific parameter attribute. The copy will be passed as
2916 /// a byval function parameter.
2917 /// Sometimes what we are copying is the end of a larger object, the part that
2918 /// does not fit in registers.
2920 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2921 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2923 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2924 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2925 false, false, MachinePointerInfo(0),
2926 MachinePointerInfo(0));
2929 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2932 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2933 SDValue Arg, SDValue PtrOff, int SPDiff,
2934 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2935 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2936 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2938 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2943 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2945 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2946 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2947 DAG.getConstant(ArgOffset, PtrVT));
2949 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2950 MachinePointerInfo(), false, false, 0));
2951 // Calculate and remember argument location.
2952 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2957 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2958 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2959 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2960 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2961 MachineFunction &MF = DAG.getMachineFunction();
2963 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2964 // might overwrite each other in case of tail call optimization.
2965 SmallVector<SDValue, 8> MemOpChains2;
2966 // Do not flag preceding copytoreg stuff together with the following stuff.
2968 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2970 if (!MemOpChains2.empty())
2971 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2972 &MemOpChains2[0], MemOpChains2.size());
2974 // Store the return address to the appropriate stack slot.
2975 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2976 isPPC64, isDarwinABI, dl);
2978 // Emit callseq_end just before tailcall node.
2979 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2980 DAG.getIntPtrConstant(0, true), InFlag);
2981 InFlag = Chain.getValue(1);
2985 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2986 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2987 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2988 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2989 const PPCSubtarget &PPCSubTarget) {
2991 bool isPPC64 = PPCSubTarget.isPPC64();
2992 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2994 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2995 NodeTys.push_back(MVT::Other); // Returns a chain
2996 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2998 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
3000 bool needIndirectCall = true;
3001 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3002 // If this is an absolute destination address, use the munged value.
3003 Callee = SDValue(Dest, 0);
3004 needIndirectCall = false;
3007 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3008 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
3009 // Use indirect calls for ALL functions calls in JIT mode, since the
3010 // far-call stubs may be outside relocation limits for a BL instruction.
3011 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
3012 unsigned OpFlags = 0;
3013 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3014 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3015 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3016 (G->getGlobal()->isDeclaration() ||
3017 G->getGlobal()->isWeakForLinker())) {
3018 // PC-relative references to external symbols should go through $stub,
3019 // unless we're building with the leopard linker or later, which
3020 // automatically synthesizes these stubs.
3021 OpFlags = PPCII::MO_DARWIN_STUB;
3024 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3025 // every direct call is) turn it into a TargetGlobalAddress /
3026 // TargetExternalSymbol node so that legalize doesn't hack it.
3027 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3028 Callee.getValueType(),
3030 needIndirectCall = false;
3034 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3035 unsigned char OpFlags = 0;
3037 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3038 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3039 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3040 // PC-relative references to external symbols should go through $stub,
3041 // unless we're building with the leopard linker or later, which
3042 // automatically synthesizes these stubs.
3043 OpFlags = PPCII::MO_DARWIN_STUB;
3046 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3048 needIndirectCall = false;
3051 if (needIndirectCall) {
3052 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3053 // to do the call, we can't use PPCISD::CALL.
3054 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3056 if (isSVR4ABI && isPPC64) {
3057 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3058 // entry point, but to the function descriptor (the function entry point
3059 // address is part of the function descriptor though).
3060 // The function descriptor is a three doubleword structure with the
3061 // following fields: function entry point, TOC base address and
3062 // environment pointer.
3063 // Thus for a call through a function pointer, the following actions need
3065 // 1. Save the TOC of the caller in the TOC save area of its stack
3066 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3067 // 2. Load the address of the function entry point from the function
3069 // 3. Load the TOC of the callee from the function descriptor into r2.
3070 // 4. Load the environment pointer from the function descriptor into
3072 // 5. Branch to the function entry point address.
3073 // 6. On return of the callee, the TOC of the caller needs to be
3074 // restored (this is done in FinishCall()).
3076 // All those operations are flagged together to ensure that no other
3077 // operations can be scheduled in between. E.g. without flagging the
3078 // operations together, a TOC access in the caller could be scheduled
3079 // between the load of the callee TOC and the branch to the callee, which
3080 // results in the TOC access going through the TOC of the callee instead
3081 // of going through the TOC of the caller, which leads to incorrect code.
3083 // Load the address of the function entry point from the function
3085 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3086 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3087 InFlag.getNode() ? 3 : 2);
3088 Chain = LoadFuncPtr.getValue(1);
3089 InFlag = LoadFuncPtr.getValue(2);
3091 // Load environment pointer into r11.
3092 // Offset of the environment pointer within the function descriptor.
3093 SDValue PtrOff = DAG.getIntPtrConstant(16);
3095 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3096 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3098 Chain = LoadEnvPtr.getValue(1);
3099 InFlag = LoadEnvPtr.getValue(2);
3101 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3103 Chain = EnvVal.getValue(0);
3104 InFlag = EnvVal.getValue(1);
3106 // Load TOC of the callee into r2. We are using a target-specific load
3107 // with r2 hard coded, because the result of a target-independent load
3108 // would never go directly into r2, since r2 is a reserved register (which
3109 // prevents the register allocator from allocating it), resulting in an
3110 // additional register being allocated and an unnecessary move instruction
3112 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3113 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3115 Chain = LoadTOCPtr.getValue(0);
3116 InFlag = LoadTOCPtr.getValue(1);
3118 MTCTROps[0] = Chain;
3119 MTCTROps[1] = LoadFuncPtr;
3120 MTCTROps[2] = InFlag;
3123 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3124 2 + (InFlag.getNode() != 0));
3125 InFlag = Chain.getValue(1);
3128 NodeTys.push_back(MVT::Other);
3129 NodeTys.push_back(MVT::Glue);
3130 Ops.push_back(Chain);
3131 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3133 // Add CTR register as callee so a bctr can be emitted later.
3135 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3138 // If this is a direct call, pass the chain and the callee.
3139 if (Callee.getNode()) {
3140 Ops.push_back(Chain);
3141 Ops.push_back(Callee);
3143 // If this is a tail call add stack pointer delta.
3145 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3147 // Add argument registers to the end of the list so that they are known live
3149 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3150 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3151 RegsToPass[i].second.getValueType()));
3157 bool isLocalCall(const SDValue &Callee)
3159 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3160 return !G->getGlobal()->isDeclaration() &&
3161 !G->getGlobal()->isWeakForLinker();
3166 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3167 CallingConv::ID CallConv, bool isVarArg,
3168 const SmallVectorImpl<ISD::InputArg> &Ins,
3169 DebugLoc dl, SelectionDAG &DAG,
3170 SmallVectorImpl<SDValue> &InVals) const {
3172 SmallVector<CCValAssign, 16> RVLocs;
3173 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3174 getTargetMachine(), RVLocs, *DAG.getContext());
3175 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3177 // Copy all of the result registers out of their specified physreg.
3178 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3179 CCValAssign &VA = RVLocs[i];
3180 assert(VA.isRegLoc() && "Can only return in registers!");
3182 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3183 VA.getLocReg(), VA.getLocVT(), InFlag);
3184 Chain = Val.getValue(1);
3185 InFlag = Val.getValue(2);
3187 switch (VA.getLocInfo()) {
3188 default: llvm_unreachable("Unknown loc info!");
3189 case CCValAssign::Full: break;
3190 case CCValAssign::AExt:
3191 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3193 case CCValAssign::ZExt:
3194 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3195 DAG.getValueType(VA.getValVT()));
3196 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3198 case CCValAssign::SExt:
3199 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3200 DAG.getValueType(VA.getValVT()));
3201 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3205 InVals.push_back(Val);
3212 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3213 bool isTailCall, bool isVarArg,
3215 SmallVector<std::pair<unsigned, SDValue>, 8>
3217 SDValue InFlag, SDValue Chain,
3219 int SPDiff, unsigned NumBytes,
3220 const SmallVectorImpl<ISD::InputArg> &Ins,
3221 SmallVectorImpl<SDValue> &InVals) const {
3222 std::vector<EVT> NodeTys;
3223 SmallVector<SDValue, 8> Ops;
3224 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3225 isTailCall, RegsToPass, Ops, NodeTys,
3228 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3229 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3230 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3232 // When performing tail call optimization the callee pops its arguments off
3233 // the stack. Account for this here so these bytes can be pushed back on in
3234 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3235 int BytesCalleePops =
3236 (CallConv == CallingConv::Fast &&
3237 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3239 // Add a register mask operand representing the call-preserved registers.
3240 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3241 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3242 assert(Mask && "Missing call preserved mask for calling convention");
3243 Ops.push_back(DAG.getRegisterMask(Mask));
3245 if (InFlag.getNode())
3246 Ops.push_back(InFlag);
3250 // If this is the first return lowered for this function, add the regs
3251 // to the liveout set for the function.
3252 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3253 SmallVector<CCValAssign, 16> RVLocs;
3254 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3255 getTargetMachine(), RVLocs, *DAG.getContext());
3256 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3257 for (unsigned i = 0; i != RVLocs.size(); ++i)
3258 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3261 assert(((Callee.getOpcode() == ISD::Register &&
3262 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3263 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3264 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3265 isa<ConstantSDNode>(Callee)) &&
3266 "Expecting an global address, external symbol, absolute value or register");
3268 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3271 // Add a NOP immediately after the branch instruction when using the 64-bit
3272 // SVR4 ABI. At link time, if caller and callee are in a different module and
3273 // thus have a different TOC, the call will be replaced with a call to a stub
3274 // function which saves the current TOC, loads the TOC of the callee and
3275 // branches to the callee. The NOP will be replaced with a load instruction
3276 // which restores the TOC of the caller from the TOC save slot of the current
3277 // stack frame. If caller and callee belong to the same module (and have the
3278 // same TOC), the NOP will remain unchanged.
3280 bool needsTOCRestore = false;
3281 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3282 if (CallOpc == PPCISD::BCTRL_SVR4) {
3283 // This is a call through a function pointer.
3284 // Restore the caller TOC from the save area into R2.
3285 // See PrepareCall() for more information about calls through function
3286 // pointers in the 64-bit SVR4 ABI.
3287 // We are using a target-specific load with r2 hard coded, because the
3288 // result of a target-independent load would never go directly into r2,
3289 // since r2 is a reserved register (which prevents the register allocator
3290 // from allocating it), resulting in an additional register being
3291 // allocated and an unnecessary move instruction being generated.
3292 needsTOCRestore = true;
3293 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3294 // Otherwise insert NOP for non-local calls.
3295 CallOpc = PPCISD::CALL_NOP_SVR4;
3299 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3300 InFlag = Chain.getValue(1);
3302 if (needsTOCRestore) {
3303 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3304 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3305 InFlag = Chain.getValue(1);
3308 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3309 DAG.getIntPtrConstant(BytesCalleePops, true),
3312 InFlag = Chain.getValue(1);
3314 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3315 Ins, dl, DAG, InVals);
3319 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3320 SmallVectorImpl<SDValue> &InVals) const {
3321 SelectionDAG &DAG = CLI.DAG;
3322 DebugLoc &dl = CLI.DL;
3323 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3324 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3325 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3326 SDValue Chain = CLI.Chain;
3327 SDValue Callee = CLI.Callee;
3328 bool &isTailCall = CLI.IsTailCall;
3329 CallingConv::ID CallConv = CLI.CallConv;
3330 bool isVarArg = CLI.IsVarArg;
3333 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3336 if (PPCSubTarget.isSVR4ABI()) {
3337 if (PPCSubTarget.isPPC64())
3338 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3339 isTailCall, Outs, OutVals, Ins,
3342 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3343 isTailCall, Outs, OutVals, Ins,
3347 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3348 isTailCall, Outs, OutVals, Ins,
3353 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3354 CallingConv::ID CallConv, bool isVarArg,
3356 const SmallVectorImpl<ISD::OutputArg> &Outs,
3357 const SmallVectorImpl<SDValue> &OutVals,
3358 const SmallVectorImpl<ISD::InputArg> &Ins,
3359 DebugLoc dl, SelectionDAG &DAG,
3360 SmallVectorImpl<SDValue> &InVals) const {
3361 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3362 // of the 32-bit SVR4 ABI stack frame layout.
3364 assert((CallConv == CallingConv::C ||
3365 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3367 unsigned PtrByteSize = 4;
3369 MachineFunction &MF = DAG.getMachineFunction();
3371 // Mark this function as potentially containing a function that contains a
3372 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3373 // and restoring the callers stack pointer in this functions epilog. This is
3374 // done because by tail calling the called function might overwrite the value
3375 // in this function's (MF) stack pointer stack slot 0(SP).
3376 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3377 CallConv == CallingConv::Fast)
3378 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3380 // Count how many bytes are to be pushed on the stack, including the linkage
3381 // area, parameter list area and the part of the local variable space which
3382 // contains copies of aggregates which are passed by value.
3384 // Assign locations to all of the outgoing arguments.
3385 SmallVector<CCValAssign, 16> ArgLocs;
3386 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3387 getTargetMachine(), ArgLocs, *DAG.getContext());
3389 // Reserve space for the linkage area on the stack.
3390 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3393 // Handle fixed and variable vector arguments differently.
3394 // Fixed vector arguments go into registers as long as registers are
3395 // available. Variable vector arguments always go into memory.
3396 unsigned NumArgs = Outs.size();
3398 for (unsigned i = 0; i != NumArgs; ++i) {
3399 MVT ArgVT = Outs[i].VT;
3400 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3403 if (Outs[i].IsFixed) {
3404 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3407 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3413 errs() << "Call operand #" << i << " has unhandled type "
3414 << EVT(ArgVT).getEVTString() << "\n";
3416 llvm_unreachable(0);
3420 // All arguments are treated the same.
3421 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3424 // Assign locations to all of the outgoing aggregate by value arguments.
3425 SmallVector<CCValAssign, 16> ByValArgLocs;
3426 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3427 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3429 // Reserve stack space for the allocations in CCInfo.
3430 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3432 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3434 // Size of the linkage area, parameter list area and the part of the local
3435 // space variable where copies of aggregates which are passed by value are
3437 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3439 // Calculate by how many bytes the stack has to be adjusted in case of tail
3440 // call optimization.
3441 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3443 // Adjust the stack pointer for the new arguments...
3444 // These operations are automatically eliminated by the prolog/epilog pass
3445 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3446 SDValue CallSeqStart = Chain;
3448 // Load the return address and frame pointer so it can be moved somewhere else
3451 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3454 // Set up a copy of the stack pointer for use loading and storing any
3455 // arguments that may not fit in the registers available for argument
3457 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3459 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3460 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3461 SmallVector<SDValue, 8> MemOpChains;
3463 bool seenFloatArg = false;
3464 // Walk the register/memloc assignments, inserting copies/loads.
3465 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3468 CCValAssign &VA = ArgLocs[i];
3469 SDValue Arg = OutVals[i];
3470 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3472 if (Flags.isByVal()) {
3473 // Argument is an aggregate which is passed by value, thus we need to
3474 // create a copy of it in the local variable space of the current stack
3475 // frame (which is the stack frame of the caller) and pass the address of
3476 // this copy to the callee.
3477 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3478 CCValAssign &ByValVA = ByValArgLocs[j++];
3479 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3481 // Memory reserved in the local variable space of the callers stack frame.
3482 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3484 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3485 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3487 // Create a copy of the argument in the local area of the current
3489 SDValue MemcpyCall =
3490 CreateCopyOfByValArgument(Arg, PtrOff,
3491 CallSeqStart.getNode()->getOperand(0),
3494 // This must go outside the CALLSEQ_START..END.
3495 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3496 CallSeqStart.getNode()->getOperand(1));
3497 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3498 NewCallSeqStart.getNode());
3499 Chain = CallSeqStart = NewCallSeqStart;
3501 // Pass the address of the aggregate copy on the stack either in a
3502 // physical register or in the parameter list area of the current stack
3503 // frame to the callee.
3507 if (VA.isRegLoc()) {
3508 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3509 // Put argument in a physical register.
3510 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3512 // Put argument in the parameter list area of the current stack frame.
3513 assert(VA.isMemLoc());
3514 unsigned LocMemOffset = VA.getLocMemOffset();
3517 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3518 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3520 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3521 MachinePointerInfo(),
3524 // Calculate and remember argument location.
3525 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3531 if (!MemOpChains.empty())
3532 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3533 &MemOpChains[0], MemOpChains.size());
3535 // Build a sequence of copy-to-reg nodes chained together with token chain
3536 // and flag operands which copy the outgoing args into the appropriate regs.
3538 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3539 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3540 RegsToPass[i].second, InFlag);
3541 InFlag = Chain.getValue(1);
3544 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3547 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3548 SDValue Ops[] = { Chain, InFlag };
3550 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3551 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3553 InFlag = Chain.getValue(1);
3557 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3558 false, TailCallArguments);
3560 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3561 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3565 // Copy an argument into memory, being careful to do this outside the
3566 // call sequence for the call to which the argument belongs.
3568 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3569 SDValue CallSeqStart,
3570 ISD::ArgFlagsTy Flags,
3572 DebugLoc dl) const {
3573 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3574 CallSeqStart.getNode()->getOperand(0),
3576 // The MEMCPY must go outside the CALLSEQ_START..END.
3577 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3578 CallSeqStart.getNode()->getOperand(1));
3579 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3580 NewCallSeqStart.getNode());
3581 return NewCallSeqStart;
3585 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3586 CallingConv::ID CallConv, bool isVarArg,
3588 const SmallVectorImpl<ISD::OutputArg> &Outs,
3589 const SmallVectorImpl<SDValue> &OutVals,
3590 const SmallVectorImpl<ISD::InputArg> &Ins,
3591 DebugLoc dl, SelectionDAG &DAG,
3592 SmallVectorImpl<SDValue> &InVals) const {
3594 unsigned NumOps = Outs.size();
3596 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3597 unsigned PtrByteSize = 8;
3599 MachineFunction &MF = DAG.getMachineFunction();
3601 // Mark this function as potentially containing a function that contains a
3602 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3603 // and restoring the callers stack pointer in this functions epilog. This is
3604 // done because by tail calling the called function might overwrite the value
3605 // in this function's (MF) stack pointer stack slot 0(SP).
3606 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3607 CallConv == CallingConv::Fast)
3608 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3610 unsigned nAltivecParamsAtEnd = 0;
3612 // Count how many bytes are to be pushed on the stack, including the linkage
3613 // area, and parameter passing area. We start with at least 48 bytes, which
3614 // is reserved space for [SP][CR][LR][3 x unused].
3615 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3618 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3619 Outs, OutVals, nAltivecParamsAtEnd);
3621 // Calculate by how many bytes the stack has to be adjusted in case of tail
3622 // call optimization.
3623 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3625 // To protect arguments on the stack from being clobbered in a tail call,
3626 // force all the loads to happen before doing any other lowering.
3628 Chain = DAG.getStackArgumentTokenFactor(Chain);
3630 // Adjust the stack pointer for the new arguments...
3631 // These operations are automatically eliminated by the prolog/epilog pass
3632 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3633 SDValue CallSeqStart = Chain;
3635 // Load the return address and frame pointer so it can be move somewhere else
3638 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3641 // Set up a copy of the stack pointer for use loading and storing any
3642 // arguments that may not fit in the registers available for argument
3644 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3646 // Figure out which arguments are going to go in registers, and which in
3647 // memory. Also, if this is a vararg function, floating point operations
3648 // must be stored to our stack, and loaded into integer regs as well, if
3649 // any integer regs are available for argument passing.
3650 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3651 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3653 static const uint16_t GPR[] = {
3654 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3655 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3657 static const uint16_t *FPR = GetFPR();
3659 static const uint16_t VR[] = {
3660 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3661 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3663 const unsigned NumGPRs = array_lengthof(GPR);
3664 const unsigned NumFPRs = 13;
3665 const unsigned NumVRs = array_lengthof(VR);
3667 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3668 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3670 SmallVector<SDValue, 8> MemOpChains;
3671 for (unsigned i = 0; i != NumOps; ++i) {
3672 SDValue Arg = OutVals[i];
3673 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3675 // PtrOff will be used to store the current argument to the stack if a
3676 // register cannot be found for it.
3679 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3681 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3683 // Promote integers to 64-bit values.
3684 if (Arg.getValueType() == MVT::i32) {
3685 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3686 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3687 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3690 // FIXME memcpy is used way more than necessary. Correctness first.
3691 // Note: "by value" is code for passing a structure by value, not
3693 if (Flags.isByVal()) {
3694 // Note: Size includes alignment padding, so
3695 // struct x { short a; char b; }
3696 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3697 // These are the proper values we need for right-justifying the
3698 // aggregate in a parameter register.
3699 unsigned Size = Flags.getByValSize();
3701 // An empty aggregate parameter takes up no storage and no
3706 // All aggregates smaller than 8 bytes must be passed right-justified.
3707 if (Size==1 || Size==2 || Size==4) {
3708 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3709 if (GPR_idx != NumGPRs) {
3710 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3711 MachinePointerInfo(), VT,
3713 MemOpChains.push_back(Load.getValue(1));
3714 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3716 ArgOffset += PtrByteSize;
3721 if (GPR_idx == NumGPRs && Size < 8) {
3722 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3723 PtrOff.getValueType());
3724 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3725 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3728 ArgOffset += PtrByteSize;
3731 // Copy entire object into memory. There are cases where gcc-generated
3732 // code assumes it is there, even if it could be put entirely into
3733 // registers. (This is not what the doc says.)
3735 // FIXME: The above statement is likely due to a misunderstanding of the
3736 // documents. All arguments must be copied into the parameter area BY
3737 // THE CALLEE in the event that the callee takes the address of any
3738 // formal argument. That has not yet been implemented. However, it is
3739 // reasonable to use the stack area as a staging area for the register
3742 // Skip this for small aggregates, as we will use the same slot for a
3743 // right-justified copy, below.
3745 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3749 // When a register is available, pass a small aggregate right-justified.
3750 if (Size < 8 && GPR_idx != NumGPRs) {
3751 // The easiest way to get this right-justified in a register
3752 // is to copy the structure into the rightmost portion of a
3753 // local variable slot, then load the whole slot into the
3755 // FIXME: The memcpy seems to produce pretty awful code for
3756 // small aggregates, particularly for packed ones.
3757 // FIXME: It would be preferable to use the slot in the
3758 // parameter save area instead of a new local variable.
3759 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3760 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3761 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3765 // Load the slot into the register.
3766 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3767 MachinePointerInfo(),
3768 false, false, false, 0);
3769 MemOpChains.push_back(Load.getValue(1));
3770 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3772 // Done with this argument.
3773 ArgOffset += PtrByteSize;
3777 // For aggregates larger than PtrByteSize, copy the pieces of the
3778 // object that fit into registers from the parameter save area.
3779 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3780 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3781 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3782 if (GPR_idx != NumGPRs) {
3783 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3784 MachinePointerInfo(),
3785 false, false, false, 0);
3786 MemOpChains.push_back(Load.getValue(1));
3787 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3788 ArgOffset += PtrByteSize;
3790 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3797 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3798 default: llvm_unreachable("Unexpected ValueType for argument!");
3801 if (GPR_idx != NumGPRs) {
3802 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3804 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3805 true, isTailCall, false, MemOpChains,
3806 TailCallArguments, dl);
3808 ArgOffset += PtrByteSize;
3812 if (FPR_idx != NumFPRs) {
3813 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3816 // A single float or an aggregate containing only a single float
3817 // must be passed right-justified in the stack doubleword, and
3818 // in the GPR, if one is available.
3820 if (Arg.getValueType().getSimpleVT().SimpleTy == MVT::f32) {
3821 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3822 StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3826 SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
3827 MachinePointerInfo(), false, false, 0);
3828 MemOpChains.push_back(Store);
3830 // Float varargs are always shadowed in available integer registers
3831 if (GPR_idx != NumGPRs) {
3832 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3833 MachinePointerInfo(), false, false,
3835 MemOpChains.push_back(Load.getValue(1));
3836 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3838 } else if (GPR_idx != NumGPRs)
3839 // If we have any FPRs remaining, we may also have GPRs remaining.
3842 // Single-precision floating-point values are mapped to the
3843 // second (rightmost) word of the stack doubleword.
3844 if (Arg.getValueType() == MVT::f32) {
3845 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3846 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3849 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3850 true, isTailCall, false, MemOpChains,
3851 TailCallArguments, dl);
3860 // These go aligned on the stack, or in the corresponding R registers
3861 // when within range. The Darwin PPC ABI doc claims they also go in
3862 // V registers; in fact gcc does this only for arguments that are
3863 // prototyped, not for those that match the ... We do it for all
3864 // arguments, seems to work.
3865 while (ArgOffset % 16 !=0) {
3866 ArgOffset += PtrByteSize;
3867 if (GPR_idx != NumGPRs)
3870 // We could elide this store in the case where the object fits
3871 // entirely in R registers. Maybe later.
3872 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3873 DAG.getConstant(ArgOffset, PtrVT));
3874 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3875 MachinePointerInfo(), false, false, 0);
3876 MemOpChains.push_back(Store);
3877 if (VR_idx != NumVRs) {
3878 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3879 MachinePointerInfo(),
3880 false, false, false, 0);
3881 MemOpChains.push_back(Load.getValue(1));
3882 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3885 for (unsigned i=0; i<16; i+=PtrByteSize) {
3886 if (GPR_idx == NumGPRs)
3888 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3889 DAG.getConstant(i, PtrVT));
3890 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3891 false, false, false, 0);
3892 MemOpChains.push_back(Load.getValue(1));
3893 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3898 // Non-varargs Altivec params generally go in registers, but have
3899 // stack space allocated at the end.
3900 if (VR_idx != NumVRs) {
3901 // Doesn't have GPR space allocated.
3902 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3904 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3905 true, isTailCall, true, MemOpChains,
3906 TailCallArguments, dl);
3913 if (!MemOpChains.empty())
3914 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3915 &MemOpChains[0], MemOpChains.size());
3917 // Check if this is an indirect call (MTCTR/BCTRL).
3918 // See PrepareCall() for more information about calls through function
3919 // pointers in the 64-bit SVR4 ABI.
3921 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3922 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3923 !isBLACompatibleAddress(Callee, DAG)) {
3924 // Load r2 into a virtual register and store it to the TOC save area.
3925 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3926 // TOC save area offset.
3927 SDValue PtrOff = DAG.getIntPtrConstant(40);
3928 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3929 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3931 // R12 must contain the address of an indirect callee. This does not
3932 // mean the MTCTR instruction must use R12; it's easier to model this
3933 // as an extra parameter, so do that.
3934 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
3937 // Build a sequence of copy-to-reg nodes chained together with token chain
3938 // and flag operands which copy the outgoing args into the appropriate regs.
3940 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3941 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3942 RegsToPass[i].second, InFlag);
3943 InFlag = Chain.getValue(1);
3947 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
3948 FPOp, true, TailCallArguments);
3950 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3951 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3956 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3957 CallingConv::ID CallConv, bool isVarArg,
3959 const SmallVectorImpl<ISD::OutputArg> &Outs,
3960 const SmallVectorImpl<SDValue> &OutVals,
3961 const SmallVectorImpl<ISD::InputArg> &Ins,
3962 DebugLoc dl, SelectionDAG &DAG,
3963 SmallVectorImpl<SDValue> &InVals) const {
3965 unsigned NumOps = Outs.size();
3967 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3968 bool isPPC64 = PtrVT == MVT::i64;
3969 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3971 MachineFunction &MF = DAG.getMachineFunction();
3973 // Mark this function as potentially containing a function that contains a
3974 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3975 // and restoring the callers stack pointer in this functions epilog. This is
3976 // done because by tail calling the called function might overwrite the value
3977 // in this function's (MF) stack pointer stack slot 0(SP).
3978 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3979 CallConv == CallingConv::Fast)
3980 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3982 unsigned nAltivecParamsAtEnd = 0;
3984 // Count how many bytes are to be pushed on the stack, including the linkage
3985 // area, and parameter passing area. We start with 24/48 bytes, which is
3986 // prereserved space for [SP][CR][LR][3 x unused].
3988 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
3990 nAltivecParamsAtEnd);
3992 // Calculate by how many bytes the stack has to be adjusted in case of tail
3993 // call optimization.
3994 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3996 // To protect arguments on the stack from being clobbered in a tail call,
3997 // force all the loads to happen before doing any other lowering.
3999 Chain = DAG.getStackArgumentTokenFactor(Chain);
4001 // Adjust the stack pointer for the new arguments...
4002 // These operations are automatically eliminated by the prolog/epilog pass
4003 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
4004 SDValue CallSeqStart = Chain;
4006 // Load the return address and frame pointer so it can be move somewhere else
4009 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4012 // Set up a copy of the stack pointer for use loading and storing any
4013 // arguments that may not fit in the registers available for argument
4017 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4019 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4021 // Figure out which arguments are going to go in registers, and which in
4022 // memory. Also, if this is a vararg function, floating point operations
4023 // must be stored to our stack, and loaded into integer regs as well, if
4024 // any integer regs are available for argument passing.
4025 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
4026 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4028 static const uint16_t GPR_32[] = { // 32-bit registers.
4029 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4030 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4032 static const uint16_t GPR_64[] = { // 64-bit registers.
4033 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4034 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4036 static const uint16_t *FPR = GetFPR();
4038 static const uint16_t VR[] = {
4039 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4040 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4042 const unsigned NumGPRs = array_lengthof(GPR_32);
4043 const unsigned NumFPRs = 13;
4044 const unsigned NumVRs = array_lengthof(VR);
4046 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
4048 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4049 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4051 SmallVector<SDValue, 8> MemOpChains;
4052 for (unsigned i = 0; i != NumOps; ++i) {
4053 SDValue Arg = OutVals[i];
4054 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4056 // PtrOff will be used to store the current argument to the stack if a
4057 // register cannot be found for it.
4060 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4062 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4064 // On PPC64, promote integers to 64-bit values.
4065 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4066 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4067 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4068 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4071 // FIXME memcpy is used way more than necessary. Correctness first.
4072 // Note: "by value" is code for passing a structure by value, not
4074 if (Flags.isByVal()) {
4075 unsigned Size = Flags.getByValSize();
4076 // Very small objects are passed right-justified. Everything else is
4077 // passed left-justified.
4078 if (Size==1 || Size==2) {
4079 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4080 if (GPR_idx != NumGPRs) {
4081 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4082 MachinePointerInfo(), VT,
4084 MemOpChains.push_back(Load.getValue(1));
4085 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4087 ArgOffset += PtrByteSize;
4089 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4090 PtrOff.getValueType());
4091 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4092 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4095 ArgOffset += PtrByteSize;
4099 // Copy entire object into memory. There are cases where gcc-generated
4100 // code assumes it is there, even if it could be put entirely into
4101 // registers. (This is not what the doc says.)
4102 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4106 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4107 // copy the pieces of the object that fit into registers from the
4108 // parameter save area.
4109 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4110 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4111 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4112 if (GPR_idx != NumGPRs) {
4113 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4114 MachinePointerInfo(),
4115 false, false, false, 0);
4116 MemOpChains.push_back(Load.getValue(1));
4117 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4118 ArgOffset += PtrByteSize;
4120 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4127 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4128 default: llvm_unreachable("Unexpected ValueType for argument!");
4131 if (GPR_idx != NumGPRs) {
4132 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4134 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4135 isPPC64, isTailCall, false, MemOpChains,
4136 TailCallArguments, dl);
4138 ArgOffset += PtrByteSize;
4142 if (FPR_idx != NumFPRs) {
4143 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4146 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4147 MachinePointerInfo(), false, false, 0);
4148 MemOpChains.push_back(Store);
4150 // Float varargs are always shadowed in available integer registers
4151 if (GPR_idx != NumGPRs) {
4152 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4153 MachinePointerInfo(), false, false,
4155 MemOpChains.push_back(Load.getValue(1));
4156 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4158 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4159 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4160 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4161 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4162 MachinePointerInfo(),
4163 false, false, false, 0);
4164 MemOpChains.push_back(Load.getValue(1));
4165 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4168 // If we have any FPRs remaining, we may also have GPRs remaining.
4169 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4171 if (GPR_idx != NumGPRs)
4173 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4174 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4178 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4179 isPPC64, isTailCall, false, MemOpChains,
4180 TailCallArguments, dl);
4184 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4191 // These go aligned on the stack, or in the corresponding R registers
4192 // when within range. The Darwin PPC ABI doc claims they also go in
4193 // V registers; in fact gcc does this only for arguments that are
4194 // prototyped, not for those that match the ... We do it for all
4195 // arguments, seems to work.
4196 while (ArgOffset % 16 !=0) {
4197 ArgOffset += PtrByteSize;
4198 if (GPR_idx != NumGPRs)
4201 // We could elide this store in the case where the object fits
4202 // entirely in R registers. Maybe later.
4203 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4204 DAG.getConstant(ArgOffset, PtrVT));
4205 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4206 MachinePointerInfo(), false, false, 0);
4207 MemOpChains.push_back(Store);
4208 if (VR_idx != NumVRs) {
4209 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4210 MachinePointerInfo(),
4211 false, false, false, 0);
4212 MemOpChains.push_back(Load.getValue(1));
4213 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4216 for (unsigned i=0; i<16; i+=PtrByteSize) {
4217 if (GPR_idx == NumGPRs)
4219 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4220 DAG.getConstant(i, PtrVT));
4221 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4222 false, false, false, 0);
4223 MemOpChains.push_back(Load.getValue(1));
4224 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4229 // Non-varargs Altivec params generally go in registers, but have
4230 // stack space allocated at the end.
4231 if (VR_idx != NumVRs) {
4232 // Doesn't have GPR space allocated.
4233 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4234 } else if (nAltivecParamsAtEnd==0) {
4235 // We are emitting Altivec params in order.
4236 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4237 isPPC64, isTailCall, true, MemOpChains,
4238 TailCallArguments, dl);
4244 // If all Altivec parameters fit in registers, as they usually do,
4245 // they get stack space following the non-Altivec parameters. We
4246 // don't track this here because nobody below needs it.
4247 // If there are more Altivec parameters than fit in registers emit
4249 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4251 // Offset is aligned; skip 1st 12 params which go in V registers.
4252 ArgOffset = ((ArgOffset+15)/16)*16;
4254 for (unsigned i = 0; i != NumOps; ++i) {
4255 SDValue Arg = OutVals[i];
4256 EVT ArgType = Outs[i].VT;
4257 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4258 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4261 // We are emitting Altivec params in order.
4262 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4263 isPPC64, isTailCall, true, MemOpChains,
4264 TailCallArguments, dl);
4271 if (!MemOpChains.empty())
4272 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4273 &MemOpChains[0], MemOpChains.size());
4275 // On Darwin, R12 must contain the address of an indirect callee. This does
4276 // not mean the MTCTR instruction must use R12; it's easier to model this as
4277 // an extra parameter, so do that.
4279 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4280 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4281 !isBLACompatibleAddress(Callee, DAG))
4282 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4283 PPC::R12), Callee));
4285 // Build a sequence of copy-to-reg nodes chained together with token chain
4286 // and flag operands which copy the outgoing args into the appropriate regs.
4288 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4289 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4290 RegsToPass[i].second, InFlag);
4291 InFlag = Chain.getValue(1);
4295 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4296 FPOp, true, TailCallArguments);
4298 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4299 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4304 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4305 MachineFunction &MF, bool isVarArg,
4306 const SmallVectorImpl<ISD::OutputArg> &Outs,
4307 LLVMContext &Context) const {
4308 SmallVector<CCValAssign, 16> RVLocs;
4309 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4311 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4315 PPCTargetLowering::LowerReturn(SDValue Chain,
4316 CallingConv::ID CallConv, bool isVarArg,
4317 const SmallVectorImpl<ISD::OutputArg> &Outs,
4318 const SmallVectorImpl<SDValue> &OutVals,
4319 DebugLoc dl, SelectionDAG &DAG) const {
4321 SmallVector<CCValAssign, 16> RVLocs;
4322 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4323 getTargetMachine(), RVLocs, *DAG.getContext());
4324 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4326 // If this is the first return lowered for this function, add the regs to the
4327 // liveout set for the function.
4328 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
4329 for (unsigned i = 0; i != RVLocs.size(); ++i)
4330 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
4335 // Copy the result values into the output registers.
4336 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4337 CCValAssign &VA = RVLocs[i];
4338 assert(VA.isRegLoc() && "Can only return in registers!");
4340 SDValue Arg = OutVals[i];
4342 switch (VA.getLocInfo()) {
4343 default: llvm_unreachable("Unknown loc info!");
4344 case CCValAssign::Full: break;
4345 case CCValAssign::AExt:
4346 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
4348 case CCValAssign::ZExt:
4349 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
4351 case CCValAssign::SExt:
4352 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
4356 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
4357 Flag = Chain.getValue(1);
4361 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
4363 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
4366 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4367 const PPCSubtarget &Subtarget) const {
4368 // When we pop the dynamic allocation we need to restore the SP link.
4369 DebugLoc dl = Op.getDebugLoc();
4371 // Get the corect type for pointers.
4372 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4374 // Construct the stack pointer operand.
4375 bool isPPC64 = Subtarget.isPPC64();
4376 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4377 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4379 // Get the operands for the STACKRESTORE.
4380 SDValue Chain = Op.getOperand(0);
4381 SDValue SaveSP = Op.getOperand(1);
4383 // Load the old link SP.
4384 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4385 MachinePointerInfo(),
4386 false, false, false, 0);
4388 // Restore the stack pointer.
4389 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4391 // Store the old link SP.
4392 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4399 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4400 MachineFunction &MF = DAG.getMachineFunction();
4401 bool isPPC64 = PPCSubTarget.isPPC64();
4402 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4403 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4405 // Get current frame pointer save index. The users of this index will be
4406 // primarily DYNALLOC instructions.
4407 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4408 int RASI = FI->getReturnAddrSaveIndex();
4410 // If the frame pointer save index hasn't been defined yet.
4412 // Find out what the fix offset of the frame pointer save area.
4413 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4414 // Allocate the frame index for frame pointer save area.
4415 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4417 FI->setReturnAddrSaveIndex(RASI);
4419 return DAG.getFrameIndex(RASI, PtrVT);
4423 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4424 MachineFunction &MF = DAG.getMachineFunction();
4425 bool isPPC64 = PPCSubTarget.isPPC64();
4426 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4427 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4429 // Get current frame pointer save index. The users of this index will be
4430 // primarily DYNALLOC instructions.
4431 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4432 int FPSI = FI->getFramePointerSaveIndex();
4434 // If the frame pointer save index hasn't been defined yet.
4436 // Find out what the fix offset of the frame pointer save area.
4437 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4440 // Allocate the frame index for frame pointer save area.
4441 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4443 FI->setFramePointerSaveIndex(FPSI);
4445 return DAG.getFrameIndex(FPSI, PtrVT);
4448 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4450 const PPCSubtarget &Subtarget) const {
4452 SDValue Chain = Op.getOperand(0);
4453 SDValue Size = Op.getOperand(1);
4454 DebugLoc dl = Op.getDebugLoc();
4456 // Get the corect type for pointers.
4457 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4459 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4460 DAG.getConstant(0, PtrVT), Size);
4461 // Construct a node for the frame pointer save index.
4462 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4463 // Build a DYNALLOC node.
4464 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4465 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4466 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4469 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4471 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4472 // Not FP? Not a fsel.
4473 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4474 !Op.getOperand(2).getValueType().isFloatingPoint())
4477 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4479 // Cannot handle SETEQ/SETNE.
4480 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4482 EVT ResVT = Op.getValueType();
4483 EVT CmpVT = Op.getOperand(0).getValueType();
4484 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4485 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4486 DebugLoc dl = Op.getDebugLoc();
4488 // If the RHS of the comparison is a 0.0, we don't need to do the
4489 // subtraction at all.
4490 if (isFloatingPointZero(RHS))
4492 default: break; // SETUO etc aren't handled by fsel.
4495 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4498 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4499 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4500 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4503 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4506 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4507 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4508 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4509 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4514 default: break; // SETUO etc aren't handled by fsel.
4517 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4518 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4519 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4520 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4523 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4524 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4525 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4526 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4529 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4530 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4531 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4532 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4535 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4536 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4537 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4538 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4543 // FIXME: Split this code up when LegalizeDAGTypes lands.
4544 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4545 DebugLoc dl) const {
4546 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4547 SDValue Src = Op.getOperand(0);
4548 if (Src.getValueType() == MVT::f32)
4549 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4552 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4553 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4555 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4560 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4564 // Convert the FP value to an int value through memory.
4565 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4567 // Emit a store to the stack slot.
4568 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4569 MachinePointerInfo(), false, false, 0);
4571 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4573 if (Op.getValueType() == MVT::i32)
4574 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4575 DAG.getConstant(4, FIPtr.getValueType()));
4576 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4577 false, false, false, 0);
4580 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4581 SelectionDAG &DAG) const {
4582 DebugLoc dl = Op.getDebugLoc();
4583 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4584 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4587 if (Op.getOperand(0).getValueType() == MVT::i64) {
4588 SDValue SINT = Op.getOperand(0);
4589 // When converting to single-precision, we actually need to convert
4590 // to double-precision first and then round to single-precision.
4591 // To avoid double-rounding effects during that operation, we have
4592 // to prepare the input operand. Bits that might be truncated when
4593 // converting to double-precision are replaced by a bit that won't
4594 // be lost at this stage, but is below the single-precision rounding
4597 // However, if -enable-unsafe-fp-math is in effect, accept double
4598 // rounding to avoid the extra overhead.
4599 if (Op.getValueType() == MVT::f32 &&
4600 !DAG.getTarget().Options.UnsafeFPMath) {
4602 // Twiddle input to make sure the low 11 bits are zero. (If this
4603 // is the case, we are guaranteed the value will fit into the 53 bit
4604 // mantissa of an IEEE double-precision value without rounding.)
4605 // If any of those low 11 bits were not zero originally, make sure
4606 // bit 12 (value 2048) is set instead, so that the final rounding
4607 // to single-precision gets the correct result.
4608 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4609 SINT, DAG.getConstant(2047, MVT::i64));
4610 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4611 Round, DAG.getConstant(2047, MVT::i64));
4612 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4613 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4614 Round, DAG.getConstant(-2048, MVT::i64));
4616 // However, we cannot use that value unconditionally: if the magnitude
4617 // of the input value is small, the bit-twiddling we did above might
4618 // end up visibly changing the output. Fortunately, in that case, we
4619 // don't need to twiddle bits since the original input will convert
4620 // exactly to double-precision floating-point already. Therefore,
4621 // construct a conditional to use the original value if the top 11
4622 // bits are all sign-bit copies, and use the rounded value computed
4624 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4625 SINT, DAG.getConstant(53, MVT::i32));
4626 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4627 Cond, DAG.getConstant(1, MVT::i64));
4628 Cond = DAG.getSetCC(dl, MVT::i32,
4629 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4631 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4633 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4634 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4635 if (Op.getValueType() == MVT::f32)
4636 FP = DAG.getNode(ISD::FP_ROUND, dl,
4637 MVT::f32, FP, DAG.getIntPtrConstant(0));
4641 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4642 "Unhandled SINT_TO_FP type in custom expander!");
4643 // Since we only generate this in 64-bit mode, we can take advantage of
4644 // 64-bit registers. In particular, sign extend the input value into the
4645 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4646 // then lfd it and fcfid it.
4647 MachineFunction &MF = DAG.getMachineFunction();
4648 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4649 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4650 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4651 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4653 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4656 // STD the extended value into the stack slot.
4657 MachineMemOperand *MMO =
4658 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4659 MachineMemOperand::MOStore, 8, 8);
4660 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4662 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4663 Ops, 4, MVT::i64, MMO);
4664 // Load the value as a double.
4665 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4666 false, false, false, 0);
4668 // FCFID it and return it.
4669 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4670 if (Op.getValueType() == MVT::f32)
4671 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4675 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4676 SelectionDAG &DAG) const {
4677 DebugLoc dl = Op.getDebugLoc();
4679 The rounding mode is in bits 30:31 of FPSR, and has the following
4686 FLT_ROUNDS, on the other hand, expects the following:
4693 To perform the conversion, we do:
4694 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4697 MachineFunction &MF = DAG.getMachineFunction();
4698 EVT VT = Op.getValueType();
4699 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4700 std::vector<EVT> NodeTys;
4701 SDValue MFFSreg, InFlag;
4703 // Save FP Control Word to register
4704 NodeTys.push_back(MVT::f64); // return register
4705 NodeTys.push_back(MVT::Glue); // unused in this context
4706 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4708 // Save FP register to stack slot
4709 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4710 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4711 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4712 StackSlot, MachinePointerInfo(), false, false,0);
4714 // Load FP Control Word from low 32 bits of stack slot.
4715 SDValue Four = DAG.getConstant(4, PtrVT);
4716 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4717 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4718 false, false, false, 0);
4720 // Transform as necessary
4722 DAG.getNode(ISD::AND, dl, MVT::i32,
4723 CWD, DAG.getConstant(3, MVT::i32));
4725 DAG.getNode(ISD::SRL, dl, MVT::i32,
4726 DAG.getNode(ISD::AND, dl, MVT::i32,
4727 DAG.getNode(ISD::XOR, dl, MVT::i32,
4728 CWD, DAG.getConstant(3, MVT::i32)),
4729 DAG.getConstant(3, MVT::i32)),
4730 DAG.getConstant(1, MVT::i32));
4733 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4735 return DAG.getNode((VT.getSizeInBits() < 16 ?
4736 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4739 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4740 EVT VT = Op.getValueType();
4741 unsigned BitWidth = VT.getSizeInBits();
4742 DebugLoc dl = Op.getDebugLoc();
4743 assert(Op.getNumOperands() == 3 &&
4744 VT == Op.getOperand(1).getValueType() &&
4747 // Expand into a bunch of logical ops. Note that these ops
4748 // depend on the PPC behavior for oversized shift amounts.
4749 SDValue Lo = Op.getOperand(0);
4750 SDValue Hi = Op.getOperand(1);
4751 SDValue Amt = Op.getOperand(2);
4752 EVT AmtVT = Amt.getValueType();
4754 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4755 DAG.getConstant(BitWidth, AmtVT), Amt);
4756 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4757 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4758 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4759 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4760 DAG.getConstant(-BitWidth, AmtVT));
4761 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4762 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4763 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4764 SDValue OutOps[] = { OutLo, OutHi };
4765 return DAG.getMergeValues(OutOps, 2, dl);
4768 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4769 EVT VT = Op.getValueType();
4770 DebugLoc dl = Op.getDebugLoc();
4771 unsigned BitWidth = VT.getSizeInBits();
4772 assert(Op.getNumOperands() == 3 &&
4773 VT == Op.getOperand(1).getValueType() &&
4776 // Expand into a bunch of logical ops. Note that these ops
4777 // depend on the PPC behavior for oversized shift amounts.
4778 SDValue Lo = Op.getOperand(0);
4779 SDValue Hi = Op.getOperand(1);
4780 SDValue Amt = Op.getOperand(2);
4781 EVT AmtVT = Amt.getValueType();
4783 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4784 DAG.getConstant(BitWidth, AmtVT), Amt);
4785 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4786 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4787 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4788 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4789 DAG.getConstant(-BitWidth, AmtVT));
4790 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4791 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4792 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4793 SDValue OutOps[] = { OutLo, OutHi };
4794 return DAG.getMergeValues(OutOps, 2, dl);
4797 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4798 DebugLoc dl = Op.getDebugLoc();
4799 EVT VT = Op.getValueType();
4800 unsigned BitWidth = VT.getSizeInBits();
4801 assert(Op.getNumOperands() == 3 &&
4802 VT == Op.getOperand(1).getValueType() &&
4805 // Expand into a bunch of logical ops, followed by a select_cc.
4806 SDValue Lo = Op.getOperand(0);
4807 SDValue Hi = Op.getOperand(1);
4808 SDValue Amt = Op.getOperand(2);
4809 EVT AmtVT = Amt.getValueType();
4811 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4812 DAG.getConstant(BitWidth, AmtVT), Amt);
4813 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4814 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4815 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4816 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4817 DAG.getConstant(-BitWidth, AmtVT));
4818 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4819 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4820 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4821 Tmp4, Tmp6, ISD::SETLE);
4822 SDValue OutOps[] = { OutLo, OutHi };
4823 return DAG.getMergeValues(OutOps, 2, dl);
4826 //===----------------------------------------------------------------------===//
4827 // Vector related lowering.
4830 /// BuildSplatI - Build a canonical splati of Val with an element size of
4831 /// SplatSize. Cast the result to VT.
4832 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4833 SelectionDAG &DAG, DebugLoc dl) {
4834 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4836 static const EVT VTys[] = { // canonical VT to use for each size.
4837 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4840 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4842 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4846 EVT CanonicalVT = VTys[SplatSize-1];
4848 // Build a canonical splat for this value.
4849 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4850 SmallVector<SDValue, 8> Ops;
4851 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4852 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4853 &Ops[0], Ops.size());
4854 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4857 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4858 /// specified intrinsic ID.
4859 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4860 SelectionDAG &DAG, DebugLoc dl,
4861 EVT DestVT = MVT::Other) {
4862 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4863 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4864 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4867 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4868 /// specified intrinsic ID.
4869 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4870 SDValue Op2, SelectionDAG &DAG,
4871 DebugLoc dl, EVT DestVT = MVT::Other) {
4872 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4873 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4874 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4878 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4879 /// amount. The result has the specified value type.
4880 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4881 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4882 // Force LHS/RHS to be the right type.
4883 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4884 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4887 for (unsigned i = 0; i != 16; ++i)
4889 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4890 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4893 // If this is a case we can't handle, return null and let the default
4894 // expansion code take care of it. If we CAN select this case, and if it
4895 // selects to a single instruction, return Op. Otherwise, if we can codegen
4896 // this case more efficiently than a constant pool load, lower it to the
4897 // sequence of ops that should be used.
4898 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4899 SelectionDAG &DAG) const {
4900 DebugLoc dl = Op.getDebugLoc();
4901 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4902 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4904 // Check if this is a splat of a constant value.
4905 APInt APSplatBits, APSplatUndef;
4906 unsigned SplatBitSize;
4908 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4909 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4912 unsigned SplatBits = APSplatBits.getZExtValue();
4913 unsigned SplatUndef = APSplatUndef.getZExtValue();
4914 unsigned SplatSize = SplatBitSize / 8;
4916 // First, handle single instruction cases.
4919 if (SplatBits == 0) {
4920 // Canonicalize all zero vectors to be v4i32.
4921 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4922 SDValue Z = DAG.getConstant(0, MVT::i32);
4923 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4924 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4929 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4930 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4932 if (SextVal >= -16 && SextVal <= 15)
4933 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4936 // Two instruction sequences.
4938 // If this value is in the range [-32,30] and is even, use:
4939 // tmp = VSPLTI[bhw], result = add tmp, tmp
4940 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4941 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4942 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4943 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4946 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4947 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4949 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4950 // Make -1 and vspltisw -1:
4951 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4953 // Make the VSLW intrinsic, computing 0x8000_0000.
4954 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4957 // xor by OnesV to invert it.
4958 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4959 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4962 // Check to see if this is a wide variety of vsplti*, binop self cases.
4963 static const signed char SplatCsts[] = {
4964 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4965 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4968 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4969 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4970 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4971 int i = SplatCsts[idx];
4973 // Figure out what shift amount will be used by altivec if shifted by i in
4975 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4977 // vsplti + shl self.
4978 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4979 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4980 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4981 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4982 Intrinsic::ppc_altivec_vslw
4984 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4985 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4988 // vsplti + srl self.
4989 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
4990 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4991 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4992 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
4993 Intrinsic::ppc_altivec_vsrw
4995 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
4996 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4999 // vsplti + sra self.
5000 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5001 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5002 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5003 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
5004 Intrinsic::ppc_altivec_vsraw
5006 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5007 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5010 // vsplti + rol self.
5011 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
5012 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
5013 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5014 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5015 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
5016 Intrinsic::ppc_altivec_vrlw
5018 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5019 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5022 // t = vsplti c, result = vsldoi t, t, 1
5023 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
5024 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5025 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
5027 // t = vsplti c, result = vsldoi t, t, 2
5028 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
5029 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5030 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
5032 // t = vsplti c, result = vsldoi t, t, 3
5033 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
5034 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5035 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
5039 // Three instruction sequences.
5041 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
5042 if (SextVal >= 0 && SextVal <= 31) {
5043 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
5044 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
5045 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
5046 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
5048 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
5049 if (SextVal >= -31 && SextVal <= 0) {
5050 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
5051 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
5052 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
5053 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
5059 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5060 /// the specified operations to build the shuffle.
5061 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5062 SDValue RHS, SelectionDAG &DAG,
5064 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5065 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5066 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5069 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5081 if (OpNum == OP_COPY) {
5082 if (LHSID == (1*9+2)*9+3) return LHS;
5083 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5087 SDValue OpLHS, OpRHS;
5088 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5089 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5093 default: llvm_unreachable("Unknown i32 permute!");
5095 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5096 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5097 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5098 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5101 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5102 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5103 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5104 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5107 for (unsigned i = 0; i != 16; ++i)
5108 ShufIdxs[i] = (i&3)+0;
5111 for (unsigned i = 0; i != 16; ++i)
5112 ShufIdxs[i] = (i&3)+4;
5115 for (unsigned i = 0; i != 16; ++i)
5116 ShufIdxs[i] = (i&3)+8;
5119 for (unsigned i = 0; i != 16; ++i)
5120 ShufIdxs[i] = (i&3)+12;
5123 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5125 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5127 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5129 EVT VT = OpLHS.getValueType();
5130 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5131 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5132 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5133 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5136 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5137 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5138 /// return the code it can be lowered into. Worst case, it can always be
5139 /// lowered into a vperm.
5140 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5141 SelectionDAG &DAG) const {
5142 DebugLoc dl = Op.getDebugLoc();
5143 SDValue V1 = Op.getOperand(0);
5144 SDValue V2 = Op.getOperand(1);
5145 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5146 EVT VT = Op.getValueType();
5148 // Cases that are handled by instructions that take permute immediates
5149 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5150 // selected by the instruction selector.
5151 if (V2.getOpcode() == ISD::UNDEF) {
5152 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5153 PPC::isSplatShuffleMask(SVOp, 2) ||
5154 PPC::isSplatShuffleMask(SVOp, 4) ||
5155 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5156 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5157 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5158 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5159 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5160 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5161 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5162 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5163 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5168 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5169 // and produce a fixed permutation. If any of these match, do not lower to
5171 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5172 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5173 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5174 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5175 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5176 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5177 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5178 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5179 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5182 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5183 // perfect shuffle table to emit an optimal matching sequence.
5184 ArrayRef<int> PermMask = SVOp->getMask();
5186 unsigned PFIndexes[4];
5187 bool isFourElementShuffle = true;
5188 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5189 unsigned EltNo = 8; // Start out undef.
5190 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5191 if (PermMask[i*4+j] < 0)
5192 continue; // Undef, ignore it.
5194 unsigned ByteSource = PermMask[i*4+j];
5195 if ((ByteSource & 3) != j) {
5196 isFourElementShuffle = false;
5201 EltNo = ByteSource/4;
5202 } else if (EltNo != ByteSource/4) {
5203 isFourElementShuffle = false;
5207 PFIndexes[i] = EltNo;
5210 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5211 // perfect shuffle vector to determine if it is cost effective to do this as
5212 // discrete instructions, or whether we should use a vperm.
5213 if (isFourElementShuffle) {
5214 // Compute the index in the perfect shuffle table.
5215 unsigned PFTableIndex =
5216 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5218 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5219 unsigned Cost = (PFEntry >> 30);
5221 // Determining when to avoid vperm is tricky. Many things affect the cost
5222 // of vperm, particularly how many times the perm mask needs to be computed.
5223 // For example, if the perm mask can be hoisted out of a loop or is already
5224 // used (perhaps because there are multiple permutes with the same shuffle
5225 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5226 // the loop requires an extra register.
5228 // As a compromise, we only emit discrete instructions if the shuffle can be
5229 // generated in 3 or fewer operations. When we have loop information
5230 // available, if this block is within a loop, we should avoid using vperm
5231 // for 3-operation perms and use a constant pool load instead.
5233 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5236 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5237 // vector that will get spilled to the constant pool.
5238 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5240 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5241 // that it is in input element units, not in bytes. Convert now.
5242 EVT EltVT = V1.getValueType().getVectorElementType();
5243 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5245 SmallVector<SDValue, 16> ResultMask;
5246 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5247 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5249 for (unsigned j = 0; j != BytesPerElement; ++j)
5250 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5254 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5255 &ResultMask[0], ResultMask.size());
5256 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5259 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5260 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5261 /// information about the intrinsic.
5262 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5264 unsigned IntrinsicID =
5265 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5268 switch (IntrinsicID) {
5269 default: return false;
5270 // Comparison predicates.
5271 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5272 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5273 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5274 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5275 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5276 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5277 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5278 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5279 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5280 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5281 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5282 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5283 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5285 // Normal Comparisons.
5286 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5287 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5288 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5289 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5290 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5291 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5292 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5293 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5294 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5295 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5296 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5297 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5298 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5303 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5304 /// lower, do it, otherwise return null.
5305 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5306 SelectionDAG &DAG) const {
5307 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5308 // opcode number of the comparison.
5309 DebugLoc dl = Op.getDebugLoc();
5312 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5313 return SDValue(); // Don't custom lower most intrinsics.
5315 // If this is a non-dot comparison, make the VCMP node and we are done.
5317 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5318 Op.getOperand(1), Op.getOperand(2),
5319 DAG.getConstant(CompareOpc, MVT::i32));
5320 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5323 // Create the PPCISD altivec 'dot' comparison node.
5325 Op.getOperand(2), // LHS
5326 Op.getOperand(3), // RHS
5327 DAG.getConstant(CompareOpc, MVT::i32)
5329 std::vector<EVT> VTs;
5330 VTs.push_back(Op.getOperand(2).getValueType());
5331 VTs.push_back(MVT::Glue);
5332 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5334 // Now that we have the comparison, emit a copy from the CR to a GPR.
5335 // This is flagged to the above dot comparison.
5336 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5337 DAG.getRegister(PPC::CR6, MVT::i32),
5338 CompNode.getValue(1));
5340 // Unpack the result based on how the target uses it.
5341 unsigned BitNo; // Bit # of CR6.
5342 bool InvertBit; // Invert result?
5343 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5344 default: // Can't happen, don't crash on invalid number though.
5345 case 0: // Return the value of the EQ bit of CR6.
5346 BitNo = 0; InvertBit = false;
5348 case 1: // Return the inverted value of the EQ bit of CR6.
5349 BitNo = 0; InvertBit = true;
5351 case 2: // Return the value of the LT bit of CR6.
5352 BitNo = 2; InvertBit = false;
5354 case 3: // Return the inverted value of the LT bit of CR6.
5355 BitNo = 2; InvertBit = true;
5359 // Shift the bit into the low position.
5360 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5361 DAG.getConstant(8-(3-BitNo), MVT::i32));
5363 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5364 DAG.getConstant(1, MVT::i32));
5366 // If we are supposed to, toggle the bit.
5368 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5369 DAG.getConstant(1, MVT::i32));
5373 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5374 SelectionDAG &DAG) const {
5375 DebugLoc dl = Op.getDebugLoc();
5376 // Create a stack slot that is 16-byte aligned.
5377 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5378 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5379 EVT PtrVT = getPointerTy();
5380 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5382 // Store the input value into Value#0 of the stack slot.
5383 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5384 Op.getOperand(0), FIdx, MachinePointerInfo(),
5387 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5388 false, false, false, 0);
5391 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5392 DebugLoc dl = Op.getDebugLoc();
5393 if (Op.getValueType() == MVT::v4i32) {
5394 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5396 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5397 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5399 SDValue RHSSwap = // = vrlw RHS, 16
5400 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5402 // Shrinkify inputs to v8i16.
5403 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5404 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5405 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5407 // Low parts multiplied together, generating 32-bit results (we ignore the
5409 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5410 LHS, RHS, DAG, dl, MVT::v4i32);
5412 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5413 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5414 // Shift the high parts up 16 bits.
5415 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5417 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5418 } else if (Op.getValueType() == MVT::v8i16) {
5419 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5421 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5423 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5424 LHS, RHS, Zero, DAG, dl);
5425 } else if (Op.getValueType() == MVT::v16i8) {
5426 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5428 // Multiply the even 8-bit parts, producing 16-bit sums.
5429 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5430 LHS, RHS, DAG, dl, MVT::v8i16);
5431 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5433 // Multiply the odd 8-bit parts, producing 16-bit sums.
5434 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5435 LHS, RHS, DAG, dl, MVT::v8i16);
5436 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5438 // Merge the results together.
5440 for (unsigned i = 0; i != 8; ++i) {
5442 Ops[i*2+1] = 2*i+1+16;
5444 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5446 llvm_unreachable("Unknown mul to lower!");
5450 /// LowerOperation - Provide custom lowering hooks for some operations.
5452 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5453 switch (Op.getOpcode()) {
5454 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5455 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5456 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5457 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5458 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5459 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5460 case ISD::SETCC: return LowerSETCC(Op, DAG);
5461 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5462 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5464 return LowerVASTART(Op, DAG, PPCSubTarget);
5467 return LowerVAARG(Op, DAG, PPCSubTarget);
5469 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5470 case ISD::DYNAMIC_STACKALLOC:
5471 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5473 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5474 case ISD::FP_TO_UINT:
5475 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5477 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5478 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5480 // Lower 64-bit shifts.
5481 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5482 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5483 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5485 // Vector-related lowering.
5486 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5487 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5488 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5489 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5490 case ISD::MUL: return LowerMUL(Op, DAG);
5492 // Frame & Return address.
5493 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5494 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5498 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5499 SmallVectorImpl<SDValue>&Results,
5500 SelectionDAG &DAG) const {
5501 const TargetMachine &TM = getTargetMachine();
5502 DebugLoc dl = N->getDebugLoc();
5503 switch (N->getOpcode()) {
5505 llvm_unreachable("Do not know how to custom type legalize this operation!");
5507 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5508 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5511 EVT VT = N->getValueType(0);
5513 if (VT == MVT::i64) {
5514 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5516 Results.push_back(NewNode);
5517 Results.push_back(NewNode.getValue(1));
5521 case ISD::FP_ROUND_INREG: {
5522 assert(N->getValueType(0) == MVT::ppcf128);
5523 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5524 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5525 MVT::f64, N->getOperand(0),
5526 DAG.getIntPtrConstant(0));
5527 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5528 MVT::f64, N->getOperand(0),
5529 DAG.getIntPtrConstant(1));
5531 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5532 // of the long double, and puts FPSCR back the way it was. We do not
5533 // actually model FPSCR.
5534 std::vector<EVT> NodeTys;
5535 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5537 NodeTys.push_back(MVT::f64); // Return register
5538 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5539 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5540 MFFSreg = Result.getValue(0);
5541 InFlag = Result.getValue(1);
5544 NodeTys.push_back(MVT::Glue); // Returns a flag
5545 Ops[0] = DAG.getConstant(31, MVT::i32);
5547 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5548 InFlag = Result.getValue(0);
5551 NodeTys.push_back(MVT::Glue); // Returns a flag
5552 Ops[0] = DAG.getConstant(30, MVT::i32);
5554 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5555 InFlag = Result.getValue(0);
5558 NodeTys.push_back(MVT::f64); // result of add
5559 NodeTys.push_back(MVT::Glue); // Returns a flag
5563 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5564 FPreg = Result.getValue(0);
5565 InFlag = Result.getValue(1);
5568 NodeTys.push_back(MVT::f64);
5569 Ops[0] = DAG.getConstant(1, MVT::i32);
5573 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5574 FPreg = Result.getValue(0);
5576 // We know the low half is about to be thrown away, so just use something
5578 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5582 case ISD::FP_TO_SINT:
5583 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5589 //===----------------------------------------------------------------------===//
5590 // Other Lowering Code
5591 //===----------------------------------------------------------------------===//
5594 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5595 bool is64bit, unsigned BinOpcode) const {
5596 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5597 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5599 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5600 MachineFunction *F = BB->getParent();
5601 MachineFunction::iterator It = BB;
5604 unsigned dest = MI->getOperand(0).getReg();
5605 unsigned ptrA = MI->getOperand(1).getReg();
5606 unsigned ptrB = MI->getOperand(2).getReg();
5607 unsigned incr = MI->getOperand(3).getReg();
5608 DebugLoc dl = MI->getDebugLoc();
5610 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5611 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5612 F->insert(It, loopMBB);
5613 F->insert(It, exitMBB);
5614 exitMBB->splice(exitMBB->begin(), BB,
5615 llvm::next(MachineBasicBlock::iterator(MI)),
5617 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5619 MachineRegisterInfo &RegInfo = F->getRegInfo();
5620 unsigned TmpReg = (!BinOpcode) ? incr :
5621 RegInfo.createVirtualRegister(
5622 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5623 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5627 // fallthrough --> loopMBB
5628 BB->addSuccessor(loopMBB);
5631 // l[wd]arx dest, ptr
5632 // add r0, dest, incr
5633 // st[wd]cx. r0, ptr
5635 // fallthrough --> exitMBB
5637 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5638 .addReg(ptrA).addReg(ptrB);
5640 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5641 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5642 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5643 BuildMI(BB, dl, TII->get(PPC::BCC))
5644 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5645 BB->addSuccessor(loopMBB);
5646 BB->addSuccessor(exitMBB);
5655 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5656 MachineBasicBlock *BB,
5657 bool is8bit, // operation
5658 unsigned BinOpcode) const {
5659 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5660 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5661 // In 64 bit mode we have to use 64 bits for addresses, even though the
5662 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5663 // registers without caring whether they're 32 or 64, but here we're
5664 // doing actual arithmetic on the addresses.
5665 bool is64bit = PPCSubTarget.isPPC64();
5666 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5668 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5669 MachineFunction *F = BB->getParent();
5670 MachineFunction::iterator It = BB;
5673 unsigned dest = MI->getOperand(0).getReg();
5674 unsigned ptrA = MI->getOperand(1).getReg();
5675 unsigned ptrB = MI->getOperand(2).getReg();
5676 unsigned incr = MI->getOperand(3).getReg();
5677 DebugLoc dl = MI->getDebugLoc();
5679 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5680 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5681 F->insert(It, loopMBB);
5682 F->insert(It, exitMBB);
5683 exitMBB->splice(exitMBB->begin(), BB,
5684 llvm::next(MachineBasicBlock::iterator(MI)),
5686 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5688 MachineRegisterInfo &RegInfo = F->getRegInfo();
5689 const TargetRegisterClass *RC =
5690 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5691 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5692 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5693 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5694 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5695 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5696 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5697 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5698 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5699 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5700 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5701 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5702 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5704 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5708 // fallthrough --> loopMBB
5709 BB->addSuccessor(loopMBB);
5711 // The 4-byte load must be aligned, while a char or short may be
5712 // anywhere in the word. Hence all this nasty bookkeeping code.
5713 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5714 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5715 // xori shift, shift1, 24 [16]
5716 // rlwinm ptr, ptr1, 0, 0, 29
5717 // slw incr2, incr, shift
5718 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5719 // slw mask, mask2, shift
5721 // lwarx tmpDest, ptr
5722 // add tmp, tmpDest, incr2
5723 // andc tmp2, tmpDest, mask
5724 // and tmp3, tmp, mask
5725 // or tmp4, tmp3, tmp2
5728 // fallthrough --> exitMBB
5729 // srw dest, tmpDest, shift
5730 if (ptrA != ZeroReg) {
5731 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5732 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5733 .addReg(ptrA).addReg(ptrB);
5737 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5738 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5739 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5740 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5742 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5743 .addReg(Ptr1Reg).addImm(0).addImm(61);
5745 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5746 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5747 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5748 .addReg(incr).addReg(ShiftReg);
5750 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5752 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5753 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5755 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5756 .addReg(Mask2Reg).addReg(ShiftReg);
5759 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5760 .addReg(ZeroReg).addReg(PtrReg);
5762 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5763 .addReg(Incr2Reg).addReg(TmpDestReg);
5764 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5765 .addReg(TmpDestReg).addReg(MaskReg);
5766 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5767 .addReg(TmpReg).addReg(MaskReg);
5768 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5769 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5770 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5771 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5772 BuildMI(BB, dl, TII->get(PPC::BCC))
5773 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5774 BB->addSuccessor(loopMBB);
5775 BB->addSuccessor(exitMBB);
5780 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5786 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5787 MachineBasicBlock *BB) const {
5788 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5790 // To "insert" these instructions we actually have to insert their
5791 // control-flow patterns.
5792 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5793 MachineFunction::iterator It = BB;
5796 MachineFunction *F = BB->getParent();
5798 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5799 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5800 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5801 PPC::ISEL8 : PPC::ISEL;
5802 unsigned SelectPred = MI->getOperand(4).getImm();
5803 DebugLoc dl = MI->getDebugLoc();
5805 // The SelectPred is ((BI << 5) | BO) for a BCC
5806 unsigned BO = SelectPred & 0xF;
5807 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5809 unsigned TrueOpNo, FalseOpNo;
5816 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5819 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5820 .addReg(MI->getOperand(TrueOpNo).getReg())
5821 .addReg(MI->getOperand(FalseOpNo).getReg())
5822 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5823 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5824 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5825 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5826 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5827 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5830 // The incoming instruction knows the destination vreg to set, the
5831 // condition code register to branch on, the true/false values to
5832 // select between, and a branch opcode to use.
5837 // cmpTY ccX, r1, r2
5839 // fallthrough --> copy0MBB
5840 MachineBasicBlock *thisMBB = BB;
5841 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5842 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5843 unsigned SelectPred = MI->getOperand(4).getImm();
5844 DebugLoc dl = MI->getDebugLoc();
5845 F->insert(It, copy0MBB);
5846 F->insert(It, sinkMBB);
5848 // Transfer the remainder of BB and its successor edges to sinkMBB.
5849 sinkMBB->splice(sinkMBB->begin(), BB,
5850 llvm::next(MachineBasicBlock::iterator(MI)),
5852 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5854 // Next, add the true and fallthrough blocks as its successors.
5855 BB->addSuccessor(copy0MBB);
5856 BB->addSuccessor(sinkMBB);
5858 BuildMI(BB, dl, TII->get(PPC::BCC))
5859 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5862 // %FalseValue = ...
5863 // # fallthrough to sinkMBB
5866 // Update machine-CFG edges
5867 BB->addSuccessor(sinkMBB);
5870 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5873 BuildMI(*BB, BB->begin(), dl,
5874 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5875 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5876 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5878 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5879 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5880 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5881 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5882 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5883 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5884 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5885 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5887 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5888 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5889 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5890 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5891 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5892 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5893 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5894 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5896 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5897 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5898 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5899 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5900 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5901 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5902 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5903 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5905 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5906 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5907 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5908 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5909 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5910 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5911 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5912 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5914 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5915 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5916 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5917 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5918 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5919 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5920 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5921 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5923 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5924 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5925 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5926 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5927 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5928 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5929 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5930 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5932 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5933 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5934 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5935 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5936 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5937 BB = EmitAtomicBinary(MI, BB, false, 0);
5938 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5939 BB = EmitAtomicBinary(MI, BB, true, 0);
5941 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5942 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5943 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5945 unsigned dest = MI->getOperand(0).getReg();
5946 unsigned ptrA = MI->getOperand(1).getReg();
5947 unsigned ptrB = MI->getOperand(2).getReg();
5948 unsigned oldval = MI->getOperand(3).getReg();
5949 unsigned newval = MI->getOperand(4).getReg();
5950 DebugLoc dl = MI->getDebugLoc();
5952 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5953 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5954 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5955 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5956 F->insert(It, loop1MBB);
5957 F->insert(It, loop2MBB);
5958 F->insert(It, midMBB);
5959 F->insert(It, exitMBB);
5960 exitMBB->splice(exitMBB->begin(), BB,
5961 llvm::next(MachineBasicBlock::iterator(MI)),
5963 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5967 // fallthrough --> loopMBB
5968 BB->addSuccessor(loop1MBB);
5971 // l[wd]arx dest, ptr
5972 // cmp[wd] dest, oldval
5975 // st[wd]cx. newval, ptr
5979 // st[wd]cx. dest, ptr
5982 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5983 .addReg(ptrA).addReg(ptrB);
5984 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
5985 .addReg(oldval).addReg(dest);
5986 BuildMI(BB, dl, TII->get(PPC::BCC))
5987 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5988 BB->addSuccessor(loop2MBB);
5989 BB->addSuccessor(midMBB);
5992 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5993 .addReg(newval).addReg(ptrA).addReg(ptrB);
5994 BuildMI(BB, dl, TII->get(PPC::BCC))
5995 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5996 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5997 BB->addSuccessor(loop1MBB);
5998 BB->addSuccessor(exitMBB);
6001 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6002 .addReg(dest).addReg(ptrA).addReg(ptrB);
6003 BB->addSuccessor(exitMBB);
6008 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
6009 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
6010 // We must use 64-bit registers for addresses when targeting 64-bit,
6011 // since we're actually doing arithmetic on them. Other registers
6013 bool is64bit = PPCSubTarget.isPPC64();
6014 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
6016 unsigned dest = MI->getOperand(0).getReg();
6017 unsigned ptrA = MI->getOperand(1).getReg();
6018 unsigned ptrB = MI->getOperand(2).getReg();
6019 unsigned oldval = MI->getOperand(3).getReg();
6020 unsigned newval = MI->getOperand(4).getReg();
6021 DebugLoc dl = MI->getDebugLoc();
6023 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6024 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6025 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6026 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6027 F->insert(It, loop1MBB);
6028 F->insert(It, loop2MBB);
6029 F->insert(It, midMBB);
6030 F->insert(It, exitMBB);
6031 exitMBB->splice(exitMBB->begin(), BB,
6032 llvm::next(MachineBasicBlock::iterator(MI)),
6034 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6036 MachineRegisterInfo &RegInfo = F->getRegInfo();
6037 const TargetRegisterClass *RC =
6038 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
6039 (const TargetRegisterClass *) &PPC::GPRCRegClass;
6040 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6041 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6042 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6043 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
6044 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
6045 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
6046 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
6047 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6048 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6049 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6050 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6051 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6052 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6054 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
6055 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
6058 // fallthrough --> loopMBB
6059 BB->addSuccessor(loop1MBB);
6061 // The 4-byte load must be aligned, while a char or short may be
6062 // anywhere in the word. Hence all this nasty bookkeeping code.
6063 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6064 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6065 // xori shift, shift1, 24 [16]
6066 // rlwinm ptr, ptr1, 0, 0, 29
6067 // slw newval2, newval, shift
6068 // slw oldval2, oldval,shift
6069 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6070 // slw mask, mask2, shift
6071 // and newval3, newval2, mask
6072 // and oldval3, oldval2, mask
6074 // lwarx tmpDest, ptr
6075 // and tmp, tmpDest, mask
6076 // cmpw tmp, oldval3
6079 // andc tmp2, tmpDest, mask
6080 // or tmp4, tmp2, newval3
6085 // stwcx. tmpDest, ptr
6087 // srw dest, tmpDest, shift
6088 if (ptrA != ZeroReg) {
6089 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6090 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6091 .addReg(ptrA).addReg(ptrB);
6095 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6096 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6097 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6098 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6100 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6101 .addReg(Ptr1Reg).addImm(0).addImm(61);
6103 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6104 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6105 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6106 .addReg(newval).addReg(ShiftReg);
6107 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6108 .addReg(oldval).addReg(ShiftReg);
6110 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6112 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6113 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6114 .addReg(Mask3Reg).addImm(65535);
6116 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6117 .addReg(Mask2Reg).addReg(ShiftReg);
6118 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6119 .addReg(NewVal2Reg).addReg(MaskReg);
6120 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6121 .addReg(OldVal2Reg).addReg(MaskReg);
6124 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6125 .addReg(ZeroReg).addReg(PtrReg);
6126 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6127 .addReg(TmpDestReg).addReg(MaskReg);
6128 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6129 .addReg(TmpReg).addReg(OldVal3Reg);
6130 BuildMI(BB, dl, TII->get(PPC::BCC))
6131 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6132 BB->addSuccessor(loop2MBB);
6133 BB->addSuccessor(midMBB);
6136 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6137 .addReg(TmpDestReg).addReg(MaskReg);
6138 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6139 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6140 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6141 .addReg(ZeroReg).addReg(PtrReg);
6142 BuildMI(BB, dl, TII->get(PPC::BCC))
6143 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6144 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6145 BB->addSuccessor(loop1MBB);
6146 BB->addSuccessor(exitMBB);
6149 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6150 .addReg(ZeroReg).addReg(PtrReg);
6151 BB->addSuccessor(exitMBB);
6156 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6159 llvm_unreachable("Unexpected instr type to insert");
6162 MI->eraseFromParent(); // The pseudo instruction is gone now.
6166 //===----------------------------------------------------------------------===//
6167 // Target Optimization Hooks
6168 //===----------------------------------------------------------------------===//
6170 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6171 DAGCombinerInfo &DCI) const {
6172 const TargetMachine &TM = getTargetMachine();
6173 SelectionDAG &DAG = DCI.DAG;
6174 DebugLoc dl = N->getDebugLoc();
6175 switch (N->getOpcode()) {
6178 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6179 if (C->isNullValue()) // 0 << V -> 0.
6180 return N->getOperand(0);
6184 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6185 if (C->isNullValue()) // 0 >>u V -> 0.
6186 return N->getOperand(0);
6190 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6191 if (C->isNullValue() || // 0 >>s V -> 0.
6192 C->isAllOnesValue()) // -1 >>s V -> -1.
6193 return N->getOperand(0);
6197 case ISD::SINT_TO_FP:
6198 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6199 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6200 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6201 // We allow the src/dst to be either f32/f64, but the intermediate
6202 // type must be i64.
6203 if (N->getOperand(0).getValueType() == MVT::i64 &&
6204 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6205 SDValue Val = N->getOperand(0).getOperand(0);
6206 if (Val.getValueType() == MVT::f32) {
6207 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6208 DCI.AddToWorklist(Val.getNode());
6211 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6212 DCI.AddToWorklist(Val.getNode());
6213 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6214 DCI.AddToWorklist(Val.getNode());
6215 if (N->getValueType(0) == MVT::f32) {
6216 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6217 DAG.getIntPtrConstant(0));
6218 DCI.AddToWorklist(Val.getNode());
6221 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6222 // If the intermediate type is i32, we can avoid the load/store here
6229 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6230 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6231 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6232 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6233 N->getOperand(1).getValueType() == MVT::i32 &&
6234 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6235 SDValue Val = N->getOperand(1).getOperand(0);
6236 if (Val.getValueType() == MVT::f32) {
6237 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6238 DCI.AddToWorklist(Val.getNode());
6240 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6241 DCI.AddToWorklist(Val.getNode());
6243 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6244 N->getOperand(2), N->getOperand(3));
6245 DCI.AddToWorklist(Val.getNode());
6249 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6250 if (cast<StoreSDNode>(N)->isUnindexed() &&
6251 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6252 N->getOperand(1).getNode()->hasOneUse() &&
6253 (N->getOperand(1).getValueType() == MVT::i32 ||
6254 N->getOperand(1).getValueType() == MVT::i16)) {
6255 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6256 // Do an any-extend to 32-bits if this is a half-word input.
6257 if (BSwapOp.getValueType() == MVT::i16)
6258 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6261 N->getOperand(0), BSwapOp, N->getOperand(2),
6262 DAG.getValueType(N->getOperand(1).getValueType())
6265 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6266 Ops, array_lengthof(Ops),
6267 cast<StoreSDNode>(N)->getMemoryVT(),
6268 cast<StoreSDNode>(N)->getMemOperand());
6272 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6273 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6274 N->getOperand(0).hasOneUse() &&
6275 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6276 SDValue Load = N->getOperand(0);
6277 LoadSDNode *LD = cast<LoadSDNode>(Load);
6278 // Create the byte-swapping load.
6280 LD->getChain(), // Chain
6281 LD->getBasePtr(), // Ptr
6282 DAG.getValueType(N->getValueType(0)) // VT
6285 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6286 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6287 LD->getMemoryVT(), LD->getMemOperand());
6289 // If this is an i16 load, insert the truncate.
6290 SDValue ResVal = BSLoad;
6291 if (N->getValueType(0) == MVT::i16)
6292 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6294 // First, combine the bswap away. This makes the value produced by the
6296 DCI.CombineTo(N, ResVal);
6298 // Next, combine the load away, we give it a bogus result value but a real
6299 // chain result. The result value is dead because the bswap is dead.
6300 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6302 // Return N so it doesn't get rechecked!
6303 return SDValue(N, 0);
6307 case PPCISD::VCMP: {
6308 // If a VCMPo node already exists with exactly the same operands as this
6309 // node, use its result instead of this node (VCMPo computes both a CR6 and
6310 // a normal output).
6312 if (!N->getOperand(0).hasOneUse() &&
6313 !N->getOperand(1).hasOneUse() &&
6314 !N->getOperand(2).hasOneUse()) {
6316 // Scan all of the users of the LHS, looking for VCMPo's that match.
6317 SDNode *VCMPoNode = 0;
6319 SDNode *LHSN = N->getOperand(0).getNode();
6320 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6322 if (UI->getOpcode() == PPCISD::VCMPo &&
6323 UI->getOperand(1) == N->getOperand(1) &&
6324 UI->getOperand(2) == N->getOperand(2) &&
6325 UI->getOperand(0) == N->getOperand(0)) {
6330 // If there is no VCMPo node, or if the flag value has a single use, don't
6332 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6335 // Look at the (necessarily single) use of the flag value. If it has a
6336 // chain, this transformation is more complex. Note that multiple things
6337 // could use the value result, which we should ignore.
6338 SDNode *FlagUser = 0;
6339 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6340 FlagUser == 0; ++UI) {
6341 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6343 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6344 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6351 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6352 // give up for right now.
6353 if (FlagUser->getOpcode() == PPCISD::MFCR)
6354 return SDValue(VCMPoNode, 0);
6359 // If this is a branch on an altivec predicate comparison, lower this so
6360 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6361 // lowering is done pre-legalize, because the legalizer lowers the predicate
6362 // compare down to code that is difficult to reassemble.
6363 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6364 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6368 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6369 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6370 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6371 assert(isDot && "Can't compare against a vector result!");
6373 // If this is a comparison against something other than 0/1, then we know
6374 // that the condition is never/always true.
6375 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6376 if (Val != 0 && Val != 1) {
6377 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6378 return N->getOperand(0);
6379 // Always !=, turn it into an unconditional branch.
6380 return DAG.getNode(ISD::BR, dl, MVT::Other,
6381 N->getOperand(0), N->getOperand(4));
6384 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6386 // Create the PPCISD altivec 'dot' comparison node.
6387 std::vector<EVT> VTs;
6389 LHS.getOperand(2), // LHS of compare
6390 LHS.getOperand(3), // RHS of compare
6391 DAG.getConstant(CompareOpc, MVT::i32)
6393 VTs.push_back(LHS.getOperand(2).getValueType());
6394 VTs.push_back(MVT::Glue);
6395 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6397 // Unpack the result based on how the target uses it.
6398 PPC::Predicate CompOpc;
6399 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6400 default: // Can't happen, don't crash on invalid number though.
6401 case 0: // Branch on the value of the EQ bit of CR6.
6402 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6404 case 1: // Branch on the inverted value of the EQ bit of CR6.
6405 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6407 case 2: // Branch on the value of the LT bit of CR6.
6408 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6410 case 3: // Branch on the inverted value of the LT bit of CR6.
6411 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6415 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6416 DAG.getConstant(CompOpc, MVT::i32),
6417 DAG.getRegister(PPC::CR6, MVT::i32),
6418 N->getOperand(4), CompNode.getValue(1));
6427 //===----------------------------------------------------------------------===//
6428 // Inline Assembly Support
6429 //===----------------------------------------------------------------------===//
6431 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6434 const SelectionDAG &DAG,
6435 unsigned Depth) const {
6436 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6437 switch (Op.getOpcode()) {
6439 case PPCISD::LBRX: {
6440 // lhbrx is known to have the top bits cleared out.
6441 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6442 KnownZero = 0xFFFF0000;
6445 case ISD::INTRINSIC_WO_CHAIN: {
6446 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6448 case Intrinsic::ppc_altivec_vcmpbfp_p:
6449 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6450 case Intrinsic::ppc_altivec_vcmpequb_p:
6451 case Intrinsic::ppc_altivec_vcmpequh_p:
6452 case Intrinsic::ppc_altivec_vcmpequw_p:
6453 case Intrinsic::ppc_altivec_vcmpgefp_p:
6454 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6455 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6456 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6457 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6458 case Intrinsic::ppc_altivec_vcmpgtub_p:
6459 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6460 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6461 KnownZero = ~1U; // All bits but the low one are known to be zero.
6469 /// getConstraintType - Given a constraint, return the type of
6470 /// constraint it is for this target.
6471 PPCTargetLowering::ConstraintType
6472 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6473 if (Constraint.size() == 1) {
6474 switch (Constraint[0]) {
6481 return C_RegisterClass;
6483 // FIXME: While Z does indicate a memory constraint, it specifically
6484 // indicates an r+r address (used in conjunction with the 'y' modifier
6485 // in the replacement string). Currently, we're forcing the base
6486 // register to be r0 in the asm printer (which is interpreted as zero)
6487 // and forming the complete address in the second register. This is
6492 return TargetLowering::getConstraintType(Constraint);
6495 /// Examine constraint type and operand type and determine a weight value.
6496 /// This object must already have been set up with the operand type
6497 /// and the current alternative constraint selected.
6498 TargetLowering::ConstraintWeight
6499 PPCTargetLowering::getSingleConstraintMatchWeight(
6500 AsmOperandInfo &info, const char *constraint) const {
6501 ConstraintWeight weight = CW_Invalid;
6502 Value *CallOperandVal = info.CallOperandVal;
6503 // If we don't have a value, we can't do a match,
6504 // but allow it at the lowest weight.
6505 if (CallOperandVal == NULL)
6507 Type *type = CallOperandVal->getType();
6508 // Look at the constraint type.
6509 switch (*constraint) {
6511 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6514 if (type->isIntegerTy())
6515 weight = CW_Register;
6518 if (type->isFloatTy())
6519 weight = CW_Register;
6522 if (type->isDoubleTy())
6523 weight = CW_Register;
6526 if (type->isVectorTy())
6527 weight = CW_Register;
6530 weight = CW_Register;
6539 std::pair<unsigned, const TargetRegisterClass*>
6540 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6542 if (Constraint.size() == 1) {
6543 // GCC RS6000 Constraint Letters
6544 switch (Constraint[0]) {
6547 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6548 return std::make_pair(0U, &PPC::G8RCRegClass);
6549 return std::make_pair(0U, &PPC::GPRCRegClass);
6551 if (VT == MVT::f32 || VT == MVT::i32)
6552 return std::make_pair(0U, &PPC::F4RCRegClass);
6553 if (VT == MVT::f64 || VT == MVT::i64)
6554 return std::make_pair(0U, &PPC::F8RCRegClass);
6557 return std::make_pair(0U, &PPC::VRRCRegClass);
6559 return std::make_pair(0U, &PPC::CRRCRegClass);
6563 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6567 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6568 /// vector. If it is invalid, don't add anything to Ops.
6569 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6570 std::string &Constraint,
6571 std::vector<SDValue>&Ops,
6572 SelectionDAG &DAG) const {
6573 SDValue Result(0,0);
6575 // Only support length 1 constraints.
6576 if (Constraint.length() > 1) return;
6578 char Letter = Constraint[0];
6589 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6590 if (!CST) return; // Must be an immediate to match.
6591 unsigned Value = CST->getZExtValue();
6593 default: llvm_unreachable("Unknown constraint letter!");
6594 case 'I': // "I" is a signed 16-bit constant.
6595 if ((short)Value == (int)Value)
6596 Result = DAG.getTargetConstant(Value, Op.getValueType());
6598 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6599 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6600 if ((short)Value == 0)
6601 Result = DAG.getTargetConstant(Value, Op.getValueType());
6603 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6604 if ((Value >> 16) == 0)
6605 Result = DAG.getTargetConstant(Value, Op.getValueType());
6607 case 'M': // "M" is a constant that is greater than 31.
6609 Result = DAG.getTargetConstant(Value, Op.getValueType());
6611 case 'N': // "N" is a positive constant that is an exact power of two.
6612 if ((int)Value > 0 && isPowerOf2_32(Value))
6613 Result = DAG.getTargetConstant(Value, Op.getValueType());
6615 case 'O': // "O" is the constant zero.
6617 Result = DAG.getTargetConstant(Value, Op.getValueType());
6619 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6620 if ((short)-Value == (int)-Value)
6621 Result = DAG.getTargetConstant(Value, Op.getValueType());
6628 if (Result.getNode()) {
6629 Ops.push_back(Result);
6633 // Handle standard constraint letters.
6634 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6637 // isLegalAddressingMode - Return true if the addressing mode represented
6638 // by AM is legal for this target, for a load/store of the specified type.
6639 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6641 // FIXME: PPC does not allow r+i addressing modes for vectors!
6643 // PPC allows a sign-extended 16-bit immediate field.
6644 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6647 // No global is ever allowed as a base.
6651 // PPC only support r+r,
6653 case 0: // "r+i" or just "i", depending on HasBaseReg.
6656 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6658 // Otherwise we have r+r or r+i.
6661 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6663 // Allow 2*r as r+r.
6666 // No other scales are supported.
6673 /// isLegalAddressImmediate - Return true if the integer value can be used
6674 /// as the offset of the target addressing mode for load / store of the
6676 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6677 // PPC allows a sign-extended 16-bit immediate field.
6678 return (V > -(1 << 16) && V < (1 << 16)-1);
6681 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6685 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6686 SelectionDAG &DAG) const {
6687 MachineFunction &MF = DAG.getMachineFunction();
6688 MachineFrameInfo *MFI = MF.getFrameInfo();
6689 MFI->setReturnAddressIsTaken(true);
6691 DebugLoc dl = Op.getDebugLoc();
6692 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6694 // Make sure the function does not optimize away the store of the RA to
6696 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6697 FuncInfo->setLRStoreRequired();
6698 bool isPPC64 = PPCSubTarget.isPPC64();
6699 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6702 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6705 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6706 isPPC64? MVT::i64 : MVT::i32);
6707 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6708 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6710 MachinePointerInfo(), false, false, false, 0);
6713 // Just load the return address off the stack.
6714 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6715 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6716 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6719 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6720 SelectionDAG &DAG) const {
6721 DebugLoc dl = Op.getDebugLoc();
6722 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6724 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6725 bool isPPC64 = PtrVT == MVT::i64;
6727 MachineFunction &MF = DAG.getMachineFunction();
6728 MachineFrameInfo *MFI = MF.getFrameInfo();
6729 MFI->setFrameAddressIsTaken(true);
6730 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6731 MFI->hasVarSizedObjects()) &&
6732 MFI->getStackSize() &&
6733 !MF.getFunction()->getFnAttributes().
6734 hasAttribute(Attributes::Naked);
6735 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6736 (is31 ? PPC::R31 : PPC::R1);
6737 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6740 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6741 FrameAddr, MachinePointerInfo(), false, false,
6747 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6748 // The PowerPC target isn't yet aware of offsets.
6752 /// getOptimalMemOpType - Returns the target specific optimal type for load
6753 /// and store operations as a result of memset, memcpy, and memmove
6754 /// lowering. If DstAlign is zero that means it's safe to destination
6755 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6756 /// means there isn't a need to check it against alignment requirement,
6757 /// probably because the source does not need to be loaded. If
6758 /// 'IsZeroVal' is true, that means it's safe to return a
6759 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6760 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6761 /// constant so it does not need to be loaded.
6762 /// It returns EVT::Other if the type should be determined using generic
6763 /// target-independent logic.
6764 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6765 unsigned DstAlign, unsigned SrcAlign,
6768 MachineFunction &MF) const {
6769 if (this->PPCSubTarget.isPPC64()) {
6776 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6777 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6778 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6779 /// is expanded to mul + add.
6780 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6784 switch (VT.getSimpleVT().SimpleTy) {
6796 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6798 return TargetLowering::getSchedulingPreference(N);