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 "MCTargetDesc/PPCPredicates.h"
16 #include "PPCMachineFunctionInfo.h"
17 #include "PPCPerfectShuffle.h"
18 #include "PPCTargetMachine.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
28 #include "llvm/Constants.h"
29 #include "llvm/DerivedTypes.h"
30 #include "llvm/Function.h"
31 #include "llvm/Intrinsics.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::FSQRT, VT, Expand);
351 setOperationAction(ISD::FLOG, VT, Expand);
352 setOperationAction(ISD::FLOG10, VT, Expand);
353 setOperationAction(ISD::FLOG2, VT, Expand);
354 setOperationAction(ISD::FEXP, VT, Expand);
355 setOperationAction(ISD::FEXP2, VT, Expand);
356 setOperationAction(ISD::FSIN, VT, Expand);
357 setOperationAction(ISD::FCOS, VT, Expand);
358 setOperationAction(ISD::FABS, VT, Expand);
359 setOperationAction(ISD::FPOWI, VT, Expand);
360 setOperationAction(ISD::FFLOOR, VT, Expand);
361 setOperationAction(ISD::FCEIL, VT, Expand);
362 setOperationAction(ISD::FTRUNC, VT, Expand);
363 setOperationAction(ISD::FRINT, VT, Expand);
364 setOperationAction(ISD::FNEARBYINT, VT, Expand);
365 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
366 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
367 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
368 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
369 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
370 setOperationAction(ISD::UDIVREM, VT, Expand);
371 setOperationAction(ISD::SDIVREM, VT, Expand);
372 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
373 setOperationAction(ISD::FPOW, VT, Expand);
374 setOperationAction(ISD::CTPOP, VT, Expand);
375 setOperationAction(ISD::CTLZ, VT, Expand);
376 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
377 setOperationAction(ISD::CTTZ, VT, Expand);
378 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
379 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
381 for (unsigned j = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
382 j <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++j) {
383 MVT::SimpleValueType InnerVT = (MVT::SimpleValueType)j;
384 setTruncStoreAction(VT, InnerVT, Expand);
386 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
387 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
388 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
391 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
392 // with merges, splats, etc.
393 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
395 setOperationAction(ISD::AND , MVT::v4i32, Legal);
396 setOperationAction(ISD::OR , MVT::v4i32, Legal);
397 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
398 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
399 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
400 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
401 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
402 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
403 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
404 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
405 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
406 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
407 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
408 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
410 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
411 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
412 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
413 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
415 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
416 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
417 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
418 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
419 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
421 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
422 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
424 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
425 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
426 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
427 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
429 // Altivec does not contain unordered floating-point compare instructions
430 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
431 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
432 setCondCodeAction(ISD::SETUGT, MVT::v4f32, Expand);
433 setCondCodeAction(ISD::SETUGE, MVT::v4f32, Expand);
434 setCondCodeAction(ISD::SETULT, MVT::v4f32, Expand);
435 setCondCodeAction(ISD::SETULE, MVT::v4f32, Expand);
438 if (Subtarget->has64BitSupport()) {
439 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
440 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
443 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
444 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
446 setBooleanContents(ZeroOrOneBooleanContent);
447 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
450 setStackPointerRegisterToSaveRestore(PPC::X1);
451 setExceptionPointerRegister(PPC::X3);
452 setExceptionSelectorRegister(PPC::X4);
454 setStackPointerRegisterToSaveRestore(PPC::R1);
455 setExceptionPointerRegister(PPC::R3);
456 setExceptionSelectorRegister(PPC::R4);
459 // We have target-specific dag combine patterns for the following nodes:
460 setTargetDAGCombine(ISD::SINT_TO_FP);
461 setTargetDAGCombine(ISD::STORE);
462 setTargetDAGCombine(ISD::BR_CC);
463 setTargetDAGCombine(ISD::BSWAP);
465 // Darwin long double math library functions have $LDBL128 appended.
466 if (Subtarget->isDarwin()) {
467 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
468 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
469 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
470 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
471 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
472 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
473 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
474 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
475 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
476 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
479 setMinFunctionAlignment(2);
480 if (PPCSubTarget.isDarwin())
481 setPrefFunctionAlignment(4);
483 if (isPPC64 && Subtarget->isJITCodeModel())
484 // Temporary workaround for the inability of PPC64 JIT to handle jump
486 setSupportJumpTables(false);
488 setInsertFencesForAtomic(true);
490 setSchedulingPreference(Sched::Hybrid);
492 computeRegisterProperties();
494 // The Freescale cores does better with aggressive inlining of memcpy and
495 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
496 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
497 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
498 maxStoresPerMemset = 32;
499 maxStoresPerMemsetOptSize = 16;
500 maxStoresPerMemcpy = 32;
501 maxStoresPerMemcpyOptSize = 8;
502 maxStoresPerMemmove = 32;
503 maxStoresPerMemmoveOptSize = 8;
505 setPrefFunctionAlignment(4);
506 benefitFromCodePlacementOpt = true;
510 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
511 /// function arguments in the caller parameter area.
512 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
513 const TargetMachine &TM = getTargetMachine();
514 // Darwin passes everything on 4 byte boundary.
515 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
518 // 16byte and wider vectors are passed on 16byte boundary.
519 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
520 if (VTy->getBitWidth() >= 128)
523 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
524 if (PPCSubTarget.isPPC64())
530 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
533 case PPCISD::FSEL: return "PPCISD::FSEL";
534 case PPCISD::FCFID: return "PPCISD::FCFID";
535 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
536 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
537 case PPCISD::STFIWX: return "PPCISD::STFIWX";
538 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
539 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
540 case PPCISD::VPERM: return "PPCISD::VPERM";
541 case PPCISD::Hi: return "PPCISD::Hi";
542 case PPCISD::Lo: return "PPCISD::Lo";
543 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
544 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
545 case PPCISD::LOAD: return "PPCISD::LOAD";
546 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
547 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
548 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
549 case PPCISD::SRL: return "PPCISD::SRL";
550 case PPCISD::SRA: return "PPCISD::SRA";
551 case PPCISD::SHL: return "PPCISD::SHL";
552 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
553 case PPCISD::STD_32: return "PPCISD::STD_32";
554 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
555 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
556 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
557 case PPCISD::NOP: return "PPCISD::NOP";
558 case PPCISD::MTCTR: return "PPCISD::MTCTR";
559 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
560 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
561 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
562 case PPCISD::MFCR: return "PPCISD::MFCR";
563 case PPCISD::VCMP: return "PPCISD::VCMP";
564 case PPCISD::VCMPo: return "PPCISD::VCMPo";
565 case PPCISD::LBRX: return "PPCISD::LBRX";
566 case PPCISD::STBRX: return "PPCISD::STBRX";
567 case PPCISD::LARX: return "PPCISD::LARX";
568 case PPCISD::STCX: return "PPCISD::STCX";
569 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
570 case PPCISD::MFFS: return "PPCISD::MFFS";
571 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
572 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
573 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
574 case PPCISD::MTFSF: return "PPCISD::MTFSF";
575 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
576 case PPCISD::CR6SET: return "PPCISD::CR6SET";
577 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
578 case PPCISD::ADDIS_TOC_HA: return "PPCISD::ADDIS_TOC_HA";
579 case PPCISD::LD_TOC_L: return "PPCISD::LD_TOC_L";
580 case PPCISD::ADDI_TOC_L: return "PPCISD::ADDI_TOC_L";
584 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
587 return VT.changeVectorElementTypeToInteger();
590 //===----------------------------------------------------------------------===//
591 // Node matching predicates, for use by the tblgen matching code.
592 //===----------------------------------------------------------------------===//
594 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
595 static bool isFloatingPointZero(SDValue Op) {
596 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
597 return CFP->getValueAPF().isZero();
598 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
599 // Maybe this has already been legalized into the constant pool?
600 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
601 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
602 return CFP->getValueAPF().isZero();
607 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
608 /// true if Op is undef or if it matches the specified value.
609 static bool isConstantOrUndef(int Op, int Val) {
610 return Op < 0 || Op == Val;
613 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
614 /// VPKUHUM instruction.
615 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
617 for (unsigned i = 0; i != 16; ++i)
618 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
621 for (unsigned i = 0; i != 8; ++i)
622 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
623 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
629 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
630 /// VPKUWUM instruction.
631 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
633 for (unsigned i = 0; i != 16; i += 2)
634 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
635 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
638 for (unsigned i = 0; i != 8; i += 2)
639 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
640 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
641 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
642 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
648 /// isVMerge - Common function, used to match vmrg* shuffles.
650 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
651 unsigned LHSStart, unsigned RHSStart) {
652 assert(N->getValueType(0) == MVT::v16i8 &&
653 "PPC only supports shuffles by bytes!");
654 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
655 "Unsupported merge size!");
657 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
658 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
659 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
660 LHSStart+j+i*UnitSize) ||
661 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
662 RHSStart+j+i*UnitSize))
668 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
669 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
670 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
673 return isVMerge(N, UnitSize, 8, 24);
674 return isVMerge(N, UnitSize, 8, 8);
677 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
678 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
679 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
682 return isVMerge(N, UnitSize, 0, 16);
683 return isVMerge(N, UnitSize, 0, 0);
687 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
688 /// amount, otherwise return -1.
689 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
690 assert(N->getValueType(0) == MVT::v16i8 &&
691 "PPC only supports shuffles by bytes!");
693 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
695 // Find the first non-undef value in the shuffle mask.
697 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
700 if (i == 16) return -1; // all undef.
702 // Otherwise, check to see if the rest of the elements are consecutively
703 // numbered from this value.
704 unsigned ShiftAmt = SVOp->getMaskElt(i);
705 if (ShiftAmt < i) return -1;
709 // Check the rest of the elements to see if they are consecutive.
710 for (++i; i != 16; ++i)
711 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
714 // Check the rest of the elements to see if they are consecutive.
715 for (++i; i != 16; ++i)
716 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
722 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
723 /// specifies a splat of a single element that is suitable for input to
724 /// VSPLTB/VSPLTH/VSPLTW.
725 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
726 assert(N->getValueType(0) == MVT::v16i8 &&
727 (EltSize == 1 || EltSize == 2 || EltSize == 4));
729 // This is a splat operation if each element of the permute is the same, and
730 // if the value doesn't reference the second vector.
731 unsigned ElementBase = N->getMaskElt(0);
733 // FIXME: Handle UNDEF elements too!
734 if (ElementBase >= 16)
737 // Check that the indices are consecutive, in the case of a multi-byte element
738 // splatted with a v16i8 mask.
739 for (unsigned i = 1; i != EltSize; ++i)
740 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
743 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
744 if (N->getMaskElt(i) < 0) continue;
745 for (unsigned j = 0; j != EltSize; ++j)
746 if (N->getMaskElt(i+j) != N->getMaskElt(j))
752 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
754 bool PPC::isAllNegativeZeroVector(SDNode *N) {
755 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
757 APInt APVal, APUndef;
761 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
762 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
763 return CFP->getValueAPF().isNegZero();
768 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
769 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
770 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
771 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
772 assert(isSplatShuffleMask(SVOp, EltSize));
773 return SVOp->getMaskElt(0) / EltSize;
776 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
777 /// by using a vspltis[bhw] instruction of the specified element size, return
778 /// the constant being splatted. The ByteSize field indicates the number of
779 /// bytes of each element [124] -> [bhw].
780 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
783 // If ByteSize of the splat is bigger than the element size of the
784 // build_vector, then we have a case where we are checking for a splat where
785 // multiple elements of the buildvector are folded together into a single
786 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
787 unsigned EltSize = 16/N->getNumOperands();
788 if (EltSize < ByteSize) {
789 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
790 SDValue UniquedVals[4];
791 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
793 // See if all of the elements in the buildvector agree across.
794 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
795 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
796 // If the element isn't a constant, bail fully out.
797 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
800 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
801 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
802 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
803 return SDValue(); // no match.
806 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
807 // either constant or undef values that are identical for each chunk. See
808 // if these chunks can form into a larger vspltis*.
810 // Check to see if all of the leading entries are either 0 or -1. If
811 // neither, then this won't fit into the immediate field.
812 bool LeadingZero = true;
813 bool LeadingOnes = true;
814 for (unsigned i = 0; i != Multiple-1; ++i) {
815 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
817 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
818 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
820 // Finally, check the least significant entry.
822 if (UniquedVals[Multiple-1].getNode() == 0)
823 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
824 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
826 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
829 if (UniquedVals[Multiple-1].getNode() == 0)
830 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
831 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
832 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
833 return DAG.getTargetConstant(Val, MVT::i32);
839 // Check to see if this buildvec has a single non-undef value in its elements.
840 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
841 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
842 if (OpVal.getNode() == 0)
843 OpVal = N->getOperand(i);
844 else if (OpVal != N->getOperand(i))
848 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
850 unsigned ValSizeInBytes = EltSize;
852 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
853 Value = CN->getZExtValue();
854 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
855 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
856 Value = FloatToBits(CN->getValueAPF().convertToFloat());
859 // If the splat value is larger than the element value, then we can never do
860 // this splat. The only case that we could fit the replicated bits into our
861 // immediate field for would be zero, and we prefer to use vxor for it.
862 if (ValSizeInBytes < ByteSize) return SDValue();
864 // If the element value is larger than the splat value, cut it in half and
865 // check to see if the two halves are equal. Continue doing this until we
866 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
867 while (ValSizeInBytes > ByteSize) {
868 ValSizeInBytes >>= 1;
870 // If the top half equals the bottom half, we're still ok.
871 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
872 (Value & ((1 << (8*ValSizeInBytes))-1)))
876 // Properly sign extend the value.
877 int MaskVal = SignExtend32(Value, ByteSize * 8);
879 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
880 if (MaskVal == 0) return SDValue();
882 // Finally, if this value fits in a 5 bit sext field, return it
883 if (SignExtend32<5>(MaskVal) == MaskVal)
884 return DAG.getTargetConstant(MaskVal, MVT::i32);
888 //===----------------------------------------------------------------------===//
889 // Addressing Mode Selection
890 //===----------------------------------------------------------------------===//
892 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
893 /// or 64-bit immediate, and if the value can be accurately represented as a
894 /// sign extension from a 16-bit value. If so, this returns true and the
896 static bool isIntS16Immediate(SDNode *N, short &Imm) {
897 if (N->getOpcode() != ISD::Constant)
900 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
901 if (N->getValueType(0) == MVT::i32)
902 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
904 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
906 static bool isIntS16Immediate(SDValue Op, short &Imm) {
907 return isIntS16Immediate(Op.getNode(), Imm);
911 /// SelectAddressRegReg - Given the specified addressed, check to see if it
912 /// can be represented as an indexed [r+r] operation. Returns false if it
913 /// can be more efficiently represented with [r+imm].
914 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
916 SelectionDAG &DAG) const {
918 if (N.getOpcode() == ISD::ADD) {
919 if (isIntS16Immediate(N.getOperand(1), imm))
921 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
924 Base = N.getOperand(0);
925 Index = N.getOperand(1);
927 } else if (N.getOpcode() == ISD::OR) {
928 if (isIntS16Immediate(N.getOperand(1), imm))
929 return false; // r+i can fold it if we can.
931 // If this is an or of disjoint bitfields, we can codegen this as an add
932 // (for better address arithmetic) if the LHS and RHS of the OR are provably
934 APInt LHSKnownZero, LHSKnownOne;
935 APInt RHSKnownZero, RHSKnownOne;
936 DAG.ComputeMaskedBits(N.getOperand(0),
937 LHSKnownZero, LHSKnownOne);
939 if (LHSKnownZero.getBoolValue()) {
940 DAG.ComputeMaskedBits(N.getOperand(1),
941 RHSKnownZero, RHSKnownOne);
942 // If all of the bits are known zero on the LHS or RHS, the add won't
944 if (~(LHSKnownZero | RHSKnownZero) == 0) {
945 Base = N.getOperand(0);
946 Index = N.getOperand(1);
955 /// Returns true if the address N can be represented by a base register plus
956 /// a signed 16-bit displacement [r+imm], and if it is not better
957 /// represented as reg+reg.
958 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
960 SelectionDAG &DAG) const {
961 // FIXME dl should come from parent load or store, not from address
962 DebugLoc dl = N.getDebugLoc();
963 // If this can be more profitably realized as r+r, fail.
964 if (SelectAddressRegReg(N, Disp, Base, DAG))
967 if (N.getOpcode() == ISD::ADD) {
969 if (isIntS16Immediate(N.getOperand(1), imm)) {
970 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
971 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
972 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
974 Base = N.getOperand(0);
976 return true; // [r+i]
977 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
978 // Match LOAD (ADD (X, Lo(G))).
979 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
980 && "Cannot handle constant offsets yet!");
981 Disp = N.getOperand(1).getOperand(0); // The global address.
982 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
983 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
984 Disp.getOpcode() == ISD::TargetConstantPool ||
985 Disp.getOpcode() == ISD::TargetJumpTable);
986 Base = N.getOperand(0);
987 return true; // [&g+r]
989 } else if (N.getOpcode() == ISD::OR) {
991 if (isIntS16Immediate(N.getOperand(1), imm)) {
992 // If this is an or of disjoint bitfields, we can codegen this as an add
993 // (for better address arithmetic) if the LHS and RHS of the OR are
994 // provably disjoint.
995 APInt LHSKnownZero, LHSKnownOne;
996 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
998 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
999 // If all of the bits are known zero on the LHS or RHS, the add won't
1001 Base = N.getOperand(0);
1002 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
1006 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1007 // Loading from a constant address.
1009 // If this address fits entirely in a 16-bit sext immediate field, codegen
1012 if (isIntS16Immediate(CN, Imm)) {
1013 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
1014 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1015 CN->getValueType(0));
1019 // Handle 32-bit sext immediates with LIS + addr mode.
1020 if (CN->getValueType(0) == MVT::i32 ||
1021 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1022 int Addr = (int)CN->getZExtValue();
1024 // Otherwise, break this down into an LIS + disp.
1025 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1027 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1028 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1029 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1034 Disp = DAG.getTargetConstant(0, getPointerTy());
1035 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1036 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1039 return true; // [r+0]
1042 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1043 /// represented as an indexed [r+r] operation.
1044 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1046 SelectionDAG &DAG) const {
1047 // Check to see if we can easily represent this as an [r+r] address. This
1048 // will fail if it thinks that the address is more profitably represented as
1049 // reg+imm, e.g. where imm = 0.
1050 if (SelectAddressRegReg(N, Base, Index, DAG))
1053 // If the operand is an addition, always emit this as [r+r], since this is
1054 // better (for code size, and execution, as the memop does the add for free)
1055 // than emitting an explicit add.
1056 if (N.getOpcode() == ISD::ADD) {
1057 Base = N.getOperand(0);
1058 Index = N.getOperand(1);
1062 // Otherwise, do it the hard way, using R0 as the base register.
1063 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1069 /// SelectAddressRegImmShift - Returns true if the address N can be
1070 /// represented by a base register plus a signed 14-bit displacement
1071 /// [r+imm*4]. Suitable for use by STD and friends.
1072 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1074 SelectionDAG &DAG) const {
1075 // FIXME dl should come from the parent load or store, not the address
1076 DebugLoc dl = N.getDebugLoc();
1077 // If this can be more profitably realized as r+r, fail.
1078 if (SelectAddressRegReg(N, Disp, Base, DAG))
1081 if (N.getOpcode() == ISD::ADD) {
1083 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1084 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1085 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1086 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1088 Base = N.getOperand(0);
1090 return true; // [r+i]
1091 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1092 // Match LOAD (ADD (X, Lo(G))).
1093 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1094 && "Cannot handle constant offsets yet!");
1095 Disp = N.getOperand(1).getOperand(0); // The global address.
1096 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1097 Disp.getOpcode() == ISD::TargetConstantPool ||
1098 Disp.getOpcode() == ISD::TargetJumpTable);
1099 Base = N.getOperand(0);
1100 return true; // [&g+r]
1102 } else if (N.getOpcode() == ISD::OR) {
1104 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1105 // If this is an or of disjoint bitfields, we can codegen this as an add
1106 // (for better address arithmetic) if the LHS and RHS of the OR are
1107 // provably disjoint.
1108 APInt LHSKnownZero, LHSKnownOne;
1109 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1110 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1111 // If all of the bits are known zero on the LHS or RHS, the add won't
1113 Base = N.getOperand(0);
1114 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1118 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1119 // Loading from a constant address. Verify low two bits are clear.
1120 if ((CN->getZExtValue() & 3) == 0) {
1121 // If this address fits entirely in a 14-bit sext immediate field, codegen
1124 if (isIntS16Immediate(CN, Imm)) {
1125 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1126 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1127 CN->getValueType(0));
1131 // Fold the low-part of 32-bit absolute addresses into addr mode.
1132 if (CN->getValueType(0) == MVT::i32 ||
1133 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1134 int Addr = (int)CN->getZExtValue();
1136 // Otherwise, break this down into an LIS + disp.
1137 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1138 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1139 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1140 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1146 Disp = DAG.getTargetConstant(0, getPointerTy());
1147 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1148 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1151 return true; // [r+0]
1155 /// getPreIndexedAddressParts - returns true by value, base pointer and
1156 /// offset pointer and addressing mode by reference if the node's address
1157 /// can be legally represented as pre-indexed load / store address.
1158 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1160 ISD::MemIndexedMode &AM,
1161 SelectionDAG &DAG) const {
1162 if (DisablePPCPreinc) return false;
1166 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1167 Ptr = LD->getBasePtr();
1168 VT = LD->getMemoryVT();
1170 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1171 Ptr = ST->getBasePtr();
1172 VT = ST->getMemoryVT();
1176 // PowerPC doesn't have preinc load/store instructions for vectors.
1180 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1185 // LDU/STU use reg+imm*4, others use reg+imm.
1186 if (VT != MVT::i64) {
1188 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1192 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1196 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1197 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1198 // sext i32 to i64 when addr mode is r+i.
1199 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1200 LD->getExtensionType() == ISD::SEXTLOAD &&
1201 isa<ConstantSDNode>(Offset))
1209 //===----------------------------------------------------------------------===//
1210 // LowerOperation implementation
1211 //===----------------------------------------------------------------------===//
1213 /// GetLabelAccessInfo - Return true if we should reference labels using a
1214 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1215 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1216 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1217 HiOpFlags = PPCII::MO_HA16;
1218 LoOpFlags = PPCII::MO_LO16;
1220 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1221 // non-darwin platform. We don't support PIC on other platforms yet.
1222 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1223 TM.getSubtarget<PPCSubtarget>().isDarwin();
1225 HiOpFlags |= PPCII::MO_PIC_FLAG;
1226 LoOpFlags |= PPCII::MO_PIC_FLAG;
1229 // If this is a reference to a global value that requires a non-lazy-ptr, make
1230 // sure that instruction lowering adds it.
1231 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1232 HiOpFlags |= PPCII::MO_NLP_FLAG;
1233 LoOpFlags |= PPCII::MO_NLP_FLAG;
1235 if (GV->hasHiddenVisibility()) {
1236 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1237 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1244 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1245 SelectionDAG &DAG) {
1246 EVT PtrVT = HiPart.getValueType();
1247 SDValue Zero = DAG.getConstant(0, PtrVT);
1248 DebugLoc DL = HiPart.getDebugLoc();
1250 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1251 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1253 // With PIC, the first instruction is actually "GR+hi(&G)".
1255 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1256 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1258 // Generate non-pic code that has direct accesses to the constant pool.
1259 // The address of the global is just (hi(&g)+lo(&g)).
1260 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1263 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1264 SelectionDAG &DAG) const {
1265 EVT PtrVT = Op.getValueType();
1266 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1267 const Constant *C = CP->getConstVal();
1269 // 64-bit SVR4 ABI code is always position-independent.
1270 // The actual address of the GlobalValue is stored in the TOC.
1271 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1272 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1273 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1274 DAG.getRegister(PPC::X2, MVT::i64));
1277 unsigned MOHiFlag, MOLoFlag;
1278 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1280 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1282 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1283 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1286 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1287 EVT PtrVT = Op.getValueType();
1288 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1290 // 64-bit SVR4 ABI code is always position-independent.
1291 // The actual address of the GlobalValue is stored in the TOC.
1292 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1293 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1294 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1295 DAG.getRegister(PPC::X2, MVT::i64));
1298 unsigned MOHiFlag, MOLoFlag;
1299 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1300 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1301 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1302 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1305 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1306 SelectionDAG &DAG) const {
1307 EVT PtrVT = Op.getValueType();
1309 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1311 unsigned MOHiFlag, MOLoFlag;
1312 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1313 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1314 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1315 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1318 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1319 SelectionDAG &DAG) const {
1321 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1322 DebugLoc dl = GA->getDebugLoc();
1323 const GlobalValue *GV = GA->getGlobal();
1324 EVT PtrVT = getPointerTy();
1325 bool is64bit = PPCSubTarget.isPPC64();
1327 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1329 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1330 PPCII::MO_TPREL16_HA);
1331 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1332 PPCII::MO_TPREL16_LO);
1334 if (model != TLSModel::LocalExec)
1335 llvm_unreachable("only local-exec TLS mode supported");
1336 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1337 is64bit ? MVT::i64 : MVT::i32);
1338 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1339 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1342 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1343 SelectionDAG &DAG) const {
1344 EVT PtrVT = Op.getValueType();
1345 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1346 DebugLoc DL = GSDN->getDebugLoc();
1347 const GlobalValue *GV = GSDN->getGlobal();
1349 // 64-bit SVR4 ABI code is always position-independent.
1350 // The actual address of the GlobalValue is stored in the TOC.
1351 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1352 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1353 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1354 DAG.getRegister(PPC::X2, MVT::i64));
1357 unsigned MOHiFlag, MOLoFlag;
1358 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1361 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1363 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1365 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1367 // If the global reference is actually to a non-lazy-pointer, we have to do an
1368 // extra load to get the address of the global.
1369 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1370 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1371 false, false, false, 0);
1375 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1376 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1377 DebugLoc dl = Op.getDebugLoc();
1379 // If we're comparing for equality to zero, expose the fact that this is
1380 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1381 // fold the new nodes.
1382 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1383 if (C->isNullValue() && CC == ISD::SETEQ) {
1384 EVT VT = Op.getOperand(0).getValueType();
1385 SDValue Zext = Op.getOperand(0);
1386 if (VT.bitsLT(MVT::i32)) {
1388 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1390 unsigned Log2b = Log2_32(VT.getSizeInBits());
1391 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1392 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1393 DAG.getConstant(Log2b, MVT::i32));
1394 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1396 // Leave comparisons against 0 and -1 alone for now, since they're usually
1397 // optimized. FIXME: revisit this when we can custom lower all setcc
1399 if (C->isAllOnesValue() || C->isNullValue())
1403 // If we have an integer seteq/setne, turn it into a compare against zero
1404 // by xor'ing the rhs with the lhs, which is faster than setting a
1405 // condition register, reading it back out, and masking the correct bit. The
1406 // normal approach here uses sub to do this instead of xor. Using xor exposes
1407 // the result to other bit-twiddling opportunities.
1408 EVT LHSVT = Op.getOperand(0).getValueType();
1409 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1410 EVT VT = Op.getValueType();
1411 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1413 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1418 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1419 const PPCSubtarget &Subtarget) const {
1420 SDNode *Node = Op.getNode();
1421 EVT VT = Node->getValueType(0);
1422 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1423 SDValue InChain = Node->getOperand(0);
1424 SDValue VAListPtr = Node->getOperand(1);
1425 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1426 DebugLoc dl = Node->getDebugLoc();
1428 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1431 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1432 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1434 InChain = GprIndex.getValue(1);
1436 if (VT == MVT::i64) {
1437 // Check if GprIndex is even
1438 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1439 DAG.getConstant(1, MVT::i32));
1440 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1441 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1442 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1443 DAG.getConstant(1, MVT::i32));
1444 // Align GprIndex to be even if it isn't
1445 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1449 // fpr index is 1 byte after gpr
1450 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1451 DAG.getConstant(1, MVT::i32));
1454 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1455 FprPtr, MachinePointerInfo(SV), MVT::i8,
1457 InChain = FprIndex.getValue(1);
1459 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1460 DAG.getConstant(8, MVT::i32));
1462 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1463 DAG.getConstant(4, MVT::i32));
1466 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1467 MachinePointerInfo(), false, false,
1469 InChain = OverflowArea.getValue(1);
1471 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1472 MachinePointerInfo(), false, false,
1474 InChain = RegSaveArea.getValue(1);
1476 // select overflow_area if index > 8
1477 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1478 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1480 // adjustment constant gpr_index * 4/8
1481 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1482 VT.isInteger() ? GprIndex : FprIndex,
1483 DAG.getConstant(VT.isInteger() ? 4 : 8,
1486 // OurReg = RegSaveArea + RegConstant
1487 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1490 // Floating types are 32 bytes into RegSaveArea
1491 if (VT.isFloatingPoint())
1492 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1493 DAG.getConstant(32, MVT::i32));
1495 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1496 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1497 VT.isInteger() ? GprIndex : FprIndex,
1498 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1501 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1502 VT.isInteger() ? VAListPtr : FprPtr,
1503 MachinePointerInfo(SV),
1504 MVT::i8, false, false, 0);
1506 // determine if we should load from reg_save_area or overflow_area
1507 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1509 // increase overflow_area by 4/8 if gpr/fpr > 8
1510 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1511 DAG.getConstant(VT.isInteger() ? 4 : 8,
1514 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1517 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1519 MachinePointerInfo(),
1520 MVT::i32, false, false, 0);
1522 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1523 false, false, false, 0);
1526 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1527 SelectionDAG &DAG) const {
1528 return Op.getOperand(0);
1531 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1532 SelectionDAG &DAG) const {
1533 SDValue Chain = Op.getOperand(0);
1534 SDValue Trmp = Op.getOperand(1); // trampoline
1535 SDValue FPtr = Op.getOperand(2); // nested function
1536 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1537 DebugLoc dl = Op.getDebugLoc();
1539 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1540 bool isPPC64 = (PtrVT == MVT::i64);
1542 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1545 TargetLowering::ArgListTy Args;
1546 TargetLowering::ArgListEntry Entry;
1548 Entry.Ty = IntPtrTy;
1549 Entry.Node = Trmp; Args.push_back(Entry);
1551 // TrampSize == (isPPC64 ? 48 : 40);
1552 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1553 isPPC64 ? MVT::i64 : MVT::i32);
1554 Args.push_back(Entry);
1556 Entry.Node = FPtr; Args.push_back(Entry);
1557 Entry.Node = Nest; Args.push_back(Entry);
1559 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1560 TargetLowering::CallLoweringInfo CLI(Chain,
1561 Type::getVoidTy(*DAG.getContext()),
1562 false, false, false, false, 0,
1564 /*isTailCall=*/false,
1565 /*doesNotRet=*/false,
1566 /*isReturnValueUsed=*/true,
1567 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1569 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1571 return CallResult.second;
1574 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1575 const PPCSubtarget &Subtarget) const {
1576 MachineFunction &MF = DAG.getMachineFunction();
1577 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1579 DebugLoc dl = Op.getDebugLoc();
1581 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1582 // vastart just stores the address of the VarArgsFrameIndex slot into the
1583 // memory location argument.
1584 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1585 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1586 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1587 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1588 MachinePointerInfo(SV),
1592 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1593 // We suppose the given va_list is already allocated.
1596 // char gpr; /* index into the array of 8 GPRs
1597 // * stored in the register save area
1598 // * gpr=0 corresponds to r3,
1599 // * gpr=1 to r4, etc.
1601 // char fpr; /* index into the array of 8 FPRs
1602 // * stored in the register save area
1603 // * fpr=0 corresponds to f1,
1604 // * fpr=1 to f2, etc.
1606 // char *overflow_arg_area;
1607 // /* location on stack that holds
1608 // * the next overflow argument
1610 // char *reg_save_area;
1611 // /* where r3:r10 and f1:f8 (if saved)
1617 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1618 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1621 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1623 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1625 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1628 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1629 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1631 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1632 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1634 uint64_t FPROffset = 1;
1635 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1637 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1639 // Store first byte : number of int regs
1640 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1642 MachinePointerInfo(SV),
1643 MVT::i8, false, false, 0);
1644 uint64_t nextOffset = FPROffset;
1645 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1648 // Store second byte : number of float regs
1649 SDValue secondStore =
1650 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1651 MachinePointerInfo(SV, nextOffset), MVT::i8,
1653 nextOffset += StackOffset;
1654 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1656 // Store second word : arguments given on stack
1657 SDValue thirdStore =
1658 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1659 MachinePointerInfo(SV, nextOffset),
1661 nextOffset += FrameOffset;
1662 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1664 // Store third word : arguments given in registers
1665 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1666 MachinePointerInfo(SV, nextOffset),
1671 #include "PPCGenCallingConv.inc"
1673 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1674 CCValAssign::LocInfo &LocInfo,
1675 ISD::ArgFlagsTy &ArgFlags,
1680 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1682 CCValAssign::LocInfo &LocInfo,
1683 ISD::ArgFlagsTy &ArgFlags,
1685 static const uint16_t ArgRegs[] = {
1686 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1687 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1689 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1691 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1693 // Skip one register if the first unallocated register has an even register
1694 // number and there are still argument registers available which have not been
1695 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1696 // need to skip a register if RegNum is odd.
1697 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1698 State.AllocateReg(ArgRegs[RegNum]);
1701 // Always return false here, as this function only makes sure that the first
1702 // unallocated register has an odd register number and does not actually
1703 // allocate a register for the current argument.
1707 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1709 CCValAssign::LocInfo &LocInfo,
1710 ISD::ArgFlagsTy &ArgFlags,
1712 static const uint16_t ArgRegs[] = {
1713 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1717 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1719 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1721 // If there is only one Floating-point register left we need to put both f64
1722 // values of a split ppc_fp128 value on the stack.
1723 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1724 State.AllocateReg(ArgRegs[RegNum]);
1727 // Always return false here, as this function only makes sure that the two f64
1728 // values a ppc_fp128 value is split into are both passed in registers or both
1729 // passed on the stack and does not actually allocate a register for the
1730 // current argument.
1734 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1736 static const uint16_t *GetFPR() {
1737 static const uint16_t FPR[] = {
1738 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1739 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1745 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1747 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1748 unsigned PtrByteSize) {
1749 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1750 if (Flags.isByVal())
1751 ArgSize = Flags.getByValSize();
1752 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1758 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1759 CallingConv::ID CallConv, bool isVarArg,
1760 const SmallVectorImpl<ISD::InputArg>
1762 DebugLoc dl, SelectionDAG &DAG,
1763 SmallVectorImpl<SDValue> &InVals)
1765 if (PPCSubTarget.isSVR4ABI()) {
1766 if (PPCSubTarget.isPPC64())
1767 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1770 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1773 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1779 PPCTargetLowering::LowerFormalArguments_32SVR4(
1781 CallingConv::ID CallConv, bool isVarArg,
1782 const SmallVectorImpl<ISD::InputArg>
1784 DebugLoc dl, SelectionDAG &DAG,
1785 SmallVectorImpl<SDValue> &InVals) const {
1787 // 32-bit SVR4 ABI Stack Frame Layout:
1788 // +-----------------------------------+
1789 // +--> | Back chain |
1790 // | +-----------------------------------+
1791 // | | Floating-point register save area |
1792 // | +-----------------------------------+
1793 // | | General register save area |
1794 // | +-----------------------------------+
1795 // | | CR save word |
1796 // | +-----------------------------------+
1797 // | | VRSAVE save word |
1798 // | +-----------------------------------+
1799 // | | Alignment padding |
1800 // | +-----------------------------------+
1801 // | | Vector register save area |
1802 // | +-----------------------------------+
1803 // | | Local variable space |
1804 // | +-----------------------------------+
1805 // | | Parameter list area |
1806 // | +-----------------------------------+
1807 // | | LR save word |
1808 // | +-----------------------------------+
1809 // SP--> +--- | Back chain |
1810 // +-----------------------------------+
1813 // System V Application Binary Interface PowerPC Processor Supplement
1814 // AltiVec Technology Programming Interface Manual
1816 MachineFunction &MF = DAG.getMachineFunction();
1817 MachineFrameInfo *MFI = MF.getFrameInfo();
1818 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1820 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1821 // Potential tail calls could cause overwriting of argument stack slots.
1822 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1823 (CallConv == CallingConv::Fast));
1824 unsigned PtrByteSize = 4;
1826 // Assign locations to all of the incoming arguments.
1827 SmallVector<CCValAssign, 16> ArgLocs;
1828 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1829 getTargetMachine(), ArgLocs, *DAG.getContext());
1831 // Reserve space for the linkage area on the stack.
1832 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1834 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1836 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1837 CCValAssign &VA = ArgLocs[i];
1839 // Arguments stored in registers.
1840 if (VA.isRegLoc()) {
1841 const TargetRegisterClass *RC;
1842 EVT ValVT = VA.getValVT();
1844 switch (ValVT.getSimpleVT().SimpleTy) {
1846 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1848 RC = &PPC::GPRCRegClass;
1851 RC = &PPC::F4RCRegClass;
1854 RC = &PPC::F8RCRegClass;
1860 RC = &PPC::VRRCRegClass;
1864 // Transform the arguments stored in physical registers into virtual ones.
1865 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1866 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1868 InVals.push_back(ArgValue);
1870 // Argument stored in memory.
1871 assert(VA.isMemLoc());
1873 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1874 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1877 // Create load nodes to retrieve arguments from the stack.
1878 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1879 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1880 MachinePointerInfo(),
1881 false, false, false, 0));
1885 // Assign locations to all of the incoming aggregate by value arguments.
1886 // Aggregates passed by value are stored in the local variable space of the
1887 // caller's stack frame, right above the parameter list area.
1888 SmallVector<CCValAssign, 16> ByValArgLocs;
1889 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1890 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1892 // Reserve stack space for the allocations in CCInfo.
1893 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1895 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1897 // Area that is at least reserved in the caller of this function.
1898 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1900 // Set the size that is at least reserved in caller of this function. Tail
1901 // call optimized function's reserved stack space needs to be aligned so that
1902 // taking the difference between two stack areas will result in an aligned
1904 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1907 std::max(MinReservedArea,
1908 PPCFrameLowering::getMinCallFrameSize(false, false));
1910 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1911 getStackAlignment();
1912 unsigned AlignMask = TargetAlign-1;
1913 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1915 FI->setMinReservedArea(MinReservedArea);
1917 SmallVector<SDValue, 8> MemOps;
1919 // If the function takes variable number of arguments, make a frame index for
1920 // the start of the first vararg value... for expansion of llvm.va_start.
1922 static const uint16_t GPArgRegs[] = {
1923 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1924 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1926 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1928 static const uint16_t FPArgRegs[] = {
1929 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1932 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1934 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1936 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1939 // Make room for NumGPArgRegs and NumFPArgRegs.
1940 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1941 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1943 FuncInfo->setVarArgsStackOffset(
1944 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1945 CCInfo.getNextStackOffset(), true));
1947 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1948 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1950 // The fixed integer arguments of a variadic function are stored to the
1951 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1952 // the result of va_next.
1953 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1954 // Get an existing live-in vreg, or add a new one.
1955 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1957 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1959 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1960 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1961 MachinePointerInfo(), false, false, 0);
1962 MemOps.push_back(Store);
1963 // Increment the address by four for the next argument to store
1964 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1965 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1968 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1970 // The double arguments are stored to the VarArgsFrameIndex
1972 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1973 // Get an existing live-in vreg, or add a new one.
1974 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1976 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1978 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1979 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1980 MachinePointerInfo(), false, false, 0);
1981 MemOps.push_back(Store);
1982 // Increment the address by eight for the next argument to store
1983 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1985 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1989 if (!MemOps.empty())
1990 Chain = DAG.getNode(ISD::TokenFactor, dl,
1991 MVT::Other, &MemOps[0], MemOps.size());
1996 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1997 // value to MVT::i64 and then truncate to the correct register size.
1999 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
2000 SelectionDAG &DAG, SDValue ArgVal,
2001 DebugLoc dl) const {
2003 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2004 DAG.getValueType(ObjectVT));
2005 else if (Flags.isZExt())
2006 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2007 DAG.getValueType(ObjectVT));
2009 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2012 // Set the size that is at least reserved in caller of this function. Tail
2013 // call optimized functions' reserved stack space needs to be aligned so that
2014 // taking the difference between two stack areas will result in an aligned
2017 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
2018 unsigned nAltivecParamsAtEnd,
2019 unsigned MinReservedArea,
2020 bool isPPC64) const {
2021 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2022 // Add the Altivec parameters at the end, if needed.
2023 if (nAltivecParamsAtEnd) {
2024 MinReservedArea = ((MinReservedArea+15)/16)*16;
2025 MinReservedArea += 16*nAltivecParamsAtEnd;
2028 std::max(MinReservedArea,
2029 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2030 unsigned TargetAlign
2031 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2032 getStackAlignment();
2033 unsigned AlignMask = TargetAlign-1;
2034 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2035 FI->setMinReservedArea(MinReservedArea);
2039 PPCTargetLowering::LowerFormalArguments_64SVR4(
2041 CallingConv::ID CallConv, bool isVarArg,
2042 const SmallVectorImpl<ISD::InputArg>
2044 DebugLoc dl, SelectionDAG &DAG,
2045 SmallVectorImpl<SDValue> &InVals) const {
2046 // TODO: add description of PPC stack frame format, or at least some docs.
2048 MachineFunction &MF = DAG.getMachineFunction();
2049 MachineFrameInfo *MFI = MF.getFrameInfo();
2050 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2052 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2053 // Potential tail calls could cause overwriting of argument stack slots.
2054 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2055 (CallConv == CallingConv::Fast));
2056 unsigned PtrByteSize = 8;
2058 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2059 // Area that is at least reserved in caller of this function.
2060 unsigned MinReservedArea = ArgOffset;
2062 static const uint16_t GPR[] = {
2063 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2064 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2067 static const uint16_t *FPR = GetFPR();
2069 static const uint16_t VR[] = {
2070 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2071 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2074 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2075 const unsigned Num_FPR_Regs = 13;
2076 const unsigned Num_VR_Regs = array_lengthof(VR);
2078 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2080 // Add DAG nodes to load the arguments or copy them out of registers. On
2081 // entry to a function on PPC, the arguments start after the linkage area,
2082 // although the first ones are often in registers.
2084 SmallVector<SDValue, 8> MemOps;
2085 unsigned nAltivecParamsAtEnd = 0;
2086 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2087 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2089 bool needsLoad = false;
2090 EVT ObjectVT = Ins[ArgNo].VT;
2091 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2092 unsigned ArgSize = ObjSize;
2093 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2095 unsigned CurArgOffset = ArgOffset;
2097 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2098 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2099 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2101 MinReservedArea = ((MinReservedArea+15)/16)*16;
2102 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2106 nAltivecParamsAtEnd++;
2108 // Calculate min reserved area.
2109 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2113 // FIXME the codegen can be much improved in some cases.
2114 // We do not have to keep everything in memory.
2115 if (Flags.isByVal()) {
2116 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2117 ObjSize = Flags.getByValSize();
2118 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2119 // Empty aggregate parameters do not take up registers. Examples:
2123 // etc. However, we have to provide a place-holder in InVals, so
2124 // pretend we have an 8-byte item at the current address for that
2127 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2128 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2129 InVals.push_back(FIN);
2132 // All aggregates smaller than 8 bytes must be passed right-justified.
2133 if (ObjSize < PtrByteSize)
2134 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2135 // The value of the object is its address.
2136 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2137 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2138 InVals.push_back(FIN);
2141 if (GPR_idx != Num_GPR_Regs) {
2142 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2143 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2146 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2147 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2148 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2149 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2150 MachinePointerInfo(FuncArg, CurArgOffset),
2151 ObjType, false, false, 0);
2153 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2154 // store the whole register as-is to the parameter save area
2155 // slot. The address of the parameter was already calculated
2156 // above (InVals.push_back(FIN)) to be the right-justified
2157 // offset within the slot. For this store, we need a new
2158 // frame index that points at the beginning of the slot.
2159 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2160 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2161 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2162 MachinePointerInfo(FuncArg, ArgOffset),
2166 MemOps.push_back(Store);
2169 // Whether we copied from a register or not, advance the offset
2170 // into the parameter save area by a full doubleword.
2171 ArgOffset += PtrByteSize;
2175 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2176 // Store whatever pieces of the object are in registers
2177 // to memory. ArgOffset will be the address of the beginning
2179 if (GPR_idx != Num_GPR_Regs) {
2181 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2182 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2183 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2184 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2185 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2186 MachinePointerInfo(FuncArg, ArgOffset),
2188 MemOps.push_back(Store);
2190 ArgOffset += PtrByteSize;
2192 ArgOffset += ArgSize - j;
2199 switch (ObjectVT.getSimpleVT().SimpleTy) {
2200 default: llvm_unreachable("Unhandled argument type!");
2203 if (GPR_idx != Num_GPR_Regs) {
2204 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2205 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2207 if (ObjectVT == MVT::i32)
2208 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2209 // value to MVT::i64 and then truncate to the correct register size.
2210 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2215 ArgSize = PtrByteSize;
2222 // Every 8 bytes of argument space consumes one of the GPRs available for
2223 // argument passing.
2224 if (GPR_idx != Num_GPR_Regs) {
2227 if (FPR_idx != Num_FPR_Regs) {
2230 if (ObjectVT == MVT::f32)
2231 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2233 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2235 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2239 ArgSize = PtrByteSize;
2248 // Note that vector arguments in registers don't reserve stack space,
2249 // except in varargs functions.
2250 if (VR_idx != Num_VR_Regs) {
2251 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2252 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2254 while ((ArgOffset % 16) != 0) {
2255 ArgOffset += PtrByteSize;
2256 if (GPR_idx != Num_GPR_Regs)
2260 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2264 // Vectors are aligned.
2265 ArgOffset = ((ArgOffset+15)/16)*16;
2266 CurArgOffset = ArgOffset;
2273 // We need to load the argument to a virtual register if we determined
2274 // above that we ran out of physical registers of the appropriate type.
2276 int FI = MFI->CreateFixedObject(ObjSize,
2277 CurArgOffset + (ArgSize - ObjSize),
2279 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2280 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2281 false, false, false, 0);
2284 InVals.push_back(ArgVal);
2287 // Set the size that is at least reserved in caller of this function. Tail
2288 // call optimized functions' reserved stack space needs to be aligned so that
2289 // taking the difference between two stack areas will result in an aligned
2291 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2293 // If the function takes variable number of arguments, make a frame index for
2294 // the start of the first vararg value... for expansion of llvm.va_start.
2296 int Depth = ArgOffset;
2298 FuncInfo->setVarArgsFrameIndex(
2299 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2300 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2302 // If this function is vararg, store any remaining integer argument regs
2303 // to their spots on the stack so that they may be loaded by deferencing the
2304 // result of va_next.
2305 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2306 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2307 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2308 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2309 MachinePointerInfo(), false, false, 0);
2310 MemOps.push_back(Store);
2311 // Increment the address by four for the next argument to store
2312 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2313 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2317 if (!MemOps.empty())
2318 Chain = DAG.getNode(ISD::TokenFactor, dl,
2319 MVT::Other, &MemOps[0], MemOps.size());
2325 PPCTargetLowering::LowerFormalArguments_Darwin(
2327 CallingConv::ID CallConv, bool isVarArg,
2328 const SmallVectorImpl<ISD::InputArg>
2330 DebugLoc dl, SelectionDAG &DAG,
2331 SmallVectorImpl<SDValue> &InVals) const {
2332 // TODO: add description of PPC stack frame format, or at least some docs.
2334 MachineFunction &MF = DAG.getMachineFunction();
2335 MachineFrameInfo *MFI = MF.getFrameInfo();
2336 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2338 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2339 bool isPPC64 = PtrVT == MVT::i64;
2340 // Potential tail calls could cause overwriting of argument stack slots.
2341 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2342 (CallConv == CallingConv::Fast));
2343 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2345 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2346 // Area that is at least reserved in caller of this function.
2347 unsigned MinReservedArea = ArgOffset;
2349 static const uint16_t GPR_32[] = { // 32-bit registers.
2350 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2351 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2353 static const uint16_t GPR_64[] = { // 64-bit registers.
2354 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2355 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2358 static const uint16_t *FPR = GetFPR();
2360 static const uint16_t VR[] = {
2361 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2362 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2365 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2366 const unsigned Num_FPR_Regs = 13;
2367 const unsigned Num_VR_Regs = array_lengthof( VR);
2369 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2371 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2373 // In 32-bit non-varargs functions, the stack space for vectors is after the
2374 // stack space for non-vectors. We do not use this space unless we have
2375 // too many vectors to fit in registers, something that only occurs in
2376 // constructed examples:), but we have to walk the arglist to figure
2377 // that out...for the pathological case, compute VecArgOffset as the
2378 // start of the vector parameter area. Computing VecArgOffset is the
2379 // entire point of the following loop.
2380 unsigned VecArgOffset = ArgOffset;
2381 if (!isVarArg && !isPPC64) {
2382 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2384 EVT ObjectVT = Ins[ArgNo].VT;
2385 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2387 if (Flags.isByVal()) {
2388 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2389 unsigned ObjSize = Flags.getByValSize();
2391 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2392 VecArgOffset += ArgSize;
2396 switch(ObjectVT.getSimpleVT().SimpleTy) {
2397 default: llvm_unreachable("Unhandled argument type!");
2402 case MVT::i64: // PPC64
2404 // FIXME: We are guaranteed to be !isPPC64 at this point.
2405 // Does MVT::i64 apply?
2412 // Nothing to do, we're only looking at Nonvector args here.
2417 // We've found where the vector parameter area in memory is. Skip the
2418 // first 12 parameters; these don't use that memory.
2419 VecArgOffset = ((VecArgOffset+15)/16)*16;
2420 VecArgOffset += 12*16;
2422 // Add DAG nodes to load the arguments or copy them out of registers. On
2423 // entry to a function on PPC, the arguments start after the linkage area,
2424 // although the first ones are often in registers.
2426 SmallVector<SDValue, 8> MemOps;
2427 unsigned nAltivecParamsAtEnd = 0;
2428 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2429 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2431 bool needsLoad = false;
2432 EVT ObjectVT = Ins[ArgNo].VT;
2433 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2434 unsigned ArgSize = ObjSize;
2435 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2437 unsigned CurArgOffset = ArgOffset;
2439 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2440 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2441 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2442 if (isVarArg || isPPC64) {
2443 MinReservedArea = ((MinReservedArea+15)/16)*16;
2444 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2447 } else nAltivecParamsAtEnd++;
2449 // Calculate min reserved area.
2450 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2454 // FIXME the codegen can be much improved in some cases.
2455 // We do not have to keep everything in memory.
2456 if (Flags.isByVal()) {
2457 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2458 ObjSize = Flags.getByValSize();
2459 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2460 // Objects of size 1 and 2 are right justified, everything else is
2461 // left justified. This means the memory address is adjusted forwards.
2462 if (ObjSize==1 || ObjSize==2) {
2463 CurArgOffset = CurArgOffset + (4 - ObjSize);
2465 // The value of the object is its address.
2466 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2467 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2468 InVals.push_back(FIN);
2469 if (ObjSize==1 || ObjSize==2) {
2470 if (GPR_idx != Num_GPR_Regs) {
2473 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2475 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2476 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2477 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2478 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2479 MachinePointerInfo(FuncArg,
2481 ObjType, false, false, 0);
2482 MemOps.push_back(Store);
2486 ArgOffset += PtrByteSize;
2490 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2491 // Store whatever pieces of the object are in registers
2492 // to memory. ArgOffset will be the address of the beginning
2494 if (GPR_idx != Num_GPR_Regs) {
2497 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2499 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2500 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2501 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2502 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2503 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2504 MachinePointerInfo(FuncArg, ArgOffset),
2506 MemOps.push_back(Store);
2508 ArgOffset += PtrByteSize;
2510 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2517 switch (ObjectVT.getSimpleVT().SimpleTy) {
2518 default: llvm_unreachable("Unhandled argument type!");
2521 if (GPR_idx != Num_GPR_Regs) {
2522 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2523 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2527 ArgSize = PtrByteSize;
2529 // All int arguments reserve stack space in the Darwin ABI.
2530 ArgOffset += PtrByteSize;
2534 case MVT::i64: // PPC64
2535 if (GPR_idx != Num_GPR_Regs) {
2536 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2537 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2539 if (ObjectVT == MVT::i32)
2540 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2541 // value to MVT::i64 and then truncate to the correct register size.
2542 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2547 ArgSize = PtrByteSize;
2549 // All int arguments reserve stack space in the Darwin ABI.
2555 // Every 4 bytes of argument space consumes one of the GPRs available for
2556 // argument passing.
2557 if (GPR_idx != Num_GPR_Regs) {
2559 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2562 if (FPR_idx != Num_FPR_Regs) {
2565 if (ObjectVT == MVT::f32)
2566 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2568 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2570 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2576 // All FP arguments reserve stack space in the Darwin ABI.
2577 ArgOffset += isPPC64 ? 8 : ObjSize;
2583 // Note that vector arguments in registers don't reserve stack space,
2584 // except in varargs functions.
2585 if (VR_idx != Num_VR_Regs) {
2586 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2587 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2589 while ((ArgOffset % 16) != 0) {
2590 ArgOffset += PtrByteSize;
2591 if (GPR_idx != Num_GPR_Regs)
2595 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2599 if (!isVarArg && !isPPC64) {
2600 // Vectors go after all the nonvectors.
2601 CurArgOffset = VecArgOffset;
2604 // Vectors are aligned.
2605 ArgOffset = ((ArgOffset+15)/16)*16;
2606 CurArgOffset = ArgOffset;
2614 // We need to load the argument to a virtual register if we determined above
2615 // that we ran out of physical registers of the appropriate type.
2617 int FI = MFI->CreateFixedObject(ObjSize,
2618 CurArgOffset + (ArgSize - ObjSize),
2620 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2621 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2622 false, false, false, 0);
2625 InVals.push_back(ArgVal);
2628 // Set the size that is at least reserved in caller of this function. Tail
2629 // call optimized functions' reserved stack space needs to be aligned so that
2630 // taking the difference between two stack areas will result in an aligned
2632 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2634 // If the function takes variable number of arguments, make a frame index for
2635 // the start of the first vararg value... for expansion of llvm.va_start.
2637 int Depth = ArgOffset;
2639 FuncInfo->setVarArgsFrameIndex(
2640 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2642 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2644 // If this function is vararg, store any remaining integer argument regs
2645 // to their spots on the stack so that they may be loaded by deferencing the
2646 // result of va_next.
2647 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2651 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2653 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2655 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2656 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2657 MachinePointerInfo(), false, false, 0);
2658 MemOps.push_back(Store);
2659 // Increment the address by four for the next argument to store
2660 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2661 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2665 if (!MemOps.empty())
2666 Chain = DAG.getNode(ISD::TokenFactor, dl,
2667 MVT::Other, &MemOps[0], MemOps.size());
2672 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2673 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2675 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2679 const SmallVectorImpl<ISD::OutputArg>
2681 const SmallVectorImpl<SDValue> &OutVals,
2682 unsigned &nAltivecParamsAtEnd) {
2683 // Count how many bytes are to be pushed on the stack, including the linkage
2684 // area, and parameter passing area. We start with 24/48 bytes, which is
2685 // prereserved space for [SP][CR][LR][3 x unused].
2686 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2687 unsigned NumOps = Outs.size();
2688 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2690 // Add up all the space actually used.
2691 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2692 // they all go in registers, but we must reserve stack space for them for
2693 // possible use by the caller. In varargs or 64-bit calls, parameters are
2694 // assigned stack space in order, with padding so Altivec parameters are
2696 nAltivecParamsAtEnd = 0;
2697 for (unsigned i = 0; i != NumOps; ++i) {
2698 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2699 EVT ArgVT = Outs[i].VT;
2700 // Varargs Altivec parameters are padded to a 16 byte boundary.
2701 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2702 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2703 if (!isVarArg && !isPPC64) {
2704 // Non-varargs Altivec parameters go after all the non-Altivec
2705 // parameters; handle those later so we know how much padding we need.
2706 nAltivecParamsAtEnd++;
2709 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2710 NumBytes = ((NumBytes+15)/16)*16;
2712 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2715 // Allow for Altivec parameters at the end, if needed.
2716 if (nAltivecParamsAtEnd) {
2717 NumBytes = ((NumBytes+15)/16)*16;
2718 NumBytes += 16*nAltivecParamsAtEnd;
2721 // The prolog code of the callee may store up to 8 GPR argument registers to
2722 // the stack, allowing va_start to index over them in memory if its varargs.
2723 // Because we cannot tell if this is needed on the caller side, we have to
2724 // conservatively assume that it is needed. As such, make sure we have at
2725 // least enough stack space for the caller to store the 8 GPRs.
2726 NumBytes = std::max(NumBytes,
2727 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2729 // Tail call needs the stack to be aligned.
2730 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2731 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2732 getFrameLowering()->getStackAlignment();
2733 unsigned AlignMask = TargetAlign-1;
2734 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2740 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2741 /// adjusted to accommodate the arguments for the tailcall.
2742 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2743 unsigned ParamSize) {
2745 if (!isTailCall) return 0;
2747 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2748 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2749 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2750 // Remember only if the new adjustement is bigger.
2751 if (SPDiff < FI->getTailCallSPDelta())
2752 FI->setTailCallSPDelta(SPDiff);
2757 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2758 /// for tail call optimization. Targets which want to do tail call
2759 /// optimization should implement this function.
2761 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2762 CallingConv::ID CalleeCC,
2764 const SmallVectorImpl<ISD::InputArg> &Ins,
2765 SelectionDAG& DAG) const {
2766 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2769 // Variable argument functions are not supported.
2773 MachineFunction &MF = DAG.getMachineFunction();
2774 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2775 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2776 // Functions containing by val parameters are not supported.
2777 for (unsigned i = 0; i != Ins.size(); i++) {
2778 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2779 if (Flags.isByVal()) return false;
2782 // Non PIC/GOT tail calls are supported.
2783 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2786 // At the moment we can only do local tail calls (in same module, hidden
2787 // or protected) if we are generating PIC.
2788 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2789 return G->getGlobal()->hasHiddenVisibility()
2790 || G->getGlobal()->hasProtectedVisibility();
2796 /// isCallCompatibleAddress - Return the immediate to use if the specified
2797 /// 32-bit value is representable in the immediate field of a BxA instruction.
2798 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2799 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2802 int Addr = C->getZExtValue();
2803 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2804 SignExtend32<26>(Addr) != Addr)
2805 return 0; // Top 6 bits have to be sext of immediate.
2807 return DAG.getConstant((int)C->getZExtValue() >> 2,
2808 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2813 struct TailCallArgumentInfo {
2818 TailCallArgumentInfo() : FrameIdx(0) {}
2823 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2825 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2827 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2828 SmallVector<SDValue, 8> &MemOpChains,
2830 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2831 SDValue Arg = TailCallArgs[i].Arg;
2832 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2833 int FI = TailCallArgs[i].FrameIdx;
2834 // Store relative to framepointer.
2835 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2836 MachinePointerInfo::getFixedStack(FI),
2841 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2842 /// the appropriate stack slot for the tail call optimized function call.
2843 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2844 MachineFunction &MF,
2853 // Calculate the new stack slot for the return address.
2854 int SlotSize = isPPC64 ? 8 : 4;
2855 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2857 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2858 NewRetAddrLoc, true);
2859 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2860 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2861 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2862 MachinePointerInfo::getFixedStack(NewRetAddr),
2865 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2866 // slot as the FP is never overwritten.
2869 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2870 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2872 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2873 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2874 MachinePointerInfo::getFixedStack(NewFPIdx),
2881 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2882 /// the position of the argument.
2884 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2885 SDValue Arg, int SPDiff, unsigned ArgOffset,
2886 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2887 int Offset = ArgOffset + SPDiff;
2888 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2889 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2890 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2891 SDValue FIN = DAG.getFrameIndex(FI, VT);
2892 TailCallArgumentInfo Info;
2894 Info.FrameIdxOp = FIN;
2896 TailCallArguments.push_back(Info);
2899 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2900 /// stack slot. Returns the chain as result and the loaded frame pointers in
2901 /// LROpOut/FPOpout. Used when tail calling.
2902 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2908 DebugLoc dl) const {
2910 // Load the LR and FP stack slot for later adjusting.
2911 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2912 LROpOut = getReturnAddrFrameIndex(DAG);
2913 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2914 false, false, false, 0);
2915 Chain = SDValue(LROpOut.getNode(), 1);
2917 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2918 // slot as the FP is never overwritten.
2920 FPOpOut = getFramePointerFrameIndex(DAG);
2921 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2922 false, false, false, 0);
2923 Chain = SDValue(FPOpOut.getNode(), 1);
2929 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2930 /// by "Src" to address "Dst" of size "Size". Alignment information is
2931 /// specified by the specific parameter attribute. The copy will be passed as
2932 /// a byval function parameter.
2933 /// Sometimes what we are copying is the end of a larger object, the part that
2934 /// does not fit in registers.
2936 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2937 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2939 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2940 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2941 false, false, MachinePointerInfo(0),
2942 MachinePointerInfo(0));
2945 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2948 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2949 SDValue Arg, SDValue PtrOff, int SPDiff,
2950 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2951 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2952 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2954 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2959 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2961 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2962 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2963 DAG.getConstant(ArgOffset, PtrVT));
2965 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2966 MachinePointerInfo(), false, false, 0));
2967 // Calculate and remember argument location.
2968 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2973 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2974 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2975 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2976 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2977 MachineFunction &MF = DAG.getMachineFunction();
2979 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2980 // might overwrite each other in case of tail call optimization.
2981 SmallVector<SDValue, 8> MemOpChains2;
2982 // Do not flag preceding copytoreg stuff together with the following stuff.
2984 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2986 if (!MemOpChains2.empty())
2987 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2988 &MemOpChains2[0], MemOpChains2.size());
2990 // Store the return address to the appropriate stack slot.
2991 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2992 isPPC64, isDarwinABI, dl);
2994 // Emit callseq_end just before tailcall node.
2995 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2996 DAG.getIntPtrConstant(0, true), InFlag);
2997 InFlag = Chain.getValue(1);
3001 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
3002 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
3003 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
3004 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
3005 const PPCSubtarget &PPCSubTarget) {
3007 bool isPPC64 = PPCSubTarget.isPPC64();
3008 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3010 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3011 NodeTys.push_back(MVT::Other); // Returns a chain
3012 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
3014 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
3016 bool needIndirectCall = true;
3017 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3018 // If this is an absolute destination address, use the munged value.
3019 Callee = SDValue(Dest, 0);
3020 needIndirectCall = false;
3023 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3024 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
3025 // Use indirect calls for ALL functions calls in JIT mode, since the
3026 // far-call stubs may be outside relocation limits for a BL instruction.
3027 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
3028 unsigned OpFlags = 0;
3029 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3030 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3031 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3032 (G->getGlobal()->isDeclaration() ||
3033 G->getGlobal()->isWeakForLinker())) {
3034 // PC-relative references to external symbols should go through $stub,
3035 // unless we're building with the leopard linker or later, which
3036 // automatically synthesizes these stubs.
3037 OpFlags = PPCII::MO_DARWIN_STUB;
3040 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3041 // every direct call is) turn it into a TargetGlobalAddress /
3042 // TargetExternalSymbol node so that legalize doesn't hack it.
3043 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3044 Callee.getValueType(),
3046 needIndirectCall = false;
3050 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3051 unsigned char OpFlags = 0;
3053 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3054 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3055 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3056 // PC-relative references to external symbols should go through $stub,
3057 // unless we're building with the leopard linker or later, which
3058 // automatically synthesizes these stubs.
3059 OpFlags = PPCII::MO_DARWIN_STUB;
3062 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3064 needIndirectCall = false;
3067 if (needIndirectCall) {
3068 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3069 // to do the call, we can't use PPCISD::CALL.
3070 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3072 if (isSVR4ABI && isPPC64) {
3073 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3074 // entry point, but to the function descriptor (the function entry point
3075 // address is part of the function descriptor though).
3076 // The function descriptor is a three doubleword structure with the
3077 // following fields: function entry point, TOC base address and
3078 // environment pointer.
3079 // Thus for a call through a function pointer, the following actions need
3081 // 1. Save the TOC of the caller in the TOC save area of its stack
3082 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3083 // 2. Load the address of the function entry point from the function
3085 // 3. Load the TOC of the callee from the function descriptor into r2.
3086 // 4. Load the environment pointer from the function descriptor into
3088 // 5. Branch to the function entry point address.
3089 // 6. On return of the callee, the TOC of the caller needs to be
3090 // restored (this is done in FinishCall()).
3092 // All those operations are flagged together to ensure that no other
3093 // operations can be scheduled in between. E.g. without flagging the
3094 // operations together, a TOC access in the caller could be scheduled
3095 // between the load of the callee TOC and the branch to the callee, which
3096 // results in the TOC access going through the TOC of the callee instead
3097 // of going through the TOC of the caller, which leads to incorrect code.
3099 // Load the address of the function entry point from the function
3101 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3102 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3103 InFlag.getNode() ? 3 : 2);
3104 Chain = LoadFuncPtr.getValue(1);
3105 InFlag = LoadFuncPtr.getValue(2);
3107 // Load environment pointer into r11.
3108 // Offset of the environment pointer within the function descriptor.
3109 SDValue PtrOff = DAG.getIntPtrConstant(16);
3111 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3112 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3114 Chain = LoadEnvPtr.getValue(1);
3115 InFlag = LoadEnvPtr.getValue(2);
3117 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3119 Chain = EnvVal.getValue(0);
3120 InFlag = EnvVal.getValue(1);
3122 // Load TOC of the callee into r2. We are using a target-specific load
3123 // with r2 hard coded, because the result of a target-independent load
3124 // would never go directly into r2, since r2 is a reserved register (which
3125 // prevents the register allocator from allocating it), resulting in an
3126 // additional register being allocated and an unnecessary move instruction
3128 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3129 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3131 Chain = LoadTOCPtr.getValue(0);
3132 InFlag = LoadTOCPtr.getValue(1);
3134 MTCTROps[0] = Chain;
3135 MTCTROps[1] = LoadFuncPtr;
3136 MTCTROps[2] = InFlag;
3139 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3140 2 + (InFlag.getNode() != 0));
3141 InFlag = Chain.getValue(1);
3144 NodeTys.push_back(MVT::Other);
3145 NodeTys.push_back(MVT::Glue);
3146 Ops.push_back(Chain);
3147 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3149 // Add CTR register as callee so a bctr can be emitted later.
3151 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3154 // If this is a direct call, pass the chain and the callee.
3155 if (Callee.getNode()) {
3156 Ops.push_back(Chain);
3157 Ops.push_back(Callee);
3159 // If this is a tail call add stack pointer delta.
3161 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3163 // Add argument registers to the end of the list so that they are known live
3165 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3166 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3167 RegsToPass[i].second.getValueType()));
3173 bool isLocalCall(const SDValue &Callee)
3175 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3176 return !G->getGlobal()->isDeclaration() &&
3177 !G->getGlobal()->isWeakForLinker();
3182 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3183 CallingConv::ID CallConv, bool isVarArg,
3184 const SmallVectorImpl<ISD::InputArg> &Ins,
3185 DebugLoc dl, SelectionDAG &DAG,
3186 SmallVectorImpl<SDValue> &InVals) const {
3188 SmallVector<CCValAssign, 16> RVLocs;
3189 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3190 getTargetMachine(), RVLocs, *DAG.getContext());
3191 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3193 // Copy all of the result registers out of their specified physreg.
3194 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3195 CCValAssign &VA = RVLocs[i];
3196 assert(VA.isRegLoc() && "Can only return in registers!");
3198 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3199 VA.getLocReg(), VA.getLocVT(), InFlag);
3200 Chain = Val.getValue(1);
3201 InFlag = Val.getValue(2);
3203 switch (VA.getLocInfo()) {
3204 default: llvm_unreachable("Unknown loc info!");
3205 case CCValAssign::Full: break;
3206 case CCValAssign::AExt:
3207 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3209 case CCValAssign::ZExt:
3210 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3211 DAG.getValueType(VA.getValVT()));
3212 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3214 case CCValAssign::SExt:
3215 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3216 DAG.getValueType(VA.getValVT()));
3217 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3221 InVals.push_back(Val);
3228 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3229 bool isTailCall, bool isVarArg,
3231 SmallVector<std::pair<unsigned, SDValue>, 8>
3233 SDValue InFlag, SDValue Chain,
3235 int SPDiff, unsigned NumBytes,
3236 const SmallVectorImpl<ISD::InputArg> &Ins,
3237 SmallVectorImpl<SDValue> &InVals) const {
3238 std::vector<EVT> NodeTys;
3239 SmallVector<SDValue, 8> Ops;
3240 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3241 isTailCall, RegsToPass, Ops, NodeTys,
3244 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3245 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3246 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3248 // When performing tail call optimization the callee pops its arguments off
3249 // the stack. Account for this here so these bytes can be pushed back on in
3250 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3251 int BytesCalleePops =
3252 (CallConv == CallingConv::Fast &&
3253 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3255 // Add a register mask operand representing the call-preserved registers.
3256 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3257 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3258 assert(Mask && "Missing call preserved mask for calling convention");
3259 Ops.push_back(DAG.getRegisterMask(Mask));
3261 if (InFlag.getNode())
3262 Ops.push_back(InFlag);
3266 // If this is the first return lowered for this function, add the regs
3267 // to the liveout set for the function.
3268 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3269 SmallVector<CCValAssign, 16> RVLocs;
3270 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3271 getTargetMachine(), RVLocs, *DAG.getContext());
3272 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3273 for (unsigned i = 0; i != RVLocs.size(); ++i)
3274 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3277 assert(((Callee.getOpcode() == ISD::Register &&
3278 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3279 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3280 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3281 isa<ConstantSDNode>(Callee)) &&
3282 "Expecting an global address, external symbol, absolute value or register");
3284 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3287 // Add a NOP immediately after the branch instruction when using the 64-bit
3288 // SVR4 ABI. At link time, if caller and callee are in a different module and
3289 // thus have a different TOC, the call will be replaced with a call to a stub
3290 // function which saves the current TOC, loads the TOC of the callee and
3291 // branches to the callee. The NOP will be replaced with a load instruction
3292 // which restores the TOC of the caller from the TOC save slot of the current
3293 // stack frame. If caller and callee belong to the same module (and have the
3294 // same TOC), the NOP will remain unchanged.
3296 bool needsTOCRestore = false;
3297 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3298 if (CallOpc == PPCISD::BCTRL_SVR4) {
3299 // This is a call through a function pointer.
3300 // Restore the caller TOC from the save area into R2.
3301 // See PrepareCall() for more information about calls through function
3302 // pointers in the 64-bit SVR4 ABI.
3303 // We are using a target-specific load with r2 hard coded, because the
3304 // result of a target-independent load would never go directly into r2,
3305 // since r2 is a reserved register (which prevents the register allocator
3306 // from allocating it), resulting in an additional register being
3307 // allocated and an unnecessary move instruction being generated.
3308 needsTOCRestore = true;
3309 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3310 // Otherwise insert NOP for non-local calls.
3311 CallOpc = PPCISD::CALL_NOP_SVR4;
3315 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3316 InFlag = Chain.getValue(1);
3318 if (needsTOCRestore) {
3319 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3320 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3321 InFlag = Chain.getValue(1);
3324 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3325 DAG.getIntPtrConstant(BytesCalleePops, true),
3328 InFlag = Chain.getValue(1);
3330 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3331 Ins, dl, DAG, InVals);
3335 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3336 SmallVectorImpl<SDValue> &InVals) const {
3337 SelectionDAG &DAG = CLI.DAG;
3338 DebugLoc &dl = CLI.DL;
3339 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3340 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3341 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3342 SDValue Chain = CLI.Chain;
3343 SDValue Callee = CLI.Callee;
3344 bool &isTailCall = CLI.IsTailCall;
3345 CallingConv::ID CallConv = CLI.CallConv;
3346 bool isVarArg = CLI.IsVarArg;
3349 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3352 if (PPCSubTarget.isSVR4ABI()) {
3353 if (PPCSubTarget.isPPC64())
3354 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3355 isTailCall, Outs, OutVals, Ins,
3358 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3359 isTailCall, Outs, OutVals, Ins,
3363 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3364 isTailCall, Outs, OutVals, Ins,
3369 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3370 CallingConv::ID CallConv, bool isVarArg,
3372 const SmallVectorImpl<ISD::OutputArg> &Outs,
3373 const SmallVectorImpl<SDValue> &OutVals,
3374 const SmallVectorImpl<ISD::InputArg> &Ins,
3375 DebugLoc dl, SelectionDAG &DAG,
3376 SmallVectorImpl<SDValue> &InVals) const {
3377 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3378 // of the 32-bit SVR4 ABI stack frame layout.
3380 assert((CallConv == CallingConv::C ||
3381 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3383 unsigned PtrByteSize = 4;
3385 MachineFunction &MF = DAG.getMachineFunction();
3387 // Mark this function as potentially containing a function that contains a
3388 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3389 // and restoring the callers stack pointer in this functions epilog. This is
3390 // done because by tail calling the called function might overwrite the value
3391 // in this function's (MF) stack pointer stack slot 0(SP).
3392 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3393 CallConv == CallingConv::Fast)
3394 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3396 // Count how many bytes are to be pushed on the stack, including the linkage
3397 // area, parameter list area and the part of the local variable space which
3398 // contains copies of aggregates which are passed by value.
3400 // Assign locations to all of the outgoing arguments.
3401 SmallVector<CCValAssign, 16> ArgLocs;
3402 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3403 getTargetMachine(), ArgLocs, *DAG.getContext());
3405 // Reserve space for the linkage area on the stack.
3406 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3409 // Handle fixed and variable vector arguments differently.
3410 // Fixed vector arguments go into registers as long as registers are
3411 // available. Variable vector arguments always go into memory.
3412 unsigned NumArgs = Outs.size();
3414 for (unsigned i = 0; i != NumArgs; ++i) {
3415 MVT ArgVT = Outs[i].VT;
3416 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3419 if (Outs[i].IsFixed) {
3420 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3423 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3429 errs() << "Call operand #" << i << " has unhandled type "
3430 << EVT(ArgVT).getEVTString() << "\n";
3432 llvm_unreachable(0);
3436 // All arguments are treated the same.
3437 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3440 // Assign locations to all of the outgoing aggregate by value arguments.
3441 SmallVector<CCValAssign, 16> ByValArgLocs;
3442 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3443 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3445 // Reserve stack space for the allocations in CCInfo.
3446 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3448 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3450 // Size of the linkage area, parameter list area and the part of the local
3451 // space variable where copies of aggregates which are passed by value are
3453 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3455 // Calculate by how many bytes the stack has to be adjusted in case of tail
3456 // call optimization.
3457 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3459 // Adjust the stack pointer for the new arguments...
3460 // These operations are automatically eliminated by the prolog/epilog pass
3461 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3462 SDValue CallSeqStart = Chain;
3464 // Load the return address and frame pointer so it can be moved somewhere else
3467 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3470 // Set up a copy of the stack pointer for use loading and storing any
3471 // arguments that may not fit in the registers available for argument
3473 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3475 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3476 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3477 SmallVector<SDValue, 8> MemOpChains;
3479 bool seenFloatArg = false;
3480 // Walk the register/memloc assignments, inserting copies/loads.
3481 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3484 CCValAssign &VA = ArgLocs[i];
3485 SDValue Arg = OutVals[i];
3486 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3488 if (Flags.isByVal()) {
3489 // Argument is an aggregate which is passed by value, thus we need to
3490 // create a copy of it in the local variable space of the current stack
3491 // frame (which is the stack frame of the caller) and pass the address of
3492 // this copy to the callee.
3493 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3494 CCValAssign &ByValVA = ByValArgLocs[j++];
3495 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3497 // Memory reserved in the local variable space of the callers stack frame.
3498 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3500 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3501 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3503 // Create a copy of the argument in the local area of the current
3505 SDValue MemcpyCall =
3506 CreateCopyOfByValArgument(Arg, PtrOff,
3507 CallSeqStart.getNode()->getOperand(0),
3510 // This must go outside the CALLSEQ_START..END.
3511 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3512 CallSeqStart.getNode()->getOperand(1));
3513 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3514 NewCallSeqStart.getNode());
3515 Chain = CallSeqStart = NewCallSeqStart;
3517 // Pass the address of the aggregate copy on the stack either in a
3518 // physical register or in the parameter list area of the current stack
3519 // frame to the callee.
3523 if (VA.isRegLoc()) {
3524 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3525 // Put argument in a physical register.
3526 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3528 // Put argument in the parameter list area of the current stack frame.
3529 assert(VA.isMemLoc());
3530 unsigned LocMemOffset = VA.getLocMemOffset();
3533 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3534 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3536 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3537 MachinePointerInfo(),
3540 // Calculate and remember argument location.
3541 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3547 if (!MemOpChains.empty())
3548 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3549 &MemOpChains[0], MemOpChains.size());
3551 // Build a sequence of copy-to-reg nodes chained together with token chain
3552 // and flag operands which copy the outgoing args into the appropriate regs.
3554 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3555 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3556 RegsToPass[i].second, InFlag);
3557 InFlag = Chain.getValue(1);
3560 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3563 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3564 SDValue Ops[] = { Chain, InFlag };
3566 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3567 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3569 InFlag = Chain.getValue(1);
3573 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3574 false, TailCallArguments);
3576 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3577 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3581 // Copy an argument into memory, being careful to do this outside the
3582 // call sequence for the call to which the argument belongs.
3584 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3585 SDValue CallSeqStart,
3586 ISD::ArgFlagsTy Flags,
3588 DebugLoc dl) const {
3589 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3590 CallSeqStart.getNode()->getOperand(0),
3592 // The MEMCPY must go outside the CALLSEQ_START..END.
3593 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3594 CallSeqStart.getNode()->getOperand(1));
3595 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3596 NewCallSeqStart.getNode());
3597 return NewCallSeqStart;
3601 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3602 CallingConv::ID CallConv, bool isVarArg,
3604 const SmallVectorImpl<ISD::OutputArg> &Outs,
3605 const SmallVectorImpl<SDValue> &OutVals,
3606 const SmallVectorImpl<ISD::InputArg> &Ins,
3607 DebugLoc dl, SelectionDAG &DAG,
3608 SmallVectorImpl<SDValue> &InVals) const {
3610 unsigned NumOps = Outs.size();
3612 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3613 unsigned PtrByteSize = 8;
3615 MachineFunction &MF = DAG.getMachineFunction();
3617 // Mark this function as potentially containing a function that contains a
3618 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3619 // and restoring the callers stack pointer in this functions epilog. This is
3620 // done because by tail calling the called function might overwrite the value
3621 // in this function's (MF) stack pointer stack slot 0(SP).
3622 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3623 CallConv == CallingConv::Fast)
3624 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3626 unsigned nAltivecParamsAtEnd = 0;
3628 // Count how many bytes are to be pushed on the stack, including the linkage
3629 // area, and parameter passing area. We start with at least 48 bytes, which
3630 // is reserved space for [SP][CR][LR][3 x unused].
3631 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3634 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3635 Outs, OutVals, nAltivecParamsAtEnd);
3637 // Calculate by how many bytes the stack has to be adjusted in case of tail
3638 // call optimization.
3639 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3641 // To protect arguments on the stack from being clobbered in a tail call,
3642 // force all the loads to happen before doing any other lowering.
3644 Chain = DAG.getStackArgumentTokenFactor(Chain);
3646 // Adjust the stack pointer for the new arguments...
3647 // These operations are automatically eliminated by the prolog/epilog pass
3648 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3649 SDValue CallSeqStart = Chain;
3651 // Load the return address and frame pointer so it can be move somewhere else
3654 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3657 // Set up a copy of the stack pointer for use loading and storing any
3658 // arguments that may not fit in the registers available for argument
3660 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3662 // Figure out which arguments are going to go in registers, and which in
3663 // memory. Also, if this is a vararg function, floating point operations
3664 // must be stored to our stack, and loaded into integer regs as well, if
3665 // any integer regs are available for argument passing.
3666 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3667 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3669 static const uint16_t GPR[] = {
3670 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3671 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3673 static const uint16_t *FPR = GetFPR();
3675 static const uint16_t VR[] = {
3676 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3677 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3679 const unsigned NumGPRs = array_lengthof(GPR);
3680 const unsigned NumFPRs = 13;
3681 const unsigned NumVRs = array_lengthof(VR);
3683 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3684 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3686 SmallVector<SDValue, 8> MemOpChains;
3687 for (unsigned i = 0; i != NumOps; ++i) {
3688 SDValue Arg = OutVals[i];
3689 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3691 // PtrOff will be used to store the current argument to the stack if a
3692 // register cannot be found for it.
3695 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3697 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3699 // Promote integers to 64-bit values.
3700 if (Arg.getValueType() == MVT::i32) {
3701 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3702 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3703 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3706 // FIXME memcpy is used way more than necessary. Correctness first.
3707 // Note: "by value" is code for passing a structure by value, not
3709 if (Flags.isByVal()) {
3710 // Note: Size includes alignment padding, so
3711 // struct x { short a; char b; }
3712 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3713 // These are the proper values we need for right-justifying the
3714 // aggregate in a parameter register.
3715 unsigned Size = Flags.getByValSize();
3717 // An empty aggregate parameter takes up no storage and no
3722 // All aggregates smaller than 8 bytes must be passed right-justified.
3723 if (Size==1 || Size==2 || Size==4) {
3724 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3725 if (GPR_idx != NumGPRs) {
3726 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3727 MachinePointerInfo(), VT,
3729 MemOpChains.push_back(Load.getValue(1));
3730 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3732 ArgOffset += PtrByteSize;
3737 if (GPR_idx == NumGPRs && Size < 8) {
3738 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3739 PtrOff.getValueType());
3740 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3741 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3744 ArgOffset += PtrByteSize;
3747 // Copy entire object into memory. There are cases where gcc-generated
3748 // code assumes it is there, even if it could be put entirely into
3749 // registers. (This is not what the doc says.)
3751 // FIXME: The above statement is likely due to a misunderstanding of the
3752 // documents. All arguments must be copied into the parameter area BY
3753 // THE CALLEE in the event that the callee takes the address of any
3754 // formal argument. That has not yet been implemented. However, it is
3755 // reasonable to use the stack area as a staging area for the register
3758 // Skip this for small aggregates, as we will use the same slot for a
3759 // right-justified copy, below.
3761 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3765 // When a register is available, pass a small aggregate right-justified.
3766 if (Size < 8 && GPR_idx != NumGPRs) {
3767 // The easiest way to get this right-justified in a register
3768 // is to copy the structure into the rightmost portion of a
3769 // local variable slot, then load the whole slot into the
3771 // FIXME: The memcpy seems to produce pretty awful code for
3772 // small aggregates, particularly for packed ones.
3773 // FIXME: It would be preferable to use the slot in the
3774 // parameter save area instead of a new local variable.
3775 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3776 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3777 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3781 // Load the slot into the register.
3782 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3783 MachinePointerInfo(),
3784 false, false, false, 0);
3785 MemOpChains.push_back(Load.getValue(1));
3786 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3788 // Done with this argument.
3789 ArgOffset += PtrByteSize;
3793 // For aggregates larger than PtrByteSize, copy the pieces of the
3794 // object that fit into registers from the parameter save area.
3795 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3796 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3797 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3798 if (GPR_idx != NumGPRs) {
3799 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3800 MachinePointerInfo(),
3801 false, false, false, 0);
3802 MemOpChains.push_back(Load.getValue(1));
3803 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3804 ArgOffset += PtrByteSize;
3806 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3813 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3814 default: llvm_unreachable("Unexpected ValueType for argument!");
3817 if (GPR_idx != NumGPRs) {
3818 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3820 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3821 true, isTailCall, false, MemOpChains,
3822 TailCallArguments, dl);
3824 ArgOffset += PtrByteSize;
3828 if (FPR_idx != NumFPRs) {
3829 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3832 // A single float or an aggregate containing only a single float
3833 // must be passed right-justified in the stack doubleword, and
3834 // in the GPR, if one is available.
3836 if (Arg.getValueType().getSimpleVT().SimpleTy == MVT::f32) {
3837 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3838 StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3842 SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
3843 MachinePointerInfo(), false, false, 0);
3844 MemOpChains.push_back(Store);
3846 // Float varargs are always shadowed in available integer registers
3847 if (GPR_idx != NumGPRs) {
3848 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3849 MachinePointerInfo(), false, false,
3851 MemOpChains.push_back(Load.getValue(1));
3852 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3854 } else if (GPR_idx != NumGPRs)
3855 // If we have any FPRs remaining, we may also have GPRs remaining.
3858 // Single-precision floating-point values are mapped to the
3859 // second (rightmost) word of the stack doubleword.
3860 if (Arg.getValueType() == MVT::f32) {
3861 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3862 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3865 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3866 true, isTailCall, false, MemOpChains,
3867 TailCallArguments, dl);
3876 // These go aligned on the stack, or in the corresponding R registers
3877 // when within range. The Darwin PPC ABI doc claims they also go in
3878 // V registers; in fact gcc does this only for arguments that are
3879 // prototyped, not for those that match the ... We do it for all
3880 // arguments, seems to work.
3881 while (ArgOffset % 16 !=0) {
3882 ArgOffset += PtrByteSize;
3883 if (GPR_idx != NumGPRs)
3886 // We could elide this store in the case where the object fits
3887 // entirely in R registers. Maybe later.
3888 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3889 DAG.getConstant(ArgOffset, PtrVT));
3890 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3891 MachinePointerInfo(), false, false, 0);
3892 MemOpChains.push_back(Store);
3893 if (VR_idx != NumVRs) {
3894 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3895 MachinePointerInfo(),
3896 false, false, false, 0);
3897 MemOpChains.push_back(Load.getValue(1));
3898 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3901 for (unsigned i=0; i<16; i+=PtrByteSize) {
3902 if (GPR_idx == NumGPRs)
3904 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3905 DAG.getConstant(i, PtrVT));
3906 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3907 false, false, false, 0);
3908 MemOpChains.push_back(Load.getValue(1));
3909 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3914 // Non-varargs Altivec params generally go in registers, but have
3915 // stack space allocated at the end.
3916 if (VR_idx != NumVRs) {
3917 // Doesn't have GPR space allocated.
3918 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3920 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3921 true, isTailCall, true, MemOpChains,
3922 TailCallArguments, dl);
3929 if (!MemOpChains.empty())
3930 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3931 &MemOpChains[0], MemOpChains.size());
3933 // Check if this is an indirect call (MTCTR/BCTRL).
3934 // See PrepareCall() for more information about calls through function
3935 // pointers in the 64-bit SVR4 ABI.
3937 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3938 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3939 !isBLACompatibleAddress(Callee, DAG)) {
3940 // Load r2 into a virtual register and store it to the TOC save area.
3941 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3942 // TOC save area offset.
3943 SDValue PtrOff = DAG.getIntPtrConstant(40);
3944 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3945 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3947 // R12 must contain the address of an indirect callee. This does not
3948 // mean the MTCTR instruction must use R12; it's easier to model this
3949 // as an extra parameter, so do that.
3950 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
3953 // Build a sequence of copy-to-reg nodes chained together with token chain
3954 // and flag operands which copy the outgoing args into the appropriate regs.
3956 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3957 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3958 RegsToPass[i].second, InFlag);
3959 InFlag = Chain.getValue(1);
3963 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
3964 FPOp, true, TailCallArguments);
3966 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3967 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3972 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3973 CallingConv::ID CallConv, bool isVarArg,
3975 const SmallVectorImpl<ISD::OutputArg> &Outs,
3976 const SmallVectorImpl<SDValue> &OutVals,
3977 const SmallVectorImpl<ISD::InputArg> &Ins,
3978 DebugLoc dl, SelectionDAG &DAG,
3979 SmallVectorImpl<SDValue> &InVals) const {
3981 unsigned NumOps = Outs.size();
3983 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3984 bool isPPC64 = PtrVT == MVT::i64;
3985 unsigned PtrByteSize = isPPC64 ? 8 : 4;
3987 MachineFunction &MF = DAG.getMachineFunction();
3989 // Mark this function as potentially containing a function that contains a
3990 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3991 // and restoring the callers stack pointer in this functions epilog. This is
3992 // done because by tail calling the called function might overwrite the value
3993 // in this function's (MF) stack pointer stack slot 0(SP).
3994 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3995 CallConv == CallingConv::Fast)
3996 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3998 unsigned nAltivecParamsAtEnd = 0;
4000 // Count how many bytes are to be pushed on the stack, including the linkage
4001 // area, and parameter passing area. We start with 24/48 bytes, which is
4002 // prereserved space for [SP][CR][LR][3 x unused].
4004 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
4006 nAltivecParamsAtEnd);
4008 // Calculate by how many bytes the stack has to be adjusted in case of tail
4009 // call optimization.
4010 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4012 // To protect arguments on the stack from being clobbered in a tail call,
4013 // force all the loads to happen before doing any other lowering.
4015 Chain = DAG.getStackArgumentTokenFactor(Chain);
4017 // Adjust the stack pointer for the new arguments...
4018 // These operations are automatically eliminated by the prolog/epilog pass
4019 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
4020 SDValue CallSeqStart = Chain;
4022 // Load the return address and frame pointer so it can be move somewhere else
4025 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4028 // Set up a copy of the stack pointer for use loading and storing any
4029 // arguments that may not fit in the registers available for argument
4033 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4035 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4037 // Figure out which arguments are going to go in registers, and which in
4038 // memory. Also, if this is a vararg function, floating point operations
4039 // must be stored to our stack, and loaded into integer regs as well, if
4040 // any integer regs are available for argument passing.
4041 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
4042 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4044 static const uint16_t GPR_32[] = { // 32-bit registers.
4045 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4046 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4048 static const uint16_t GPR_64[] = { // 64-bit registers.
4049 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4050 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4052 static const uint16_t *FPR = GetFPR();
4054 static const uint16_t VR[] = {
4055 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4056 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4058 const unsigned NumGPRs = array_lengthof(GPR_32);
4059 const unsigned NumFPRs = 13;
4060 const unsigned NumVRs = array_lengthof(VR);
4062 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
4064 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4065 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4067 SmallVector<SDValue, 8> MemOpChains;
4068 for (unsigned i = 0; i != NumOps; ++i) {
4069 SDValue Arg = OutVals[i];
4070 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4072 // PtrOff will be used to store the current argument to the stack if a
4073 // register cannot be found for it.
4076 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4078 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4080 // On PPC64, promote integers to 64-bit values.
4081 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4082 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4083 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4084 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4087 // FIXME memcpy is used way more than necessary. Correctness first.
4088 // Note: "by value" is code for passing a structure by value, not
4090 if (Flags.isByVal()) {
4091 unsigned Size = Flags.getByValSize();
4092 // Very small objects are passed right-justified. Everything else is
4093 // passed left-justified.
4094 if (Size==1 || Size==2) {
4095 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4096 if (GPR_idx != NumGPRs) {
4097 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4098 MachinePointerInfo(), VT,
4100 MemOpChains.push_back(Load.getValue(1));
4101 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4103 ArgOffset += PtrByteSize;
4105 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4106 PtrOff.getValueType());
4107 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4108 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4111 ArgOffset += PtrByteSize;
4115 // Copy entire object into memory. There are cases where gcc-generated
4116 // code assumes it is there, even if it could be put entirely into
4117 // registers. (This is not what the doc says.)
4118 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4122 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4123 // copy the pieces of the object that fit into registers from the
4124 // parameter save area.
4125 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4126 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4127 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4128 if (GPR_idx != NumGPRs) {
4129 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4130 MachinePointerInfo(),
4131 false, false, false, 0);
4132 MemOpChains.push_back(Load.getValue(1));
4133 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4134 ArgOffset += PtrByteSize;
4136 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4143 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4144 default: llvm_unreachable("Unexpected ValueType for argument!");
4147 if (GPR_idx != NumGPRs) {
4148 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4150 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4151 isPPC64, isTailCall, false, MemOpChains,
4152 TailCallArguments, dl);
4154 ArgOffset += PtrByteSize;
4158 if (FPR_idx != NumFPRs) {
4159 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4162 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4163 MachinePointerInfo(), false, false, 0);
4164 MemOpChains.push_back(Store);
4166 // Float varargs are always shadowed in available integer registers
4167 if (GPR_idx != NumGPRs) {
4168 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4169 MachinePointerInfo(), false, false,
4171 MemOpChains.push_back(Load.getValue(1));
4172 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4174 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4175 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4176 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4177 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4178 MachinePointerInfo(),
4179 false, false, false, 0);
4180 MemOpChains.push_back(Load.getValue(1));
4181 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4184 // If we have any FPRs remaining, we may also have GPRs remaining.
4185 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4187 if (GPR_idx != NumGPRs)
4189 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4190 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4194 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4195 isPPC64, isTailCall, false, MemOpChains,
4196 TailCallArguments, dl);
4200 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4207 // These go aligned on the stack, or in the corresponding R registers
4208 // when within range. The Darwin PPC ABI doc claims they also go in
4209 // V registers; in fact gcc does this only for arguments that are
4210 // prototyped, not for those that match the ... We do it for all
4211 // arguments, seems to work.
4212 while (ArgOffset % 16 !=0) {
4213 ArgOffset += PtrByteSize;
4214 if (GPR_idx != NumGPRs)
4217 // We could elide this store in the case where the object fits
4218 // entirely in R registers. Maybe later.
4219 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4220 DAG.getConstant(ArgOffset, PtrVT));
4221 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4222 MachinePointerInfo(), false, false, 0);
4223 MemOpChains.push_back(Store);
4224 if (VR_idx != NumVRs) {
4225 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4226 MachinePointerInfo(),
4227 false, false, false, 0);
4228 MemOpChains.push_back(Load.getValue(1));
4229 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4232 for (unsigned i=0; i<16; i+=PtrByteSize) {
4233 if (GPR_idx == NumGPRs)
4235 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4236 DAG.getConstant(i, PtrVT));
4237 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4238 false, false, false, 0);
4239 MemOpChains.push_back(Load.getValue(1));
4240 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4245 // Non-varargs Altivec params generally go in registers, but have
4246 // stack space allocated at the end.
4247 if (VR_idx != NumVRs) {
4248 // Doesn't have GPR space allocated.
4249 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4250 } else if (nAltivecParamsAtEnd==0) {
4251 // We are emitting Altivec params in order.
4252 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4253 isPPC64, isTailCall, true, MemOpChains,
4254 TailCallArguments, dl);
4260 // If all Altivec parameters fit in registers, as they usually do,
4261 // they get stack space following the non-Altivec parameters. We
4262 // don't track this here because nobody below needs it.
4263 // If there are more Altivec parameters than fit in registers emit
4265 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4267 // Offset is aligned; skip 1st 12 params which go in V registers.
4268 ArgOffset = ((ArgOffset+15)/16)*16;
4270 for (unsigned i = 0; i != NumOps; ++i) {
4271 SDValue Arg = OutVals[i];
4272 EVT ArgType = Outs[i].VT;
4273 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4274 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4277 // We are emitting Altivec params in order.
4278 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4279 isPPC64, isTailCall, true, MemOpChains,
4280 TailCallArguments, dl);
4287 if (!MemOpChains.empty())
4288 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4289 &MemOpChains[0], MemOpChains.size());
4291 // On Darwin, R12 must contain the address of an indirect callee. This does
4292 // not mean the MTCTR instruction must use R12; it's easier to model this as
4293 // an extra parameter, so do that.
4295 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4296 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4297 !isBLACompatibleAddress(Callee, DAG))
4298 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4299 PPC::R12), Callee));
4301 // Build a sequence of copy-to-reg nodes chained together with token chain
4302 // and flag operands which copy the outgoing args into the appropriate regs.
4304 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4305 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4306 RegsToPass[i].second, InFlag);
4307 InFlag = Chain.getValue(1);
4311 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4312 FPOp, true, TailCallArguments);
4314 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4315 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4320 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4321 MachineFunction &MF, bool isVarArg,
4322 const SmallVectorImpl<ISD::OutputArg> &Outs,
4323 LLVMContext &Context) const {
4324 SmallVector<CCValAssign, 16> RVLocs;
4325 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4327 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4331 PPCTargetLowering::LowerReturn(SDValue Chain,
4332 CallingConv::ID CallConv, bool isVarArg,
4333 const SmallVectorImpl<ISD::OutputArg> &Outs,
4334 const SmallVectorImpl<SDValue> &OutVals,
4335 DebugLoc dl, SelectionDAG &DAG) const {
4337 SmallVector<CCValAssign, 16> RVLocs;
4338 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4339 getTargetMachine(), RVLocs, *DAG.getContext());
4340 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4342 // If this is the first return lowered for this function, add the regs to the
4343 // liveout set for the function.
4344 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
4345 for (unsigned i = 0; i != RVLocs.size(); ++i)
4346 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
4351 // Copy the result values into the output registers.
4352 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4353 CCValAssign &VA = RVLocs[i];
4354 assert(VA.isRegLoc() && "Can only return in registers!");
4356 SDValue Arg = OutVals[i];
4358 switch (VA.getLocInfo()) {
4359 default: llvm_unreachable("Unknown loc info!");
4360 case CCValAssign::Full: break;
4361 case CCValAssign::AExt:
4362 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
4364 case CCValAssign::ZExt:
4365 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
4367 case CCValAssign::SExt:
4368 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
4372 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
4373 Flag = Chain.getValue(1);
4377 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
4379 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
4382 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4383 const PPCSubtarget &Subtarget) const {
4384 // When we pop the dynamic allocation we need to restore the SP link.
4385 DebugLoc dl = Op.getDebugLoc();
4387 // Get the corect type for pointers.
4388 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4390 // Construct the stack pointer operand.
4391 bool isPPC64 = Subtarget.isPPC64();
4392 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4393 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4395 // Get the operands for the STACKRESTORE.
4396 SDValue Chain = Op.getOperand(0);
4397 SDValue SaveSP = Op.getOperand(1);
4399 // Load the old link SP.
4400 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4401 MachinePointerInfo(),
4402 false, false, false, 0);
4404 // Restore the stack pointer.
4405 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4407 // Store the old link SP.
4408 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4415 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4416 MachineFunction &MF = DAG.getMachineFunction();
4417 bool isPPC64 = PPCSubTarget.isPPC64();
4418 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4419 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4421 // Get current frame pointer save index. The users of this index will be
4422 // primarily DYNALLOC instructions.
4423 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4424 int RASI = FI->getReturnAddrSaveIndex();
4426 // If the frame pointer save index hasn't been defined yet.
4428 // Find out what the fix offset of the frame pointer save area.
4429 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4430 // Allocate the frame index for frame pointer save area.
4431 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4433 FI->setReturnAddrSaveIndex(RASI);
4435 return DAG.getFrameIndex(RASI, PtrVT);
4439 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4440 MachineFunction &MF = DAG.getMachineFunction();
4441 bool isPPC64 = PPCSubTarget.isPPC64();
4442 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4443 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4445 // Get current frame pointer save index. The users of this index will be
4446 // primarily DYNALLOC instructions.
4447 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4448 int FPSI = FI->getFramePointerSaveIndex();
4450 // If the frame pointer save index hasn't been defined yet.
4452 // Find out what the fix offset of the frame pointer save area.
4453 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4456 // Allocate the frame index for frame pointer save area.
4457 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4459 FI->setFramePointerSaveIndex(FPSI);
4461 return DAG.getFrameIndex(FPSI, PtrVT);
4464 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4466 const PPCSubtarget &Subtarget) const {
4468 SDValue Chain = Op.getOperand(0);
4469 SDValue Size = Op.getOperand(1);
4470 DebugLoc dl = Op.getDebugLoc();
4472 // Get the corect type for pointers.
4473 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4475 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4476 DAG.getConstant(0, PtrVT), Size);
4477 // Construct a node for the frame pointer save index.
4478 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4479 // Build a DYNALLOC node.
4480 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4481 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4482 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4485 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4487 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4488 // Not FP? Not a fsel.
4489 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4490 !Op.getOperand(2).getValueType().isFloatingPoint())
4493 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4495 // Cannot handle SETEQ/SETNE.
4496 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4498 EVT ResVT = Op.getValueType();
4499 EVT CmpVT = Op.getOperand(0).getValueType();
4500 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4501 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4502 DebugLoc dl = Op.getDebugLoc();
4504 // If the RHS of the comparison is a 0.0, we don't need to do the
4505 // subtraction at all.
4506 if (isFloatingPointZero(RHS))
4508 default: break; // SETUO etc aren't handled by fsel.
4511 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4514 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4515 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4516 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4519 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4522 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4523 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4524 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4525 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4530 default: break; // SETUO etc aren't handled by fsel.
4533 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4534 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4535 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4536 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4539 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4540 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4541 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4542 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4545 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4546 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4547 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4548 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4551 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4552 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4553 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4554 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4559 // FIXME: Split this code up when LegalizeDAGTypes lands.
4560 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4561 DebugLoc dl) const {
4562 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4563 SDValue Src = Op.getOperand(0);
4564 if (Src.getValueType() == MVT::f32)
4565 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4568 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4569 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4571 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4576 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4580 // Convert the FP value to an int value through memory.
4581 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4583 // Emit a store to the stack slot.
4584 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4585 MachinePointerInfo(), false, false, 0);
4587 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4589 if (Op.getValueType() == MVT::i32)
4590 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4591 DAG.getConstant(4, FIPtr.getValueType()));
4592 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4593 false, false, false, 0);
4596 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4597 SelectionDAG &DAG) const {
4598 DebugLoc dl = Op.getDebugLoc();
4599 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4600 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4603 if (Op.getOperand(0).getValueType() == MVT::i64) {
4604 SDValue SINT = Op.getOperand(0);
4605 // When converting to single-precision, we actually need to convert
4606 // to double-precision first and then round to single-precision.
4607 // To avoid double-rounding effects during that operation, we have
4608 // to prepare the input operand. Bits that might be truncated when
4609 // converting to double-precision are replaced by a bit that won't
4610 // be lost at this stage, but is below the single-precision rounding
4613 // However, if -enable-unsafe-fp-math is in effect, accept double
4614 // rounding to avoid the extra overhead.
4615 if (Op.getValueType() == MVT::f32 &&
4616 !DAG.getTarget().Options.UnsafeFPMath) {
4618 // Twiddle input to make sure the low 11 bits are zero. (If this
4619 // is the case, we are guaranteed the value will fit into the 53 bit
4620 // mantissa of an IEEE double-precision value without rounding.)
4621 // If any of those low 11 bits were not zero originally, make sure
4622 // bit 12 (value 2048) is set instead, so that the final rounding
4623 // to single-precision gets the correct result.
4624 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4625 SINT, DAG.getConstant(2047, MVT::i64));
4626 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4627 Round, DAG.getConstant(2047, MVT::i64));
4628 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4629 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4630 Round, DAG.getConstant(-2048, MVT::i64));
4632 // However, we cannot use that value unconditionally: if the magnitude
4633 // of the input value is small, the bit-twiddling we did above might
4634 // end up visibly changing the output. Fortunately, in that case, we
4635 // don't need to twiddle bits since the original input will convert
4636 // exactly to double-precision floating-point already. Therefore,
4637 // construct a conditional to use the original value if the top 11
4638 // bits are all sign-bit copies, and use the rounded value computed
4640 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4641 SINT, DAG.getConstant(53, MVT::i32));
4642 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4643 Cond, DAG.getConstant(1, MVT::i64));
4644 Cond = DAG.getSetCC(dl, MVT::i32,
4645 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4647 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4649 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4650 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4651 if (Op.getValueType() == MVT::f32)
4652 FP = DAG.getNode(ISD::FP_ROUND, dl,
4653 MVT::f32, FP, DAG.getIntPtrConstant(0));
4657 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4658 "Unhandled SINT_TO_FP type in custom expander!");
4659 // Since we only generate this in 64-bit mode, we can take advantage of
4660 // 64-bit registers. In particular, sign extend the input value into the
4661 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4662 // then lfd it and fcfid it.
4663 MachineFunction &MF = DAG.getMachineFunction();
4664 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4665 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4666 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4667 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4669 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4672 // STD the extended value into the stack slot.
4673 MachineMemOperand *MMO =
4674 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4675 MachineMemOperand::MOStore, 8, 8);
4676 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4678 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4679 Ops, 4, MVT::i64, MMO);
4680 // Load the value as a double.
4681 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4682 false, false, false, 0);
4684 // FCFID it and return it.
4685 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4686 if (Op.getValueType() == MVT::f32)
4687 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4691 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4692 SelectionDAG &DAG) const {
4693 DebugLoc dl = Op.getDebugLoc();
4695 The rounding mode is in bits 30:31 of FPSR, and has the following
4702 FLT_ROUNDS, on the other hand, expects the following:
4709 To perform the conversion, we do:
4710 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4713 MachineFunction &MF = DAG.getMachineFunction();
4714 EVT VT = Op.getValueType();
4715 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4716 std::vector<EVT> NodeTys;
4717 SDValue MFFSreg, InFlag;
4719 // Save FP Control Word to register
4720 NodeTys.push_back(MVT::f64); // return register
4721 NodeTys.push_back(MVT::Glue); // unused in this context
4722 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4724 // Save FP register to stack slot
4725 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4726 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4727 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4728 StackSlot, MachinePointerInfo(), false, false,0);
4730 // Load FP Control Word from low 32 bits of stack slot.
4731 SDValue Four = DAG.getConstant(4, PtrVT);
4732 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4733 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4734 false, false, false, 0);
4736 // Transform as necessary
4738 DAG.getNode(ISD::AND, dl, MVT::i32,
4739 CWD, DAG.getConstant(3, MVT::i32));
4741 DAG.getNode(ISD::SRL, dl, MVT::i32,
4742 DAG.getNode(ISD::AND, dl, MVT::i32,
4743 DAG.getNode(ISD::XOR, dl, MVT::i32,
4744 CWD, DAG.getConstant(3, MVT::i32)),
4745 DAG.getConstant(3, MVT::i32)),
4746 DAG.getConstant(1, MVT::i32));
4749 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4751 return DAG.getNode((VT.getSizeInBits() < 16 ?
4752 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4755 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4756 EVT VT = Op.getValueType();
4757 unsigned BitWidth = VT.getSizeInBits();
4758 DebugLoc dl = Op.getDebugLoc();
4759 assert(Op.getNumOperands() == 3 &&
4760 VT == Op.getOperand(1).getValueType() &&
4763 // Expand into a bunch of logical ops. Note that these ops
4764 // depend on the PPC behavior for oversized shift amounts.
4765 SDValue Lo = Op.getOperand(0);
4766 SDValue Hi = Op.getOperand(1);
4767 SDValue Amt = Op.getOperand(2);
4768 EVT AmtVT = Amt.getValueType();
4770 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4771 DAG.getConstant(BitWidth, AmtVT), Amt);
4772 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4773 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4774 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4775 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4776 DAG.getConstant(-BitWidth, AmtVT));
4777 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4778 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4779 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4780 SDValue OutOps[] = { OutLo, OutHi };
4781 return DAG.getMergeValues(OutOps, 2, dl);
4784 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4785 EVT VT = Op.getValueType();
4786 DebugLoc dl = Op.getDebugLoc();
4787 unsigned BitWidth = VT.getSizeInBits();
4788 assert(Op.getNumOperands() == 3 &&
4789 VT == Op.getOperand(1).getValueType() &&
4792 // Expand into a bunch of logical ops. Note that these ops
4793 // depend on the PPC behavior for oversized shift amounts.
4794 SDValue Lo = Op.getOperand(0);
4795 SDValue Hi = Op.getOperand(1);
4796 SDValue Amt = Op.getOperand(2);
4797 EVT AmtVT = Amt.getValueType();
4799 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4800 DAG.getConstant(BitWidth, AmtVT), Amt);
4801 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4802 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4803 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4804 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4805 DAG.getConstant(-BitWidth, AmtVT));
4806 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4807 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4808 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4809 SDValue OutOps[] = { OutLo, OutHi };
4810 return DAG.getMergeValues(OutOps, 2, dl);
4813 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4814 DebugLoc dl = Op.getDebugLoc();
4815 EVT VT = Op.getValueType();
4816 unsigned BitWidth = VT.getSizeInBits();
4817 assert(Op.getNumOperands() == 3 &&
4818 VT == Op.getOperand(1).getValueType() &&
4821 // Expand into a bunch of logical ops, followed by a select_cc.
4822 SDValue Lo = Op.getOperand(0);
4823 SDValue Hi = Op.getOperand(1);
4824 SDValue Amt = Op.getOperand(2);
4825 EVT AmtVT = Amt.getValueType();
4827 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4828 DAG.getConstant(BitWidth, AmtVT), Amt);
4829 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4830 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4831 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4832 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4833 DAG.getConstant(-BitWidth, AmtVT));
4834 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4835 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4836 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4837 Tmp4, Tmp6, ISD::SETLE);
4838 SDValue OutOps[] = { OutLo, OutHi };
4839 return DAG.getMergeValues(OutOps, 2, dl);
4842 //===----------------------------------------------------------------------===//
4843 // Vector related lowering.
4846 /// BuildSplatI - Build a canonical splati of Val with an element size of
4847 /// SplatSize. Cast the result to VT.
4848 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4849 SelectionDAG &DAG, DebugLoc dl) {
4850 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4852 static const EVT VTys[] = { // canonical VT to use for each size.
4853 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4856 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4858 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4862 EVT CanonicalVT = VTys[SplatSize-1];
4864 // Build a canonical splat for this value.
4865 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4866 SmallVector<SDValue, 8> Ops;
4867 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4868 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4869 &Ops[0], Ops.size());
4870 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4873 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4874 /// specified intrinsic ID.
4875 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4876 SelectionDAG &DAG, DebugLoc dl,
4877 EVT DestVT = MVT::Other) {
4878 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4879 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4880 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4883 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4884 /// specified intrinsic ID.
4885 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4886 SDValue Op2, SelectionDAG &DAG,
4887 DebugLoc dl, EVT DestVT = MVT::Other) {
4888 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4889 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4890 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4894 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4895 /// amount. The result has the specified value type.
4896 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4897 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4898 // Force LHS/RHS to be the right type.
4899 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4900 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4903 for (unsigned i = 0; i != 16; ++i)
4905 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4906 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4909 // If this is a case we can't handle, return null and let the default
4910 // expansion code take care of it. If we CAN select this case, and if it
4911 // selects to a single instruction, return Op. Otherwise, if we can codegen
4912 // this case more efficiently than a constant pool load, lower it to the
4913 // sequence of ops that should be used.
4914 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4915 SelectionDAG &DAG) const {
4916 DebugLoc dl = Op.getDebugLoc();
4917 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4918 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4920 // Check if this is a splat of a constant value.
4921 APInt APSplatBits, APSplatUndef;
4922 unsigned SplatBitSize;
4924 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4925 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4928 unsigned SplatBits = APSplatBits.getZExtValue();
4929 unsigned SplatUndef = APSplatUndef.getZExtValue();
4930 unsigned SplatSize = SplatBitSize / 8;
4932 // First, handle single instruction cases.
4935 if (SplatBits == 0) {
4936 // Canonicalize all zero vectors to be v4i32.
4937 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4938 SDValue Z = DAG.getConstant(0, MVT::i32);
4939 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4940 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4945 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4946 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4948 if (SextVal >= -16 && SextVal <= 15)
4949 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4952 // Two instruction sequences.
4954 // If this value is in the range [-32,30] and is even, use:
4955 // tmp = VSPLTI[bhw], result = add tmp, tmp
4956 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4957 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4958 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4959 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4962 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4963 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4965 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4966 // Make -1 and vspltisw -1:
4967 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4969 // Make the VSLW intrinsic, computing 0x8000_0000.
4970 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4973 // xor by OnesV to invert it.
4974 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4975 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4978 // Check to see if this is a wide variety of vsplti*, binop self cases.
4979 static const signed char SplatCsts[] = {
4980 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4981 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
4984 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
4985 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
4986 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
4987 int i = SplatCsts[idx];
4989 // Figure out what shift amount will be used by altivec if shifted by i in
4991 unsigned TypeShiftAmt = i & (SplatBitSize-1);
4993 // vsplti + shl self.
4994 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
4995 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
4996 static const unsigned IIDs[] = { // Intrinsic to use for each size.
4997 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
4998 Intrinsic::ppc_altivec_vslw
5000 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5001 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5004 // vsplti + srl self.
5005 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5006 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5007 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5008 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
5009 Intrinsic::ppc_altivec_vsrw
5011 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5012 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5015 // vsplti + sra self.
5016 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5017 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5018 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5019 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
5020 Intrinsic::ppc_altivec_vsraw
5022 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5023 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5026 // vsplti + rol self.
5027 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
5028 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
5029 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5030 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5031 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
5032 Intrinsic::ppc_altivec_vrlw
5034 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5035 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5038 // t = vsplti c, result = vsldoi t, t, 1
5039 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
5040 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5041 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
5043 // t = vsplti c, result = vsldoi t, t, 2
5044 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
5045 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5046 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
5048 // t = vsplti c, result = vsldoi t, t, 3
5049 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
5050 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5051 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
5055 // Three instruction sequences.
5057 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
5058 if (SextVal >= 0 && SextVal <= 31) {
5059 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
5060 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
5061 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
5062 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
5064 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
5065 if (SextVal >= -31 && SextVal <= 0) {
5066 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
5067 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
5068 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
5069 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
5075 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5076 /// the specified operations to build the shuffle.
5077 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5078 SDValue RHS, SelectionDAG &DAG,
5080 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5081 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5082 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5085 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5097 if (OpNum == OP_COPY) {
5098 if (LHSID == (1*9+2)*9+3) return LHS;
5099 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5103 SDValue OpLHS, OpRHS;
5104 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5105 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5109 default: llvm_unreachable("Unknown i32 permute!");
5111 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5112 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5113 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5114 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5117 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5118 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5119 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5120 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5123 for (unsigned i = 0; i != 16; ++i)
5124 ShufIdxs[i] = (i&3)+0;
5127 for (unsigned i = 0; i != 16; ++i)
5128 ShufIdxs[i] = (i&3)+4;
5131 for (unsigned i = 0; i != 16; ++i)
5132 ShufIdxs[i] = (i&3)+8;
5135 for (unsigned i = 0; i != 16; ++i)
5136 ShufIdxs[i] = (i&3)+12;
5139 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5141 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5143 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5145 EVT VT = OpLHS.getValueType();
5146 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5147 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5148 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5149 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5152 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5153 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5154 /// return the code it can be lowered into. Worst case, it can always be
5155 /// lowered into a vperm.
5156 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5157 SelectionDAG &DAG) const {
5158 DebugLoc dl = Op.getDebugLoc();
5159 SDValue V1 = Op.getOperand(0);
5160 SDValue V2 = Op.getOperand(1);
5161 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5162 EVT VT = Op.getValueType();
5164 // Cases that are handled by instructions that take permute immediates
5165 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5166 // selected by the instruction selector.
5167 if (V2.getOpcode() == ISD::UNDEF) {
5168 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5169 PPC::isSplatShuffleMask(SVOp, 2) ||
5170 PPC::isSplatShuffleMask(SVOp, 4) ||
5171 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5172 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5173 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5174 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5175 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5176 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5177 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5178 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5179 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5184 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5185 // and produce a fixed permutation. If any of these match, do not lower to
5187 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5188 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5189 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5190 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5191 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5192 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5193 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5194 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5195 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5198 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5199 // perfect shuffle table to emit an optimal matching sequence.
5200 ArrayRef<int> PermMask = SVOp->getMask();
5202 unsigned PFIndexes[4];
5203 bool isFourElementShuffle = true;
5204 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5205 unsigned EltNo = 8; // Start out undef.
5206 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5207 if (PermMask[i*4+j] < 0)
5208 continue; // Undef, ignore it.
5210 unsigned ByteSource = PermMask[i*4+j];
5211 if ((ByteSource & 3) != j) {
5212 isFourElementShuffle = false;
5217 EltNo = ByteSource/4;
5218 } else if (EltNo != ByteSource/4) {
5219 isFourElementShuffle = false;
5223 PFIndexes[i] = EltNo;
5226 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5227 // perfect shuffle vector to determine if it is cost effective to do this as
5228 // discrete instructions, or whether we should use a vperm.
5229 if (isFourElementShuffle) {
5230 // Compute the index in the perfect shuffle table.
5231 unsigned PFTableIndex =
5232 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5234 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5235 unsigned Cost = (PFEntry >> 30);
5237 // Determining when to avoid vperm is tricky. Many things affect the cost
5238 // of vperm, particularly how many times the perm mask needs to be computed.
5239 // For example, if the perm mask can be hoisted out of a loop or is already
5240 // used (perhaps because there are multiple permutes with the same shuffle
5241 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5242 // the loop requires an extra register.
5244 // As a compromise, we only emit discrete instructions if the shuffle can be
5245 // generated in 3 or fewer operations. When we have loop information
5246 // available, if this block is within a loop, we should avoid using vperm
5247 // for 3-operation perms and use a constant pool load instead.
5249 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5252 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5253 // vector that will get spilled to the constant pool.
5254 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5256 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5257 // that it is in input element units, not in bytes. Convert now.
5258 EVT EltVT = V1.getValueType().getVectorElementType();
5259 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5261 SmallVector<SDValue, 16> ResultMask;
5262 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5263 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5265 for (unsigned j = 0; j != BytesPerElement; ++j)
5266 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5270 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5271 &ResultMask[0], ResultMask.size());
5272 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5275 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5276 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5277 /// information about the intrinsic.
5278 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5280 unsigned IntrinsicID =
5281 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5284 switch (IntrinsicID) {
5285 default: return false;
5286 // Comparison predicates.
5287 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5288 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5289 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5290 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5291 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5292 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5293 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5294 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5295 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5296 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5297 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5298 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5299 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5301 // Normal Comparisons.
5302 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5303 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5304 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5305 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5306 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5307 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5308 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5309 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5310 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5311 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5312 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5313 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5314 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5319 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5320 /// lower, do it, otherwise return null.
5321 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5322 SelectionDAG &DAG) const {
5323 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5324 // opcode number of the comparison.
5325 DebugLoc dl = Op.getDebugLoc();
5328 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5329 return SDValue(); // Don't custom lower most intrinsics.
5331 // If this is a non-dot comparison, make the VCMP node and we are done.
5333 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5334 Op.getOperand(1), Op.getOperand(2),
5335 DAG.getConstant(CompareOpc, MVT::i32));
5336 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5339 // Create the PPCISD altivec 'dot' comparison node.
5341 Op.getOperand(2), // LHS
5342 Op.getOperand(3), // RHS
5343 DAG.getConstant(CompareOpc, MVT::i32)
5345 std::vector<EVT> VTs;
5346 VTs.push_back(Op.getOperand(2).getValueType());
5347 VTs.push_back(MVT::Glue);
5348 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5350 // Now that we have the comparison, emit a copy from the CR to a GPR.
5351 // This is flagged to the above dot comparison.
5352 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5353 DAG.getRegister(PPC::CR6, MVT::i32),
5354 CompNode.getValue(1));
5356 // Unpack the result based on how the target uses it.
5357 unsigned BitNo; // Bit # of CR6.
5358 bool InvertBit; // Invert result?
5359 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5360 default: // Can't happen, don't crash on invalid number though.
5361 case 0: // Return the value of the EQ bit of CR6.
5362 BitNo = 0; InvertBit = false;
5364 case 1: // Return the inverted value of the EQ bit of CR6.
5365 BitNo = 0; InvertBit = true;
5367 case 2: // Return the value of the LT bit of CR6.
5368 BitNo = 2; InvertBit = false;
5370 case 3: // Return the inverted value of the LT bit of CR6.
5371 BitNo = 2; InvertBit = true;
5375 // Shift the bit into the low position.
5376 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5377 DAG.getConstant(8-(3-BitNo), MVT::i32));
5379 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5380 DAG.getConstant(1, MVT::i32));
5382 // If we are supposed to, toggle the bit.
5384 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5385 DAG.getConstant(1, MVT::i32));
5389 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5390 SelectionDAG &DAG) const {
5391 DebugLoc dl = Op.getDebugLoc();
5392 // Create a stack slot that is 16-byte aligned.
5393 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5394 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5395 EVT PtrVT = getPointerTy();
5396 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5398 // Store the input value into Value#0 of the stack slot.
5399 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5400 Op.getOperand(0), FIdx, MachinePointerInfo(),
5403 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5404 false, false, false, 0);
5407 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5408 DebugLoc dl = Op.getDebugLoc();
5409 if (Op.getValueType() == MVT::v4i32) {
5410 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5412 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5413 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5415 SDValue RHSSwap = // = vrlw RHS, 16
5416 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5418 // Shrinkify inputs to v8i16.
5419 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5420 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5421 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5423 // Low parts multiplied together, generating 32-bit results (we ignore the
5425 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5426 LHS, RHS, DAG, dl, MVT::v4i32);
5428 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5429 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5430 // Shift the high parts up 16 bits.
5431 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5433 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5434 } else if (Op.getValueType() == MVT::v8i16) {
5435 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5437 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5439 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5440 LHS, RHS, Zero, DAG, dl);
5441 } else if (Op.getValueType() == MVT::v16i8) {
5442 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5444 // Multiply the even 8-bit parts, producing 16-bit sums.
5445 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5446 LHS, RHS, DAG, dl, MVT::v8i16);
5447 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5449 // Multiply the odd 8-bit parts, producing 16-bit sums.
5450 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5451 LHS, RHS, DAG, dl, MVT::v8i16);
5452 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5454 // Merge the results together.
5456 for (unsigned i = 0; i != 8; ++i) {
5458 Ops[i*2+1] = 2*i+1+16;
5460 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5462 llvm_unreachable("Unknown mul to lower!");
5466 /// LowerOperation - Provide custom lowering hooks for some operations.
5468 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5469 switch (Op.getOpcode()) {
5470 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5471 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5472 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5473 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5474 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5475 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5476 case ISD::SETCC: return LowerSETCC(Op, DAG);
5477 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5478 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5480 return LowerVASTART(Op, DAG, PPCSubTarget);
5483 return LowerVAARG(Op, DAG, PPCSubTarget);
5485 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5486 case ISD::DYNAMIC_STACKALLOC:
5487 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5489 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5490 case ISD::FP_TO_UINT:
5491 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5493 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5494 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5496 // Lower 64-bit shifts.
5497 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5498 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5499 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5501 // Vector-related lowering.
5502 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5503 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5504 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5505 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5506 case ISD::MUL: return LowerMUL(Op, DAG);
5508 // Frame & Return address.
5509 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5510 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5514 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5515 SmallVectorImpl<SDValue>&Results,
5516 SelectionDAG &DAG) const {
5517 const TargetMachine &TM = getTargetMachine();
5518 DebugLoc dl = N->getDebugLoc();
5519 switch (N->getOpcode()) {
5521 llvm_unreachable("Do not know how to custom type legalize this operation!");
5523 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5524 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5527 EVT VT = N->getValueType(0);
5529 if (VT == MVT::i64) {
5530 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5532 Results.push_back(NewNode);
5533 Results.push_back(NewNode.getValue(1));
5537 case ISD::FP_ROUND_INREG: {
5538 assert(N->getValueType(0) == MVT::ppcf128);
5539 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5540 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5541 MVT::f64, N->getOperand(0),
5542 DAG.getIntPtrConstant(0));
5543 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5544 MVT::f64, N->getOperand(0),
5545 DAG.getIntPtrConstant(1));
5547 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5548 // of the long double, and puts FPSCR back the way it was. We do not
5549 // actually model FPSCR.
5550 std::vector<EVT> NodeTys;
5551 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5553 NodeTys.push_back(MVT::f64); // Return register
5554 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5555 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5556 MFFSreg = Result.getValue(0);
5557 InFlag = Result.getValue(1);
5560 NodeTys.push_back(MVT::Glue); // Returns a flag
5561 Ops[0] = DAG.getConstant(31, MVT::i32);
5563 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5564 InFlag = Result.getValue(0);
5567 NodeTys.push_back(MVT::Glue); // Returns a flag
5568 Ops[0] = DAG.getConstant(30, MVT::i32);
5570 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5571 InFlag = Result.getValue(0);
5574 NodeTys.push_back(MVT::f64); // result of add
5575 NodeTys.push_back(MVT::Glue); // Returns a flag
5579 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5580 FPreg = Result.getValue(0);
5581 InFlag = Result.getValue(1);
5584 NodeTys.push_back(MVT::f64);
5585 Ops[0] = DAG.getConstant(1, MVT::i32);
5589 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5590 FPreg = Result.getValue(0);
5592 // We know the low half is about to be thrown away, so just use something
5594 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5598 case ISD::FP_TO_SINT:
5599 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5605 //===----------------------------------------------------------------------===//
5606 // Other Lowering Code
5607 //===----------------------------------------------------------------------===//
5610 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5611 bool is64bit, unsigned BinOpcode) const {
5612 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5613 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5615 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5616 MachineFunction *F = BB->getParent();
5617 MachineFunction::iterator It = BB;
5620 unsigned dest = MI->getOperand(0).getReg();
5621 unsigned ptrA = MI->getOperand(1).getReg();
5622 unsigned ptrB = MI->getOperand(2).getReg();
5623 unsigned incr = MI->getOperand(3).getReg();
5624 DebugLoc dl = MI->getDebugLoc();
5626 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5627 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5628 F->insert(It, loopMBB);
5629 F->insert(It, exitMBB);
5630 exitMBB->splice(exitMBB->begin(), BB,
5631 llvm::next(MachineBasicBlock::iterator(MI)),
5633 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5635 MachineRegisterInfo &RegInfo = F->getRegInfo();
5636 unsigned TmpReg = (!BinOpcode) ? incr :
5637 RegInfo.createVirtualRegister(
5638 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5639 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5643 // fallthrough --> loopMBB
5644 BB->addSuccessor(loopMBB);
5647 // l[wd]arx dest, ptr
5648 // add r0, dest, incr
5649 // st[wd]cx. r0, ptr
5651 // fallthrough --> exitMBB
5653 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5654 .addReg(ptrA).addReg(ptrB);
5656 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5657 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5658 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5659 BuildMI(BB, dl, TII->get(PPC::BCC))
5660 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5661 BB->addSuccessor(loopMBB);
5662 BB->addSuccessor(exitMBB);
5671 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5672 MachineBasicBlock *BB,
5673 bool is8bit, // operation
5674 unsigned BinOpcode) const {
5675 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5676 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5677 // In 64 bit mode we have to use 64 bits for addresses, even though the
5678 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5679 // registers without caring whether they're 32 or 64, but here we're
5680 // doing actual arithmetic on the addresses.
5681 bool is64bit = PPCSubTarget.isPPC64();
5682 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5684 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5685 MachineFunction *F = BB->getParent();
5686 MachineFunction::iterator It = BB;
5689 unsigned dest = MI->getOperand(0).getReg();
5690 unsigned ptrA = MI->getOperand(1).getReg();
5691 unsigned ptrB = MI->getOperand(2).getReg();
5692 unsigned incr = MI->getOperand(3).getReg();
5693 DebugLoc dl = MI->getDebugLoc();
5695 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5696 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5697 F->insert(It, loopMBB);
5698 F->insert(It, exitMBB);
5699 exitMBB->splice(exitMBB->begin(), BB,
5700 llvm::next(MachineBasicBlock::iterator(MI)),
5702 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5704 MachineRegisterInfo &RegInfo = F->getRegInfo();
5705 const TargetRegisterClass *RC =
5706 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5707 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5708 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5709 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5710 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5711 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5712 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5713 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5714 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5715 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5716 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5717 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5718 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5720 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5724 // fallthrough --> loopMBB
5725 BB->addSuccessor(loopMBB);
5727 // The 4-byte load must be aligned, while a char or short may be
5728 // anywhere in the word. Hence all this nasty bookkeeping code.
5729 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5730 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5731 // xori shift, shift1, 24 [16]
5732 // rlwinm ptr, ptr1, 0, 0, 29
5733 // slw incr2, incr, shift
5734 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5735 // slw mask, mask2, shift
5737 // lwarx tmpDest, ptr
5738 // add tmp, tmpDest, incr2
5739 // andc tmp2, tmpDest, mask
5740 // and tmp3, tmp, mask
5741 // or tmp4, tmp3, tmp2
5744 // fallthrough --> exitMBB
5745 // srw dest, tmpDest, shift
5746 if (ptrA != ZeroReg) {
5747 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5748 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5749 .addReg(ptrA).addReg(ptrB);
5753 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5754 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5755 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5756 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5758 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5759 .addReg(Ptr1Reg).addImm(0).addImm(61);
5761 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5762 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5763 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5764 .addReg(incr).addReg(ShiftReg);
5766 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5768 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5769 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5771 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5772 .addReg(Mask2Reg).addReg(ShiftReg);
5775 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5776 .addReg(ZeroReg).addReg(PtrReg);
5778 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5779 .addReg(Incr2Reg).addReg(TmpDestReg);
5780 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5781 .addReg(TmpDestReg).addReg(MaskReg);
5782 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5783 .addReg(TmpReg).addReg(MaskReg);
5784 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5785 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5786 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5787 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5788 BuildMI(BB, dl, TII->get(PPC::BCC))
5789 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5790 BB->addSuccessor(loopMBB);
5791 BB->addSuccessor(exitMBB);
5796 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5802 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5803 MachineBasicBlock *BB) const {
5804 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5806 // To "insert" these instructions we actually have to insert their
5807 // control-flow patterns.
5808 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5809 MachineFunction::iterator It = BB;
5812 MachineFunction *F = BB->getParent();
5814 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5815 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5816 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5817 PPC::ISEL8 : PPC::ISEL;
5818 unsigned SelectPred = MI->getOperand(4).getImm();
5819 DebugLoc dl = MI->getDebugLoc();
5821 // The SelectPred is ((BI << 5) | BO) for a BCC
5822 unsigned BO = SelectPred & 0xF;
5823 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5825 unsigned TrueOpNo, FalseOpNo;
5832 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5835 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5836 .addReg(MI->getOperand(TrueOpNo).getReg())
5837 .addReg(MI->getOperand(FalseOpNo).getReg())
5838 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5839 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5840 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5841 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5842 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5843 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5846 // The incoming instruction knows the destination vreg to set, the
5847 // condition code register to branch on, the true/false values to
5848 // select between, and a branch opcode to use.
5853 // cmpTY ccX, r1, r2
5855 // fallthrough --> copy0MBB
5856 MachineBasicBlock *thisMBB = BB;
5857 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5858 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5859 unsigned SelectPred = MI->getOperand(4).getImm();
5860 DebugLoc dl = MI->getDebugLoc();
5861 F->insert(It, copy0MBB);
5862 F->insert(It, sinkMBB);
5864 // Transfer the remainder of BB and its successor edges to sinkMBB.
5865 sinkMBB->splice(sinkMBB->begin(), BB,
5866 llvm::next(MachineBasicBlock::iterator(MI)),
5868 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5870 // Next, add the true and fallthrough blocks as its successors.
5871 BB->addSuccessor(copy0MBB);
5872 BB->addSuccessor(sinkMBB);
5874 BuildMI(BB, dl, TII->get(PPC::BCC))
5875 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5878 // %FalseValue = ...
5879 // # fallthrough to sinkMBB
5882 // Update machine-CFG edges
5883 BB->addSuccessor(sinkMBB);
5886 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5889 BuildMI(*BB, BB->begin(), dl,
5890 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5891 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5892 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5894 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5895 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5896 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5897 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5898 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5899 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5900 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5901 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5903 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5904 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5905 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5906 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5907 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5908 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5909 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5910 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5912 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5913 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5914 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5915 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5916 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5917 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5918 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5919 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5921 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5922 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5923 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5924 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5925 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5926 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5927 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5928 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5930 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5931 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5932 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5933 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5934 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5935 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5936 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5937 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5939 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5940 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5941 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5942 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5943 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5944 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5945 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5946 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5948 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5949 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5950 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5951 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5952 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5953 BB = EmitAtomicBinary(MI, BB, false, 0);
5954 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5955 BB = EmitAtomicBinary(MI, BB, true, 0);
5957 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5958 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5959 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5961 unsigned dest = MI->getOperand(0).getReg();
5962 unsigned ptrA = MI->getOperand(1).getReg();
5963 unsigned ptrB = MI->getOperand(2).getReg();
5964 unsigned oldval = MI->getOperand(3).getReg();
5965 unsigned newval = MI->getOperand(4).getReg();
5966 DebugLoc dl = MI->getDebugLoc();
5968 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5969 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5970 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5971 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5972 F->insert(It, loop1MBB);
5973 F->insert(It, loop2MBB);
5974 F->insert(It, midMBB);
5975 F->insert(It, exitMBB);
5976 exitMBB->splice(exitMBB->begin(), BB,
5977 llvm::next(MachineBasicBlock::iterator(MI)),
5979 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5983 // fallthrough --> loopMBB
5984 BB->addSuccessor(loop1MBB);
5987 // l[wd]arx dest, ptr
5988 // cmp[wd] dest, oldval
5991 // st[wd]cx. newval, ptr
5995 // st[wd]cx. dest, ptr
5998 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5999 .addReg(ptrA).addReg(ptrB);
6000 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
6001 .addReg(oldval).addReg(dest);
6002 BuildMI(BB, dl, TII->get(PPC::BCC))
6003 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6004 BB->addSuccessor(loop2MBB);
6005 BB->addSuccessor(midMBB);
6008 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6009 .addReg(newval).addReg(ptrA).addReg(ptrB);
6010 BuildMI(BB, dl, TII->get(PPC::BCC))
6011 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6012 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6013 BB->addSuccessor(loop1MBB);
6014 BB->addSuccessor(exitMBB);
6017 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6018 .addReg(dest).addReg(ptrA).addReg(ptrB);
6019 BB->addSuccessor(exitMBB);
6024 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
6025 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
6026 // We must use 64-bit registers for addresses when targeting 64-bit,
6027 // since we're actually doing arithmetic on them. Other registers
6029 bool is64bit = PPCSubTarget.isPPC64();
6030 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
6032 unsigned dest = MI->getOperand(0).getReg();
6033 unsigned ptrA = MI->getOperand(1).getReg();
6034 unsigned ptrB = MI->getOperand(2).getReg();
6035 unsigned oldval = MI->getOperand(3).getReg();
6036 unsigned newval = MI->getOperand(4).getReg();
6037 DebugLoc dl = MI->getDebugLoc();
6039 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6040 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6041 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6042 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6043 F->insert(It, loop1MBB);
6044 F->insert(It, loop2MBB);
6045 F->insert(It, midMBB);
6046 F->insert(It, exitMBB);
6047 exitMBB->splice(exitMBB->begin(), BB,
6048 llvm::next(MachineBasicBlock::iterator(MI)),
6050 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6052 MachineRegisterInfo &RegInfo = F->getRegInfo();
6053 const TargetRegisterClass *RC =
6054 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
6055 (const TargetRegisterClass *) &PPC::GPRCRegClass;
6056 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6057 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6058 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6059 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
6060 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
6061 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
6062 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
6063 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6064 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6065 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6066 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6067 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6068 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6070 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
6071 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
6074 // fallthrough --> loopMBB
6075 BB->addSuccessor(loop1MBB);
6077 // The 4-byte load must be aligned, while a char or short may be
6078 // anywhere in the word. Hence all this nasty bookkeeping code.
6079 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6080 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6081 // xori shift, shift1, 24 [16]
6082 // rlwinm ptr, ptr1, 0, 0, 29
6083 // slw newval2, newval, shift
6084 // slw oldval2, oldval,shift
6085 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6086 // slw mask, mask2, shift
6087 // and newval3, newval2, mask
6088 // and oldval3, oldval2, mask
6090 // lwarx tmpDest, ptr
6091 // and tmp, tmpDest, mask
6092 // cmpw tmp, oldval3
6095 // andc tmp2, tmpDest, mask
6096 // or tmp4, tmp2, newval3
6101 // stwcx. tmpDest, ptr
6103 // srw dest, tmpDest, shift
6104 if (ptrA != ZeroReg) {
6105 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6106 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6107 .addReg(ptrA).addReg(ptrB);
6111 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6112 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6113 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6114 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6116 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6117 .addReg(Ptr1Reg).addImm(0).addImm(61);
6119 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6120 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6121 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6122 .addReg(newval).addReg(ShiftReg);
6123 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6124 .addReg(oldval).addReg(ShiftReg);
6126 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6128 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6129 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6130 .addReg(Mask3Reg).addImm(65535);
6132 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6133 .addReg(Mask2Reg).addReg(ShiftReg);
6134 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6135 .addReg(NewVal2Reg).addReg(MaskReg);
6136 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6137 .addReg(OldVal2Reg).addReg(MaskReg);
6140 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6141 .addReg(ZeroReg).addReg(PtrReg);
6142 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6143 .addReg(TmpDestReg).addReg(MaskReg);
6144 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6145 .addReg(TmpReg).addReg(OldVal3Reg);
6146 BuildMI(BB, dl, TII->get(PPC::BCC))
6147 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6148 BB->addSuccessor(loop2MBB);
6149 BB->addSuccessor(midMBB);
6152 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6153 .addReg(TmpDestReg).addReg(MaskReg);
6154 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6155 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6156 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6157 .addReg(ZeroReg).addReg(PtrReg);
6158 BuildMI(BB, dl, TII->get(PPC::BCC))
6159 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6160 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6161 BB->addSuccessor(loop1MBB);
6162 BB->addSuccessor(exitMBB);
6165 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6166 .addReg(ZeroReg).addReg(PtrReg);
6167 BB->addSuccessor(exitMBB);
6172 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6175 llvm_unreachable("Unexpected instr type to insert");
6178 MI->eraseFromParent(); // The pseudo instruction is gone now.
6182 //===----------------------------------------------------------------------===//
6183 // Target Optimization Hooks
6184 //===----------------------------------------------------------------------===//
6186 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6187 DAGCombinerInfo &DCI) const {
6188 const TargetMachine &TM = getTargetMachine();
6189 SelectionDAG &DAG = DCI.DAG;
6190 DebugLoc dl = N->getDebugLoc();
6191 switch (N->getOpcode()) {
6194 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6195 if (C->isNullValue()) // 0 << V -> 0.
6196 return N->getOperand(0);
6200 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6201 if (C->isNullValue()) // 0 >>u V -> 0.
6202 return N->getOperand(0);
6206 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6207 if (C->isNullValue() || // 0 >>s V -> 0.
6208 C->isAllOnesValue()) // -1 >>s V -> -1.
6209 return N->getOperand(0);
6213 case ISD::SINT_TO_FP:
6214 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6215 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6216 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6217 // We allow the src/dst to be either f32/f64, but the intermediate
6218 // type must be i64.
6219 if (N->getOperand(0).getValueType() == MVT::i64 &&
6220 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6221 SDValue Val = N->getOperand(0).getOperand(0);
6222 if (Val.getValueType() == MVT::f32) {
6223 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6224 DCI.AddToWorklist(Val.getNode());
6227 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6228 DCI.AddToWorklist(Val.getNode());
6229 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6230 DCI.AddToWorklist(Val.getNode());
6231 if (N->getValueType(0) == MVT::f32) {
6232 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6233 DAG.getIntPtrConstant(0));
6234 DCI.AddToWorklist(Val.getNode());
6237 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6238 // If the intermediate type is i32, we can avoid the load/store here
6245 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6246 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6247 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6248 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6249 N->getOperand(1).getValueType() == MVT::i32 &&
6250 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6251 SDValue Val = N->getOperand(1).getOperand(0);
6252 if (Val.getValueType() == MVT::f32) {
6253 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6254 DCI.AddToWorklist(Val.getNode());
6256 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6257 DCI.AddToWorklist(Val.getNode());
6259 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6260 N->getOperand(2), N->getOperand(3));
6261 DCI.AddToWorklist(Val.getNode());
6265 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6266 if (cast<StoreSDNode>(N)->isUnindexed() &&
6267 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6268 N->getOperand(1).getNode()->hasOneUse() &&
6269 (N->getOperand(1).getValueType() == MVT::i32 ||
6270 N->getOperand(1).getValueType() == MVT::i16)) {
6271 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6272 // Do an any-extend to 32-bits if this is a half-word input.
6273 if (BSwapOp.getValueType() == MVT::i16)
6274 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6277 N->getOperand(0), BSwapOp, N->getOperand(2),
6278 DAG.getValueType(N->getOperand(1).getValueType())
6281 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6282 Ops, array_lengthof(Ops),
6283 cast<StoreSDNode>(N)->getMemoryVT(),
6284 cast<StoreSDNode>(N)->getMemOperand());
6288 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6289 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6290 N->getOperand(0).hasOneUse() &&
6291 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6292 SDValue Load = N->getOperand(0);
6293 LoadSDNode *LD = cast<LoadSDNode>(Load);
6294 // Create the byte-swapping load.
6296 LD->getChain(), // Chain
6297 LD->getBasePtr(), // Ptr
6298 DAG.getValueType(N->getValueType(0)) // VT
6301 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6302 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6303 LD->getMemoryVT(), LD->getMemOperand());
6305 // If this is an i16 load, insert the truncate.
6306 SDValue ResVal = BSLoad;
6307 if (N->getValueType(0) == MVT::i16)
6308 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6310 // First, combine the bswap away. This makes the value produced by the
6312 DCI.CombineTo(N, ResVal);
6314 // Next, combine the load away, we give it a bogus result value but a real
6315 // chain result. The result value is dead because the bswap is dead.
6316 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6318 // Return N so it doesn't get rechecked!
6319 return SDValue(N, 0);
6323 case PPCISD::VCMP: {
6324 // If a VCMPo node already exists with exactly the same operands as this
6325 // node, use its result instead of this node (VCMPo computes both a CR6 and
6326 // a normal output).
6328 if (!N->getOperand(0).hasOneUse() &&
6329 !N->getOperand(1).hasOneUse() &&
6330 !N->getOperand(2).hasOneUse()) {
6332 // Scan all of the users of the LHS, looking for VCMPo's that match.
6333 SDNode *VCMPoNode = 0;
6335 SDNode *LHSN = N->getOperand(0).getNode();
6336 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6338 if (UI->getOpcode() == PPCISD::VCMPo &&
6339 UI->getOperand(1) == N->getOperand(1) &&
6340 UI->getOperand(2) == N->getOperand(2) &&
6341 UI->getOperand(0) == N->getOperand(0)) {
6346 // If there is no VCMPo node, or if the flag value has a single use, don't
6348 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6351 // Look at the (necessarily single) use of the flag value. If it has a
6352 // chain, this transformation is more complex. Note that multiple things
6353 // could use the value result, which we should ignore.
6354 SDNode *FlagUser = 0;
6355 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6356 FlagUser == 0; ++UI) {
6357 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6359 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6360 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6367 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6368 // give up for right now.
6369 if (FlagUser->getOpcode() == PPCISD::MFCR)
6370 return SDValue(VCMPoNode, 0);
6375 // If this is a branch on an altivec predicate comparison, lower this so
6376 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6377 // lowering is done pre-legalize, because the legalizer lowers the predicate
6378 // compare down to code that is difficult to reassemble.
6379 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6380 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6384 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6385 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6386 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6387 assert(isDot && "Can't compare against a vector result!");
6389 // If this is a comparison against something other than 0/1, then we know
6390 // that the condition is never/always true.
6391 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6392 if (Val != 0 && Val != 1) {
6393 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6394 return N->getOperand(0);
6395 // Always !=, turn it into an unconditional branch.
6396 return DAG.getNode(ISD::BR, dl, MVT::Other,
6397 N->getOperand(0), N->getOperand(4));
6400 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6402 // Create the PPCISD altivec 'dot' comparison node.
6403 std::vector<EVT> VTs;
6405 LHS.getOperand(2), // LHS of compare
6406 LHS.getOperand(3), // RHS of compare
6407 DAG.getConstant(CompareOpc, MVT::i32)
6409 VTs.push_back(LHS.getOperand(2).getValueType());
6410 VTs.push_back(MVT::Glue);
6411 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6413 // Unpack the result based on how the target uses it.
6414 PPC::Predicate CompOpc;
6415 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6416 default: // Can't happen, don't crash on invalid number though.
6417 case 0: // Branch on the value of the EQ bit of CR6.
6418 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6420 case 1: // Branch on the inverted value of the EQ bit of CR6.
6421 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6423 case 2: // Branch on the value of the LT bit of CR6.
6424 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6426 case 3: // Branch on the inverted value of the LT bit of CR6.
6427 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6431 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6432 DAG.getConstant(CompOpc, MVT::i32),
6433 DAG.getRegister(PPC::CR6, MVT::i32),
6434 N->getOperand(4), CompNode.getValue(1));
6443 //===----------------------------------------------------------------------===//
6444 // Inline Assembly Support
6445 //===----------------------------------------------------------------------===//
6447 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6450 const SelectionDAG &DAG,
6451 unsigned Depth) const {
6452 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6453 switch (Op.getOpcode()) {
6455 case PPCISD::LBRX: {
6456 // lhbrx is known to have the top bits cleared out.
6457 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6458 KnownZero = 0xFFFF0000;
6461 case ISD::INTRINSIC_WO_CHAIN: {
6462 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6464 case Intrinsic::ppc_altivec_vcmpbfp_p:
6465 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6466 case Intrinsic::ppc_altivec_vcmpequb_p:
6467 case Intrinsic::ppc_altivec_vcmpequh_p:
6468 case Intrinsic::ppc_altivec_vcmpequw_p:
6469 case Intrinsic::ppc_altivec_vcmpgefp_p:
6470 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6471 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6472 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6473 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6474 case Intrinsic::ppc_altivec_vcmpgtub_p:
6475 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6476 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6477 KnownZero = ~1U; // All bits but the low one are known to be zero.
6485 /// getConstraintType - Given a constraint, return the type of
6486 /// constraint it is for this target.
6487 PPCTargetLowering::ConstraintType
6488 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6489 if (Constraint.size() == 1) {
6490 switch (Constraint[0]) {
6497 return C_RegisterClass;
6499 // FIXME: While Z does indicate a memory constraint, it specifically
6500 // indicates an r+r address (used in conjunction with the 'y' modifier
6501 // in the replacement string). Currently, we're forcing the base
6502 // register to be r0 in the asm printer (which is interpreted as zero)
6503 // and forming the complete address in the second register. This is
6508 return TargetLowering::getConstraintType(Constraint);
6511 /// Examine constraint type and operand type and determine a weight value.
6512 /// This object must already have been set up with the operand type
6513 /// and the current alternative constraint selected.
6514 TargetLowering::ConstraintWeight
6515 PPCTargetLowering::getSingleConstraintMatchWeight(
6516 AsmOperandInfo &info, const char *constraint) const {
6517 ConstraintWeight weight = CW_Invalid;
6518 Value *CallOperandVal = info.CallOperandVal;
6519 // If we don't have a value, we can't do a match,
6520 // but allow it at the lowest weight.
6521 if (CallOperandVal == NULL)
6523 Type *type = CallOperandVal->getType();
6524 // Look at the constraint type.
6525 switch (*constraint) {
6527 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6530 if (type->isIntegerTy())
6531 weight = CW_Register;
6534 if (type->isFloatTy())
6535 weight = CW_Register;
6538 if (type->isDoubleTy())
6539 weight = CW_Register;
6542 if (type->isVectorTy())
6543 weight = CW_Register;
6546 weight = CW_Register;
6555 std::pair<unsigned, const TargetRegisterClass*>
6556 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6558 if (Constraint.size() == 1) {
6559 // GCC RS6000 Constraint Letters
6560 switch (Constraint[0]) {
6563 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6564 return std::make_pair(0U, &PPC::G8RCRegClass);
6565 return std::make_pair(0U, &PPC::GPRCRegClass);
6567 if (VT == MVT::f32 || VT == MVT::i32)
6568 return std::make_pair(0U, &PPC::F4RCRegClass);
6569 if (VT == MVT::f64 || VT == MVT::i64)
6570 return std::make_pair(0U, &PPC::F8RCRegClass);
6573 return std::make_pair(0U, &PPC::VRRCRegClass);
6575 return std::make_pair(0U, &PPC::CRRCRegClass);
6579 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6583 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6584 /// vector. If it is invalid, don't add anything to Ops.
6585 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6586 std::string &Constraint,
6587 std::vector<SDValue>&Ops,
6588 SelectionDAG &DAG) const {
6589 SDValue Result(0,0);
6591 // Only support length 1 constraints.
6592 if (Constraint.length() > 1) return;
6594 char Letter = Constraint[0];
6605 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6606 if (!CST) return; // Must be an immediate to match.
6607 unsigned Value = CST->getZExtValue();
6609 default: llvm_unreachable("Unknown constraint letter!");
6610 case 'I': // "I" is a signed 16-bit constant.
6611 if ((short)Value == (int)Value)
6612 Result = DAG.getTargetConstant(Value, Op.getValueType());
6614 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6615 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6616 if ((short)Value == 0)
6617 Result = DAG.getTargetConstant(Value, Op.getValueType());
6619 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6620 if ((Value >> 16) == 0)
6621 Result = DAG.getTargetConstant(Value, Op.getValueType());
6623 case 'M': // "M" is a constant that is greater than 31.
6625 Result = DAG.getTargetConstant(Value, Op.getValueType());
6627 case 'N': // "N" is a positive constant that is an exact power of two.
6628 if ((int)Value > 0 && isPowerOf2_32(Value))
6629 Result = DAG.getTargetConstant(Value, Op.getValueType());
6631 case 'O': // "O" is the constant zero.
6633 Result = DAG.getTargetConstant(Value, Op.getValueType());
6635 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6636 if ((short)-Value == (int)-Value)
6637 Result = DAG.getTargetConstant(Value, Op.getValueType());
6644 if (Result.getNode()) {
6645 Ops.push_back(Result);
6649 // Handle standard constraint letters.
6650 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6653 // isLegalAddressingMode - Return true if the addressing mode represented
6654 // by AM is legal for this target, for a load/store of the specified type.
6655 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6657 // FIXME: PPC does not allow r+i addressing modes for vectors!
6659 // PPC allows a sign-extended 16-bit immediate field.
6660 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6663 // No global is ever allowed as a base.
6667 // PPC only support r+r,
6669 case 0: // "r+i" or just "i", depending on HasBaseReg.
6672 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6674 // Otherwise we have r+r or r+i.
6677 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6679 // Allow 2*r as r+r.
6682 // No other scales are supported.
6689 /// isLegalAddressImmediate - Return true if the integer value can be used
6690 /// as the offset of the target addressing mode for load / store of the
6692 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6693 // PPC allows a sign-extended 16-bit immediate field.
6694 return (V > -(1 << 16) && V < (1 << 16)-1);
6697 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6701 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6702 SelectionDAG &DAG) const {
6703 MachineFunction &MF = DAG.getMachineFunction();
6704 MachineFrameInfo *MFI = MF.getFrameInfo();
6705 MFI->setReturnAddressIsTaken(true);
6707 DebugLoc dl = Op.getDebugLoc();
6708 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6710 // Make sure the function does not optimize away the store of the RA to
6712 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6713 FuncInfo->setLRStoreRequired();
6714 bool isPPC64 = PPCSubTarget.isPPC64();
6715 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6718 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6721 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6722 isPPC64? MVT::i64 : MVT::i32);
6723 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6724 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6726 MachinePointerInfo(), false, false, false, 0);
6729 // Just load the return address off the stack.
6730 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6731 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6732 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6735 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6736 SelectionDAG &DAG) const {
6737 DebugLoc dl = Op.getDebugLoc();
6738 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6740 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6741 bool isPPC64 = PtrVT == MVT::i64;
6743 MachineFunction &MF = DAG.getMachineFunction();
6744 MachineFrameInfo *MFI = MF.getFrameInfo();
6745 MFI->setFrameAddressIsTaken(true);
6746 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6747 MFI->hasVarSizedObjects()) &&
6748 MFI->getStackSize() &&
6749 !MF.getFunction()->getFnAttributes().
6750 hasAttribute(Attributes::Naked);
6751 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6752 (is31 ? PPC::R31 : PPC::R1);
6753 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6756 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6757 FrameAddr, MachinePointerInfo(), false, false,
6763 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6764 // The PowerPC target isn't yet aware of offsets.
6768 /// getOptimalMemOpType - Returns the target specific optimal type for load
6769 /// and store operations as a result of memset, memcpy, and memmove
6770 /// lowering. If DstAlign is zero that means it's safe to destination
6771 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6772 /// means there isn't a need to check it against alignment requirement,
6773 /// probably because the source does not need to be loaded. If
6774 /// 'IsZeroVal' is true, that means it's safe to return a
6775 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6776 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6777 /// constant so it does not need to be loaded.
6778 /// It returns EVT::Other if the type should be determined using generic
6779 /// target-independent logic.
6780 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6781 unsigned DstAlign, unsigned SrcAlign,
6784 MachineFunction &MF) const {
6785 if (this->PPCSubTarget.isPPC64()) {
6792 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6793 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6794 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6795 /// is expanded to mul + add.
6796 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6800 switch (VT.getSimpleVT().SimpleTy) {
6812 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6814 return TargetLowering::getSchedulingPreference(N);