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";
581 case PPCISD::LD_GOT_TPREL: return "PPCISD::LD_GOT_TPREL";
582 case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS";
586 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
589 return VT.changeVectorElementTypeToInteger();
592 //===----------------------------------------------------------------------===//
593 // Node matching predicates, for use by the tblgen matching code.
594 //===----------------------------------------------------------------------===//
596 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
597 static bool isFloatingPointZero(SDValue Op) {
598 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
599 return CFP->getValueAPF().isZero();
600 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
601 // Maybe this has already been legalized into the constant pool?
602 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
603 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
604 return CFP->getValueAPF().isZero();
609 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
610 /// true if Op is undef or if it matches the specified value.
611 static bool isConstantOrUndef(int Op, int Val) {
612 return Op < 0 || Op == Val;
615 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
616 /// VPKUHUM instruction.
617 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
619 for (unsigned i = 0; i != 16; ++i)
620 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
623 for (unsigned i = 0; i != 8; ++i)
624 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
625 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
631 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
632 /// VPKUWUM instruction.
633 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
635 for (unsigned i = 0; i != 16; i += 2)
636 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
637 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
640 for (unsigned i = 0; i != 8; i += 2)
641 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
642 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
643 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
644 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
650 /// isVMerge - Common function, used to match vmrg* shuffles.
652 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
653 unsigned LHSStart, unsigned RHSStart) {
654 assert(N->getValueType(0) == MVT::v16i8 &&
655 "PPC only supports shuffles by bytes!");
656 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
657 "Unsupported merge size!");
659 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
660 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
661 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
662 LHSStart+j+i*UnitSize) ||
663 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
664 RHSStart+j+i*UnitSize))
670 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
671 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
672 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
675 return isVMerge(N, UnitSize, 8, 24);
676 return isVMerge(N, UnitSize, 8, 8);
679 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
680 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
681 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
684 return isVMerge(N, UnitSize, 0, 16);
685 return isVMerge(N, UnitSize, 0, 0);
689 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
690 /// amount, otherwise return -1.
691 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
692 assert(N->getValueType(0) == MVT::v16i8 &&
693 "PPC only supports shuffles by bytes!");
695 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
697 // Find the first non-undef value in the shuffle mask.
699 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
702 if (i == 16) return -1; // all undef.
704 // Otherwise, check to see if the rest of the elements are consecutively
705 // numbered from this value.
706 unsigned ShiftAmt = SVOp->getMaskElt(i);
707 if (ShiftAmt < i) return -1;
711 // Check the rest of the elements to see if they are consecutive.
712 for (++i; i != 16; ++i)
713 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
716 // Check the rest of the elements to see if they are consecutive.
717 for (++i; i != 16; ++i)
718 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
724 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
725 /// specifies a splat of a single element that is suitable for input to
726 /// VSPLTB/VSPLTH/VSPLTW.
727 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
728 assert(N->getValueType(0) == MVT::v16i8 &&
729 (EltSize == 1 || EltSize == 2 || EltSize == 4));
731 // This is a splat operation if each element of the permute is the same, and
732 // if the value doesn't reference the second vector.
733 unsigned ElementBase = N->getMaskElt(0);
735 // FIXME: Handle UNDEF elements too!
736 if (ElementBase >= 16)
739 // Check that the indices are consecutive, in the case of a multi-byte element
740 // splatted with a v16i8 mask.
741 for (unsigned i = 1; i != EltSize; ++i)
742 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
745 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
746 if (N->getMaskElt(i) < 0) continue;
747 for (unsigned j = 0; j != EltSize; ++j)
748 if (N->getMaskElt(i+j) != N->getMaskElt(j))
754 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
756 bool PPC::isAllNegativeZeroVector(SDNode *N) {
757 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
759 APInt APVal, APUndef;
763 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
764 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
765 return CFP->getValueAPF().isNegZero();
770 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
771 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
772 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
773 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
774 assert(isSplatShuffleMask(SVOp, EltSize));
775 return SVOp->getMaskElt(0) / EltSize;
778 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
779 /// by using a vspltis[bhw] instruction of the specified element size, return
780 /// the constant being splatted. The ByteSize field indicates the number of
781 /// bytes of each element [124] -> [bhw].
782 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
785 // If ByteSize of the splat is bigger than the element size of the
786 // build_vector, then we have a case where we are checking for a splat where
787 // multiple elements of the buildvector are folded together into a single
788 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
789 unsigned EltSize = 16/N->getNumOperands();
790 if (EltSize < ByteSize) {
791 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
792 SDValue UniquedVals[4];
793 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
795 // See if all of the elements in the buildvector agree across.
796 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
797 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
798 // If the element isn't a constant, bail fully out.
799 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
802 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
803 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
804 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
805 return SDValue(); // no match.
808 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
809 // either constant or undef values that are identical for each chunk. See
810 // if these chunks can form into a larger vspltis*.
812 // Check to see if all of the leading entries are either 0 or -1. If
813 // neither, then this won't fit into the immediate field.
814 bool LeadingZero = true;
815 bool LeadingOnes = true;
816 for (unsigned i = 0; i != Multiple-1; ++i) {
817 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
819 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
820 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
822 // Finally, check the least significant entry.
824 if (UniquedVals[Multiple-1].getNode() == 0)
825 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
826 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
828 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
831 if (UniquedVals[Multiple-1].getNode() == 0)
832 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
833 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
834 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
835 return DAG.getTargetConstant(Val, MVT::i32);
841 // Check to see if this buildvec has a single non-undef value in its elements.
842 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
843 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
844 if (OpVal.getNode() == 0)
845 OpVal = N->getOperand(i);
846 else if (OpVal != N->getOperand(i))
850 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
852 unsigned ValSizeInBytes = EltSize;
854 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
855 Value = CN->getZExtValue();
856 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
857 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
858 Value = FloatToBits(CN->getValueAPF().convertToFloat());
861 // If the splat value is larger than the element value, then we can never do
862 // this splat. The only case that we could fit the replicated bits into our
863 // immediate field for would be zero, and we prefer to use vxor for it.
864 if (ValSizeInBytes < ByteSize) return SDValue();
866 // If the element value is larger than the splat value, cut it in half and
867 // check to see if the two halves are equal. Continue doing this until we
868 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
869 while (ValSizeInBytes > ByteSize) {
870 ValSizeInBytes >>= 1;
872 // If the top half equals the bottom half, we're still ok.
873 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
874 (Value & ((1 << (8*ValSizeInBytes))-1)))
878 // Properly sign extend the value.
879 int MaskVal = SignExtend32(Value, ByteSize * 8);
881 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
882 if (MaskVal == 0) return SDValue();
884 // Finally, if this value fits in a 5 bit sext field, return it
885 if (SignExtend32<5>(MaskVal) == MaskVal)
886 return DAG.getTargetConstant(MaskVal, MVT::i32);
890 //===----------------------------------------------------------------------===//
891 // Addressing Mode Selection
892 //===----------------------------------------------------------------------===//
894 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
895 /// or 64-bit immediate, and if the value can be accurately represented as a
896 /// sign extension from a 16-bit value. If so, this returns true and the
898 static bool isIntS16Immediate(SDNode *N, short &Imm) {
899 if (N->getOpcode() != ISD::Constant)
902 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
903 if (N->getValueType(0) == MVT::i32)
904 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
906 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
908 static bool isIntS16Immediate(SDValue Op, short &Imm) {
909 return isIntS16Immediate(Op.getNode(), Imm);
913 /// SelectAddressRegReg - Given the specified addressed, check to see if it
914 /// can be represented as an indexed [r+r] operation. Returns false if it
915 /// can be more efficiently represented with [r+imm].
916 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
918 SelectionDAG &DAG) const {
920 if (N.getOpcode() == ISD::ADD) {
921 if (isIntS16Immediate(N.getOperand(1), imm))
923 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
926 Base = N.getOperand(0);
927 Index = N.getOperand(1);
929 } else if (N.getOpcode() == ISD::OR) {
930 if (isIntS16Immediate(N.getOperand(1), imm))
931 return false; // r+i can fold it if we can.
933 // If this is an or of disjoint bitfields, we can codegen this as an add
934 // (for better address arithmetic) if the LHS and RHS of the OR are provably
936 APInt LHSKnownZero, LHSKnownOne;
937 APInt RHSKnownZero, RHSKnownOne;
938 DAG.ComputeMaskedBits(N.getOperand(0),
939 LHSKnownZero, LHSKnownOne);
941 if (LHSKnownZero.getBoolValue()) {
942 DAG.ComputeMaskedBits(N.getOperand(1),
943 RHSKnownZero, RHSKnownOne);
944 // If all of the bits are known zero on the LHS or RHS, the add won't
946 if (~(LHSKnownZero | RHSKnownZero) == 0) {
947 Base = N.getOperand(0);
948 Index = N.getOperand(1);
957 /// Returns true if the address N can be represented by a base register plus
958 /// a signed 16-bit displacement [r+imm], and if it is not better
959 /// represented as reg+reg.
960 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
962 SelectionDAG &DAG) const {
963 // FIXME dl should come from parent load or store, not from address
964 DebugLoc dl = N.getDebugLoc();
965 // If this can be more profitably realized as r+r, fail.
966 if (SelectAddressRegReg(N, Disp, Base, DAG))
969 if (N.getOpcode() == ISD::ADD) {
971 if (isIntS16Immediate(N.getOperand(1), imm)) {
972 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
973 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
974 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
976 Base = N.getOperand(0);
978 return true; // [r+i]
979 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
980 // Match LOAD (ADD (X, Lo(G))).
981 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
982 && "Cannot handle constant offsets yet!");
983 Disp = N.getOperand(1).getOperand(0); // The global address.
984 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
985 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
986 Disp.getOpcode() == ISD::TargetConstantPool ||
987 Disp.getOpcode() == ISD::TargetJumpTable);
988 Base = N.getOperand(0);
989 return true; // [&g+r]
991 } else if (N.getOpcode() == ISD::OR) {
993 if (isIntS16Immediate(N.getOperand(1), imm)) {
994 // If this is an or of disjoint bitfields, we can codegen this as an add
995 // (for better address arithmetic) if the LHS and RHS of the OR are
996 // provably disjoint.
997 APInt LHSKnownZero, LHSKnownOne;
998 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1000 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1001 // If all of the bits are known zero on the LHS or RHS, the add won't
1003 Base = N.getOperand(0);
1004 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
1008 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1009 // Loading from a constant address.
1011 // If this address fits entirely in a 16-bit sext immediate field, codegen
1014 if (isIntS16Immediate(CN, Imm)) {
1015 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
1016 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1017 CN->getValueType(0));
1021 // Handle 32-bit sext immediates with LIS + addr mode.
1022 if (CN->getValueType(0) == MVT::i32 ||
1023 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1024 int Addr = (int)CN->getZExtValue();
1026 // Otherwise, break this down into an LIS + disp.
1027 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1029 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1030 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1031 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1036 Disp = DAG.getTargetConstant(0, getPointerTy());
1037 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1038 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1041 return true; // [r+0]
1044 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1045 /// represented as an indexed [r+r] operation.
1046 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1048 SelectionDAG &DAG) const {
1049 // Check to see if we can easily represent this as an [r+r] address. This
1050 // will fail if it thinks that the address is more profitably represented as
1051 // reg+imm, e.g. where imm = 0.
1052 if (SelectAddressRegReg(N, Base, Index, DAG))
1055 // If the operand is an addition, always emit this as [r+r], since this is
1056 // better (for code size, and execution, as the memop does the add for free)
1057 // than emitting an explicit add.
1058 if (N.getOpcode() == ISD::ADD) {
1059 Base = N.getOperand(0);
1060 Index = N.getOperand(1);
1064 // Otherwise, do it the hard way, using R0 as the base register.
1065 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1071 /// SelectAddressRegImmShift - Returns true if the address N can be
1072 /// represented by a base register plus a signed 14-bit displacement
1073 /// [r+imm*4]. Suitable for use by STD and friends.
1074 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1076 SelectionDAG &DAG) const {
1077 // FIXME dl should come from the parent load or store, not the address
1078 DebugLoc dl = N.getDebugLoc();
1079 // If this can be more profitably realized as r+r, fail.
1080 if (SelectAddressRegReg(N, Disp, Base, DAG))
1083 if (N.getOpcode() == ISD::ADD) {
1085 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1086 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1087 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1088 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1090 Base = N.getOperand(0);
1092 return true; // [r+i]
1093 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1094 // Match LOAD (ADD (X, Lo(G))).
1095 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1096 && "Cannot handle constant offsets yet!");
1097 Disp = N.getOperand(1).getOperand(0); // The global address.
1098 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1099 Disp.getOpcode() == ISD::TargetConstantPool ||
1100 Disp.getOpcode() == ISD::TargetJumpTable);
1101 Base = N.getOperand(0);
1102 return true; // [&g+r]
1104 } else if (N.getOpcode() == ISD::OR) {
1106 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1107 // If this is an or of disjoint bitfields, we can codegen this as an add
1108 // (for better address arithmetic) if the LHS and RHS of the OR are
1109 // provably disjoint.
1110 APInt LHSKnownZero, LHSKnownOne;
1111 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1112 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1113 // If all of the bits are known zero on the LHS or RHS, the add won't
1115 Base = N.getOperand(0);
1116 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1120 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1121 // Loading from a constant address. Verify low two bits are clear.
1122 if ((CN->getZExtValue() & 3) == 0) {
1123 // If this address fits entirely in a 14-bit sext immediate field, codegen
1126 if (isIntS16Immediate(CN, Imm)) {
1127 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1128 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1129 CN->getValueType(0));
1133 // Fold the low-part of 32-bit absolute addresses into addr mode.
1134 if (CN->getValueType(0) == MVT::i32 ||
1135 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1136 int Addr = (int)CN->getZExtValue();
1138 // Otherwise, break this down into an LIS + disp.
1139 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1140 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1141 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1142 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1148 Disp = DAG.getTargetConstant(0, getPointerTy());
1149 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1150 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1153 return true; // [r+0]
1157 /// getPreIndexedAddressParts - returns true by value, base pointer and
1158 /// offset pointer and addressing mode by reference if the node's address
1159 /// can be legally represented as pre-indexed load / store address.
1160 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1162 ISD::MemIndexedMode &AM,
1163 SelectionDAG &DAG) const {
1164 if (DisablePPCPreinc) return false;
1168 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1169 Ptr = LD->getBasePtr();
1170 VT = LD->getMemoryVT();
1172 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1173 Ptr = ST->getBasePtr();
1174 VT = ST->getMemoryVT();
1178 // PowerPC doesn't have preinc load/store instructions for vectors.
1182 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1187 // LDU/STU use reg+imm*4, others use reg+imm.
1188 if (VT != MVT::i64) {
1190 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1194 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1198 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1199 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1200 // sext i32 to i64 when addr mode is r+i.
1201 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1202 LD->getExtensionType() == ISD::SEXTLOAD &&
1203 isa<ConstantSDNode>(Offset))
1211 //===----------------------------------------------------------------------===//
1212 // LowerOperation implementation
1213 //===----------------------------------------------------------------------===//
1215 /// GetLabelAccessInfo - Return true if we should reference labels using a
1216 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1217 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1218 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1219 HiOpFlags = PPCII::MO_HA16;
1220 LoOpFlags = PPCII::MO_LO16;
1222 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1223 // non-darwin platform. We don't support PIC on other platforms yet.
1224 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1225 TM.getSubtarget<PPCSubtarget>().isDarwin();
1227 HiOpFlags |= PPCII::MO_PIC_FLAG;
1228 LoOpFlags |= PPCII::MO_PIC_FLAG;
1231 // If this is a reference to a global value that requires a non-lazy-ptr, make
1232 // sure that instruction lowering adds it.
1233 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1234 HiOpFlags |= PPCII::MO_NLP_FLAG;
1235 LoOpFlags |= PPCII::MO_NLP_FLAG;
1237 if (GV->hasHiddenVisibility()) {
1238 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1239 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1246 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1247 SelectionDAG &DAG) {
1248 EVT PtrVT = HiPart.getValueType();
1249 SDValue Zero = DAG.getConstant(0, PtrVT);
1250 DebugLoc DL = HiPart.getDebugLoc();
1252 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1253 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1255 // With PIC, the first instruction is actually "GR+hi(&G)".
1257 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1258 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1260 // Generate non-pic code that has direct accesses to the constant pool.
1261 // The address of the global is just (hi(&g)+lo(&g)).
1262 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1265 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1266 SelectionDAG &DAG) const {
1267 EVT PtrVT = Op.getValueType();
1268 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1269 const Constant *C = CP->getConstVal();
1271 // 64-bit SVR4 ABI code is always position-independent.
1272 // The actual address of the GlobalValue is stored in the TOC.
1273 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1274 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1275 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1276 DAG.getRegister(PPC::X2, MVT::i64));
1279 unsigned MOHiFlag, MOLoFlag;
1280 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1282 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1284 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1285 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1288 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1289 EVT PtrVT = Op.getValueType();
1290 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1292 // 64-bit SVR4 ABI code is always position-independent.
1293 // The actual address of the GlobalValue is stored in the TOC.
1294 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1295 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1296 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1297 DAG.getRegister(PPC::X2, MVT::i64));
1300 unsigned MOHiFlag, MOLoFlag;
1301 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1302 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1303 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1304 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1307 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1308 SelectionDAG &DAG) const {
1309 EVT PtrVT = Op.getValueType();
1311 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1313 unsigned MOHiFlag, MOLoFlag;
1314 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1315 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1316 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1317 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1320 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1321 SelectionDAG &DAG) const {
1323 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1324 DebugLoc dl = GA->getDebugLoc();
1325 const GlobalValue *GV = GA->getGlobal();
1326 EVT PtrVT = getPointerTy();
1327 bool is64bit = PPCSubTarget.isPPC64();
1329 TLSModel::Model Model = getTargetMachine().getTLSModel(GV);
1331 if (Model == TLSModel::LocalExec) {
1332 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1333 PPCII::MO_TPREL16_HA);
1334 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1335 PPCII::MO_TPREL16_LO);
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);
1343 llvm_unreachable("only local-exec is currently supported for ppc32");
1345 if (Model != TLSModel::InitialExec)
1346 llvm_unreachable("only local-exec and initial-exec TLS modes supported");
1348 SDValue GOTOffset = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1349 PPCII::MO_GOT_TPREL16_DS);
1350 SDValue TPReg = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1352 SDValue GOTReg = DAG.getRegister(is64bit ? PPC::X2 : PPC::R2,
1353 is64bit ? MVT::i64 : MVT::i32);
1354 SDValue TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL, dl, PtrVT,
1356 return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TPReg);
1359 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1360 SelectionDAG &DAG) const {
1361 EVT PtrVT = Op.getValueType();
1362 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1363 DebugLoc DL = GSDN->getDebugLoc();
1364 const GlobalValue *GV = GSDN->getGlobal();
1366 // 64-bit SVR4 ABI code is always position-independent.
1367 // The actual address of the GlobalValue is stored in the TOC.
1368 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1369 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1370 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1371 DAG.getRegister(PPC::X2, MVT::i64));
1374 unsigned MOHiFlag, MOLoFlag;
1375 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1378 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1380 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1382 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1384 // If the global reference is actually to a non-lazy-pointer, we have to do an
1385 // extra load to get the address of the global.
1386 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1387 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1388 false, false, false, 0);
1392 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1393 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1394 DebugLoc dl = Op.getDebugLoc();
1396 // If we're comparing for equality to zero, expose the fact that this is
1397 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1398 // fold the new nodes.
1399 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1400 if (C->isNullValue() && CC == ISD::SETEQ) {
1401 EVT VT = Op.getOperand(0).getValueType();
1402 SDValue Zext = Op.getOperand(0);
1403 if (VT.bitsLT(MVT::i32)) {
1405 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1407 unsigned Log2b = Log2_32(VT.getSizeInBits());
1408 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1409 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1410 DAG.getConstant(Log2b, MVT::i32));
1411 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1413 // Leave comparisons against 0 and -1 alone for now, since they're usually
1414 // optimized. FIXME: revisit this when we can custom lower all setcc
1416 if (C->isAllOnesValue() || C->isNullValue())
1420 // If we have an integer seteq/setne, turn it into a compare against zero
1421 // by xor'ing the rhs with the lhs, which is faster than setting a
1422 // condition register, reading it back out, and masking the correct bit. The
1423 // normal approach here uses sub to do this instead of xor. Using xor exposes
1424 // the result to other bit-twiddling opportunities.
1425 EVT LHSVT = Op.getOperand(0).getValueType();
1426 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1427 EVT VT = Op.getValueType();
1428 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1430 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1435 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1436 const PPCSubtarget &Subtarget) const {
1437 SDNode *Node = Op.getNode();
1438 EVT VT = Node->getValueType(0);
1439 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1440 SDValue InChain = Node->getOperand(0);
1441 SDValue VAListPtr = Node->getOperand(1);
1442 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1443 DebugLoc dl = Node->getDebugLoc();
1445 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1448 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1449 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1451 InChain = GprIndex.getValue(1);
1453 if (VT == MVT::i64) {
1454 // Check if GprIndex is even
1455 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1456 DAG.getConstant(1, MVT::i32));
1457 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1458 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1459 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1460 DAG.getConstant(1, MVT::i32));
1461 // Align GprIndex to be even if it isn't
1462 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1466 // fpr index is 1 byte after gpr
1467 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1468 DAG.getConstant(1, MVT::i32));
1471 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1472 FprPtr, MachinePointerInfo(SV), MVT::i8,
1474 InChain = FprIndex.getValue(1);
1476 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1477 DAG.getConstant(8, MVT::i32));
1479 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1480 DAG.getConstant(4, MVT::i32));
1483 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1484 MachinePointerInfo(), false, false,
1486 InChain = OverflowArea.getValue(1);
1488 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1489 MachinePointerInfo(), false, false,
1491 InChain = RegSaveArea.getValue(1);
1493 // select overflow_area if index > 8
1494 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1495 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1497 // adjustment constant gpr_index * 4/8
1498 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1499 VT.isInteger() ? GprIndex : FprIndex,
1500 DAG.getConstant(VT.isInteger() ? 4 : 8,
1503 // OurReg = RegSaveArea + RegConstant
1504 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1507 // Floating types are 32 bytes into RegSaveArea
1508 if (VT.isFloatingPoint())
1509 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1510 DAG.getConstant(32, MVT::i32));
1512 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1513 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1514 VT.isInteger() ? GprIndex : FprIndex,
1515 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1518 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1519 VT.isInteger() ? VAListPtr : FprPtr,
1520 MachinePointerInfo(SV),
1521 MVT::i8, false, false, 0);
1523 // determine if we should load from reg_save_area or overflow_area
1524 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1526 // increase overflow_area by 4/8 if gpr/fpr > 8
1527 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1528 DAG.getConstant(VT.isInteger() ? 4 : 8,
1531 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1534 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1536 MachinePointerInfo(),
1537 MVT::i32, false, false, 0);
1539 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1540 false, false, false, 0);
1543 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1544 SelectionDAG &DAG) const {
1545 return Op.getOperand(0);
1548 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1549 SelectionDAG &DAG) const {
1550 SDValue Chain = Op.getOperand(0);
1551 SDValue Trmp = Op.getOperand(1); // trampoline
1552 SDValue FPtr = Op.getOperand(2); // nested function
1553 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1554 DebugLoc dl = Op.getDebugLoc();
1556 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1557 bool isPPC64 = (PtrVT == MVT::i64);
1559 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1562 TargetLowering::ArgListTy Args;
1563 TargetLowering::ArgListEntry Entry;
1565 Entry.Ty = IntPtrTy;
1566 Entry.Node = Trmp; Args.push_back(Entry);
1568 // TrampSize == (isPPC64 ? 48 : 40);
1569 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1570 isPPC64 ? MVT::i64 : MVT::i32);
1571 Args.push_back(Entry);
1573 Entry.Node = FPtr; Args.push_back(Entry);
1574 Entry.Node = Nest; Args.push_back(Entry);
1576 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1577 TargetLowering::CallLoweringInfo CLI(Chain,
1578 Type::getVoidTy(*DAG.getContext()),
1579 false, false, false, false, 0,
1581 /*isTailCall=*/false,
1582 /*doesNotRet=*/false,
1583 /*isReturnValueUsed=*/true,
1584 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1586 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1588 return CallResult.second;
1591 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1592 const PPCSubtarget &Subtarget) const {
1593 MachineFunction &MF = DAG.getMachineFunction();
1594 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1596 DebugLoc dl = Op.getDebugLoc();
1598 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1599 // vastart just stores the address of the VarArgsFrameIndex slot into the
1600 // memory location argument.
1601 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1602 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1603 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1604 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1605 MachinePointerInfo(SV),
1609 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1610 // We suppose the given va_list is already allocated.
1613 // char gpr; /* index into the array of 8 GPRs
1614 // * stored in the register save area
1615 // * gpr=0 corresponds to r3,
1616 // * gpr=1 to r4, etc.
1618 // char fpr; /* index into the array of 8 FPRs
1619 // * stored in the register save area
1620 // * fpr=0 corresponds to f1,
1621 // * fpr=1 to f2, etc.
1623 // char *overflow_arg_area;
1624 // /* location on stack that holds
1625 // * the next overflow argument
1627 // char *reg_save_area;
1628 // /* where r3:r10 and f1:f8 (if saved)
1634 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1635 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1638 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1640 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1642 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1645 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1646 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1648 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1649 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1651 uint64_t FPROffset = 1;
1652 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1654 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1656 // Store first byte : number of int regs
1657 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1659 MachinePointerInfo(SV),
1660 MVT::i8, false, false, 0);
1661 uint64_t nextOffset = FPROffset;
1662 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1665 // Store second byte : number of float regs
1666 SDValue secondStore =
1667 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1668 MachinePointerInfo(SV, nextOffset), MVT::i8,
1670 nextOffset += StackOffset;
1671 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1673 // Store second word : arguments given on stack
1674 SDValue thirdStore =
1675 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1676 MachinePointerInfo(SV, nextOffset),
1678 nextOffset += FrameOffset;
1679 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1681 // Store third word : arguments given in registers
1682 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1683 MachinePointerInfo(SV, nextOffset),
1688 #include "PPCGenCallingConv.inc"
1690 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1691 CCValAssign::LocInfo &LocInfo,
1692 ISD::ArgFlagsTy &ArgFlags,
1697 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1699 CCValAssign::LocInfo &LocInfo,
1700 ISD::ArgFlagsTy &ArgFlags,
1702 static const uint16_t ArgRegs[] = {
1703 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1704 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1706 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1708 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1710 // Skip one register if the first unallocated register has an even register
1711 // number and there are still argument registers available which have not been
1712 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1713 // need to skip a register if RegNum is odd.
1714 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1715 State.AllocateReg(ArgRegs[RegNum]);
1718 // Always return false here, as this function only makes sure that the first
1719 // unallocated register has an odd register number and does not actually
1720 // allocate a register for the current argument.
1724 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1726 CCValAssign::LocInfo &LocInfo,
1727 ISD::ArgFlagsTy &ArgFlags,
1729 static const uint16_t ArgRegs[] = {
1730 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1734 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1736 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1738 // If there is only one Floating-point register left we need to put both f64
1739 // values of a split ppc_fp128 value on the stack.
1740 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1741 State.AllocateReg(ArgRegs[RegNum]);
1744 // Always return false here, as this function only makes sure that the two f64
1745 // values a ppc_fp128 value is split into are both passed in registers or both
1746 // passed on the stack and does not actually allocate a register for the
1747 // current argument.
1751 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1753 static const uint16_t *GetFPR() {
1754 static const uint16_t FPR[] = {
1755 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1756 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1762 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1764 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1765 unsigned PtrByteSize) {
1766 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1767 if (Flags.isByVal())
1768 ArgSize = Flags.getByValSize();
1769 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1775 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1776 CallingConv::ID CallConv, bool isVarArg,
1777 const SmallVectorImpl<ISD::InputArg>
1779 DebugLoc dl, SelectionDAG &DAG,
1780 SmallVectorImpl<SDValue> &InVals)
1782 if (PPCSubTarget.isSVR4ABI()) {
1783 if (PPCSubTarget.isPPC64())
1784 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1787 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1790 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1796 PPCTargetLowering::LowerFormalArguments_32SVR4(
1798 CallingConv::ID CallConv, bool isVarArg,
1799 const SmallVectorImpl<ISD::InputArg>
1801 DebugLoc dl, SelectionDAG &DAG,
1802 SmallVectorImpl<SDValue> &InVals) const {
1804 // 32-bit SVR4 ABI Stack Frame Layout:
1805 // +-----------------------------------+
1806 // +--> | Back chain |
1807 // | +-----------------------------------+
1808 // | | Floating-point register save area |
1809 // | +-----------------------------------+
1810 // | | General register save area |
1811 // | +-----------------------------------+
1812 // | | CR save word |
1813 // | +-----------------------------------+
1814 // | | VRSAVE save word |
1815 // | +-----------------------------------+
1816 // | | Alignment padding |
1817 // | +-----------------------------------+
1818 // | | Vector register save area |
1819 // | +-----------------------------------+
1820 // | | Local variable space |
1821 // | +-----------------------------------+
1822 // | | Parameter list area |
1823 // | +-----------------------------------+
1824 // | | LR save word |
1825 // | +-----------------------------------+
1826 // SP--> +--- | Back chain |
1827 // +-----------------------------------+
1830 // System V Application Binary Interface PowerPC Processor Supplement
1831 // AltiVec Technology Programming Interface Manual
1833 MachineFunction &MF = DAG.getMachineFunction();
1834 MachineFrameInfo *MFI = MF.getFrameInfo();
1835 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1837 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1838 // Potential tail calls could cause overwriting of argument stack slots.
1839 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1840 (CallConv == CallingConv::Fast));
1841 unsigned PtrByteSize = 4;
1843 // Assign locations to all of the incoming arguments.
1844 SmallVector<CCValAssign, 16> ArgLocs;
1845 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1846 getTargetMachine(), ArgLocs, *DAG.getContext());
1848 // Reserve space for the linkage area on the stack.
1849 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1851 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1853 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1854 CCValAssign &VA = ArgLocs[i];
1856 // Arguments stored in registers.
1857 if (VA.isRegLoc()) {
1858 const TargetRegisterClass *RC;
1859 EVT ValVT = VA.getValVT();
1861 switch (ValVT.getSimpleVT().SimpleTy) {
1863 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1865 RC = &PPC::GPRCRegClass;
1868 RC = &PPC::F4RCRegClass;
1871 RC = &PPC::F8RCRegClass;
1877 RC = &PPC::VRRCRegClass;
1881 // Transform the arguments stored in physical registers into virtual ones.
1882 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1883 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1885 InVals.push_back(ArgValue);
1887 // Argument stored in memory.
1888 assert(VA.isMemLoc());
1890 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1891 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1894 // Create load nodes to retrieve arguments from the stack.
1895 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1896 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1897 MachinePointerInfo(),
1898 false, false, false, 0));
1902 // Assign locations to all of the incoming aggregate by value arguments.
1903 // Aggregates passed by value are stored in the local variable space of the
1904 // caller's stack frame, right above the parameter list area.
1905 SmallVector<CCValAssign, 16> ByValArgLocs;
1906 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1907 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1909 // Reserve stack space for the allocations in CCInfo.
1910 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1912 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1914 // Area that is at least reserved in the caller of this function.
1915 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1917 // Set the size that is at least reserved in caller of this function. Tail
1918 // call optimized function's reserved stack space needs to be aligned so that
1919 // taking the difference between two stack areas will result in an aligned
1921 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1924 std::max(MinReservedArea,
1925 PPCFrameLowering::getMinCallFrameSize(false, false));
1927 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1928 getStackAlignment();
1929 unsigned AlignMask = TargetAlign-1;
1930 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1932 FI->setMinReservedArea(MinReservedArea);
1934 SmallVector<SDValue, 8> MemOps;
1936 // If the function takes variable number of arguments, make a frame index for
1937 // the start of the first vararg value... for expansion of llvm.va_start.
1939 static const uint16_t GPArgRegs[] = {
1940 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1941 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1943 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1945 static const uint16_t FPArgRegs[] = {
1946 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1949 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1951 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1953 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1956 // Make room for NumGPArgRegs and NumFPArgRegs.
1957 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1958 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1960 FuncInfo->setVarArgsStackOffset(
1961 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1962 CCInfo.getNextStackOffset(), true));
1964 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1965 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1967 // The fixed integer arguments of a variadic function are stored to the
1968 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1969 // the result of va_next.
1970 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1971 // Get an existing live-in vreg, or add a new one.
1972 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1974 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1976 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1977 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1978 MachinePointerInfo(), false, false, 0);
1979 MemOps.push_back(Store);
1980 // Increment the address by four for the next argument to store
1981 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1982 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1985 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1987 // The double arguments are stored to the VarArgsFrameIndex
1989 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1990 // Get an existing live-in vreg, or add a new one.
1991 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1993 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1995 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1996 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1997 MachinePointerInfo(), false, false, 0);
1998 MemOps.push_back(Store);
1999 // Increment the address by eight for the next argument to store
2000 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
2002 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2006 if (!MemOps.empty())
2007 Chain = DAG.getNode(ISD::TokenFactor, dl,
2008 MVT::Other, &MemOps[0], MemOps.size());
2013 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2014 // value to MVT::i64 and then truncate to the correct register size.
2016 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
2017 SelectionDAG &DAG, SDValue ArgVal,
2018 DebugLoc dl) const {
2020 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2021 DAG.getValueType(ObjectVT));
2022 else if (Flags.isZExt())
2023 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2024 DAG.getValueType(ObjectVT));
2026 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2029 // Set the size that is at least reserved in caller of this function. Tail
2030 // call optimized functions' reserved stack space needs to be aligned so that
2031 // taking the difference between two stack areas will result in an aligned
2034 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
2035 unsigned nAltivecParamsAtEnd,
2036 unsigned MinReservedArea,
2037 bool isPPC64) const {
2038 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2039 // Add the Altivec parameters at the end, if needed.
2040 if (nAltivecParamsAtEnd) {
2041 MinReservedArea = ((MinReservedArea+15)/16)*16;
2042 MinReservedArea += 16*nAltivecParamsAtEnd;
2045 std::max(MinReservedArea,
2046 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2047 unsigned TargetAlign
2048 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2049 getStackAlignment();
2050 unsigned AlignMask = TargetAlign-1;
2051 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2052 FI->setMinReservedArea(MinReservedArea);
2056 PPCTargetLowering::LowerFormalArguments_64SVR4(
2058 CallingConv::ID CallConv, bool isVarArg,
2059 const SmallVectorImpl<ISD::InputArg>
2061 DebugLoc dl, SelectionDAG &DAG,
2062 SmallVectorImpl<SDValue> &InVals) const {
2063 // TODO: add description of PPC stack frame format, or at least some docs.
2065 MachineFunction &MF = DAG.getMachineFunction();
2066 MachineFrameInfo *MFI = MF.getFrameInfo();
2067 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2069 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2070 // Potential tail calls could cause overwriting of argument stack slots.
2071 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2072 (CallConv == CallingConv::Fast));
2073 unsigned PtrByteSize = 8;
2075 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2076 // Area that is at least reserved in caller of this function.
2077 unsigned MinReservedArea = ArgOffset;
2079 static const uint16_t GPR[] = {
2080 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2081 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2084 static const uint16_t *FPR = GetFPR();
2086 static const uint16_t VR[] = {
2087 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2088 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2091 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2092 const unsigned Num_FPR_Regs = 13;
2093 const unsigned Num_VR_Regs = array_lengthof(VR);
2095 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2097 // Add DAG nodes to load the arguments or copy them out of registers. On
2098 // entry to a function on PPC, the arguments start after the linkage area,
2099 // although the first ones are often in registers.
2101 SmallVector<SDValue, 8> MemOps;
2102 unsigned nAltivecParamsAtEnd = 0;
2103 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2104 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2106 bool needsLoad = false;
2107 EVT ObjectVT = Ins[ArgNo].VT;
2108 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2109 unsigned ArgSize = ObjSize;
2110 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2112 unsigned CurArgOffset = ArgOffset;
2114 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2115 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2116 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2118 MinReservedArea = ((MinReservedArea+15)/16)*16;
2119 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2123 nAltivecParamsAtEnd++;
2125 // Calculate min reserved area.
2126 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2130 // FIXME the codegen can be much improved in some cases.
2131 // We do not have to keep everything in memory.
2132 if (Flags.isByVal()) {
2133 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2134 ObjSize = Flags.getByValSize();
2135 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2136 // Empty aggregate parameters do not take up registers. Examples:
2140 // etc. However, we have to provide a place-holder in InVals, so
2141 // pretend we have an 8-byte item at the current address for that
2144 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2145 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2146 InVals.push_back(FIN);
2149 // All aggregates smaller than 8 bytes must be passed right-justified.
2150 if (ObjSize < PtrByteSize)
2151 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2152 // The value of the object is its address.
2153 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2154 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2155 InVals.push_back(FIN);
2158 if (GPR_idx != Num_GPR_Regs) {
2159 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2160 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2163 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2164 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2165 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2166 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2167 MachinePointerInfo(FuncArg, CurArgOffset),
2168 ObjType, false, false, 0);
2170 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2171 // store the whole register as-is to the parameter save area
2172 // slot. The address of the parameter was already calculated
2173 // above (InVals.push_back(FIN)) to be the right-justified
2174 // offset within the slot. For this store, we need a new
2175 // frame index that points at the beginning of the slot.
2176 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2177 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2178 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2179 MachinePointerInfo(FuncArg, ArgOffset),
2183 MemOps.push_back(Store);
2186 // Whether we copied from a register or not, advance the offset
2187 // into the parameter save area by a full doubleword.
2188 ArgOffset += PtrByteSize;
2192 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2193 // Store whatever pieces of the object are in registers
2194 // to memory. ArgOffset will be the address of the beginning
2196 if (GPR_idx != Num_GPR_Regs) {
2198 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2199 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2200 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2201 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2202 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2203 MachinePointerInfo(FuncArg, ArgOffset),
2205 MemOps.push_back(Store);
2207 ArgOffset += PtrByteSize;
2209 ArgOffset += ArgSize - j;
2216 switch (ObjectVT.getSimpleVT().SimpleTy) {
2217 default: llvm_unreachable("Unhandled argument type!");
2220 if (GPR_idx != Num_GPR_Regs) {
2221 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2222 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2224 if (ObjectVT == MVT::i32)
2225 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2226 // value to MVT::i64 and then truncate to the correct register size.
2227 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2232 ArgSize = PtrByteSize;
2239 // Every 8 bytes of argument space consumes one of the GPRs available for
2240 // argument passing.
2241 if (GPR_idx != Num_GPR_Regs) {
2244 if (FPR_idx != Num_FPR_Regs) {
2247 if (ObjectVT == MVT::f32)
2248 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2250 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2252 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2256 ArgSize = PtrByteSize;
2265 // Note that vector arguments in registers don't reserve stack space,
2266 // except in varargs functions.
2267 if (VR_idx != Num_VR_Regs) {
2268 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2269 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2271 while ((ArgOffset % 16) != 0) {
2272 ArgOffset += PtrByteSize;
2273 if (GPR_idx != Num_GPR_Regs)
2277 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2281 // Vectors are aligned.
2282 ArgOffset = ((ArgOffset+15)/16)*16;
2283 CurArgOffset = ArgOffset;
2290 // We need to load the argument to a virtual register if we determined
2291 // above that we ran out of physical registers of the appropriate type.
2293 int FI = MFI->CreateFixedObject(ObjSize,
2294 CurArgOffset + (ArgSize - ObjSize),
2296 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2297 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2298 false, false, false, 0);
2301 InVals.push_back(ArgVal);
2304 // Set the size that is at least reserved in caller of this function. Tail
2305 // call optimized functions' reserved stack space needs to be aligned so that
2306 // taking the difference between two stack areas will result in an aligned
2308 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2310 // If the function takes variable number of arguments, make a frame index for
2311 // the start of the first vararg value... for expansion of llvm.va_start.
2313 int Depth = ArgOffset;
2315 FuncInfo->setVarArgsFrameIndex(
2316 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2317 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2319 // If this function is vararg, store any remaining integer argument regs
2320 // to their spots on the stack so that they may be loaded by deferencing the
2321 // result of va_next.
2322 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2323 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2324 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2325 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2326 MachinePointerInfo(), false, false, 0);
2327 MemOps.push_back(Store);
2328 // Increment the address by four for the next argument to store
2329 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2330 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2334 if (!MemOps.empty())
2335 Chain = DAG.getNode(ISD::TokenFactor, dl,
2336 MVT::Other, &MemOps[0], MemOps.size());
2342 PPCTargetLowering::LowerFormalArguments_Darwin(
2344 CallingConv::ID CallConv, bool isVarArg,
2345 const SmallVectorImpl<ISD::InputArg>
2347 DebugLoc dl, SelectionDAG &DAG,
2348 SmallVectorImpl<SDValue> &InVals) const {
2349 // TODO: add description of PPC stack frame format, or at least some docs.
2351 MachineFunction &MF = DAG.getMachineFunction();
2352 MachineFrameInfo *MFI = MF.getFrameInfo();
2353 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2355 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2356 bool isPPC64 = PtrVT == MVT::i64;
2357 // Potential tail calls could cause overwriting of argument stack slots.
2358 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2359 (CallConv == CallingConv::Fast));
2360 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2362 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2363 // Area that is at least reserved in caller of this function.
2364 unsigned MinReservedArea = ArgOffset;
2366 static const uint16_t GPR_32[] = { // 32-bit registers.
2367 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2368 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2370 static const uint16_t GPR_64[] = { // 64-bit registers.
2371 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2372 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2375 static const uint16_t *FPR = GetFPR();
2377 static const uint16_t VR[] = {
2378 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2379 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2382 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2383 const unsigned Num_FPR_Regs = 13;
2384 const unsigned Num_VR_Regs = array_lengthof( VR);
2386 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2388 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2390 // In 32-bit non-varargs functions, the stack space for vectors is after the
2391 // stack space for non-vectors. We do not use this space unless we have
2392 // too many vectors to fit in registers, something that only occurs in
2393 // constructed examples:), but we have to walk the arglist to figure
2394 // that out...for the pathological case, compute VecArgOffset as the
2395 // start of the vector parameter area. Computing VecArgOffset is the
2396 // entire point of the following loop.
2397 unsigned VecArgOffset = ArgOffset;
2398 if (!isVarArg && !isPPC64) {
2399 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2401 EVT ObjectVT = Ins[ArgNo].VT;
2402 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2404 if (Flags.isByVal()) {
2405 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2406 unsigned ObjSize = Flags.getByValSize();
2408 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2409 VecArgOffset += ArgSize;
2413 switch(ObjectVT.getSimpleVT().SimpleTy) {
2414 default: llvm_unreachable("Unhandled argument type!");
2419 case MVT::i64: // PPC64
2421 // FIXME: We are guaranteed to be !isPPC64 at this point.
2422 // Does MVT::i64 apply?
2429 // Nothing to do, we're only looking at Nonvector args here.
2434 // We've found where the vector parameter area in memory is. Skip the
2435 // first 12 parameters; these don't use that memory.
2436 VecArgOffset = ((VecArgOffset+15)/16)*16;
2437 VecArgOffset += 12*16;
2439 // Add DAG nodes to load the arguments or copy them out of registers. On
2440 // entry to a function on PPC, the arguments start after the linkage area,
2441 // although the first ones are often in registers.
2443 SmallVector<SDValue, 8> MemOps;
2444 unsigned nAltivecParamsAtEnd = 0;
2445 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2446 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2448 bool needsLoad = false;
2449 EVT ObjectVT = Ins[ArgNo].VT;
2450 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2451 unsigned ArgSize = ObjSize;
2452 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2454 unsigned CurArgOffset = ArgOffset;
2456 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2457 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2458 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2459 if (isVarArg || isPPC64) {
2460 MinReservedArea = ((MinReservedArea+15)/16)*16;
2461 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2464 } else nAltivecParamsAtEnd++;
2466 // Calculate min reserved area.
2467 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2471 // FIXME the codegen can be much improved in some cases.
2472 // We do not have to keep everything in memory.
2473 if (Flags.isByVal()) {
2474 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2475 ObjSize = Flags.getByValSize();
2476 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2477 // Objects of size 1 and 2 are right justified, everything else is
2478 // left justified. This means the memory address is adjusted forwards.
2479 if (ObjSize==1 || ObjSize==2) {
2480 CurArgOffset = CurArgOffset + (4 - ObjSize);
2482 // The value of the object is its address.
2483 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2484 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2485 InVals.push_back(FIN);
2486 if (ObjSize==1 || ObjSize==2) {
2487 if (GPR_idx != Num_GPR_Regs) {
2490 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2492 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2493 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2494 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2495 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2496 MachinePointerInfo(FuncArg,
2498 ObjType, false, false, 0);
2499 MemOps.push_back(Store);
2503 ArgOffset += PtrByteSize;
2507 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2508 // Store whatever pieces of the object are in registers
2509 // to memory. ArgOffset will be the address of the beginning
2511 if (GPR_idx != Num_GPR_Regs) {
2514 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2516 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2517 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2518 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2519 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2520 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2521 MachinePointerInfo(FuncArg, ArgOffset),
2523 MemOps.push_back(Store);
2525 ArgOffset += PtrByteSize;
2527 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2534 switch (ObjectVT.getSimpleVT().SimpleTy) {
2535 default: llvm_unreachable("Unhandled argument type!");
2538 if (GPR_idx != Num_GPR_Regs) {
2539 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2540 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2544 ArgSize = PtrByteSize;
2546 // All int arguments reserve stack space in the Darwin ABI.
2547 ArgOffset += PtrByteSize;
2551 case MVT::i64: // PPC64
2552 if (GPR_idx != Num_GPR_Regs) {
2553 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2554 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2556 if (ObjectVT == MVT::i32)
2557 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2558 // value to MVT::i64 and then truncate to the correct register size.
2559 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2564 ArgSize = PtrByteSize;
2566 // All int arguments reserve stack space in the Darwin ABI.
2572 // Every 4 bytes of argument space consumes one of the GPRs available for
2573 // argument passing.
2574 if (GPR_idx != Num_GPR_Regs) {
2576 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2579 if (FPR_idx != Num_FPR_Regs) {
2582 if (ObjectVT == MVT::f32)
2583 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2585 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2587 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2593 // All FP arguments reserve stack space in the Darwin ABI.
2594 ArgOffset += isPPC64 ? 8 : ObjSize;
2600 // Note that vector arguments in registers don't reserve stack space,
2601 // except in varargs functions.
2602 if (VR_idx != Num_VR_Regs) {
2603 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2604 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2606 while ((ArgOffset % 16) != 0) {
2607 ArgOffset += PtrByteSize;
2608 if (GPR_idx != Num_GPR_Regs)
2612 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2616 if (!isVarArg && !isPPC64) {
2617 // Vectors go after all the nonvectors.
2618 CurArgOffset = VecArgOffset;
2621 // Vectors are aligned.
2622 ArgOffset = ((ArgOffset+15)/16)*16;
2623 CurArgOffset = ArgOffset;
2631 // We need to load the argument to a virtual register if we determined above
2632 // that we ran out of physical registers of the appropriate type.
2634 int FI = MFI->CreateFixedObject(ObjSize,
2635 CurArgOffset + (ArgSize - ObjSize),
2637 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2638 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2639 false, false, false, 0);
2642 InVals.push_back(ArgVal);
2645 // Set the size that is at least reserved in caller of this function. Tail
2646 // call optimized functions' reserved stack space needs to be aligned so that
2647 // taking the difference between two stack areas will result in an aligned
2649 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2651 // If the function takes variable number of arguments, make a frame index for
2652 // the start of the first vararg value... for expansion of llvm.va_start.
2654 int Depth = ArgOffset;
2656 FuncInfo->setVarArgsFrameIndex(
2657 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2659 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2661 // If this function is vararg, store any remaining integer argument regs
2662 // to their spots on the stack so that they may be loaded by deferencing the
2663 // result of va_next.
2664 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2668 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2670 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2672 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2673 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2674 MachinePointerInfo(), false, false, 0);
2675 MemOps.push_back(Store);
2676 // Increment the address by four for the next argument to store
2677 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2678 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2682 if (!MemOps.empty())
2683 Chain = DAG.getNode(ISD::TokenFactor, dl,
2684 MVT::Other, &MemOps[0], MemOps.size());
2689 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2690 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2692 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2696 const SmallVectorImpl<ISD::OutputArg>
2698 const SmallVectorImpl<SDValue> &OutVals,
2699 unsigned &nAltivecParamsAtEnd) {
2700 // Count how many bytes are to be pushed on the stack, including the linkage
2701 // area, and parameter passing area. We start with 24/48 bytes, which is
2702 // prereserved space for [SP][CR][LR][3 x unused].
2703 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2704 unsigned NumOps = Outs.size();
2705 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2707 // Add up all the space actually used.
2708 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2709 // they all go in registers, but we must reserve stack space for them for
2710 // possible use by the caller. In varargs or 64-bit calls, parameters are
2711 // assigned stack space in order, with padding so Altivec parameters are
2713 nAltivecParamsAtEnd = 0;
2714 for (unsigned i = 0; i != NumOps; ++i) {
2715 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2716 EVT ArgVT = Outs[i].VT;
2717 // Varargs Altivec parameters are padded to a 16 byte boundary.
2718 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2719 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2720 if (!isVarArg && !isPPC64) {
2721 // Non-varargs Altivec parameters go after all the non-Altivec
2722 // parameters; handle those later so we know how much padding we need.
2723 nAltivecParamsAtEnd++;
2726 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2727 NumBytes = ((NumBytes+15)/16)*16;
2729 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2732 // Allow for Altivec parameters at the end, if needed.
2733 if (nAltivecParamsAtEnd) {
2734 NumBytes = ((NumBytes+15)/16)*16;
2735 NumBytes += 16*nAltivecParamsAtEnd;
2738 // The prolog code of the callee may store up to 8 GPR argument registers to
2739 // the stack, allowing va_start to index over them in memory if its varargs.
2740 // Because we cannot tell if this is needed on the caller side, we have to
2741 // conservatively assume that it is needed. As such, make sure we have at
2742 // least enough stack space for the caller to store the 8 GPRs.
2743 NumBytes = std::max(NumBytes,
2744 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2746 // Tail call needs the stack to be aligned.
2747 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2748 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2749 getFrameLowering()->getStackAlignment();
2750 unsigned AlignMask = TargetAlign-1;
2751 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2757 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2758 /// adjusted to accommodate the arguments for the tailcall.
2759 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2760 unsigned ParamSize) {
2762 if (!isTailCall) return 0;
2764 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2765 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2766 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2767 // Remember only if the new adjustement is bigger.
2768 if (SPDiff < FI->getTailCallSPDelta())
2769 FI->setTailCallSPDelta(SPDiff);
2774 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2775 /// for tail call optimization. Targets which want to do tail call
2776 /// optimization should implement this function.
2778 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2779 CallingConv::ID CalleeCC,
2781 const SmallVectorImpl<ISD::InputArg> &Ins,
2782 SelectionDAG& DAG) const {
2783 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2786 // Variable argument functions are not supported.
2790 MachineFunction &MF = DAG.getMachineFunction();
2791 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2792 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2793 // Functions containing by val parameters are not supported.
2794 for (unsigned i = 0; i != Ins.size(); i++) {
2795 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2796 if (Flags.isByVal()) return false;
2799 // Non PIC/GOT tail calls are supported.
2800 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2803 // At the moment we can only do local tail calls (in same module, hidden
2804 // or protected) if we are generating PIC.
2805 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2806 return G->getGlobal()->hasHiddenVisibility()
2807 || G->getGlobal()->hasProtectedVisibility();
2813 /// isCallCompatibleAddress - Return the immediate to use if the specified
2814 /// 32-bit value is representable in the immediate field of a BxA instruction.
2815 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2816 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2819 int Addr = C->getZExtValue();
2820 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2821 SignExtend32<26>(Addr) != Addr)
2822 return 0; // Top 6 bits have to be sext of immediate.
2824 return DAG.getConstant((int)C->getZExtValue() >> 2,
2825 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2830 struct TailCallArgumentInfo {
2835 TailCallArgumentInfo() : FrameIdx(0) {}
2840 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2842 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2844 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2845 SmallVector<SDValue, 8> &MemOpChains,
2847 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2848 SDValue Arg = TailCallArgs[i].Arg;
2849 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2850 int FI = TailCallArgs[i].FrameIdx;
2851 // Store relative to framepointer.
2852 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2853 MachinePointerInfo::getFixedStack(FI),
2858 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2859 /// the appropriate stack slot for the tail call optimized function call.
2860 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2861 MachineFunction &MF,
2870 // Calculate the new stack slot for the return address.
2871 int SlotSize = isPPC64 ? 8 : 4;
2872 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2874 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2875 NewRetAddrLoc, true);
2876 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2877 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2878 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2879 MachinePointerInfo::getFixedStack(NewRetAddr),
2882 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2883 // slot as the FP is never overwritten.
2886 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2887 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2889 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2890 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2891 MachinePointerInfo::getFixedStack(NewFPIdx),
2898 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2899 /// the position of the argument.
2901 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2902 SDValue Arg, int SPDiff, unsigned ArgOffset,
2903 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2904 int Offset = ArgOffset + SPDiff;
2905 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2906 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2907 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2908 SDValue FIN = DAG.getFrameIndex(FI, VT);
2909 TailCallArgumentInfo Info;
2911 Info.FrameIdxOp = FIN;
2913 TailCallArguments.push_back(Info);
2916 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2917 /// stack slot. Returns the chain as result and the loaded frame pointers in
2918 /// LROpOut/FPOpout. Used when tail calling.
2919 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2925 DebugLoc dl) const {
2927 // Load the LR and FP stack slot for later adjusting.
2928 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2929 LROpOut = getReturnAddrFrameIndex(DAG);
2930 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2931 false, false, false, 0);
2932 Chain = SDValue(LROpOut.getNode(), 1);
2934 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2935 // slot as the FP is never overwritten.
2937 FPOpOut = getFramePointerFrameIndex(DAG);
2938 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2939 false, false, false, 0);
2940 Chain = SDValue(FPOpOut.getNode(), 1);
2946 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2947 /// by "Src" to address "Dst" of size "Size". Alignment information is
2948 /// specified by the specific parameter attribute. The copy will be passed as
2949 /// a byval function parameter.
2950 /// Sometimes what we are copying is the end of a larger object, the part that
2951 /// does not fit in registers.
2953 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2954 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2956 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2957 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2958 false, false, MachinePointerInfo(0),
2959 MachinePointerInfo(0));
2962 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2965 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2966 SDValue Arg, SDValue PtrOff, int SPDiff,
2967 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2968 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2969 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2971 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2976 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2978 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2979 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2980 DAG.getConstant(ArgOffset, PtrVT));
2982 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2983 MachinePointerInfo(), false, false, 0));
2984 // Calculate and remember argument location.
2985 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2990 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2991 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2992 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2993 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2994 MachineFunction &MF = DAG.getMachineFunction();
2996 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2997 // might overwrite each other in case of tail call optimization.
2998 SmallVector<SDValue, 8> MemOpChains2;
2999 // Do not flag preceding copytoreg stuff together with the following stuff.
3001 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
3003 if (!MemOpChains2.empty())
3004 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3005 &MemOpChains2[0], MemOpChains2.size());
3007 // Store the return address to the appropriate stack slot.
3008 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
3009 isPPC64, isDarwinABI, dl);
3011 // Emit callseq_end just before tailcall node.
3012 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3013 DAG.getIntPtrConstant(0, true), InFlag);
3014 InFlag = Chain.getValue(1);
3018 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
3019 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
3020 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
3021 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
3022 const PPCSubtarget &PPCSubTarget) {
3024 bool isPPC64 = PPCSubTarget.isPPC64();
3025 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3027 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3028 NodeTys.push_back(MVT::Other); // Returns a chain
3029 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
3031 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
3033 bool needIndirectCall = true;
3034 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3035 // If this is an absolute destination address, use the munged value.
3036 Callee = SDValue(Dest, 0);
3037 needIndirectCall = false;
3040 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3041 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
3042 // Use indirect calls for ALL functions calls in JIT mode, since the
3043 // far-call stubs may be outside relocation limits for a BL instruction.
3044 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
3045 unsigned OpFlags = 0;
3046 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3047 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3048 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3049 (G->getGlobal()->isDeclaration() ||
3050 G->getGlobal()->isWeakForLinker())) {
3051 // PC-relative references to external symbols should go through $stub,
3052 // unless we're building with the leopard linker or later, which
3053 // automatically synthesizes these stubs.
3054 OpFlags = PPCII::MO_DARWIN_STUB;
3057 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3058 // every direct call is) turn it into a TargetGlobalAddress /
3059 // TargetExternalSymbol node so that legalize doesn't hack it.
3060 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3061 Callee.getValueType(),
3063 needIndirectCall = false;
3067 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3068 unsigned char OpFlags = 0;
3070 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3071 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3072 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3073 // PC-relative references to external symbols should go through $stub,
3074 // unless we're building with the leopard linker or later, which
3075 // automatically synthesizes these stubs.
3076 OpFlags = PPCII::MO_DARWIN_STUB;
3079 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3081 needIndirectCall = false;
3084 if (needIndirectCall) {
3085 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3086 // to do the call, we can't use PPCISD::CALL.
3087 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3089 if (isSVR4ABI && isPPC64) {
3090 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3091 // entry point, but to the function descriptor (the function entry point
3092 // address is part of the function descriptor though).
3093 // The function descriptor is a three doubleword structure with the
3094 // following fields: function entry point, TOC base address and
3095 // environment pointer.
3096 // Thus for a call through a function pointer, the following actions need
3098 // 1. Save the TOC of the caller in the TOC save area of its stack
3099 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3100 // 2. Load the address of the function entry point from the function
3102 // 3. Load the TOC of the callee from the function descriptor into r2.
3103 // 4. Load the environment pointer from the function descriptor into
3105 // 5. Branch to the function entry point address.
3106 // 6. On return of the callee, the TOC of the caller needs to be
3107 // restored (this is done in FinishCall()).
3109 // All those operations are flagged together to ensure that no other
3110 // operations can be scheduled in between. E.g. without flagging the
3111 // operations together, a TOC access in the caller could be scheduled
3112 // between the load of the callee TOC and the branch to the callee, which
3113 // results in the TOC access going through the TOC of the callee instead
3114 // of going through the TOC of the caller, which leads to incorrect code.
3116 // Load the address of the function entry point from the function
3118 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3119 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3120 InFlag.getNode() ? 3 : 2);
3121 Chain = LoadFuncPtr.getValue(1);
3122 InFlag = LoadFuncPtr.getValue(2);
3124 // Load environment pointer into r11.
3125 // Offset of the environment pointer within the function descriptor.
3126 SDValue PtrOff = DAG.getIntPtrConstant(16);
3128 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3129 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3131 Chain = LoadEnvPtr.getValue(1);
3132 InFlag = LoadEnvPtr.getValue(2);
3134 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3136 Chain = EnvVal.getValue(0);
3137 InFlag = EnvVal.getValue(1);
3139 // Load TOC of the callee into r2. We are using a target-specific load
3140 // with r2 hard coded, because the result of a target-independent load
3141 // would never go directly into r2, since r2 is a reserved register (which
3142 // prevents the register allocator from allocating it), resulting in an
3143 // additional register being allocated and an unnecessary move instruction
3145 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3146 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3148 Chain = LoadTOCPtr.getValue(0);
3149 InFlag = LoadTOCPtr.getValue(1);
3151 MTCTROps[0] = Chain;
3152 MTCTROps[1] = LoadFuncPtr;
3153 MTCTROps[2] = InFlag;
3156 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3157 2 + (InFlag.getNode() != 0));
3158 InFlag = Chain.getValue(1);
3161 NodeTys.push_back(MVT::Other);
3162 NodeTys.push_back(MVT::Glue);
3163 Ops.push_back(Chain);
3164 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3166 // Add CTR register as callee so a bctr can be emitted later.
3168 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3171 // If this is a direct call, pass the chain and the callee.
3172 if (Callee.getNode()) {
3173 Ops.push_back(Chain);
3174 Ops.push_back(Callee);
3176 // If this is a tail call add stack pointer delta.
3178 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3180 // Add argument registers to the end of the list so that they are known live
3182 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3183 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3184 RegsToPass[i].second.getValueType()));
3190 bool isLocalCall(const SDValue &Callee)
3192 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3193 return !G->getGlobal()->isDeclaration() &&
3194 !G->getGlobal()->isWeakForLinker();
3199 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3200 CallingConv::ID CallConv, bool isVarArg,
3201 const SmallVectorImpl<ISD::InputArg> &Ins,
3202 DebugLoc dl, SelectionDAG &DAG,
3203 SmallVectorImpl<SDValue> &InVals) const {
3205 SmallVector<CCValAssign, 16> RVLocs;
3206 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3207 getTargetMachine(), RVLocs, *DAG.getContext());
3208 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3210 // Copy all of the result registers out of their specified physreg.
3211 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3212 CCValAssign &VA = RVLocs[i];
3213 assert(VA.isRegLoc() && "Can only return in registers!");
3215 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3216 VA.getLocReg(), VA.getLocVT(), InFlag);
3217 Chain = Val.getValue(1);
3218 InFlag = Val.getValue(2);
3220 switch (VA.getLocInfo()) {
3221 default: llvm_unreachable("Unknown loc info!");
3222 case CCValAssign::Full: break;
3223 case CCValAssign::AExt:
3224 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3226 case CCValAssign::ZExt:
3227 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3228 DAG.getValueType(VA.getValVT()));
3229 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3231 case CCValAssign::SExt:
3232 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3233 DAG.getValueType(VA.getValVT()));
3234 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3238 InVals.push_back(Val);
3245 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3246 bool isTailCall, bool isVarArg,
3248 SmallVector<std::pair<unsigned, SDValue>, 8>
3250 SDValue InFlag, SDValue Chain,
3252 int SPDiff, unsigned NumBytes,
3253 const SmallVectorImpl<ISD::InputArg> &Ins,
3254 SmallVectorImpl<SDValue> &InVals) const {
3255 std::vector<EVT> NodeTys;
3256 SmallVector<SDValue, 8> Ops;
3257 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3258 isTailCall, RegsToPass, Ops, NodeTys,
3261 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3262 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3263 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3265 // When performing tail call optimization the callee pops its arguments off
3266 // the stack. Account for this here so these bytes can be pushed back on in
3267 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3268 int BytesCalleePops =
3269 (CallConv == CallingConv::Fast &&
3270 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3272 // Add a register mask operand representing the call-preserved registers.
3273 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3274 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3275 assert(Mask && "Missing call preserved mask for calling convention");
3276 Ops.push_back(DAG.getRegisterMask(Mask));
3278 if (InFlag.getNode())
3279 Ops.push_back(InFlag);
3283 // If this is the first return lowered for this function, add the regs
3284 // to the liveout set for the function.
3285 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3286 SmallVector<CCValAssign, 16> RVLocs;
3287 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3288 getTargetMachine(), RVLocs, *DAG.getContext());
3289 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3290 for (unsigned i = 0; i != RVLocs.size(); ++i)
3291 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3294 assert(((Callee.getOpcode() == ISD::Register &&
3295 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3296 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3297 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3298 isa<ConstantSDNode>(Callee)) &&
3299 "Expecting an global address, external symbol, absolute value or register");
3301 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3304 // Add a NOP immediately after the branch instruction when using the 64-bit
3305 // SVR4 ABI. At link time, if caller and callee are in a different module and
3306 // thus have a different TOC, the call will be replaced with a call to a stub
3307 // function which saves the current TOC, loads the TOC of the callee and
3308 // branches to the callee. The NOP will be replaced with a load instruction
3309 // which restores the TOC of the caller from the TOC save slot of the current
3310 // stack frame. If caller and callee belong to the same module (and have the
3311 // same TOC), the NOP will remain unchanged.
3313 bool needsTOCRestore = false;
3314 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3315 if (CallOpc == PPCISD::BCTRL_SVR4) {
3316 // This is a call through a function pointer.
3317 // Restore the caller TOC from the save area into R2.
3318 // See PrepareCall() for more information about calls through function
3319 // pointers in the 64-bit SVR4 ABI.
3320 // We are using a target-specific load with r2 hard coded, because the
3321 // result of a target-independent load would never go directly into r2,
3322 // since r2 is a reserved register (which prevents the register allocator
3323 // from allocating it), resulting in an additional register being
3324 // allocated and an unnecessary move instruction being generated.
3325 needsTOCRestore = true;
3326 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3327 // Otherwise insert NOP for non-local calls.
3328 CallOpc = PPCISD::CALL_NOP_SVR4;
3332 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3333 InFlag = Chain.getValue(1);
3335 if (needsTOCRestore) {
3336 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3337 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3338 InFlag = Chain.getValue(1);
3341 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3342 DAG.getIntPtrConstant(BytesCalleePops, true),
3345 InFlag = Chain.getValue(1);
3347 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3348 Ins, dl, DAG, InVals);
3352 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3353 SmallVectorImpl<SDValue> &InVals) const {
3354 SelectionDAG &DAG = CLI.DAG;
3355 DebugLoc &dl = CLI.DL;
3356 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3357 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3358 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3359 SDValue Chain = CLI.Chain;
3360 SDValue Callee = CLI.Callee;
3361 bool &isTailCall = CLI.IsTailCall;
3362 CallingConv::ID CallConv = CLI.CallConv;
3363 bool isVarArg = CLI.IsVarArg;
3366 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3369 if (PPCSubTarget.isSVR4ABI()) {
3370 if (PPCSubTarget.isPPC64())
3371 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3372 isTailCall, Outs, OutVals, Ins,
3375 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3376 isTailCall, Outs, OutVals, Ins,
3380 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3381 isTailCall, Outs, OutVals, Ins,
3386 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3387 CallingConv::ID CallConv, bool isVarArg,
3389 const SmallVectorImpl<ISD::OutputArg> &Outs,
3390 const SmallVectorImpl<SDValue> &OutVals,
3391 const SmallVectorImpl<ISD::InputArg> &Ins,
3392 DebugLoc dl, SelectionDAG &DAG,
3393 SmallVectorImpl<SDValue> &InVals) const {
3394 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3395 // of the 32-bit SVR4 ABI stack frame layout.
3397 assert((CallConv == CallingConv::C ||
3398 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3400 unsigned PtrByteSize = 4;
3402 MachineFunction &MF = DAG.getMachineFunction();
3404 // Mark this function as potentially containing a function that contains a
3405 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3406 // and restoring the callers stack pointer in this functions epilog. This is
3407 // done because by tail calling the called function might overwrite the value
3408 // in this function's (MF) stack pointer stack slot 0(SP).
3409 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3410 CallConv == CallingConv::Fast)
3411 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3413 // Count how many bytes are to be pushed on the stack, including the linkage
3414 // area, parameter list area and the part of the local variable space which
3415 // contains copies of aggregates which are passed by value.
3417 // Assign locations to all of the outgoing arguments.
3418 SmallVector<CCValAssign, 16> ArgLocs;
3419 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3420 getTargetMachine(), ArgLocs, *DAG.getContext());
3422 // Reserve space for the linkage area on the stack.
3423 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3426 // Handle fixed and variable vector arguments differently.
3427 // Fixed vector arguments go into registers as long as registers are
3428 // available. Variable vector arguments always go into memory.
3429 unsigned NumArgs = Outs.size();
3431 for (unsigned i = 0; i != NumArgs; ++i) {
3432 MVT ArgVT = Outs[i].VT;
3433 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3436 if (Outs[i].IsFixed) {
3437 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3440 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3446 errs() << "Call operand #" << i << " has unhandled type "
3447 << EVT(ArgVT).getEVTString() << "\n";
3449 llvm_unreachable(0);
3453 // All arguments are treated the same.
3454 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
3457 // Assign locations to all of the outgoing aggregate by value arguments.
3458 SmallVector<CCValAssign, 16> ByValArgLocs;
3459 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3460 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3462 // Reserve stack space for the allocations in CCInfo.
3463 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3465 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
3467 // Size of the linkage area, parameter list area and the part of the local
3468 // space variable where copies of aggregates which are passed by value are
3470 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3472 // Calculate by how many bytes the stack has to be adjusted in case of tail
3473 // call optimization.
3474 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3476 // Adjust the stack pointer for the new arguments...
3477 // These operations are automatically eliminated by the prolog/epilog pass
3478 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3479 SDValue CallSeqStart = Chain;
3481 // Load the return address and frame pointer so it can be moved somewhere else
3484 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3487 // Set up a copy of the stack pointer for use loading and storing any
3488 // arguments that may not fit in the registers available for argument
3490 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3492 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3493 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3494 SmallVector<SDValue, 8> MemOpChains;
3496 bool seenFloatArg = false;
3497 // Walk the register/memloc assignments, inserting copies/loads.
3498 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3501 CCValAssign &VA = ArgLocs[i];
3502 SDValue Arg = OutVals[i];
3503 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3505 if (Flags.isByVal()) {
3506 // Argument is an aggregate which is passed by value, thus we need to
3507 // create a copy of it in the local variable space of the current stack
3508 // frame (which is the stack frame of the caller) and pass the address of
3509 // this copy to the callee.
3510 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3511 CCValAssign &ByValVA = ByValArgLocs[j++];
3512 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3514 // Memory reserved in the local variable space of the callers stack frame.
3515 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3517 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3518 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3520 // Create a copy of the argument in the local area of the current
3522 SDValue MemcpyCall =
3523 CreateCopyOfByValArgument(Arg, PtrOff,
3524 CallSeqStart.getNode()->getOperand(0),
3527 // This must go outside the CALLSEQ_START..END.
3528 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3529 CallSeqStart.getNode()->getOperand(1));
3530 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3531 NewCallSeqStart.getNode());
3532 Chain = CallSeqStart = NewCallSeqStart;
3534 // Pass the address of the aggregate copy on the stack either in a
3535 // physical register or in the parameter list area of the current stack
3536 // frame to the callee.
3540 if (VA.isRegLoc()) {
3541 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3542 // Put argument in a physical register.
3543 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3545 // Put argument in the parameter list area of the current stack frame.
3546 assert(VA.isMemLoc());
3547 unsigned LocMemOffset = VA.getLocMemOffset();
3550 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3551 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3553 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3554 MachinePointerInfo(),
3557 // Calculate and remember argument location.
3558 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3564 if (!MemOpChains.empty())
3565 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3566 &MemOpChains[0], MemOpChains.size());
3568 // Build a sequence of copy-to-reg nodes chained together with token chain
3569 // and flag operands which copy the outgoing args into the appropriate regs.
3571 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3572 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3573 RegsToPass[i].second, InFlag);
3574 InFlag = Chain.getValue(1);
3577 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3580 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3581 SDValue Ops[] = { Chain, InFlag };
3583 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3584 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3586 InFlag = Chain.getValue(1);
3590 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3591 false, TailCallArguments);
3593 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3594 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3598 // Copy an argument into memory, being careful to do this outside the
3599 // call sequence for the call to which the argument belongs.
3601 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3602 SDValue CallSeqStart,
3603 ISD::ArgFlagsTy Flags,
3605 DebugLoc dl) const {
3606 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3607 CallSeqStart.getNode()->getOperand(0),
3609 // The MEMCPY must go outside the CALLSEQ_START..END.
3610 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3611 CallSeqStart.getNode()->getOperand(1));
3612 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3613 NewCallSeqStart.getNode());
3614 return NewCallSeqStart;
3618 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3619 CallingConv::ID CallConv, bool isVarArg,
3621 const SmallVectorImpl<ISD::OutputArg> &Outs,
3622 const SmallVectorImpl<SDValue> &OutVals,
3623 const SmallVectorImpl<ISD::InputArg> &Ins,
3624 DebugLoc dl, SelectionDAG &DAG,
3625 SmallVectorImpl<SDValue> &InVals) const {
3627 unsigned NumOps = Outs.size();
3629 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3630 unsigned PtrByteSize = 8;
3632 MachineFunction &MF = DAG.getMachineFunction();
3634 // Mark this function as potentially containing a function that contains a
3635 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3636 // and restoring the callers stack pointer in this functions epilog. This is
3637 // done because by tail calling the called function might overwrite the value
3638 // in this function's (MF) stack pointer stack slot 0(SP).
3639 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3640 CallConv == CallingConv::Fast)
3641 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3643 unsigned nAltivecParamsAtEnd = 0;
3645 // Count how many bytes are to be pushed on the stack, including the linkage
3646 // area, and parameter passing area. We start with at least 48 bytes, which
3647 // is reserved space for [SP][CR][LR][3 x unused].
3648 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3651 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3652 Outs, OutVals, nAltivecParamsAtEnd);
3654 // Calculate by how many bytes the stack has to be adjusted in case of tail
3655 // call optimization.
3656 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3658 // To protect arguments on the stack from being clobbered in a tail call,
3659 // force all the loads to happen before doing any other lowering.
3661 Chain = DAG.getStackArgumentTokenFactor(Chain);
3663 // Adjust the stack pointer for the new arguments...
3664 // These operations are automatically eliminated by the prolog/epilog pass
3665 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3666 SDValue CallSeqStart = Chain;
3668 // Load the return address and frame pointer so it can be move somewhere else
3671 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3674 // Set up a copy of the stack pointer for use loading and storing any
3675 // arguments that may not fit in the registers available for argument
3677 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3679 // Figure out which arguments are going to go in registers, and which in
3680 // memory. Also, if this is a vararg function, floating point operations
3681 // must be stored to our stack, and loaded into integer regs as well, if
3682 // any integer regs are available for argument passing.
3683 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3684 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3686 static const uint16_t GPR[] = {
3687 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3688 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3690 static const uint16_t *FPR = GetFPR();
3692 static const uint16_t VR[] = {
3693 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3694 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3696 const unsigned NumGPRs = array_lengthof(GPR);
3697 const unsigned NumFPRs = 13;
3698 const unsigned NumVRs = array_lengthof(VR);
3700 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3701 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3703 SmallVector<SDValue, 8> MemOpChains;
3704 for (unsigned i = 0; i != NumOps; ++i) {
3705 SDValue Arg = OutVals[i];
3706 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3708 // PtrOff will be used to store the current argument to the stack if a
3709 // register cannot be found for it.
3712 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3714 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3716 // Promote integers to 64-bit values.
3717 if (Arg.getValueType() == MVT::i32) {
3718 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3719 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3720 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3723 // FIXME memcpy is used way more than necessary. Correctness first.
3724 // Note: "by value" is code for passing a structure by value, not
3726 if (Flags.isByVal()) {
3727 // Note: Size includes alignment padding, so
3728 // struct x { short a; char b; }
3729 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3730 // These are the proper values we need for right-justifying the
3731 // aggregate in a parameter register.
3732 unsigned Size = Flags.getByValSize();
3734 // An empty aggregate parameter takes up no storage and no
3739 // All aggregates smaller than 8 bytes must be passed right-justified.
3740 if (Size==1 || Size==2 || Size==4) {
3741 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3742 if (GPR_idx != NumGPRs) {
3743 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3744 MachinePointerInfo(), VT,
3746 MemOpChains.push_back(Load.getValue(1));
3747 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3749 ArgOffset += PtrByteSize;
3754 if (GPR_idx == NumGPRs && Size < 8) {
3755 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3756 PtrOff.getValueType());
3757 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3758 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3761 ArgOffset += PtrByteSize;
3764 // Copy entire object into memory. There are cases where gcc-generated
3765 // code assumes it is there, even if it could be put entirely into
3766 // registers. (This is not what the doc says.)
3768 // FIXME: The above statement is likely due to a misunderstanding of the
3769 // documents. All arguments must be copied into the parameter area BY
3770 // THE CALLEE in the event that the callee takes the address of any
3771 // formal argument. That has not yet been implemented. However, it is
3772 // reasonable to use the stack area as a staging area for the register
3775 // Skip this for small aggregates, as we will use the same slot for a
3776 // right-justified copy, below.
3778 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3782 // When a register is available, pass a small aggregate right-justified.
3783 if (Size < 8 && GPR_idx != NumGPRs) {
3784 // The easiest way to get this right-justified in a register
3785 // is to copy the structure into the rightmost portion of a
3786 // local variable slot, then load the whole slot into the
3788 // FIXME: The memcpy seems to produce pretty awful code for
3789 // small aggregates, particularly for packed ones.
3790 // FIXME: It would be preferable to use the slot in the
3791 // parameter save area instead of a new local variable.
3792 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3793 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3794 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3798 // Load the slot into the register.
3799 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
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));
3805 // Done with this argument.
3806 ArgOffset += PtrByteSize;
3810 // For aggregates larger than PtrByteSize, copy the pieces of the
3811 // object that fit into registers from the parameter save area.
3812 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3813 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3814 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3815 if (GPR_idx != NumGPRs) {
3816 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3817 MachinePointerInfo(),
3818 false, false, false, 0);
3819 MemOpChains.push_back(Load.getValue(1));
3820 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3821 ArgOffset += PtrByteSize;
3823 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3830 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3831 default: llvm_unreachable("Unexpected ValueType for argument!");
3834 if (GPR_idx != NumGPRs) {
3835 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3837 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3838 true, isTailCall, false, MemOpChains,
3839 TailCallArguments, dl);
3841 ArgOffset += PtrByteSize;
3845 if (FPR_idx != NumFPRs) {
3846 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3849 // A single float or an aggregate containing only a single float
3850 // must be passed right-justified in the stack doubleword, and
3851 // in the GPR, if one is available.
3853 if (Arg.getValueType().getSimpleVT().SimpleTy == MVT::f32) {
3854 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3855 StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3859 SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
3860 MachinePointerInfo(), false, false, 0);
3861 MemOpChains.push_back(Store);
3863 // Float varargs are always shadowed in available integer registers
3864 if (GPR_idx != NumGPRs) {
3865 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3866 MachinePointerInfo(), false, false,
3868 MemOpChains.push_back(Load.getValue(1));
3869 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3871 } else if (GPR_idx != NumGPRs)
3872 // If we have any FPRs remaining, we may also have GPRs remaining.
3875 // Single-precision floating-point values are mapped to the
3876 // second (rightmost) word of the stack doubleword.
3877 if (Arg.getValueType() == MVT::f32) {
3878 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3879 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3882 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3883 true, isTailCall, false, MemOpChains,
3884 TailCallArguments, dl);
3893 // These go aligned on the stack, or in the corresponding R registers
3894 // when within range. The Darwin PPC ABI doc claims they also go in
3895 // V registers; in fact gcc does this only for arguments that are
3896 // prototyped, not for those that match the ... We do it for all
3897 // arguments, seems to work.
3898 while (ArgOffset % 16 !=0) {
3899 ArgOffset += PtrByteSize;
3900 if (GPR_idx != NumGPRs)
3903 // We could elide this store in the case where the object fits
3904 // entirely in R registers. Maybe later.
3905 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3906 DAG.getConstant(ArgOffset, PtrVT));
3907 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3908 MachinePointerInfo(), false, false, 0);
3909 MemOpChains.push_back(Store);
3910 if (VR_idx != NumVRs) {
3911 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3912 MachinePointerInfo(),
3913 false, false, false, 0);
3914 MemOpChains.push_back(Load.getValue(1));
3915 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3918 for (unsigned i=0; i<16; i+=PtrByteSize) {
3919 if (GPR_idx == NumGPRs)
3921 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3922 DAG.getConstant(i, PtrVT));
3923 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3924 false, false, false, 0);
3925 MemOpChains.push_back(Load.getValue(1));
3926 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3931 // Non-varargs Altivec params generally go in registers, but have
3932 // stack space allocated at the end.
3933 if (VR_idx != NumVRs) {
3934 // Doesn't have GPR space allocated.
3935 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3937 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3938 true, isTailCall, true, MemOpChains,
3939 TailCallArguments, dl);
3946 if (!MemOpChains.empty())
3947 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3948 &MemOpChains[0], MemOpChains.size());
3950 // Check if this is an indirect call (MTCTR/BCTRL).
3951 // See PrepareCall() for more information about calls through function
3952 // pointers in the 64-bit SVR4 ABI.
3954 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3955 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3956 !isBLACompatibleAddress(Callee, DAG)) {
3957 // Load r2 into a virtual register and store it to the TOC save area.
3958 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3959 // TOC save area offset.
3960 SDValue PtrOff = DAG.getIntPtrConstant(40);
3961 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3962 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3964 // R12 must contain the address of an indirect callee. This does not
3965 // mean the MTCTR instruction must use R12; it's easier to model this
3966 // as an extra parameter, so do that.
3967 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
3970 // Build a sequence of copy-to-reg nodes chained together with token chain
3971 // and flag operands which copy the outgoing args into the appropriate regs.
3973 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3974 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3975 RegsToPass[i].second, InFlag);
3976 InFlag = Chain.getValue(1);
3980 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
3981 FPOp, true, TailCallArguments);
3983 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3984 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3989 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
3990 CallingConv::ID CallConv, bool isVarArg,
3992 const SmallVectorImpl<ISD::OutputArg> &Outs,
3993 const SmallVectorImpl<SDValue> &OutVals,
3994 const SmallVectorImpl<ISD::InputArg> &Ins,
3995 DebugLoc dl, SelectionDAG &DAG,
3996 SmallVectorImpl<SDValue> &InVals) const {
3998 unsigned NumOps = Outs.size();
4000 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4001 bool isPPC64 = PtrVT == MVT::i64;
4002 unsigned PtrByteSize = isPPC64 ? 8 : 4;
4004 MachineFunction &MF = DAG.getMachineFunction();
4006 // Mark this function as potentially containing a function that contains a
4007 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4008 // and restoring the callers stack pointer in this functions epilog. This is
4009 // done because by tail calling the called function might overwrite the value
4010 // in this function's (MF) stack pointer stack slot 0(SP).
4011 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4012 CallConv == CallingConv::Fast)
4013 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4015 unsigned nAltivecParamsAtEnd = 0;
4017 // Count how many bytes are to be pushed on the stack, including the linkage
4018 // area, and parameter passing area. We start with 24/48 bytes, which is
4019 // prereserved space for [SP][CR][LR][3 x unused].
4021 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
4023 nAltivecParamsAtEnd);
4025 // Calculate by how many bytes the stack has to be adjusted in case of tail
4026 // call optimization.
4027 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4029 // To protect arguments on the stack from being clobbered in a tail call,
4030 // force all the loads to happen before doing any other lowering.
4032 Chain = DAG.getStackArgumentTokenFactor(Chain);
4034 // Adjust the stack pointer for the new arguments...
4035 // These operations are automatically eliminated by the prolog/epilog pass
4036 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
4037 SDValue CallSeqStart = Chain;
4039 // Load the return address and frame pointer so it can be move somewhere else
4042 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4045 // Set up a copy of the stack pointer for use loading and storing any
4046 // arguments that may not fit in the registers available for argument
4050 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4052 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4054 // Figure out which arguments are going to go in registers, and which in
4055 // memory. Also, if this is a vararg function, floating point operations
4056 // must be stored to our stack, and loaded into integer regs as well, if
4057 // any integer regs are available for argument passing.
4058 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
4059 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4061 static const uint16_t GPR_32[] = { // 32-bit registers.
4062 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4063 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4065 static const uint16_t GPR_64[] = { // 64-bit registers.
4066 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4067 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4069 static const uint16_t *FPR = GetFPR();
4071 static const uint16_t VR[] = {
4072 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4073 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4075 const unsigned NumGPRs = array_lengthof(GPR_32);
4076 const unsigned NumFPRs = 13;
4077 const unsigned NumVRs = array_lengthof(VR);
4079 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
4081 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4082 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4084 SmallVector<SDValue, 8> MemOpChains;
4085 for (unsigned i = 0; i != NumOps; ++i) {
4086 SDValue Arg = OutVals[i];
4087 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4089 // PtrOff will be used to store the current argument to the stack if a
4090 // register cannot be found for it.
4093 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4095 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4097 // On PPC64, promote integers to 64-bit values.
4098 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4099 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4100 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4101 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4104 // FIXME memcpy is used way more than necessary. Correctness first.
4105 // Note: "by value" is code for passing a structure by value, not
4107 if (Flags.isByVal()) {
4108 unsigned Size = Flags.getByValSize();
4109 // Very small objects are passed right-justified. Everything else is
4110 // passed left-justified.
4111 if (Size==1 || Size==2) {
4112 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4113 if (GPR_idx != NumGPRs) {
4114 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4115 MachinePointerInfo(), VT,
4117 MemOpChains.push_back(Load.getValue(1));
4118 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4120 ArgOffset += PtrByteSize;
4122 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4123 PtrOff.getValueType());
4124 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4125 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4128 ArgOffset += PtrByteSize;
4132 // Copy entire object into memory. There are cases where gcc-generated
4133 // code assumes it is there, even if it could be put entirely into
4134 // registers. (This is not what the doc says.)
4135 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4139 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4140 // copy the pieces of the object that fit into registers from the
4141 // parameter save area.
4142 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4143 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4144 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4145 if (GPR_idx != NumGPRs) {
4146 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4147 MachinePointerInfo(),
4148 false, false, false, 0);
4149 MemOpChains.push_back(Load.getValue(1));
4150 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4151 ArgOffset += PtrByteSize;
4153 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4160 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4161 default: llvm_unreachable("Unexpected ValueType for argument!");
4164 if (GPR_idx != NumGPRs) {
4165 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4167 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4168 isPPC64, isTailCall, false, MemOpChains,
4169 TailCallArguments, dl);
4171 ArgOffset += PtrByteSize;
4175 if (FPR_idx != NumFPRs) {
4176 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4179 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4180 MachinePointerInfo(), false, false, 0);
4181 MemOpChains.push_back(Store);
4183 // Float varargs are always shadowed in available integer registers
4184 if (GPR_idx != NumGPRs) {
4185 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4186 MachinePointerInfo(), false, false,
4188 MemOpChains.push_back(Load.getValue(1));
4189 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4191 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4192 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4193 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4194 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4195 MachinePointerInfo(),
4196 false, false, false, 0);
4197 MemOpChains.push_back(Load.getValue(1));
4198 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4201 // If we have any FPRs remaining, we may also have GPRs remaining.
4202 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4204 if (GPR_idx != NumGPRs)
4206 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4207 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4211 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4212 isPPC64, isTailCall, false, MemOpChains,
4213 TailCallArguments, dl);
4217 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4224 // These go aligned on the stack, or in the corresponding R registers
4225 // when within range. The Darwin PPC ABI doc claims they also go in
4226 // V registers; in fact gcc does this only for arguments that are
4227 // prototyped, not for those that match the ... We do it for all
4228 // arguments, seems to work.
4229 while (ArgOffset % 16 !=0) {
4230 ArgOffset += PtrByteSize;
4231 if (GPR_idx != NumGPRs)
4234 // We could elide this store in the case where the object fits
4235 // entirely in R registers. Maybe later.
4236 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4237 DAG.getConstant(ArgOffset, PtrVT));
4238 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4239 MachinePointerInfo(), false, false, 0);
4240 MemOpChains.push_back(Store);
4241 if (VR_idx != NumVRs) {
4242 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4243 MachinePointerInfo(),
4244 false, false, false, 0);
4245 MemOpChains.push_back(Load.getValue(1));
4246 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4249 for (unsigned i=0; i<16; i+=PtrByteSize) {
4250 if (GPR_idx == NumGPRs)
4252 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4253 DAG.getConstant(i, PtrVT));
4254 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4255 false, false, false, 0);
4256 MemOpChains.push_back(Load.getValue(1));
4257 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4262 // Non-varargs Altivec params generally go in registers, but have
4263 // stack space allocated at the end.
4264 if (VR_idx != NumVRs) {
4265 // Doesn't have GPR space allocated.
4266 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4267 } else if (nAltivecParamsAtEnd==0) {
4268 // We are emitting Altivec params in order.
4269 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4270 isPPC64, isTailCall, true, MemOpChains,
4271 TailCallArguments, dl);
4277 // If all Altivec parameters fit in registers, as they usually do,
4278 // they get stack space following the non-Altivec parameters. We
4279 // don't track this here because nobody below needs it.
4280 // If there are more Altivec parameters than fit in registers emit
4282 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4284 // Offset is aligned; skip 1st 12 params which go in V registers.
4285 ArgOffset = ((ArgOffset+15)/16)*16;
4287 for (unsigned i = 0; i != NumOps; ++i) {
4288 SDValue Arg = OutVals[i];
4289 EVT ArgType = Outs[i].VT;
4290 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4291 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4294 // We are emitting Altivec params in order.
4295 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4296 isPPC64, isTailCall, true, MemOpChains,
4297 TailCallArguments, dl);
4304 if (!MemOpChains.empty())
4305 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4306 &MemOpChains[0], MemOpChains.size());
4308 // On Darwin, R12 must contain the address of an indirect callee. This does
4309 // not mean the MTCTR instruction must use R12; it's easier to model this as
4310 // an extra parameter, so do that.
4312 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4313 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4314 !isBLACompatibleAddress(Callee, DAG))
4315 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4316 PPC::R12), Callee));
4318 // Build a sequence of copy-to-reg nodes chained together with token chain
4319 // and flag operands which copy the outgoing args into the appropriate regs.
4321 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4322 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4323 RegsToPass[i].second, InFlag);
4324 InFlag = Chain.getValue(1);
4328 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4329 FPOp, true, TailCallArguments);
4331 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4332 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4337 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4338 MachineFunction &MF, bool isVarArg,
4339 const SmallVectorImpl<ISD::OutputArg> &Outs,
4340 LLVMContext &Context) const {
4341 SmallVector<CCValAssign, 16> RVLocs;
4342 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4344 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4348 PPCTargetLowering::LowerReturn(SDValue Chain,
4349 CallingConv::ID CallConv, bool isVarArg,
4350 const SmallVectorImpl<ISD::OutputArg> &Outs,
4351 const SmallVectorImpl<SDValue> &OutVals,
4352 DebugLoc dl, SelectionDAG &DAG) const {
4354 SmallVector<CCValAssign, 16> RVLocs;
4355 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4356 getTargetMachine(), RVLocs, *DAG.getContext());
4357 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4359 // If this is the first return lowered for this function, add the regs to the
4360 // liveout set for the function.
4361 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
4362 for (unsigned i = 0; i != RVLocs.size(); ++i)
4363 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
4368 // Copy the result values into the output registers.
4369 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4370 CCValAssign &VA = RVLocs[i];
4371 assert(VA.isRegLoc() && "Can only return in registers!");
4373 SDValue Arg = OutVals[i];
4375 switch (VA.getLocInfo()) {
4376 default: llvm_unreachable("Unknown loc info!");
4377 case CCValAssign::Full: break;
4378 case CCValAssign::AExt:
4379 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
4381 case CCValAssign::ZExt:
4382 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
4384 case CCValAssign::SExt:
4385 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
4389 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
4390 Flag = Chain.getValue(1);
4394 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
4396 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
4399 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4400 const PPCSubtarget &Subtarget) const {
4401 // When we pop the dynamic allocation we need to restore the SP link.
4402 DebugLoc dl = Op.getDebugLoc();
4404 // Get the corect type for pointers.
4405 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4407 // Construct the stack pointer operand.
4408 bool isPPC64 = Subtarget.isPPC64();
4409 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4410 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4412 // Get the operands for the STACKRESTORE.
4413 SDValue Chain = Op.getOperand(0);
4414 SDValue SaveSP = Op.getOperand(1);
4416 // Load the old link SP.
4417 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4418 MachinePointerInfo(),
4419 false, false, false, 0);
4421 // Restore the stack pointer.
4422 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4424 // Store the old link SP.
4425 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4432 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4433 MachineFunction &MF = DAG.getMachineFunction();
4434 bool isPPC64 = PPCSubTarget.isPPC64();
4435 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4436 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4438 // Get current frame pointer save index. The users of this index will be
4439 // primarily DYNALLOC instructions.
4440 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4441 int RASI = FI->getReturnAddrSaveIndex();
4443 // If the frame pointer save index hasn't been defined yet.
4445 // Find out what the fix offset of the frame pointer save area.
4446 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4447 // Allocate the frame index for frame pointer save area.
4448 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4450 FI->setReturnAddrSaveIndex(RASI);
4452 return DAG.getFrameIndex(RASI, PtrVT);
4456 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4457 MachineFunction &MF = DAG.getMachineFunction();
4458 bool isPPC64 = PPCSubTarget.isPPC64();
4459 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4460 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4462 // Get current frame pointer save index. The users of this index will be
4463 // primarily DYNALLOC instructions.
4464 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4465 int FPSI = FI->getFramePointerSaveIndex();
4467 // If the frame pointer save index hasn't been defined yet.
4469 // Find out what the fix offset of the frame pointer save area.
4470 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4473 // Allocate the frame index for frame pointer save area.
4474 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4476 FI->setFramePointerSaveIndex(FPSI);
4478 return DAG.getFrameIndex(FPSI, PtrVT);
4481 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4483 const PPCSubtarget &Subtarget) const {
4485 SDValue Chain = Op.getOperand(0);
4486 SDValue Size = Op.getOperand(1);
4487 DebugLoc dl = Op.getDebugLoc();
4489 // Get the corect type for pointers.
4490 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4492 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4493 DAG.getConstant(0, PtrVT), Size);
4494 // Construct a node for the frame pointer save index.
4495 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4496 // Build a DYNALLOC node.
4497 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4498 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4499 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4502 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4504 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4505 // Not FP? Not a fsel.
4506 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4507 !Op.getOperand(2).getValueType().isFloatingPoint())
4510 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4512 // Cannot handle SETEQ/SETNE.
4513 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4515 EVT ResVT = Op.getValueType();
4516 EVT CmpVT = Op.getOperand(0).getValueType();
4517 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4518 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4519 DebugLoc dl = Op.getDebugLoc();
4521 // If the RHS of the comparison is a 0.0, we don't need to do the
4522 // subtraction at all.
4523 if (isFloatingPointZero(RHS))
4525 default: break; // SETUO etc aren't handled by fsel.
4528 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4531 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4532 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4533 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4536 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4539 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4540 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4541 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4542 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4547 default: break; // SETUO etc aren't handled by fsel.
4550 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4551 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4552 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4553 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4556 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4557 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4558 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4559 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4562 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4563 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4564 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4565 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4568 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4569 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4570 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4571 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4576 // FIXME: Split this code up when LegalizeDAGTypes lands.
4577 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4578 DebugLoc dl) const {
4579 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4580 SDValue Src = Op.getOperand(0);
4581 if (Src.getValueType() == MVT::f32)
4582 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4585 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4586 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4588 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4593 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4597 // Convert the FP value to an int value through memory.
4598 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4600 // Emit a store to the stack slot.
4601 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4602 MachinePointerInfo(), false, false, 0);
4604 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4606 if (Op.getValueType() == MVT::i32)
4607 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4608 DAG.getConstant(4, FIPtr.getValueType()));
4609 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4610 false, false, false, 0);
4613 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4614 SelectionDAG &DAG) const {
4615 DebugLoc dl = Op.getDebugLoc();
4616 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4617 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4620 if (Op.getOperand(0).getValueType() == MVT::i64) {
4621 SDValue SINT = Op.getOperand(0);
4622 // When converting to single-precision, we actually need to convert
4623 // to double-precision first and then round to single-precision.
4624 // To avoid double-rounding effects during that operation, we have
4625 // to prepare the input operand. Bits that might be truncated when
4626 // converting to double-precision are replaced by a bit that won't
4627 // be lost at this stage, but is below the single-precision rounding
4630 // However, if -enable-unsafe-fp-math is in effect, accept double
4631 // rounding to avoid the extra overhead.
4632 if (Op.getValueType() == MVT::f32 &&
4633 !DAG.getTarget().Options.UnsafeFPMath) {
4635 // Twiddle input to make sure the low 11 bits are zero. (If this
4636 // is the case, we are guaranteed the value will fit into the 53 bit
4637 // mantissa of an IEEE double-precision value without rounding.)
4638 // If any of those low 11 bits were not zero originally, make sure
4639 // bit 12 (value 2048) is set instead, so that the final rounding
4640 // to single-precision gets the correct result.
4641 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4642 SINT, DAG.getConstant(2047, MVT::i64));
4643 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4644 Round, DAG.getConstant(2047, MVT::i64));
4645 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4646 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4647 Round, DAG.getConstant(-2048, MVT::i64));
4649 // However, we cannot use that value unconditionally: if the magnitude
4650 // of the input value is small, the bit-twiddling we did above might
4651 // end up visibly changing the output. Fortunately, in that case, we
4652 // don't need to twiddle bits since the original input will convert
4653 // exactly to double-precision floating-point already. Therefore,
4654 // construct a conditional to use the original value if the top 11
4655 // bits are all sign-bit copies, and use the rounded value computed
4657 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4658 SINT, DAG.getConstant(53, MVT::i32));
4659 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4660 Cond, DAG.getConstant(1, MVT::i64));
4661 Cond = DAG.getSetCC(dl, MVT::i32,
4662 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4664 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4666 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4667 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4668 if (Op.getValueType() == MVT::f32)
4669 FP = DAG.getNode(ISD::FP_ROUND, dl,
4670 MVT::f32, FP, DAG.getIntPtrConstant(0));
4674 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4675 "Unhandled SINT_TO_FP type in custom expander!");
4676 // Since we only generate this in 64-bit mode, we can take advantage of
4677 // 64-bit registers. In particular, sign extend the input value into the
4678 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4679 // then lfd it and fcfid it.
4680 MachineFunction &MF = DAG.getMachineFunction();
4681 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4682 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4683 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4684 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4686 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4689 // STD the extended value into the stack slot.
4690 MachineMemOperand *MMO =
4691 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4692 MachineMemOperand::MOStore, 8, 8);
4693 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4695 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4696 Ops, 4, MVT::i64, MMO);
4697 // Load the value as a double.
4698 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4699 false, false, false, 0);
4701 // FCFID it and return it.
4702 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4703 if (Op.getValueType() == MVT::f32)
4704 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4708 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4709 SelectionDAG &DAG) const {
4710 DebugLoc dl = Op.getDebugLoc();
4712 The rounding mode is in bits 30:31 of FPSR, and has the following
4719 FLT_ROUNDS, on the other hand, expects the following:
4726 To perform the conversion, we do:
4727 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4730 MachineFunction &MF = DAG.getMachineFunction();
4731 EVT VT = Op.getValueType();
4732 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4733 std::vector<EVT> NodeTys;
4734 SDValue MFFSreg, InFlag;
4736 // Save FP Control Word to register
4737 NodeTys.push_back(MVT::f64); // return register
4738 NodeTys.push_back(MVT::Glue); // unused in this context
4739 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4741 // Save FP register to stack slot
4742 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4743 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4744 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4745 StackSlot, MachinePointerInfo(), false, false,0);
4747 // Load FP Control Word from low 32 bits of stack slot.
4748 SDValue Four = DAG.getConstant(4, PtrVT);
4749 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4750 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4751 false, false, false, 0);
4753 // Transform as necessary
4755 DAG.getNode(ISD::AND, dl, MVT::i32,
4756 CWD, DAG.getConstant(3, MVT::i32));
4758 DAG.getNode(ISD::SRL, dl, MVT::i32,
4759 DAG.getNode(ISD::AND, dl, MVT::i32,
4760 DAG.getNode(ISD::XOR, dl, MVT::i32,
4761 CWD, DAG.getConstant(3, MVT::i32)),
4762 DAG.getConstant(3, MVT::i32)),
4763 DAG.getConstant(1, MVT::i32));
4766 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4768 return DAG.getNode((VT.getSizeInBits() < 16 ?
4769 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4772 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4773 EVT VT = Op.getValueType();
4774 unsigned BitWidth = VT.getSizeInBits();
4775 DebugLoc dl = Op.getDebugLoc();
4776 assert(Op.getNumOperands() == 3 &&
4777 VT == Op.getOperand(1).getValueType() &&
4780 // Expand into a bunch of logical ops. Note that these ops
4781 // depend on the PPC behavior for oversized shift amounts.
4782 SDValue Lo = Op.getOperand(0);
4783 SDValue Hi = Op.getOperand(1);
4784 SDValue Amt = Op.getOperand(2);
4785 EVT AmtVT = Amt.getValueType();
4787 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4788 DAG.getConstant(BitWidth, AmtVT), Amt);
4789 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4790 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4791 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4792 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4793 DAG.getConstant(-BitWidth, AmtVT));
4794 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4795 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4796 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4797 SDValue OutOps[] = { OutLo, OutHi };
4798 return DAG.getMergeValues(OutOps, 2, dl);
4801 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4802 EVT VT = Op.getValueType();
4803 DebugLoc dl = Op.getDebugLoc();
4804 unsigned BitWidth = VT.getSizeInBits();
4805 assert(Op.getNumOperands() == 3 &&
4806 VT == Op.getOperand(1).getValueType() &&
4809 // Expand into a bunch of logical ops. Note that these ops
4810 // depend on the PPC behavior for oversized shift amounts.
4811 SDValue Lo = Op.getOperand(0);
4812 SDValue Hi = Op.getOperand(1);
4813 SDValue Amt = Op.getOperand(2);
4814 EVT AmtVT = Amt.getValueType();
4816 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4817 DAG.getConstant(BitWidth, AmtVT), Amt);
4818 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4819 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4820 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4821 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4822 DAG.getConstant(-BitWidth, AmtVT));
4823 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4824 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4825 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4826 SDValue OutOps[] = { OutLo, OutHi };
4827 return DAG.getMergeValues(OutOps, 2, dl);
4830 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4831 DebugLoc dl = Op.getDebugLoc();
4832 EVT VT = Op.getValueType();
4833 unsigned BitWidth = VT.getSizeInBits();
4834 assert(Op.getNumOperands() == 3 &&
4835 VT == Op.getOperand(1).getValueType() &&
4838 // Expand into a bunch of logical ops, followed by a select_cc.
4839 SDValue Lo = Op.getOperand(0);
4840 SDValue Hi = Op.getOperand(1);
4841 SDValue Amt = Op.getOperand(2);
4842 EVT AmtVT = Amt.getValueType();
4844 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4845 DAG.getConstant(BitWidth, AmtVT), Amt);
4846 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4847 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4848 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4849 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4850 DAG.getConstant(-BitWidth, AmtVT));
4851 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4852 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4853 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4854 Tmp4, Tmp6, ISD::SETLE);
4855 SDValue OutOps[] = { OutLo, OutHi };
4856 return DAG.getMergeValues(OutOps, 2, dl);
4859 //===----------------------------------------------------------------------===//
4860 // Vector related lowering.
4863 /// BuildSplatI - Build a canonical splati of Val with an element size of
4864 /// SplatSize. Cast the result to VT.
4865 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4866 SelectionDAG &DAG, DebugLoc dl) {
4867 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4869 static const EVT VTys[] = { // canonical VT to use for each size.
4870 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4873 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4875 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4879 EVT CanonicalVT = VTys[SplatSize-1];
4881 // Build a canonical splat for this value.
4882 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4883 SmallVector<SDValue, 8> Ops;
4884 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4885 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4886 &Ops[0], Ops.size());
4887 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4890 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4891 /// specified intrinsic ID.
4892 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4893 SelectionDAG &DAG, DebugLoc dl,
4894 EVT DestVT = MVT::Other) {
4895 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4896 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4897 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4900 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4901 /// specified intrinsic ID.
4902 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4903 SDValue Op2, SelectionDAG &DAG,
4904 DebugLoc dl, EVT DestVT = MVT::Other) {
4905 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4906 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4907 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4911 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4912 /// amount. The result has the specified value type.
4913 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4914 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4915 // Force LHS/RHS to be the right type.
4916 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4917 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4920 for (unsigned i = 0; i != 16; ++i)
4922 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4923 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4926 // If this is a case we can't handle, return null and let the default
4927 // expansion code take care of it. If we CAN select this case, and if it
4928 // selects to a single instruction, return Op. Otherwise, if we can codegen
4929 // this case more efficiently than a constant pool load, lower it to the
4930 // sequence of ops that should be used.
4931 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4932 SelectionDAG &DAG) const {
4933 DebugLoc dl = Op.getDebugLoc();
4934 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4935 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4937 // Check if this is a splat of a constant value.
4938 APInt APSplatBits, APSplatUndef;
4939 unsigned SplatBitSize;
4941 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4942 HasAnyUndefs, 0, true) || SplatBitSize > 32)
4945 unsigned SplatBits = APSplatBits.getZExtValue();
4946 unsigned SplatUndef = APSplatUndef.getZExtValue();
4947 unsigned SplatSize = SplatBitSize / 8;
4949 // First, handle single instruction cases.
4952 if (SplatBits == 0) {
4953 // Canonicalize all zero vectors to be v4i32.
4954 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
4955 SDValue Z = DAG.getConstant(0, MVT::i32);
4956 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
4957 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
4962 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
4963 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
4965 if (SextVal >= -16 && SextVal <= 15)
4966 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
4969 // Two instruction sequences.
4971 // If this value is in the range [-32,30] and is even, use:
4972 // tmp = VSPLTI[bhw], result = add tmp, tmp
4973 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
4974 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
4975 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
4976 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4979 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
4980 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
4982 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
4983 // Make -1 and vspltisw -1:
4984 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
4986 // Make the VSLW intrinsic, computing 0x8000_0000.
4987 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
4990 // xor by OnesV to invert it.
4991 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
4992 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
4995 // Check to see if this is a wide variety of vsplti*, binop self cases.
4996 static const signed char SplatCsts[] = {
4997 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
4998 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
5001 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
5002 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
5003 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
5004 int i = SplatCsts[idx];
5006 // Figure out what shift amount will be used by altivec if shifted by i in
5008 unsigned TypeShiftAmt = i & (SplatBitSize-1);
5010 // vsplti + shl self.
5011 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
5012 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5013 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5014 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
5015 Intrinsic::ppc_altivec_vslw
5017 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5018 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5021 // vsplti + srl self.
5022 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5023 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5024 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5025 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
5026 Intrinsic::ppc_altivec_vsrw
5028 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5029 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5032 // vsplti + sra self.
5033 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5034 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5035 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5036 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
5037 Intrinsic::ppc_altivec_vsraw
5039 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5040 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5043 // vsplti + rol self.
5044 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
5045 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
5046 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5047 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5048 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
5049 Intrinsic::ppc_altivec_vrlw
5051 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5052 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5055 // t = vsplti c, result = vsldoi t, t, 1
5056 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
5057 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5058 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
5060 // t = vsplti c, result = vsldoi t, t, 2
5061 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
5062 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5063 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
5065 // t = vsplti c, result = vsldoi t, t, 3
5066 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
5067 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5068 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
5072 // Three instruction sequences.
5074 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
5075 if (SextVal >= 0 && SextVal <= 31) {
5076 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
5077 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
5078 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
5079 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
5081 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
5082 if (SextVal >= -31 && SextVal <= 0) {
5083 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
5084 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
5085 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
5086 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
5092 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5093 /// the specified operations to build the shuffle.
5094 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5095 SDValue RHS, SelectionDAG &DAG,
5097 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5098 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5099 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5102 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5114 if (OpNum == OP_COPY) {
5115 if (LHSID == (1*9+2)*9+3) return LHS;
5116 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5120 SDValue OpLHS, OpRHS;
5121 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5122 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5126 default: llvm_unreachable("Unknown i32 permute!");
5128 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5129 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5130 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5131 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5134 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5135 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5136 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5137 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5140 for (unsigned i = 0; i != 16; ++i)
5141 ShufIdxs[i] = (i&3)+0;
5144 for (unsigned i = 0; i != 16; ++i)
5145 ShufIdxs[i] = (i&3)+4;
5148 for (unsigned i = 0; i != 16; ++i)
5149 ShufIdxs[i] = (i&3)+8;
5152 for (unsigned i = 0; i != 16; ++i)
5153 ShufIdxs[i] = (i&3)+12;
5156 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5158 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5160 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5162 EVT VT = OpLHS.getValueType();
5163 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5164 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5165 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5166 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5169 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5170 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5171 /// return the code it can be lowered into. Worst case, it can always be
5172 /// lowered into a vperm.
5173 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5174 SelectionDAG &DAG) const {
5175 DebugLoc dl = Op.getDebugLoc();
5176 SDValue V1 = Op.getOperand(0);
5177 SDValue V2 = Op.getOperand(1);
5178 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5179 EVT VT = Op.getValueType();
5181 // Cases that are handled by instructions that take permute immediates
5182 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5183 // selected by the instruction selector.
5184 if (V2.getOpcode() == ISD::UNDEF) {
5185 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5186 PPC::isSplatShuffleMask(SVOp, 2) ||
5187 PPC::isSplatShuffleMask(SVOp, 4) ||
5188 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5189 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5190 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5191 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5192 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5193 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5194 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5195 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5196 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5201 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5202 // and produce a fixed permutation. If any of these match, do not lower to
5204 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5205 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5206 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5207 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5208 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5209 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5210 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5211 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5212 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5215 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5216 // perfect shuffle table to emit an optimal matching sequence.
5217 ArrayRef<int> PermMask = SVOp->getMask();
5219 unsigned PFIndexes[4];
5220 bool isFourElementShuffle = true;
5221 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5222 unsigned EltNo = 8; // Start out undef.
5223 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5224 if (PermMask[i*4+j] < 0)
5225 continue; // Undef, ignore it.
5227 unsigned ByteSource = PermMask[i*4+j];
5228 if ((ByteSource & 3) != j) {
5229 isFourElementShuffle = false;
5234 EltNo = ByteSource/4;
5235 } else if (EltNo != ByteSource/4) {
5236 isFourElementShuffle = false;
5240 PFIndexes[i] = EltNo;
5243 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5244 // perfect shuffle vector to determine if it is cost effective to do this as
5245 // discrete instructions, or whether we should use a vperm.
5246 if (isFourElementShuffle) {
5247 // Compute the index in the perfect shuffle table.
5248 unsigned PFTableIndex =
5249 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5251 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5252 unsigned Cost = (PFEntry >> 30);
5254 // Determining when to avoid vperm is tricky. Many things affect the cost
5255 // of vperm, particularly how many times the perm mask needs to be computed.
5256 // For example, if the perm mask can be hoisted out of a loop or is already
5257 // used (perhaps because there are multiple permutes with the same shuffle
5258 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5259 // the loop requires an extra register.
5261 // As a compromise, we only emit discrete instructions if the shuffle can be
5262 // generated in 3 or fewer operations. When we have loop information
5263 // available, if this block is within a loop, we should avoid using vperm
5264 // for 3-operation perms and use a constant pool load instead.
5266 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5269 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5270 // vector that will get spilled to the constant pool.
5271 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5273 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5274 // that it is in input element units, not in bytes. Convert now.
5275 EVT EltVT = V1.getValueType().getVectorElementType();
5276 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5278 SmallVector<SDValue, 16> ResultMask;
5279 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5280 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5282 for (unsigned j = 0; j != BytesPerElement; ++j)
5283 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5287 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5288 &ResultMask[0], ResultMask.size());
5289 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5292 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5293 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5294 /// information about the intrinsic.
5295 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5297 unsigned IntrinsicID =
5298 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5301 switch (IntrinsicID) {
5302 default: return false;
5303 // Comparison predicates.
5304 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5305 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5306 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5307 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5308 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5309 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5310 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5311 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5312 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5313 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5314 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5315 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5316 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5318 // Normal Comparisons.
5319 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5320 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5321 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5322 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5323 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5324 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5325 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5326 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5327 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5328 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5329 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5330 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5331 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5336 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5337 /// lower, do it, otherwise return null.
5338 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5339 SelectionDAG &DAG) const {
5340 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5341 // opcode number of the comparison.
5342 DebugLoc dl = Op.getDebugLoc();
5345 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5346 return SDValue(); // Don't custom lower most intrinsics.
5348 // If this is a non-dot comparison, make the VCMP node and we are done.
5350 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5351 Op.getOperand(1), Op.getOperand(2),
5352 DAG.getConstant(CompareOpc, MVT::i32));
5353 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5356 // Create the PPCISD altivec 'dot' comparison node.
5358 Op.getOperand(2), // LHS
5359 Op.getOperand(3), // RHS
5360 DAG.getConstant(CompareOpc, MVT::i32)
5362 std::vector<EVT> VTs;
5363 VTs.push_back(Op.getOperand(2).getValueType());
5364 VTs.push_back(MVT::Glue);
5365 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5367 // Now that we have the comparison, emit a copy from the CR to a GPR.
5368 // This is flagged to the above dot comparison.
5369 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5370 DAG.getRegister(PPC::CR6, MVT::i32),
5371 CompNode.getValue(1));
5373 // Unpack the result based on how the target uses it.
5374 unsigned BitNo; // Bit # of CR6.
5375 bool InvertBit; // Invert result?
5376 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5377 default: // Can't happen, don't crash on invalid number though.
5378 case 0: // Return the value of the EQ bit of CR6.
5379 BitNo = 0; InvertBit = false;
5381 case 1: // Return the inverted value of the EQ bit of CR6.
5382 BitNo = 0; InvertBit = true;
5384 case 2: // Return the value of the LT bit of CR6.
5385 BitNo = 2; InvertBit = false;
5387 case 3: // Return the inverted value of the LT bit of CR6.
5388 BitNo = 2; InvertBit = true;
5392 // Shift the bit into the low position.
5393 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5394 DAG.getConstant(8-(3-BitNo), MVT::i32));
5396 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5397 DAG.getConstant(1, MVT::i32));
5399 // If we are supposed to, toggle the bit.
5401 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5402 DAG.getConstant(1, MVT::i32));
5406 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5407 SelectionDAG &DAG) const {
5408 DebugLoc dl = Op.getDebugLoc();
5409 // Create a stack slot that is 16-byte aligned.
5410 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5411 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5412 EVT PtrVT = getPointerTy();
5413 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5415 // Store the input value into Value#0 of the stack slot.
5416 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5417 Op.getOperand(0), FIdx, MachinePointerInfo(),
5420 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5421 false, false, false, 0);
5424 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5425 DebugLoc dl = Op.getDebugLoc();
5426 if (Op.getValueType() == MVT::v4i32) {
5427 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5429 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5430 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5432 SDValue RHSSwap = // = vrlw RHS, 16
5433 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5435 // Shrinkify inputs to v8i16.
5436 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5437 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5438 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5440 // Low parts multiplied together, generating 32-bit results (we ignore the
5442 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5443 LHS, RHS, DAG, dl, MVT::v4i32);
5445 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5446 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5447 // Shift the high parts up 16 bits.
5448 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5450 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5451 } else if (Op.getValueType() == MVT::v8i16) {
5452 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5454 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5456 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5457 LHS, RHS, Zero, DAG, dl);
5458 } else if (Op.getValueType() == MVT::v16i8) {
5459 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5461 // Multiply the even 8-bit parts, producing 16-bit sums.
5462 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5463 LHS, RHS, DAG, dl, MVT::v8i16);
5464 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5466 // Multiply the odd 8-bit parts, producing 16-bit sums.
5467 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5468 LHS, RHS, DAG, dl, MVT::v8i16);
5469 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5471 // Merge the results together.
5473 for (unsigned i = 0; i != 8; ++i) {
5475 Ops[i*2+1] = 2*i+1+16;
5477 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5479 llvm_unreachable("Unknown mul to lower!");
5483 /// LowerOperation - Provide custom lowering hooks for some operations.
5485 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5486 switch (Op.getOpcode()) {
5487 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5488 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5489 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5490 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5491 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5492 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5493 case ISD::SETCC: return LowerSETCC(Op, DAG);
5494 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5495 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5497 return LowerVASTART(Op, DAG, PPCSubTarget);
5500 return LowerVAARG(Op, DAG, PPCSubTarget);
5502 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5503 case ISD::DYNAMIC_STACKALLOC:
5504 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5506 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5507 case ISD::FP_TO_UINT:
5508 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5510 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5511 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5513 // Lower 64-bit shifts.
5514 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5515 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5516 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5518 // Vector-related lowering.
5519 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5520 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5521 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5522 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5523 case ISD::MUL: return LowerMUL(Op, DAG);
5525 // Frame & Return address.
5526 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5527 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5531 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5532 SmallVectorImpl<SDValue>&Results,
5533 SelectionDAG &DAG) const {
5534 const TargetMachine &TM = getTargetMachine();
5535 DebugLoc dl = N->getDebugLoc();
5536 switch (N->getOpcode()) {
5538 llvm_unreachable("Do not know how to custom type legalize this operation!");
5540 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5541 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5544 EVT VT = N->getValueType(0);
5546 if (VT == MVT::i64) {
5547 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5549 Results.push_back(NewNode);
5550 Results.push_back(NewNode.getValue(1));
5554 case ISD::FP_ROUND_INREG: {
5555 assert(N->getValueType(0) == MVT::ppcf128);
5556 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5557 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5558 MVT::f64, N->getOperand(0),
5559 DAG.getIntPtrConstant(0));
5560 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5561 MVT::f64, N->getOperand(0),
5562 DAG.getIntPtrConstant(1));
5564 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5565 // of the long double, and puts FPSCR back the way it was. We do not
5566 // actually model FPSCR.
5567 std::vector<EVT> NodeTys;
5568 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5570 NodeTys.push_back(MVT::f64); // Return register
5571 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5572 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5573 MFFSreg = Result.getValue(0);
5574 InFlag = Result.getValue(1);
5577 NodeTys.push_back(MVT::Glue); // Returns a flag
5578 Ops[0] = DAG.getConstant(31, MVT::i32);
5580 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5581 InFlag = Result.getValue(0);
5584 NodeTys.push_back(MVT::Glue); // Returns a flag
5585 Ops[0] = DAG.getConstant(30, MVT::i32);
5587 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5588 InFlag = Result.getValue(0);
5591 NodeTys.push_back(MVT::f64); // result of add
5592 NodeTys.push_back(MVT::Glue); // Returns a flag
5596 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5597 FPreg = Result.getValue(0);
5598 InFlag = Result.getValue(1);
5601 NodeTys.push_back(MVT::f64);
5602 Ops[0] = DAG.getConstant(1, MVT::i32);
5606 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5607 FPreg = Result.getValue(0);
5609 // We know the low half is about to be thrown away, so just use something
5611 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5615 case ISD::FP_TO_SINT:
5616 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5622 //===----------------------------------------------------------------------===//
5623 // Other Lowering Code
5624 //===----------------------------------------------------------------------===//
5627 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5628 bool is64bit, unsigned BinOpcode) const {
5629 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5630 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5632 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5633 MachineFunction *F = BB->getParent();
5634 MachineFunction::iterator It = BB;
5637 unsigned dest = MI->getOperand(0).getReg();
5638 unsigned ptrA = MI->getOperand(1).getReg();
5639 unsigned ptrB = MI->getOperand(2).getReg();
5640 unsigned incr = MI->getOperand(3).getReg();
5641 DebugLoc dl = MI->getDebugLoc();
5643 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5644 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5645 F->insert(It, loopMBB);
5646 F->insert(It, exitMBB);
5647 exitMBB->splice(exitMBB->begin(), BB,
5648 llvm::next(MachineBasicBlock::iterator(MI)),
5650 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5652 MachineRegisterInfo &RegInfo = F->getRegInfo();
5653 unsigned TmpReg = (!BinOpcode) ? incr :
5654 RegInfo.createVirtualRegister(
5655 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5656 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5660 // fallthrough --> loopMBB
5661 BB->addSuccessor(loopMBB);
5664 // l[wd]arx dest, ptr
5665 // add r0, dest, incr
5666 // st[wd]cx. r0, ptr
5668 // fallthrough --> exitMBB
5670 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5671 .addReg(ptrA).addReg(ptrB);
5673 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5674 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5675 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5676 BuildMI(BB, dl, TII->get(PPC::BCC))
5677 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5678 BB->addSuccessor(loopMBB);
5679 BB->addSuccessor(exitMBB);
5688 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5689 MachineBasicBlock *BB,
5690 bool is8bit, // operation
5691 unsigned BinOpcode) const {
5692 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5693 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5694 // In 64 bit mode we have to use 64 bits for addresses, even though the
5695 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5696 // registers without caring whether they're 32 or 64, but here we're
5697 // doing actual arithmetic on the addresses.
5698 bool is64bit = PPCSubTarget.isPPC64();
5699 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5701 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5702 MachineFunction *F = BB->getParent();
5703 MachineFunction::iterator It = BB;
5706 unsigned dest = MI->getOperand(0).getReg();
5707 unsigned ptrA = MI->getOperand(1).getReg();
5708 unsigned ptrB = MI->getOperand(2).getReg();
5709 unsigned incr = MI->getOperand(3).getReg();
5710 DebugLoc dl = MI->getDebugLoc();
5712 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5713 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5714 F->insert(It, loopMBB);
5715 F->insert(It, exitMBB);
5716 exitMBB->splice(exitMBB->begin(), BB,
5717 llvm::next(MachineBasicBlock::iterator(MI)),
5719 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5721 MachineRegisterInfo &RegInfo = F->getRegInfo();
5722 const TargetRegisterClass *RC =
5723 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5724 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5725 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5726 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5727 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5728 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5729 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5730 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5731 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5732 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5733 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5734 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5735 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5737 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5741 // fallthrough --> loopMBB
5742 BB->addSuccessor(loopMBB);
5744 // The 4-byte load must be aligned, while a char or short may be
5745 // anywhere in the word. Hence all this nasty bookkeeping code.
5746 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5747 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5748 // xori shift, shift1, 24 [16]
5749 // rlwinm ptr, ptr1, 0, 0, 29
5750 // slw incr2, incr, shift
5751 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5752 // slw mask, mask2, shift
5754 // lwarx tmpDest, ptr
5755 // add tmp, tmpDest, incr2
5756 // andc tmp2, tmpDest, mask
5757 // and tmp3, tmp, mask
5758 // or tmp4, tmp3, tmp2
5761 // fallthrough --> exitMBB
5762 // srw dest, tmpDest, shift
5763 if (ptrA != ZeroReg) {
5764 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5765 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5766 .addReg(ptrA).addReg(ptrB);
5770 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5771 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5772 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5773 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5775 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5776 .addReg(Ptr1Reg).addImm(0).addImm(61);
5778 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5779 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5780 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5781 .addReg(incr).addReg(ShiftReg);
5783 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5785 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5786 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5788 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5789 .addReg(Mask2Reg).addReg(ShiftReg);
5792 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5793 .addReg(ZeroReg).addReg(PtrReg);
5795 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5796 .addReg(Incr2Reg).addReg(TmpDestReg);
5797 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5798 .addReg(TmpDestReg).addReg(MaskReg);
5799 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5800 .addReg(TmpReg).addReg(MaskReg);
5801 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5802 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5803 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5804 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5805 BuildMI(BB, dl, TII->get(PPC::BCC))
5806 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5807 BB->addSuccessor(loopMBB);
5808 BB->addSuccessor(exitMBB);
5813 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5819 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5820 MachineBasicBlock *BB) const {
5821 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5823 // To "insert" these instructions we actually have to insert their
5824 // control-flow patterns.
5825 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5826 MachineFunction::iterator It = BB;
5829 MachineFunction *F = BB->getParent();
5831 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5832 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5833 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5834 PPC::ISEL8 : PPC::ISEL;
5835 unsigned SelectPred = MI->getOperand(4).getImm();
5836 DebugLoc dl = MI->getDebugLoc();
5838 // The SelectPred is ((BI << 5) | BO) for a BCC
5839 unsigned BO = SelectPred & 0xF;
5840 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5842 unsigned TrueOpNo, FalseOpNo;
5849 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5852 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5853 .addReg(MI->getOperand(TrueOpNo).getReg())
5854 .addReg(MI->getOperand(FalseOpNo).getReg())
5855 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5856 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5857 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5858 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5859 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5860 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5863 // The incoming instruction knows the destination vreg to set, the
5864 // condition code register to branch on, the true/false values to
5865 // select between, and a branch opcode to use.
5870 // cmpTY ccX, r1, r2
5872 // fallthrough --> copy0MBB
5873 MachineBasicBlock *thisMBB = BB;
5874 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5875 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5876 unsigned SelectPred = MI->getOperand(4).getImm();
5877 DebugLoc dl = MI->getDebugLoc();
5878 F->insert(It, copy0MBB);
5879 F->insert(It, sinkMBB);
5881 // Transfer the remainder of BB and its successor edges to sinkMBB.
5882 sinkMBB->splice(sinkMBB->begin(), BB,
5883 llvm::next(MachineBasicBlock::iterator(MI)),
5885 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5887 // Next, add the true and fallthrough blocks as its successors.
5888 BB->addSuccessor(copy0MBB);
5889 BB->addSuccessor(sinkMBB);
5891 BuildMI(BB, dl, TII->get(PPC::BCC))
5892 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5895 // %FalseValue = ...
5896 // # fallthrough to sinkMBB
5899 // Update machine-CFG edges
5900 BB->addSuccessor(sinkMBB);
5903 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5906 BuildMI(*BB, BB->begin(), dl,
5907 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5908 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5909 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5911 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5912 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5913 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5914 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5915 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5916 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5917 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5918 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5920 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5921 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5922 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5923 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5924 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5925 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5926 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5927 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5929 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5930 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5931 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5932 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5933 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5934 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5935 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5936 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5938 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5939 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5940 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5941 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5942 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5943 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5944 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5945 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5947 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5948 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5949 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5950 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5951 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
5952 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
5953 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
5954 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
5956 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
5957 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
5958 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
5959 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
5960 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
5961 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
5962 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
5963 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
5965 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
5966 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
5967 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
5968 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
5969 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
5970 BB = EmitAtomicBinary(MI, BB, false, 0);
5971 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
5972 BB = EmitAtomicBinary(MI, BB, true, 0);
5974 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
5975 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
5976 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
5978 unsigned dest = MI->getOperand(0).getReg();
5979 unsigned ptrA = MI->getOperand(1).getReg();
5980 unsigned ptrB = MI->getOperand(2).getReg();
5981 unsigned oldval = MI->getOperand(3).getReg();
5982 unsigned newval = MI->getOperand(4).getReg();
5983 DebugLoc dl = MI->getDebugLoc();
5985 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
5986 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
5987 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
5988 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5989 F->insert(It, loop1MBB);
5990 F->insert(It, loop2MBB);
5991 F->insert(It, midMBB);
5992 F->insert(It, exitMBB);
5993 exitMBB->splice(exitMBB->begin(), BB,
5994 llvm::next(MachineBasicBlock::iterator(MI)),
5996 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6000 // fallthrough --> loopMBB
6001 BB->addSuccessor(loop1MBB);
6004 // l[wd]arx dest, ptr
6005 // cmp[wd] dest, oldval
6008 // st[wd]cx. newval, ptr
6012 // st[wd]cx. dest, ptr
6015 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
6016 .addReg(ptrA).addReg(ptrB);
6017 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
6018 .addReg(oldval).addReg(dest);
6019 BuildMI(BB, dl, TII->get(PPC::BCC))
6020 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6021 BB->addSuccessor(loop2MBB);
6022 BB->addSuccessor(midMBB);
6025 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6026 .addReg(newval).addReg(ptrA).addReg(ptrB);
6027 BuildMI(BB, dl, TII->get(PPC::BCC))
6028 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6029 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6030 BB->addSuccessor(loop1MBB);
6031 BB->addSuccessor(exitMBB);
6034 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6035 .addReg(dest).addReg(ptrA).addReg(ptrB);
6036 BB->addSuccessor(exitMBB);
6041 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
6042 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
6043 // We must use 64-bit registers for addresses when targeting 64-bit,
6044 // since we're actually doing arithmetic on them. Other registers
6046 bool is64bit = PPCSubTarget.isPPC64();
6047 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
6049 unsigned dest = MI->getOperand(0).getReg();
6050 unsigned ptrA = MI->getOperand(1).getReg();
6051 unsigned ptrB = MI->getOperand(2).getReg();
6052 unsigned oldval = MI->getOperand(3).getReg();
6053 unsigned newval = MI->getOperand(4).getReg();
6054 DebugLoc dl = MI->getDebugLoc();
6056 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6057 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6058 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6059 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6060 F->insert(It, loop1MBB);
6061 F->insert(It, loop2MBB);
6062 F->insert(It, midMBB);
6063 F->insert(It, exitMBB);
6064 exitMBB->splice(exitMBB->begin(), BB,
6065 llvm::next(MachineBasicBlock::iterator(MI)),
6067 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6069 MachineRegisterInfo &RegInfo = F->getRegInfo();
6070 const TargetRegisterClass *RC =
6071 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
6072 (const TargetRegisterClass *) &PPC::GPRCRegClass;
6073 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6074 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6075 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6076 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
6077 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
6078 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
6079 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
6080 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6081 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6082 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6083 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6084 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6085 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6087 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
6088 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
6091 // fallthrough --> loopMBB
6092 BB->addSuccessor(loop1MBB);
6094 // The 4-byte load must be aligned, while a char or short may be
6095 // anywhere in the word. Hence all this nasty bookkeeping code.
6096 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6097 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6098 // xori shift, shift1, 24 [16]
6099 // rlwinm ptr, ptr1, 0, 0, 29
6100 // slw newval2, newval, shift
6101 // slw oldval2, oldval,shift
6102 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6103 // slw mask, mask2, shift
6104 // and newval3, newval2, mask
6105 // and oldval3, oldval2, mask
6107 // lwarx tmpDest, ptr
6108 // and tmp, tmpDest, mask
6109 // cmpw tmp, oldval3
6112 // andc tmp2, tmpDest, mask
6113 // or tmp4, tmp2, newval3
6118 // stwcx. tmpDest, ptr
6120 // srw dest, tmpDest, shift
6121 if (ptrA != ZeroReg) {
6122 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6123 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6124 .addReg(ptrA).addReg(ptrB);
6128 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6129 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6130 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6131 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6133 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6134 .addReg(Ptr1Reg).addImm(0).addImm(61);
6136 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6137 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6138 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6139 .addReg(newval).addReg(ShiftReg);
6140 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6141 .addReg(oldval).addReg(ShiftReg);
6143 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6145 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6146 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6147 .addReg(Mask3Reg).addImm(65535);
6149 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6150 .addReg(Mask2Reg).addReg(ShiftReg);
6151 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6152 .addReg(NewVal2Reg).addReg(MaskReg);
6153 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6154 .addReg(OldVal2Reg).addReg(MaskReg);
6157 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6158 .addReg(ZeroReg).addReg(PtrReg);
6159 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6160 .addReg(TmpDestReg).addReg(MaskReg);
6161 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6162 .addReg(TmpReg).addReg(OldVal3Reg);
6163 BuildMI(BB, dl, TII->get(PPC::BCC))
6164 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6165 BB->addSuccessor(loop2MBB);
6166 BB->addSuccessor(midMBB);
6169 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6170 .addReg(TmpDestReg).addReg(MaskReg);
6171 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6172 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6173 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6174 .addReg(ZeroReg).addReg(PtrReg);
6175 BuildMI(BB, dl, TII->get(PPC::BCC))
6176 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6177 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6178 BB->addSuccessor(loop1MBB);
6179 BB->addSuccessor(exitMBB);
6182 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6183 .addReg(ZeroReg).addReg(PtrReg);
6184 BB->addSuccessor(exitMBB);
6189 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6192 llvm_unreachable("Unexpected instr type to insert");
6195 MI->eraseFromParent(); // The pseudo instruction is gone now.
6199 //===----------------------------------------------------------------------===//
6200 // Target Optimization Hooks
6201 //===----------------------------------------------------------------------===//
6203 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6204 DAGCombinerInfo &DCI) const {
6205 const TargetMachine &TM = getTargetMachine();
6206 SelectionDAG &DAG = DCI.DAG;
6207 DebugLoc dl = N->getDebugLoc();
6208 switch (N->getOpcode()) {
6211 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6212 if (C->isNullValue()) // 0 << V -> 0.
6213 return N->getOperand(0);
6217 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6218 if (C->isNullValue()) // 0 >>u V -> 0.
6219 return N->getOperand(0);
6223 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6224 if (C->isNullValue() || // 0 >>s V -> 0.
6225 C->isAllOnesValue()) // -1 >>s V -> -1.
6226 return N->getOperand(0);
6230 case ISD::SINT_TO_FP:
6231 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6232 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6233 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6234 // We allow the src/dst to be either f32/f64, but the intermediate
6235 // type must be i64.
6236 if (N->getOperand(0).getValueType() == MVT::i64 &&
6237 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6238 SDValue Val = N->getOperand(0).getOperand(0);
6239 if (Val.getValueType() == MVT::f32) {
6240 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6241 DCI.AddToWorklist(Val.getNode());
6244 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6245 DCI.AddToWorklist(Val.getNode());
6246 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6247 DCI.AddToWorklist(Val.getNode());
6248 if (N->getValueType(0) == MVT::f32) {
6249 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6250 DAG.getIntPtrConstant(0));
6251 DCI.AddToWorklist(Val.getNode());
6254 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6255 // If the intermediate type is i32, we can avoid the load/store here
6262 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6263 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6264 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6265 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6266 N->getOperand(1).getValueType() == MVT::i32 &&
6267 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6268 SDValue Val = N->getOperand(1).getOperand(0);
6269 if (Val.getValueType() == MVT::f32) {
6270 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6271 DCI.AddToWorklist(Val.getNode());
6273 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6274 DCI.AddToWorklist(Val.getNode());
6276 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6277 N->getOperand(2), N->getOperand(3));
6278 DCI.AddToWorklist(Val.getNode());
6282 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6283 if (cast<StoreSDNode>(N)->isUnindexed() &&
6284 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6285 N->getOperand(1).getNode()->hasOneUse() &&
6286 (N->getOperand(1).getValueType() == MVT::i32 ||
6287 N->getOperand(1).getValueType() == MVT::i16)) {
6288 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6289 // Do an any-extend to 32-bits if this is a half-word input.
6290 if (BSwapOp.getValueType() == MVT::i16)
6291 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6294 N->getOperand(0), BSwapOp, N->getOperand(2),
6295 DAG.getValueType(N->getOperand(1).getValueType())
6298 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6299 Ops, array_lengthof(Ops),
6300 cast<StoreSDNode>(N)->getMemoryVT(),
6301 cast<StoreSDNode>(N)->getMemOperand());
6305 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6306 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6307 N->getOperand(0).hasOneUse() &&
6308 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6309 SDValue Load = N->getOperand(0);
6310 LoadSDNode *LD = cast<LoadSDNode>(Load);
6311 // Create the byte-swapping load.
6313 LD->getChain(), // Chain
6314 LD->getBasePtr(), // Ptr
6315 DAG.getValueType(N->getValueType(0)) // VT
6318 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6319 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6320 LD->getMemoryVT(), LD->getMemOperand());
6322 // If this is an i16 load, insert the truncate.
6323 SDValue ResVal = BSLoad;
6324 if (N->getValueType(0) == MVT::i16)
6325 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6327 // First, combine the bswap away. This makes the value produced by the
6329 DCI.CombineTo(N, ResVal);
6331 // Next, combine the load away, we give it a bogus result value but a real
6332 // chain result. The result value is dead because the bswap is dead.
6333 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6335 // Return N so it doesn't get rechecked!
6336 return SDValue(N, 0);
6340 case PPCISD::VCMP: {
6341 // If a VCMPo node already exists with exactly the same operands as this
6342 // node, use its result instead of this node (VCMPo computes both a CR6 and
6343 // a normal output).
6345 if (!N->getOperand(0).hasOneUse() &&
6346 !N->getOperand(1).hasOneUse() &&
6347 !N->getOperand(2).hasOneUse()) {
6349 // Scan all of the users of the LHS, looking for VCMPo's that match.
6350 SDNode *VCMPoNode = 0;
6352 SDNode *LHSN = N->getOperand(0).getNode();
6353 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6355 if (UI->getOpcode() == PPCISD::VCMPo &&
6356 UI->getOperand(1) == N->getOperand(1) &&
6357 UI->getOperand(2) == N->getOperand(2) &&
6358 UI->getOperand(0) == N->getOperand(0)) {
6363 // If there is no VCMPo node, or if the flag value has a single use, don't
6365 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6368 // Look at the (necessarily single) use of the flag value. If it has a
6369 // chain, this transformation is more complex. Note that multiple things
6370 // could use the value result, which we should ignore.
6371 SDNode *FlagUser = 0;
6372 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6373 FlagUser == 0; ++UI) {
6374 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6376 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6377 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6384 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6385 // give up for right now.
6386 if (FlagUser->getOpcode() == PPCISD::MFCR)
6387 return SDValue(VCMPoNode, 0);
6392 // If this is a branch on an altivec predicate comparison, lower this so
6393 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6394 // lowering is done pre-legalize, because the legalizer lowers the predicate
6395 // compare down to code that is difficult to reassemble.
6396 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6397 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6401 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6402 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6403 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6404 assert(isDot && "Can't compare against a vector result!");
6406 // If this is a comparison against something other than 0/1, then we know
6407 // that the condition is never/always true.
6408 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6409 if (Val != 0 && Val != 1) {
6410 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6411 return N->getOperand(0);
6412 // Always !=, turn it into an unconditional branch.
6413 return DAG.getNode(ISD::BR, dl, MVT::Other,
6414 N->getOperand(0), N->getOperand(4));
6417 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6419 // Create the PPCISD altivec 'dot' comparison node.
6420 std::vector<EVT> VTs;
6422 LHS.getOperand(2), // LHS of compare
6423 LHS.getOperand(3), // RHS of compare
6424 DAG.getConstant(CompareOpc, MVT::i32)
6426 VTs.push_back(LHS.getOperand(2).getValueType());
6427 VTs.push_back(MVT::Glue);
6428 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6430 // Unpack the result based on how the target uses it.
6431 PPC::Predicate CompOpc;
6432 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6433 default: // Can't happen, don't crash on invalid number though.
6434 case 0: // Branch on the value of the EQ bit of CR6.
6435 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6437 case 1: // Branch on the inverted value of the EQ bit of CR6.
6438 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6440 case 2: // Branch on the value of the LT bit of CR6.
6441 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6443 case 3: // Branch on the inverted value of the LT bit of CR6.
6444 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6448 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6449 DAG.getConstant(CompOpc, MVT::i32),
6450 DAG.getRegister(PPC::CR6, MVT::i32),
6451 N->getOperand(4), CompNode.getValue(1));
6460 //===----------------------------------------------------------------------===//
6461 // Inline Assembly Support
6462 //===----------------------------------------------------------------------===//
6464 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6467 const SelectionDAG &DAG,
6468 unsigned Depth) const {
6469 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6470 switch (Op.getOpcode()) {
6472 case PPCISD::LBRX: {
6473 // lhbrx is known to have the top bits cleared out.
6474 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6475 KnownZero = 0xFFFF0000;
6478 case ISD::INTRINSIC_WO_CHAIN: {
6479 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6481 case Intrinsic::ppc_altivec_vcmpbfp_p:
6482 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6483 case Intrinsic::ppc_altivec_vcmpequb_p:
6484 case Intrinsic::ppc_altivec_vcmpequh_p:
6485 case Intrinsic::ppc_altivec_vcmpequw_p:
6486 case Intrinsic::ppc_altivec_vcmpgefp_p:
6487 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6488 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6489 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6490 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6491 case Intrinsic::ppc_altivec_vcmpgtub_p:
6492 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6493 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6494 KnownZero = ~1U; // All bits but the low one are known to be zero.
6502 /// getConstraintType - Given a constraint, return the type of
6503 /// constraint it is for this target.
6504 PPCTargetLowering::ConstraintType
6505 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6506 if (Constraint.size() == 1) {
6507 switch (Constraint[0]) {
6514 return C_RegisterClass;
6516 // FIXME: While Z does indicate a memory constraint, it specifically
6517 // indicates an r+r address (used in conjunction with the 'y' modifier
6518 // in the replacement string). Currently, we're forcing the base
6519 // register to be r0 in the asm printer (which is interpreted as zero)
6520 // and forming the complete address in the second register. This is
6525 return TargetLowering::getConstraintType(Constraint);
6528 /// Examine constraint type and operand type and determine a weight value.
6529 /// This object must already have been set up with the operand type
6530 /// and the current alternative constraint selected.
6531 TargetLowering::ConstraintWeight
6532 PPCTargetLowering::getSingleConstraintMatchWeight(
6533 AsmOperandInfo &info, const char *constraint) const {
6534 ConstraintWeight weight = CW_Invalid;
6535 Value *CallOperandVal = info.CallOperandVal;
6536 // If we don't have a value, we can't do a match,
6537 // but allow it at the lowest weight.
6538 if (CallOperandVal == NULL)
6540 Type *type = CallOperandVal->getType();
6541 // Look at the constraint type.
6542 switch (*constraint) {
6544 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6547 if (type->isIntegerTy())
6548 weight = CW_Register;
6551 if (type->isFloatTy())
6552 weight = CW_Register;
6555 if (type->isDoubleTy())
6556 weight = CW_Register;
6559 if (type->isVectorTy())
6560 weight = CW_Register;
6563 weight = CW_Register;
6572 std::pair<unsigned, const TargetRegisterClass*>
6573 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6575 if (Constraint.size() == 1) {
6576 // GCC RS6000 Constraint Letters
6577 switch (Constraint[0]) {
6580 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6581 return std::make_pair(0U, &PPC::G8RCRegClass);
6582 return std::make_pair(0U, &PPC::GPRCRegClass);
6584 if (VT == MVT::f32 || VT == MVT::i32)
6585 return std::make_pair(0U, &PPC::F4RCRegClass);
6586 if (VT == MVT::f64 || VT == MVT::i64)
6587 return std::make_pair(0U, &PPC::F8RCRegClass);
6590 return std::make_pair(0U, &PPC::VRRCRegClass);
6592 return std::make_pair(0U, &PPC::CRRCRegClass);
6596 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6600 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6601 /// vector. If it is invalid, don't add anything to Ops.
6602 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6603 std::string &Constraint,
6604 std::vector<SDValue>&Ops,
6605 SelectionDAG &DAG) const {
6606 SDValue Result(0,0);
6608 // Only support length 1 constraints.
6609 if (Constraint.length() > 1) return;
6611 char Letter = Constraint[0];
6622 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6623 if (!CST) return; // Must be an immediate to match.
6624 unsigned Value = CST->getZExtValue();
6626 default: llvm_unreachable("Unknown constraint letter!");
6627 case 'I': // "I" is a signed 16-bit constant.
6628 if ((short)Value == (int)Value)
6629 Result = DAG.getTargetConstant(Value, Op.getValueType());
6631 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6632 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6633 if ((short)Value == 0)
6634 Result = DAG.getTargetConstant(Value, Op.getValueType());
6636 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6637 if ((Value >> 16) == 0)
6638 Result = DAG.getTargetConstant(Value, Op.getValueType());
6640 case 'M': // "M" is a constant that is greater than 31.
6642 Result = DAG.getTargetConstant(Value, Op.getValueType());
6644 case 'N': // "N" is a positive constant that is an exact power of two.
6645 if ((int)Value > 0 && isPowerOf2_32(Value))
6646 Result = DAG.getTargetConstant(Value, Op.getValueType());
6648 case 'O': // "O" is the constant zero.
6650 Result = DAG.getTargetConstant(Value, Op.getValueType());
6652 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6653 if ((short)-Value == (int)-Value)
6654 Result = DAG.getTargetConstant(Value, Op.getValueType());
6661 if (Result.getNode()) {
6662 Ops.push_back(Result);
6666 // Handle standard constraint letters.
6667 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6670 // isLegalAddressingMode - Return true if the addressing mode represented
6671 // by AM is legal for this target, for a load/store of the specified type.
6672 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6674 // FIXME: PPC does not allow r+i addressing modes for vectors!
6676 // PPC allows a sign-extended 16-bit immediate field.
6677 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6680 // No global is ever allowed as a base.
6684 // PPC only support r+r,
6686 case 0: // "r+i" or just "i", depending on HasBaseReg.
6689 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6691 // Otherwise we have r+r or r+i.
6694 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6696 // Allow 2*r as r+r.
6699 // No other scales are supported.
6706 /// isLegalAddressImmediate - Return true if the integer value can be used
6707 /// as the offset of the target addressing mode for load / store of the
6709 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6710 // PPC allows a sign-extended 16-bit immediate field.
6711 return (V > -(1 << 16) && V < (1 << 16)-1);
6714 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6718 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6719 SelectionDAG &DAG) const {
6720 MachineFunction &MF = DAG.getMachineFunction();
6721 MachineFrameInfo *MFI = MF.getFrameInfo();
6722 MFI->setReturnAddressIsTaken(true);
6724 DebugLoc dl = Op.getDebugLoc();
6725 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6727 // Make sure the function does not optimize away the store of the RA to
6729 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6730 FuncInfo->setLRStoreRequired();
6731 bool isPPC64 = PPCSubTarget.isPPC64();
6732 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6735 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6738 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6739 isPPC64? MVT::i64 : MVT::i32);
6740 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6741 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6743 MachinePointerInfo(), false, false, false, 0);
6746 // Just load the return address off the stack.
6747 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6748 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6749 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6752 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6753 SelectionDAG &DAG) const {
6754 DebugLoc dl = Op.getDebugLoc();
6755 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6757 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6758 bool isPPC64 = PtrVT == MVT::i64;
6760 MachineFunction &MF = DAG.getMachineFunction();
6761 MachineFrameInfo *MFI = MF.getFrameInfo();
6762 MFI->setFrameAddressIsTaken(true);
6763 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6764 MFI->hasVarSizedObjects()) &&
6765 MFI->getStackSize() &&
6766 !MF.getFunction()->getFnAttributes().
6767 hasAttribute(Attributes::Naked);
6768 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6769 (is31 ? PPC::R31 : PPC::R1);
6770 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6773 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6774 FrameAddr, MachinePointerInfo(), false, false,
6780 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6781 // The PowerPC target isn't yet aware of offsets.
6785 /// getOptimalMemOpType - Returns the target specific optimal type for load
6786 /// and store operations as a result of memset, memcpy, and memmove
6787 /// lowering. If DstAlign is zero that means it's safe to destination
6788 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6789 /// means there isn't a need to check it against alignment requirement,
6790 /// probably because the source does not need to be loaded. If
6791 /// 'IsZeroVal' is true, that means it's safe to return a
6792 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
6793 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
6794 /// constant so it does not need to be loaded.
6795 /// It returns EVT::Other if the type should be determined using generic
6796 /// target-independent logic.
6797 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6798 unsigned DstAlign, unsigned SrcAlign,
6801 MachineFunction &MF) const {
6802 if (this->PPCSubTarget.isPPC64()) {
6809 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6810 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6811 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6812 /// is expanded to mul + add.
6813 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6817 switch (VT.getSimpleVT().SimpleTy) {
6829 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6831 return TargetLowering::getSchedulingPreference(N);