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/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
27 #include "llvm/IR/CallingConv.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/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_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
40 CCValAssign::LocInfo &LocInfo,
41 ISD::ArgFlagsTy &ArgFlags,
43 static bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
45 CCValAssign::LocInfo &LocInfo,
46 ISD::ArgFlagsTy &ArgFlags,
48 static bool CC_PPC32_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::FSINCOS, MVT::f64, Expand);
136 setOperationAction(ISD::FREM , MVT::f64, Expand);
137 setOperationAction(ISD::FPOW , MVT::f64, Expand);
138 setOperationAction(ISD::FMA , MVT::f64, Legal);
139 setOperationAction(ISD::FSIN , MVT::f32, Expand);
140 setOperationAction(ISD::FCOS , MVT::f32, Expand);
141 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
142 setOperationAction(ISD::FREM , MVT::f32, Expand);
143 setOperationAction(ISD::FPOW , MVT::f32, Expand);
144 setOperationAction(ISD::FMA , MVT::f32, Legal);
146 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
148 // If we're enabling GP optimizations, use hardware square root
149 if (!Subtarget->hasFSQRT()) {
150 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
151 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
154 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
155 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
157 // PowerPC does not have BSWAP, CTPOP or CTTZ
158 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
159 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
160 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
162 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
163 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
164 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
165 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
166 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
167 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
169 // PowerPC does not have ROTR
170 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
171 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
173 // PowerPC does not have Select
174 setOperationAction(ISD::SELECT, MVT::i32, Expand);
175 setOperationAction(ISD::SELECT, MVT::i64, Expand);
176 setOperationAction(ISD::SELECT, MVT::f32, Expand);
177 setOperationAction(ISD::SELECT, MVT::f64, Expand);
179 // PowerPC wants to turn select_cc of FP into fsel when possible.
180 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
181 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
183 // PowerPC wants to optimize integer setcc a bit
184 setOperationAction(ISD::SETCC, MVT::i32, Custom);
186 // PowerPC does not have BRCOND which requires SetCC
187 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
189 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
191 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
192 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
194 // PowerPC does not have [U|S]INT_TO_FP
195 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
196 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
198 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
199 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
200 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
201 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
203 // We cannot sextinreg(i1). Expand to shifts.
204 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
206 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
207 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
208 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
209 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
212 // We want to legalize GlobalAddress and ConstantPool nodes into the
213 // appropriate instructions to materialize the address.
214 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
215 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
216 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
217 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
218 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
219 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
220 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
221 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
222 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
223 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
226 setOperationAction(ISD::TRAP, MVT::Other, Legal);
228 // TRAMPOLINE is custom lowered.
229 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
230 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
232 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
233 setOperationAction(ISD::VASTART , MVT::Other, Custom);
235 if (Subtarget->isSVR4ABI()) {
237 // VAARG always uses double-word chunks, so promote anything smaller.
238 setOperationAction(ISD::VAARG, MVT::i1, Promote);
239 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
240 setOperationAction(ISD::VAARG, MVT::i8, Promote);
241 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
242 setOperationAction(ISD::VAARG, MVT::i16, Promote);
243 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
244 setOperationAction(ISD::VAARG, MVT::i32, Promote);
245 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
246 setOperationAction(ISD::VAARG, MVT::Other, Expand);
248 // VAARG is custom lowered with the 32-bit SVR4 ABI.
249 setOperationAction(ISD::VAARG, MVT::Other, Custom);
250 setOperationAction(ISD::VAARG, MVT::i64, Custom);
253 setOperationAction(ISD::VAARG, MVT::Other, Expand);
255 // Use the default implementation.
256 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
257 setOperationAction(ISD::VAEND , MVT::Other, Expand);
258 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
259 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
260 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
261 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
263 // We want to custom lower some of our intrinsics.
264 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
266 // Comparisons that require checking two conditions.
267 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
270 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
271 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
272 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
273 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
274 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
275 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
276 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
277 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
278 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
280 if (Subtarget->has64BitSupport()) {
281 // They also have instructions for converting between i64 and fp.
282 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
283 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
284 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
285 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
286 // This is just the low 32 bits of a (signed) fp->i64 conversion.
287 // We cannot do this with Promote because i64 is not a legal type.
288 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
290 // FIXME: disable this lowered code. This generates 64-bit register values,
291 // and we don't model the fact that the top part is clobbered by calls. We
292 // need to flag these together so that the value isn't live across a call.
293 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
295 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
296 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
299 if (Subtarget->use64BitRegs()) {
300 // 64-bit PowerPC implementations can support i64 types directly
301 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
302 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
303 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
304 // 64-bit PowerPC wants to expand i128 shifts itself.
305 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
306 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
307 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
309 // 32-bit PowerPC wants to expand i64 shifts itself.
310 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
311 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
312 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
315 if (Subtarget->hasAltivec()) {
316 // First set operation action for all vector types to expand. Then we
317 // will selectively turn on ones that can be effectively codegen'd.
318 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
319 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
320 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
322 // add/sub are legal for all supported vector VT's.
323 setOperationAction(ISD::ADD , VT, Legal);
324 setOperationAction(ISD::SUB , VT, Legal);
326 // We promote all shuffles to v16i8.
327 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
328 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
330 // We promote all non-typed operations to v4i32.
331 setOperationAction(ISD::AND , VT, Promote);
332 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
333 setOperationAction(ISD::OR , VT, Promote);
334 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
335 setOperationAction(ISD::XOR , VT, Promote);
336 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
337 setOperationAction(ISD::LOAD , VT, Promote);
338 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
339 setOperationAction(ISD::SELECT, VT, Promote);
340 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
341 setOperationAction(ISD::STORE, VT, Promote);
342 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
344 // No other operations are legal.
345 setOperationAction(ISD::MUL , VT, Expand);
346 setOperationAction(ISD::SDIV, VT, Expand);
347 setOperationAction(ISD::SREM, VT, Expand);
348 setOperationAction(ISD::UDIV, VT, Expand);
349 setOperationAction(ISD::UREM, VT, Expand);
350 setOperationAction(ISD::FDIV, VT, Expand);
351 setOperationAction(ISD::FNEG, VT, Expand);
352 setOperationAction(ISD::FSQRT, VT, Expand);
353 setOperationAction(ISD::FLOG, VT, Expand);
354 setOperationAction(ISD::FLOG10, VT, Expand);
355 setOperationAction(ISD::FLOG2, VT, Expand);
356 setOperationAction(ISD::FEXP, VT, Expand);
357 setOperationAction(ISD::FEXP2, VT, Expand);
358 setOperationAction(ISD::FSIN, VT, Expand);
359 setOperationAction(ISD::FCOS, VT, Expand);
360 setOperationAction(ISD::FABS, VT, Expand);
361 setOperationAction(ISD::FPOWI, VT, Expand);
362 setOperationAction(ISD::FFLOOR, VT, Expand);
363 setOperationAction(ISD::FCEIL, VT, Expand);
364 setOperationAction(ISD::FTRUNC, VT, Expand);
365 setOperationAction(ISD::FRINT, VT, Expand);
366 setOperationAction(ISD::FNEARBYINT, VT, Expand);
367 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
368 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
369 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
370 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
371 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
372 setOperationAction(ISD::UDIVREM, VT, Expand);
373 setOperationAction(ISD::SDIVREM, VT, Expand);
374 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
375 setOperationAction(ISD::FPOW, VT, Expand);
376 setOperationAction(ISD::CTPOP, VT, Expand);
377 setOperationAction(ISD::CTLZ, VT, Expand);
378 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
379 setOperationAction(ISD::CTTZ, VT, Expand);
380 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
381 setOperationAction(ISD::VSELECT, VT, Expand);
382 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
384 for (unsigned j = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
385 j <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++j) {
386 MVT::SimpleValueType InnerVT = (MVT::SimpleValueType)j;
387 setTruncStoreAction(VT, InnerVT, Expand);
389 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
390 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
391 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
394 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
395 // with merges, splats, etc.
396 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
398 setOperationAction(ISD::AND , MVT::v4i32, Legal);
399 setOperationAction(ISD::OR , MVT::v4i32, Legal);
400 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
401 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
402 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
403 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
404 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
405 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
406 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
407 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
408 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
409 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
410 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
411 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
413 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
414 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
415 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
416 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
418 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
419 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
420 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
421 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
422 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
424 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
425 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
427 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
428 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
429 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
430 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
432 // Altivec does not contain unordered floating-point compare instructions
433 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
434 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
435 setCondCodeAction(ISD::SETUGT, MVT::v4f32, Expand);
436 setCondCodeAction(ISD::SETUGE, MVT::v4f32, Expand);
437 setCondCodeAction(ISD::SETULT, MVT::v4f32, Expand);
438 setCondCodeAction(ISD::SETULE, MVT::v4f32, Expand);
441 if (Subtarget->has64BitSupport()) {
442 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
443 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
446 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
447 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
448 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
449 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
451 setBooleanContents(ZeroOrOneBooleanContent);
452 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
455 setStackPointerRegisterToSaveRestore(PPC::X1);
456 setExceptionPointerRegister(PPC::X3);
457 setExceptionSelectorRegister(PPC::X4);
459 setStackPointerRegisterToSaveRestore(PPC::R1);
460 setExceptionPointerRegister(PPC::R3);
461 setExceptionSelectorRegister(PPC::R4);
464 // We have target-specific dag combine patterns for the following nodes:
465 setTargetDAGCombine(ISD::SINT_TO_FP);
466 setTargetDAGCombine(ISD::STORE);
467 setTargetDAGCombine(ISD::BR_CC);
468 setTargetDAGCombine(ISD::BSWAP);
470 // Darwin long double math library functions have $LDBL128 appended.
471 if (Subtarget->isDarwin()) {
472 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
473 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
474 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
475 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
476 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
477 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
478 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
479 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
480 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
481 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
484 setMinFunctionAlignment(2);
485 if (PPCSubTarget.isDarwin())
486 setPrefFunctionAlignment(4);
488 if (isPPC64 && Subtarget->isJITCodeModel())
489 // Temporary workaround for the inability of PPC64 JIT to handle jump
491 setSupportJumpTables(false);
493 setInsertFencesForAtomic(true);
495 setSchedulingPreference(Sched::Hybrid);
497 computeRegisterProperties();
499 // The Freescale cores does better with aggressive inlining of memcpy and
500 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
501 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
502 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
503 MaxStoresPerMemset = 32;
504 MaxStoresPerMemsetOptSize = 16;
505 MaxStoresPerMemcpy = 32;
506 MaxStoresPerMemcpyOptSize = 8;
507 MaxStoresPerMemmove = 32;
508 MaxStoresPerMemmoveOptSize = 8;
510 setPrefFunctionAlignment(4);
511 BenefitFromCodePlacementOpt = true;
515 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
516 /// function arguments in the caller parameter area.
517 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
518 const TargetMachine &TM = getTargetMachine();
519 // Darwin passes everything on 4 byte boundary.
520 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
523 // 16byte and wider vectors are passed on 16byte boundary.
524 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
525 if (VTy->getBitWidth() >= 128)
528 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
529 if (PPCSubTarget.isPPC64())
535 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
538 case PPCISD::FSEL: return "PPCISD::FSEL";
539 case PPCISD::FCFID: return "PPCISD::FCFID";
540 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
541 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
542 case PPCISD::STFIWX: return "PPCISD::STFIWX";
543 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
544 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
545 case PPCISD::VPERM: return "PPCISD::VPERM";
546 case PPCISD::Hi: return "PPCISD::Hi";
547 case PPCISD::Lo: return "PPCISD::Lo";
548 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
549 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
550 case PPCISD::LOAD: return "PPCISD::LOAD";
551 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
552 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
553 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
554 case PPCISD::SRL: return "PPCISD::SRL";
555 case PPCISD::SRA: return "PPCISD::SRA";
556 case PPCISD::SHL: return "PPCISD::SHL";
557 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
558 case PPCISD::STD_32: return "PPCISD::STD_32";
559 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
560 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
561 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
562 case PPCISD::NOP: return "PPCISD::NOP";
563 case PPCISD::MTCTR: return "PPCISD::MTCTR";
564 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
565 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
566 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
567 case PPCISD::MFCR: return "PPCISD::MFCR";
568 case PPCISD::VCMP: return "PPCISD::VCMP";
569 case PPCISD::VCMPo: return "PPCISD::VCMPo";
570 case PPCISD::LBRX: return "PPCISD::LBRX";
571 case PPCISD::STBRX: return "PPCISD::STBRX";
572 case PPCISD::LARX: return "PPCISD::LARX";
573 case PPCISD::STCX: return "PPCISD::STCX";
574 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
575 case PPCISD::MFFS: return "PPCISD::MFFS";
576 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
577 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
578 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
579 case PPCISD::MTFSF: return "PPCISD::MTFSF";
580 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
581 case PPCISD::CR6SET: return "PPCISD::CR6SET";
582 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
583 case PPCISD::ADDIS_TOC_HA: return "PPCISD::ADDIS_TOC_HA";
584 case PPCISD::LD_TOC_L: return "PPCISD::LD_TOC_L";
585 case PPCISD::ADDI_TOC_L: return "PPCISD::ADDI_TOC_L";
586 case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA";
587 case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L";
588 case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS";
589 case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA";
590 case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L";
591 case PPCISD::GET_TLS_ADDR: return "PPCISD::GET_TLS_ADDR";
592 case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA";
593 case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L";
594 case PPCISD::GET_TLSLD_ADDR: return "PPCISD::GET_TLSLD_ADDR";
595 case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
596 case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L";
597 case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT";
601 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
604 return VT.changeVectorElementTypeToInteger();
607 //===----------------------------------------------------------------------===//
608 // Node matching predicates, for use by the tblgen matching code.
609 //===----------------------------------------------------------------------===//
611 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
612 static bool isFloatingPointZero(SDValue Op) {
613 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
614 return CFP->getValueAPF().isZero();
615 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
616 // Maybe this has already been legalized into the constant pool?
617 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
618 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
619 return CFP->getValueAPF().isZero();
624 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
625 /// true if Op is undef or if it matches the specified value.
626 static bool isConstantOrUndef(int Op, int Val) {
627 return Op < 0 || Op == Val;
630 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
631 /// VPKUHUM instruction.
632 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
634 for (unsigned i = 0; i != 16; ++i)
635 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
638 for (unsigned i = 0; i != 8; ++i)
639 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
640 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
646 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
647 /// VPKUWUM instruction.
648 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
650 for (unsigned i = 0; i != 16; i += 2)
651 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
652 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
655 for (unsigned i = 0; i != 8; i += 2)
656 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
657 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
658 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
659 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
665 /// isVMerge - Common function, used to match vmrg* shuffles.
667 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
668 unsigned LHSStart, unsigned RHSStart) {
669 assert(N->getValueType(0) == MVT::v16i8 &&
670 "PPC only supports shuffles by bytes!");
671 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
672 "Unsupported merge size!");
674 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
675 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
676 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
677 LHSStart+j+i*UnitSize) ||
678 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
679 RHSStart+j+i*UnitSize))
685 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
686 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
687 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
690 return isVMerge(N, UnitSize, 8, 24);
691 return isVMerge(N, UnitSize, 8, 8);
694 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
695 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
696 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
699 return isVMerge(N, UnitSize, 0, 16);
700 return isVMerge(N, UnitSize, 0, 0);
704 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
705 /// amount, otherwise return -1.
706 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
707 assert(N->getValueType(0) == MVT::v16i8 &&
708 "PPC only supports shuffles by bytes!");
710 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
712 // Find the first non-undef value in the shuffle mask.
714 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
717 if (i == 16) return -1; // all undef.
719 // Otherwise, check to see if the rest of the elements are consecutively
720 // numbered from this value.
721 unsigned ShiftAmt = SVOp->getMaskElt(i);
722 if (ShiftAmt < i) return -1;
726 // Check the rest of the elements to see if they are consecutive.
727 for (++i; i != 16; ++i)
728 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
731 // Check the rest of the elements to see if they are consecutive.
732 for (++i; i != 16; ++i)
733 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
739 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
740 /// specifies a splat of a single element that is suitable for input to
741 /// VSPLTB/VSPLTH/VSPLTW.
742 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
743 assert(N->getValueType(0) == MVT::v16i8 &&
744 (EltSize == 1 || EltSize == 2 || EltSize == 4));
746 // This is a splat operation if each element of the permute is the same, and
747 // if the value doesn't reference the second vector.
748 unsigned ElementBase = N->getMaskElt(0);
750 // FIXME: Handle UNDEF elements too!
751 if (ElementBase >= 16)
754 // Check that the indices are consecutive, in the case of a multi-byte element
755 // splatted with a v16i8 mask.
756 for (unsigned i = 1; i != EltSize; ++i)
757 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
760 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
761 if (N->getMaskElt(i) < 0) continue;
762 for (unsigned j = 0; j != EltSize; ++j)
763 if (N->getMaskElt(i+j) != N->getMaskElt(j))
769 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
771 bool PPC::isAllNegativeZeroVector(SDNode *N) {
772 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
774 APInt APVal, APUndef;
778 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
779 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
780 return CFP->getValueAPF().isNegZero();
785 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
786 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
787 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
788 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
789 assert(isSplatShuffleMask(SVOp, EltSize));
790 return SVOp->getMaskElt(0) / EltSize;
793 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
794 /// by using a vspltis[bhw] instruction of the specified element size, return
795 /// the constant being splatted. The ByteSize field indicates the number of
796 /// bytes of each element [124] -> [bhw].
797 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
800 // If ByteSize of the splat is bigger than the element size of the
801 // build_vector, then we have a case where we are checking for a splat where
802 // multiple elements of the buildvector are folded together into a single
803 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
804 unsigned EltSize = 16/N->getNumOperands();
805 if (EltSize < ByteSize) {
806 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
807 SDValue UniquedVals[4];
808 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
810 // See if all of the elements in the buildvector agree across.
811 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
812 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
813 // If the element isn't a constant, bail fully out.
814 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
817 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
818 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
819 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
820 return SDValue(); // no match.
823 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
824 // either constant or undef values that are identical for each chunk. See
825 // if these chunks can form into a larger vspltis*.
827 // Check to see if all of the leading entries are either 0 or -1. If
828 // neither, then this won't fit into the immediate field.
829 bool LeadingZero = true;
830 bool LeadingOnes = true;
831 for (unsigned i = 0; i != Multiple-1; ++i) {
832 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
834 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
835 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
837 // Finally, check the least significant entry.
839 if (UniquedVals[Multiple-1].getNode() == 0)
840 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
841 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
843 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
846 if (UniquedVals[Multiple-1].getNode() == 0)
847 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
848 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
849 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
850 return DAG.getTargetConstant(Val, MVT::i32);
856 // Check to see if this buildvec has a single non-undef value in its elements.
857 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
858 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
859 if (OpVal.getNode() == 0)
860 OpVal = N->getOperand(i);
861 else if (OpVal != N->getOperand(i))
865 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
867 unsigned ValSizeInBytes = EltSize;
869 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
870 Value = CN->getZExtValue();
871 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
872 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
873 Value = FloatToBits(CN->getValueAPF().convertToFloat());
876 // If the splat value is larger than the element value, then we can never do
877 // this splat. The only case that we could fit the replicated bits into our
878 // immediate field for would be zero, and we prefer to use vxor for it.
879 if (ValSizeInBytes < ByteSize) return SDValue();
881 // If the element value is larger than the splat value, cut it in half and
882 // check to see if the two halves are equal. Continue doing this until we
883 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
884 while (ValSizeInBytes > ByteSize) {
885 ValSizeInBytes >>= 1;
887 // If the top half equals the bottom half, we're still ok.
888 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
889 (Value & ((1 << (8*ValSizeInBytes))-1)))
893 // Properly sign extend the value.
894 int MaskVal = SignExtend32(Value, ByteSize * 8);
896 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
897 if (MaskVal == 0) return SDValue();
899 // Finally, if this value fits in a 5 bit sext field, return it
900 if (SignExtend32<5>(MaskVal) == MaskVal)
901 return DAG.getTargetConstant(MaskVal, MVT::i32);
905 //===----------------------------------------------------------------------===//
906 // Addressing Mode Selection
907 //===----------------------------------------------------------------------===//
909 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
910 /// or 64-bit immediate, and if the value can be accurately represented as a
911 /// sign extension from a 16-bit value. If so, this returns true and the
913 static bool isIntS16Immediate(SDNode *N, short &Imm) {
914 if (N->getOpcode() != ISD::Constant)
917 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
918 if (N->getValueType(0) == MVT::i32)
919 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
921 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
923 static bool isIntS16Immediate(SDValue Op, short &Imm) {
924 return isIntS16Immediate(Op.getNode(), Imm);
928 /// SelectAddressRegReg - Given the specified addressed, check to see if it
929 /// can be represented as an indexed [r+r] operation. Returns false if it
930 /// can be more efficiently represented with [r+imm].
931 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
933 SelectionDAG &DAG) const {
935 if (N.getOpcode() == ISD::ADD) {
936 if (isIntS16Immediate(N.getOperand(1), imm))
938 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
941 Base = N.getOperand(0);
942 Index = N.getOperand(1);
944 } else if (N.getOpcode() == ISD::OR) {
945 if (isIntS16Immediate(N.getOperand(1), imm))
946 return false; // r+i can fold it if we can.
948 // If this is an or of disjoint bitfields, we can codegen this as an add
949 // (for better address arithmetic) if the LHS and RHS of the OR are provably
951 APInt LHSKnownZero, LHSKnownOne;
952 APInt RHSKnownZero, RHSKnownOne;
953 DAG.ComputeMaskedBits(N.getOperand(0),
954 LHSKnownZero, LHSKnownOne);
956 if (LHSKnownZero.getBoolValue()) {
957 DAG.ComputeMaskedBits(N.getOperand(1),
958 RHSKnownZero, RHSKnownOne);
959 // If all of the bits are known zero on the LHS or RHS, the add won't
961 if (~(LHSKnownZero | RHSKnownZero) == 0) {
962 Base = N.getOperand(0);
963 Index = N.getOperand(1);
972 /// Returns true if the address N can be represented by a base register plus
973 /// a signed 16-bit displacement [r+imm], and if it is not better
974 /// represented as reg+reg.
975 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
977 SelectionDAG &DAG) const {
978 // FIXME dl should come from parent load or store, not from address
979 DebugLoc dl = N.getDebugLoc();
980 // If this can be more profitably realized as r+r, fail.
981 if (SelectAddressRegReg(N, Disp, Base, DAG))
984 if (N.getOpcode() == ISD::ADD) {
986 if (isIntS16Immediate(N.getOperand(1), imm)) {
987 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
988 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
989 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
991 Base = N.getOperand(0);
993 return true; // [r+i]
994 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
995 // Match LOAD (ADD (X, Lo(G))).
996 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
997 && "Cannot handle constant offsets yet!");
998 Disp = N.getOperand(1).getOperand(0); // The global address.
999 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1000 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
1001 Disp.getOpcode() == ISD::TargetConstantPool ||
1002 Disp.getOpcode() == ISD::TargetJumpTable);
1003 Base = N.getOperand(0);
1004 return true; // [&g+r]
1006 } else if (N.getOpcode() == ISD::OR) {
1008 if (isIntS16Immediate(N.getOperand(1), imm)) {
1009 // If this is an or of disjoint bitfields, we can codegen this as an add
1010 // (for better address arithmetic) if the LHS and RHS of the OR are
1011 // provably disjoint.
1012 APInt LHSKnownZero, LHSKnownOne;
1013 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1015 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1016 // If all of the bits are known zero on the LHS or RHS, the add won't
1018 Base = N.getOperand(0);
1019 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
1023 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1024 // Loading from a constant address.
1026 // If this address fits entirely in a 16-bit sext immediate field, codegen
1029 if (isIntS16Immediate(CN, Imm)) {
1030 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
1031 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1032 CN->getValueType(0));
1036 // Handle 32-bit sext immediates with LIS + addr mode.
1037 if (CN->getValueType(0) == MVT::i32 ||
1038 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1039 int Addr = (int)CN->getZExtValue();
1041 // Otherwise, break this down into an LIS + disp.
1042 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1044 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1045 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1046 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1051 Disp = DAG.getTargetConstant(0, getPointerTy());
1052 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1053 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1056 return true; // [r+0]
1059 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1060 /// represented as an indexed [r+r] operation.
1061 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1063 SelectionDAG &DAG) const {
1064 // Check to see if we can easily represent this as an [r+r] address. This
1065 // will fail if it thinks that the address is more profitably represented as
1066 // reg+imm, e.g. where imm = 0.
1067 if (SelectAddressRegReg(N, Base, Index, DAG))
1070 // If the operand is an addition, always emit this as [r+r], since this is
1071 // better (for code size, and execution, as the memop does the add for free)
1072 // than emitting an explicit add.
1073 if (N.getOpcode() == ISD::ADD) {
1074 Base = N.getOperand(0);
1075 Index = N.getOperand(1);
1079 // Otherwise, do it the hard way, using R0 as the base register.
1080 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1086 /// SelectAddressRegImmShift - Returns true if the address N can be
1087 /// represented by a base register plus a signed 14-bit displacement
1088 /// [r+imm*4]. Suitable for use by STD and friends.
1089 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1091 SelectionDAG &DAG) const {
1092 // FIXME dl should come from the parent load or store, not the address
1093 DebugLoc dl = N.getDebugLoc();
1094 // If this can be more profitably realized as r+r, fail.
1095 if (SelectAddressRegReg(N, Disp, Base, DAG))
1098 if (N.getOpcode() == ISD::ADD) {
1100 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1101 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1102 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1103 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1105 Base = N.getOperand(0);
1107 return true; // [r+i]
1108 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1109 // Match LOAD (ADD (X, Lo(G))).
1110 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1111 && "Cannot handle constant offsets yet!");
1112 Disp = N.getOperand(1).getOperand(0); // The global address.
1113 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1114 Disp.getOpcode() == ISD::TargetConstantPool ||
1115 Disp.getOpcode() == ISD::TargetJumpTable);
1116 Base = N.getOperand(0);
1117 return true; // [&g+r]
1119 } else if (N.getOpcode() == ISD::OR) {
1121 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1122 // If this is an or of disjoint bitfields, we can codegen this as an add
1123 // (for better address arithmetic) if the LHS and RHS of the OR are
1124 // provably disjoint.
1125 APInt LHSKnownZero, LHSKnownOne;
1126 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1127 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1128 // If all of the bits are known zero on the LHS or RHS, the add won't
1130 Base = N.getOperand(0);
1131 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1135 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1136 // Loading from a constant address. Verify low two bits are clear.
1137 if ((CN->getZExtValue() & 3) == 0) {
1138 // If this address fits entirely in a 14-bit sext immediate field, codegen
1141 if (isIntS16Immediate(CN, Imm)) {
1142 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1143 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1144 CN->getValueType(0));
1148 // Fold the low-part of 32-bit absolute addresses into addr mode.
1149 if (CN->getValueType(0) == MVT::i32 ||
1150 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1151 int Addr = (int)CN->getZExtValue();
1153 // Otherwise, break this down into an LIS + disp.
1154 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1155 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1156 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1157 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1163 Disp = DAG.getTargetConstant(0, getPointerTy());
1164 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1165 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1168 return true; // [r+0]
1172 /// getPreIndexedAddressParts - returns true by value, base pointer and
1173 /// offset pointer and addressing mode by reference if the node's address
1174 /// can be legally represented as pre-indexed load / store address.
1175 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1177 ISD::MemIndexedMode &AM,
1178 SelectionDAG &DAG) const {
1179 if (DisablePPCPreinc) return false;
1183 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1184 Ptr = LD->getBasePtr();
1185 VT = LD->getMemoryVT();
1187 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1188 Ptr = ST->getBasePtr();
1189 VT = ST->getMemoryVT();
1193 // PowerPC doesn't have preinc load/store instructions for vectors.
1197 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1202 // LDU/STU use reg+imm*4, others use reg+imm.
1203 if (VT != MVT::i64) {
1205 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1209 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1213 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1214 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1215 // sext i32 to i64 when addr mode is r+i.
1216 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1217 LD->getExtensionType() == ISD::SEXTLOAD &&
1218 isa<ConstantSDNode>(Offset))
1226 //===----------------------------------------------------------------------===//
1227 // LowerOperation implementation
1228 //===----------------------------------------------------------------------===//
1230 /// GetLabelAccessInfo - Return true if we should reference labels using a
1231 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1232 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1233 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1234 HiOpFlags = PPCII::MO_HA16;
1235 LoOpFlags = PPCII::MO_LO16;
1237 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1238 // non-darwin platform. We don't support PIC on other platforms yet.
1239 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1240 TM.getSubtarget<PPCSubtarget>().isDarwin();
1242 HiOpFlags |= PPCII::MO_PIC_FLAG;
1243 LoOpFlags |= PPCII::MO_PIC_FLAG;
1246 // If this is a reference to a global value that requires a non-lazy-ptr, make
1247 // sure that instruction lowering adds it.
1248 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1249 HiOpFlags |= PPCII::MO_NLP_FLAG;
1250 LoOpFlags |= PPCII::MO_NLP_FLAG;
1252 if (GV->hasHiddenVisibility()) {
1253 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1254 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1261 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1262 SelectionDAG &DAG) {
1263 EVT PtrVT = HiPart.getValueType();
1264 SDValue Zero = DAG.getConstant(0, PtrVT);
1265 DebugLoc DL = HiPart.getDebugLoc();
1267 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1268 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1270 // With PIC, the first instruction is actually "GR+hi(&G)".
1272 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1273 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1275 // Generate non-pic code that has direct accesses to the constant pool.
1276 // The address of the global is just (hi(&g)+lo(&g)).
1277 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1280 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1281 SelectionDAG &DAG) const {
1282 EVT PtrVT = Op.getValueType();
1283 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1284 const Constant *C = CP->getConstVal();
1286 // 64-bit SVR4 ABI code is always position-independent.
1287 // The actual address of the GlobalValue is stored in the TOC.
1288 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1289 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1290 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1291 DAG.getRegister(PPC::X2, MVT::i64));
1294 unsigned MOHiFlag, MOLoFlag;
1295 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1297 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1299 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1300 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1303 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1304 EVT PtrVT = Op.getValueType();
1305 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1307 // 64-bit SVR4 ABI code is always position-independent.
1308 // The actual address of the GlobalValue is stored in the TOC.
1309 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1310 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1311 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1312 DAG.getRegister(PPC::X2, MVT::i64));
1315 unsigned MOHiFlag, MOLoFlag;
1316 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1317 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1318 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1319 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1322 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1323 SelectionDAG &DAG) const {
1324 EVT PtrVT = Op.getValueType();
1326 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1328 unsigned MOHiFlag, MOLoFlag;
1329 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1330 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1331 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1332 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1335 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1336 SelectionDAG &DAG) const {
1338 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1339 DebugLoc dl = GA->getDebugLoc();
1340 const GlobalValue *GV = GA->getGlobal();
1341 EVT PtrVT = getPointerTy();
1342 bool is64bit = PPCSubTarget.isPPC64();
1344 TLSModel::Model Model = getTargetMachine().getTLSModel(GV);
1346 if (Model == TLSModel::LocalExec) {
1347 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1348 PPCII::MO_TPREL16_HA);
1349 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1350 PPCII::MO_TPREL16_LO);
1351 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1352 is64bit ? MVT::i64 : MVT::i32);
1353 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1354 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1358 llvm_unreachable("only local-exec is currently supported for ppc32");
1360 if (Model == TLSModel::InitialExec) {
1361 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1362 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1363 SDValue TPOffsetHi = DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl,
1364 PtrVT, GOTReg, TGA);
1365 SDValue TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl,
1366 PtrVT, TGA, TPOffsetHi);
1367 return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGA);
1370 if (Model == TLSModel::GeneralDynamic) {
1371 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1372 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1373 SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
1375 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSGD_L, dl, PtrVT,
1378 // We need a chain node, and don't have one handy. The underlying
1379 // call has no side effects, so using the function entry node
1381 SDValue Chain = DAG.getEntryNode();
1382 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
1383 SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
1384 SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLS_ADDR, dl,
1385 PtrVT, ParmReg, TGA);
1386 // The return value from GET_TLS_ADDR really is in X3 already, but
1387 // some hacks are needed here to tie everything together. The extra
1388 // copies dissolve during subsequent transforms.
1389 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
1390 return DAG.getCopyFromReg(Chain, dl, PPC::X3, PtrVT);
1393 if (Model == TLSModel::LocalDynamic) {
1394 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1395 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1396 SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
1398 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSLD_L, dl, PtrVT,
1401 // We need a chain node, and don't have one handy. The underlying
1402 // call has no side effects, so using the function entry node
1404 SDValue Chain = DAG.getEntryNode();
1405 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
1406 SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
1407 SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLSLD_ADDR, dl,
1408 PtrVT, ParmReg, TGA);
1409 // The return value from GET_TLSLD_ADDR really is in X3 already, but
1410 // some hacks are needed here to tie everything together. The extra
1411 // copies dissolve during subsequent transforms.
1412 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
1413 SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, PtrVT,
1414 Chain, ParmReg, TGA);
1415 return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
1418 llvm_unreachable("Unknown TLS model!");
1421 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1422 SelectionDAG &DAG) const {
1423 EVT PtrVT = Op.getValueType();
1424 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1425 DebugLoc DL = GSDN->getDebugLoc();
1426 const GlobalValue *GV = GSDN->getGlobal();
1428 // 64-bit SVR4 ABI code is always position-independent.
1429 // The actual address of the GlobalValue is stored in the TOC.
1430 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1431 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1432 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1433 DAG.getRegister(PPC::X2, MVT::i64));
1436 unsigned MOHiFlag, MOLoFlag;
1437 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1440 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1442 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1444 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1446 // If the global reference is actually to a non-lazy-pointer, we have to do an
1447 // extra load to get the address of the global.
1448 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1449 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1450 false, false, false, 0);
1454 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1455 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1456 DebugLoc dl = Op.getDebugLoc();
1458 // If we're comparing for equality to zero, expose the fact that this is
1459 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1460 // fold the new nodes.
1461 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1462 if (C->isNullValue() && CC == ISD::SETEQ) {
1463 EVT VT = Op.getOperand(0).getValueType();
1464 SDValue Zext = Op.getOperand(0);
1465 if (VT.bitsLT(MVT::i32)) {
1467 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1469 unsigned Log2b = Log2_32(VT.getSizeInBits());
1470 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1471 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1472 DAG.getConstant(Log2b, MVT::i32));
1473 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1475 // Leave comparisons against 0 and -1 alone for now, since they're usually
1476 // optimized. FIXME: revisit this when we can custom lower all setcc
1478 if (C->isAllOnesValue() || C->isNullValue())
1482 // If we have an integer seteq/setne, turn it into a compare against zero
1483 // by xor'ing the rhs with the lhs, which is faster than setting a
1484 // condition register, reading it back out, and masking the correct bit. The
1485 // normal approach here uses sub to do this instead of xor. Using xor exposes
1486 // the result to other bit-twiddling opportunities.
1487 EVT LHSVT = Op.getOperand(0).getValueType();
1488 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1489 EVT VT = Op.getValueType();
1490 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1492 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1497 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1498 const PPCSubtarget &Subtarget) const {
1499 SDNode *Node = Op.getNode();
1500 EVT VT = Node->getValueType(0);
1501 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1502 SDValue InChain = Node->getOperand(0);
1503 SDValue VAListPtr = Node->getOperand(1);
1504 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1505 DebugLoc dl = Node->getDebugLoc();
1507 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1510 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1511 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1513 InChain = GprIndex.getValue(1);
1515 if (VT == MVT::i64) {
1516 // Check if GprIndex is even
1517 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1518 DAG.getConstant(1, MVT::i32));
1519 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1520 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1521 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1522 DAG.getConstant(1, MVT::i32));
1523 // Align GprIndex to be even if it isn't
1524 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1528 // fpr index is 1 byte after gpr
1529 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1530 DAG.getConstant(1, MVT::i32));
1533 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1534 FprPtr, MachinePointerInfo(SV), MVT::i8,
1536 InChain = FprIndex.getValue(1);
1538 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1539 DAG.getConstant(8, MVT::i32));
1541 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1542 DAG.getConstant(4, MVT::i32));
1545 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1546 MachinePointerInfo(), false, false,
1548 InChain = OverflowArea.getValue(1);
1550 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1551 MachinePointerInfo(), false, false,
1553 InChain = RegSaveArea.getValue(1);
1555 // select overflow_area if index > 8
1556 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1557 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1559 // adjustment constant gpr_index * 4/8
1560 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1561 VT.isInteger() ? GprIndex : FprIndex,
1562 DAG.getConstant(VT.isInteger() ? 4 : 8,
1565 // OurReg = RegSaveArea + RegConstant
1566 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1569 // Floating types are 32 bytes into RegSaveArea
1570 if (VT.isFloatingPoint())
1571 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1572 DAG.getConstant(32, MVT::i32));
1574 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1575 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1576 VT.isInteger() ? GprIndex : FprIndex,
1577 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1580 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1581 VT.isInteger() ? VAListPtr : FprPtr,
1582 MachinePointerInfo(SV),
1583 MVT::i8, false, false, 0);
1585 // determine if we should load from reg_save_area or overflow_area
1586 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1588 // increase overflow_area by 4/8 if gpr/fpr > 8
1589 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1590 DAG.getConstant(VT.isInteger() ? 4 : 8,
1593 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1596 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1598 MachinePointerInfo(),
1599 MVT::i32, false, false, 0);
1601 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1602 false, false, false, 0);
1605 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1606 SelectionDAG &DAG) const {
1607 return Op.getOperand(0);
1610 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1611 SelectionDAG &DAG) const {
1612 SDValue Chain = Op.getOperand(0);
1613 SDValue Trmp = Op.getOperand(1); // trampoline
1614 SDValue FPtr = Op.getOperand(2); // nested function
1615 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1616 DebugLoc dl = Op.getDebugLoc();
1618 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1619 bool isPPC64 = (PtrVT == MVT::i64);
1621 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1624 TargetLowering::ArgListTy Args;
1625 TargetLowering::ArgListEntry Entry;
1627 Entry.Ty = IntPtrTy;
1628 Entry.Node = Trmp; Args.push_back(Entry);
1630 // TrampSize == (isPPC64 ? 48 : 40);
1631 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1632 isPPC64 ? MVT::i64 : MVT::i32);
1633 Args.push_back(Entry);
1635 Entry.Node = FPtr; Args.push_back(Entry);
1636 Entry.Node = Nest; Args.push_back(Entry);
1638 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1639 TargetLowering::CallLoweringInfo CLI(Chain,
1640 Type::getVoidTy(*DAG.getContext()),
1641 false, false, false, false, 0,
1643 /*isTailCall=*/false,
1644 /*doesNotRet=*/false,
1645 /*isReturnValueUsed=*/true,
1646 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1648 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1650 return CallResult.second;
1653 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1654 const PPCSubtarget &Subtarget) const {
1655 MachineFunction &MF = DAG.getMachineFunction();
1656 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1658 DebugLoc dl = Op.getDebugLoc();
1660 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1661 // vastart just stores the address of the VarArgsFrameIndex slot into the
1662 // memory location argument.
1663 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1664 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1665 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1666 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1667 MachinePointerInfo(SV),
1671 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1672 // We suppose the given va_list is already allocated.
1675 // char gpr; /* index into the array of 8 GPRs
1676 // * stored in the register save area
1677 // * gpr=0 corresponds to r3,
1678 // * gpr=1 to r4, etc.
1680 // char fpr; /* index into the array of 8 FPRs
1681 // * stored in the register save area
1682 // * fpr=0 corresponds to f1,
1683 // * fpr=1 to f2, etc.
1685 // char *overflow_arg_area;
1686 // /* location on stack that holds
1687 // * the next overflow argument
1689 // char *reg_save_area;
1690 // /* where r3:r10 and f1:f8 (if saved)
1696 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1697 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1700 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1702 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1704 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1707 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1708 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1710 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1711 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1713 uint64_t FPROffset = 1;
1714 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1716 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1718 // Store first byte : number of int regs
1719 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1721 MachinePointerInfo(SV),
1722 MVT::i8, false, false, 0);
1723 uint64_t nextOffset = FPROffset;
1724 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1727 // Store second byte : number of float regs
1728 SDValue secondStore =
1729 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1730 MachinePointerInfo(SV, nextOffset), MVT::i8,
1732 nextOffset += StackOffset;
1733 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1735 // Store second word : arguments given on stack
1736 SDValue thirdStore =
1737 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1738 MachinePointerInfo(SV, nextOffset),
1740 nextOffset += FrameOffset;
1741 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1743 // Store third word : arguments given in registers
1744 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1745 MachinePointerInfo(SV, nextOffset),
1750 #include "PPCGenCallingConv.inc"
1752 static bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1753 CCValAssign::LocInfo &LocInfo,
1754 ISD::ArgFlagsTy &ArgFlags,
1759 static bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1761 CCValAssign::LocInfo &LocInfo,
1762 ISD::ArgFlagsTy &ArgFlags,
1764 static const uint16_t ArgRegs[] = {
1765 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1766 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1768 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1770 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1772 // Skip one register if the first unallocated register has an even register
1773 // number and there are still argument registers available which have not been
1774 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1775 // need to skip a register if RegNum is odd.
1776 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1777 State.AllocateReg(ArgRegs[RegNum]);
1780 // Always return false here, as this function only makes sure that the first
1781 // unallocated register has an odd register number and does not actually
1782 // allocate a register for the current argument.
1786 static bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1788 CCValAssign::LocInfo &LocInfo,
1789 ISD::ArgFlagsTy &ArgFlags,
1791 static const uint16_t ArgRegs[] = {
1792 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1796 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1798 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1800 // If there is only one Floating-point register left we need to put both f64
1801 // values of a split ppc_fp128 value on the stack.
1802 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1803 State.AllocateReg(ArgRegs[RegNum]);
1806 // Always return false here, as this function only makes sure that the two f64
1807 // values a ppc_fp128 value is split into are both passed in registers or both
1808 // passed on the stack and does not actually allocate a register for the
1809 // current argument.
1813 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1815 static const uint16_t *GetFPR() {
1816 static const uint16_t FPR[] = {
1817 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1818 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1824 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1826 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1827 unsigned PtrByteSize) {
1828 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1829 if (Flags.isByVal())
1830 ArgSize = Flags.getByValSize();
1831 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1837 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1838 CallingConv::ID CallConv, bool isVarArg,
1839 const SmallVectorImpl<ISD::InputArg>
1841 DebugLoc dl, SelectionDAG &DAG,
1842 SmallVectorImpl<SDValue> &InVals)
1844 if (PPCSubTarget.isSVR4ABI()) {
1845 if (PPCSubTarget.isPPC64())
1846 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1849 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1852 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1858 PPCTargetLowering::LowerFormalArguments_32SVR4(
1860 CallingConv::ID CallConv, bool isVarArg,
1861 const SmallVectorImpl<ISD::InputArg>
1863 DebugLoc dl, SelectionDAG &DAG,
1864 SmallVectorImpl<SDValue> &InVals) const {
1866 // 32-bit SVR4 ABI Stack Frame Layout:
1867 // +-----------------------------------+
1868 // +--> | Back chain |
1869 // | +-----------------------------------+
1870 // | | Floating-point register save area |
1871 // | +-----------------------------------+
1872 // | | General register save area |
1873 // | +-----------------------------------+
1874 // | | CR save word |
1875 // | +-----------------------------------+
1876 // | | VRSAVE save word |
1877 // | +-----------------------------------+
1878 // | | Alignment padding |
1879 // | +-----------------------------------+
1880 // | | Vector register save area |
1881 // | +-----------------------------------+
1882 // | | Local variable space |
1883 // | +-----------------------------------+
1884 // | | Parameter list area |
1885 // | +-----------------------------------+
1886 // | | LR save word |
1887 // | +-----------------------------------+
1888 // SP--> +--- | Back chain |
1889 // +-----------------------------------+
1892 // System V Application Binary Interface PowerPC Processor Supplement
1893 // AltiVec Technology Programming Interface Manual
1895 MachineFunction &MF = DAG.getMachineFunction();
1896 MachineFrameInfo *MFI = MF.getFrameInfo();
1897 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1899 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1900 // Potential tail calls could cause overwriting of argument stack slots.
1901 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1902 (CallConv == CallingConv::Fast));
1903 unsigned PtrByteSize = 4;
1905 // Assign locations to all of the incoming arguments.
1906 SmallVector<CCValAssign, 16> ArgLocs;
1907 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1908 getTargetMachine(), ArgLocs, *DAG.getContext());
1910 // Reserve space for the linkage area on the stack.
1911 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1913 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4);
1915 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1916 CCValAssign &VA = ArgLocs[i];
1918 // Arguments stored in registers.
1919 if (VA.isRegLoc()) {
1920 const TargetRegisterClass *RC;
1921 EVT ValVT = VA.getValVT();
1923 switch (ValVT.getSimpleVT().SimpleTy) {
1925 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1927 RC = &PPC::GPRCRegClass;
1930 RC = &PPC::F4RCRegClass;
1933 RC = &PPC::F8RCRegClass;
1939 RC = &PPC::VRRCRegClass;
1943 // Transform the arguments stored in physical registers into virtual ones.
1944 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1945 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1947 InVals.push_back(ArgValue);
1949 // Argument stored in memory.
1950 assert(VA.isMemLoc());
1952 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1953 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1956 // Create load nodes to retrieve arguments from the stack.
1957 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1958 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1959 MachinePointerInfo(),
1960 false, false, false, 0));
1964 // Assign locations to all of the incoming aggregate by value arguments.
1965 // Aggregates passed by value are stored in the local variable space of the
1966 // caller's stack frame, right above the parameter list area.
1967 SmallVector<CCValAssign, 16> ByValArgLocs;
1968 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1969 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1971 // Reserve stack space for the allocations in CCInfo.
1972 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1974 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal);
1976 // Area that is at least reserved in the caller of this function.
1977 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1979 // Set the size that is at least reserved in caller of this function. Tail
1980 // call optimized function's reserved stack space needs to be aligned so that
1981 // taking the difference between two stack areas will result in an aligned
1983 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1986 std::max(MinReservedArea,
1987 PPCFrameLowering::getMinCallFrameSize(false, false));
1989 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1990 getStackAlignment();
1991 unsigned AlignMask = TargetAlign-1;
1992 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1994 FI->setMinReservedArea(MinReservedArea);
1996 SmallVector<SDValue, 8> MemOps;
1998 // If the function takes variable number of arguments, make a frame index for
1999 // the start of the first vararg value... for expansion of llvm.va_start.
2001 static const uint16_t GPArgRegs[] = {
2002 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2003 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2005 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
2007 static const uint16_t FPArgRegs[] = {
2008 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2011 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
2013 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
2015 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
2018 // Make room for NumGPArgRegs and NumFPArgRegs.
2019 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
2020 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
2022 FuncInfo->setVarArgsStackOffset(
2023 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2024 CCInfo.getNextStackOffset(), true));
2026 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
2027 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2029 // The fixed integer arguments of a variadic function are stored to the
2030 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
2031 // the result of va_next.
2032 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
2033 // Get an existing live-in vreg, or add a new one.
2034 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
2036 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
2038 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2039 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2040 MachinePointerInfo(), false, false, 0);
2041 MemOps.push_back(Store);
2042 // Increment the address by four for the next argument to store
2043 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2044 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2047 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
2049 // The double arguments are stored to the VarArgsFrameIndex
2051 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
2052 // Get an existing live-in vreg, or add a new one.
2053 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
2055 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
2057 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
2058 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2059 MachinePointerInfo(), false, false, 0);
2060 MemOps.push_back(Store);
2061 // Increment the address by eight for the next argument to store
2062 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
2064 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2068 if (!MemOps.empty())
2069 Chain = DAG.getNode(ISD::TokenFactor, dl,
2070 MVT::Other, &MemOps[0], MemOps.size());
2075 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2076 // value to MVT::i64 and then truncate to the correct register size.
2078 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
2079 SelectionDAG &DAG, SDValue ArgVal,
2080 DebugLoc dl) const {
2082 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2083 DAG.getValueType(ObjectVT));
2084 else if (Flags.isZExt())
2085 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2086 DAG.getValueType(ObjectVT));
2088 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2091 // Set the size that is at least reserved in caller of this function. Tail
2092 // call optimized functions' reserved stack space needs to be aligned so that
2093 // taking the difference between two stack areas will result in an aligned
2096 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
2097 unsigned nAltivecParamsAtEnd,
2098 unsigned MinReservedArea,
2099 bool isPPC64) const {
2100 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2101 // Add the Altivec parameters at the end, if needed.
2102 if (nAltivecParamsAtEnd) {
2103 MinReservedArea = ((MinReservedArea+15)/16)*16;
2104 MinReservedArea += 16*nAltivecParamsAtEnd;
2107 std::max(MinReservedArea,
2108 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2109 unsigned TargetAlign
2110 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2111 getStackAlignment();
2112 unsigned AlignMask = TargetAlign-1;
2113 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2114 FI->setMinReservedArea(MinReservedArea);
2118 PPCTargetLowering::LowerFormalArguments_64SVR4(
2120 CallingConv::ID CallConv, bool isVarArg,
2121 const SmallVectorImpl<ISD::InputArg>
2123 DebugLoc dl, SelectionDAG &DAG,
2124 SmallVectorImpl<SDValue> &InVals) const {
2125 // TODO: add description of PPC stack frame format, or at least some docs.
2127 MachineFunction &MF = DAG.getMachineFunction();
2128 MachineFrameInfo *MFI = MF.getFrameInfo();
2129 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2131 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2132 // Potential tail calls could cause overwriting of argument stack slots.
2133 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2134 (CallConv == CallingConv::Fast));
2135 unsigned PtrByteSize = 8;
2137 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2138 // Area that is at least reserved in caller of this function.
2139 unsigned MinReservedArea = ArgOffset;
2141 static const uint16_t GPR[] = {
2142 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2143 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2146 static const uint16_t *FPR = GetFPR();
2148 static const uint16_t VR[] = {
2149 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2150 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2153 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2154 const unsigned Num_FPR_Regs = 13;
2155 const unsigned Num_VR_Regs = array_lengthof(VR);
2157 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2159 // Add DAG nodes to load the arguments or copy them out of registers. On
2160 // entry to a function on PPC, the arguments start after the linkage area,
2161 // although the first ones are often in registers.
2163 SmallVector<SDValue, 8> MemOps;
2164 unsigned nAltivecParamsAtEnd = 0;
2165 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2166 unsigned CurArgIdx = 0;
2167 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2169 bool needsLoad = false;
2170 EVT ObjectVT = Ins[ArgNo].VT;
2171 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2172 unsigned ArgSize = ObjSize;
2173 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2174 std::advance(FuncArg, Ins[ArgNo].OrigArgIndex - CurArgIdx);
2175 CurArgIdx = Ins[ArgNo].OrigArgIndex;
2177 unsigned CurArgOffset = ArgOffset;
2179 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2180 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2181 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2183 MinReservedArea = ((MinReservedArea+15)/16)*16;
2184 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2188 nAltivecParamsAtEnd++;
2190 // Calculate min reserved area.
2191 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2195 // FIXME the codegen can be much improved in some cases.
2196 // We do not have to keep everything in memory.
2197 if (Flags.isByVal()) {
2198 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2199 ObjSize = Flags.getByValSize();
2200 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2201 // Empty aggregate parameters do not take up registers. Examples:
2205 // etc. However, we have to provide a place-holder in InVals, so
2206 // pretend we have an 8-byte item at the current address for that
2209 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2210 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2211 InVals.push_back(FIN);
2214 // All aggregates smaller than 8 bytes must be passed right-justified.
2215 if (ObjSize < PtrByteSize)
2216 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2217 // The value of the object is its address.
2218 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2219 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2220 InVals.push_back(FIN);
2223 if (GPR_idx != Num_GPR_Regs) {
2224 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2225 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2228 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2229 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2230 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2231 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2232 MachinePointerInfo(FuncArg, CurArgOffset),
2233 ObjType, false, false, 0);
2235 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2236 // store the whole register as-is to the parameter save area
2237 // slot. The address of the parameter was already calculated
2238 // above (InVals.push_back(FIN)) to be the right-justified
2239 // offset within the slot. For this store, we need a new
2240 // frame index that points at the beginning of the slot.
2241 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2242 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2243 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2244 MachinePointerInfo(FuncArg, ArgOffset),
2248 MemOps.push_back(Store);
2251 // Whether we copied from a register or not, advance the offset
2252 // into the parameter save area by a full doubleword.
2253 ArgOffset += PtrByteSize;
2257 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2258 // Store whatever pieces of the object are in registers
2259 // to memory. ArgOffset will be the address of the beginning
2261 if (GPR_idx != Num_GPR_Regs) {
2263 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2264 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2265 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2266 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2267 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2268 MachinePointerInfo(FuncArg, ArgOffset),
2270 MemOps.push_back(Store);
2272 ArgOffset += PtrByteSize;
2274 ArgOffset += ArgSize - j;
2281 switch (ObjectVT.getSimpleVT().SimpleTy) {
2282 default: llvm_unreachable("Unhandled argument type!");
2285 if (GPR_idx != Num_GPR_Regs) {
2286 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2287 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2289 if (ObjectVT == MVT::i32)
2290 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2291 // value to MVT::i64 and then truncate to the correct register size.
2292 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2297 ArgSize = PtrByteSize;
2304 // Every 8 bytes of argument space consumes one of the GPRs available for
2305 // argument passing.
2306 if (GPR_idx != Num_GPR_Regs) {
2309 if (FPR_idx != Num_FPR_Regs) {
2312 if (ObjectVT == MVT::f32)
2313 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2315 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2317 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2321 ArgSize = PtrByteSize;
2330 // Note that vector arguments in registers don't reserve stack space,
2331 // except in varargs functions.
2332 if (VR_idx != Num_VR_Regs) {
2333 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2334 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2336 while ((ArgOffset % 16) != 0) {
2337 ArgOffset += PtrByteSize;
2338 if (GPR_idx != Num_GPR_Regs)
2342 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2346 // Vectors are aligned.
2347 ArgOffset = ((ArgOffset+15)/16)*16;
2348 CurArgOffset = ArgOffset;
2355 // We need to load the argument to a virtual register if we determined
2356 // above that we ran out of physical registers of the appropriate type.
2358 int FI = MFI->CreateFixedObject(ObjSize,
2359 CurArgOffset + (ArgSize - ObjSize),
2361 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2362 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2363 false, false, false, 0);
2366 InVals.push_back(ArgVal);
2369 // Set the size that is at least reserved in caller of this function. Tail
2370 // call optimized functions' reserved stack space needs to be aligned so that
2371 // taking the difference between two stack areas will result in an aligned
2373 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2375 // If the function takes variable number of arguments, make a frame index for
2376 // the start of the first vararg value... for expansion of llvm.va_start.
2378 int Depth = ArgOffset;
2380 FuncInfo->setVarArgsFrameIndex(
2381 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2382 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2384 // If this function is vararg, store any remaining integer argument regs
2385 // to their spots on the stack so that they may be loaded by deferencing the
2386 // result of va_next.
2387 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2388 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2389 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2390 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2391 MachinePointerInfo(), false, false, 0);
2392 MemOps.push_back(Store);
2393 // Increment the address by four for the next argument to store
2394 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2395 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2399 if (!MemOps.empty())
2400 Chain = DAG.getNode(ISD::TokenFactor, dl,
2401 MVT::Other, &MemOps[0], MemOps.size());
2407 PPCTargetLowering::LowerFormalArguments_Darwin(
2409 CallingConv::ID CallConv, bool isVarArg,
2410 const SmallVectorImpl<ISD::InputArg>
2412 DebugLoc dl, SelectionDAG &DAG,
2413 SmallVectorImpl<SDValue> &InVals) const {
2414 // TODO: add description of PPC stack frame format, or at least some docs.
2416 MachineFunction &MF = DAG.getMachineFunction();
2417 MachineFrameInfo *MFI = MF.getFrameInfo();
2418 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2420 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2421 bool isPPC64 = PtrVT == MVT::i64;
2422 // Potential tail calls could cause overwriting of argument stack slots.
2423 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2424 (CallConv == CallingConv::Fast));
2425 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2427 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2428 // Area that is at least reserved in caller of this function.
2429 unsigned MinReservedArea = ArgOffset;
2431 static const uint16_t GPR_32[] = { // 32-bit registers.
2432 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2433 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2435 static const uint16_t GPR_64[] = { // 64-bit registers.
2436 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2437 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2440 static const uint16_t *FPR = GetFPR();
2442 static const uint16_t VR[] = {
2443 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2444 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2447 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2448 const unsigned Num_FPR_Regs = 13;
2449 const unsigned Num_VR_Regs = array_lengthof( VR);
2451 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2453 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2455 // In 32-bit non-varargs functions, the stack space for vectors is after the
2456 // stack space for non-vectors. We do not use this space unless we have
2457 // too many vectors to fit in registers, something that only occurs in
2458 // constructed examples:), but we have to walk the arglist to figure
2459 // that out...for the pathological case, compute VecArgOffset as the
2460 // start of the vector parameter area. Computing VecArgOffset is the
2461 // entire point of the following loop.
2462 unsigned VecArgOffset = ArgOffset;
2463 if (!isVarArg && !isPPC64) {
2464 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2466 EVT ObjectVT = Ins[ArgNo].VT;
2467 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2469 if (Flags.isByVal()) {
2470 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2471 unsigned ObjSize = Flags.getByValSize();
2473 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2474 VecArgOffset += ArgSize;
2478 switch(ObjectVT.getSimpleVT().SimpleTy) {
2479 default: llvm_unreachable("Unhandled argument type!");
2484 case MVT::i64: // PPC64
2486 // FIXME: We are guaranteed to be !isPPC64 at this point.
2487 // Does MVT::i64 apply?
2494 // Nothing to do, we're only looking at Nonvector args here.
2499 // We've found where the vector parameter area in memory is. Skip the
2500 // first 12 parameters; these don't use that memory.
2501 VecArgOffset = ((VecArgOffset+15)/16)*16;
2502 VecArgOffset += 12*16;
2504 // Add DAG nodes to load the arguments or copy them out of registers. On
2505 // entry to a function on PPC, the arguments start after the linkage area,
2506 // although the first ones are often in registers.
2508 SmallVector<SDValue, 8> MemOps;
2509 unsigned nAltivecParamsAtEnd = 0;
2510 // FIXME: FuncArg and Ins[ArgNo] must reference the same argument.
2511 // When passing anonymous aggregates, this is currently not true.
2512 // See LowerFormalArguments_64SVR4 for a fix.
2513 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2514 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2516 bool needsLoad = false;
2517 EVT ObjectVT = Ins[ArgNo].VT;
2518 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2519 unsigned ArgSize = ObjSize;
2520 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2522 unsigned CurArgOffset = ArgOffset;
2524 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2525 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2526 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2527 if (isVarArg || isPPC64) {
2528 MinReservedArea = ((MinReservedArea+15)/16)*16;
2529 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2532 } else nAltivecParamsAtEnd++;
2534 // Calculate min reserved area.
2535 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2539 // FIXME the codegen can be much improved in some cases.
2540 // We do not have to keep everything in memory.
2541 if (Flags.isByVal()) {
2542 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2543 ObjSize = Flags.getByValSize();
2544 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2545 // Objects of size 1 and 2 are right justified, everything else is
2546 // left justified. This means the memory address is adjusted forwards.
2547 if (ObjSize==1 || ObjSize==2) {
2548 CurArgOffset = CurArgOffset + (4 - ObjSize);
2550 // The value of the object is its address.
2551 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2552 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2553 InVals.push_back(FIN);
2554 if (ObjSize==1 || ObjSize==2) {
2555 if (GPR_idx != Num_GPR_Regs) {
2558 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2560 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2561 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2562 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2563 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2564 MachinePointerInfo(FuncArg,
2566 ObjType, false, false, 0);
2567 MemOps.push_back(Store);
2571 ArgOffset += PtrByteSize;
2575 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2576 // Store whatever pieces of the object are in registers
2577 // to memory. ArgOffset will be the address of the beginning
2579 if (GPR_idx != Num_GPR_Regs) {
2582 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2584 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2585 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2586 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2587 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2588 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2589 MachinePointerInfo(FuncArg, ArgOffset),
2591 MemOps.push_back(Store);
2593 ArgOffset += PtrByteSize;
2595 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2602 switch (ObjectVT.getSimpleVT().SimpleTy) {
2603 default: llvm_unreachable("Unhandled argument type!");
2606 if (GPR_idx != Num_GPR_Regs) {
2607 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2608 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2612 ArgSize = PtrByteSize;
2614 // All int arguments reserve stack space in the Darwin ABI.
2615 ArgOffset += PtrByteSize;
2619 case MVT::i64: // PPC64
2620 if (GPR_idx != Num_GPR_Regs) {
2621 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2622 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2624 if (ObjectVT == MVT::i32)
2625 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2626 // value to MVT::i64 and then truncate to the correct register size.
2627 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2632 ArgSize = PtrByteSize;
2634 // All int arguments reserve stack space in the Darwin ABI.
2640 // Every 4 bytes of argument space consumes one of the GPRs available for
2641 // argument passing.
2642 if (GPR_idx != Num_GPR_Regs) {
2644 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2647 if (FPR_idx != Num_FPR_Regs) {
2650 if (ObjectVT == MVT::f32)
2651 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2653 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2655 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2661 // All FP arguments reserve stack space in the Darwin ABI.
2662 ArgOffset += isPPC64 ? 8 : ObjSize;
2668 // Note that vector arguments in registers don't reserve stack space,
2669 // except in varargs functions.
2670 if (VR_idx != Num_VR_Regs) {
2671 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2672 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2674 while ((ArgOffset % 16) != 0) {
2675 ArgOffset += PtrByteSize;
2676 if (GPR_idx != Num_GPR_Regs)
2680 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2684 if (!isVarArg && !isPPC64) {
2685 // Vectors go after all the nonvectors.
2686 CurArgOffset = VecArgOffset;
2689 // Vectors are aligned.
2690 ArgOffset = ((ArgOffset+15)/16)*16;
2691 CurArgOffset = ArgOffset;
2699 // We need to load the argument to a virtual register if we determined above
2700 // that we ran out of physical registers of the appropriate type.
2702 int FI = MFI->CreateFixedObject(ObjSize,
2703 CurArgOffset + (ArgSize - ObjSize),
2705 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2706 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2707 false, false, false, 0);
2710 InVals.push_back(ArgVal);
2713 // Set the size that is at least reserved in caller of this function. Tail
2714 // call optimized functions' reserved stack space needs to be aligned so that
2715 // taking the difference between two stack areas will result in an aligned
2717 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2719 // If the function takes variable number of arguments, make a frame index for
2720 // the start of the first vararg value... for expansion of llvm.va_start.
2722 int Depth = ArgOffset;
2724 FuncInfo->setVarArgsFrameIndex(
2725 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2727 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2729 // If this function is vararg, store any remaining integer argument regs
2730 // to their spots on the stack so that they may be loaded by deferencing the
2731 // result of va_next.
2732 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2736 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2738 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2740 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2741 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2742 MachinePointerInfo(), false, false, 0);
2743 MemOps.push_back(Store);
2744 // Increment the address by four for the next argument to store
2745 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2746 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2750 if (!MemOps.empty())
2751 Chain = DAG.getNode(ISD::TokenFactor, dl,
2752 MVT::Other, &MemOps[0], MemOps.size());
2757 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2758 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2760 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2764 const SmallVectorImpl<ISD::OutputArg>
2766 const SmallVectorImpl<SDValue> &OutVals,
2767 unsigned &nAltivecParamsAtEnd) {
2768 // Count how many bytes are to be pushed on the stack, including the linkage
2769 // area, and parameter passing area. We start with 24/48 bytes, which is
2770 // prereserved space for [SP][CR][LR][3 x unused].
2771 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2772 unsigned NumOps = Outs.size();
2773 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2775 // Add up all the space actually used.
2776 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2777 // they all go in registers, but we must reserve stack space for them for
2778 // possible use by the caller. In varargs or 64-bit calls, parameters are
2779 // assigned stack space in order, with padding so Altivec parameters are
2781 nAltivecParamsAtEnd = 0;
2782 for (unsigned i = 0; i != NumOps; ++i) {
2783 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2784 EVT ArgVT = Outs[i].VT;
2785 // Varargs Altivec parameters are padded to a 16 byte boundary.
2786 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2787 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2788 if (!isVarArg && !isPPC64) {
2789 // Non-varargs Altivec parameters go after all the non-Altivec
2790 // parameters; handle those later so we know how much padding we need.
2791 nAltivecParamsAtEnd++;
2794 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2795 NumBytes = ((NumBytes+15)/16)*16;
2797 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2800 // Allow for Altivec parameters at the end, if needed.
2801 if (nAltivecParamsAtEnd) {
2802 NumBytes = ((NumBytes+15)/16)*16;
2803 NumBytes += 16*nAltivecParamsAtEnd;
2806 // The prolog code of the callee may store up to 8 GPR argument registers to
2807 // the stack, allowing va_start to index over them in memory if its varargs.
2808 // Because we cannot tell if this is needed on the caller side, we have to
2809 // conservatively assume that it is needed. As such, make sure we have at
2810 // least enough stack space for the caller to store the 8 GPRs.
2811 NumBytes = std::max(NumBytes,
2812 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2814 // Tail call needs the stack to be aligned.
2815 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2816 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2817 getFrameLowering()->getStackAlignment();
2818 unsigned AlignMask = TargetAlign-1;
2819 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2825 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2826 /// adjusted to accommodate the arguments for the tailcall.
2827 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2828 unsigned ParamSize) {
2830 if (!isTailCall) return 0;
2832 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2833 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2834 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2835 // Remember only if the new adjustement is bigger.
2836 if (SPDiff < FI->getTailCallSPDelta())
2837 FI->setTailCallSPDelta(SPDiff);
2842 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2843 /// for tail call optimization. Targets which want to do tail call
2844 /// optimization should implement this function.
2846 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2847 CallingConv::ID CalleeCC,
2849 const SmallVectorImpl<ISD::InputArg> &Ins,
2850 SelectionDAG& DAG) const {
2851 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2854 // Variable argument functions are not supported.
2858 MachineFunction &MF = DAG.getMachineFunction();
2859 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2860 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2861 // Functions containing by val parameters are not supported.
2862 for (unsigned i = 0; i != Ins.size(); i++) {
2863 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2864 if (Flags.isByVal()) return false;
2867 // Non PIC/GOT tail calls are supported.
2868 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2871 // At the moment we can only do local tail calls (in same module, hidden
2872 // or protected) if we are generating PIC.
2873 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2874 return G->getGlobal()->hasHiddenVisibility()
2875 || G->getGlobal()->hasProtectedVisibility();
2881 /// isCallCompatibleAddress - Return the immediate to use if the specified
2882 /// 32-bit value is representable in the immediate field of a BxA instruction.
2883 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2884 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2887 int Addr = C->getZExtValue();
2888 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2889 SignExtend32<26>(Addr) != Addr)
2890 return 0; // Top 6 bits have to be sext of immediate.
2892 return DAG.getConstant((int)C->getZExtValue() >> 2,
2893 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2898 struct TailCallArgumentInfo {
2903 TailCallArgumentInfo() : FrameIdx(0) {}
2908 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2910 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2912 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2913 SmallVector<SDValue, 8> &MemOpChains,
2915 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2916 SDValue Arg = TailCallArgs[i].Arg;
2917 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2918 int FI = TailCallArgs[i].FrameIdx;
2919 // Store relative to framepointer.
2920 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2921 MachinePointerInfo::getFixedStack(FI),
2926 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2927 /// the appropriate stack slot for the tail call optimized function call.
2928 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2929 MachineFunction &MF,
2938 // Calculate the new stack slot for the return address.
2939 int SlotSize = isPPC64 ? 8 : 4;
2940 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2942 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2943 NewRetAddrLoc, true);
2944 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2945 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2946 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2947 MachinePointerInfo::getFixedStack(NewRetAddr),
2950 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2951 // slot as the FP is never overwritten.
2954 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2955 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2957 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2958 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2959 MachinePointerInfo::getFixedStack(NewFPIdx),
2966 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2967 /// the position of the argument.
2969 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2970 SDValue Arg, int SPDiff, unsigned ArgOffset,
2971 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2972 int Offset = ArgOffset + SPDiff;
2973 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2974 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2975 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2976 SDValue FIN = DAG.getFrameIndex(FI, VT);
2977 TailCallArgumentInfo Info;
2979 Info.FrameIdxOp = FIN;
2981 TailCallArguments.push_back(Info);
2984 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2985 /// stack slot. Returns the chain as result and the loaded frame pointers in
2986 /// LROpOut/FPOpout. Used when tail calling.
2987 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2993 DebugLoc dl) const {
2995 // Load the LR and FP stack slot for later adjusting.
2996 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2997 LROpOut = getReturnAddrFrameIndex(DAG);
2998 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2999 false, false, false, 0);
3000 Chain = SDValue(LROpOut.getNode(), 1);
3002 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
3003 // slot as the FP is never overwritten.
3005 FPOpOut = getFramePointerFrameIndex(DAG);
3006 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
3007 false, false, false, 0);
3008 Chain = SDValue(FPOpOut.getNode(), 1);
3014 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
3015 /// by "Src" to address "Dst" of size "Size". Alignment information is
3016 /// specified by the specific parameter attribute. The copy will be passed as
3017 /// a byval function parameter.
3018 /// Sometimes what we are copying is the end of a larger object, the part that
3019 /// does not fit in registers.
3021 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
3022 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
3024 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
3025 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
3026 false, false, MachinePointerInfo(0),
3027 MachinePointerInfo(0));
3030 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
3033 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
3034 SDValue Arg, SDValue PtrOff, int SPDiff,
3035 unsigned ArgOffset, bool isPPC64, bool isTailCall,
3036 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
3037 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
3039 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3044 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3046 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3047 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3048 DAG.getConstant(ArgOffset, PtrVT));
3050 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3051 MachinePointerInfo(), false, false, 0));
3052 // Calculate and remember argument location.
3053 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
3058 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
3059 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
3060 SDValue LROp, SDValue FPOp, bool isDarwinABI,
3061 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
3062 MachineFunction &MF = DAG.getMachineFunction();
3064 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
3065 // might overwrite each other in case of tail call optimization.
3066 SmallVector<SDValue, 8> MemOpChains2;
3067 // Do not flag preceding copytoreg stuff together with the following stuff.
3069 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
3071 if (!MemOpChains2.empty())
3072 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3073 &MemOpChains2[0], MemOpChains2.size());
3075 // Store the return address to the appropriate stack slot.
3076 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
3077 isPPC64, isDarwinABI, dl);
3079 // Emit callseq_end just before tailcall node.
3080 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3081 DAG.getIntPtrConstant(0, true), InFlag);
3082 InFlag = Chain.getValue(1);
3086 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
3087 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
3088 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
3089 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
3090 const PPCSubtarget &PPCSubTarget) {
3092 bool isPPC64 = PPCSubTarget.isPPC64();
3093 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3095 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3096 NodeTys.push_back(MVT::Other); // Returns a chain
3097 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
3099 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
3101 bool needIndirectCall = true;
3102 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3103 // If this is an absolute destination address, use the munged value.
3104 Callee = SDValue(Dest, 0);
3105 needIndirectCall = false;
3108 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3109 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
3110 // Use indirect calls for ALL functions calls in JIT mode, since the
3111 // far-call stubs may be outside relocation limits for a BL instruction.
3112 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
3113 unsigned OpFlags = 0;
3114 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3115 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3116 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3117 (G->getGlobal()->isDeclaration() ||
3118 G->getGlobal()->isWeakForLinker())) {
3119 // PC-relative references to external symbols should go through $stub,
3120 // unless we're building with the leopard linker or later, which
3121 // automatically synthesizes these stubs.
3122 OpFlags = PPCII::MO_DARWIN_STUB;
3125 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3126 // every direct call is) turn it into a TargetGlobalAddress /
3127 // TargetExternalSymbol node so that legalize doesn't hack it.
3128 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3129 Callee.getValueType(),
3131 needIndirectCall = false;
3135 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3136 unsigned char OpFlags = 0;
3138 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3139 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3140 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3141 // PC-relative references to external symbols should go through $stub,
3142 // unless we're building with the leopard linker or later, which
3143 // automatically synthesizes these stubs.
3144 OpFlags = PPCII::MO_DARWIN_STUB;
3147 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3149 needIndirectCall = false;
3152 if (needIndirectCall) {
3153 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3154 // to do the call, we can't use PPCISD::CALL.
3155 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3157 if (isSVR4ABI && isPPC64) {
3158 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3159 // entry point, but to the function descriptor (the function entry point
3160 // address is part of the function descriptor though).
3161 // The function descriptor is a three doubleword structure with the
3162 // following fields: function entry point, TOC base address and
3163 // environment pointer.
3164 // Thus for a call through a function pointer, the following actions need
3166 // 1. Save the TOC of the caller in the TOC save area of its stack
3167 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3168 // 2. Load the address of the function entry point from the function
3170 // 3. Load the TOC of the callee from the function descriptor into r2.
3171 // 4. Load the environment pointer from the function descriptor into
3173 // 5. Branch to the function entry point address.
3174 // 6. On return of the callee, the TOC of the caller needs to be
3175 // restored (this is done in FinishCall()).
3177 // All those operations are flagged together to ensure that no other
3178 // operations can be scheduled in between. E.g. without flagging the
3179 // operations together, a TOC access in the caller could be scheduled
3180 // between the load of the callee TOC and the branch to the callee, which
3181 // results in the TOC access going through the TOC of the callee instead
3182 // of going through the TOC of the caller, which leads to incorrect code.
3184 // Load the address of the function entry point from the function
3186 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3187 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3188 InFlag.getNode() ? 3 : 2);
3189 Chain = LoadFuncPtr.getValue(1);
3190 InFlag = LoadFuncPtr.getValue(2);
3192 // Load environment pointer into r11.
3193 // Offset of the environment pointer within the function descriptor.
3194 SDValue PtrOff = DAG.getIntPtrConstant(16);
3196 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3197 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3199 Chain = LoadEnvPtr.getValue(1);
3200 InFlag = LoadEnvPtr.getValue(2);
3202 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3204 Chain = EnvVal.getValue(0);
3205 InFlag = EnvVal.getValue(1);
3207 // Load TOC of the callee into r2. We are using a target-specific load
3208 // with r2 hard coded, because the result of a target-independent load
3209 // would never go directly into r2, since r2 is a reserved register (which
3210 // prevents the register allocator from allocating it), resulting in an
3211 // additional register being allocated and an unnecessary move instruction
3213 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3214 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3216 Chain = LoadTOCPtr.getValue(0);
3217 InFlag = LoadTOCPtr.getValue(1);
3219 MTCTROps[0] = Chain;
3220 MTCTROps[1] = LoadFuncPtr;
3221 MTCTROps[2] = InFlag;
3224 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3225 2 + (InFlag.getNode() != 0));
3226 InFlag = Chain.getValue(1);
3229 NodeTys.push_back(MVT::Other);
3230 NodeTys.push_back(MVT::Glue);
3231 Ops.push_back(Chain);
3232 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3234 // Add CTR register as callee so a bctr can be emitted later.
3236 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3239 // If this is a direct call, pass the chain and the callee.
3240 if (Callee.getNode()) {
3241 Ops.push_back(Chain);
3242 Ops.push_back(Callee);
3244 // If this is a tail call add stack pointer delta.
3246 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3248 // Add argument registers to the end of the list so that they are known live
3250 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3251 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3252 RegsToPass[i].second.getValueType()));
3258 bool isLocalCall(const SDValue &Callee)
3260 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3261 return !G->getGlobal()->isDeclaration() &&
3262 !G->getGlobal()->isWeakForLinker();
3267 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3268 CallingConv::ID CallConv, bool isVarArg,
3269 const SmallVectorImpl<ISD::InputArg> &Ins,
3270 DebugLoc dl, SelectionDAG &DAG,
3271 SmallVectorImpl<SDValue> &InVals) const {
3273 SmallVector<CCValAssign, 16> RVLocs;
3274 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3275 getTargetMachine(), RVLocs, *DAG.getContext());
3276 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3278 // Copy all of the result registers out of their specified physreg.
3279 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3280 CCValAssign &VA = RVLocs[i];
3281 assert(VA.isRegLoc() && "Can only return in registers!");
3283 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3284 VA.getLocReg(), VA.getLocVT(), InFlag);
3285 Chain = Val.getValue(1);
3286 InFlag = Val.getValue(2);
3288 switch (VA.getLocInfo()) {
3289 default: llvm_unreachable("Unknown loc info!");
3290 case CCValAssign::Full: break;
3291 case CCValAssign::AExt:
3292 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3294 case CCValAssign::ZExt:
3295 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3296 DAG.getValueType(VA.getValVT()));
3297 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3299 case CCValAssign::SExt:
3300 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3301 DAG.getValueType(VA.getValVT()));
3302 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3306 InVals.push_back(Val);
3313 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3314 bool isTailCall, bool isVarArg,
3316 SmallVector<std::pair<unsigned, SDValue>, 8>
3318 SDValue InFlag, SDValue Chain,
3320 int SPDiff, unsigned NumBytes,
3321 const SmallVectorImpl<ISD::InputArg> &Ins,
3322 SmallVectorImpl<SDValue> &InVals) const {
3323 std::vector<EVT> NodeTys;
3324 SmallVector<SDValue, 8> Ops;
3325 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3326 isTailCall, RegsToPass, Ops, NodeTys,
3329 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3330 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3331 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3333 // When performing tail call optimization the callee pops its arguments off
3334 // the stack. Account for this here so these bytes can be pushed back on in
3335 // PPCFrameLowering::eliminateCallFramePseudoInstr.
3336 int BytesCalleePops =
3337 (CallConv == CallingConv::Fast &&
3338 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3340 // Add a register mask operand representing the call-preserved registers.
3341 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3342 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3343 assert(Mask && "Missing call preserved mask for calling convention");
3344 Ops.push_back(DAG.getRegisterMask(Mask));
3346 if (InFlag.getNode())
3347 Ops.push_back(InFlag);
3351 assert(((Callee.getOpcode() == ISD::Register &&
3352 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3353 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3354 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3355 isa<ConstantSDNode>(Callee)) &&
3356 "Expecting an global address, external symbol, absolute value or register");
3358 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3361 // Add a NOP immediately after the branch instruction when using the 64-bit
3362 // SVR4 ABI. At link time, if caller and callee are in a different module and
3363 // thus have a different TOC, the call will be replaced with a call to a stub
3364 // function which saves the current TOC, loads the TOC of the callee and
3365 // branches to the callee. The NOP will be replaced with a load instruction
3366 // which restores the TOC of the caller from the TOC save slot of the current
3367 // stack frame. If caller and callee belong to the same module (and have the
3368 // same TOC), the NOP will remain unchanged.
3370 bool needsTOCRestore = false;
3371 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3372 if (CallOpc == PPCISD::BCTRL_SVR4) {
3373 // This is a call through a function pointer.
3374 // Restore the caller TOC from the save area into R2.
3375 // See PrepareCall() for more information about calls through function
3376 // pointers in the 64-bit SVR4 ABI.
3377 // We are using a target-specific load with r2 hard coded, because the
3378 // result of a target-independent load would never go directly into r2,
3379 // since r2 is a reserved register (which prevents the register allocator
3380 // from allocating it), resulting in an additional register being
3381 // allocated and an unnecessary move instruction being generated.
3382 needsTOCRestore = true;
3383 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3384 // Otherwise insert NOP for non-local calls.
3385 CallOpc = PPCISD::CALL_NOP_SVR4;
3389 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3390 InFlag = Chain.getValue(1);
3392 if (needsTOCRestore) {
3393 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3394 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3395 InFlag = Chain.getValue(1);
3398 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3399 DAG.getIntPtrConstant(BytesCalleePops, true),
3402 InFlag = Chain.getValue(1);
3404 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3405 Ins, dl, DAG, InVals);
3409 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3410 SmallVectorImpl<SDValue> &InVals) const {
3411 SelectionDAG &DAG = CLI.DAG;
3412 DebugLoc &dl = CLI.DL;
3413 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3414 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3415 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3416 SDValue Chain = CLI.Chain;
3417 SDValue Callee = CLI.Callee;
3418 bool &isTailCall = CLI.IsTailCall;
3419 CallingConv::ID CallConv = CLI.CallConv;
3420 bool isVarArg = CLI.IsVarArg;
3423 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3426 if (PPCSubTarget.isSVR4ABI()) {
3427 if (PPCSubTarget.isPPC64())
3428 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3429 isTailCall, Outs, OutVals, Ins,
3432 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3433 isTailCall, Outs, OutVals, Ins,
3437 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3438 isTailCall, Outs, OutVals, Ins,
3443 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3444 CallingConv::ID CallConv, bool isVarArg,
3446 const SmallVectorImpl<ISD::OutputArg> &Outs,
3447 const SmallVectorImpl<SDValue> &OutVals,
3448 const SmallVectorImpl<ISD::InputArg> &Ins,
3449 DebugLoc dl, SelectionDAG &DAG,
3450 SmallVectorImpl<SDValue> &InVals) const {
3451 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3452 // of the 32-bit SVR4 ABI stack frame layout.
3454 assert((CallConv == CallingConv::C ||
3455 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3457 unsigned PtrByteSize = 4;
3459 MachineFunction &MF = DAG.getMachineFunction();
3461 // Mark this function as potentially containing a function that contains a
3462 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3463 // and restoring the callers stack pointer in this functions epilog. This is
3464 // done because by tail calling the called function might overwrite the value
3465 // in this function's (MF) stack pointer stack slot 0(SP).
3466 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3467 CallConv == CallingConv::Fast)
3468 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3470 // Count how many bytes are to be pushed on the stack, including the linkage
3471 // area, parameter list area and the part of the local variable space which
3472 // contains copies of aggregates which are passed by value.
3474 // Assign locations to all of the outgoing arguments.
3475 SmallVector<CCValAssign, 16> ArgLocs;
3476 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3477 getTargetMachine(), ArgLocs, *DAG.getContext());
3479 // Reserve space for the linkage area on the stack.
3480 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3483 // Handle fixed and variable vector arguments differently.
3484 // Fixed vector arguments go into registers as long as registers are
3485 // available. Variable vector arguments always go into memory.
3486 unsigned NumArgs = Outs.size();
3488 for (unsigned i = 0; i != NumArgs; ++i) {
3489 MVT ArgVT = Outs[i].VT;
3490 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3493 if (Outs[i].IsFixed) {
3494 Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3497 Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3503 errs() << "Call operand #" << i << " has unhandled type "
3504 << EVT(ArgVT).getEVTString() << "\n";
3506 llvm_unreachable(0);
3510 // All arguments are treated the same.
3511 CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4);
3514 // Assign locations to all of the outgoing aggregate by value arguments.
3515 SmallVector<CCValAssign, 16> ByValArgLocs;
3516 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3517 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3519 // Reserve stack space for the allocations in CCInfo.
3520 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3522 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal);
3524 // Size of the linkage area, parameter list area and the part of the local
3525 // space variable where copies of aggregates which are passed by value are
3527 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3529 // Calculate by how many bytes the stack has to be adjusted in case of tail
3530 // call optimization.
3531 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3533 // Adjust the stack pointer for the new arguments...
3534 // These operations are automatically eliminated by the prolog/epilog pass
3535 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3536 SDValue CallSeqStart = Chain;
3538 // Load the return address and frame pointer so it can be moved somewhere else
3541 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3544 // Set up a copy of the stack pointer for use loading and storing any
3545 // arguments that may not fit in the registers available for argument
3547 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3549 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3550 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3551 SmallVector<SDValue, 8> MemOpChains;
3553 bool seenFloatArg = false;
3554 // Walk the register/memloc assignments, inserting copies/loads.
3555 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3558 CCValAssign &VA = ArgLocs[i];
3559 SDValue Arg = OutVals[i];
3560 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3562 if (Flags.isByVal()) {
3563 // Argument is an aggregate which is passed by value, thus we need to
3564 // create a copy of it in the local variable space of the current stack
3565 // frame (which is the stack frame of the caller) and pass the address of
3566 // this copy to the callee.
3567 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3568 CCValAssign &ByValVA = ByValArgLocs[j++];
3569 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3571 // Memory reserved in the local variable space of the callers stack frame.
3572 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3574 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3575 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3577 // Create a copy of the argument in the local area of the current
3579 SDValue MemcpyCall =
3580 CreateCopyOfByValArgument(Arg, PtrOff,
3581 CallSeqStart.getNode()->getOperand(0),
3584 // This must go outside the CALLSEQ_START..END.
3585 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3586 CallSeqStart.getNode()->getOperand(1));
3587 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3588 NewCallSeqStart.getNode());
3589 Chain = CallSeqStart = NewCallSeqStart;
3591 // Pass the address of the aggregate copy on the stack either in a
3592 // physical register or in the parameter list area of the current stack
3593 // frame to the callee.
3597 if (VA.isRegLoc()) {
3598 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3599 // Put argument in a physical register.
3600 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3602 // Put argument in the parameter list area of the current stack frame.
3603 assert(VA.isMemLoc());
3604 unsigned LocMemOffset = VA.getLocMemOffset();
3607 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3608 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3610 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3611 MachinePointerInfo(),
3614 // Calculate and remember argument location.
3615 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3621 if (!MemOpChains.empty())
3622 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3623 &MemOpChains[0], MemOpChains.size());
3625 // Build a sequence of copy-to-reg nodes chained together with token chain
3626 // and flag operands which copy the outgoing args into the appropriate regs.
3628 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3629 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3630 RegsToPass[i].second, InFlag);
3631 InFlag = Chain.getValue(1);
3634 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3637 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3638 SDValue Ops[] = { Chain, InFlag };
3640 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3641 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3643 InFlag = Chain.getValue(1);
3647 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3648 false, TailCallArguments);
3650 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3651 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3655 // Copy an argument into memory, being careful to do this outside the
3656 // call sequence for the call to which the argument belongs.
3658 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3659 SDValue CallSeqStart,
3660 ISD::ArgFlagsTy Flags,
3662 DebugLoc dl) const {
3663 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3664 CallSeqStart.getNode()->getOperand(0),
3666 // The MEMCPY must go outside the CALLSEQ_START..END.
3667 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3668 CallSeqStart.getNode()->getOperand(1));
3669 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3670 NewCallSeqStart.getNode());
3671 return NewCallSeqStart;
3675 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3676 CallingConv::ID CallConv, bool isVarArg,
3678 const SmallVectorImpl<ISD::OutputArg> &Outs,
3679 const SmallVectorImpl<SDValue> &OutVals,
3680 const SmallVectorImpl<ISD::InputArg> &Ins,
3681 DebugLoc dl, SelectionDAG &DAG,
3682 SmallVectorImpl<SDValue> &InVals) const {
3684 unsigned NumOps = Outs.size();
3686 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3687 unsigned PtrByteSize = 8;
3689 MachineFunction &MF = DAG.getMachineFunction();
3691 // Mark this function as potentially containing a function that contains a
3692 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3693 // and restoring the callers stack pointer in this functions epilog. This is
3694 // done because by tail calling the called function might overwrite the value
3695 // in this function's (MF) stack pointer stack slot 0(SP).
3696 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3697 CallConv == CallingConv::Fast)
3698 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3700 unsigned nAltivecParamsAtEnd = 0;
3702 // Count how many bytes are to be pushed on the stack, including the linkage
3703 // area, and parameter passing area. We start with at least 48 bytes, which
3704 // is reserved space for [SP][CR][LR][3 x unused].
3705 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3708 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3709 Outs, OutVals, nAltivecParamsAtEnd);
3711 // Calculate by how many bytes the stack has to be adjusted in case of tail
3712 // call optimization.
3713 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3715 // To protect arguments on the stack from being clobbered in a tail call,
3716 // force all the loads to happen before doing any other lowering.
3718 Chain = DAG.getStackArgumentTokenFactor(Chain);
3720 // Adjust the stack pointer for the new arguments...
3721 // These operations are automatically eliminated by the prolog/epilog pass
3722 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3723 SDValue CallSeqStart = Chain;
3725 // Load the return address and frame pointer so it can be move somewhere else
3728 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3731 // Set up a copy of the stack pointer for use loading and storing any
3732 // arguments that may not fit in the registers available for argument
3734 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3736 // Figure out which arguments are going to go in registers, and which in
3737 // memory. Also, if this is a vararg function, floating point operations
3738 // must be stored to our stack, and loaded into integer regs as well, if
3739 // any integer regs are available for argument passing.
3740 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3741 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3743 static const uint16_t GPR[] = {
3744 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3745 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3747 static const uint16_t *FPR = GetFPR();
3749 static const uint16_t VR[] = {
3750 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3751 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3753 const unsigned NumGPRs = array_lengthof(GPR);
3754 const unsigned NumFPRs = 13;
3755 const unsigned NumVRs = array_lengthof(VR);
3757 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3758 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3760 SmallVector<SDValue, 8> MemOpChains;
3761 for (unsigned i = 0; i != NumOps; ++i) {
3762 SDValue Arg = OutVals[i];
3763 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3765 // PtrOff will be used to store the current argument to the stack if a
3766 // register cannot be found for it.
3769 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3771 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3773 // Promote integers to 64-bit values.
3774 if (Arg.getValueType() == MVT::i32) {
3775 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3776 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3777 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3780 // FIXME memcpy is used way more than necessary. Correctness first.
3781 // Note: "by value" is code for passing a structure by value, not
3783 if (Flags.isByVal()) {
3784 // Note: Size includes alignment padding, so
3785 // struct x { short a; char b; }
3786 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3787 // These are the proper values we need for right-justifying the
3788 // aggregate in a parameter register.
3789 unsigned Size = Flags.getByValSize();
3791 // An empty aggregate parameter takes up no storage and no
3796 // All aggregates smaller than 8 bytes must be passed right-justified.
3797 if (Size==1 || Size==2 || Size==4) {
3798 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3799 if (GPR_idx != NumGPRs) {
3800 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3801 MachinePointerInfo(), VT,
3803 MemOpChains.push_back(Load.getValue(1));
3804 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3806 ArgOffset += PtrByteSize;
3811 if (GPR_idx == NumGPRs && Size < 8) {
3812 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3813 PtrOff.getValueType());
3814 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3815 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3818 ArgOffset += PtrByteSize;
3821 // Copy entire object into memory. There are cases where gcc-generated
3822 // code assumes it is there, even if it could be put entirely into
3823 // registers. (This is not what the doc says.)
3825 // FIXME: The above statement is likely due to a misunderstanding of the
3826 // documents. All arguments must be copied into the parameter area BY
3827 // THE CALLEE in the event that the callee takes the address of any
3828 // formal argument. That has not yet been implemented. However, it is
3829 // reasonable to use the stack area as a staging area for the register
3832 // Skip this for small aggregates, as we will use the same slot for a
3833 // right-justified copy, below.
3835 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3839 // When a register is available, pass a small aggregate right-justified.
3840 if (Size < 8 && GPR_idx != NumGPRs) {
3841 // The easiest way to get this right-justified in a register
3842 // is to copy the structure into the rightmost portion of a
3843 // local variable slot, then load the whole slot into the
3845 // FIXME: The memcpy seems to produce pretty awful code for
3846 // small aggregates, particularly for packed ones.
3847 // FIXME: It would be preferable to use the slot in the
3848 // parameter save area instead of a new local variable.
3849 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3850 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3851 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3855 // Load the slot into the register.
3856 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3857 MachinePointerInfo(),
3858 false, false, false, 0);
3859 MemOpChains.push_back(Load.getValue(1));
3860 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3862 // Done with this argument.
3863 ArgOffset += PtrByteSize;
3867 // For aggregates larger than PtrByteSize, copy the pieces of the
3868 // object that fit into registers from the parameter save area.
3869 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3870 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3871 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3872 if (GPR_idx != NumGPRs) {
3873 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3874 MachinePointerInfo(),
3875 false, false, false, 0);
3876 MemOpChains.push_back(Load.getValue(1));
3877 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3878 ArgOffset += PtrByteSize;
3880 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3887 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3888 default: llvm_unreachable("Unexpected ValueType for argument!");
3891 if (GPR_idx != NumGPRs) {
3892 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3894 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3895 true, isTailCall, false, MemOpChains,
3896 TailCallArguments, dl);
3898 ArgOffset += PtrByteSize;
3902 if (FPR_idx != NumFPRs) {
3903 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3906 // A single float or an aggregate containing only a single float
3907 // must be passed right-justified in the stack doubleword, and
3908 // in the GPR, if one is available.
3910 if (Arg.getValueType().getSimpleVT().SimpleTy == MVT::f32) {
3911 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3912 StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3916 SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
3917 MachinePointerInfo(), false, false, 0);
3918 MemOpChains.push_back(Store);
3920 // Float varargs are always shadowed in available integer registers
3921 if (GPR_idx != NumGPRs) {
3922 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3923 MachinePointerInfo(), false, false,
3925 MemOpChains.push_back(Load.getValue(1));
3926 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3928 } else if (GPR_idx != NumGPRs)
3929 // If we have any FPRs remaining, we may also have GPRs remaining.
3932 // Single-precision floating-point values are mapped to the
3933 // second (rightmost) word of the stack doubleword.
3934 if (Arg.getValueType() == MVT::f32) {
3935 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3936 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3939 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3940 true, isTailCall, false, MemOpChains,
3941 TailCallArguments, dl);
3950 // These go aligned on the stack, or in the corresponding R registers
3951 // when within range. The Darwin PPC ABI doc claims they also go in
3952 // V registers; in fact gcc does this only for arguments that are
3953 // prototyped, not for those that match the ... We do it for all
3954 // arguments, seems to work.
3955 while (ArgOffset % 16 !=0) {
3956 ArgOffset += PtrByteSize;
3957 if (GPR_idx != NumGPRs)
3960 // We could elide this store in the case where the object fits
3961 // entirely in R registers. Maybe later.
3962 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3963 DAG.getConstant(ArgOffset, PtrVT));
3964 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3965 MachinePointerInfo(), false, false, 0);
3966 MemOpChains.push_back(Store);
3967 if (VR_idx != NumVRs) {
3968 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3969 MachinePointerInfo(),
3970 false, false, false, 0);
3971 MemOpChains.push_back(Load.getValue(1));
3972 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3975 for (unsigned i=0; i<16; i+=PtrByteSize) {
3976 if (GPR_idx == NumGPRs)
3978 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3979 DAG.getConstant(i, PtrVT));
3980 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3981 false, false, false, 0);
3982 MemOpChains.push_back(Load.getValue(1));
3983 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3988 // Non-varargs Altivec params generally go in registers, but have
3989 // stack space allocated at the end.
3990 if (VR_idx != NumVRs) {
3991 // Doesn't have GPR space allocated.
3992 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3994 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3995 true, isTailCall, true, MemOpChains,
3996 TailCallArguments, dl);
4003 if (!MemOpChains.empty())
4004 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4005 &MemOpChains[0], MemOpChains.size());
4007 // Check if this is an indirect call (MTCTR/BCTRL).
4008 // See PrepareCall() for more information about calls through function
4009 // pointers in the 64-bit SVR4 ABI.
4011 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4012 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4013 !isBLACompatibleAddress(Callee, DAG)) {
4014 // Load r2 into a virtual register and store it to the TOC save area.
4015 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
4016 // TOC save area offset.
4017 SDValue PtrOff = DAG.getIntPtrConstant(40);
4018 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4019 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
4021 // R12 must contain the address of an indirect callee. This does not
4022 // mean the MTCTR instruction must use R12; it's easier to model this
4023 // as an extra parameter, so do that.
4024 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
4027 // Build a sequence of copy-to-reg nodes chained together with token chain
4028 // and flag operands which copy the outgoing args into the appropriate regs.
4030 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4031 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4032 RegsToPass[i].second, InFlag);
4033 InFlag = Chain.getValue(1);
4037 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
4038 FPOp, true, TailCallArguments);
4040 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4041 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4046 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
4047 CallingConv::ID CallConv, bool isVarArg,
4049 const SmallVectorImpl<ISD::OutputArg> &Outs,
4050 const SmallVectorImpl<SDValue> &OutVals,
4051 const SmallVectorImpl<ISD::InputArg> &Ins,
4052 DebugLoc dl, SelectionDAG &DAG,
4053 SmallVectorImpl<SDValue> &InVals) const {
4055 unsigned NumOps = Outs.size();
4057 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4058 bool isPPC64 = PtrVT == MVT::i64;
4059 unsigned PtrByteSize = isPPC64 ? 8 : 4;
4061 MachineFunction &MF = DAG.getMachineFunction();
4063 // Mark this function as potentially containing a function that contains a
4064 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4065 // and restoring the callers stack pointer in this functions epilog. This is
4066 // done because by tail calling the called function might overwrite the value
4067 // in this function's (MF) stack pointer stack slot 0(SP).
4068 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4069 CallConv == CallingConv::Fast)
4070 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4072 unsigned nAltivecParamsAtEnd = 0;
4074 // Count how many bytes are to be pushed on the stack, including the linkage
4075 // area, and parameter passing area. We start with 24/48 bytes, which is
4076 // prereserved space for [SP][CR][LR][3 x unused].
4078 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
4080 nAltivecParamsAtEnd);
4082 // Calculate by how many bytes the stack has to be adjusted in case of tail
4083 // call optimization.
4084 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4086 // To protect arguments on the stack from being clobbered in a tail call,
4087 // force all the loads to happen before doing any other lowering.
4089 Chain = DAG.getStackArgumentTokenFactor(Chain);
4091 // Adjust the stack pointer for the new arguments...
4092 // These operations are automatically eliminated by the prolog/epilog pass
4093 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
4094 SDValue CallSeqStart = Chain;
4096 // Load the return address and frame pointer so it can be move somewhere else
4099 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4102 // Set up a copy of the stack pointer for use loading and storing any
4103 // arguments that may not fit in the registers available for argument
4107 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4109 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4111 // Figure out which arguments are going to go in registers, and which in
4112 // memory. Also, if this is a vararg function, floating point operations
4113 // must be stored to our stack, and loaded into integer regs as well, if
4114 // any integer regs are available for argument passing.
4115 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
4116 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4118 static const uint16_t GPR_32[] = { // 32-bit registers.
4119 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4120 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4122 static const uint16_t GPR_64[] = { // 64-bit registers.
4123 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4124 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4126 static const uint16_t *FPR = GetFPR();
4128 static const uint16_t VR[] = {
4129 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4130 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4132 const unsigned NumGPRs = array_lengthof(GPR_32);
4133 const unsigned NumFPRs = 13;
4134 const unsigned NumVRs = array_lengthof(VR);
4136 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
4138 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4139 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4141 SmallVector<SDValue, 8> MemOpChains;
4142 for (unsigned i = 0; i != NumOps; ++i) {
4143 SDValue Arg = OutVals[i];
4144 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4146 // PtrOff will be used to store the current argument to the stack if a
4147 // register cannot be found for it.
4150 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4152 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4154 // On PPC64, promote integers to 64-bit values.
4155 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4156 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4157 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4158 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4161 // FIXME memcpy is used way more than necessary. Correctness first.
4162 // Note: "by value" is code for passing a structure by value, not
4164 if (Flags.isByVal()) {
4165 unsigned Size = Flags.getByValSize();
4166 // Very small objects are passed right-justified. Everything else is
4167 // passed left-justified.
4168 if (Size==1 || Size==2) {
4169 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4170 if (GPR_idx != NumGPRs) {
4171 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4172 MachinePointerInfo(), VT,
4174 MemOpChains.push_back(Load.getValue(1));
4175 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4177 ArgOffset += PtrByteSize;
4179 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4180 PtrOff.getValueType());
4181 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4182 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4185 ArgOffset += PtrByteSize;
4189 // Copy entire object into memory. There are cases where gcc-generated
4190 // code assumes it is there, even if it could be put entirely into
4191 // registers. (This is not what the doc says.)
4192 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4196 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4197 // copy the pieces of the object that fit into registers from the
4198 // parameter save area.
4199 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4200 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4201 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4202 if (GPR_idx != NumGPRs) {
4203 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4204 MachinePointerInfo(),
4205 false, false, false, 0);
4206 MemOpChains.push_back(Load.getValue(1));
4207 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4208 ArgOffset += PtrByteSize;
4210 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4217 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4218 default: llvm_unreachable("Unexpected ValueType for argument!");
4221 if (GPR_idx != NumGPRs) {
4222 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4224 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4225 isPPC64, isTailCall, false, MemOpChains,
4226 TailCallArguments, dl);
4228 ArgOffset += PtrByteSize;
4232 if (FPR_idx != NumFPRs) {
4233 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4236 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4237 MachinePointerInfo(), false, false, 0);
4238 MemOpChains.push_back(Store);
4240 // Float varargs are always shadowed in available integer registers
4241 if (GPR_idx != NumGPRs) {
4242 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4243 MachinePointerInfo(), false, false,
4245 MemOpChains.push_back(Load.getValue(1));
4246 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4248 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4249 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4250 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4251 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4252 MachinePointerInfo(),
4253 false, false, false, 0);
4254 MemOpChains.push_back(Load.getValue(1));
4255 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4258 // If we have any FPRs remaining, we may also have GPRs remaining.
4259 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4261 if (GPR_idx != NumGPRs)
4263 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4264 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4268 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4269 isPPC64, isTailCall, false, MemOpChains,
4270 TailCallArguments, dl);
4274 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4281 // These go aligned on the stack, or in the corresponding R registers
4282 // when within range. The Darwin PPC ABI doc claims they also go in
4283 // V registers; in fact gcc does this only for arguments that are
4284 // prototyped, not for those that match the ... We do it for all
4285 // arguments, seems to work.
4286 while (ArgOffset % 16 !=0) {
4287 ArgOffset += PtrByteSize;
4288 if (GPR_idx != NumGPRs)
4291 // We could elide this store in the case where the object fits
4292 // entirely in R registers. Maybe later.
4293 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4294 DAG.getConstant(ArgOffset, PtrVT));
4295 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4296 MachinePointerInfo(), false, false, 0);
4297 MemOpChains.push_back(Store);
4298 if (VR_idx != NumVRs) {
4299 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4300 MachinePointerInfo(),
4301 false, false, false, 0);
4302 MemOpChains.push_back(Load.getValue(1));
4303 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4306 for (unsigned i=0; i<16; i+=PtrByteSize) {
4307 if (GPR_idx == NumGPRs)
4309 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4310 DAG.getConstant(i, PtrVT));
4311 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4312 false, false, false, 0);
4313 MemOpChains.push_back(Load.getValue(1));
4314 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4319 // Non-varargs Altivec params generally go in registers, but have
4320 // stack space allocated at the end.
4321 if (VR_idx != NumVRs) {
4322 // Doesn't have GPR space allocated.
4323 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4324 } else if (nAltivecParamsAtEnd==0) {
4325 // We are emitting Altivec params in order.
4326 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4327 isPPC64, isTailCall, true, MemOpChains,
4328 TailCallArguments, dl);
4334 // If all Altivec parameters fit in registers, as they usually do,
4335 // they get stack space following the non-Altivec parameters. We
4336 // don't track this here because nobody below needs it.
4337 // If there are more Altivec parameters than fit in registers emit
4339 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4341 // Offset is aligned; skip 1st 12 params which go in V registers.
4342 ArgOffset = ((ArgOffset+15)/16)*16;
4344 for (unsigned i = 0; i != NumOps; ++i) {
4345 SDValue Arg = OutVals[i];
4346 EVT ArgType = Outs[i].VT;
4347 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4348 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4351 // We are emitting Altivec params in order.
4352 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4353 isPPC64, isTailCall, true, MemOpChains,
4354 TailCallArguments, dl);
4361 if (!MemOpChains.empty())
4362 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4363 &MemOpChains[0], MemOpChains.size());
4365 // On Darwin, R12 must contain the address of an indirect callee. This does
4366 // not mean the MTCTR instruction must use R12; it's easier to model this as
4367 // an extra parameter, so do that.
4369 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4370 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4371 !isBLACompatibleAddress(Callee, DAG))
4372 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4373 PPC::R12), Callee));
4375 // Build a sequence of copy-to-reg nodes chained together with token chain
4376 // and flag operands which copy the outgoing args into the appropriate regs.
4378 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4379 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4380 RegsToPass[i].second, InFlag);
4381 InFlag = Chain.getValue(1);
4385 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4386 FPOp, true, TailCallArguments);
4388 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4389 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4394 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4395 MachineFunction &MF, bool isVarArg,
4396 const SmallVectorImpl<ISD::OutputArg> &Outs,
4397 LLVMContext &Context) const {
4398 SmallVector<CCValAssign, 16> RVLocs;
4399 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4401 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4405 PPCTargetLowering::LowerReturn(SDValue Chain,
4406 CallingConv::ID CallConv, bool isVarArg,
4407 const SmallVectorImpl<ISD::OutputArg> &Outs,
4408 const SmallVectorImpl<SDValue> &OutVals,
4409 DebugLoc dl, SelectionDAG &DAG) const {
4411 SmallVector<CCValAssign, 16> RVLocs;
4412 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4413 getTargetMachine(), RVLocs, *DAG.getContext());
4414 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4417 SmallVector<SDValue, 4> RetOps(1, Chain);
4419 // Copy the result values into the output registers.
4420 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4421 CCValAssign &VA = RVLocs[i];
4422 assert(VA.isRegLoc() && "Can only return in registers!");
4424 SDValue Arg = OutVals[i];
4426 switch (VA.getLocInfo()) {
4427 default: llvm_unreachable("Unknown loc info!");
4428 case CCValAssign::Full: break;
4429 case CCValAssign::AExt:
4430 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
4432 case CCValAssign::ZExt:
4433 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
4435 case CCValAssign::SExt:
4436 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
4440 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
4441 Flag = Chain.getValue(1);
4442 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
4445 RetOps[0] = Chain; // Update chain.
4447 // Add the flag if we have it.
4449 RetOps.push_back(Flag);
4451 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other,
4452 &RetOps[0], RetOps.size());
4455 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4456 const PPCSubtarget &Subtarget) const {
4457 // When we pop the dynamic allocation we need to restore the SP link.
4458 DebugLoc dl = Op.getDebugLoc();
4460 // Get the corect type for pointers.
4461 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4463 // Construct the stack pointer operand.
4464 bool isPPC64 = Subtarget.isPPC64();
4465 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4466 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4468 // Get the operands for the STACKRESTORE.
4469 SDValue Chain = Op.getOperand(0);
4470 SDValue SaveSP = Op.getOperand(1);
4472 // Load the old link SP.
4473 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4474 MachinePointerInfo(),
4475 false, false, false, 0);
4477 // Restore the stack pointer.
4478 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4480 // Store the old link SP.
4481 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4488 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4489 MachineFunction &MF = DAG.getMachineFunction();
4490 bool isPPC64 = PPCSubTarget.isPPC64();
4491 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4492 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4494 // Get current frame pointer save index. The users of this index will be
4495 // primarily DYNALLOC instructions.
4496 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4497 int RASI = FI->getReturnAddrSaveIndex();
4499 // If the frame pointer save index hasn't been defined yet.
4501 // Find out what the fix offset of the frame pointer save area.
4502 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4503 // Allocate the frame index for frame pointer save area.
4504 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4506 FI->setReturnAddrSaveIndex(RASI);
4508 return DAG.getFrameIndex(RASI, PtrVT);
4512 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4513 MachineFunction &MF = DAG.getMachineFunction();
4514 bool isPPC64 = PPCSubTarget.isPPC64();
4515 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4516 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4518 // Get current frame pointer save index. The users of this index will be
4519 // primarily DYNALLOC instructions.
4520 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4521 int FPSI = FI->getFramePointerSaveIndex();
4523 // If the frame pointer save index hasn't been defined yet.
4525 // Find out what the fix offset of the frame pointer save area.
4526 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4529 // Allocate the frame index for frame pointer save area.
4530 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4532 FI->setFramePointerSaveIndex(FPSI);
4534 return DAG.getFrameIndex(FPSI, PtrVT);
4537 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4539 const PPCSubtarget &Subtarget) const {
4541 SDValue Chain = Op.getOperand(0);
4542 SDValue Size = Op.getOperand(1);
4543 DebugLoc dl = Op.getDebugLoc();
4545 // Get the corect type for pointers.
4546 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4548 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4549 DAG.getConstant(0, PtrVT), Size);
4550 // Construct a node for the frame pointer save index.
4551 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4552 // Build a DYNALLOC node.
4553 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4554 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4555 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4558 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4560 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4561 // Not FP? Not a fsel.
4562 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4563 !Op.getOperand(2).getValueType().isFloatingPoint())
4566 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4568 // Cannot handle SETEQ/SETNE.
4569 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4571 EVT ResVT = Op.getValueType();
4572 EVT CmpVT = Op.getOperand(0).getValueType();
4573 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4574 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4575 DebugLoc dl = Op.getDebugLoc();
4577 // If the RHS of the comparison is a 0.0, we don't need to do the
4578 // subtraction at all.
4579 if (isFloatingPointZero(RHS))
4581 default: break; // SETUO etc aren't handled by fsel.
4584 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4587 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4588 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4589 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4592 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4595 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4596 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4597 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4598 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4603 default: break; // SETUO etc aren't handled by fsel.
4606 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4607 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4608 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4609 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4612 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4613 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4614 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4615 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4618 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4619 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4620 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4621 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4624 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4625 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4626 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4627 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4632 // FIXME: Split this code up when LegalizeDAGTypes lands.
4633 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4634 DebugLoc dl) const {
4635 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4636 SDValue Src = Op.getOperand(0);
4637 if (Src.getValueType() == MVT::f32)
4638 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4641 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4642 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4644 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4649 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4653 // Convert the FP value to an int value through memory.
4654 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4656 // Emit a store to the stack slot.
4657 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4658 MachinePointerInfo(), false, false, 0);
4660 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4662 if (Op.getValueType() == MVT::i32)
4663 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4664 DAG.getConstant(4, FIPtr.getValueType()));
4665 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4666 false, false, false, 0);
4669 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4670 SelectionDAG &DAG) const {
4671 DebugLoc dl = Op.getDebugLoc();
4672 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4673 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4676 if (Op.getOperand(0).getValueType() == MVT::i64) {
4677 SDValue SINT = Op.getOperand(0);
4678 // When converting to single-precision, we actually need to convert
4679 // to double-precision first and then round to single-precision.
4680 // To avoid double-rounding effects during that operation, we have
4681 // to prepare the input operand. Bits that might be truncated when
4682 // converting to double-precision are replaced by a bit that won't
4683 // be lost at this stage, but is below the single-precision rounding
4686 // However, if -enable-unsafe-fp-math is in effect, accept double
4687 // rounding to avoid the extra overhead.
4688 if (Op.getValueType() == MVT::f32 &&
4689 !DAG.getTarget().Options.UnsafeFPMath) {
4691 // Twiddle input to make sure the low 11 bits are zero. (If this
4692 // is the case, we are guaranteed the value will fit into the 53 bit
4693 // mantissa of an IEEE double-precision value without rounding.)
4694 // If any of those low 11 bits were not zero originally, make sure
4695 // bit 12 (value 2048) is set instead, so that the final rounding
4696 // to single-precision gets the correct result.
4697 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4698 SINT, DAG.getConstant(2047, MVT::i64));
4699 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4700 Round, DAG.getConstant(2047, MVT::i64));
4701 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4702 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4703 Round, DAG.getConstant(-2048, MVT::i64));
4705 // However, we cannot use that value unconditionally: if the magnitude
4706 // of the input value is small, the bit-twiddling we did above might
4707 // end up visibly changing the output. Fortunately, in that case, we
4708 // don't need to twiddle bits since the original input will convert
4709 // exactly to double-precision floating-point already. Therefore,
4710 // construct a conditional to use the original value if the top 11
4711 // bits are all sign-bit copies, and use the rounded value computed
4713 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4714 SINT, DAG.getConstant(53, MVT::i32));
4715 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4716 Cond, DAG.getConstant(1, MVT::i64));
4717 Cond = DAG.getSetCC(dl, MVT::i32,
4718 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4720 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4722 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4723 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4724 if (Op.getValueType() == MVT::f32)
4725 FP = DAG.getNode(ISD::FP_ROUND, dl,
4726 MVT::f32, FP, DAG.getIntPtrConstant(0));
4730 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4731 "Unhandled SINT_TO_FP type in custom expander!");
4732 // Since we only generate this in 64-bit mode, we can take advantage of
4733 // 64-bit registers. In particular, sign extend the input value into the
4734 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4735 // then lfd it and fcfid it.
4736 MachineFunction &MF = DAG.getMachineFunction();
4737 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4738 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4739 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4740 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4742 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4745 // STD the extended value into the stack slot.
4746 MachineMemOperand *MMO =
4747 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4748 MachineMemOperand::MOStore, 8, 8);
4749 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4751 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4752 Ops, 4, MVT::i64, MMO);
4753 // Load the value as a double.
4754 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4755 false, false, false, 0);
4757 // FCFID it and return it.
4758 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4759 if (Op.getValueType() == MVT::f32)
4760 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4764 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4765 SelectionDAG &DAG) const {
4766 DebugLoc dl = Op.getDebugLoc();
4768 The rounding mode is in bits 30:31 of FPSR, and has the following
4775 FLT_ROUNDS, on the other hand, expects the following:
4782 To perform the conversion, we do:
4783 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4786 MachineFunction &MF = DAG.getMachineFunction();
4787 EVT VT = Op.getValueType();
4788 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4789 SDValue MFFSreg, InFlag;
4791 // Save FP Control Word to register
4793 MVT::f64, // return register
4794 MVT::Glue // unused in this context
4796 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4798 // Save FP register to stack slot
4799 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4800 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4801 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4802 StackSlot, MachinePointerInfo(), false, false,0);
4804 // Load FP Control Word from low 32 bits of stack slot.
4805 SDValue Four = DAG.getConstant(4, PtrVT);
4806 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4807 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4808 false, false, false, 0);
4810 // Transform as necessary
4812 DAG.getNode(ISD::AND, dl, MVT::i32,
4813 CWD, DAG.getConstant(3, MVT::i32));
4815 DAG.getNode(ISD::SRL, dl, MVT::i32,
4816 DAG.getNode(ISD::AND, dl, MVT::i32,
4817 DAG.getNode(ISD::XOR, dl, MVT::i32,
4818 CWD, DAG.getConstant(3, MVT::i32)),
4819 DAG.getConstant(3, MVT::i32)),
4820 DAG.getConstant(1, MVT::i32));
4823 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4825 return DAG.getNode((VT.getSizeInBits() < 16 ?
4826 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4829 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4830 EVT VT = Op.getValueType();
4831 unsigned BitWidth = VT.getSizeInBits();
4832 DebugLoc dl = Op.getDebugLoc();
4833 assert(Op.getNumOperands() == 3 &&
4834 VT == Op.getOperand(1).getValueType() &&
4837 // Expand into a bunch of logical ops. Note that these ops
4838 // depend on the PPC behavior for oversized shift amounts.
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::SHL, dl, VT, Hi, Amt);
4847 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, 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::SHL, dl, VT, Lo, Tmp5);
4852 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4853 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4854 SDValue OutOps[] = { OutLo, OutHi };
4855 return DAG.getMergeValues(OutOps, 2, dl);
4858 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4859 EVT VT = Op.getValueType();
4860 DebugLoc dl = Op.getDebugLoc();
4861 unsigned BitWidth = VT.getSizeInBits();
4862 assert(Op.getNumOperands() == 3 &&
4863 VT == Op.getOperand(1).getValueType() &&
4866 // Expand into a bunch of logical ops. Note that these ops
4867 // depend on the PPC behavior for oversized shift amounts.
4868 SDValue Lo = Op.getOperand(0);
4869 SDValue Hi = Op.getOperand(1);
4870 SDValue Amt = Op.getOperand(2);
4871 EVT AmtVT = Amt.getValueType();
4873 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4874 DAG.getConstant(BitWidth, AmtVT), Amt);
4875 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4876 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4877 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4878 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4879 DAG.getConstant(-BitWidth, AmtVT));
4880 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4881 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4882 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4883 SDValue OutOps[] = { OutLo, OutHi };
4884 return DAG.getMergeValues(OutOps, 2, dl);
4887 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4888 DebugLoc dl = Op.getDebugLoc();
4889 EVT VT = Op.getValueType();
4890 unsigned BitWidth = VT.getSizeInBits();
4891 assert(Op.getNumOperands() == 3 &&
4892 VT == Op.getOperand(1).getValueType() &&
4895 // Expand into a bunch of logical ops, followed by a select_cc.
4896 SDValue Lo = Op.getOperand(0);
4897 SDValue Hi = Op.getOperand(1);
4898 SDValue Amt = Op.getOperand(2);
4899 EVT AmtVT = Amt.getValueType();
4901 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4902 DAG.getConstant(BitWidth, AmtVT), Amt);
4903 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4904 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4905 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4906 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4907 DAG.getConstant(-BitWidth, AmtVT));
4908 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4909 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4910 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4911 Tmp4, Tmp6, ISD::SETLE);
4912 SDValue OutOps[] = { OutLo, OutHi };
4913 return DAG.getMergeValues(OutOps, 2, dl);
4916 //===----------------------------------------------------------------------===//
4917 // Vector related lowering.
4920 /// BuildSplatI - Build a canonical splati of Val with an element size of
4921 /// SplatSize. Cast the result to VT.
4922 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4923 SelectionDAG &DAG, DebugLoc dl) {
4924 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4926 static const EVT VTys[] = { // canonical VT to use for each size.
4927 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4930 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4932 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4936 EVT CanonicalVT = VTys[SplatSize-1];
4938 // Build a canonical splat for this value.
4939 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4940 SmallVector<SDValue, 8> Ops;
4941 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4942 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4943 &Ops[0], Ops.size());
4944 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4947 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4948 /// specified intrinsic ID.
4949 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4950 SelectionDAG &DAG, DebugLoc dl,
4951 EVT DestVT = MVT::Other) {
4952 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4953 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4954 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4957 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4958 /// specified intrinsic ID.
4959 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4960 SDValue Op2, SelectionDAG &DAG,
4961 DebugLoc dl, EVT DestVT = MVT::Other) {
4962 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4963 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4964 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4968 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4969 /// amount. The result has the specified value type.
4970 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4971 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4972 // Force LHS/RHS to be the right type.
4973 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4974 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4977 for (unsigned i = 0; i != 16; ++i)
4979 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4980 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4983 // If this is a case we can't handle, return null and let the default
4984 // expansion code take care of it. If we CAN select this case, and if it
4985 // selects to a single instruction, return Op. Otherwise, if we can codegen
4986 // this case more efficiently than a constant pool load, lower it to the
4987 // sequence of ops that should be used.
4988 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4989 SelectionDAG &DAG) const {
4990 DebugLoc dl = Op.getDebugLoc();
4991 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4992 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4994 // Check if this is a splat of a constant value.
4995 APInt APSplatBits, APSplatUndef;
4996 unsigned SplatBitSize;
4998 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
4999 HasAnyUndefs, 0, true) || SplatBitSize > 32)
5002 unsigned SplatBits = APSplatBits.getZExtValue();
5003 unsigned SplatUndef = APSplatUndef.getZExtValue();
5004 unsigned SplatSize = SplatBitSize / 8;
5006 // First, handle single instruction cases.
5009 if (SplatBits == 0) {
5010 // Canonicalize all zero vectors to be v4i32.
5011 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
5012 SDValue Z = DAG.getConstant(0, MVT::i32);
5013 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
5014 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
5019 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
5020 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
5022 if (SextVal >= -16 && SextVal <= 15)
5023 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
5026 // Two instruction sequences.
5028 // If this value is in the range [-32,30] and is even, use:
5029 // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2)
5030 // If this value is in the range [17,31] and is odd, use:
5031 // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16)
5032 // If this value is in the range [-31,-17] and is odd, use:
5033 // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16)
5034 // Note the last two are three-instruction sequences.
5035 if (SextVal >= -32 && SextVal <= 31) {
5036 // To avoid having these optimizations undone by constant folding,
5037 // we convert to a pseudo that will be expanded later into one of
5039 SDValue Elt = DAG.getConstant(SextVal, MVT::i32);
5040 EVT VT = Op.getValueType();
5041 int Size = VT == MVT::v16i8 ? 1 : (VT == MVT::v8i16 ? 2 : 4);
5042 SDValue EltSize = DAG.getConstant(Size, MVT::i32);
5043 return DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
5046 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
5047 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
5049 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
5050 // Make -1 and vspltisw -1:
5051 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
5053 // Make the VSLW intrinsic, computing 0x8000_0000.
5054 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
5057 // xor by OnesV to invert it.
5058 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
5059 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5062 // Check to see if this is a wide variety of vsplti*, binop self cases.
5063 static const signed char SplatCsts[] = {
5064 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
5065 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
5068 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
5069 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
5070 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
5071 int i = SplatCsts[idx];
5073 // Figure out what shift amount will be used by altivec if shifted by i in
5075 unsigned TypeShiftAmt = i & (SplatBitSize-1);
5077 // vsplti + shl self.
5078 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
5079 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5080 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5081 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
5082 Intrinsic::ppc_altivec_vslw
5084 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5085 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5088 // vsplti + srl self.
5089 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5090 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5091 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5092 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
5093 Intrinsic::ppc_altivec_vsrw
5095 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5096 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5099 // vsplti + sra self.
5100 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5101 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5102 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5103 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
5104 Intrinsic::ppc_altivec_vsraw
5106 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5107 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5110 // vsplti + rol self.
5111 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
5112 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
5113 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5114 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5115 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
5116 Intrinsic::ppc_altivec_vrlw
5118 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5119 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5122 // t = vsplti c, result = vsldoi t, t, 1
5123 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
5124 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5125 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
5127 // t = vsplti c, result = vsldoi t, t, 2
5128 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
5129 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5130 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
5132 // t = vsplti c, result = vsldoi t, t, 3
5133 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
5134 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5135 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
5142 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5143 /// the specified operations to build the shuffle.
5144 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5145 SDValue RHS, SelectionDAG &DAG,
5147 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5148 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5149 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5152 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5164 if (OpNum == OP_COPY) {
5165 if (LHSID == (1*9+2)*9+3) return LHS;
5166 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5170 SDValue OpLHS, OpRHS;
5171 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5172 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5176 default: llvm_unreachable("Unknown i32 permute!");
5178 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5179 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5180 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5181 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5184 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5185 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5186 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5187 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5190 for (unsigned i = 0; i != 16; ++i)
5191 ShufIdxs[i] = (i&3)+0;
5194 for (unsigned i = 0; i != 16; ++i)
5195 ShufIdxs[i] = (i&3)+4;
5198 for (unsigned i = 0; i != 16; ++i)
5199 ShufIdxs[i] = (i&3)+8;
5202 for (unsigned i = 0; i != 16; ++i)
5203 ShufIdxs[i] = (i&3)+12;
5206 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5208 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5210 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5212 EVT VT = OpLHS.getValueType();
5213 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5214 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5215 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5216 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5219 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5220 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5221 /// return the code it can be lowered into. Worst case, it can always be
5222 /// lowered into a vperm.
5223 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5224 SelectionDAG &DAG) const {
5225 DebugLoc dl = Op.getDebugLoc();
5226 SDValue V1 = Op.getOperand(0);
5227 SDValue V2 = Op.getOperand(1);
5228 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5229 EVT VT = Op.getValueType();
5231 // Cases that are handled by instructions that take permute immediates
5232 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5233 // selected by the instruction selector.
5234 if (V2.getOpcode() == ISD::UNDEF) {
5235 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5236 PPC::isSplatShuffleMask(SVOp, 2) ||
5237 PPC::isSplatShuffleMask(SVOp, 4) ||
5238 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5239 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5240 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5241 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5242 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5243 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5244 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5245 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5246 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5251 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5252 // and produce a fixed permutation. If any of these match, do not lower to
5254 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5255 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5256 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5257 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5258 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5259 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5260 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5261 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5262 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5265 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5266 // perfect shuffle table to emit an optimal matching sequence.
5267 ArrayRef<int> PermMask = SVOp->getMask();
5269 unsigned PFIndexes[4];
5270 bool isFourElementShuffle = true;
5271 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5272 unsigned EltNo = 8; // Start out undef.
5273 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5274 if (PermMask[i*4+j] < 0)
5275 continue; // Undef, ignore it.
5277 unsigned ByteSource = PermMask[i*4+j];
5278 if ((ByteSource & 3) != j) {
5279 isFourElementShuffle = false;
5284 EltNo = ByteSource/4;
5285 } else if (EltNo != ByteSource/4) {
5286 isFourElementShuffle = false;
5290 PFIndexes[i] = EltNo;
5293 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5294 // perfect shuffle vector to determine if it is cost effective to do this as
5295 // discrete instructions, or whether we should use a vperm.
5296 if (isFourElementShuffle) {
5297 // Compute the index in the perfect shuffle table.
5298 unsigned PFTableIndex =
5299 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5301 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5302 unsigned Cost = (PFEntry >> 30);
5304 // Determining when to avoid vperm is tricky. Many things affect the cost
5305 // of vperm, particularly how many times the perm mask needs to be computed.
5306 // For example, if the perm mask can be hoisted out of a loop or is already
5307 // used (perhaps because there are multiple permutes with the same shuffle
5308 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5309 // the loop requires an extra register.
5311 // As a compromise, we only emit discrete instructions if the shuffle can be
5312 // generated in 3 or fewer operations. When we have loop information
5313 // available, if this block is within a loop, we should avoid using vperm
5314 // for 3-operation perms and use a constant pool load instead.
5316 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5319 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5320 // vector that will get spilled to the constant pool.
5321 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5323 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5324 // that it is in input element units, not in bytes. Convert now.
5325 EVT EltVT = V1.getValueType().getVectorElementType();
5326 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5328 SmallVector<SDValue, 16> ResultMask;
5329 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5330 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5332 for (unsigned j = 0; j != BytesPerElement; ++j)
5333 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5337 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5338 &ResultMask[0], ResultMask.size());
5339 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5342 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5343 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5344 /// information about the intrinsic.
5345 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5347 unsigned IntrinsicID =
5348 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5351 switch (IntrinsicID) {
5352 default: return false;
5353 // Comparison predicates.
5354 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5355 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5356 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5357 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5358 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5359 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5360 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5361 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5362 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5363 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5364 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5365 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5366 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5368 // Normal Comparisons.
5369 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5370 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5371 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5372 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5373 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5374 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5375 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5376 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5377 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5378 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5379 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5380 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5381 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5386 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5387 /// lower, do it, otherwise return null.
5388 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5389 SelectionDAG &DAG) const {
5390 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5391 // opcode number of the comparison.
5392 DebugLoc dl = Op.getDebugLoc();
5395 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5396 return SDValue(); // Don't custom lower most intrinsics.
5398 // If this is a non-dot comparison, make the VCMP node and we are done.
5400 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5401 Op.getOperand(1), Op.getOperand(2),
5402 DAG.getConstant(CompareOpc, MVT::i32));
5403 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5406 // Create the PPCISD altivec 'dot' comparison node.
5408 Op.getOperand(2), // LHS
5409 Op.getOperand(3), // RHS
5410 DAG.getConstant(CompareOpc, MVT::i32)
5412 EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue };
5413 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5415 // Now that we have the comparison, emit a copy from the CR to a GPR.
5416 // This is flagged to the above dot comparison.
5417 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5418 DAG.getRegister(PPC::CR6, MVT::i32),
5419 CompNode.getValue(1));
5421 // Unpack the result based on how the target uses it.
5422 unsigned BitNo; // Bit # of CR6.
5423 bool InvertBit; // Invert result?
5424 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5425 default: // Can't happen, don't crash on invalid number though.
5426 case 0: // Return the value of the EQ bit of CR6.
5427 BitNo = 0; InvertBit = false;
5429 case 1: // Return the inverted value of the EQ bit of CR6.
5430 BitNo = 0; InvertBit = true;
5432 case 2: // Return the value of the LT bit of CR6.
5433 BitNo = 2; InvertBit = false;
5435 case 3: // Return the inverted value of the LT bit of CR6.
5436 BitNo = 2; InvertBit = true;
5440 // Shift the bit into the low position.
5441 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5442 DAG.getConstant(8-(3-BitNo), MVT::i32));
5444 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5445 DAG.getConstant(1, MVT::i32));
5447 // If we are supposed to, toggle the bit.
5449 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5450 DAG.getConstant(1, MVT::i32));
5454 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5455 SelectionDAG &DAG) const {
5456 DebugLoc dl = Op.getDebugLoc();
5457 // Create a stack slot that is 16-byte aligned.
5458 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5459 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5460 EVT PtrVT = getPointerTy();
5461 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5463 // Store the input value into Value#0 of the stack slot.
5464 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5465 Op.getOperand(0), FIdx, MachinePointerInfo(),
5468 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5469 false, false, false, 0);
5472 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5473 DebugLoc dl = Op.getDebugLoc();
5474 if (Op.getValueType() == MVT::v4i32) {
5475 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5477 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5478 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5480 SDValue RHSSwap = // = vrlw RHS, 16
5481 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5483 // Shrinkify inputs to v8i16.
5484 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5485 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5486 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5488 // Low parts multiplied together, generating 32-bit results (we ignore the
5490 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5491 LHS, RHS, DAG, dl, MVT::v4i32);
5493 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5494 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5495 // Shift the high parts up 16 bits.
5496 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5498 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5499 } else if (Op.getValueType() == MVT::v8i16) {
5500 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5502 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5504 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5505 LHS, RHS, Zero, DAG, dl);
5506 } else if (Op.getValueType() == MVT::v16i8) {
5507 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5509 // Multiply the even 8-bit parts, producing 16-bit sums.
5510 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5511 LHS, RHS, DAG, dl, MVT::v8i16);
5512 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5514 // Multiply the odd 8-bit parts, producing 16-bit sums.
5515 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5516 LHS, RHS, DAG, dl, MVT::v8i16);
5517 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5519 // Merge the results together.
5521 for (unsigned i = 0; i != 8; ++i) {
5523 Ops[i*2+1] = 2*i+1+16;
5525 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5527 llvm_unreachable("Unknown mul to lower!");
5531 /// LowerOperation - Provide custom lowering hooks for some operations.
5533 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5534 switch (Op.getOpcode()) {
5535 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5536 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5537 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5538 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5539 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5540 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5541 case ISD::SETCC: return LowerSETCC(Op, DAG);
5542 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5543 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5545 return LowerVASTART(Op, DAG, PPCSubTarget);
5548 return LowerVAARG(Op, DAG, PPCSubTarget);
5550 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5551 case ISD::DYNAMIC_STACKALLOC:
5552 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5554 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5555 case ISD::FP_TO_UINT:
5556 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5558 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5559 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5561 // Lower 64-bit shifts.
5562 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5563 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5564 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5566 // Vector-related lowering.
5567 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5568 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5569 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5570 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5571 case ISD::MUL: return LowerMUL(Op, DAG);
5573 // Frame & Return address.
5574 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5575 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5579 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5580 SmallVectorImpl<SDValue>&Results,
5581 SelectionDAG &DAG) const {
5582 const TargetMachine &TM = getTargetMachine();
5583 DebugLoc dl = N->getDebugLoc();
5584 switch (N->getOpcode()) {
5586 llvm_unreachable("Do not know how to custom type legalize this operation!");
5588 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5589 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5592 EVT VT = N->getValueType(0);
5594 if (VT == MVT::i64) {
5595 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5597 Results.push_back(NewNode);
5598 Results.push_back(NewNode.getValue(1));
5602 case ISD::FP_ROUND_INREG: {
5603 assert(N->getValueType(0) == MVT::ppcf128);
5604 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5605 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5606 MVT::f64, N->getOperand(0),
5607 DAG.getIntPtrConstant(0));
5608 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5609 MVT::f64, N->getOperand(0),
5610 DAG.getIntPtrConstant(1));
5612 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5613 // of the long double, and puts FPSCR back the way it was. We do not
5614 // actually model FPSCR.
5615 std::vector<EVT> NodeTys;
5616 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5618 NodeTys.push_back(MVT::f64); // Return register
5619 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5620 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5621 MFFSreg = Result.getValue(0);
5622 InFlag = Result.getValue(1);
5625 NodeTys.push_back(MVT::Glue); // Returns a flag
5626 Ops[0] = DAG.getConstant(31, MVT::i32);
5628 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5629 InFlag = Result.getValue(0);
5632 NodeTys.push_back(MVT::Glue); // Returns a flag
5633 Ops[0] = DAG.getConstant(30, MVT::i32);
5635 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5636 InFlag = Result.getValue(0);
5639 NodeTys.push_back(MVT::f64); // result of add
5640 NodeTys.push_back(MVT::Glue); // Returns a flag
5644 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5645 FPreg = Result.getValue(0);
5646 InFlag = Result.getValue(1);
5649 NodeTys.push_back(MVT::f64);
5650 Ops[0] = DAG.getConstant(1, MVT::i32);
5654 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5655 FPreg = Result.getValue(0);
5657 // We know the low half is about to be thrown away, so just use something
5659 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5663 case ISD::FP_TO_SINT:
5664 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5670 //===----------------------------------------------------------------------===//
5671 // Other Lowering Code
5672 //===----------------------------------------------------------------------===//
5675 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5676 bool is64bit, unsigned BinOpcode) const {
5677 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5678 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5680 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5681 MachineFunction *F = BB->getParent();
5682 MachineFunction::iterator It = BB;
5685 unsigned dest = MI->getOperand(0).getReg();
5686 unsigned ptrA = MI->getOperand(1).getReg();
5687 unsigned ptrB = MI->getOperand(2).getReg();
5688 unsigned incr = MI->getOperand(3).getReg();
5689 DebugLoc dl = MI->getDebugLoc();
5691 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5692 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5693 F->insert(It, loopMBB);
5694 F->insert(It, exitMBB);
5695 exitMBB->splice(exitMBB->begin(), BB,
5696 llvm::next(MachineBasicBlock::iterator(MI)),
5698 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5700 MachineRegisterInfo &RegInfo = F->getRegInfo();
5701 unsigned TmpReg = (!BinOpcode) ? incr :
5702 RegInfo.createVirtualRegister(
5703 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5704 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5708 // fallthrough --> loopMBB
5709 BB->addSuccessor(loopMBB);
5712 // l[wd]arx dest, ptr
5713 // add r0, dest, incr
5714 // st[wd]cx. r0, ptr
5716 // fallthrough --> exitMBB
5718 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5719 .addReg(ptrA).addReg(ptrB);
5721 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5722 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5723 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5724 BuildMI(BB, dl, TII->get(PPC::BCC))
5725 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5726 BB->addSuccessor(loopMBB);
5727 BB->addSuccessor(exitMBB);
5736 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5737 MachineBasicBlock *BB,
5738 bool is8bit, // operation
5739 unsigned BinOpcode) const {
5740 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5741 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5742 // In 64 bit mode we have to use 64 bits for addresses, even though the
5743 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5744 // registers without caring whether they're 32 or 64, but here we're
5745 // doing actual arithmetic on the addresses.
5746 bool is64bit = PPCSubTarget.isPPC64();
5747 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5749 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5750 MachineFunction *F = BB->getParent();
5751 MachineFunction::iterator It = BB;
5754 unsigned dest = MI->getOperand(0).getReg();
5755 unsigned ptrA = MI->getOperand(1).getReg();
5756 unsigned ptrB = MI->getOperand(2).getReg();
5757 unsigned incr = MI->getOperand(3).getReg();
5758 DebugLoc dl = MI->getDebugLoc();
5760 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5761 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5762 F->insert(It, loopMBB);
5763 F->insert(It, exitMBB);
5764 exitMBB->splice(exitMBB->begin(), BB,
5765 llvm::next(MachineBasicBlock::iterator(MI)),
5767 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5769 MachineRegisterInfo &RegInfo = F->getRegInfo();
5770 const TargetRegisterClass *RC =
5771 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5772 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5773 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5774 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5775 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5776 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5777 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5778 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5779 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5780 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5781 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5782 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5783 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5785 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5789 // fallthrough --> loopMBB
5790 BB->addSuccessor(loopMBB);
5792 // The 4-byte load must be aligned, while a char or short may be
5793 // anywhere in the word. Hence all this nasty bookkeeping code.
5794 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5795 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5796 // xori shift, shift1, 24 [16]
5797 // rlwinm ptr, ptr1, 0, 0, 29
5798 // slw incr2, incr, shift
5799 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5800 // slw mask, mask2, shift
5802 // lwarx tmpDest, ptr
5803 // add tmp, tmpDest, incr2
5804 // andc tmp2, tmpDest, mask
5805 // and tmp3, tmp, mask
5806 // or tmp4, tmp3, tmp2
5809 // fallthrough --> exitMBB
5810 // srw dest, tmpDest, shift
5811 if (ptrA != ZeroReg) {
5812 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5813 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5814 .addReg(ptrA).addReg(ptrB);
5818 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5819 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5820 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5821 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5823 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5824 .addReg(Ptr1Reg).addImm(0).addImm(61);
5826 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5827 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5828 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5829 .addReg(incr).addReg(ShiftReg);
5831 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5833 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5834 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5836 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5837 .addReg(Mask2Reg).addReg(ShiftReg);
5840 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5841 .addReg(ZeroReg).addReg(PtrReg);
5843 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5844 .addReg(Incr2Reg).addReg(TmpDestReg);
5845 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5846 .addReg(TmpDestReg).addReg(MaskReg);
5847 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5848 .addReg(TmpReg).addReg(MaskReg);
5849 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5850 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5851 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5852 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5853 BuildMI(BB, dl, TII->get(PPC::BCC))
5854 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5855 BB->addSuccessor(loopMBB);
5856 BB->addSuccessor(exitMBB);
5861 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5867 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5868 MachineBasicBlock *BB) const {
5869 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5871 // To "insert" these instructions we actually have to insert their
5872 // control-flow patterns.
5873 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5874 MachineFunction::iterator It = BB;
5877 MachineFunction *F = BB->getParent();
5879 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5880 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5881 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5882 PPC::ISEL8 : PPC::ISEL;
5883 unsigned SelectPred = MI->getOperand(4).getImm();
5884 DebugLoc dl = MI->getDebugLoc();
5886 // The SelectPred is ((BI << 5) | BO) for a BCC
5887 unsigned BO = SelectPred & 0xF;
5888 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5890 unsigned TrueOpNo, FalseOpNo;
5897 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5900 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5901 .addReg(MI->getOperand(TrueOpNo).getReg())
5902 .addReg(MI->getOperand(FalseOpNo).getReg())
5903 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5904 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5905 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5906 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5907 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5908 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5911 // The incoming instruction knows the destination vreg to set, the
5912 // condition code register to branch on, the true/false values to
5913 // select between, and a branch opcode to use.
5918 // cmpTY ccX, r1, r2
5920 // fallthrough --> copy0MBB
5921 MachineBasicBlock *thisMBB = BB;
5922 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5923 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5924 unsigned SelectPred = MI->getOperand(4).getImm();
5925 DebugLoc dl = MI->getDebugLoc();
5926 F->insert(It, copy0MBB);
5927 F->insert(It, sinkMBB);
5929 // Transfer the remainder of BB and its successor edges to sinkMBB.
5930 sinkMBB->splice(sinkMBB->begin(), BB,
5931 llvm::next(MachineBasicBlock::iterator(MI)),
5933 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5935 // Next, add the true and fallthrough blocks as its successors.
5936 BB->addSuccessor(copy0MBB);
5937 BB->addSuccessor(sinkMBB);
5939 BuildMI(BB, dl, TII->get(PPC::BCC))
5940 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5943 // %FalseValue = ...
5944 // # fallthrough to sinkMBB
5947 // Update machine-CFG edges
5948 BB->addSuccessor(sinkMBB);
5951 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5954 BuildMI(*BB, BB->begin(), dl,
5955 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5956 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5957 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5959 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5960 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5961 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5962 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5963 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5964 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5965 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5966 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5968 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5969 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5970 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5971 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5972 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5973 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5974 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5975 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5977 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5978 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5979 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5980 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5981 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5982 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5983 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5984 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5986 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5987 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5988 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5989 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5990 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5991 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5992 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5993 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5995 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5996 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
5997 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
5998 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
5999 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
6000 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
6001 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
6002 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
6004 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
6005 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
6006 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
6007 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
6008 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
6009 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
6010 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
6011 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
6013 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
6014 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
6015 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
6016 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
6017 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
6018 BB = EmitAtomicBinary(MI, BB, false, 0);
6019 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
6020 BB = EmitAtomicBinary(MI, BB, true, 0);
6022 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
6023 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
6024 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
6026 unsigned dest = MI->getOperand(0).getReg();
6027 unsigned ptrA = MI->getOperand(1).getReg();
6028 unsigned ptrB = MI->getOperand(2).getReg();
6029 unsigned oldval = MI->getOperand(3).getReg();
6030 unsigned newval = MI->getOperand(4).getReg();
6031 DebugLoc dl = MI->getDebugLoc();
6033 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6034 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6035 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6036 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6037 F->insert(It, loop1MBB);
6038 F->insert(It, loop2MBB);
6039 F->insert(It, midMBB);
6040 F->insert(It, exitMBB);
6041 exitMBB->splice(exitMBB->begin(), BB,
6042 llvm::next(MachineBasicBlock::iterator(MI)),
6044 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6048 // fallthrough --> loopMBB
6049 BB->addSuccessor(loop1MBB);
6052 // l[wd]arx dest, ptr
6053 // cmp[wd] dest, oldval
6056 // st[wd]cx. newval, ptr
6060 // st[wd]cx. dest, ptr
6063 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
6064 .addReg(ptrA).addReg(ptrB);
6065 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
6066 .addReg(oldval).addReg(dest);
6067 BuildMI(BB, dl, TII->get(PPC::BCC))
6068 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6069 BB->addSuccessor(loop2MBB);
6070 BB->addSuccessor(midMBB);
6073 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6074 .addReg(newval).addReg(ptrA).addReg(ptrB);
6075 BuildMI(BB, dl, TII->get(PPC::BCC))
6076 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6077 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6078 BB->addSuccessor(loop1MBB);
6079 BB->addSuccessor(exitMBB);
6082 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6083 .addReg(dest).addReg(ptrA).addReg(ptrB);
6084 BB->addSuccessor(exitMBB);
6089 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
6090 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
6091 // We must use 64-bit registers for addresses when targeting 64-bit,
6092 // since we're actually doing arithmetic on them. Other registers
6094 bool is64bit = PPCSubTarget.isPPC64();
6095 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
6097 unsigned dest = MI->getOperand(0).getReg();
6098 unsigned ptrA = MI->getOperand(1).getReg();
6099 unsigned ptrB = MI->getOperand(2).getReg();
6100 unsigned oldval = MI->getOperand(3).getReg();
6101 unsigned newval = MI->getOperand(4).getReg();
6102 DebugLoc dl = MI->getDebugLoc();
6104 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6105 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6106 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6107 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6108 F->insert(It, loop1MBB);
6109 F->insert(It, loop2MBB);
6110 F->insert(It, midMBB);
6111 F->insert(It, exitMBB);
6112 exitMBB->splice(exitMBB->begin(), BB,
6113 llvm::next(MachineBasicBlock::iterator(MI)),
6115 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6117 MachineRegisterInfo &RegInfo = F->getRegInfo();
6118 const TargetRegisterClass *RC =
6119 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
6120 (const TargetRegisterClass *) &PPC::GPRCRegClass;
6121 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6122 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6123 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6124 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
6125 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
6126 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
6127 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
6128 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6129 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6130 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6131 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6132 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6133 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6135 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
6136 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
6139 // fallthrough --> loopMBB
6140 BB->addSuccessor(loop1MBB);
6142 // The 4-byte load must be aligned, while a char or short may be
6143 // anywhere in the word. Hence all this nasty bookkeeping code.
6144 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6145 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6146 // xori shift, shift1, 24 [16]
6147 // rlwinm ptr, ptr1, 0, 0, 29
6148 // slw newval2, newval, shift
6149 // slw oldval2, oldval,shift
6150 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6151 // slw mask, mask2, shift
6152 // and newval3, newval2, mask
6153 // and oldval3, oldval2, mask
6155 // lwarx tmpDest, ptr
6156 // and tmp, tmpDest, mask
6157 // cmpw tmp, oldval3
6160 // andc tmp2, tmpDest, mask
6161 // or tmp4, tmp2, newval3
6166 // stwcx. tmpDest, ptr
6168 // srw dest, tmpDest, shift
6169 if (ptrA != ZeroReg) {
6170 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6171 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6172 .addReg(ptrA).addReg(ptrB);
6176 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6177 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6178 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6179 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6181 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6182 .addReg(Ptr1Reg).addImm(0).addImm(61);
6184 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6185 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6186 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6187 .addReg(newval).addReg(ShiftReg);
6188 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6189 .addReg(oldval).addReg(ShiftReg);
6191 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6193 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6194 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6195 .addReg(Mask3Reg).addImm(65535);
6197 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6198 .addReg(Mask2Reg).addReg(ShiftReg);
6199 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6200 .addReg(NewVal2Reg).addReg(MaskReg);
6201 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6202 .addReg(OldVal2Reg).addReg(MaskReg);
6205 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6206 .addReg(ZeroReg).addReg(PtrReg);
6207 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6208 .addReg(TmpDestReg).addReg(MaskReg);
6209 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6210 .addReg(TmpReg).addReg(OldVal3Reg);
6211 BuildMI(BB, dl, TII->get(PPC::BCC))
6212 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6213 BB->addSuccessor(loop2MBB);
6214 BB->addSuccessor(midMBB);
6217 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6218 .addReg(TmpDestReg).addReg(MaskReg);
6219 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6220 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6221 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6222 .addReg(ZeroReg).addReg(PtrReg);
6223 BuildMI(BB, dl, TII->get(PPC::BCC))
6224 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6225 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6226 BB->addSuccessor(loop1MBB);
6227 BB->addSuccessor(exitMBB);
6230 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6231 .addReg(ZeroReg).addReg(PtrReg);
6232 BB->addSuccessor(exitMBB);
6237 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6240 llvm_unreachable("Unexpected instr type to insert");
6243 MI->eraseFromParent(); // The pseudo instruction is gone now.
6247 //===----------------------------------------------------------------------===//
6248 // Target Optimization Hooks
6249 //===----------------------------------------------------------------------===//
6251 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6252 DAGCombinerInfo &DCI) const {
6253 const TargetMachine &TM = getTargetMachine();
6254 SelectionDAG &DAG = DCI.DAG;
6255 DebugLoc dl = N->getDebugLoc();
6256 switch (N->getOpcode()) {
6259 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6260 if (C->isNullValue()) // 0 << V -> 0.
6261 return N->getOperand(0);
6265 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6266 if (C->isNullValue()) // 0 >>u V -> 0.
6267 return N->getOperand(0);
6271 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6272 if (C->isNullValue() || // 0 >>s V -> 0.
6273 C->isAllOnesValue()) // -1 >>s V -> -1.
6274 return N->getOperand(0);
6278 case ISD::SINT_TO_FP:
6279 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6280 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6281 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6282 // We allow the src/dst to be either f32/f64, but the intermediate
6283 // type must be i64.
6284 if (N->getOperand(0).getValueType() == MVT::i64 &&
6285 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6286 SDValue Val = N->getOperand(0).getOperand(0);
6287 if (Val.getValueType() == MVT::f32) {
6288 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6289 DCI.AddToWorklist(Val.getNode());
6292 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6293 DCI.AddToWorklist(Val.getNode());
6294 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6295 DCI.AddToWorklist(Val.getNode());
6296 if (N->getValueType(0) == MVT::f32) {
6297 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6298 DAG.getIntPtrConstant(0));
6299 DCI.AddToWorklist(Val.getNode());
6302 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6303 // If the intermediate type is i32, we can avoid the load/store here
6310 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6311 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6312 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6313 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6314 N->getOperand(1).getValueType() == MVT::i32 &&
6315 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6316 SDValue Val = N->getOperand(1).getOperand(0);
6317 if (Val.getValueType() == MVT::f32) {
6318 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6319 DCI.AddToWorklist(Val.getNode());
6321 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6322 DCI.AddToWorklist(Val.getNode());
6324 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6325 N->getOperand(2), N->getOperand(3));
6326 DCI.AddToWorklist(Val.getNode());
6330 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6331 if (cast<StoreSDNode>(N)->isUnindexed() &&
6332 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6333 N->getOperand(1).getNode()->hasOneUse() &&
6334 (N->getOperand(1).getValueType() == MVT::i32 ||
6335 N->getOperand(1).getValueType() == MVT::i16)) {
6336 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6337 // Do an any-extend to 32-bits if this is a half-word input.
6338 if (BSwapOp.getValueType() == MVT::i16)
6339 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6342 N->getOperand(0), BSwapOp, N->getOperand(2),
6343 DAG.getValueType(N->getOperand(1).getValueType())
6346 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6347 Ops, array_lengthof(Ops),
6348 cast<StoreSDNode>(N)->getMemoryVT(),
6349 cast<StoreSDNode>(N)->getMemOperand());
6353 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6354 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6355 N->getOperand(0).hasOneUse() &&
6356 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6357 SDValue Load = N->getOperand(0);
6358 LoadSDNode *LD = cast<LoadSDNode>(Load);
6359 // Create the byte-swapping load.
6361 LD->getChain(), // Chain
6362 LD->getBasePtr(), // Ptr
6363 DAG.getValueType(N->getValueType(0)) // VT
6366 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6367 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6368 LD->getMemoryVT(), LD->getMemOperand());
6370 // If this is an i16 load, insert the truncate.
6371 SDValue ResVal = BSLoad;
6372 if (N->getValueType(0) == MVT::i16)
6373 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6375 // First, combine the bswap away. This makes the value produced by the
6377 DCI.CombineTo(N, ResVal);
6379 // Next, combine the load away, we give it a bogus result value but a real
6380 // chain result. The result value is dead because the bswap is dead.
6381 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6383 // Return N so it doesn't get rechecked!
6384 return SDValue(N, 0);
6388 case PPCISD::VCMP: {
6389 // If a VCMPo node already exists with exactly the same operands as this
6390 // node, use its result instead of this node (VCMPo computes both a CR6 and
6391 // a normal output).
6393 if (!N->getOperand(0).hasOneUse() &&
6394 !N->getOperand(1).hasOneUse() &&
6395 !N->getOperand(2).hasOneUse()) {
6397 // Scan all of the users of the LHS, looking for VCMPo's that match.
6398 SDNode *VCMPoNode = 0;
6400 SDNode *LHSN = N->getOperand(0).getNode();
6401 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6403 if (UI->getOpcode() == PPCISD::VCMPo &&
6404 UI->getOperand(1) == N->getOperand(1) &&
6405 UI->getOperand(2) == N->getOperand(2) &&
6406 UI->getOperand(0) == N->getOperand(0)) {
6411 // If there is no VCMPo node, or if the flag value has a single use, don't
6413 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6416 // Look at the (necessarily single) use of the flag value. If it has a
6417 // chain, this transformation is more complex. Note that multiple things
6418 // could use the value result, which we should ignore.
6419 SDNode *FlagUser = 0;
6420 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6421 FlagUser == 0; ++UI) {
6422 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6424 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6425 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6432 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6433 // give up for right now.
6434 if (FlagUser->getOpcode() == PPCISD::MFCR)
6435 return SDValue(VCMPoNode, 0);
6440 // If this is a branch on an altivec predicate comparison, lower this so
6441 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6442 // lowering is done pre-legalize, because the legalizer lowers the predicate
6443 // compare down to code that is difficult to reassemble.
6444 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6445 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6449 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6450 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6451 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6452 assert(isDot && "Can't compare against a vector result!");
6454 // If this is a comparison against something other than 0/1, then we know
6455 // that the condition is never/always true.
6456 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6457 if (Val != 0 && Val != 1) {
6458 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6459 return N->getOperand(0);
6460 // Always !=, turn it into an unconditional branch.
6461 return DAG.getNode(ISD::BR, dl, MVT::Other,
6462 N->getOperand(0), N->getOperand(4));
6465 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6467 // Create the PPCISD altivec 'dot' comparison node.
6469 LHS.getOperand(2), // LHS of compare
6470 LHS.getOperand(3), // RHS of compare
6471 DAG.getConstant(CompareOpc, MVT::i32)
6473 EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue };
6474 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6476 // Unpack the result based on how the target uses it.
6477 PPC::Predicate CompOpc;
6478 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6479 default: // Can't happen, don't crash on invalid number though.
6480 case 0: // Branch on the value of the EQ bit of CR6.
6481 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6483 case 1: // Branch on the inverted value of the EQ bit of CR6.
6484 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6486 case 2: // Branch on the value of the LT bit of CR6.
6487 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6489 case 3: // Branch on the inverted value of the LT bit of CR6.
6490 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6494 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6495 DAG.getConstant(CompOpc, MVT::i32),
6496 DAG.getRegister(PPC::CR6, MVT::i32),
6497 N->getOperand(4), CompNode.getValue(1));
6506 //===----------------------------------------------------------------------===//
6507 // Inline Assembly Support
6508 //===----------------------------------------------------------------------===//
6510 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6513 const SelectionDAG &DAG,
6514 unsigned Depth) const {
6515 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6516 switch (Op.getOpcode()) {
6518 case PPCISD::LBRX: {
6519 // lhbrx is known to have the top bits cleared out.
6520 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6521 KnownZero = 0xFFFF0000;
6524 case ISD::INTRINSIC_WO_CHAIN: {
6525 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6527 case Intrinsic::ppc_altivec_vcmpbfp_p:
6528 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6529 case Intrinsic::ppc_altivec_vcmpequb_p:
6530 case Intrinsic::ppc_altivec_vcmpequh_p:
6531 case Intrinsic::ppc_altivec_vcmpequw_p:
6532 case Intrinsic::ppc_altivec_vcmpgefp_p:
6533 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6534 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6535 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6536 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6537 case Intrinsic::ppc_altivec_vcmpgtub_p:
6538 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6539 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6540 KnownZero = ~1U; // All bits but the low one are known to be zero.
6548 /// getConstraintType - Given a constraint, return the type of
6549 /// constraint it is for this target.
6550 PPCTargetLowering::ConstraintType
6551 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6552 if (Constraint.size() == 1) {
6553 switch (Constraint[0]) {
6560 return C_RegisterClass;
6562 // FIXME: While Z does indicate a memory constraint, it specifically
6563 // indicates an r+r address (used in conjunction with the 'y' modifier
6564 // in the replacement string). Currently, we're forcing the base
6565 // register to be r0 in the asm printer (which is interpreted as zero)
6566 // and forming the complete address in the second register. This is
6571 return TargetLowering::getConstraintType(Constraint);
6574 /// Examine constraint type and operand type and determine a weight value.
6575 /// This object must already have been set up with the operand type
6576 /// and the current alternative constraint selected.
6577 TargetLowering::ConstraintWeight
6578 PPCTargetLowering::getSingleConstraintMatchWeight(
6579 AsmOperandInfo &info, const char *constraint) const {
6580 ConstraintWeight weight = CW_Invalid;
6581 Value *CallOperandVal = info.CallOperandVal;
6582 // If we don't have a value, we can't do a match,
6583 // but allow it at the lowest weight.
6584 if (CallOperandVal == NULL)
6586 Type *type = CallOperandVal->getType();
6587 // Look at the constraint type.
6588 switch (*constraint) {
6590 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6593 if (type->isIntegerTy())
6594 weight = CW_Register;
6597 if (type->isFloatTy())
6598 weight = CW_Register;
6601 if (type->isDoubleTy())
6602 weight = CW_Register;
6605 if (type->isVectorTy())
6606 weight = CW_Register;
6609 weight = CW_Register;
6618 std::pair<unsigned, const TargetRegisterClass*>
6619 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6621 if (Constraint.size() == 1) {
6622 // GCC RS6000 Constraint Letters
6623 switch (Constraint[0]) {
6626 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6627 return std::make_pair(0U, &PPC::G8RCRegClass);
6628 return std::make_pair(0U, &PPC::GPRCRegClass);
6630 if (VT == MVT::f32 || VT == MVT::i32)
6631 return std::make_pair(0U, &PPC::F4RCRegClass);
6632 if (VT == MVT::f64 || VT == MVT::i64)
6633 return std::make_pair(0U, &PPC::F8RCRegClass);
6636 return std::make_pair(0U, &PPC::VRRCRegClass);
6638 return std::make_pair(0U, &PPC::CRRCRegClass);
6642 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6646 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6647 /// vector. If it is invalid, don't add anything to Ops.
6648 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6649 std::string &Constraint,
6650 std::vector<SDValue>&Ops,
6651 SelectionDAG &DAG) const {
6652 SDValue Result(0,0);
6654 // Only support length 1 constraints.
6655 if (Constraint.length() > 1) return;
6657 char Letter = Constraint[0];
6668 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6669 if (!CST) return; // Must be an immediate to match.
6670 unsigned Value = CST->getZExtValue();
6672 default: llvm_unreachable("Unknown constraint letter!");
6673 case 'I': // "I" is a signed 16-bit constant.
6674 if ((short)Value == (int)Value)
6675 Result = DAG.getTargetConstant(Value, Op.getValueType());
6677 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6678 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6679 if ((short)Value == 0)
6680 Result = DAG.getTargetConstant(Value, Op.getValueType());
6682 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6683 if ((Value >> 16) == 0)
6684 Result = DAG.getTargetConstant(Value, Op.getValueType());
6686 case 'M': // "M" is a constant that is greater than 31.
6688 Result = DAG.getTargetConstant(Value, Op.getValueType());
6690 case 'N': // "N" is a positive constant that is an exact power of two.
6691 if ((int)Value > 0 && isPowerOf2_32(Value))
6692 Result = DAG.getTargetConstant(Value, Op.getValueType());
6694 case 'O': // "O" is the constant zero.
6696 Result = DAG.getTargetConstant(Value, Op.getValueType());
6698 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6699 if ((short)-Value == (int)-Value)
6700 Result = DAG.getTargetConstant(Value, Op.getValueType());
6707 if (Result.getNode()) {
6708 Ops.push_back(Result);
6712 // Handle standard constraint letters.
6713 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6716 // isLegalAddressingMode - Return true if the addressing mode represented
6717 // by AM is legal for this target, for a load/store of the specified type.
6718 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6720 // FIXME: PPC does not allow r+i addressing modes for vectors!
6722 // PPC allows a sign-extended 16-bit immediate field.
6723 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6726 // No global is ever allowed as a base.
6730 // PPC only support r+r,
6732 case 0: // "r+i" or just "i", depending on HasBaseReg.
6735 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6737 // Otherwise we have r+r or r+i.
6740 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6742 // Allow 2*r as r+r.
6745 // No other scales are supported.
6752 /// isLegalAddressImmediate - Return true if the integer value can be used
6753 /// as the offset of the target addressing mode for load / store of the
6755 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6756 // PPC allows a sign-extended 16-bit immediate field.
6757 return (V > -(1 << 16) && V < (1 << 16)-1);
6760 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6764 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6765 SelectionDAG &DAG) const {
6766 MachineFunction &MF = DAG.getMachineFunction();
6767 MachineFrameInfo *MFI = MF.getFrameInfo();
6768 MFI->setReturnAddressIsTaken(true);
6770 DebugLoc dl = Op.getDebugLoc();
6771 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6773 // Make sure the function does not optimize away the store of the RA to
6775 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6776 FuncInfo->setLRStoreRequired();
6777 bool isPPC64 = PPCSubTarget.isPPC64();
6778 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6781 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6784 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6785 isPPC64? MVT::i64 : MVT::i32);
6786 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6787 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6789 MachinePointerInfo(), false, false, false, 0);
6792 // Just load the return address off the stack.
6793 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6794 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6795 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6798 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6799 SelectionDAG &DAG) const {
6800 DebugLoc dl = Op.getDebugLoc();
6801 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6803 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6804 bool isPPC64 = PtrVT == MVT::i64;
6806 MachineFunction &MF = DAG.getMachineFunction();
6807 MachineFrameInfo *MFI = MF.getFrameInfo();
6808 MFI->setFrameAddressIsTaken(true);
6809 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6810 MFI->hasVarSizedObjects()) &&
6811 MFI->getStackSize() &&
6812 !MF.getFunction()->getAttributes().
6813 hasAttribute(AttributeSet::FunctionIndex, Attribute::Naked);
6814 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6815 (is31 ? PPC::R31 : PPC::R1);
6816 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6819 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6820 FrameAddr, MachinePointerInfo(), false, false,
6826 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6827 // The PowerPC target isn't yet aware of offsets.
6831 /// getOptimalMemOpType - Returns the target specific optimal type for load
6832 /// and store operations as a result of memset, memcpy, and memmove
6833 /// lowering. If DstAlign is zero that means it's safe to destination
6834 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6835 /// means there isn't a need to check it against alignment requirement,
6836 /// probably because the source does not need to be loaded. If 'IsMemset' is
6837 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
6838 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
6839 /// source is constant so it does not need to be loaded.
6840 /// It returns EVT::Other if the type should be determined using generic
6841 /// target-independent logic.
6842 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6843 unsigned DstAlign, unsigned SrcAlign,
6844 bool IsMemset, bool ZeroMemset,
6846 MachineFunction &MF) const {
6847 if (this->PPCSubTarget.isPPC64()) {
6854 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6855 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6856 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6857 /// is expanded to mul + add.
6858 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6862 switch (VT.getSimpleVT().SimpleTy) {
6874 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6876 return TargetLowering::getSchedulingPreference(N);