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 cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned",
61 cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden);
63 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
64 if (TM.getSubtargetImpl()->isDarwin())
65 return new TargetLoweringObjectFileMachO();
67 return new TargetLoweringObjectFileELF();
70 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
71 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
72 const PPCSubtarget *Subtarget = &TM.getSubtarget<PPCSubtarget>();
76 // Use _setjmp/_longjmp instead of setjmp/longjmp.
77 setUseUnderscoreSetJmp(true);
78 setUseUnderscoreLongJmp(true);
80 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
81 // arguments are at least 4/8 bytes aligned.
82 bool isPPC64 = Subtarget->isPPC64();
83 setMinStackArgumentAlignment(isPPC64 ? 8:4);
85 // Set up the register classes.
86 addRegisterClass(MVT::i32, &PPC::GPRCRegClass);
87 addRegisterClass(MVT::f32, &PPC::F4RCRegClass);
88 addRegisterClass(MVT::f64, &PPC::F8RCRegClass);
90 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
91 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
92 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
94 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
96 // PowerPC has pre-inc load and store's.
97 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
98 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
99 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
100 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
101 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
102 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
103 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
104 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
105 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
106 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
108 // This is used in the ppcf128->int sequence. Note it has different semantics
109 // from FP_ROUND: that rounds to nearest, this rounds to zero.
110 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
112 // We do not currently implement these libm ops for PowerPC.
113 setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand);
114 setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand);
115 setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand);
116 setOperationAction(ISD::FRINT, MVT::ppcf128, Expand);
117 setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand);
119 // PowerPC has no SREM/UREM instructions
120 setOperationAction(ISD::SREM, MVT::i32, Expand);
121 setOperationAction(ISD::UREM, MVT::i32, Expand);
122 setOperationAction(ISD::SREM, MVT::i64, Expand);
123 setOperationAction(ISD::UREM, MVT::i64, Expand);
125 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
126 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
127 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
128 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
129 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
130 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
131 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
132 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
133 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
135 // We don't support sin/cos/sqrt/fmod/pow
136 setOperationAction(ISD::FSIN , MVT::f64, Expand);
137 setOperationAction(ISD::FCOS , MVT::f64, Expand);
138 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
139 setOperationAction(ISD::FREM , MVT::f64, Expand);
140 setOperationAction(ISD::FPOW , MVT::f64, Expand);
141 setOperationAction(ISD::FMA , MVT::f64, Legal);
142 setOperationAction(ISD::FSIN , MVT::f32, Expand);
143 setOperationAction(ISD::FCOS , MVT::f32, Expand);
144 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
145 setOperationAction(ISD::FREM , MVT::f32, Expand);
146 setOperationAction(ISD::FPOW , MVT::f32, Expand);
147 setOperationAction(ISD::FMA , MVT::f32, Legal);
149 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
151 // If we're enabling GP optimizations, use hardware square root
152 if (!Subtarget->hasFSQRT()) {
153 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
154 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
157 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
158 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
160 // PowerPC does not have BSWAP, CTPOP or CTTZ
161 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
162 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
163 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
164 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
165 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
166 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
167 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
168 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
169 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
170 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
172 // PowerPC does not have ROTR
173 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
174 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
176 // PowerPC does not have Select
177 setOperationAction(ISD::SELECT, MVT::i32, Expand);
178 setOperationAction(ISD::SELECT, MVT::i64, Expand);
179 setOperationAction(ISD::SELECT, MVT::f32, Expand);
180 setOperationAction(ISD::SELECT, MVT::f64, Expand);
182 // PowerPC wants to turn select_cc of FP into fsel when possible.
183 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
184 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
186 // PowerPC wants to optimize integer setcc a bit
187 setOperationAction(ISD::SETCC, MVT::i32, Custom);
189 // PowerPC does not have BRCOND which requires SetCC
190 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
192 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
194 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
195 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
197 // PowerPC does not have [U|S]INT_TO_FP
198 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
199 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
201 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
202 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
203 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
204 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
206 // We cannot sextinreg(i1). Expand to shifts.
207 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
209 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
210 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
211 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
212 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
215 // We want to legalize GlobalAddress and ConstantPool nodes into the
216 // appropriate instructions to materialize the address.
217 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
218 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
219 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
220 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
221 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
222 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
223 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
224 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
225 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
226 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
229 setOperationAction(ISD::TRAP, MVT::Other, Legal);
231 // TRAMPOLINE is custom lowered.
232 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
233 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
235 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
236 setOperationAction(ISD::VASTART , MVT::Other, Custom);
238 if (Subtarget->isSVR4ABI()) {
240 // VAARG always uses double-word chunks, so promote anything smaller.
241 setOperationAction(ISD::VAARG, MVT::i1, Promote);
242 AddPromotedToType (ISD::VAARG, MVT::i1, MVT::i64);
243 setOperationAction(ISD::VAARG, MVT::i8, Promote);
244 AddPromotedToType (ISD::VAARG, MVT::i8, MVT::i64);
245 setOperationAction(ISD::VAARG, MVT::i16, Promote);
246 AddPromotedToType (ISD::VAARG, MVT::i16, MVT::i64);
247 setOperationAction(ISD::VAARG, MVT::i32, Promote);
248 AddPromotedToType (ISD::VAARG, MVT::i32, MVT::i64);
249 setOperationAction(ISD::VAARG, MVT::Other, Expand);
251 // VAARG is custom lowered with the 32-bit SVR4 ABI.
252 setOperationAction(ISD::VAARG, MVT::Other, Custom);
253 setOperationAction(ISD::VAARG, MVT::i64, Custom);
256 setOperationAction(ISD::VAARG, MVT::Other, Expand);
258 // Use the default implementation.
259 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
260 setOperationAction(ISD::VAEND , MVT::Other, Expand);
261 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
262 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
263 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
264 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
266 // We want to custom lower some of our intrinsics.
267 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
269 // Comparisons that require checking two conditions.
270 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
271 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
272 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
273 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
274 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
275 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
276 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
277 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
278 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
279 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
280 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
281 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
283 if (Subtarget->has64BitSupport()) {
284 // They also have instructions for converting between i64 and fp.
285 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
286 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
287 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
288 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
289 // This is just the low 32 bits of a (signed) fp->i64 conversion.
290 // We cannot do this with Promote because i64 is not a legal type.
291 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
293 // FIXME: disable this lowered code. This generates 64-bit register values,
294 // and we don't model the fact that the top part is clobbered by calls. We
295 // need to flag these together so that the value isn't live across a call.
296 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
298 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
299 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
302 if (Subtarget->use64BitRegs()) {
303 // 64-bit PowerPC implementations can support i64 types directly
304 addRegisterClass(MVT::i64, &PPC::G8RCRegClass);
305 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
306 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
307 // 64-bit PowerPC wants to expand i128 shifts itself.
308 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
309 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
310 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
312 // 32-bit PowerPC wants to expand i64 shifts itself.
313 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
314 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
315 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
318 if (Subtarget->hasAltivec()) {
319 // First set operation action for all vector types to expand. Then we
320 // will selectively turn on ones that can be effectively codegen'd.
321 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
322 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
323 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
325 // add/sub are legal for all supported vector VT's.
326 setOperationAction(ISD::ADD , VT, Legal);
327 setOperationAction(ISD::SUB , VT, Legal);
329 // We promote all shuffles to v16i8.
330 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
331 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
333 // We promote all non-typed operations to v4i32.
334 setOperationAction(ISD::AND , VT, Promote);
335 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
336 setOperationAction(ISD::OR , VT, Promote);
337 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
338 setOperationAction(ISD::XOR , VT, Promote);
339 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
340 setOperationAction(ISD::LOAD , VT, Promote);
341 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
342 setOperationAction(ISD::SELECT, VT, Promote);
343 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
344 setOperationAction(ISD::STORE, VT, Promote);
345 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
347 // No other operations are legal.
348 setOperationAction(ISD::MUL , VT, Expand);
349 setOperationAction(ISD::SDIV, VT, Expand);
350 setOperationAction(ISD::SREM, VT, Expand);
351 setOperationAction(ISD::UDIV, VT, Expand);
352 setOperationAction(ISD::UREM, VT, Expand);
353 setOperationAction(ISD::FDIV, VT, Expand);
354 setOperationAction(ISD::FNEG, VT, Expand);
355 setOperationAction(ISD::FSQRT, VT, Expand);
356 setOperationAction(ISD::FLOG, VT, Expand);
357 setOperationAction(ISD::FLOG10, VT, Expand);
358 setOperationAction(ISD::FLOG2, VT, Expand);
359 setOperationAction(ISD::FEXP, VT, Expand);
360 setOperationAction(ISD::FEXP2, VT, Expand);
361 setOperationAction(ISD::FSIN, VT, Expand);
362 setOperationAction(ISD::FCOS, VT, Expand);
363 setOperationAction(ISD::FABS, VT, Expand);
364 setOperationAction(ISD::FPOWI, VT, Expand);
365 setOperationAction(ISD::FFLOOR, VT, Expand);
366 setOperationAction(ISD::FCEIL, VT, Expand);
367 setOperationAction(ISD::FTRUNC, VT, Expand);
368 setOperationAction(ISD::FRINT, VT, Expand);
369 setOperationAction(ISD::FNEARBYINT, VT, Expand);
370 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
371 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
372 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
373 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
374 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
375 setOperationAction(ISD::UDIVREM, VT, Expand);
376 setOperationAction(ISD::SDIVREM, VT, Expand);
377 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
378 setOperationAction(ISD::FPOW, VT, Expand);
379 setOperationAction(ISD::CTPOP, VT, Expand);
380 setOperationAction(ISD::CTLZ, VT, Expand);
381 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
382 setOperationAction(ISD::CTTZ, VT, Expand);
383 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
384 setOperationAction(ISD::VSELECT, VT, Expand);
385 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
387 for (unsigned j = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
388 j <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++j) {
389 MVT::SimpleValueType InnerVT = (MVT::SimpleValueType)j;
390 setTruncStoreAction(VT, InnerVT, Expand);
392 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
393 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
394 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
397 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
398 // with merges, splats, etc.
399 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
401 setOperationAction(ISD::AND , MVT::v4i32, Legal);
402 setOperationAction(ISD::OR , MVT::v4i32, Legal);
403 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
404 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
405 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
406 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
407 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
408 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
409 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
410 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
411 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
412 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
413 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
414 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
416 addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass);
417 addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass);
418 addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass);
419 addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass);
421 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
422 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
423 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
424 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
425 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
427 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
428 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
430 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
431 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
432 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
433 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
435 // Altivec does not contain unordered floating-point compare instructions
436 setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
437 setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
438 setCondCodeAction(ISD::SETUGT, MVT::v4f32, Expand);
439 setCondCodeAction(ISD::SETUGE, MVT::v4f32, Expand);
440 setCondCodeAction(ISD::SETULT, MVT::v4f32, Expand);
441 setCondCodeAction(ISD::SETULE, MVT::v4f32, Expand);
444 if (Subtarget->has64BitSupport()) {
445 setOperationAction(ISD::PREFETCH, MVT::Other, Legal);
446 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);
449 setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Expand);
450 setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Expand);
451 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
452 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
454 setBooleanContents(ZeroOrOneBooleanContent);
455 setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct?
458 setStackPointerRegisterToSaveRestore(PPC::X1);
459 setExceptionPointerRegister(PPC::X3);
460 setExceptionSelectorRegister(PPC::X4);
462 setStackPointerRegisterToSaveRestore(PPC::R1);
463 setExceptionPointerRegister(PPC::R3);
464 setExceptionSelectorRegister(PPC::R4);
467 // We have target-specific dag combine patterns for the following nodes:
468 setTargetDAGCombine(ISD::SINT_TO_FP);
469 setTargetDAGCombine(ISD::STORE);
470 setTargetDAGCombine(ISD::BR_CC);
471 setTargetDAGCombine(ISD::BSWAP);
473 // Darwin long double math library functions have $LDBL128 appended.
474 if (Subtarget->isDarwin()) {
475 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
476 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
477 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
478 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
479 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
480 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
481 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
482 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
483 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
484 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
487 setMinFunctionAlignment(2);
488 if (PPCSubTarget.isDarwin())
489 setPrefFunctionAlignment(4);
491 if (isPPC64 && Subtarget->isJITCodeModel())
492 // Temporary workaround for the inability of PPC64 JIT to handle jump
494 setSupportJumpTables(false);
496 setInsertFencesForAtomic(true);
498 setSchedulingPreference(Sched::Hybrid);
500 computeRegisterProperties();
502 // The Freescale cores does better with aggressive inlining of memcpy and
503 // friends. Gcc uses same threshold of 128 bytes (= 32 word stores).
504 if (Subtarget->getDarwinDirective() == PPC::DIR_E500mc ||
505 Subtarget->getDarwinDirective() == PPC::DIR_E5500) {
506 MaxStoresPerMemset = 32;
507 MaxStoresPerMemsetOptSize = 16;
508 MaxStoresPerMemcpy = 32;
509 MaxStoresPerMemcpyOptSize = 8;
510 MaxStoresPerMemmove = 32;
511 MaxStoresPerMemmoveOptSize = 8;
513 setPrefFunctionAlignment(4);
514 BenefitFromCodePlacementOpt = true;
518 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
519 /// function arguments in the caller parameter area.
520 unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty) const {
521 const TargetMachine &TM = getTargetMachine();
522 // Darwin passes everything on 4 byte boundary.
523 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
526 // 16byte and wider vectors are passed on 16byte boundary.
527 if (VectorType *VTy = dyn_cast<VectorType>(Ty))
528 if (VTy->getBitWidth() >= 128)
531 // The rest is 8 on PPC64 and 4 on PPC32 boundary.
532 if (PPCSubTarget.isPPC64())
538 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
541 case PPCISD::FSEL: return "PPCISD::FSEL";
542 case PPCISD::FCFID: return "PPCISD::FCFID";
543 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
544 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
545 case PPCISD::STFIWX: return "PPCISD::STFIWX";
546 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
547 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
548 case PPCISD::VPERM: return "PPCISD::VPERM";
549 case PPCISD::Hi: return "PPCISD::Hi";
550 case PPCISD::Lo: return "PPCISD::Lo";
551 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
552 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
553 case PPCISD::LOAD: return "PPCISD::LOAD";
554 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
555 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
556 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
557 case PPCISD::SRL: return "PPCISD::SRL";
558 case PPCISD::SRA: return "PPCISD::SRA";
559 case PPCISD::SHL: return "PPCISD::SHL";
560 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
561 case PPCISD::STD_32: return "PPCISD::STD_32";
562 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
563 case PPCISD::CALL_NOP_SVR4: return "PPCISD::CALL_NOP_SVR4";
564 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
565 case PPCISD::NOP: return "PPCISD::NOP";
566 case PPCISD::MTCTR: return "PPCISD::MTCTR";
567 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
568 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
569 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
570 case PPCISD::MFCR: return "PPCISD::MFCR";
571 case PPCISD::VCMP: return "PPCISD::VCMP";
572 case PPCISD::VCMPo: return "PPCISD::VCMPo";
573 case PPCISD::LBRX: return "PPCISD::LBRX";
574 case PPCISD::STBRX: return "PPCISD::STBRX";
575 case PPCISD::LARX: return "PPCISD::LARX";
576 case PPCISD::STCX: return "PPCISD::STCX";
577 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
578 case PPCISD::MFFS: return "PPCISD::MFFS";
579 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
580 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
581 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
582 case PPCISD::MTFSF: return "PPCISD::MTFSF";
583 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
584 case PPCISD::CR6SET: return "PPCISD::CR6SET";
585 case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET";
586 case PPCISD::ADDIS_TOC_HA: return "PPCISD::ADDIS_TOC_HA";
587 case PPCISD::LD_TOC_L: return "PPCISD::LD_TOC_L";
588 case PPCISD::ADDI_TOC_L: return "PPCISD::ADDI_TOC_L";
589 case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA";
590 case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L";
591 case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS";
592 case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA";
593 case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L";
594 case PPCISD::GET_TLS_ADDR: return "PPCISD::GET_TLS_ADDR";
595 case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA";
596 case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L";
597 case PPCISD::GET_TLSLD_ADDR: return "PPCISD::GET_TLSLD_ADDR";
598 case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA";
599 case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L";
600 case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT";
604 EVT PPCTargetLowering::getSetCCResultType(EVT VT) const {
607 return VT.changeVectorElementTypeToInteger();
610 //===----------------------------------------------------------------------===//
611 // Node matching predicates, for use by the tblgen matching code.
612 //===----------------------------------------------------------------------===//
614 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
615 static bool isFloatingPointZero(SDValue Op) {
616 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
617 return CFP->getValueAPF().isZero();
618 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
619 // Maybe this has already been legalized into the constant pool?
620 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
621 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
622 return CFP->getValueAPF().isZero();
627 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
628 /// true if Op is undef or if it matches the specified value.
629 static bool isConstantOrUndef(int Op, int Val) {
630 return Op < 0 || Op == Val;
633 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
634 /// VPKUHUM instruction.
635 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
637 for (unsigned i = 0; i != 16; ++i)
638 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
641 for (unsigned i = 0; i != 8; ++i)
642 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
643 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
649 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
650 /// VPKUWUM instruction.
651 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
653 for (unsigned i = 0; i != 16; i += 2)
654 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
655 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
658 for (unsigned i = 0; i != 8; i += 2)
659 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
660 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
661 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
662 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
668 /// isVMerge - Common function, used to match vmrg* shuffles.
670 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
671 unsigned LHSStart, unsigned RHSStart) {
672 assert(N->getValueType(0) == MVT::v16i8 &&
673 "PPC only supports shuffles by bytes!");
674 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
675 "Unsupported merge size!");
677 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
678 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
679 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
680 LHSStart+j+i*UnitSize) ||
681 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
682 RHSStart+j+i*UnitSize))
688 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
689 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
690 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
693 return isVMerge(N, UnitSize, 8, 24);
694 return isVMerge(N, UnitSize, 8, 8);
697 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
698 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
699 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
702 return isVMerge(N, UnitSize, 0, 16);
703 return isVMerge(N, UnitSize, 0, 0);
707 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
708 /// amount, otherwise return -1.
709 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
710 assert(N->getValueType(0) == MVT::v16i8 &&
711 "PPC only supports shuffles by bytes!");
713 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
715 // Find the first non-undef value in the shuffle mask.
717 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
720 if (i == 16) return -1; // all undef.
722 // Otherwise, check to see if the rest of the elements are consecutively
723 // numbered from this value.
724 unsigned ShiftAmt = SVOp->getMaskElt(i);
725 if (ShiftAmt < i) return -1;
729 // Check the rest of the elements to see if they are consecutive.
730 for (++i; i != 16; ++i)
731 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
734 // Check the rest of the elements to see if they are consecutive.
735 for (++i; i != 16; ++i)
736 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
742 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
743 /// specifies a splat of a single element that is suitable for input to
744 /// VSPLTB/VSPLTH/VSPLTW.
745 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
746 assert(N->getValueType(0) == MVT::v16i8 &&
747 (EltSize == 1 || EltSize == 2 || EltSize == 4));
749 // This is a splat operation if each element of the permute is the same, and
750 // if the value doesn't reference the second vector.
751 unsigned ElementBase = N->getMaskElt(0);
753 // FIXME: Handle UNDEF elements too!
754 if (ElementBase >= 16)
757 // Check that the indices are consecutive, in the case of a multi-byte element
758 // splatted with a v16i8 mask.
759 for (unsigned i = 1; i != EltSize; ++i)
760 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
763 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
764 if (N->getMaskElt(i) < 0) continue;
765 for (unsigned j = 0; j != EltSize; ++j)
766 if (N->getMaskElt(i+j) != N->getMaskElt(j))
772 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
774 bool PPC::isAllNegativeZeroVector(SDNode *N) {
775 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
777 APInt APVal, APUndef;
781 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
782 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
783 return CFP->getValueAPF().isNegZero();
788 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
789 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
790 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
791 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
792 assert(isSplatShuffleMask(SVOp, EltSize));
793 return SVOp->getMaskElt(0) / EltSize;
796 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
797 /// by using a vspltis[bhw] instruction of the specified element size, return
798 /// the constant being splatted. The ByteSize field indicates the number of
799 /// bytes of each element [124] -> [bhw].
800 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
803 // If ByteSize of the splat is bigger than the element size of the
804 // build_vector, then we have a case where we are checking for a splat where
805 // multiple elements of the buildvector are folded together into a single
806 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
807 unsigned EltSize = 16/N->getNumOperands();
808 if (EltSize < ByteSize) {
809 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
810 SDValue UniquedVals[4];
811 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
813 // See if all of the elements in the buildvector agree across.
814 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
815 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
816 // If the element isn't a constant, bail fully out.
817 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
820 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
821 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
822 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
823 return SDValue(); // no match.
826 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
827 // either constant or undef values that are identical for each chunk. See
828 // if these chunks can form into a larger vspltis*.
830 // Check to see if all of the leading entries are either 0 or -1. If
831 // neither, then this won't fit into the immediate field.
832 bool LeadingZero = true;
833 bool LeadingOnes = true;
834 for (unsigned i = 0; i != Multiple-1; ++i) {
835 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
837 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
838 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
840 // Finally, check the least significant entry.
842 if (UniquedVals[Multiple-1].getNode() == 0)
843 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
844 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
846 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
849 if (UniquedVals[Multiple-1].getNode() == 0)
850 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
851 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
852 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
853 return DAG.getTargetConstant(Val, MVT::i32);
859 // Check to see if this buildvec has a single non-undef value in its elements.
860 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
861 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
862 if (OpVal.getNode() == 0)
863 OpVal = N->getOperand(i);
864 else if (OpVal != N->getOperand(i))
868 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
870 unsigned ValSizeInBytes = EltSize;
872 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
873 Value = CN->getZExtValue();
874 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
875 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
876 Value = FloatToBits(CN->getValueAPF().convertToFloat());
879 // If the splat value is larger than the element value, then we can never do
880 // this splat. The only case that we could fit the replicated bits into our
881 // immediate field for would be zero, and we prefer to use vxor for it.
882 if (ValSizeInBytes < ByteSize) return SDValue();
884 // If the element value is larger than the splat value, cut it in half and
885 // check to see if the two halves are equal. Continue doing this until we
886 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
887 while (ValSizeInBytes > ByteSize) {
888 ValSizeInBytes >>= 1;
890 // If the top half equals the bottom half, we're still ok.
891 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
892 (Value & ((1 << (8*ValSizeInBytes))-1)))
896 // Properly sign extend the value.
897 int MaskVal = SignExtend32(Value, ByteSize * 8);
899 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
900 if (MaskVal == 0) return SDValue();
902 // Finally, if this value fits in a 5 bit sext field, return it
903 if (SignExtend32<5>(MaskVal) == MaskVal)
904 return DAG.getTargetConstant(MaskVal, MVT::i32);
908 //===----------------------------------------------------------------------===//
909 // Addressing Mode Selection
910 //===----------------------------------------------------------------------===//
912 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
913 /// or 64-bit immediate, and if the value can be accurately represented as a
914 /// sign extension from a 16-bit value. If so, this returns true and the
916 static bool isIntS16Immediate(SDNode *N, short &Imm) {
917 if (N->getOpcode() != ISD::Constant)
920 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
921 if (N->getValueType(0) == MVT::i32)
922 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
924 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
926 static bool isIntS16Immediate(SDValue Op, short &Imm) {
927 return isIntS16Immediate(Op.getNode(), Imm);
931 /// SelectAddressRegReg - Given the specified addressed, check to see if it
932 /// can be represented as an indexed [r+r] operation. Returns false if it
933 /// can be more efficiently represented with [r+imm].
934 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
936 SelectionDAG &DAG) const {
938 if (N.getOpcode() == ISD::ADD) {
939 if (isIntS16Immediate(N.getOperand(1), imm))
941 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
944 Base = N.getOperand(0);
945 Index = N.getOperand(1);
947 } else if (N.getOpcode() == ISD::OR) {
948 if (isIntS16Immediate(N.getOperand(1), imm))
949 return false; // r+i can fold it if we can.
951 // If this is an or of disjoint bitfields, we can codegen this as an add
952 // (for better address arithmetic) if the LHS and RHS of the OR are provably
954 APInt LHSKnownZero, LHSKnownOne;
955 APInt RHSKnownZero, RHSKnownOne;
956 DAG.ComputeMaskedBits(N.getOperand(0),
957 LHSKnownZero, LHSKnownOne);
959 if (LHSKnownZero.getBoolValue()) {
960 DAG.ComputeMaskedBits(N.getOperand(1),
961 RHSKnownZero, RHSKnownOne);
962 // If all of the bits are known zero on the LHS or RHS, the add won't
964 if (~(LHSKnownZero | RHSKnownZero) == 0) {
965 Base = N.getOperand(0);
966 Index = N.getOperand(1);
975 /// Returns true if the address N can be represented by a base register plus
976 /// a signed 16-bit displacement [r+imm], and if it is not better
977 /// represented as reg+reg.
978 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
980 SelectionDAG &DAG) const {
981 // FIXME dl should come from parent load or store, not from address
982 DebugLoc dl = N.getDebugLoc();
983 // If this can be more profitably realized as r+r, fail.
984 if (SelectAddressRegReg(N, Disp, Base, DAG))
987 if (N.getOpcode() == ISD::ADD) {
989 if (isIntS16Immediate(N.getOperand(1), imm)) {
990 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
991 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
992 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
994 Base = N.getOperand(0);
996 return true; // [r+i]
997 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
998 // Match LOAD (ADD (X, Lo(G))).
999 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1000 && "Cannot handle constant offsets yet!");
1001 Disp = N.getOperand(1).getOperand(0); // The global address.
1002 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1003 Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||
1004 Disp.getOpcode() == ISD::TargetConstantPool ||
1005 Disp.getOpcode() == ISD::TargetJumpTable);
1006 Base = N.getOperand(0);
1007 return true; // [&g+r]
1009 } else if (N.getOpcode() == ISD::OR) {
1011 if (isIntS16Immediate(N.getOperand(1), imm)) {
1012 // If this is an or of disjoint bitfields, we can codegen this as an add
1013 // (for better address arithmetic) if the LHS and RHS of the OR are
1014 // provably disjoint.
1015 APInt LHSKnownZero, LHSKnownOne;
1016 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1018 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1019 // If all of the bits are known zero on the LHS or RHS, the add won't
1021 Base = N.getOperand(0);
1022 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
1026 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1027 // Loading from a constant address.
1029 // If this address fits entirely in a 16-bit sext immediate field, codegen
1032 if (isIntS16Immediate(CN, Imm)) {
1033 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
1034 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1035 CN->getValueType(0));
1039 // Handle 32-bit sext immediates with LIS + addr mode.
1040 if (CN->getValueType(0) == MVT::i32 ||
1041 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1042 int Addr = (int)CN->getZExtValue();
1044 // Otherwise, break this down into an LIS + disp.
1045 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
1047 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
1048 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1049 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
1054 Disp = DAG.getTargetConstant(0, getPointerTy());
1055 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1056 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1059 return true; // [r+0]
1062 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
1063 /// represented as an indexed [r+r] operation.
1064 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
1066 SelectionDAG &DAG) const {
1067 // Check to see if we can easily represent this as an [r+r] address. This
1068 // will fail if it thinks that the address is more profitably represented as
1069 // reg+imm, e.g. where imm = 0.
1070 if (SelectAddressRegReg(N, Base, Index, DAG))
1073 // If the operand is an addition, always emit this as [r+r], since this is
1074 // better (for code size, and execution, as the memop does the add for free)
1075 // than emitting an explicit add.
1076 if (N.getOpcode() == ISD::ADD) {
1077 Base = N.getOperand(0);
1078 Index = N.getOperand(1);
1082 // Otherwise, do it the hard way, using R0 as the base register.
1083 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1089 /// SelectAddressRegImmShift - Returns true if the address N can be
1090 /// represented by a base register plus a signed 14-bit displacement
1091 /// [r+imm*4]. Suitable for use by STD and friends.
1092 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
1094 SelectionDAG &DAG) const {
1095 // FIXME dl should come from the parent load or store, not the address
1096 DebugLoc dl = N.getDebugLoc();
1097 // If this can be more profitably realized as r+r, fail.
1098 if (SelectAddressRegReg(N, Disp, Base, DAG))
1101 if (N.getOpcode() == ISD::ADD) {
1103 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1104 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1105 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
1106 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1108 Base = N.getOperand(0);
1110 return true; // [r+i]
1111 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
1112 // Match LOAD (ADD (X, Lo(G))).
1113 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
1114 && "Cannot handle constant offsets yet!");
1115 Disp = N.getOperand(1).getOperand(0); // The global address.
1116 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
1117 Disp.getOpcode() == ISD::TargetConstantPool ||
1118 Disp.getOpcode() == ISD::TargetJumpTable);
1119 Base = N.getOperand(0);
1120 return true; // [&g+r]
1122 } else if (N.getOpcode() == ISD::OR) {
1124 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
1125 // If this is an or of disjoint bitfields, we can codegen this as an add
1126 // (for better address arithmetic) if the LHS and RHS of the OR are
1127 // provably disjoint.
1128 APInt LHSKnownZero, LHSKnownOne;
1129 DAG.ComputeMaskedBits(N.getOperand(0), LHSKnownZero, LHSKnownOne);
1130 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1131 // If all of the bits are known zero on the LHS or RHS, the add won't
1133 Base = N.getOperand(0);
1134 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1138 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1139 // Loading from a constant address. Verify low two bits are clear.
1140 if ((CN->getZExtValue() & 3) == 0) {
1141 // If this address fits entirely in a 14-bit sext immediate field, codegen
1144 if (isIntS16Immediate(CN, Imm)) {
1145 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1146 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1147 CN->getValueType(0));
1151 // Fold the low-part of 32-bit absolute addresses into addr mode.
1152 if (CN->getValueType(0) == MVT::i32 ||
1153 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1154 int Addr = (int)CN->getZExtValue();
1156 // Otherwise, break this down into an LIS + disp.
1157 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1158 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1159 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1160 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1166 Disp = DAG.getTargetConstant(0, getPointerTy());
1167 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1168 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1171 return true; // [r+0]
1175 /// getPreIndexedAddressParts - returns true by value, base pointer and
1176 /// offset pointer and addressing mode by reference if the node's address
1177 /// can be legally represented as pre-indexed load / store address.
1178 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1180 ISD::MemIndexedMode &AM,
1181 SelectionDAG &DAG) const {
1182 if (DisablePPCPreinc) return false;
1186 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1187 Ptr = LD->getBasePtr();
1188 VT = LD->getMemoryVT();
1190 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1191 Ptr = ST->getBasePtr();
1192 VT = ST->getMemoryVT();
1196 // PowerPC doesn't have preinc load/store instructions for vectors.
1200 if (SelectAddressRegReg(Ptr, Offset, Base, DAG)) {
1205 // LDU/STU use reg+imm*4, others use reg+imm.
1206 if (VT != MVT::i64) {
1208 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1212 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1216 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1217 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1218 // sext i32 to i64 when addr mode is r+i.
1219 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1220 LD->getExtensionType() == ISD::SEXTLOAD &&
1221 isa<ConstantSDNode>(Offset))
1229 //===----------------------------------------------------------------------===//
1230 // LowerOperation implementation
1231 //===----------------------------------------------------------------------===//
1233 /// GetLabelAccessInfo - Return true if we should reference labels using a
1234 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1235 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1236 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1237 HiOpFlags = PPCII::MO_HA16;
1238 LoOpFlags = PPCII::MO_LO16;
1240 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1241 // non-darwin platform. We don't support PIC on other platforms yet.
1242 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1243 TM.getSubtarget<PPCSubtarget>().isDarwin();
1245 HiOpFlags |= PPCII::MO_PIC_FLAG;
1246 LoOpFlags |= PPCII::MO_PIC_FLAG;
1249 // If this is a reference to a global value that requires a non-lazy-ptr, make
1250 // sure that instruction lowering adds it.
1251 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1252 HiOpFlags |= PPCII::MO_NLP_FLAG;
1253 LoOpFlags |= PPCII::MO_NLP_FLAG;
1255 if (GV->hasHiddenVisibility()) {
1256 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1257 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1264 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1265 SelectionDAG &DAG) {
1266 EVT PtrVT = HiPart.getValueType();
1267 SDValue Zero = DAG.getConstant(0, PtrVT);
1268 DebugLoc DL = HiPart.getDebugLoc();
1270 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1271 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1273 // With PIC, the first instruction is actually "GR+hi(&G)".
1275 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1276 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1278 // Generate non-pic code that has direct accesses to the constant pool.
1279 // The address of the global is just (hi(&g)+lo(&g)).
1280 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1283 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1284 SelectionDAG &DAG) const {
1285 EVT PtrVT = Op.getValueType();
1286 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1287 const Constant *C = CP->getConstVal();
1289 // 64-bit SVR4 ABI code is always position-independent.
1290 // The actual address of the GlobalValue is stored in the TOC.
1291 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1292 SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0);
1293 return DAG.getNode(PPCISD::TOC_ENTRY, CP->getDebugLoc(), MVT::i64, GA,
1294 DAG.getRegister(PPC::X2, MVT::i64));
1297 unsigned MOHiFlag, MOLoFlag;
1298 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1300 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1302 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1303 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1306 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1307 EVT PtrVT = Op.getValueType();
1308 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1310 // 64-bit SVR4 ABI code is always position-independent.
1311 // The actual address of the GlobalValue is stored in the TOC.
1312 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1313 SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1314 return DAG.getNode(PPCISD::TOC_ENTRY, JT->getDebugLoc(), MVT::i64, GA,
1315 DAG.getRegister(PPC::X2, MVT::i64));
1318 unsigned MOHiFlag, MOLoFlag;
1319 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1320 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1321 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1322 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1325 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1326 SelectionDAG &DAG) const {
1327 EVT PtrVT = Op.getValueType();
1329 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1331 unsigned MOHiFlag, MOLoFlag;
1332 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1333 SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag);
1334 SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag);
1335 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1338 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1339 SelectionDAG &DAG) const {
1341 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1342 DebugLoc dl = GA->getDebugLoc();
1343 const GlobalValue *GV = GA->getGlobal();
1344 EVT PtrVT = getPointerTy();
1345 bool is64bit = PPCSubTarget.isPPC64();
1347 TLSModel::Model Model = getTargetMachine().getTLSModel(GV);
1349 if (Model == TLSModel::LocalExec) {
1350 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1351 PPCII::MO_TPREL16_HA);
1352 SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1353 PPCII::MO_TPREL16_LO);
1354 SDValue TLSReg = DAG.getRegister(is64bit ? PPC::X13 : PPC::R2,
1355 is64bit ? MVT::i64 : MVT::i32);
1356 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg);
1357 return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi);
1361 llvm_unreachable("only local-exec is currently supported for ppc32");
1363 if (Model == TLSModel::InitialExec) {
1364 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1365 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1366 SDValue TPOffsetHi = DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl,
1367 PtrVT, GOTReg, TGA);
1368 SDValue TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl,
1369 PtrVT, TGA, TPOffsetHi);
1370 return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGA);
1373 if (Model == TLSModel::GeneralDynamic) {
1374 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1375 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1376 SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
1378 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSGD_L, dl, PtrVT,
1381 // We need a chain node, and don't have one handy. The underlying
1382 // call has no side effects, so using the function entry node
1384 SDValue Chain = DAG.getEntryNode();
1385 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
1386 SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
1387 SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLS_ADDR, dl,
1388 PtrVT, ParmReg, TGA);
1389 // The return value from GET_TLS_ADDR really is in X3 already, but
1390 // some hacks are needed here to tie everything together. The extra
1391 // copies dissolve during subsequent transforms.
1392 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
1393 return DAG.getCopyFromReg(Chain, dl, PPC::X3, PtrVT);
1396 if (Model == TLSModel::LocalDynamic) {
1397 SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
1398 SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
1399 SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
1401 SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSLD_L, dl, PtrVT,
1404 // We need a chain node, and don't have one handy. The underlying
1405 // call has no side effects, so using the function entry node
1407 SDValue Chain = DAG.getEntryNode();
1408 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
1409 SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
1410 SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLSLD_ADDR, dl,
1411 PtrVT, ParmReg, TGA);
1412 // The return value from GET_TLSLD_ADDR really is in X3 already, but
1413 // some hacks are needed here to tie everything together. The extra
1414 // copies dissolve during subsequent transforms.
1415 Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
1416 SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, PtrVT,
1417 Chain, ParmReg, TGA);
1418 return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
1421 llvm_unreachable("Unknown TLS model!");
1424 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1425 SelectionDAG &DAG) const {
1426 EVT PtrVT = Op.getValueType();
1427 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1428 DebugLoc DL = GSDN->getDebugLoc();
1429 const GlobalValue *GV = GSDN->getGlobal();
1431 // 64-bit SVR4 ABI code is always position-independent.
1432 // The actual address of the GlobalValue is stored in the TOC.
1433 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1434 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1435 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1436 DAG.getRegister(PPC::X2, MVT::i64));
1439 unsigned MOHiFlag, MOLoFlag;
1440 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1443 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1445 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1447 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1449 // If the global reference is actually to a non-lazy-pointer, we have to do an
1450 // extra load to get the address of the global.
1451 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1452 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1453 false, false, false, 0);
1457 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1458 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1459 DebugLoc dl = Op.getDebugLoc();
1461 // If we're comparing for equality to zero, expose the fact that this is
1462 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1463 // fold the new nodes.
1464 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1465 if (C->isNullValue() && CC == ISD::SETEQ) {
1466 EVT VT = Op.getOperand(0).getValueType();
1467 SDValue Zext = Op.getOperand(0);
1468 if (VT.bitsLT(MVT::i32)) {
1470 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1472 unsigned Log2b = Log2_32(VT.getSizeInBits());
1473 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1474 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1475 DAG.getConstant(Log2b, MVT::i32));
1476 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1478 // Leave comparisons against 0 and -1 alone for now, since they're usually
1479 // optimized. FIXME: revisit this when we can custom lower all setcc
1481 if (C->isAllOnesValue() || C->isNullValue())
1485 // If we have an integer seteq/setne, turn it into a compare against zero
1486 // by xor'ing the rhs with the lhs, which is faster than setting a
1487 // condition register, reading it back out, and masking the correct bit. The
1488 // normal approach here uses sub to do this instead of xor. Using xor exposes
1489 // the result to other bit-twiddling opportunities.
1490 EVT LHSVT = Op.getOperand(0).getValueType();
1491 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1492 EVT VT = Op.getValueType();
1493 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1495 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1500 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1501 const PPCSubtarget &Subtarget) const {
1502 SDNode *Node = Op.getNode();
1503 EVT VT = Node->getValueType(0);
1504 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1505 SDValue InChain = Node->getOperand(0);
1506 SDValue VAListPtr = Node->getOperand(1);
1507 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1508 DebugLoc dl = Node->getDebugLoc();
1510 assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only");
1513 SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1514 VAListPtr, MachinePointerInfo(SV), MVT::i8,
1516 InChain = GprIndex.getValue(1);
1518 if (VT == MVT::i64) {
1519 // Check if GprIndex is even
1520 SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex,
1521 DAG.getConstant(1, MVT::i32));
1522 SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd,
1523 DAG.getConstant(0, MVT::i32), ISD::SETNE);
1524 SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex,
1525 DAG.getConstant(1, MVT::i32));
1526 // Align GprIndex to be even if it isn't
1527 GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne,
1531 // fpr index is 1 byte after gpr
1532 SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1533 DAG.getConstant(1, MVT::i32));
1536 SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
1537 FprPtr, MachinePointerInfo(SV), MVT::i8,
1539 InChain = FprIndex.getValue(1);
1541 SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1542 DAG.getConstant(8, MVT::i32));
1544 SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
1545 DAG.getConstant(4, MVT::i32));
1548 SDValue OverflowArea = DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr,
1549 MachinePointerInfo(), false, false,
1551 InChain = OverflowArea.getValue(1);
1553 SDValue RegSaveArea = DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr,
1554 MachinePointerInfo(), false, false,
1556 InChain = RegSaveArea.getValue(1);
1558 // select overflow_area if index > 8
1559 SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex,
1560 DAG.getConstant(8, MVT::i32), ISD::SETLT);
1562 // adjustment constant gpr_index * 4/8
1563 SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32,
1564 VT.isInteger() ? GprIndex : FprIndex,
1565 DAG.getConstant(VT.isInteger() ? 4 : 8,
1568 // OurReg = RegSaveArea + RegConstant
1569 SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea,
1572 // Floating types are 32 bytes into RegSaveArea
1573 if (VT.isFloatingPoint())
1574 OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg,
1575 DAG.getConstant(32, MVT::i32));
1577 // increase {f,g}pr_index by 1 (or 2 if VT is i64)
1578 SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32,
1579 VT.isInteger() ? GprIndex : FprIndex,
1580 DAG.getConstant(VT == MVT::i64 ? 2 : 1,
1583 InChain = DAG.getTruncStore(InChain, dl, IndexPlus1,
1584 VT.isInteger() ? VAListPtr : FprPtr,
1585 MachinePointerInfo(SV),
1586 MVT::i8, false, false, 0);
1588 // determine if we should load from reg_save_area or overflow_area
1589 SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea);
1591 // increase overflow_area by 4/8 if gpr/fpr > 8
1592 SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea,
1593 DAG.getConstant(VT.isInteger() ? 4 : 8,
1596 OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea,
1599 InChain = DAG.getTruncStore(InChain, dl, OverflowArea,
1601 MachinePointerInfo(),
1602 MVT::i32, false, false, 0);
1604 return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo(),
1605 false, false, false, 0);
1608 SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
1609 SelectionDAG &DAG) const {
1610 return Op.getOperand(0);
1613 SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
1614 SelectionDAG &DAG) const {
1615 SDValue Chain = Op.getOperand(0);
1616 SDValue Trmp = Op.getOperand(1); // trampoline
1617 SDValue FPtr = Op.getOperand(2); // nested function
1618 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1619 DebugLoc dl = Op.getDebugLoc();
1621 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1622 bool isPPC64 = (PtrVT == MVT::i64);
1624 DAG.getTargetLoweringInfo().getDataLayout()->getIntPtrType(
1627 TargetLowering::ArgListTy Args;
1628 TargetLowering::ArgListEntry Entry;
1630 Entry.Ty = IntPtrTy;
1631 Entry.Node = Trmp; Args.push_back(Entry);
1633 // TrampSize == (isPPC64 ? 48 : 40);
1634 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1635 isPPC64 ? MVT::i64 : MVT::i32);
1636 Args.push_back(Entry);
1638 Entry.Node = FPtr; Args.push_back(Entry);
1639 Entry.Node = Nest; Args.push_back(Entry);
1641 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1642 TargetLowering::CallLoweringInfo CLI(Chain,
1643 Type::getVoidTy(*DAG.getContext()),
1644 false, false, false, false, 0,
1646 /*isTailCall=*/false,
1647 /*doesNotRet=*/false,
1648 /*isReturnValueUsed=*/true,
1649 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1651 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1653 return CallResult.second;
1656 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1657 const PPCSubtarget &Subtarget) const {
1658 MachineFunction &MF = DAG.getMachineFunction();
1659 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1661 DebugLoc dl = Op.getDebugLoc();
1663 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1664 // vastart just stores the address of the VarArgsFrameIndex slot into the
1665 // memory location argument.
1666 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1667 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1668 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1669 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1670 MachinePointerInfo(SV),
1674 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1675 // We suppose the given va_list is already allocated.
1678 // char gpr; /* index into the array of 8 GPRs
1679 // * stored in the register save area
1680 // * gpr=0 corresponds to r3,
1681 // * gpr=1 to r4, etc.
1683 // char fpr; /* index into the array of 8 FPRs
1684 // * stored in the register save area
1685 // * fpr=0 corresponds to f1,
1686 // * fpr=1 to f2, etc.
1688 // char *overflow_arg_area;
1689 // /* location on stack that holds
1690 // * the next overflow argument
1692 // char *reg_save_area;
1693 // /* where r3:r10 and f1:f8 (if saved)
1699 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1700 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1703 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1705 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1707 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1710 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1711 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1713 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1714 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1716 uint64_t FPROffset = 1;
1717 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1719 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1721 // Store first byte : number of int regs
1722 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1724 MachinePointerInfo(SV),
1725 MVT::i8, false, false, 0);
1726 uint64_t nextOffset = FPROffset;
1727 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1730 // Store second byte : number of float regs
1731 SDValue secondStore =
1732 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1733 MachinePointerInfo(SV, nextOffset), MVT::i8,
1735 nextOffset += StackOffset;
1736 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1738 // Store second word : arguments given on stack
1739 SDValue thirdStore =
1740 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1741 MachinePointerInfo(SV, nextOffset),
1743 nextOffset += FrameOffset;
1744 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1746 // Store third word : arguments given in registers
1747 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1748 MachinePointerInfo(SV, nextOffset),
1753 #include "PPCGenCallingConv.inc"
1755 static bool CC_PPC32_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1756 CCValAssign::LocInfo &LocInfo,
1757 ISD::ArgFlagsTy &ArgFlags,
1762 static bool CC_PPC32_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1764 CCValAssign::LocInfo &LocInfo,
1765 ISD::ArgFlagsTy &ArgFlags,
1767 static const uint16_t ArgRegs[] = {
1768 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1769 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1771 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1773 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1775 // Skip one register if the first unallocated register has an even register
1776 // number and there are still argument registers available which have not been
1777 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1778 // need to skip a register if RegNum is odd.
1779 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1780 State.AllocateReg(ArgRegs[RegNum]);
1783 // Always return false here, as this function only makes sure that the first
1784 // unallocated register has an odd register number and does not actually
1785 // allocate a register for the current argument.
1789 static bool CC_PPC32_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1791 CCValAssign::LocInfo &LocInfo,
1792 ISD::ArgFlagsTy &ArgFlags,
1794 static const uint16_t ArgRegs[] = {
1795 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1799 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1801 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1803 // If there is only one Floating-point register left we need to put both f64
1804 // values of a split ppc_fp128 value on the stack.
1805 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1806 State.AllocateReg(ArgRegs[RegNum]);
1809 // Always return false here, as this function only makes sure that the two f64
1810 // values a ppc_fp128 value is split into are both passed in registers or both
1811 // passed on the stack and does not actually allocate a register for the
1812 // current argument.
1816 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1818 static const uint16_t *GetFPR() {
1819 static const uint16_t FPR[] = {
1820 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1821 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1827 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1829 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1830 unsigned PtrByteSize) {
1831 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1832 if (Flags.isByVal())
1833 ArgSize = Flags.getByValSize();
1834 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1840 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1841 CallingConv::ID CallConv, bool isVarArg,
1842 const SmallVectorImpl<ISD::InputArg>
1844 DebugLoc dl, SelectionDAG &DAG,
1845 SmallVectorImpl<SDValue> &InVals)
1847 if (PPCSubTarget.isSVR4ABI()) {
1848 if (PPCSubTarget.isPPC64())
1849 return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins,
1852 return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins,
1855 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1861 PPCTargetLowering::LowerFormalArguments_32SVR4(
1863 CallingConv::ID CallConv, bool isVarArg,
1864 const SmallVectorImpl<ISD::InputArg>
1866 DebugLoc dl, SelectionDAG &DAG,
1867 SmallVectorImpl<SDValue> &InVals) const {
1869 // 32-bit SVR4 ABI Stack Frame Layout:
1870 // +-----------------------------------+
1871 // +--> | Back chain |
1872 // | +-----------------------------------+
1873 // | | Floating-point register save area |
1874 // | +-----------------------------------+
1875 // | | General register save area |
1876 // | +-----------------------------------+
1877 // | | CR save word |
1878 // | +-----------------------------------+
1879 // | | VRSAVE save word |
1880 // | +-----------------------------------+
1881 // | | Alignment padding |
1882 // | +-----------------------------------+
1883 // | | Vector register save area |
1884 // | +-----------------------------------+
1885 // | | Local variable space |
1886 // | +-----------------------------------+
1887 // | | Parameter list area |
1888 // | +-----------------------------------+
1889 // | | LR save word |
1890 // | +-----------------------------------+
1891 // SP--> +--- | Back chain |
1892 // +-----------------------------------+
1895 // System V Application Binary Interface PowerPC Processor Supplement
1896 // AltiVec Technology Programming Interface Manual
1898 MachineFunction &MF = DAG.getMachineFunction();
1899 MachineFrameInfo *MFI = MF.getFrameInfo();
1900 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1902 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1903 // Potential tail calls could cause overwriting of argument stack slots.
1904 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
1905 (CallConv == CallingConv::Fast));
1906 unsigned PtrByteSize = 4;
1908 // Assign locations to all of the incoming arguments.
1909 SmallVector<CCValAssign, 16> ArgLocs;
1910 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1911 getTargetMachine(), ArgLocs, *DAG.getContext());
1913 // Reserve space for the linkage area on the stack.
1914 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1916 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4);
1918 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1919 CCValAssign &VA = ArgLocs[i];
1921 // Arguments stored in registers.
1922 if (VA.isRegLoc()) {
1923 const TargetRegisterClass *RC;
1924 EVT ValVT = VA.getValVT();
1926 switch (ValVT.getSimpleVT().SimpleTy) {
1928 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1930 RC = &PPC::GPRCRegClass;
1933 RC = &PPC::F4RCRegClass;
1936 RC = &PPC::F8RCRegClass;
1942 RC = &PPC::VRRCRegClass;
1946 // Transform the arguments stored in physical registers into virtual ones.
1947 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1948 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1950 InVals.push_back(ArgValue);
1952 // Argument stored in memory.
1953 assert(VA.isMemLoc());
1955 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1956 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1959 // Create load nodes to retrieve arguments from the stack.
1960 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1961 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1962 MachinePointerInfo(),
1963 false, false, false, 0));
1967 // Assign locations to all of the incoming aggregate by value arguments.
1968 // Aggregates passed by value are stored in the local variable space of the
1969 // caller's stack frame, right above the parameter list area.
1970 SmallVector<CCValAssign, 16> ByValArgLocs;
1971 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1972 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1974 // Reserve stack space for the allocations in CCInfo.
1975 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1977 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal);
1979 // Area that is at least reserved in the caller of this function.
1980 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1982 // Set the size that is at least reserved in caller of this function. Tail
1983 // call optimized function's reserved stack space needs to be aligned so that
1984 // taking the difference between two stack areas will result in an aligned
1986 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1989 std::max(MinReservedArea,
1990 PPCFrameLowering::getMinCallFrameSize(false, false));
1992 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1993 getStackAlignment();
1994 unsigned AlignMask = TargetAlign-1;
1995 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1997 FI->setMinReservedArea(MinReservedArea);
1999 SmallVector<SDValue, 8> MemOps;
2001 // If the function takes variable number of arguments, make a frame index for
2002 // the start of the first vararg value... for expansion of llvm.va_start.
2004 static const uint16_t GPArgRegs[] = {
2005 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2006 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2008 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
2010 static const uint16_t FPArgRegs[] = {
2011 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
2014 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
2016 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
2018 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
2021 // Make room for NumGPArgRegs and NumFPArgRegs.
2022 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
2023 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
2025 FuncInfo->setVarArgsStackOffset(
2026 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2027 CCInfo.getNextStackOffset(), true));
2029 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
2030 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2032 // The fixed integer arguments of a variadic function are stored to the
2033 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
2034 // the result of va_next.
2035 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
2036 // Get an existing live-in vreg, or add a new one.
2037 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
2039 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
2041 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2042 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2043 MachinePointerInfo(), false, false, 0);
2044 MemOps.push_back(Store);
2045 // Increment the address by four for the next argument to store
2046 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2047 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2050 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
2052 // The double arguments are stored to the VarArgsFrameIndex
2054 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
2055 // Get an existing live-in vreg, or add a new one.
2056 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
2058 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
2060 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
2061 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2062 MachinePointerInfo(), false, false, 0);
2063 MemOps.push_back(Store);
2064 // Increment the address by eight for the next argument to store
2065 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
2067 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2071 if (!MemOps.empty())
2072 Chain = DAG.getNode(ISD::TokenFactor, dl,
2073 MVT::Other, &MemOps[0], MemOps.size());
2078 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2079 // value to MVT::i64 and then truncate to the correct register size.
2081 PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, EVT ObjectVT,
2082 SelectionDAG &DAG, SDValue ArgVal,
2083 DebugLoc dl) const {
2085 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
2086 DAG.getValueType(ObjectVT));
2087 else if (Flags.isZExt())
2088 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
2089 DAG.getValueType(ObjectVT));
2091 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
2094 // Set the size that is at least reserved in caller of this function. Tail
2095 // call optimized functions' reserved stack space needs to be aligned so that
2096 // taking the difference between two stack areas will result in an aligned
2099 PPCTargetLowering::setMinReservedArea(MachineFunction &MF, SelectionDAG &DAG,
2100 unsigned nAltivecParamsAtEnd,
2101 unsigned MinReservedArea,
2102 bool isPPC64) const {
2103 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2104 // Add the Altivec parameters at the end, if needed.
2105 if (nAltivecParamsAtEnd) {
2106 MinReservedArea = ((MinReservedArea+15)/16)*16;
2107 MinReservedArea += 16*nAltivecParamsAtEnd;
2110 std::max(MinReservedArea,
2111 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2112 unsigned TargetAlign
2113 = DAG.getMachineFunction().getTarget().getFrameLowering()->
2114 getStackAlignment();
2115 unsigned AlignMask = TargetAlign-1;
2116 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2117 FI->setMinReservedArea(MinReservedArea);
2121 PPCTargetLowering::LowerFormalArguments_64SVR4(
2123 CallingConv::ID CallConv, bool isVarArg,
2124 const SmallVectorImpl<ISD::InputArg>
2126 DebugLoc dl, SelectionDAG &DAG,
2127 SmallVectorImpl<SDValue> &InVals) const {
2128 // TODO: add description of PPC stack frame format, or at least some docs.
2130 MachineFunction &MF = DAG.getMachineFunction();
2131 MachineFrameInfo *MFI = MF.getFrameInfo();
2132 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2134 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2135 // Potential tail calls could cause overwriting of argument stack slots.
2136 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2137 (CallConv == CallingConv::Fast));
2138 unsigned PtrByteSize = 8;
2140 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
2141 // Area that is at least reserved in caller of this function.
2142 unsigned MinReservedArea = ArgOffset;
2144 static const uint16_t GPR[] = {
2145 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2146 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2149 static const uint16_t *FPR = GetFPR();
2151 static const uint16_t VR[] = {
2152 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2153 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2156 const unsigned Num_GPR_Regs = array_lengthof(GPR);
2157 const unsigned Num_FPR_Regs = 13;
2158 const unsigned Num_VR_Regs = array_lengthof(VR);
2160 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2162 // Add DAG nodes to load the arguments or copy them out of registers. On
2163 // entry to a function on PPC, the arguments start after the linkage area,
2164 // although the first ones are often in registers.
2166 SmallVector<SDValue, 8> MemOps;
2167 unsigned nAltivecParamsAtEnd = 0;
2168 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2169 unsigned CurArgIdx = 0;
2170 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
2172 bool needsLoad = false;
2173 EVT ObjectVT = Ins[ArgNo].VT;
2174 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2175 unsigned ArgSize = ObjSize;
2176 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2177 std::advance(FuncArg, Ins[ArgNo].OrigArgIndex - CurArgIdx);
2178 CurArgIdx = Ins[ArgNo].OrigArgIndex;
2180 unsigned CurArgOffset = ArgOffset;
2182 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2183 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2184 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2186 MinReservedArea = ((MinReservedArea+15)/16)*16;
2187 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2191 nAltivecParamsAtEnd++;
2193 // Calculate min reserved area.
2194 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2198 // FIXME the codegen can be much improved in some cases.
2199 // We do not have to keep everything in memory.
2200 if (Flags.isByVal()) {
2201 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2202 ObjSize = Flags.getByValSize();
2203 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2204 // Empty aggregate parameters do not take up registers. Examples:
2208 // etc. However, we have to provide a place-holder in InVals, so
2209 // pretend we have an 8-byte item at the current address for that
2212 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2213 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2214 InVals.push_back(FIN);
2217 // All aggregates smaller than 8 bytes must be passed right-justified.
2218 if (ObjSize < PtrByteSize)
2219 CurArgOffset = CurArgOffset + (PtrByteSize - ObjSize);
2220 // The value of the object is its address.
2221 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2222 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2223 InVals.push_back(FIN);
2226 if (GPR_idx != Num_GPR_Regs) {
2227 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2228 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2231 if (ObjSize==1 || ObjSize==2 || ObjSize==4) {
2232 EVT ObjType = (ObjSize == 1 ? MVT::i8 :
2233 (ObjSize == 2 ? MVT::i16 : MVT::i32));
2234 Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2235 MachinePointerInfo(FuncArg, CurArgOffset),
2236 ObjType, false, false, 0);
2238 // For sizes that don't fit a truncating store (3, 5, 6, 7),
2239 // store the whole register as-is to the parameter save area
2240 // slot. The address of the parameter was already calculated
2241 // above (InVals.push_back(FIN)) to be the right-justified
2242 // offset within the slot. For this store, we need a new
2243 // frame index that points at the beginning of the slot.
2244 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2245 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2246 Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2247 MachinePointerInfo(FuncArg, ArgOffset),
2251 MemOps.push_back(Store);
2254 // Whether we copied from a register or not, advance the offset
2255 // into the parameter save area by a full doubleword.
2256 ArgOffset += PtrByteSize;
2260 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2261 // Store whatever pieces of the object are in registers
2262 // to memory. ArgOffset will be the address of the beginning
2264 if (GPR_idx != Num_GPR_Regs) {
2266 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2267 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2268 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2269 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2270 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2271 MachinePointerInfo(FuncArg, ArgOffset),
2273 MemOps.push_back(Store);
2275 ArgOffset += PtrByteSize;
2277 ArgOffset += ArgSize - j;
2284 switch (ObjectVT.getSimpleVT().SimpleTy) {
2285 default: llvm_unreachable("Unhandled argument type!");
2288 if (GPR_idx != Num_GPR_Regs) {
2289 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2290 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2292 if (ObjectVT == MVT::i32)
2293 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2294 // value to MVT::i64 and then truncate to the correct register size.
2295 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2300 ArgSize = PtrByteSize;
2307 // Every 8 bytes of argument space consumes one of the GPRs available for
2308 // argument passing.
2309 if (GPR_idx != Num_GPR_Regs) {
2312 if (FPR_idx != Num_FPR_Regs) {
2315 if (ObjectVT == MVT::f32)
2316 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2318 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2320 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2324 ArgSize = PtrByteSize;
2333 // Note that vector arguments in registers don't reserve stack space,
2334 // except in varargs functions.
2335 if (VR_idx != Num_VR_Regs) {
2336 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2337 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2339 while ((ArgOffset % 16) != 0) {
2340 ArgOffset += PtrByteSize;
2341 if (GPR_idx != Num_GPR_Regs)
2345 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2349 // Vectors are aligned.
2350 ArgOffset = ((ArgOffset+15)/16)*16;
2351 CurArgOffset = ArgOffset;
2358 // We need to load the argument to a virtual register if we determined
2359 // above that we ran out of physical registers of the appropriate type.
2361 int FI = MFI->CreateFixedObject(ObjSize,
2362 CurArgOffset + (ArgSize - ObjSize),
2364 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2365 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2366 false, false, false, 0);
2369 InVals.push_back(ArgVal);
2372 // Set the size that is at least reserved in caller of this function. Tail
2373 // call optimized functions' reserved stack space needs to be aligned so that
2374 // taking the difference between two stack areas will result in an aligned
2376 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, true);
2378 // If the function takes variable number of arguments, make a frame index for
2379 // the start of the first vararg value... for expansion of llvm.va_start.
2381 int Depth = ArgOffset;
2383 FuncInfo->setVarArgsFrameIndex(
2384 MFI->CreateFixedObject(PtrByteSize, Depth, true));
2385 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2387 // If this function is vararg, store any remaining integer argument regs
2388 // to their spots on the stack so that they may be loaded by deferencing the
2389 // result of va_next.
2390 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2391 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2392 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2393 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2394 MachinePointerInfo(), false, false, 0);
2395 MemOps.push_back(Store);
2396 // Increment the address by four for the next argument to store
2397 SDValue PtrOff = DAG.getConstant(PtrByteSize, PtrVT);
2398 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2402 if (!MemOps.empty())
2403 Chain = DAG.getNode(ISD::TokenFactor, dl,
2404 MVT::Other, &MemOps[0], MemOps.size());
2410 PPCTargetLowering::LowerFormalArguments_Darwin(
2412 CallingConv::ID CallConv, bool isVarArg,
2413 const SmallVectorImpl<ISD::InputArg>
2415 DebugLoc dl, SelectionDAG &DAG,
2416 SmallVectorImpl<SDValue> &InVals) const {
2417 // TODO: add description of PPC stack frame format, or at least some docs.
2419 MachineFunction &MF = DAG.getMachineFunction();
2420 MachineFrameInfo *MFI = MF.getFrameInfo();
2421 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
2423 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2424 bool isPPC64 = PtrVT == MVT::i64;
2425 // Potential tail calls could cause overwriting of argument stack slots.
2426 bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt &&
2427 (CallConv == CallingConv::Fast));
2428 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2430 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2431 // Area that is at least reserved in caller of this function.
2432 unsigned MinReservedArea = ArgOffset;
2434 static const uint16_t GPR_32[] = { // 32-bit registers.
2435 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2436 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2438 static const uint16_t GPR_64[] = { // 64-bit registers.
2439 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2440 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2443 static const uint16_t *FPR = GetFPR();
2445 static const uint16_t VR[] = {
2446 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2447 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2450 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
2451 const unsigned Num_FPR_Regs = 13;
2452 const unsigned Num_VR_Regs = array_lengthof( VR);
2454 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2456 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
2458 // In 32-bit non-varargs functions, the stack space for vectors is after the
2459 // stack space for non-vectors. We do not use this space unless we have
2460 // too many vectors to fit in registers, something that only occurs in
2461 // constructed examples:), but we have to walk the arglist to figure
2462 // that out...for the pathological case, compute VecArgOffset as the
2463 // start of the vector parameter area. Computing VecArgOffset is the
2464 // entire point of the following loop.
2465 unsigned VecArgOffset = ArgOffset;
2466 if (!isVarArg && !isPPC64) {
2467 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
2469 EVT ObjectVT = Ins[ArgNo].VT;
2470 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2472 if (Flags.isByVal()) {
2473 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
2474 unsigned ObjSize = Flags.getByValSize();
2476 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2477 VecArgOffset += ArgSize;
2481 switch(ObjectVT.getSimpleVT().SimpleTy) {
2482 default: llvm_unreachable("Unhandled argument type!");
2487 case MVT::i64: // PPC64
2489 // FIXME: We are guaranteed to be !isPPC64 at this point.
2490 // Does MVT::i64 apply?
2497 // Nothing to do, we're only looking at Nonvector args here.
2502 // We've found where the vector parameter area in memory is. Skip the
2503 // first 12 parameters; these don't use that memory.
2504 VecArgOffset = ((VecArgOffset+15)/16)*16;
2505 VecArgOffset += 12*16;
2507 // Add DAG nodes to load the arguments or copy them out of registers. On
2508 // entry to a function on PPC, the arguments start after the linkage area,
2509 // although the first ones are often in registers.
2511 SmallVector<SDValue, 8> MemOps;
2512 unsigned nAltivecParamsAtEnd = 0;
2513 // FIXME: FuncArg and Ins[ArgNo] must reference the same argument.
2514 // When passing anonymous aggregates, this is currently not true.
2515 // See LowerFormalArguments_64SVR4 for a fix.
2516 Function::const_arg_iterator FuncArg = MF.getFunction()->arg_begin();
2517 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {
2519 bool needsLoad = false;
2520 EVT ObjectVT = Ins[ArgNo].VT;
2521 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
2522 unsigned ArgSize = ObjSize;
2523 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
2525 unsigned CurArgOffset = ArgOffset;
2527 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
2528 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
2529 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
2530 if (isVarArg || isPPC64) {
2531 MinReservedArea = ((MinReservedArea+15)/16)*16;
2532 MinReservedArea += CalculateStackSlotSize(ObjectVT,
2535 } else nAltivecParamsAtEnd++;
2537 // Calculate min reserved area.
2538 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
2542 // FIXME the codegen can be much improved in some cases.
2543 // We do not have to keep everything in memory.
2544 if (Flags.isByVal()) {
2545 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
2546 ObjSize = Flags.getByValSize();
2547 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
2548 // Objects of size 1 and 2 are right justified, everything else is
2549 // left justified. This means the memory address is adjusted forwards.
2550 if (ObjSize==1 || ObjSize==2) {
2551 CurArgOffset = CurArgOffset + (4 - ObjSize);
2553 // The value of the object is its address.
2554 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
2555 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2556 InVals.push_back(FIN);
2557 if (ObjSize==1 || ObjSize==2) {
2558 if (GPR_idx != Num_GPR_Regs) {
2561 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2563 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2564 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2565 EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16;
2566 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
2567 MachinePointerInfo(FuncArg,
2569 ObjType, false, false, 0);
2570 MemOps.push_back(Store);
2574 ArgOffset += PtrByteSize;
2578 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
2579 // Store whatever pieces of the object are in registers
2580 // to memory. ArgOffset will be the address of the beginning
2582 if (GPR_idx != Num_GPR_Regs) {
2585 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2587 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2588 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
2589 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2590 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2591 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2592 MachinePointerInfo(FuncArg, ArgOffset),
2594 MemOps.push_back(Store);
2596 ArgOffset += PtrByteSize;
2598 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
2605 switch (ObjectVT.getSimpleVT().SimpleTy) {
2606 default: llvm_unreachable("Unhandled argument type!");
2609 if (GPR_idx != Num_GPR_Regs) {
2610 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2611 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2615 ArgSize = PtrByteSize;
2617 // All int arguments reserve stack space in the Darwin ABI.
2618 ArgOffset += PtrByteSize;
2622 case MVT::i64: // PPC64
2623 if (GPR_idx != Num_GPR_Regs) {
2624 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2625 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
2627 if (ObjectVT == MVT::i32)
2628 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
2629 // value to MVT::i64 and then truncate to the correct register size.
2630 ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl);
2635 ArgSize = PtrByteSize;
2637 // All int arguments reserve stack space in the Darwin ABI.
2643 // Every 4 bytes of argument space consumes one of the GPRs available for
2644 // argument passing.
2645 if (GPR_idx != Num_GPR_Regs) {
2647 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
2650 if (FPR_idx != Num_FPR_Regs) {
2653 if (ObjectVT == MVT::f32)
2654 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
2656 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2658 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2664 // All FP arguments reserve stack space in the Darwin ABI.
2665 ArgOffset += isPPC64 ? 8 : ObjSize;
2671 // Note that vector arguments in registers don't reserve stack space,
2672 // except in varargs functions.
2673 if (VR_idx != Num_VR_Regs) {
2674 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2675 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2677 while ((ArgOffset % 16) != 0) {
2678 ArgOffset += PtrByteSize;
2679 if (GPR_idx != Num_GPR_Regs)
2683 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2687 if (!isVarArg && !isPPC64) {
2688 // Vectors go after all the nonvectors.
2689 CurArgOffset = VecArgOffset;
2692 // Vectors are aligned.
2693 ArgOffset = ((ArgOffset+15)/16)*16;
2694 CurArgOffset = ArgOffset;
2702 // We need to load the argument to a virtual register if we determined above
2703 // that we ran out of physical registers of the appropriate type.
2705 int FI = MFI->CreateFixedObject(ObjSize,
2706 CurArgOffset + (ArgSize - ObjSize),
2708 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2709 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2710 false, false, false, 0);
2713 InVals.push_back(ArgVal);
2716 // Set the size that is at least reserved in caller of this function. Tail
2717 // call optimized functions' reserved stack space needs to be aligned so that
2718 // taking the difference between two stack areas will result in an aligned
2720 setMinReservedArea(MF, DAG, nAltivecParamsAtEnd, MinReservedArea, isPPC64);
2722 // If the function takes variable number of arguments, make a frame index for
2723 // the start of the first vararg value... for expansion of llvm.va_start.
2725 int Depth = ArgOffset;
2727 FuncInfo->setVarArgsFrameIndex(
2728 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2730 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2732 // If this function is vararg, store any remaining integer argument regs
2733 // to their spots on the stack so that they may be loaded by deferencing the
2734 // result of va_next.
2735 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2739 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2741 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2743 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2744 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2745 MachinePointerInfo(), false, false, 0);
2746 MemOps.push_back(Store);
2747 // Increment the address by four for the next argument to store
2748 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2749 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2753 if (!MemOps.empty())
2754 Chain = DAG.getNode(ISD::TokenFactor, dl,
2755 MVT::Other, &MemOps[0], MemOps.size());
2760 /// CalculateParameterAndLinkageAreaSize - Get the size of the parameter plus
2761 /// linkage area for the Darwin ABI, or the 64-bit SVR4 ABI.
2763 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2767 const SmallVectorImpl<ISD::OutputArg>
2769 const SmallVectorImpl<SDValue> &OutVals,
2770 unsigned &nAltivecParamsAtEnd) {
2771 // Count how many bytes are to be pushed on the stack, including the linkage
2772 // area, and parameter passing area. We start with 24/48 bytes, which is
2773 // prereserved space for [SP][CR][LR][3 x unused].
2774 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2775 unsigned NumOps = Outs.size();
2776 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2778 // Add up all the space actually used.
2779 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2780 // they all go in registers, but we must reserve stack space for them for
2781 // possible use by the caller. In varargs or 64-bit calls, parameters are
2782 // assigned stack space in order, with padding so Altivec parameters are
2784 nAltivecParamsAtEnd = 0;
2785 for (unsigned i = 0; i != NumOps; ++i) {
2786 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2787 EVT ArgVT = Outs[i].VT;
2788 // Varargs Altivec parameters are padded to a 16 byte boundary.
2789 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2790 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2791 if (!isVarArg && !isPPC64) {
2792 // Non-varargs Altivec parameters go after all the non-Altivec
2793 // parameters; handle those later so we know how much padding we need.
2794 nAltivecParamsAtEnd++;
2797 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2798 NumBytes = ((NumBytes+15)/16)*16;
2800 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2803 // Allow for Altivec parameters at the end, if needed.
2804 if (nAltivecParamsAtEnd) {
2805 NumBytes = ((NumBytes+15)/16)*16;
2806 NumBytes += 16*nAltivecParamsAtEnd;
2809 // The prolog code of the callee may store up to 8 GPR argument registers to
2810 // the stack, allowing va_start to index over them in memory if its varargs.
2811 // Because we cannot tell if this is needed on the caller side, we have to
2812 // conservatively assume that it is needed. As such, make sure we have at
2813 // least enough stack space for the caller to store the 8 GPRs.
2814 NumBytes = std::max(NumBytes,
2815 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2817 // Tail call needs the stack to be aligned.
2818 if (CC == CallingConv::Fast && DAG.getTarget().Options.GuaranteedTailCallOpt){
2819 unsigned TargetAlign = DAG.getMachineFunction().getTarget().
2820 getFrameLowering()->getStackAlignment();
2821 unsigned AlignMask = TargetAlign-1;
2822 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2828 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2829 /// adjusted to accommodate the arguments for the tailcall.
2830 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2831 unsigned ParamSize) {
2833 if (!isTailCall) return 0;
2835 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2836 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2837 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2838 // Remember only if the new adjustement is bigger.
2839 if (SPDiff < FI->getTailCallSPDelta())
2840 FI->setTailCallSPDelta(SPDiff);
2845 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2846 /// for tail call optimization. Targets which want to do tail call
2847 /// optimization should implement this function.
2849 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2850 CallingConv::ID CalleeCC,
2852 const SmallVectorImpl<ISD::InputArg> &Ins,
2853 SelectionDAG& DAG) const {
2854 if (!getTargetMachine().Options.GuaranteedTailCallOpt)
2857 // Variable argument functions are not supported.
2861 MachineFunction &MF = DAG.getMachineFunction();
2862 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2863 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2864 // Functions containing by val parameters are not supported.
2865 for (unsigned i = 0; i != Ins.size(); i++) {
2866 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2867 if (Flags.isByVal()) return false;
2870 // Non PIC/GOT tail calls are supported.
2871 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2874 // At the moment we can only do local tail calls (in same module, hidden
2875 // or protected) if we are generating PIC.
2876 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2877 return G->getGlobal()->hasHiddenVisibility()
2878 || G->getGlobal()->hasProtectedVisibility();
2884 /// isCallCompatibleAddress - Return the immediate to use if the specified
2885 /// 32-bit value is representable in the immediate field of a BxA instruction.
2886 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2887 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2890 int Addr = C->getZExtValue();
2891 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2892 SignExtend32<26>(Addr) != Addr)
2893 return 0; // Top 6 bits have to be sext of immediate.
2895 return DAG.getConstant((int)C->getZExtValue() >> 2,
2896 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2901 struct TailCallArgumentInfo {
2906 TailCallArgumentInfo() : FrameIdx(0) {}
2911 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2913 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2915 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2916 SmallVector<SDValue, 8> &MemOpChains,
2918 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2919 SDValue Arg = TailCallArgs[i].Arg;
2920 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2921 int FI = TailCallArgs[i].FrameIdx;
2922 // Store relative to framepointer.
2923 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2924 MachinePointerInfo::getFixedStack(FI),
2929 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2930 /// the appropriate stack slot for the tail call optimized function call.
2931 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2932 MachineFunction &MF,
2941 // Calculate the new stack slot for the return address.
2942 int SlotSize = isPPC64 ? 8 : 4;
2943 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2945 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2946 NewRetAddrLoc, true);
2947 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2948 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2949 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2950 MachinePointerInfo::getFixedStack(NewRetAddr),
2953 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2954 // slot as the FP is never overwritten.
2957 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2958 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2960 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2961 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2962 MachinePointerInfo::getFixedStack(NewFPIdx),
2969 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2970 /// the position of the argument.
2972 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2973 SDValue Arg, int SPDiff, unsigned ArgOffset,
2974 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2975 int Offset = ArgOffset + SPDiff;
2976 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2977 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2978 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2979 SDValue FIN = DAG.getFrameIndex(FI, VT);
2980 TailCallArgumentInfo Info;
2982 Info.FrameIdxOp = FIN;
2984 TailCallArguments.push_back(Info);
2987 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2988 /// stack slot. Returns the chain as result and the loaded frame pointers in
2989 /// LROpOut/FPOpout. Used when tail calling.
2990 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2996 DebugLoc dl) const {
2998 // Load the LR and FP stack slot for later adjusting.
2999 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
3000 LROpOut = getReturnAddrFrameIndex(DAG);
3001 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
3002 false, false, false, 0);
3003 Chain = SDValue(LROpOut.getNode(), 1);
3005 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
3006 // slot as the FP is never overwritten.
3008 FPOpOut = getFramePointerFrameIndex(DAG);
3009 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
3010 false, false, false, 0);
3011 Chain = SDValue(FPOpOut.getNode(), 1);
3017 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
3018 /// by "Src" to address "Dst" of size "Size". Alignment information is
3019 /// specified by the specific parameter attribute. The copy will be passed as
3020 /// a byval function parameter.
3021 /// Sometimes what we are copying is the end of a larger object, the part that
3022 /// does not fit in registers.
3024 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
3025 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
3027 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
3028 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
3029 false, false, MachinePointerInfo(0),
3030 MachinePointerInfo(0));
3033 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
3036 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
3037 SDValue Arg, SDValue PtrOff, int SPDiff,
3038 unsigned ArgOffset, bool isPPC64, bool isTailCall,
3039 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
3040 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
3042 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3047 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3049 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3050 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3051 DAG.getConstant(ArgOffset, PtrVT));
3053 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3054 MachinePointerInfo(), false, false, 0));
3055 // Calculate and remember argument location.
3056 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
3061 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
3062 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
3063 SDValue LROp, SDValue FPOp, bool isDarwinABI,
3064 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
3065 MachineFunction &MF = DAG.getMachineFunction();
3067 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
3068 // might overwrite each other in case of tail call optimization.
3069 SmallVector<SDValue, 8> MemOpChains2;
3070 // Do not flag preceding copytoreg stuff together with the following stuff.
3072 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
3074 if (!MemOpChains2.empty())
3075 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3076 &MemOpChains2[0], MemOpChains2.size());
3078 // Store the return address to the appropriate stack slot.
3079 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
3080 isPPC64, isDarwinABI, dl);
3082 // Emit callseq_end just before tailcall node.
3083 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3084 DAG.getIntPtrConstant(0, true), InFlag);
3085 InFlag = Chain.getValue(1);
3089 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
3090 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
3091 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
3092 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
3093 const PPCSubtarget &PPCSubTarget) {
3095 bool isPPC64 = PPCSubTarget.isPPC64();
3096 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
3098 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3099 NodeTys.push_back(MVT::Other); // Returns a chain
3100 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
3102 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
3104 bool needIndirectCall = true;
3105 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
3106 // If this is an absolute destination address, use the munged value.
3107 Callee = SDValue(Dest, 0);
3108 needIndirectCall = false;
3111 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
3112 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
3113 // Use indirect calls for ALL functions calls in JIT mode, since the
3114 // far-call stubs may be outside relocation limits for a BL instruction.
3115 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
3116 unsigned OpFlags = 0;
3117 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3118 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3119 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
3120 (G->getGlobal()->isDeclaration() ||
3121 G->getGlobal()->isWeakForLinker())) {
3122 // PC-relative references to external symbols should go through $stub,
3123 // unless we're building with the leopard linker or later, which
3124 // automatically synthesizes these stubs.
3125 OpFlags = PPCII::MO_DARWIN_STUB;
3128 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
3129 // every direct call is) turn it into a TargetGlobalAddress /
3130 // TargetExternalSymbol node so that legalize doesn't hack it.
3131 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
3132 Callee.getValueType(),
3134 needIndirectCall = false;
3138 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3139 unsigned char OpFlags = 0;
3141 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
3142 (PPCSubTarget.getTargetTriple().isMacOSX() &&
3143 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
3144 // PC-relative references to external symbols should go through $stub,
3145 // unless we're building with the leopard linker or later, which
3146 // automatically synthesizes these stubs.
3147 OpFlags = PPCII::MO_DARWIN_STUB;
3150 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
3152 needIndirectCall = false;
3155 if (needIndirectCall) {
3156 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3157 // to do the call, we can't use PPCISD::CALL.
3158 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3160 if (isSVR4ABI && isPPC64) {
3161 // Function pointers in the 64-bit SVR4 ABI do not point to the function
3162 // entry point, but to the function descriptor (the function entry point
3163 // address is part of the function descriptor though).
3164 // The function descriptor is a three doubleword structure with the
3165 // following fields: function entry point, TOC base address and
3166 // environment pointer.
3167 // Thus for a call through a function pointer, the following actions need
3169 // 1. Save the TOC of the caller in the TOC save area of its stack
3170 // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()).
3171 // 2. Load the address of the function entry point from the function
3173 // 3. Load the TOC of the callee from the function descriptor into r2.
3174 // 4. Load the environment pointer from the function descriptor into
3176 // 5. Branch to the function entry point address.
3177 // 6. On return of the callee, the TOC of the caller needs to be
3178 // restored (this is done in FinishCall()).
3180 // All those operations are flagged together to ensure that no other
3181 // operations can be scheduled in between. E.g. without flagging the
3182 // operations together, a TOC access in the caller could be scheduled
3183 // between the load of the callee TOC and the branch to the callee, which
3184 // results in the TOC access going through the TOC of the callee instead
3185 // of going through the TOC of the caller, which leads to incorrect code.
3187 // Load the address of the function entry point from the function
3189 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
3190 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
3191 InFlag.getNode() ? 3 : 2);
3192 Chain = LoadFuncPtr.getValue(1);
3193 InFlag = LoadFuncPtr.getValue(2);
3195 // Load environment pointer into r11.
3196 // Offset of the environment pointer within the function descriptor.
3197 SDValue PtrOff = DAG.getIntPtrConstant(16);
3199 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
3200 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
3202 Chain = LoadEnvPtr.getValue(1);
3203 InFlag = LoadEnvPtr.getValue(2);
3205 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
3207 Chain = EnvVal.getValue(0);
3208 InFlag = EnvVal.getValue(1);
3210 // Load TOC of the callee into r2. We are using a target-specific load
3211 // with r2 hard coded, because the result of a target-independent load
3212 // would never go directly into r2, since r2 is a reserved register (which
3213 // prevents the register allocator from allocating it), resulting in an
3214 // additional register being allocated and an unnecessary move instruction
3216 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3217 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
3219 Chain = LoadTOCPtr.getValue(0);
3220 InFlag = LoadTOCPtr.getValue(1);
3222 MTCTROps[0] = Chain;
3223 MTCTROps[1] = LoadFuncPtr;
3224 MTCTROps[2] = InFlag;
3227 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3228 2 + (InFlag.getNode() != 0));
3229 InFlag = Chain.getValue(1);
3232 NodeTys.push_back(MVT::Other);
3233 NodeTys.push_back(MVT::Glue);
3234 Ops.push_back(Chain);
3235 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
3237 // Add CTR register as callee so a bctr can be emitted later.
3239 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
3242 // If this is a direct call, pass the chain and the callee.
3243 if (Callee.getNode()) {
3244 Ops.push_back(Chain);
3245 Ops.push_back(Callee);
3247 // If this is a tail call add stack pointer delta.
3249 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3251 // Add argument registers to the end of the list so that they are known live
3253 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3254 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3255 RegsToPass[i].second.getValueType()));
3261 bool isLocalCall(const SDValue &Callee)
3263 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3264 return !G->getGlobal()->isDeclaration() &&
3265 !G->getGlobal()->isWeakForLinker();
3270 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
3271 CallingConv::ID CallConv, bool isVarArg,
3272 const SmallVectorImpl<ISD::InputArg> &Ins,
3273 DebugLoc dl, SelectionDAG &DAG,
3274 SmallVectorImpl<SDValue> &InVals) const {
3276 SmallVector<CCValAssign, 16> RVLocs;
3277 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3278 getTargetMachine(), RVLocs, *DAG.getContext());
3279 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
3281 // Copy all of the result registers out of their specified physreg.
3282 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3283 CCValAssign &VA = RVLocs[i];
3284 assert(VA.isRegLoc() && "Can only return in registers!");
3286 SDValue Val = DAG.getCopyFromReg(Chain, dl,
3287 VA.getLocReg(), VA.getLocVT(), InFlag);
3288 Chain = Val.getValue(1);
3289 InFlag = Val.getValue(2);
3291 switch (VA.getLocInfo()) {
3292 default: llvm_unreachable("Unknown loc info!");
3293 case CCValAssign::Full: break;
3294 case CCValAssign::AExt:
3295 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3297 case CCValAssign::ZExt:
3298 Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val,
3299 DAG.getValueType(VA.getValVT()));
3300 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3302 case CCValAssign::SExt:
3303 Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val,
3304 DAG.getValueType(VA.getValVT()));
3305 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
3309 InVals.push_back(Val);
3316 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
3317 bool isTailCall, bool isVarArg,
3319 SmallVector<std::pair<unsigned, SDValue>, 8>
3321 SDValue InFlag, SDValue Chain,
3323 int SPDiff, unsigned NumBytes,
3324 const SmallVectorImpl<ISD::InputArg> &Ins,
3325 SmallVectorImpl<SDValue> &InVals) const {
3326 std::vector<EVT> NodeTys;
3327 SmallVector<SDValue, 8> Ops;
3328 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
3329 isTailCall, RegsToPass, Ops, NodeTys,
3332 // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls
3333 if (isVarArg && PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
3334 Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32));
3336 // When performing tail call optimization the callee pops its arguments off
3337 // the stack. Account for this here so these bytes can be pushed back on in
3338 // PPCFrameLowering::eliminateCallFramePseudoInstr.
3339 int BytesCalleePops =
3340 (CallConv == CallingConv::Fast &&
3341 getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
3343 // Add a register mask operand representing the call-preserved registers.
3344 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
3345 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3346 assert(Mask && "Missing call preserved mask for calling convention");
3347 Ops.push_back(DAG.getRegisterMask(Mask));
3349 if (InFlag.getNode())
3350 Ops.push_back(InFlag);
3354 assert(((Callee.getOpcode() == ISD::Register &&
3355 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
3356 Callee.getOpcode() == ISD::TargetExternalSymbol ||
3357 Callee.getOpcode() == ISD::TargetGlobalAddress ||
3358 isa<ConstantSDNode>(Callee)) &&
3359 "Expecting an global address, external symbol, absolute value or register");
3361 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
3364 // Add a NOP immediately after the branch instruction when using the 64-bit
3365 // SVR4 ABI. At link time, if caller and callee are in a different module and
3366 // thus have a different TOC, the call will be replaced with a call to a stub
3367 // function which saves the current TOC, loads the TOC of the callee and
3368 // branches to the callee. The NOP will be replaced with a load instruction
3369 // which restores the TOC of the caller from the TOC save slot of the current
3370 // stack frame. If caller and callee belong to the same module (and have the
3371 // same TOC), the NOP will remain unchanged.
3373 bool needsTOCRestore = false;
3374 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
3375 if (CallOpc == PPCISD::BCTRL_SVR4) {
3376 // This is a call through a function pointer.
3377 // Restore the caller TOC from the save area into R2.
3378 // See PrepareCall() for more information about calls through function
3379 // pointers in the 64-bit SVR4 ABI.
3380 // We are using a target-specific load with r2 hard coded, because the
3381 // result of a target-independent load would never go directly into r2,
3382 // since r2 is a reserved register (which prevents the register allocator
3383 // from allocating it), resulting in an additional register being
3384 // allocated and an unnecessary move instruction being generated.
3385 needsTOCRestore = true;
3386 } else if ((CallOpc == PPCISD::CALL_SVR4) && !isLocalCall(Callee)) {
3387 // Otherwise insert NOP for non-local calls.
3388 CallOpc = PPCISD::CALL_NOP_SVR4;
3392 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3393 InFlag = Chain.getValue(1);
3395 if (needsTOCRestore) {
3396 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3397 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
3398 InFlag = Chain.getValue(1);
3401 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3402 DAG.getIntPtrConstant(BytesCalleePops, true),
3405 InFlag = Chain.getValue(1);
3407 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3408 Ins, dl, DAG, InVals);
3412 PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3413 SmallVectorImpl<SDValue> &InVals) const {
3414 SelectionDAG &DAG = CLI.DAG;
3415 DebugLoc &dl = CLI.DL;
3416 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
3417 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
3418 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
3419 SDValue Chain = CLI.Chain;
3420 SDValue Callee = CLI.Callee;
3421 bool &isTailCall = CLI.IsTailCall;
3422 CallingConv::ID CallConv = CLI.CallConv;
3423 bool isVarArg = CLI.IsVarArg;
3426 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
3429 if (PPCSubTarget.isSVR4ABI()) {
3430 if (PPCSubTarget.isPPC64())
3431 return LowerCall_64SVR4(Chain, Callee, CallConv, isVarArg,
3432 isTailCall, Outs, OutVals, Ins,
3435 return LowerCall_32SVR4(Chain, Callee, CallConv, isVarArg,
3436 isTailCall, Outs, OutVals, Ins,
3440 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
3441 isTailCall, Outs, OutVals, Ins,
3446 PPCTargetLowering::LowerCall_32SVR4(SDValue Chain, SDValue Callee,
3447 CallingConv::ID CallConv, bool isVarArg,
3449 const SmallVectorImpl<ISD::OutputArg> &Outs,
3450 const SmallVectorImpl<SDValue> &OutVals,
3451 const SmallVectorImpl<ISD::InputArg> &Ins,
3452 DebugLoc dl, SelectionDAG &DAG,
3453 SmallVectorImpl<SDValue> &InVals) const {
3454 // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description
3455 // of the 32-bit SVR4 ABI stack frame layout.
3457 assert((CallConv == CallingConv::C ||
3458 CallConv == CallingConv::Fast) && "Unknown calling convention!");
3460 unsigned PtrByteSize = 4;
3462 MachineFunction &MF = DAG.getMachineFunction();
3464 // Mark this function as potentially containing a function that contains a
3465 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3466 // and restoring the callers stack pointer in this functions epilog. This is
3467 // done because by tail calling the called function might overwrite the value
3468 // in this function's (MF) stack pointer stack slot 0(SP).
3469 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3470 CallConv == CallingConv::Fast)
3471 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3473 // Count how many bytes are to be pushed on the stack, including the linkage
3474 // area, parameter list area and the part of the local variable space which
3475 // contains copies of aggregates which are passed by value.
3477 // Assign locations to all of the outgoing arguments.
3478 SmallVector<CCValAssign, 16> ArgLocs;
3479 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3480 getTargetMachine(), ArgLocs, *DAG.getContext());
3482 // Reserve space for the linkage area on the stack.
3483 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
3486 // Handle fixed and variable vector arguments differently.
3487 // Fixed vector arguments go into registers as long as registers are
3488 // available. Variable vector arguments always go into memory.
3489 unsigned NumArgs = Outs.size();
3491 for (unsigned i = 0; i != NumArgs; ++i) {
3492 MVT ArgVT = Outs[i].VT;
3493 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
3496 if (Outs[i].IsFixed) {
3497 Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
3500 Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
3506 errs() << "Call operand #" << i << " has unhandled type "
3507 << EVT(ArgVT).getEVTString() << "\n";
3509 llvm_unreachable(0);
3513 // All arguments are treated the same.
3514 CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4);
3517 // Assign locations to all of the outgoing aggregate by value arguments.
3518 SmallVector<CCValAssign, 16> ByValArgLocs;
3519 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3520 getTargetMachine(), ByValArgLocs, *DAG.getContext());
3522 // Reserve stack space for the allocations in CCInfo.
3523 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
3525 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal);
3527 // Size of the linkage area, parameter list area and the part of the local
3528 // space variable where copies of aggregates which are passed by value are
3530 unsigned NumBytes = CCByValInfo.getNextStackOffset();
3532 // Calculate by how many bytes the stack has to be adjusted in case of tail
3533 // call optimization.
3534 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3536 // Adjust the stack pointer for the new arguments...
3537 // These operations are automatically eliminated by the prolog/epilog pass
3538 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3539 SDValue CallSeqStart = Chain;
3541 // Load the return address and frame pointer so it can be moved somewhere else
3544 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
3547 // Set up a copy of the stack pointer for use loading and storing any
3548 // arguments that may not fit in the registers available for argument
3550 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3552 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3553 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3554 SmallVector<SDValue, 8> MemOpChains;
3556 bool seenFloatArg = false;
3557 // Walk the register/memloc assignments, inserting copies/loads.
3558 for (unsigned i = 0, j = 0, e = ArgLocs.size();
3561 CCValAssign &VA = ArgLocs[i];
3562 SDValue Arg = OutVals[i];
3563 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3565 if (Flags.isByVal()) {
3566 // Argument is an aggregate which is passed by value, thus we need to
3567 // create a copy of it in the local variable space of the current stack
3568 // frame (which is the stack frame of the caller) and pass the address of
3569 // this copy to the callee.
3570 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
3571 CCValAssign &ByValVA = ByValArgLocs[j++];
3572 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
3574 // Memory reserved in the local variable space of the callers stack frame.
3575 unsigned LocMemOffset = ByValVA.getLocMemOffset();
3577 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3578 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3580 // Create a copy of the argument in the local area of the current
3582 SDValue MemcpyCall =
3583 CreateCopyOfByValArgument(Arg, PtrOff,
3584 CallSeqStart.getNode()->getOperand(0),
3587 // This must go outside the CALLSEQ_START..END.
3588 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3589 CallSeqStart.getNode()->getOperand(1));
3590 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3591 NewCallSeqStart.getNode());
3592 Chain = CallSeqStart = NewCallSeqStart;
3594 // Pass the address of the aggregate copy on the stack either in a
3595 // physical register or in the parameter list area of the current stack
3596 // frame to the callee.
3600 if (VA.isRegLoc()) {
3601 seenFloatArg |= VA.getLocVT().isFloatingPoint();
3602 // Put argument in a physical register.
3603 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3605 // Put argument in the parameter list area of the current stack frame.
3606 assert(VA.isMemLoc());
3607 unsigned LocMemOffset = VA.getLocMemOffset();
3610 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
3611 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
3613 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
3614 MachinePointerInfo(),
3617 // Calculate and remember argument location.
3618 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
3624 if (!MemOpChains.empty())
3625 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3626 &MemOpChains[0], MemOpChains.size());
3628 // Build a sequence of copy-to-reg nodes chained together with token chain
3629 // and flag operands which copy the outgoing args into the appropriate regs.
3631 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3632 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3633 RegsToPass[i].second, InFlag);
3634 InFlag = Chain.getValue(1);
3637 // Set CR bit 6 to true if this is a vararg call with floating args passed in
3640 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
3641 SDValue Ops[] = { Chain, InFlag };
3643 Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET,
3644 dl, VTs, Ops, InFlag.getNode() ? 2 : 1);
3646 InFlag = Chain.getValue(1);
3650 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
3651 false, TailCallArguments);
3653 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3654 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3658 // Copy an argument into memory, being careful to do this outside the
3659 // call sequence for the call to which the argument belongs.
3661 PPCTargetLowering::createMemcpyOutsideCallSeq(SDValue Arg, SDValue PtrOff,
3662 SDValue CallSeqStart,
3663 ISD::ArgFlagsTy Flags,
3665 DebugLoc dl) const {
3666 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3667 CallSeqStart.getNode()->getOperand(0),
3669 // The MEMCPY must go outside the CALLSEQ_START..END.
3670 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3671 CallSeqStart.getNode()->getOperand(1));
3672 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3673 NewCallSeqStart.getNode());
3674 return NewCallSeqStart;
3678 PPCTargetLowering::LowerCall_64SVR4(SDValue Chain, SDValue Callee,
3679 CallingConv::ID CallConv, bool isVarArg,
3681 const SmallVectorImpl<ISD::OutputArg> &Outs,
3682 const SmallVectorImpl<SDValue> &OutVals,
3683 const SmallVectorImpl<ISD::InputArg> &Ins,
3684 DebugLoc dl, SelectionDAG &DAG,
3685 SmallVectorImpl<SDValue> &InVals) const {
3687 unsigned NumOps = Outs.size();
3689 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3690 unsigned PtrByteSize = 8;
3692 MachineFunction &MF = DAG.getMachineFunction();
3694 // Mark this function as potentially containing a function that contains a
3695 // tail call. As a consequence the frame pointer will be used for dynamicalloc
3696 // and restoring the callers stack pointer in this functions epilog. This is
3697 // done because by tail calling the called function might overwrite the value
3698 // in this function's (MF) stack pointer stack slot 0(SP).
3699 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
3700 CallConv == CallingConv::Fast)
3701 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
3703 unsigned nAltivecParamsAtEnd = 0;
3705 // Count how many bytes are to be pushed on the stack, including the linkage
3706 // area, and parameter passing area. We start with at least 48 bytes, which
3707 // is reserved space for [SP][CR][LR][3 x unused].
3708 // NOTE: For PPC64, nAltivecParamsAtEnd always remains zero as a result
3711 CalculateParameterAndLinkageAreaSize(DAG, true, isVarArg, CallConv,
3712 Outs, OutVals, nAltivecParamsAtEnd);
3714 // Calculate by how many bytes the stack has to be adjusted in case of tail
3715 // call optimization.
3716 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
3718 // To protect arguments on the stack from being clobbered in a tail call,
3719 // force all the loads to happen before doing any other lowering.
3721 Chain = DAG.getStackArgumentTokenFactor(Chain);
3723 // Adjust the stack pointer for the new arguments...
3724 // These operations are automatically eliminated by the prolog/epilog pass
3725 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
3726 SDValue CallSeqStart = Chain;
3728 // Load the return address and frame pointer so it can be move somewhere else
3731 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
3734 // Set up a copy of the stack pointer for use loading and storing any
3735 // arguments that may not fit in the registers available for argument
3737 SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3739 // Figure out which arguments are going to go in registers, and which in
3740 // memory. Also, if this is a vararg function, floating point operations
3741 // must be stored to our stack, and loaded into integer regs as well, if
3742 // any integer regs are available for argument passing.
3743 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(true, true);
3744 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3746 static const uint16_t GPR[] = {
3747 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3748 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3750 static const uint16_t *FPR = GetFPR();
3752 static const uint16_t VR[] = {
3753 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3754 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3756 const unsigned NumGPRs = array_lengthof(GPR);
3757 const unsigned NumFPRs = 13;
3758 const unsigned NumVRs = array_lengthof(VR);
3760 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3761 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3763 SmallVector<SDValue, 8> MemOpChains;
3764 for (unsigned i = 0; i != NumOps; ++i) {
3765 SDValue Arg = OutVals[i];
3766 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3768 // PtrOff will be used to store the current argument to the stack if a
3769 // register cannot be found for it.
3772 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3774 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3776 // Promote integers to 64-bit values.
3777 if (Arg.getValueType() == MVT::i32) {
3778 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3779 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3780 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3783 // FIXME memcpy is used way more than necessary. Correctness first.
3784 // Note: "by value" is code for passing a structure by value, not
3786 if (Flags.isByVal()) {
3787 // Note: Size includes alignment padding, so
3788 // struct x { short a; char b; }
3789 // will have Size = 4. With #pragma pack(1), it will have Size = 3.
3790 // These are the proper values we need for right-justifying the
3791 // aggregate in a parameter register.
3792 unsigned Size = Flags.getByValSize();
3794 // An empty aggregate parameter takes up no storage and no
3799 // All aggregates smaller than 8 bytes must be passed right-justified.
3800 if (Size==1 || Size==2 || Size==4) {
3801 EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32);
3802 if (GPR_idx != NumGPRs) {
3803 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3804 MachinePointerInfo(), VT,
3806 MemOpChains.push_back(Load.getValue(1));
3807 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3809 ArgOffset += PtrByteSize;
3814 if (GPR_idx == NumGPRs && Size < 8) {
3815 SDValue Const = DAG.getConstant(PtrByteSize - Size,
3816 PtrOff.getValueType());
3817 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3818 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3821 ArgOffset += PtrByteSize;
3824 // Copy entire object into memory. There are cases where gcc-generated
3825 // code assumes it is there, even if it could be put entirely into
3826 // registers. (This is not what the doc says.)
3828 // FIXME: The above statement is likely due to a misunderstanding of the
3829 // documents. All arguments must be copied into the parameter area BY
3830 // THE CALLEE in the event that the callee takes the address of any
3831 // formal argument. That has not yet been implemented. However, it is
3832 // reasonable to use the stack area as a staging area for the register
3835 // Skip this for small aggregates, as we will use the same slot for a
3836 // right-justified copy, below.
3838 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
3842 // When a register is available, pass a small aggregate right-justified.
3843 if (Size < 8 && GPR_idx != NumGPRs) {
3844 // The easiest way to get this right-justified in a register
3845 // is to copy the structure into the rightmost portion of a
3846 // local variable slot, then load the whole slot into the
3848 // FIXME: The memcpy seems to produce pretty awful code for
3849 // small aggregates, particularly for packed ones.
3850 // FIXME: It would be preferable to use the slot in the
3851 // parameter save area instead of a new local variable.
3852 SDValue Const = DAG.getConstant(8 - Size, PtrOff.getValueType());
3853 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3854 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
3858 // Load the slot into the register.
3859 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, PtrOff,
3860 MachinePointerInfo(),
3861 false, false, false, 0);
3862 MemOpChains.push_back(Load.getValue(1));
3863 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3865 // Done with this argument.
3866 ArgOffset += PtrByteSize;
3870 // For aggregates larger than PtrByteSize, copy the pieces of the
3871 // object that fit into registers from the parameter save area.
3872 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3873 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3874 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3875 if (GPR_idx != NumGPRs) {
3876 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3877 MachinePointerInfo(),
3878 false, false, false, 0);
3879 MemOpChains.push_back(Load.getValue(1));
3880 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3881 ArgOffset += PtrByteSize;
3883 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3890 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3891 default: llvm_unreachable("Unexpected ValueType for argument!");
3894 if (GPR_idx != NumGPRs) {
3895 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3897 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3898 true, isTailCall, false, MemOpChains,
3899 TailCallArguments, dl);
3901 ArgOffset += PtrByteSize;
3905 if (FPR_idx != NumFPRs) {
3906 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3909 // A single float or an aggregate containing only a single float
3910 // must be passed right-justified in the stack doubleword, and
3911 // in the GPR, if one is available.
3913 if (Arg.getValueType().getSimpleVT().SimpleTy == MVT::f32) {
3914 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3915 StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3919 SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
3920 MachinePointerInfo(), false, false, 0);
3921 MemOpChains.push_back(Store);
3923 // Float varargs are always shadowed in available integer registers
3924 if (GPR_idx != NumGPRs) {
3925 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3926 MachinePointerInfo(), false, false,
3928 MemOpChains.push_back(Load.getValue(1));
3929 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3931 } else if (GPR_idx != NumGPRs)
3932 // If we have any FPRs remaining, we may also have GPRs remaining.
3935 // Single-precision floating-point values are mapped to the
3936 // second (rightmost) word of the stack doubleword.
3937 if (Arg.getValueType() == MVT::f32) {
3938 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3939 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3942 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3943 true, isTailCall, false, MemOpChains,
3944 TailCallArguments, dl);
3953 // These go aligned on the stack, or in the corresponding R registers
3954 // when within range. The Darwin PPC ABI doc claims they also go in
3955 // V registers; in fact gcc does this only for arguments that are
3956 // prototyped, not for those that match the ... We do it for all
3957 // arguments, seems to work.
3958 while (ArgOffset % 16 !=0) {
3959 ArgOffset += PtrByteSize;
3960 if (GPR_idx != NumGPRs)
3963 // We could elide this store in the case where the object fits
3964 // entirely in R registers. Maybe later.
3965 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3966 DAG.getConstant(ArgOffset, PtrVT));
3967 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3968 MachinePointerInfo(), false, false, 0);
3969 MemOpChains.push_back(Store);
3970 if (VR_idx != NumVRs) {
3971 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3972 MachinePointerInfo(),
3973 false, false, false, 0);
3974 MemOpChains.push_back(Load.getValue(1));
3975 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3978 for (unsigned i=0; i<16; i+=PtrByteSize) {
3979 if (GPR_idx == NumGPRs)
3981 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3982 DAG.getConstant(i, PtrVT));
3983 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3984 false, false, false, 0);
3985 MemOpChains.push_back(Load.getValue(1));
3986 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3991 // Non-varargs Altivec params generally go in registers, but have
3992 // stack space allocated at the end.
3993 if (VR_idx != NumVRs) {
3994 // Doesn't have GPR space allocated.
3995 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3997 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3998 true, isTailCall, true, MemOpChains,
3999 TailCallArguments, dl);
4006 if (!MemOpChains.empty())
4007 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4008 &MemOpChains[0], MemOpChains.size());
4010 // Check if this is an indirect call (MTCTR/BCTRL).
4011 // See PrepareCall() for more information about calls through function
4012 // pointers in the 64-bit SVR4 ABI.
4014 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4015 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4016 !isBLACompatibleAddress(Callee, DAG)) {
4017 // Load r2 into a virtual register and store it to the TOC save area.
4018 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
4019 // TOC save area offset.
4020 SDValue PtrOff = DAG.getIntPtrConstant(40);
4021 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4022 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
4024 // R12 must contain the address of an indirect callee. This does not
4025 // mean the MTCTR instruction must use R12; it's easier to model this
4026 // as an extra parameter, so do that.
4027 RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee));
4030 // Build a sequence of copy-to-reg nodes chained together with token chain
4031 // and flag operands which copy the outgoing args into the appropriate regs.
4033 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4034 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4035 RegsToPass[i].second, InFlag);
4036 InFlag = Chain.getValue(1);
4040 PrepareTailCall(DAG, InFlag, Chain, dl, true, SPDiff, NumBytes, LROp,
4041 FPOp, true, TailCallArguments);
4043 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4044 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4049 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
4050 CallingConv::ID CallConv, bool isVarArg,
4052 const SmallVectorImpl<ISD::OutputArg> &Outs,
4053 const SmallVectorImpl<SDValue> &OutVals,
4054 const SmallVectorImpl<ISD::InputArg> &Ins,
4055 DebugLoc dl, SelectionDAG &DAG,
4056 SmallVectorImpl<SDValue> &InVals) const {
4058 unsigned NumOps = Outs.size();
4060 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4061 bool isPPC64 = PtrVT == MVT::i64;
4062 unsigned PtrByteSize = isPPC64 ? 8 : 4;
4064 MachineFunction &MF = DAG.getMachineFunction();
4066 // Mark this function as potentially containing a function that contains a
4067 // tail call. As a consequence the frame pointer will be used for dynamicalloc
4068 // and restoring the callers stack pointer in this functions epilog. This is
4069 // done because by tail calling the called function might overwrite the value
4070 // in this function's (MF) stack pointer stack slot 0(SP).
4071 if (getTargetMachine().Options.GuaranteedTailCallOpt &&
4072 CallConv == CallingConv::Fast)
4073 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
4075 unsigned nAltivecParamsAtEnd = 0;
4077 // Count how many bytes are to be pushed on the stack, including the linkage
4078 // area, and parameter passing area. We start with 24/48 bytes, which is
4079 // prereserved space for [SP][CR][LR][3 x unused].
4081 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
4083 nAltivecParamsAtEnd);
4085 // Calculate by how many bytes the stack has to be adjusted in case of tail
4086 // call optimization.
4087 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
4089 // To protect arguments on the stack from being clobbered in a tail call,
4090 // force all the loads to happen before doing any other lowering.
4092 Chain = DAG.getStackArgumentTokenFactor(Chain);
4094 // Adjust the stack pointer for the new arguments...
4095 // These operations are automatically eliminated by the prolog/epilog pass
4096 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
4097 SDValue CallSeqStart = Chain;
4099 // Load the return address and frame pointer so it can be move somewhere else
4102 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
4105 // Set up a copy of the stack pointer for use loading and storing any
4106 // arguments that may not fit in the registers available for argument
4110 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
4112 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
4114 // Figure out which arguments are going to go in registers, and which in
4115 // memory. Also, if this is a vararg function, floating point operations
4116 // must be stored to our stack, and loaded into integer regs as well, if
4117 // any integer regs are available for argument passing.
4118 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
4119 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
4121 static const uint16_t GPR_32[] = { // 32-bit registers.
4122 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
4123 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
4125 static const uint16_t GPR_64[] = { // 64-bit registers.
4126 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
4127 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
4129 static const uint16_t *FPR = GetFPR();
4131 static const uint16_t VR[] = {
4132 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
4133 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
4135 const unsigned NumGPRs = array_lengthof(GPR_32);
4136 const unsigned NumFPRs = 13;
4137 const unsigned NumVRs = array_lengthof(VR);
4139 const uint16_t *GPR = isPPC64 ? GPR_64 : GPR_32;
4141 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
4142 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
4144 SmallVector<SDValue, 8> MemOpChains;
4145 for (unsigned i = 0; i != NumOps; ++i) {
4146 SDValue Arg = OutVals[i];
4147 ISD::ArgFlagsTy Flags = Outs[i].Flags;
4149 // PtrOff will be used to store the current argument to the stack if a
4150 // register cannot be found for it.
4153 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
4155 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
4157 // On PPC64, promote integers to 64-bit values.
4158 if (isPPC64 && Arg.getValueType() == MVT::i32) {
4159 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
4160 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4161 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
4164 // FIXME memcpy is used way more than necessary. Correctness first.
4165 // Note: "by value" is code for passing a structure by value, not
4167 if (Flags.isByVal()) {
4168 unsigned Size = Flags.getByValSize();
4169 // Very small objects are passed right-justified. Everything else is
4170 // passed left-justified.
4171 if (Size==1 || Size==2) {
4172 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
4173 if (GPR_idx != NumGPRs) {
4174 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
4175 MachinePointerInfo(), VT,
4177 MemOpChains.push_back(Load.getValue(1));
4178 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4180 ArgOffset += PtrByteSize;
4182 SDValue Const = DAG.getConstant(PtrByteSize - Size,
4183 PtrOff.getValueType());
4184 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
4185 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr,
4188 ArgOffset += PtrByteSize;
4192 // Copy entire object into memory. There are cases where gcc-generated
4193 // code assumes it is there, even if it could be put entirely into
4194 // registers. (This is not what the doc says.)
4195 Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff,
4199 // For small aggregates (Darwin only) and aggregates >= PtrByteSize,
4200 // copy the pieces of the object that fit into registers from the
4201 // parameter save area.
4202 for (unsigned j=0; j<Size; j+=PtrByteSize) {
4203 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
4204 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
4205 if (GPR_idx != NumGPRs) {
4206 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
4207 MachinePointerInfo(),
4208 false, false, false, 0);
4209 MemOpChains.push_back(Load.getValue(1));
4210 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4211 ArgOffset += PtrByteSize;
4213 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
4220 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
4221 default: llvm_unreachable("Unexpected ValueType for argument!");
4224 if (GPR_idx != NumGPRs) {
4225 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
4227 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4228 isPPC64, isTailCall, false, MemOpChains,
4229 TailCallArguments, dl);
4231 ArgOffset += PtrByteSize;
4235 if (FPR_idx != NumFPRs) {
4236 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
4239 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4240 MachinePointerInfo(), false, false, 0);
4241 MemOpChains.push_back(Store);
4243 // Float varargs are always shadowed in available integer registers
4244 if (GPR_idx != NumGPRs) {
4245 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4246 MachinePointerInfo(), false, false,
4248 MemOpChains.push_back(Load.getValue(1));
4249 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4251 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
4252 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
4253 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
4254 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
4255 MachinePointerInfo(),
4256 false, false, false, 0);
4257 MemOpChains.push_back(Load.getValue(1));
4258 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4261 // If we have any FPRs remaining, we may also have GPRs remaining.
4262 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
4264 if (GPR_idx != NumGPRs)
4266 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
4267 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
4271 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4272 isPPC64, isTailCall, false, MemOpChains,
4273 TailCallArguments, dl);
4277 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
4284 // These go aligned on the stack, or in the corresponding R registers
4285 // when within range. The Darwin PPC ABI doc claims they also go in
4286 // V registers; in fact gcc does this only for arguments that are
4287 // prototyped, not for those that match the ... We do it for all
4288 // arguments, seems to work.
4289 while (ArgOffset % 16 !=0) {
4290 ArgOffset += PtrByteSize;
4291 if (GPR_idx != NumGPRs)
4294 // We could elide this store in the case where the object fits
4295 // entirely in R registers. Maybe later.
4296 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
4297 DAG.getConstant(ArgOffset, PtrVT));
4298 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
4299 MachinePointerInfo(), false, false, 0);
4300 MemOpChains.push_back(Store);
4301 if (VR_idx != NumVRs) {
4302 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
4303 MachinePointerInfo(),
4304 false, false, false, 0);
4305 MemOpChains.push_back(Load.getValue(1));
4306 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
4309 for (unsigned i=0; i<16; i+=PtrByteSize) {
4310 if (GPR_idx == NumGPRs)
4312 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
4313 DAG.getConstant(i, PtrVT));
4314 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
4315 false, false, false, 0);
4316 MemOpChains.push_back(Load.getValue(1));
4317 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
4322 // Non-varargs Altivec params generally go in registers, but have
4323 // stack space allocated at the end.
4324 if (VR_idx != NumVRs) {
4325 // Doesn't have GPR space allocated.
4326 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
4327 } else if (nAltivecParamsAtEnd==0) {
4328 // We are emitting Altivec params in order.
4329 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4330 isPPC64, isTailCall, true, MemOpChains,
4331 TailCallArguments, dl);
4337 // If all Altivec parameters fit in registers, as they usually do,
4338 // they get stack space following the non-Altivec parameters. We
4339 // don't track this here because nobody below needs it.
4340 // If there are more Altivec parameters than fit in registers emit
4342 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
4344 // Offset is aligned; skip 1st 12 params which go in V registers.
4345 ArgOffset = ((ArgOffset+15)/16)*16;
4347 for (unsigned i = 0; i != NumOps; ++i) {
4348 SDValue Arg = OutVals[i];
4349 EVT ArgType = Outs[i].VT;
4350 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
4351 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
4354 // We are emitting Altivec params in order.
4355 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
4356 isPPC64, isTailCall, true, MemOpChains,
4357 TailCallArguments, dl);
4364 if (!MemOpChains.empty())
4365 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
4366 &MemOpChains[0], MemOpChains.size());
4368 // On Darwin, R12 must contain the address of an indirect callee. This does
4369 // not mean the MTCTR instruction must use R12; it's easier to model this as
4370 // an extra parameter, so do that.
4372 !dyn_cast<GlobalAddressSDNode>(Callee) &&
4373 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
4374 !isBLACompatibleAddress(Callee, DAG))
4375 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
4376 PPC::R12), Callee));
4378 // Build a sequence of copy-to-reg nodes chained together with token chain
4379 // and flag operands which copy the outgoing args into the appropriate regs.
4381 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
4382 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
4383 RegsToPass[i].second, InFlag);
4384 InFlag = Chain.getValue(1);
4388 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
4389 FPOp, true, TailCallArguments);
4391 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
4392 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
4397 PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
4398 MachineFunction &MF, bool isVarArg,
4399 const SmallVectorImpl<ISD::OutputArg> &Outs,
4400 LLVMContext &Context) const {
4401 SmallVector<CCValAssign, 16> RVLocs;
4402 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
4404 return CCInfo.CheckReturn(Outs, RetCC_PPC);
4408 PPCTargetLowering::LowerReturn(SDValue Chain,
4409 CallingConv::ID CallConv, bool isVarArg,
4410 const SmallVectorImpl<ISD::OutputArg> &Outs,
4411 const SmallVectorImpl<SDValue> &OutVals,
4412 DebugLoc dl, SelectionDAG &DAG) const {
4414 SmallVector<CCValAssign, 16> RVLocs;
4415 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
4416 getTargetMachine(), RVLocs, *DAG.getContext());
4417 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
4420 SmallVector<SDValue, 4> RetOps(1, Chain);
4422 // Copy the result values into the output registers.
4423 for (unsigned i = 0; i != RVLocs.size(); ++i) {
4424 CCValAssign &VA = RVLocs[i];
4425 assert(VA.isRegLoc() && "Can only return in registers!");
4427 SDValue Arg = OutVals[i];
4429 switch (VA.getLocInfo()) {
4430 default: llvm_unreachable("Unknown loc info!");
4431 case CCValAssign::Full: break;
4432 case CCValAssign::AExt:
4433 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
4435 case CCValAssign::ZExt:
4436 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
4438 case CCValAssign::SExt:
4439 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
4443 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
4444 Flag = Chain.getValue(1);
4445 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
4448 RetOps[0] = Chain; // Update chain.
4450 // Add the flag if we have it.
4452 RetOps.push_back(Flag);
4454 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other,
4455 &RetOps[0], RetOps.size());
4458 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
4459 const PPCSubtarget &Subtarget) const {
4460 // When we pop the dynamic allocation we need to restore the SP link.
4461 DebugLoc dl = Op.getDebugLoc();
4463 // Get the corect type for pointers.
4464 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4466 // Construct the stack pointer operand.
4467 bool isPPC64 = Subtarget.isPPC64();
4468 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
4469 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
4471 // Get the operands for the STACKRESTORE.
4472 SDValue Chain = Op.getOperand(0);
4473 SDValue SaveSP = Op.getOperand(1);
4475 // Load the old link SP.
4476 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
4477 MachinePointerInfo(),
4478 false, false, false, 0);
4480 // Restore the stack pointer.
4481 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
4483 // Store the old link SP.
4484 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
4491 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
4492 MachineFunction &MF = DAG.getMachineFunction();
4493 bool isPPC64 = PPCSubTarget.isPPC64();
4494 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4495 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4497 // Get current frame pointer save index. The users of this index will be
4498 // primarily DYNALLOC instructions.
4499 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4500 int RASI = FI->getReturnAddrSaveIndex();
4502 // If the frame pointer save index hasn't been defined yet.
4504 // Find out what the fix offset of the frame pointer save area.
4505 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
4506 // Allocate the frame index for frame pointer save area.
4507 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
4509 FI->setReturnAddrSaveIndex(RASI);
4511 return DAG.getFrameIndex(RASI, PtrVT);
4515 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
4516 MachineFunction &MF = DAG.getMachineFunction();
4517 bool isPPC64 = PPCSubTarget.isPPC64();
4518 bool isDarwinABI = PPCSubTarget.isDarwinABI();
4519 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4521 // Get current frame pointer save index. The users of this index will be
4522 // primarily DYNALLOC instructions.
4523 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
4524 int FPSI = FI->getFramePointerSaveIndex();
4526 // If the frame pointer save index hasn't been defined yet.
4528 // Find out what the fix offset of the frame pointer save area.
4529 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
4532 // Allocate the frame index for frame pointer save area.
4533 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
4535 FI->setFramePointerSaveIndex(FPSI);
4537 return DAG.getFrameIndex(FPSI, PtrVT);
4540 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
4542 const PPCSubtarget &Subtarget) const {
4544 SDValue Chain = Op.getOperand(0);
4545 SDValue Size = Op.getOperand(1);
4546 DebugLoc dl = Op.getDebugLoc();
4548 // Get the corect type for pointers.
4549 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4551 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
4552 DAG.getConstant(0, PtrVT), Size);
4553 // Construct a node for the frame pointer save index.
4554 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
4555 // Build a DYNALLOC node.
4556 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
4557 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
4558 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
4561 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
4563 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
4564 // Not FP? Not a fsel.
4565 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
4566 !Op.getOperand(2).getValueType().isFloatingPoint())
4569 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
4571 // Cannot handle SETEQ/SETNE.
4572 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
4574 EVT ResVT = Op.getValueType();
4575 EVT CmpVT = Op.getOperand(0).getValueType();
4576 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4577 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
4578 DebugLoc dl = Op.getDebugLoc();
4580 // If the RHS of the comparison is a 0.0, we don't need to do the
4581 // subtraction at all.
4582 if (isFloatingPointZero(RHS))
4584 default: break; // SETUO etc aren't handled by fsel.
4587 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4590 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4591 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4592 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
4595 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
4598 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
4599 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
4600 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
4601 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
4606 default: break; // SETUO etc aren't handled by fsel.
4609 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4610 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4611 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4612 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4615 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
4616 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4617 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4618 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4621 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4622 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4623 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4624 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
4627 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
4628 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
4629 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
4630 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
4635 // FIXME: Split this code up when LegalizeDAGTypes lands.
4636 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
4637 DebugLoc dl) const {
4638 assert(Op.getOperand(0).getValueType().isFloatingPoint());
4639 SDValue Src = Op.getOperand(0);
4640 if (Src.getValueType() == MVT::f32)
4641 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
4644 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4645 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
4647 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
4652 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
4656 // Convert the FP value to an int value through memory.
4657 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
4659 // Emit a store to the stack slot.
4660 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
4661 MachinePointerInfo(), false, false, 0);
4663 // Result is a load from the stack slot. If loading 4 bytes, make sure to
4665 if (Op.getValueType() == MVT::i32)
4666 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
4667 DAG.getConstant(4, FIPtr.getValueType()));
4668 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
4669 false, false, false, 0);
4672 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
4673 SelectionDAG &DAG) const {
4674 DebugLoc dl = Op.getDebugLoc();
4675 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
4676 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
4679 if (Op.getOperand(0).getValueType() == MVT::i64) {
4680 SDValue SINT = Op.getOperand(0);
4681 // When converting to single-precision, we actually need to convert
4682 // to double-precision first and then round to single-precision.
4683 // To avoid double-rounding effects during that operation, we have
4684 // to prepare the input operand. Bits that might be truncated when
4685 // converting to double-precision are replaced by a bit that won't
4686 // be lost at this stage, but is below the single-precision rounding
4689 // However, if -enable-unsafe-fp-math is in effect, accept double
4690 // rounding to avoid the extra overhead.
4691 if (Op.getValueType() == MVT::f32 &&
4692 !DAG.getTarget().Options.UnsafeFPMath) {
4694 // Twiddle input to make sure the low 11 bits are zero. (If this
4695 // is the case, we are guaranteed the value will fit into the 53 bit
4696 // mantissa of an IEEE double-precision value without rounding.)
4697 // If any of those low 11 bits were not zero originally, make sure
4698 // bit 12 (value 2048) is set instead, so that the final rounding
4699 // to single-precision gets the correct result.
4700 SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4701 SINT, DAG.getConstant(2047, MVT::i64));
4702 Round = DAG.getNode(ISD::ADD, dl, MVT::i64,
4703 Round, DAG.getConstant(2047, MVT::i64));
4704 Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT);
4705 Round = DAG.getNode(ISD::AND, dl, MVT::i64,
4706 Round, DAG.getConstant(-2048, MVT::i64));
4708 // However, we cannot use that value unconditionally: if the magnitude
4709 // of the input value is small, the bit-twiddling we did above might
4710 // end up visibly changing the output. Fortunately, in that case, we
4711 // don't need to twiddle bits since the original input will convert
4712 // exactly to double-precision floating-point already. Therefore,
4713 // construct a conditional to use the original value if the top 11
4714 // bits are all sign-bit copies, and use the rounded value computed
4716 SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64,
4717 SINT, DAG.getConstant(53, MVT::i32));
4718 Cond = DAG.getNode(ISD::ADD, dl, MVT::i64,
4719 Cond, DAG.getConstant(1, MVT::i64));
4720 Cond = DAG.getSetCC(dl, MVT::i32,
4721 Cond, DAG.getConstant(1, MVT::i64), ISD::SETUGT);
4723 SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT);
4725 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT);
4726 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
4727 if (Op.getValueType() == MVT::f32)
4728 FP = DAG.getNode(ISD::FP_ROUND, dl,
4729 MVT::f32, FP, DAG.getIntPtrConstant(0));
4733 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
4734 "Unhandled SINT_TO_FP type in custom expander!");
4735 // Since we only generate this in 64-bit mode, we can take advantage of
4736 // 64-bit registers. In particular, sign extend the input value into the
4737 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
4738 // then lfd it and fcfid it.
4739 MachineFunction &MF = DAG.getMachineFunction();
4740 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
4741 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
4742 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4743 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4745 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
4748 // STD the extended value into the stack slot.
4749 MachineMemOperand *MMO =
4750 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
4751 MachineMemOperand::MOStore, 8, 8);
4752 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
4754 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
4755 Ops, 4, MVT::i64, MMO);
4756 // Load the value as a double.
4757 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
4758 false, false, false, 0);
4760 // FCFID it and return it.
4761 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
4762 if (Op.getValueType() == MVT::f32)
4763 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
4767 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
4768 SelectionDAG &DAG) const {
4769 DebugLoc dl = Op.getDebugLoc();
4771 The rounding mode is in bits 30:31 of FPSR, and has the following
4778 FLT_ROUNDS, on the other hand, expects the following:
4785 To perform the conversion, we do:
4786 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
4789 MachineFunction &MF = DAG.getMachineFunction();
4790 EVT VT = Op.getValueType();
4791 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4792 SDValue MFFSreg, InFlag;
4794 // Save FP Control Word to register
4796 MVT::f64, // return register
4797 MVT::Glue // unused in this context
4799 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4801 // Save FP register to stack slot
4802 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
4803 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
4804 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
4805 StackSlot, MachinePointerInfo(), false, false,0);
4807 // Load FP Control Word from low 32 bits of stack slot.
4808 SDValue Four = DAG.getConstant(4, PtrVT);
4809 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
4810 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
4811 false, false, false, 0);
4813 // Transform as necessary
4815 DAG.getNode(ISD::AND, dl, MVT::i32,
4816 CWD, DAG.getConstant(3, MVT::i32));
4818 DAG.getNode(ISD::SRL, dl, MVT::i32,
4819 DAG.getNode(ISD::AND, dl, MVT::i32,
4820 DAG.getNode(ISD::XOR, dl, MVT::i32,
4821 CWD, DAG.getConstant(3, MVT::i32)),
4822 DAG.getConstant(3, MVT::i32)),
4823 DAG.getConstant(1, MVT::i32));
4826 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
4828 return DAG.getNode((VT.getSizeInBits() < 16 ?
4829 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
4832 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4833 EVT VT = Op.getValueType();
4834 unsigned BitWidth = VT.getSizeInBits();
4835 DebugLoc dl = Op.getDebugLoc();
4836 assert(Op.getNumOperands() == 3 &&
4837 VT == Op.getOperand(1).getValueType() &&
4840 // Expand into a bunch of logical ops. Note that these ops
4841 // depend on the PPC behavior for oversized shift amounts.
4842 SDValue Lo = Op.getOperand(0);
4843 SDValue Hi = Op.getOperand(1);
4844 SDValue Amt = Op.getOperand(2);
4845 EVT AmtVT = Amt.getValueType();
4847 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4848 DAG.getConstant(BitWidth, AmtVT), Amt);
4849 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
4850 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
4851 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
4852 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4853 DAG.getConstant(-BitWidth, AmtVT));
4854 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
4855 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4856 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
4857 SDValue OutOps[] = { OutLo, OutHi };
4858 return DAG.getMergeValues(OutOps, 2, dl);
4861 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
4862 EVT VT = Op.getValueType();
4863 DebugLoc dl = Op.getDebugLoc();
4864 unsigned BitWidth = VT.getSizeInBits();
4865 assert(Op.getNumOperands() == 3 &&
4866 VT == Op.getOperand(1).getValueType() &&
4869 // Expand into a bunch of logical ops. Note that these ops
4870 // depend on the PPC behavior for oversized shift amounts.
4871 SDValue Lo = Op.getOperand(0);
4872 SDValue Hi = Op.getOperand(1);
4873 SDValue Amt = Op.getOperand(2);
4874 EVT AmtVT = Amt.getValueType();
4876 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4877 DAG.getConstant(BitWidth, AmtVT), Amt);
4878 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4879 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4880 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4881 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4882 DAG.getConstant(-BitWidth, AmtVT));
4883 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
4884 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
4885 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
4886 SDValue OutOps[] = { OutLo, OutHi };
4887 return DAG.getMergeValues(OutOps, 2, dl);
4890 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
4891 DebugLoc dl = Op.getDebugLoc();
4892 EVT VT = Op.getValueType();
4893 unsigned BitWidth = VT.getSizeInBits();
4894 assert(Op.getNumOperands() == 3 &&
4895 VT == Op.getOperand(1).getValueType() &&
4898 // Expand into a bunch of logical ops, followed by a select_cc.
4899 SDValue Lo = Op.getOperand(0);
4900 SDValue Hi = Op.getOperand(1);
4901 SDValue Amt = Op.getOperand(2);
4902 EVT AmtVT = Amt.getValueType();
4904 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
4905 DAG.getConstant(BitWidth, AmtVT), Amt);
4906 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
4907 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
4908 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
4909 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
4910 DAG.getConstant(-BitWidth, AmtVT));
4911 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
4912 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
4913 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
4914 Tmp4, Tmp6, ISD::SETLE);
4915 SDValue OutOps[] = { OutLo, OutHi };
4916 return DAG.getMergeValues(OutOps, 2, dl);
4919 //===----------------------------------------------------------------------===//
4920 // Vector related lowering.
4923 /// BuildSplatI - Build a canonical splati of Val with an element size of
4924 /// SplatSize. Cast the result to VT.
4925 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
4926 SelectionDAG &DAG, DebugLoc dl) {
4927 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
4929 static const EVT VTys[] = { // canonical VT to use for each size.
4930 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
4933 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
4935 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
4939 EVT CanonicalVT = VTys[SplatSize-1];
4941 // Build a canonical splat for this value.
4942 SDValue Elt = DAG.getConstant(Val, MVT::i32);
4943 SmallVector<SDValue, 8> Ops;
4944 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
4945 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
4946 &Ops[0], Ops.size());
4947 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
4950 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
4951 /// specified intrinsic ID.
4952 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
4953 SelectionDAG &DAG, DebugLoc dl,
4954 EVT DestVT = MVT::Other) {
4955 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
4956 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4957 DAG.getConstant(IID, MVT::i32), LHS, RHS);
4960 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
4961 /// specified intrinsic ID.
4962 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
4963 SDValue Op2, SelectionDAG &DAG,
4964 DebugLoc dl, EVT DestVT = MVT::Other) {
4965 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
4966 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
4967 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
4971 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
4972 /// amount. The result has the specified value type.
4973 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
4974 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4975 // Force LHS/RHS to be the right type.
4976 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
4977 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
4980 for (unsigned i = 0; i != 16; ++i)
4982 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
4983 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4986 // If this is a case we can't handle, return null and let the default
4987 // expansion code take care of it. If we CAN select this case, and if it
4988 // selects to a single instruction, return Op. Otherwise, if we can codegen
4989 // this case more efficiently than a constant pool load, lower it to the
4990 // sequence of ops that should be used.
4991 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
4992 SelectionDAG &DAG) const {
4993 DebugLoc dl = Op.getDebugLoc();
4994 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
4995 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
4997 // Check if this is a splat of a constant value.
4998 APInt APSplatBits, APSplatUndef;
4999 unsigned SplatBitSize;
5001 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
5002 HasAnyUndefs, 0, true) || SplatBitSize > 32)
5005 unsigned SplatBits = APSplatBits.getZExtValue();
5006 unsigned SplatUndef = APSplatUndef.getZExtValue();
5007 unsigned SplatSize = SplatBitSize / 8;
5009 // First, handle single instruction cases.
5012 if (SplatBits == 0) {
5013 // Canonicalize all zero vectors to be v4i32.
5014 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
5015 SDValue Z = DAG.getConstant(0, MVT::i32);
5016 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
5017 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
5022 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
5023 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
5025 if (SextVal >= -16 && SextVal <= 15)
5026 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
5029 // Two instruction sequences.
5031 // If this value is in the range [-32,30] and is even, use:
5032 // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2)
5033 // If this value is in the range [17,31] and is odd, use:
5034 // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16)
5035 // If this value is in the range [-31,-17] and is odd, use:
5036 // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16)
5037 // Note the last two are three-instruction sequences.
5038 if (SextVal >= -32 && SextVal <= 31) {
5039 // To avoid having these optimizations undone by constant folding,
5040 // we convert to a pseudo that will be expanded later into one of
5042 SDValue Elt = DAG.getConstant(SextVal, MVT::i32);
5043 EVT VT = Op.getValueType();
5044 int Size = VT == MVT::v16i8 ? 1 : (VT == MVT::v8i16 ? 2 : 4);
5045 SDValue EltSize = DAG.getConstant(Size, MVT::i32);
5046 return DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize);
5049 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
5050 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
5052 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
5053 // Make -1 and vspltisw -1:
5054 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
5056 // Make the VSLW intrinsic, computing 0x8000_0000.
5057 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
5060 // xor by OnesV to invert it.
5061 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
5062 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5065 // Check to see if this is a wide variety of vsplti*, binop self cases.
5066 static const signed char SplatCsts[] = {
5067 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
5068 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
5071 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
5072 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
5073 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
5074 int i = SplatCsts[idx];
5076 // Figure out what shift amount will be used by altivec if shifted by i in
5078 unsigned TypeShiftAmt = i & (SplatBitSize-1);
5080 // vsplti + shl self.
5081 if (SextVal == (int)((unsigned)i << TypeShiftAmt)) {
5082 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5083 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5084 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
5085 Intrinsic::ppc_altivec_vslw
5087 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5088 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5091 // vsplti + srl self.
5092 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5093 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5094 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5095 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
5096 Intrinsic::ppc_altivec_vsrw
5098 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5099 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5102 // vsplti + sra self.
5103 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
5104 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5105 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5106 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
5107 Intrinsic::ppc_altivec_vsraw
5109 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5110 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5113 // vsplti + rol self.
5114 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
5115 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
5116 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
5117 static const unsigned IIDs[] = { // Intrinsic to use for each size.
5118 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
5119 Intrinsic::ppc_altivec_vrlw
5121 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
5122 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
5125 // t = vsplti c, result = vsldoi t, t, 1
5126 if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) {
5127 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5128 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
5130 // t = vsplti c, result = vsldoi t, t, 2
5131 if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) {
5132 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5133 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
5135 // t = vsplti c, result = vsldoi t, t, 3
5136 if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
5137 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
5138 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
5145 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
5146 /// the specified operations to build the shuffle.
5147 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
5148 SDValue RHS, SelectionDAG &DAG,
5150 unsigned OpNum = (PFEntry >> 26) & 0x0F;
5151 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
5152 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
5155 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
5167 if (OpNum == OP_COPY) {
5168 if (LHSID == (1*9+2)*9+3) return LHS;
5169 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
5173 SDValue OpLHS, OpRHS;
5174 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
5175 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
5179 default: llvm_unreachable("Unknown i32 permute!");
5181 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
5182 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
5183 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
5184 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
5187 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
5188 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
5189 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
5190 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
5193 for (unsigned i = 0; i != 16; ++i)
5194 ShufIdxs[i] = (i&3)+0;
5197 for (unsigned i = 0; i != 16; ++i)
5198 ShufIdxs[i] = (i&3)+4;
5201 for (unsigned i = 0; i != 16; ++i)
5202 ShufIdxs[i] = (i&3)+8;
5205 for (unsigned i = 0; i != 16; ++i)
5206 ShufIdxs[i] = (i&3)+12;
5209 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
5211 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
5213 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
5215 EVT VT = OpLHS.getValueType();
5216 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
5217 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
5218 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
5219 return DAG.getNode(ISD::BITCAST, dl, VT, T);
5222 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
5223 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
5224 /// return the code it can be lowered into. Worst case, it can always be
5225 /// lowered into a vperm.
5226 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5227 SelectionDAG &DAG) const {
5228 DebugLoc dl = Op.getDebugLoc();
5229 SDValue V1 = Op.getOperand(0);
5230 SDValue V2 = Op.getOperand(1);
5231 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5232 EVT VT = Op.getValueType();
5234 // Cases that are handled by instructions that take permute immediates
5235 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
5236 // selected by the instruction selector.
5237 if (V2.getOpcode() == ISD::UNDEF) {
5238 if (PPC::isSplatShuffleMask(SVOp, 1) ||
5239 PPC::isSplatShuffleMask(SVOp, 2) ||
5240 PPC::isSplatShuffleMask(SVOp, 4) ||
5241 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
5242 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
5243 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
5244 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
5245 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
5246 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
5247 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
5248 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
5249 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
5254 // Altivec has a variety of "shuffle immediates" that take two vector inputs
5255 // and produce a fixed permutation. If any of these match, do not lower to
5257 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
5258 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
5259 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
5260 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
5261 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
5262 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
5263 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
5264 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
5265 PPC::isVMRGHShuffleMask(SVOp, 4, false))
5268 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
5269 // perfect shuffle table to emit an optimal matching sequence.
5270 ArrayRef<int> PermMask = SVOp->getMask();
5272 unsigned PFIndexes[4];
5273 bool isFourElementShuffle = true;
5274 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
5275 unsigned EltNo = 8; // Start out undef.
5276 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
5277 if (PermMask[i*4+j] < 0)
5278 continue; // Undef, ignore it.
5280 unsigned ByteSource = PermMask[i*4+j];
5281 if ((ByteSource & 3) != j) {
5282 isFourElementShuffle = false;
5287 EltNo = ByteSource/4;
5288 } else if (EltNo != ByteSource/4) {
5289 isFourElementShuffle = false;
5293 PFIndexes[i] = EltNo;
5296 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
5297 // perfect shuffle vector to determine if it is cost effective to do this as
5298 // discrete instructions, or whether we should use a vperm.
5299 if (isFourElementShuffle) {
5300 // Compute the index in the perfect shuffle table.
5301 unsigned PFTableIndex =
5302 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
5304 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
5305 unsigned Cost = (PFEntry >> 30);
5307 // Determining when to avoid vperm is tricky. Many things affect the cost
5308 // of vperm, particularly how many times the perm mask needs to be computed.
5309 // For example, if the perm mask can be hoisted out of a loop or is already
5310 // used (perhaps because there are multiple permutes with the same shuffle
5311 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
5312 // the loop requires an extra register.
5314 // As a compromise, we only emit discrete instructions if the shuffle can be
5315 // generated in 3 or fewer operations. When we have loop information
5316 // available, if this block is within a loop, we should avoid using vperm
5317 // for 3-operation perms and use a constant pool load instead.
5319 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
5322 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
5323 // vector that will get spilled to the constant pool.
5324 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
5326 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
5327 // that it is in input element units, not in bytes. Convert now.
5328 EVT EltVT = V1.getValueType().getVectorElementType();
5329 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
5331 SmallVector<SDValue, 16> ResultMask;
5332 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
5333 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
5335 for (unsigned j = 0; j != BytesPerElement; ++j)
5336 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
5340 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
5341 &ResultMask[0], ResultMask.size());
5342 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
5345 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
5346 /// altivec comparison. If it is, return true and fill in Opc/isDot with
5347 /// information about the intrinsic.
5348 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
5350 unsigned IntrinsicID =
5351 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
5354 switch (IntrinsicID) {
5355 default: return false;
5356 // Comparison predicates.
5357 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
5358 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
5359 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
5360 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
5361 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
5362 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
5363 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
5364 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
5365 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
5366 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
5367 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
5368 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
5369 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
5371 // Normal Comparisons.
5372 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
5373 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
5374 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
5375 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
5376 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
5377 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
5378 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
5379 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
5380 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
5381 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
5382 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
5383 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
5384 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
5389 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
5390 /// lower, do it, otherwise return null.
5391 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
5392 SelectionDAG &DAG) const {
5393 // If this is a lowered altivec predicate compare, CompareOpc is set to the
5394 // opcode number of the comparison.
5395 DebugLoc dl = Op.getDebugLoc();
5398 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
5399 return SDValue(); // Don't custom lower most intrinsics.
5401 // If this is a non-dot comparison, make the VCMP node and we are done.
5403 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
5404 Op.getOperand(1), Op.getOperand(2),
5405 DAG.getConstant(CompareOpc, MVT::i32));
5406 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
5409 // Create the PPCISD altivec 'dot' comparison node.
5411 Op.getOperand(2), // LHS
5412 Op.getOperand(3), // RHS
5413 DAG.getConstant(CompareOpc, MVT::i32)
5415 EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue };
5416 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5418 // Now that we have the comparison, emit a copy from the CR to a GPR.
5419 // This is flagged to the above dot comparison.
5420 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
5421 DAG.getRegister(PPC::CR6, MVT::i32),
5422 CompNode.getValue(1));
5424 // Unpack the result based on how the target uses it.
5425 unsigned BitNo; // Bit # of CR6.
5426 bool InvertBit; // Invert result?
5427 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
5428 default: // Can't happen, don't crash on invalid number though.
5429 case 0: // Return the value of the EQ bit of CR6.
5430 BitNo = 0; InvertBit = false;
5432 case 1: // Return the inverted value of the EQ bit of CR6.
5433 BitNo = 0; InvertBit = true;
5435 case 2: // Return the value of the LT bit of CR6.
5436 BitNo = 2; InvertBit = false;
5438 case 3: // Return the inverted value of the LT bit of CR6.
5439 BitNo = 2; InvertBit = true;
5443 // Shift the bit into the low position.
5444 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
5445 DAG.getConstant(8-(3-BitNo), MVT::i32));
5447 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
5448 DAG.getConstant(1, MVT::i32));
5450 // If we are supposed to, toggle the bit.
5452 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
5453 DAG.getConstant(1, MVT::i32));
5457 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
5458 SelectionDAG &DAG) const {
5459 DebugLoc dl = Op.getDebugLoc();
5460 // Create a stack slot that is 16-byte aligned.
5461 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
5462 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
5463 EVT PtrVT = getPointerTy();
5464 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
5466 // Store the input value into Value#0 of the stack slot.
5467 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
5468 Op.getOperand(0), FIdx, MachinePointerInfo(),
5471 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
5472 false, false, false, 0);
5475 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
5476 DebugLoc dl = Op.getDebugLoc();
5477 if (Op.getValueType() == MVT::v4i32) {
5478 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5480 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
5481 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
5483 SDValue RHSSwap = // = vrlw RHS, 16
5484 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
5486 // Shrinkify inputs to v8i16.
5487 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
5488 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
5489 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
5491 // Low parts multiplied together, generating 32-bit results (we ignore the
5493 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
5494 LHS, RHS, DAG, dl, MVT::v4i32);
5496 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
5497 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
5498 // Shift the high parts up 16 bits.
5499 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
5501 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
5502 } else if (Op.getValueType() == MVT::v8i16) {
5503 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5505 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
5507 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
5508 LHS, RHS, Zero, DAG, dl);
5509 } else if (Op.getValueType() == MVT::v16i8) {
5510 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
5512 // Multiply the even 8-bit parts, producing 16-bit sums.
5513 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
5514 LHS, RHS, DAG, dl, MVT::v8i16);
5515 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
5517 // Multiply the odd 8-bit parts, producing 16-bit sums.
5518 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
5519 LHS, RHS, DAG, dl, MVT::v8i16);
5520 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
5522 // Merge the results together.
5524 for (unsigned i = 0; i != 8; ++i) {
5526 Ops[i*2+1] = 2*i+1+16;
5528 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
5530 llvm_unreachable("Unknown mul to lower!");
5534 /// LowerOperation - Provide custom lowering hooks for some operations.
5536 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5537 switch (Op.getOpcode()) {
5538 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
5539 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5540 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5541 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
5542 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5543 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
5544 case ISD::SETCC: return LowerSETCC(Op, DAG);
5545 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
5546 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
5548 return LowerVASTART(Op, DAG, PPCSubTarget);
5551 return LowerVAARG(Op, DAG, PPCSubTarget);
5553 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
5554 case ISD::DYNAMIC_STACKALLOC:
5555 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
5557 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5558 case ISD::FP_TO_UINT:
5559 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
5561 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
5562 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5564 // Lower 64-bit shifts.
5565 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
5566 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
5567 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
5569 // Vector-related lowering.
5570 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
5571 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5572 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
5573 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
5574 case ISD::MUL: return LowerMUL(Op, DAG);
5576 // Frame & Return address.
5577 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5578 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5582 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
5583 SmallVectorImpl<SDValue>&Results,
5584 SelectionDAG &DAG) const {
5585 const TargetMachine &TM = getTargetMachine();
5586 DebugLoc dl = N->getDebugLoc();
5587 switch (N->getOpcode()) {
5589 llvm_unreachable("Do not know how to custom type legalize this operation!");
5591 if (!TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
5592 || TM.getSubtarget<PPCSubtarget>().isPPC64())
5595 EVT VT = N->getValueType(0);
5597 if (VT == MVT::i64) {
5598 SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG, PPCSubTarget);
5600 Results.push_back(NewNode);
5601 Results.push_back(NewNode.getValue(1));
5605 case ISD::FP_ROUND_INREG: {
5606 assert(N->getValueType(0) == MVT::ppcf128);
5607 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
5608 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5609 MVT::f64, N->getOperand(0),
5610 DAG.getIntPtrConstant(0));
5611 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
5612 MVT::f64, N->getOperand(0),
5613 DAG.getIntPtrConstant(1));
5615 // This sequence changes FPSCR to do round-to-zero, adds the two halves
5616 // of the long double, and puts FPSCR back the way it was. We do not
5617 // actually model FPSCR.
5618 std::vector<EVT> NodeTys;
5619 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
5621 NodeTys.push_back(MVT::f64); // Return register
5622 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
5623 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
5624 MFFSreg = Result.getValue(0);
5625 InFlag = Result.getValue(1);
5628 NodeTys.push_back(MVT::Glue); // Returns a flag
5629 Ops[0] = DAG.getConstant(31, MVT::i32);
5631 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
5632 InFlag = Result.getValue(0);
5635 NodeTys.push_back(MVT::Glue); // Returns a flag
5636 Ops[0] = DAG.getConstant(30, MVT::i32);
5638 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
5639 InFlag = Result.getValue(0);
5642 NodeTys.push_back(MVT::f64); // result of add
5643 NodeTys.push_back(MVT::Glue); // Returns a flag
5647 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
5648 FPreg = Result.getValue(0);
5649 InFlag = Result.getValue(1);
5652 NodeTys.push_back(MVT::f64);
5653 Ops[0] = DAG.getConstant(1, MVT::i32);
5657 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
5658 FPreg = Result.getValue(0);
5660 // We know the low half is about to be thrown away, so just use something
5662 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
5666 case ISD::FP_TO_SINT:
5667 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
5673 //===----------------------------------------------------------------------===//
5674 // Other Lowering Code
5675 //===----------------------------------------------------------------------===//
5678 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5679 bool is64bit, unsigned BinOpcode) const {
5680 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5681 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5683 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5684 MachineFunction *F = BB->getParent();
5685 MachineFunction::iterator It = BB;
5688 unsigned dest = MI->getOperand(0).getReg();
5689 unsigned ptrA = MI->getOperand(1).getReg();
5690 unsigned ptrB = MI->getOperand(2).getReg();
5691 unsigned incr = MI->getOperand(3).getReg();
5692 DebugLoc dl = MI->getDebugLoc();
5694 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5695 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5696 F->insert(It, loopMBB);
5697 F->insert(It, exitMBB);
5698 exitMBB->splice(exitMBB->begin(), BB,
5699 llvm::next(MachineBasicBlock::iterator(MI)),
5701 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5703 MachineRegisterInfo &RegInfo = F->getRegInfo();
5704 unsigned TmpReg = (!BinOpcode) ? incr :
5705 RegInfo.createVirtualRegister(
5706 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5707 (const TargetRegisterClass *) &PPC::GPRCRegClass);
5711 // fallthrough --> loopMBB
5712 BB->addSuccessor(loopMBB);
5715 // l[wd]arx dest, ptr
5716 // add r0, dest, incr
5717 // st[wd]cx. r0, ptr
5719 // fallthrough --> exitMBB
5721 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
5722 .addReg(ptrA).addReg(ptrB);
5724 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
5725 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5726 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
5727 BuildMI(BB, dl, TII->get(PPC::BCC))
5728 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5729 BB->addSuccessor(loopMBB);
5730 BB->addSuccessor(exitMBB);
5739 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
5740 MachineBasicBlock *BB,
5741 bool is8bit, // operation
5742 unsigned BinOpcode) const {
5743 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5744 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5745 // In 64 bit mode we have to use 64 bits for addresses, even though the
5746 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
5747 // registers without caring whether they're 32 or 64, but here we're
5748 // doing actual arithmetic on the addresses.
5749 bool is64bit = PPCSubTarget.isPPC64();
5750 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
5752 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5753 MachineFunction *F = BB->getParent();
5754 MachineFunction::iterator It = BB;
5757 unsigned dest = MI->getOperand(0).getReg();
5758 unsigned ptrA = MI->getOperand(1).getReg();
5759 unsigned ptrB = MI->getOperand(2).getReg();
5760 unsigned incr = MI->getOperand(3).getReg();
5761 DebugLoc dl = MI->getDebugLoc();
5763 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
5764 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
5765 F->insert(It, loopMBB);
5766 F->insert(It, exitMBB);
5767 exitMBB->splice(exitMBB->begin(), BB,
5768 llvm::next(MachineBasicBlock::iterator(MI)),
5770 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5772 MachineRegisterInfo &RegInfo = F->getRegInfo();
5773 const TargetRegisterClass *RC =
5774 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
5775 (const TargetRegisterClass *) &PPC::GPRCRegClass;
5776 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
5777 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
5778 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
5779 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
5780 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
5781 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
5782 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
5783 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
5784 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
5785 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
5786 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
5788 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
5792 // fallthrough --> loopMBB
5793 BB->addSuccessor(loopMBB);
5795 // The 4-byte load must be aligned, while a char or short may be
5796 // anywhere in the word. Hence all this nasty bookkeeping code.
5797 // add ptr1, ptrA, ptrB [copy if ptrA==0]
5798 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
5799 // xori shift, shift1, 24 [16]
5800 // rlwinm ptr, ptr1, 0, 0, 29
5801 // slw incr2, incr, shift
5802 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
5803 // slw mask, mask2, shift
5805 // lwarx tmpDest, ptr
5806 // add tmp, tmpDest, incr2
5807 // andc tmp2, tmpDest, mask
5808 // and tmp3, tmp, mask
5809 // or tmp4, tmp3, tmp2
5812 // fallthrough --> exitMBB
5813 // srw dest, tmpDest, shift
5814 if (ptrA != ZeroReg) {
5815 Ptr1Reg = RegInfo.createVirtualRegister(RC);
5816 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
5817 .addReg(ptrA).addReg(ptrB);
5821 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
5822 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
5823 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
5824 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
5826 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
5827 .addReg(Ptr1Reg).addImm(0).addImm(61);
5829 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
5830 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
5831 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
5832 .addReg(incr).addReg(ShiftReg);
5834 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5836 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5837 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
5839 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5840 .addReg(Mask2Reg).addReg(ShiftReg);
5843 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5844 .addReg(ZeroReg).addReg(PtrReg);
5846 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
5847 .addReg(Incr2Reg).addReg(TmpDestReg);
5848 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
5849 .addReg(TmpDestReg).addReg(MaskReg);
5850 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
5851 .addReg(TmpReg).addReg(MaskReg);
5852 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
5853 .addReg(Tmp3Reg).addReg(Tmp2Reg);
5854 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
5855 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
5856 BuildMI(BB, dl, TII->get(PPC::BCC))
5857 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
5858 BB->addSuccessor(loopMBB);
5859 BB->addSuccessor(exitMBB);
5864 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
5870 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
5871 MachineBasicBlock *BB) const {
5872 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5874 // To "insert" these instructions we actually have to insert their
5875 // control-flow patterns.
5876 const BasicBlock *LLVM_BB = BB->getBasicBlock();
5877 MachineFunction::iterator It = BB;
5880 MachineFunction *F = BB->getParent();
5882 if (PPCSubTarget.hasISEL() && (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5883 MI->getOpcode() == PPC::SELECT_CC_I8)) {
5884 unsigned OpCode = MI->getOpcode() == PPC::SELECT_CC_I8 ?
5885 PPC::ISEL8 : PPC::ISEL;
5886 unsigned SelectPred = MI->getOperand(4).getImm();
5887 DebugLoc dl = MI->getDebugLoc();
5889 // The SelectPred is ((BI << 5) | BO) for a BCC
5890 unsigned BO = SelectPred & 0xF;
5891 assert((BO == 12 || BO == 4) && "invalid predicate BO field for isel");
5893 unsigned TrueOpNo, FalseOpNo;
5900 SelectPred = PPC::InvertPredicate((PPC::Predicate)SelectPred);
5903 BuildMI(*BB, MI, dl, TII->get(OpCode), MI->getOperand(0).getReg())
5904 .addReg(MI->getOperand(TrueOpNo).getReg())
5905 .addReg(MI->getOperand(FalseOpNo).getReg())
5906 .addImm(SelectPred).addReg(MI->getOperand(1).getReg());
5907 } else if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
5908 MI->getOpcode() == PPC::SELECT_CC_I8 ||
5909 MI->getOpcode() == PPC::SELECT_CC_F4 ||
5910 MI->getOpcode() == PPC::SELECT_CC_F8 ||
5911 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
5914 // The incoming instruction knows the destination vreg to set, the
5915 // condition code register to branch on, the true/false values to
5916 // select between, and a branch opcode to use.
5921 // cmpTY ccX, r1, r2
5923 // fallthrough --> copy0MBB
5924 MachineBasicBlock *thisMBB = BB;
5925 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
5926 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
5927 unsigned SelectPred = MI->getOperand(4).getImm();
5928 DebugLoc dl = MI->getDebugLoc();
5929 F->insert(It, copy0MBB);
5930 F->insert(It, sinkMBB);
5932 // Transfer the remainder of BB and its successor edges to sinkMBB.
5933 sinkMBB->splice(sinkMBB->begin(), BB,
5934 llvm::next(MachineBasicBlock::iterator(MI)),
5936 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
5938 // Next, add the true and fallthrough blocks as its successors.
5939 BB->addSuccessor(copy0MBB);
5940 BB->addSuccessor(sinkMBB);
5942 BuildMI(BB, dl, TII->get(PPC::BCC))
5943 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
5946 // %FalseValue = ...
5947 // # fallthrough to sinkMBB
5950 // Update machine-CFG edges
5951 BB->addSuccessor(sinkMBB);
5954 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
5957 BuildMI(*BB, BB->begin(), dl,
5958 TII->get(PPC::PHI), MI->getOperand(0).getReg())
5959 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
5960 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
5962 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
5963 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
5964 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
5965 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
5966 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
5967 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
5968 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
5969 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
5971 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
5972 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
5973 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
5974 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
5975 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
5976 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
5977 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
5978 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
5980 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
5981 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
5982 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
5983 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
5984 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
5985 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
5986 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
5987 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
5989 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
5990 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
5991 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
5992 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
5993 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
5994 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
5995 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
5996 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
5998 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
5999 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
6000 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
6001 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
6002 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
6003 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
6004 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
6005 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
6007 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
6008 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
6009 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
6010 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
6011 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
6012 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
6013 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
6014 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
6016 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
6017 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
6018 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
6019 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
6020 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
6021 BB = EmitAtomicBinary(MI, BB, false, 0);
6022 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
6023 BB = EmitAtomicBinary(MI, BB, true, 0);
6025 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
6026 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
6027 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
6029 unsigned dest = MI->getOperand(0).getReg();
6030 unsigned ptrA = MI->getOperand(1).getReg();
6031 unsigned ptrB = MI->getOperand(2).getReg();
6032 unsigned oldval = MI->getOperand(3).getReg();
6033 unsigned newval = MI->getOperand(4).getReg();
6034 DebugLoc dl = MI->getDebugLoc();
6036 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6037 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6038 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6039 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6040 F->insert(It, loop1MBB);
6041 F->insert(It, loop2MBB);
6042 F->insert(It, midMBB);
6043 F->insert(It, exitMBB);
6044 exitMBB->splice(exitMBB->begin(), BB,
6045 llvm::next(MachineBasicBlock::iterator(MI)),
6047 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6051 // fallthrough --> loopMBB
6052 BB->addSuccessor(loop1MBB);
6055 // l[wd]arx dest, ptr
6056 // cmp[wd] dest, oldval
6059 // st[wd]cx. newval, ptr
6063 // st[wd]cx. dest, ptr
6066 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
6067 .addReg(ptrA).addReg(ptrB);
6068 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
6069 .addReg(oldval).addReg(dest);
6070 BuildMI(BB, dl, TII->get(PPC::BCC))
6071 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6072 BB->addSuccessor(loop2MBB);
6073 BB->addSuccessor(midMBB);
6076 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6077 .addReg(newval).addReg(ptrA).addReg(ptrB);
6078 BuildMI(BB, dl, TII->get(PPC::BCC))
6079 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6080 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6081 BB->addSuccessor(loop1MBB);
6082 BB->addSuccessor(exitMBB);
6085 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
6086 .addReg(dest).addReg(ptrA).addReg(ptrB);
6087 BB->addSuccessor(exitMBB);
6092 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
6093 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
6094 // We must use 64-bit registers for addresses when targeting 64-bit,
6095 // since we're actually doing arithmetic on them. Other registers
6097 bool is64bit = PPCSubTarget.isPPC64();
6098 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
6100 unsigned dest = MI->getOperand(0).getReg();
6101 unsigned ptrA = MI->getOperand(1).getReg();
6102 unsigned ptrB = MI->getOperand(2).getReg();
6103 unsigned oldval = MI->getOperand(3).getReg();
6104 unsigned newval = MI->getOperand(4).getReg();
6105 DebugLoc dl = MI->getDebugLoc();
6107 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
6108 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
6109 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
6110 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
6111 F->insert(It, loop1MBB);
6112 F->insert(It, loop2MBB);
6113 F->insert(It, midMBB);
6114 F->insert(It, exitMBB);
6115 exitMBB->splice(exitMBB->begin(), BB,
6116 llvm::next(MachineBasicBlock::iterator(MI)),
6118 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
6120 MachineRegisterInfo &RegInfo = F->getRegInfo();
6121 const TargetRegisterClass *RC =
6122 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
6123 (const TargetRegisterClass *) &PPC::GPRCRegClass;
6124 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
6125 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
6126 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
6127 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
6128 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
6129 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
6130 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
6131 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
6132 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
6133 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
6134 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
6135 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
6136 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
6138 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
6139 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
6142 // fallthrough --> loopMBB
6143 BB->addSuccessor(loop1MBB);
6145 // The 4-byte load must be aligned, while a char or short may be
6146 // anywhere in the word. Hence all this nasty bookkeeping code.
6147 // add ptr1, ptrA, ptrB [copy if ptrA==0]
6148 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
6149 // xori shift, shift1, 24 [16]
6150 // rlwinm ptr, ptr1, 0, 0, 29
6151 // slw newval2, newval, shift
6152 // slw oldval2, oldval,shift
6153 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
6154 // slw mask, mask2, shift
6155 // and newval3, newval2, mask
6156 // and oldval3, oldval2, mask
6158 // lwarx tmpDest, ptr
6159 // and tmp, tmpDest, mask
6160 // cmpw tmp, oldval3
6163 // andc tmp2, tmpDest, mask
6164 // or tmp4, tmp2, newval3
6169 // stwcx. tmpDest, ptr
6171 // srw dest, tmpDest, shift
6172 if (ptrA != ZeroReg) {
6173 Ptr1Reg = RegInfo.createVirtualRegister(RC);
6174 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
6175 .addReg(ptrA).addReg(ptrB);
6179 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
6180 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
6181 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
6182 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
6184 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
6185 .addReg(Ptr1Reg).addImm(0).addImm(61);
6187 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
6188 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
6189 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
6190 .addReg(newval).addReg(ShiftReg);
6191 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
6192 .addReg(oldval).addReg(ShiftReg);
6194 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
6196 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
6197 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
6198 .addReg(Mask3Reg).addImm(65535);
6200 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
6201 .addReg(Mask2Reg).addReg(ShiftReg);
6202 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
6203 .addReg(NewVal2Reg).addReg(MaskReg);
6204 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
6205 .addReg(OldVal2Reg).addReg(MaskReg);
6208 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
6209 .addReg(ZeroReg).addReg(PtrReg);
6210 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
6211 .addReg(TmpDestReg).addReg(MaskReg);
6212 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
6213 .addReg(TmpReg).addReg(OldVal3Reg);
6214 BuildMI(BB, dl, TII->get(PPC::BCC))
6215 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
6216 BB->addSuccessor(loop2MBB);
6217 BB->addSuccessor(midMBB);
6220 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
6221 .addReg(TmpDestReg).addReg(MaskReg);
6222 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
6223 .addReg(Tmp2Reg).addReg(NewVal3Reg);
6224 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
6225 .addReg(ZeroReg).addReg(PtrReg);
6226 BuildMI(BB, dl, TII->get(PPC::BCC))
6227 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
6228 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
6229 BB->addSuccessor(loop1MBB);
6230 BB->addSuccessor(exitMBB);
6233 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
6234 .addReg(ZeroReg).addReg(PtrReg);
6235 BB->addSuccessor(exitMBB);
6240 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
6243 llvm_unreachable("Unexpected instr type to insert");
6246 MI->eraseFromParent(); // The pseudo instruction is gone now.
6250 //===----------------------------------------------------------------------===//
6251 // Target Optimization Hooks
6252 //===----------------------------------------------------------------------===//
6254 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
6255 DAGCombinerInfo &DCI) const {
6256 const TargetMachine &TM = getTargetMachine();
6257 SelectionDAG &DAG = DCI.DAG;
6258 DebugLoc dl = N->getDebugLoc();
6259 switch (N->getOpcode()) {
6262 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6263 if (C->isNullValue()) // 0 << V -> 0.
6264 return N->getOperand(0);
6268 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6269 if (C->isNullValue()) // 0 >>u V -> 0.
6270 return N->getOperand(0);
6274 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
6275 if (C->isNullValue() || // 0 >>s V -> 0.
6276 C->isAllOnesValue()) // -1 >>s V -> -1.
6277 return N->getOperand(0);
6281 case ISD::SINT_TO_FP:
6282 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
6283 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
6284 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
6285 // We allow the src/dst to be either f32/f64, but the intermediate
6286 // type must be i64.
6287 if (N->getOperand(0).getValueType() == MVT::i64 &&
6288 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
6289 SDValue Val = N->getOperand(0).getOperand(0);
6290 if (Val.getValueType() == MVT::f32) {
6291 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6292 DCI.AddToWorklist(Val.getNode());
6295 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
6296 DCI.AddToWorklist(Val.getNode());
6297 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
6298 DCI.AddToWorklist(Val.getNode());
6299 if (N->getValueType(0) == MVT::f32) {
6300 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
6301 DAG.getIntPtrConstant(0));
6302 DCI.AddToWorklist(Val.getNode());
6305 } else if (N->getOperand(0).getValueType() == MVT::i32) {
6306 // If the intermediate type is i32, we can avoid the load/store here
6313 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
6314 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
6315 !cast<StoreSDNode>(N)->isTruncatingStore() &&
6316 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
6317 N->getOperand(1).getValueType() == MVT::i32 &&
6318 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
6319 SDValue Val = N->getOperand(1).getOperand(0);
6320 if (Val.getValueType() == MVT::f32) {
6321 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
6322 DCI.AddToWorklist(Val.getNode());
6324 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
6325 DCI.AddToWorklist(Val.getNode());
6327 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
6328 N->getOperand(2), N->getOperand(3));
6329 DCI.AddToWorklist(Val.getNode());
6333 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
6334 if (cast<StoreSDNode>(N)->isUnindexed() &&
6335 N->getOperand(1).getOpcode() == ISD::BSWAP &&
6336 N->getOperand(1).getNode()->hasOneUse() &&
6337 (N->getOperand(1).getValueType() == MVT::i32 ||
6338 N->getOperand(1).getValueType() == MVT::i16)) {
6339 SDValue BSwapOp = N->getOperand(1).getOperand(0);
6340 // Do an any-extend to 32-bits if this is a half-word input.
6341 if (BSwapOp.getValueType() == MVT::i16)
6342 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
6345 N->getOperand(0), BSwapOp, N->getOperand(2),
6346 DAG.getValueType(N->getOperand(1).getValueType())
6349 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
6350 Ops, array_lengthof(Ops),
6351 cast<StoreSDNode>(N)->getMemoryVT(),
6352 cast<StoreSDNode>(N)->getMemOperand());
6356 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
6357 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
6358 N->getOperand(0).hasOneUse() &&
6359 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
6360 SDValue Load = N->getOperand(0);
6361 LoadSDNode *LD = cast<LoadSDNode>(Load);
6362 // Create the byte-swapping load.
6364 LD->getChain(), // Chain
6365 LD->getBasePtr(), // Ptr
6366 DAG.getValueType(N->getValueType(0)) // VT
6369 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
6370 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
6371 LD->getMemoryVT(), LD->getMemOperand());
6373 // If this is an i16 load, insert the truncate.
6374 SDValue ResVal = BSLoad;
6375 if (N->getValueType(0) == MVT::i16)
6376 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
6378 // First, combine the bswap away. This makes the value produced by the
6380 DCI.CombineTo(N, ResVal);
6382 // Next, combine the load away, we give it a bogus result value but a real
6383 // chain result. The result value is dead because the bswap is dead.
6384 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
6386 // Return N so it doesn't get rechecked!
6387 return SDValue(N, 0);
6391 case PPCISD::VCMP: {
6392 // If a VCMPo node already exists with exactly the same operands as this
6393 // node, use its result instead of this node (VCMPo computes both a CR6 and
6394 // a normal output).
6396 if (!N->getOperand(0).hasOneUse() &&
6397 !N->getOperand(1).hasOneUse() &&
6398 !N->getOperand(2).hasOneUse()) {
6400 // Scan all of the users of the LHS, looking for VCMPo's that match.
6401 SDNode *VCMPoNode = 0;
6403 SDNode *LHSN = N->getOperand(0).getNode();
6404 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
6406 if (UI->getOpcode() == PPCISD::VCMPo &&
6407 UI->getOperand(1) == N->getOperand(1) &&
6408 UI->getOperand(2) == N->getOperand(2) &&
6409 UI->getOperand(0) == N->getOperand(0)) {
6414 // If there is no VCMPo node, or if the flag value has a single use, don't
6416 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
6419 // Look at the (necessarily single) use of the flag value. If it has a
6420 // chain, this transformation is more complex. Note that multiple things
6421 // could use the value result, which we should ignore.
6422 SDNode *FlagUser = 0;
6423 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
6424 FlagUser == 0; ++UI) {
6425 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
6427 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
6428 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
6435 // If the user is a MFCR instruction, we know this is safe. Otherwise we
6436 // give up for right now.
6437 if (FlagUser->getOpcode() == PPCISD::MFCR)
6438 return SDValue(VCMPoNode, 0);
6443 // If this is a branch on an altivec predicate comparison, lower this so
6444 // that we don't have to do a MFCR: instead, branch directly on CR6. This
6445 // lowering is done pre-legalize, because the legalizer lowers the predicate
6446 // compare down to code that is difficult to reassemble.
6447 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
6448 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
6452 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
6453 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
6454 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
6455 assert(isDot && "Can't compare against a vector result!");
6457 // If this is a comparison against something other than 0/1, then we know
6458 // that the condition is never/always true.
6459 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
6460 if (Val != 0 && Val != 1) {
6461 if (CC == ISD::SETEQ) // Cond never true, remove branch.
6462 return N->getOperand(0);
6463 // Always !=, turn it into an unconditional branch.
6464 return DAG.getNode(ISD::BR, dl, MVT::Other,
6465 N->getOperand(0), N->getOperand(4));
6468 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
6470 // Create the PPCISD altivec 'dot' comparison node.
6472 LHS.getOperand(2), // LHS of compare
6473 LHS.getOperand(3), // RHS of compare
6474 DAG.getConstant(CompareOpc, MVT::i32)
6476 EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue };
6477 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
6479 // Unpack the result based on how the target uses it.
6480 PPC::Predicate CompOpc;
6481 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
6482 default: // Can't happen, don't crash on invalid number though.
6483 case 0: // Branch on the value of the EQ bit of CR6.
6484 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
6486 case 1: // Branch on the inverted value of the EQ bit of CR6.
6487 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
6489 case 2: // Branch on the value of the LT bit of CR6.
6490 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
6492 case 3: // Branch on the inverted value of the LT bit of CR6.
6493 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
6497 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
6498 DAG.getConstant(CompOpc, MVT::i32),
6499 DAG.getRegister(PPC::CR6, MVT::i32),
6500 N->getOperand(4), CompNode.getValue(1));
6509 //===----------------------------------------------------------------------===//
6510 // Inline Assembly Support
6511 //===----------------------------------------------------------------------===//
6513 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
6516 const SelectionDAG &DAG,
6517 unsigned Depth) const {
6518 KnownZero = KnownOne = APInt(KnownZero.getBitWidth(), 0);
6519 switch (Op.getOpcode()) {
6521 case PPCISD::LBRX: {
6522 // lhbrx is known to have the top bits cleared out.
6523 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
6524 KnownZero = 0xFFFF0000;
6527 case ISD::INTRINSIC_WO_CHAIN: {
6528 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
6530 case Intrinsic::ppc_altivec_vcmpbfp_p:
6531 case Intrinsic::ppc_altivec_vcmpeqfp_p:
6532 case Intrinsic::ppc_altivec_vcmpequb_p:
6533 case Intrinsic::ppc_altivec_vcmpequh_p:
6534 case Intrinsic::ppc_altivec_vcmpequw_p:
6535 case Intrinsic::ppc_altivec_vcmpgefp_p:
6536 case Intrinsic::ppc_altivec_vcmpgtfp_p:
6537 case Intrinsic::ppc_altivec_vcmpgtsb_p:
6538 case Intrinsic::ppc_altivec_vcmpgtsh_p:
6539 case Intrinsic::ppc_altivec_vcmpgtsw_p:
6540 case Intrinsic::ppc_altivec_vcmpgtub_p:
6541 case Intrinsic::ppc_altivec_vcmpgtuh_p:
6542 case Intrinsic::ppc_altivec_vcmpgtuw_p:
6543 KnownZero = ~1U; // All bits but the low one are known to be zero.
6551 /// getConstraintType - Given a constraint, return the type of
6552 /// constraint it is for this target.
6553 PPCTargetLowering::ConstraintType
6554 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
6555 if (Constraint.size() == 1) {
6556 switch (Constraint[0]) {
6563 return C_RegisterClass;
6565 // FIXME: While Z does indicate a memory constraint, it specifically
6566 // indicates an r+r address (used in conjunction with the 'y' modifier
6567 // in the replacement string). Currently, we're forcing the base
6568 // register to be r0 in the asm printer (which is interpreted as zero)
6569 // and forming the complete address in the second register. This is
6574 return TargetLowering::getConstraintType(Constraint);
6577 /// Examine constraint type and operand type and determine a weight value.
6578 /// This object must already have been set up with the operand type
6579 /// and the current alternative constraint selected.
6580 TargetLowering::ConstraintWeight
6581 PPCTargetLowering::getSingleConstraintMatchWeight(
6582 AsmOperandInfo &info, const char *constraint) const {
6583 ConstraintWeight weight = CW_Invalid;
6584 Value *CallOperandVal = info.CallOperandVal;
6585 // If we don't have a value, we can't do a match,
6586 // but allow it at the lowest weight.
6587 if (CallOperandVal == NULL)
6589 Type *type = CallOperandVal->getType();
6590 // Look at the constraint type.
6591 switch (*constraint) {
6593 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
6596 if (type->isIntegerTy())
6597 weight = CW_Register;
6600 if (type->isFloatTy())
6601 weight = CW_Register;
6604 if (type->isDoubleTy())
6605 weight = CW_Register;
6608 if (type->isVectorTy())
6609 weight = CW_Register;
6612 weight = CW_Register;
6621 std::pair<unsigned, const TargetRegisterClass*>
6622 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
6624 if (Constraint.size() == 1) {
6625 // GCC RS6000 Constraint Letters
6626 switch (Constraint[0]) {
6629 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
6630 return std::make_pair(0U, &PPC::G8RCRegClass);
6631 return std::make_pair(0U, &PPC::GPRCRegClass);
6633 if (VT == MVT::f32 || VT == MVT::i32)
6634 return std::make_pair(0U, &PPC::F4RCRegClass);
6635 if (VT == MVT::f64 || VT == MVT::i64)
6636 return std::make_pair(0U, &PPC::F8RCRegClass);
6639 return std::make_pair(0U, &PPC::VRRCRegClass);
6641 return std::make_pair(0U, &PPC::CRRCRegClass);
6645 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
6649 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
6650 /// vector. If it is invalid, don't add anything to Ops.
6651 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
6652 std::string &Constraint,
6653 std::vector<SDValue>&Ops,
6654 SelectionDAG &DAG) const {
6655 SDValue Result(0,0);
6657 // Only support length 1 constraints.
6658 if (Constraint.length() > 1) return;
6660 char Letter = Constraint[0];
6671 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
6672 if (!CST) return; // Must be an immediate to match.
6673 unsigned Value = CST->getZExtValue();
6675 default: llvm_unreachable("Unknown constraint letter!");
6676 case 'I': // "I" is a signed 16-bit constant.
6677 if ((short)Value == (int)Value)
6678 Result = DAG.getTargetConstant(Value, Op.getValueType());
6680 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
6681 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
6682 if ((short)Value == 0)
6683 Result = DAG.getTargetConstant(Value, Op.getValueType());
6685 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
6686 if ((Value >> 16) == 0)
6687 Result = DAG.getTargetConstant(Value, Op.getValueType());
6689 case 'M': // "M" is a constant that is greater than 31.
6691 Result = DAG.getTargetConstant(Value, Op.getValueType());
6693 case 'N': // "N" is a positive constant that is an exact power of two.
6694 if ((int)Value > 0 && isPowerOf2_32(Value))
6695 Result = DAG.getTargetConstant(Value, Op.getValueType());
6697 case 'O': // "O" is the constant zero.
6699 Result = DAG.getTargetConstant(Value, Op.getValueType());
6701 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
6702 if ((short)-Value == (int)-Value)
6703 Result = DAG.getTargetConstant(Value, Op.getValueType());
6710 if (Result.getNode()) {
6711 Ops.push_back(Result);
6715 // Handle standard constraint letters.
6716 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
6719 // isLegalAddressingMode - Return true if the addressing mode represented
6720 // by AM is legal for this target, for a load/store of the specified type.
6721 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
6723 // FIXME: PPC does not allow r+i addressing modes for vectors!
6725 // PPC allows a sign-extended 16-bit immediate field.
6726 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
6729 // No global is ever allowed as a base.
6733 // PPC only support r+r,
6735 case 0: // "r+i" or just "i", depending on HasBaseReg.
6738 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
6740 // Otherwise we have r+r or r+i.
6743 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
6745 // Allow 2*r as r+r.
6748 // No other scales are supported.
6755 /// isLegalAddressImmediate - Return true if the integer value can be used
6756 /// as the offset of the target addressing mode for load / store of the
6758 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,Type *Ty) const{
6759 // PPC allows a sign-extended 16-bit immediate field.
6760 return (V > -(1 << 16) && V < (1 << 16)-1);
6763 bool PPCTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
6767 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
6768 SelectionDAG &DAG) const {
6769 MachineFunction &MF = DAG.getMachineFunction();
6770 MachineFrameInfo *MFI = MF.getFrameInfo();
6771 MFI->setReturnAddressIsTaken(true);
6773 DebugLoc dl = Op.getDebugLoc();
6774 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6776 // Make sure the function does not optimize away the store of the RA to
6778 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
6779 FuncInfo->setLRStoreRequired();
6780 bool isPPC64 = PPCSubTarget.isPPC64();
6781 bool isDarwinABI = PPCSubTarget.isDarwinABI();
6784 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6787 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
6788 isPPC64? MVT::i64 : MVT::i32);
6789 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6790 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6792 MachinePointerInfo(), false, false, false, 0);
6795 // Just load the return address off the stack.
6796 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
6797 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6798 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
6801 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
6802 SelectionDAG &DAG) const {
6803 DebugLoc dl = Op.getDebugLoc();
6804 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6806 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
6807 bool isPPC64 = PtrVT == MVT::i64;
6809 MachineFunction &MF = DAG.getMachineFunction();
6810 MachineFrameInfo *MFI = MF.getFrameInfo();
6811 MFI->setFrameAddressIsTaken(true);
6812 bool is31 = (getTargetMachine().Options.DisableFramePointerElim(MF) ||
6813 MFI->hasVarSizedObjects()) &&
6814 MFI->getStackSize() &&
6815 !MF.getFunction()->getAttributes().
6816 hasAttribute(AttributeSet::FunctionIndex, Attribute::Naked);
6817 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
6818 (is31 ? PPC::R31 : PPC::R1);
6819 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
6822 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
6823 FrameAddr, MachinePointerInfo(), false, false,
6829 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
6830 // The PowerPC target isn't yet aware of offsets.
6834 /// getOptimalMemOpType - Returns the target specific optimal type for load
6835 /// and store operations as a result of memset, memcpy, and memmove
6836 /// lowering. If DstAlign is zero that means it's safe to destination
6837 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
6838 /// means there isn't a need to check it against alignment requirement,
6839 /// probably because the source does not need to be loaded. If 'IsMemset' is
6840 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
6841 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
6842 /// source is constant so it does not need to be loaded.
6843 /// It returns EVT::Other if the type should be determined using generic
6844 /// target-independent logic.
6845 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
6846 unsigned DstAlign, unsigned SrcAlign,
6847 bool IsMemset, bool ZeroMemset,
6849 MachineFunction &MF) const {
6850 if (this->PPCSubTarget.isPPC64()) {
6857 bool PPCTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
6859 if (DisablePPCUnaligned)
6862 // PowerPC supports unaligned memory access for simple non-vector types.
6863 // Although accessing unaligned addresses is not as efficient as accessing
6864 // aligned addresses, it is generally more efficient than manual expansion,
6865 // and generally only traps for software emulation when crossing page
6871 if (VT.getSimpleVT().isVector())
6874 if (VT == MVT::ppcf128)
6883 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
6884 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
6885 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
6886 /// is expanded to mul + add.
6887 bool PPCTargetLowering::isFMAFasterThanMulAndAdd(EVT VT) const {
6891 switch (VT.getSimpleVT().SimpleTy) {
6903 Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const {
6905 return TargetLowering::getSchedulingPreference(N);