1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCMachineFunctionInfo.h"
16 #include "PPCPerfectShuffle.h"
17 #include "PPCPredicates.h"
18 #include "PPCTargetMachine.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/VectorExtras.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
29 #include "llvm/CallingConv.h"
30 #include "llvm/Constants.h"
31 #include "llvm/Function.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/ErrorHandling.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/DerivedTypes.h"
41 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
42 CCValAssign::LocInfo &LocInfo,
43 ISD::ArgFlagsTy &ArgFlags,
45 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
47 CCValAssign::LocInfo &LocInfo,
48 ISD::ArgFlagsTy &ArgFlags,
50 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
52 CCValAssign::LocInfo &LocInfo,
53 ISD::ArgFlagsTy &ArgFlags,
56 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
57 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
60 static TargetLoweringObjectFile *CreateTLOF(const PPCTargetMachine &TM) {
61 if (TM.getSubtargetImpl()->isDarwin())
62 return new TargetLoweringObjectFileMachO();
64 return new TargetLoweringObjectFileELF();
67 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
68 : TargetLowering(TM, CreateTLOF(TM)), PPCSubTarget(*TM.getSubtargetImpl()) {
72 // Use _setjmp/_longjmp instead of setjmp/longjmp.
73 setUseUnderscoreSetJmp(true);
74 setUseUnderscoreLongJmp(true);
76 // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all
77 // arguments are at least 4/8 bytes aligned.
78 setMinStackArgumentAlignment(TM.getSubtarget<PPCSubtarget>().isPPC64() ? 8:4);
80 // Set up the register classes.
81 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
82 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
83 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
85 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
86 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
87 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
89 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
91 // PowerPC has pre-inc load and store's.
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
93 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
94 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
95 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
96 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
98 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
99 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
100 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
101 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
103 // This is used in the ppcf128->int sequence. Note it has different semantics
104 // from FP_ROUND: that rounds to nearest, this rounds to zero.
105 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
107 // PowerPC has no SREM/UREM instructions
108 setOperationAction(ISD::SREM, MVT::i32, Expand);
109 setOperationAction(ISD::UREM, MVT::i32, Expand);
110 setOperationAction(ISD::SREM, MVT::i64, Expand);
111 setOperationAction(ISD::UREM, MVT::i64, Expand);
113 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
114 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
115 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
116 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
117 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
118 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
119 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
120 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
121 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
123 // We don't support sin/cos/sqrt/fmod/pow
124 setOperationAction(ISD::FSIN , MVT::f64, Expand);
125 setOperationAction(ISD::FCOS , MVT::f64, Expand);
126 setOperationAction(ISD::FREM , MVT::f64, Expand);
127 setOperationAction(ISD::FPOW , MVT::f64, Expand);
128 setOperationAction(ISD::FSIN , MVT::f32, Expand);
129 setOperationAction(ISD::FCOS , MVT::f32, Expand);
130 setOperationAction(ISD::FREM , MVT::f32, Expand);
131 setOperationAction(ISD::FPOW , MVT::f32, Expand);
133 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
135 // If we're enabling GP optimizations, use hardware square root
136 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
137 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
138 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
141 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
142 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
144 // PowerPC does not have BSWAP, CTPOP or CTTZ
145 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
146 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
147 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
148 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
149 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
150 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
152 // PowerPC does not have ROTR
153 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
154 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
156 // PowerPC does not have Select
157 setOperationAction(ISD::SELECT, MVT::i32, Expand);
158 setOperationAction(ISD::SELECT, MVT::i64, Expand);
159 setOperationAction(ISD::SELECT, MVT::f32, Expand);
160 setOperationAction(ISD::SELECT, MVT::f64, Expand);
162 // PowerPC wants to turn select_cc of FP into fsel when possible.
163 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
164 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
166 // PowerPC wants to optimize integer setcc a bit
167 setOperationAction(ISD::SETCC, MVT::i32, Custom);
169 // PowerPC does not have BRCOND which requires SetCC
170 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
172 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
174 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
175 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
177 // PowerPC does not have [U|S]INT_TO_FP
178 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
179 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
181 setOperationAction(ISD::BITCAST, MVT::f32, Expand);
182 setOperationAction(ISD::BITCAST, MVT::i32, Expand);
183 setOperationAction(ISD::BITCAST, MVT::i64, Expand);
184 setOperationAction(ISD::BITCAST, MVT::f64, Expand);
186 // We cannot sextinreg(i1). Expand to shifts.
187 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
189 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
190 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
191 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
192 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
195 // We want to legalize GlobalAddress and ConstantPool nodes into the
196 // appropriate instructions to materialize the address.
197 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
198 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
199 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
200 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
201 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
202 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
203 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
204 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
205 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
206 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
209 setOperationAction(ISD::TRAP, MVT::Other, Legal);
211 // TRAMPOLINE is custom lowered.
212 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
214 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
215 setOperationAction(ISD::VASTART , MVT::Other, Custom);
217 // VAARG is custom lowered with the 32-bit SVR4 ABI.
218 if ( TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
219 && !TM.getSubtarget<PPCSubtarget>().isPPC64())
220 setOperationAction(ISD::VAARG, MVT::Other, Custom);
222 setOperationAction(ISD::VAARG, MVT::Other, Expand);
224 // Use the default implementation.
225 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
226 setOperationAction(ISD::VAEND , MVT::Other, Expand);
227 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
228 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
229 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
230 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
232 // We want to custom lower some of our intrinsics.
233 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
235 // Comparisons that require checking two conditions.
236 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
237 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
238 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
239 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
240 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
241 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
242 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
243 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
244 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
245 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
246 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
247 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
249 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
250 // They also have instructions for converting between i64 and fp.
251 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
252 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
253 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
254 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
255 // This is just the low 32 bits of a (signed) fp->i64 conversion.
256 // We cannot do this with Promote because i64 is not a legal type.
257 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
259 // FIXME: disable this lowered code. This generates 64-bit register values,
260 // and we don't model the fact that the top part is clobbered by calls. We
261 // need to flag these together so that the value isn't live across a call.
262 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
264 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
265 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
268 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
269 // 64-bit PowerPC implementations can support i64 types directly
270 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
271 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
272 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
273 // 64-bit PowerPC wants to expand i128 shifts itself.
274 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
275 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
276 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
278 // 32-bit PowerPC wants to expand i64 shifts itself.
279 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
280 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
281 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
284 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
285 // First set operation action for all vector types to expand. Then we
286 // will selectively turn on ones that can be effectively codegen'd.
287 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
288 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
289 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
291 // add/sub are legal for all supported vector VT's.
292 setOperationAction(ISD::ADD , VT, Legal);
293 setOperationAction(ISD::SUB , VT, Legal);
295 // We promote all shuffles to v16i8.
296 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
297 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
299 // We promote all non-typed operations to v4i32.
300 setOperationAction(ISD::AND , VT, Promote);
301 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
302 setOperationAction(ISD::OR , VT, Promote);
303 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
304 setOperationAction(ISD::XOR , VT, Promote);
305 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
306 setOperationAction(ISD::LOAD , VT, Promote);
307 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
308 setOperationAction(ISD::SELECT, VT, Promote);
309 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
310 setOperationAction(ISD::STORE, VT, Promote);
311 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
313 // No other operations are legal.
314 setOperationAction(ISD::MUL , VT, Expand);
315 setOperationAction(ISD::SDIV, VT, Expand);
316 setOperationAction(ISD::SREM, VT, Expand);
317 setOperationAction(ISD::UDIV, VT, Expand);
318 setOperationAction(ISD::UREM, VT, Expand);
319 setOperationAction(ISD::FDIV, VT, Expand);
320 setOperationAction(ISD::FNEG, VT, Expand);
321 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
322 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
323 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
324 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
325 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
326 setOperationAction(ISD::UDIVREM, VT, Expand);
327 setOperationAction(ISD::SDIVREM, VT, Expand);
328 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
329 setOperationAction(ISD::FPOW, VT, Expand);
330 setOperationAction(ISD::CTPOP, VT, Expand);
331 setOperationAction(ISD::CTLZ, VT, Expand);
332 setOperationAction(ISD::CTTZ, VT, Expand);
335 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
336 // with merges, splats, etc.
337 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
339 setOperationAction(ISD::AND , MVT::v4i32, Legal);
340 setOperationAction(ISD::OR , MVT::v4i32, Legal);
341 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
342 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
343 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
344 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
346 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
347 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
348 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
349 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
351 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
352 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
353 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
354 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
356 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
357 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
359 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
360 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
361 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
362 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
365 setBooleanContents(ZeroOrOneBooleanContent);
367 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
368 setStackPointerRegisterToSaveRestore(PPC::X1);
369 setExceptionPointerRegister(PPC::X3);
370 setExceptionSelectorRegister(PPC::X4);
372 setStackPointerRegisterToSaveRestore(PPC::R1);
373 setExceptionPointerRegister(PPC::R3);
374 setExceptionSelectorRegister(PPC::R4);
377 // We have target-specific dag combine patterns for the following nodes:
378 setTargetDAGCombine(ISD::SINT_TO_FP);
379 setTargetDAGCombine(ISD::STORE);
380 setTargetDAGCombine(ISD::BR_CC);
381 setTargetDAGCombine(ISD::BSWAP);
383 // Darwin long double math library functions have $LDBL128 appended.
384 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
385 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
386 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
387 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
388 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
389 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
390 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
391 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
392 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
393 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
394 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
397 computeRegisterProperties();
400 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
401 /// function arguments in the caller parameter area.
402 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
403 const TargetMachine &TM = getTargetMachine();
404 // Darwin passes everything on 4 byte boundary.
405 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
411 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
414 case PPCISD::FSEL: return "PPCISD::FSEL";
415 case PPCISD::FCFID: return "PPCISD::FCFID";
416 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
417 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
418 case PPCISD::STFIWX: return "PPCISD::STFIWX";
419 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
420 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
421 case PPCISD::VPERM: return "PPCISD::VPERM";
422 case PPCISD::Hi: return "PPCISD::Hi";
423 case PPCISD::Lo: return "PPCISD::Lo";
424 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
425 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
426 case PPCISD::LOAD: return "PPCISD::LOAD";
427 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
428 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
429 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
430 case PPCISD::SRL: return "PPCISD::SRL";
431 case PPCISD::SRA: return "PPCISD::SRA";
432 case PPCISD::SHL: return "PPCISD::SHL";
433 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
434 case PPCISD::STD_32: return "PPCISD::STD_32";
435 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
436 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
437 case PPCISD::NOP: return "PPCISD::NOP";
438 case PPCISD::MTCTR: return "PPCISD::MTCTR";
439 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
440 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
441 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
442 case PPCISD::MFCR: return "PPCISD::MFCR";
443 case PPCISD::VCMP: return "PPCISD::VCMP";
444 case PPCISD::VCMPo: return "PPCISD::VCMPo";
445 case PPCISD::LBRX: return "PPCISD::LBRX";
446 case PPCISD::STBRX: return "PPCISD::STBRX";
447 case PPCISD::LARX: return "PPCISD::LARX";
448 case PPCISD::STCX: return "PPCISD::STCX";
449 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
450 case PPCISD::MFFS: return "PPCISD::MFFS";
451 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
452 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
453 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
454 case PPCISD::MTFSF: return "PPCISD::MTFSF";
455 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
459 MVT::SimpleValueType PPCTargetLowering::getSetCCResultType(EVT VT) const {
463 /// getFunctionAlignment - Return the Log2 alignment of this function.
464 unsigned PPCTargetLowering::getFunctionAlignment(const Function *F) const {
465 if (getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin())
466 return F->hasFnAttr(Attribute::OptimizeForSize) ? 2 : 4;
471 //===----------------------------------------------------------------------===//
472 // Node matching predicates, for use by the tblgen matching code.
473 //===----------------------------------------------------------------------===//
475 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
476 static bool isFloatingPointZero(SDValue Op) {
477 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
478 return CFP->getValueAPF().isZero();
479 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
480 // Maybe this has already been legalized into the constant pool?
481 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
482 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
483 return CFP->getValueAPF().isZero();
488 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
489 /// true if Op is undef or if it matches the specified value.
490 static bool isConstantOrUndef(int Op, int Val) {
491 return Op < 0 || Op == Val;
494 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
495 /// VPKUHUM instruction.
496 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
498 for (unsigned i = 0; i != 16; ++i)
499 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
502 for (unsigned i = 0; i != 8; ++i)
503 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
504 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
510 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
511 /// VPKUWUM instruction.
512 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
514 for (unsigned i = 0; i != 16; i += 2)
515 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
516 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
519 for (unsigned i = 0; i != 8; i += 2)
520 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
521 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
522 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
523 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
529 /// isVMerge - Common function, used to match vmrg* shuffles.
531 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
532 unsigned LHSStart, unsigned RHSStart) {
533 assert(N->getValueType(0) == MVT::v16i8 &&
534 "PPC only supports shuffles by bytes!");
535 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
536 "Unsupported merge size!");
538 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
539 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
540 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
541 LHSStart+j+i*UnitSize) ||
542 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
543 RHSStart+j+i*UnitSize))
549 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
550 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
551 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
554 return isVMerge(N, UnitSize, 8, 24);
555 return isVMerge(N, UnitSize, 8, 8);
558 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
559 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
560 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
563 return isVMerge(N, UnitSize, 0, 16);
564 return isVMerge(N, UnitSize, 0, 0);
568 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
569 /// amount, otherwise return -1.
570 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
571 assert(N->getValueType(0) == MVT::v16i8 &&
572 "PPC only supports shuffles by bytes!");
574 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
576 // Find the first non-undef value in the shuffle mask.
578 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
581 if (i == 16) return -1; // all undef.
583 // Otherwise, check to see if the rest of the elements are consecutively
584 // numbered from this value.
585 unsigned ShiftAmt = SVOp->getMaskElt(i);
586 if (ShiftAmt < i) return -1;
590 // Check the rest of the elements to see if they are consecutive.
591 for (++i; i != 16; ++i)
592 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
595 // Check the rest of the elements to see if they are consecutive.
596 for (++i; i != 16; ++i)
597 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
603 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
604 /// specifies a splat of a single element that is suitable for input to
605 /// VSPLTB/VSPLTH/VSPLTW.
606 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
607 assert(N->getValueType(0) == MVT::v16i8 &&
608 (EltSize == 1 || EltSize == 2 || EltSize == 4));
610 // This is a splat operation if each element of the permute is the same, and
611 // if the value doesn't reference the second vector.
612 unsigned ElementBase = N->getMaskElt(0);
614 // FIXME: Handle UNDEF elements too!
615 if (ElementBase >= 16)
618 // Check that the indices are consecutive, in the case of a multi-byte element
619 // splatted with a v16i8 mask.
620 for (unsigned i = 1; i != EltSize; ++i)
621 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
624 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
625 if (N->getMaskElt(i) < 0) continue;
626 for (unsigned j = 0; j != EltSize; ++j)
627 if (N->getMaskElt(i+j) != N->getMaskElt(j))
633 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
635 bool PPC::isAllNegativeZeroVector(SDNode *N) {
636 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
638 APInt APVal, APUndef;
642 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
643 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
644 return CFP->getValueAPF().isNegZero();
649 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
650 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
651 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
652 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
653 assert(isSplatShuffleMask(SVOp, EltSize));
654 return SVOp->getMaskElt(0) / EltSize;
657 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
658 /// by using a vspltis[bhw] instruction of the specified element size, return
659 /// the constant being splatted. The ByteSize field indicates the number of
660 /// bytes of each element [124] -> [bhw].
661 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
664 // If ByteSize of the splat is bigger than the element size of the
665 // build_vector, then we have a case where we are checking for a splat where
666 // multiple elements of the buildvector are folded together into a single
667 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
668 unsigned EltSize = 16/N->getNumOperands();
669 if (EltSize < ByteSize) {
670 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
671 SDValue UniquedVals[4];
672 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
674 // See if all of the elements in the buildvector agree across.
675 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
676 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
677 // If the element isn't a constant, bail fully out.
678 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
681 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
682 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
683 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
684 return SDValue(); // no match.
687 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
688 // either constant or undef values that are identical for each chunk. See
689 // if these chunks can form into a larger vspltis*.
691 // Check to see if all of the leading entries are either 0 or -1. If
692 // neither, then this won't fit into the immediate field.
693 bool LeadingZero = true;
694 bool LeadingOnes = true;
695 for (unsigned i = 0; i != Multiple-1; ++i) {
696 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
698 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
699 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
701 // Finally, check the least significant entry.
703 if (UniquedVals[Multiple-1].getNode() == 0)
704 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
705 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
707 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
710 if (UniquedVals[Multiple-1].getNode() == 0)
711 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
712 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
713 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
714 return DAG.getTargetConstant(Val, MVT::i32);
720 // Check to see if this buildvec has a single non-undef value in its elements.
721 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
722 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
723 if (OpVal.getNode() == 0)
724 OpVal = N->getOperand(i);
725 else if (OpVal != N->getOperand(i))
729 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
731 unsigned ValSizeInBytes = EltSize;
733 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
734 Value = CN->getZExtValue();
735 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
736 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
737 Value = FloatToBits(CN->getValueAPF().convertToFloat());
740 // If the splat value is larger than the element value, then we can never do
741 // this splat. The only case that we could fit the replicated bits into our
742 // immediate field for would be zero, and we prefer to use vxor for it.
743 if (ValSizeInBytes < ByteSize) return SDValue();
745 // If the element value is larger than the splat value, cut it in half and
746 // check to see if the two halves are equal. Continue doing this until we
747 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
748 while (ValSizeInBytes > ByteSize) {
749 ValSizeInBytes >>= 1;
751 // If the top half equals the bottom half, we're still ok.
752 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
753 (Value & ((1 << (8*ValSizeInBytes))-1)))
757 // Properly sign extend the value.
758 int ShAmt = (4-ByteSize)*8;
759 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
761 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
762 if (MaskVal == 0) return SDValue();
764 // Finally, if this value fits in a 5 bit sext field, return it
765 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
766 return DAG.getTargetConstant(MaskVal, MVT::i32);
770 //===----------------------------------------------------------------------===//
771 // Addressing Mode Selection
772 //===----------------------------------------------------------------------===//
774 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
775 /// or 64-bit immediate, and if the value can be accurately represented as a
776 /// sign extension from a 16-bit value. If so, this returns true and the
778 static bool isIntS16Immediate(SDNode *N, short &Imm) {
779 if (N->getOpcode() != ISD::Constant)
782 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
783 if (N->getValueType(0) == MVT::i32)
784 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
786 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
788 static bool isIntS16Immediate(SDValue Op, short &Imm) {
789 return isIntS16Immediate(Op.getNode(), Imm);
793 /// SelectAddressRegReg - Given the specified addressed, check to see if it
794 /// can be represented as an indexed [r+r] operation. Returns false if it
795 /// can be more efficiently represented with [r+imm].
796 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
798 SelectionDAG &DAG) const {
800 if (N.getOpcode() == ISD::ADD) {
801 if (isIntS16Immediate(N.getOperand(1), imm))
803 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
806 Base = N.getOperand(0);
807 Index = N.getOperand(1);
809 } else if (N.getOpcode() == ISD::OR) {
810 if (isIntS16Immediate(N.getOperand(1), imm))
811 return false; // r+i can fold it if we can.
813 // If this is an or of disjoint bitfields, we can codegen this as an add
814 // (for better address arithmetic) if the LHS and RHS of the OR are provably
816 APInt LHSKnownZero, LHSKnownOne;
817 APInt RHSKnownZero, RHSKnownOne;
818 DAG.ComputeMaskedBits(N.getOperand(0),
819 APInt::getAllOnesValue(N.getOperand(0)
820 .getValueSizeInBits()),
821 LHSKnownZero, LHSKnownOne);
823 if (LHSKnownZero.getBoolValue()) {
824 DAG.ComputeMaskedBits(N.getOperand(1),
825 APInt::getAllOnesValue(N.getOperand(1)
826 .getValueSizeInBits()),
827 RHSKnownZero, RHSKnownOne);
828 // If all of the bits are known zero on the LHS or RHS, the add won't
830 if (~(LHSKnownZero | RHSKnownZero) == 0) {
831 Base = N.getOperand(0);
832 Index = N.getOperand(1);
841 /// Returns true if the address N can be represented by a base register plus
842 /// a signed 16-bit displacement [r+imm], and if it is not better
843 /// represented as reg+reg.
844 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
846 SelectionDAG &DAG) const {
847 // FIXME dl should come from parent load or store, not from address
848 DebugLoc dl = N.getDebugLoc();
849 // If this can be more profitably realized as r+r, fail.
850 if (SelectAddressRegReg(N, Disp, Base, DAG))
853 if (N.getOpcode() == ISD::ADD) {
855 if (isIntS16Immediate(N.getOperand(1), imm)) {
856 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
857 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
858 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
860 Base = N.getOperand(0);
862 return true; // [r+i]
863 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
864 // Match LOAD (ADD (X, Lo(G))).
865 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
866 && "Cannot handle constant offsets yet!");
867 Disp = N.getOperand(1).getOperand(0); // The global address.
868 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
869 Disp.getOpcode() == ISD::TargetConstantPool ||
870 Disp.getOpcode() == ISD::TargetJumpTable);
871 Base = N.getOperand(0);
872 return true; // [&g+r]
874 } else if (N.getOpcode() == ISD::OR) {
876 if (isIntS16Immediate(N.getOperand(1), imm)) {
877 // If this is an or of disjoint bitfields, we can codegen this as an add
878 // (for better address arithmetic) if the LHS and RHS of the OR are
879 // provably disjoint.
880 APInt LHSKnownZero, LHSKnownOne;
881 DAG.ComputeMaskedBits(N.getOperand(0),
882 APInt::getAllOnesValue(N.getOperand(0)
883 .getValueSizeInBits()),
884 LHSKnownZero, LHSKnownOne);
886 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
887 // If all of the bits are known zero on the LHS or RHS, the add won't
889 Base = N.getOperand(0);
890 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
894 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
895 // Loading from a constant address.
897 // If this address fits entirely in a 16-bit sext immediate field, codegen
900 if (isIntS16Immediate(CN, Imm)) {
901 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
902 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
906 // Handle 32-bit sext immediates with LIS + addr mode.
907 if (CN->getValueType(0) == MVT::i32 ||
908 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
909 int Addr = (int)CN->getZExtValue();
911 // Otherwise, break this down into an LIS + disp.
912 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
914 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
915 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
916 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
921 Disp = DAG.getTargetConstant(0, getPointerTy());
922 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
923 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
926 return true; // [r+0]
929 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
930 /// represented as an indexed [r+r] operation.
931 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
933 SelectionDAG &DAG) const {
934 // Check to see if we can easily represent this as an [r+r] address. This
935 // will fail if it thinks that the address is more profitably represented as
936 // reg+imm, e.g. where imm = 0.
937 if (SelectAddressRegReg(N, Base, Index, DAG))
940 // If the operand is an addition, always emit this as [r+r], since this is
941 // better (for code size, and execution, as the memop does the add for free)
942 // than emitting an explicit add.
943 if (N.getOpcode() == ISD::ADD) {
944 Base = N.getOperand(0);
945 Index = N.getOperand(1);
949 // Otherwise, do it the hard way, using R0 as the base register.
950 Base = DAG.getRegister(PPC::R0, N.getValueType());
955 /// SelectAddressRegImmShift - Returns true if the address N can be
956 /// represented by a base register plus a signed 14-bit displacement
957 /// [r+imm*4]. Suitable for use by STD and friends.
958 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
960 SelectionDAG &DAG) const {
961 // FIXME dl should come from the parent load or store, not the address
962 DebugLoc dl = N.getDebugLoc();
963 // If this can be more profitably realized as r+r, fail.
964 if (SelectAddressRegReg(N, Disp, Base, DAG))
967 if (N.getOpcode() == ISD::ADD) {
969 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
970 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
971 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
972 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
974 Base = N.getOperand(0);
976 return true; // [r+i]
977 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
978 // Match LOAD (ADD (X, Lo(G))).
979 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
980 && "Cannot handle constant offsets yet!");
981 Disp = N.getOperand(1).getOperand(0); // The global address.
982 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
983 Disp.getOpcode() == ISD::TargetConstantPool ||
984 Disp.getOpcode() == ISD::TargetJumpTable);
985 Base = N.getOperand(0);
986 return true; // [&g+r]
988 } else if (N.getOpcode() == ISD::OR) {
990 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
991 // If this is an or of disjoint bitfields, we can codegen this as an add
992 // (for better address arithmetic) if the LHS and RHS of the OR are
993 // provably disjoint.
994 APInt LHSKnownZero, LHSKnownOne;
995 DAG.ComputeMaskedBits(N.getOperand(0),
996 APInt::getAllOnesValue(N.getOperand(0)
997 .getValueSizeInBits()),
998 LHSKnownZero, LHSKnownOne);
999 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
1000 // If all of the bits are known zero on the LHS or RHS, the add won't
1002 Base = N.getOperand(0);
1003 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1007 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1008 // Loading from a constant address. Verify low two bits are clear.
1009 if ((CN->getZExtValue() & 3) == 0) {
1010 // If this address fits entirely in a 14-bit sext immediate field, codegen
1013 if (isIntS16Immediate(CN, Imm)) {
1014 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1015 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
1019 // Fold the low-part of 32-bit absolute addresses into addr mode.
1020 if (CN->getValueType(0) == MVT::i32 ||
1021 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1022 int Addr = (int)CN->getZExtValue();
1024 // Otherwise, break this down into an LIS + disp.
1025 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1026 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1027 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1028 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1034 Disp = DAG.getTargetConstant(0, getPointerTy());
1035 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1036 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1039 return true; // [r+0]
1043 /// getPreIndexedAddressParts - returns true by value, base pointer and
1044 /// offset pointer and addressing mode by reference if the node's address
1045 /// can be legally represented as pre-indexed load / store address.
1046 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1048 ISD::MemIndexedMode &AM,
1049 SelectionDAG &DAG) const {
1050 // Disabled by default for now.
1051 if (!EnablePPCPreinc) return false;
1055 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1056 Ptr = LD->getBasePtr();
1057 VT = LD->getMemoryVT();
1059 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1060 Ptr = ST->getBasePtr();
1061 VT = ST->getMemoryVT();
1065 // PowerPC doesn't have preinc load/store instructions for vectors.
1069 // TODO: Check reg+reg first.
1071 // LDU/STU use reg+imm*4, others use reg+imm.
1072 if (VT != MVT::i64) {
1074 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1078 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1082 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1083 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1084 // sext i32 to i64 when addr mode is r+i.
1085 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1086 LD->getExtensionType() == ISD::SEXTLOAD &&
1087 isa<ConstantSDNode>(Offset))
1095 //===----------------------------------------------------------------------===//
1096 // LowerOperation implementation
1097 //===----------------------------------------------------------------------===//
1099 /// GetLabelAccessInfo - Return true if we should reference labels using a
1100 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1101 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1102 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1103 HiOpFlags = PPCII::MO_HA16;
1104 LoOpFlags = PPCII::MO_LO16;
1106 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1107 // non-darwin platform. We don't support PIC on other platforms yet.
1108 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1109 TM.getSubtarget<PPCSubtarget>().isDarwin();
1111 HiOpFlags |= PPCII::MO_PIC_FLAG;
1112 LoOpFlags |= PPCII::MO_PIC_FLAG;
1115 // If this is a reference to a global value that requires a non-lazy-ptr, make
1116 // sure that instruction lowering adds it.
1117 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1118 HiOpFlags |= PPCII::MO_NLP_FLAG;
1119 LoOpFlags |= PPCII::MO_NLP_FLAG;
1121 if (GV->hasHiddenVisibility()) {
1122 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1123 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1130 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1131 SelectionDAG &DAG) {
1132 EVT PtrVT = HiPart.getValueType();
1133 SDValue Zero = DAG.getConstant(0, PtrVT);
1134 DebugLoc DL = HiPart.getDebugLoc();
1136 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1137 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1139 // With PIC, the first instruction is actually "GR+hi(&G)".
1141 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1142 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1144 // Generate non-pic code that has direct accesses to the constant pool.
1145 // The address of the global is just (hi(&g)+lo(&g)).
1146 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1149 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1150 SelectionDAG &DAG) const {
1151 EVT PtrVT = Op.getValueType();
1152 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1153 const Constant *C = CP->getConstVal();
1155 unsigned MOHiFlag, MOLoFlag;
1156 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1158 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1160 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1161 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1164 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1165 EVT PtrVT = Op.getValueType();
1166 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1168 unsigned MOHiFlag, MOLoFlag;
1169 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1170 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1171 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1172 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1175 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1176 SelectionDAG &DAG) const {
1177 EVT PtrVT = Op.getValueType();
1179 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1181 unsigned MOHiFlag, MOLoFlag;
1182 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1183 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1184 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1185 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1188 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1189 SelectionDAG &DAG) const {
1190 EVT PtrVT = Op.getValueType();
1191 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1192 DebugLoc DL = GSDN->getDebugLoc();
1193 const GlobalValue *GV = GSDN->getGlobal();
1195 // 64-bit SVR4 ABI code is always position-independent.
1196 // The actual address of the GlobalValue is stored in the TOC.
1197 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1198 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1199 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1200 DAG.getRegister(PPC::X2, MVT::i64));
1203 unsigned MOHiFlag, MOLoFlag;
1204 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1207 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1209 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1211 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1213 // If the global reference is actually to a non-lazy-pointer, we have to do an
1214 // extra load to get the address of the global.
1215 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1216 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1221 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1222 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1223 DebugLoc dl = Op.getDebugLoc();
1225 // If we're comparing for equality to zero, expose the fact that this is
1226 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1227 // fold the new nodes.
1228 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1229 if (C->isNullValue() && CC == ISD::SETEQ) {
1230 EVT VT = Op.getOperand(0).getValueType();
1231 SDValue Zext = Op.getOperand(0);
1232 if (VT.bitsLT(MVT::i32)) {
1234 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1236 unsigned Log2b = Log2_32(VT.getSizeInBits());
1237 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1238 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1239 DAG.getConstant(Log2b, MVT::i32));
1240 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1242 // Leave comparisons against 0 and -1 alone for now, since they're usually
1243 // optimized. FIXME: revisit this when we can custom lower all setcc
1245 if (C->isAllOnesValue() || C->isNullValue())
1249 // If we have an integer seteq/setne, turn it into a compare against zero
1250 // by xor'ing the rhs with the lhs, which is faster than setting a
1251 // condition register, reading it back out, and masking the correct bit. The
1252 // normal approach here uses sub to do this instead of xor. Using xor exposes
1253 // the result to other bit-twiddling opportunities.
1254 EVT LHSVT = Op.getOperand(0).getValueType();
1255 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1256 EVT VT = Op.getValueType();
1257 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1259 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1264 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1265 const PPCSubtarget &Subtarget) const {
1267 llvm_unreachable("VAARG not yet implemented for the SVR4 ABI!");
1268 return SDValue(); // Not reached
1271 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op,
1272 SelectionDAG &DAG) const {
1273 SDValue Chain = Op.getOperand(0);
1274 SDValue Trmp = Op.getOperand(1); // trampoline
1275 SDValue FPtr = Op.getOperand(2); // nested function
1276 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1277 DebugLoc dl = Op.getDebugLoc();
1279 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1280 bool isPPC64 = (PtrVT == MVT::i64);
1281 const Type *IntPtrTy =
1282 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1285 TargetLowering::ArgListTy Args;
1286 TargetLowering::ArgListEntry Entry;
1288 Entry.Ty = IntPtrTy;
1289 Entry.Node = Trmp; Args.push_back(Entry);
1291 // TrampSize == (isPPC64 ? 48 : 40);
1292 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1293 isPPC64 ? MVT::i64 : MVT::i32);
1294 Args.push_back(Entry);
1296 Entry.Node = FPtr; Args.push_back(Entry);
1297 Entry.Node = Nest; Args.push_back(Entry);
1299 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1300 std::pair<SDValue, SDValue> CallResult =
1301 LowerCallTo(Chain, Op.getValueType().getTypeForEVT(*DAG.getContext()),
1302 false, false, false, false, 0, CallingConv::C, false,
1303 /*isReturnValueUsed=*/true,
1304 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1308 { CallResult.first, CallResult.second };
1310 return DAG.getMergeValues(Ops, 2, dl);
1313 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1314 const PPCSubtarget &Subtarget) const {
1315 MachineFunction &MF = DAG.getMachineFunction();
1316 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1318 DebugLoc dl = Op.getDebugLoc();
1320 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1321 // vastart just stores the address of the VarArgsFrameIndex slot into the
1322 // memory location argument.
1323 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1324 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1325 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1326 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1327 MachinePointerInfo(SV),
1331 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1332 // We suppose the given va_list is already allocated.
1335 // char gpr; /* index into the array of 8 GPRs
1336 // * stored in the register save area
1337 // * gpr=0 corresponds to r3,
1338 // * gpr=1 to r4, etc.
1340 // char fpr; /* index into the array of 8 FPRs
1341 // * stored in the register save area
1342 // * fpr=0 corresponds to f1,
1343 // * fpr=1 to f2, etc.
1345 // char *overflow_arg_area;
1346 // /* location on stack that holds
1347 // * the next overflow argument
1349 // char *reg_save_area;
1350 // /* where r3:r10 and f1:f8 (if saved)
1356 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1357 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1360 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1362 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1364 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1367 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1368 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1370 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1371 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1373 uint64_t FPROffset = 1;
1374 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1376 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1378 // Store first byte : number of int regs
1379 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1381 MachinePointerInfo(SV),
1382 MVT::i8, false, false, 0);
1383 uint64_t nextOffset = FPROffset;
1384 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1387 // Store second byte : number of float regs
1388 SDValue secondStore =
1389 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1390 MachinePointerInfo(SV, nextOffset), MVT::i8,
1392 nextOffset += StackOffset;
1393 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1395 // Store second word : arguments given on stack
1396 SDValue thirdStore =
1397 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1398 MachinePointerInfo(SV, nextOffset),
1400 nextOffset += FrameOffset;
1401 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1403 // Store third word : arguments given in registers
1404 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1405 MachinePointerInfo(SV, nextOffset),
1410 #include "PPCGenCallingConv.inc"
1412 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1413 CCValAssign::LocInfo &LocInfo,
1414 ISD::ArgFlagsTy &ArgFlags,
1419 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1421 CCValAssign::LocInfo &LocInfo,
1422 ISD::ArgFlagsTy &ArgFlags,
1424 static const unsigned ArgRegs[] = {
1425 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1426 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1428 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1430 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1432 // Skip one register if the first unallocated register has an even register
1433 // number and there are still argument registers available which have not been
1434 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1435 // need to skip a register if RegNum is odd.
1436 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1437 State.AllocateReg(ArgRegs[RegNum]);
1440 // Always return false here, as this function only makes sure that the first
1441 // unallocated register has an odd register number and does not actually
1442 // allocate a register for the current argument.
1446 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1448 CCValAssign::LocInfo &LocInfo,
1449 ISD::ArgFlagsTy &ArgFlags,
1451 static const unsigned ArgRegs[] = {
1452 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1456 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1458 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1460 // If there is only one Floating-point register left we need to put both f64
1461 // values of a split ppc_fp128 value on the stack.
1462 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1463 State.AllocateReg(ArgRegs[RegNum]);
1466 // Always return false here, as this function only makes sure that the two f64
1467 // values a ppc_fp128 value is split into are both passed in registers or both
1468 // passed on the stack and does not actually allocate a register for the
1469 // current argument.
1473 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1475 static const unsigned *GetFPR() {
1476 static const unsigned FPR[] = {
1477 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1478 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1484 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1486 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1487 unsigned PtrByteSize) {
1488 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1489 if (Flags.isByVal())
1490 ArgSize = Flags.getByValSize();
1491 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1497 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1498 CallingConv::ID CallConv, bool isVarArg,
1499 const SmallVectorImpl<ISD::InputArg>
1501 DebugLoc dl, SelectionDAG &DAG,
1502 SmallVectorImpl<SDValue> &InVals)
1504 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1505 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1508 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1514 PPCTargetLowering::LowerFormalArguments_SVR4(
1516 CallingConv::ID CallConv, bool isVarArg,
1517 const SmallVectorImpl<ISD::InputArg>
1519 DebugLoc dl, SelectionDAG &DAG,
1520 SmallVectorImpl<SDValue> &InVals) const {
1522 // 32-bit SVR4 ABI Stack Frame Layout:
1523 // +-----------------------------------+
1524 // +--> | Back chain |
1525 // | +-----------------------------------+
1526 // | | Floating-point register save area |
1527 // | +-----------------------------------+
1528 // | | General register save area |
1529 // | +-----------------------------------+
1530 // | | CR save word |
1531 // | +-----------------------------------+
1532 // | | VRSAVE save word |
1533 // | +-----------------------------------+
1534 // | | Alignment padding |
1535 // | +-----------------------------------+
1536 // | | Vector register save area |
1537 // | +-----------------------------------+
1538 // | | Local variable space |
1539 // | +-----------------------------------+
1540 // | | Parameter list area |
1541 // | +-----------------------------------+
1542 // | | LR save word |
1543 // | +-----------------------------------+
1544 // SP--> +--- | Back chain |
1545 // +-----------------------------------+
1548 // System V Application Binary Interface PowerPC Processor Supplement
1549 // AltiVec Technology Programming Interface Manual
1551 MachineFunction &MF = DAG.getMachineFunction();
1552 MachineFrameInfo *MFI = MF.getFrameInfo();
1553 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1555 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1556 // Potential tail calls could cause overwriting of argument stack slots.
1557 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1558 unsigned PtrByteSize = 4;
1560 // Assign locations to all of the incoming arguments.
1561 SmallVector<CCValAssign, 16> ArgLocs;
1562 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1565 // Reserve space for the linkage area on the stack.
1566 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1568 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1570 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1571 CCValAssign &VA = ArgLocs[i];
1573 // Arguments stored in registers.
1574 if (VA.isRegLoc()) {
1575 TargetRegisterClass *RC;
1576 EVT ValVT = VA.getValVT();
1578 switch (ValVT.getSimpleVT().SimpleTy) {
1580 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1582 RC = PPC::GPRCRegisterClass;
1585 RC = PPC::F4RCRegisterClass;
1588 RC = PPC::F8RCRegisterClass;
1594 RC = PPC::VRRCRegisterClass;
1598 // Transform the arguments stored in physical registers into virtual ones.
1599 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1600 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1602 InVals.push_back(ArgValue);
1604 // Argument stored in memory.
1605 assert(VA.isMemLoc());
1607 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1608 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1611 // Create load nodes to retrieve arguments from the stack.
1612 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1613 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1614 MachinePointerInfo(),
1619 // Assign locations to all of the incoming aggregate by value arguments.
1620 // Aggregates passed by value are stored in the local variable space of the
1621 // caller's stack frame, right above the parameter list area.
1622 SmallVector<CCValAssign, 16> ByValArgLocs;
1623 CCState CCByValInfo(CallConv, isVarArg, getTargetMachine(),
1624 ByValArgLocs, *DAG.getContext());
1626 // Reserve stack space for the allocations in CCInfo.
1627 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1629 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1631 // Area that is at least reserved in the caller of this function.
1632 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1634 // Set the size that is at least reserved in caller of this function. Tail
1635 // call optimized function's reserved stack space needs to be aligned so that
1636 // taking the difference between two stack areas will result in an aligned
1638 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1641 std::max(MinReservedArea,
1642 PPCFrameLowering::getMinCallFrameSize(false, false));
1644 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1645 getStackAlignment();
1646 unsigned AlignMask = TargetAlign-1;
1647 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1649 FI->setMinReservedArea(MinReservedArea);
1651 SmallVector<SDValue, 8> MemOps;
1653 // If the function takes variable number of arguments, make a frame index for
1654 // the start of the first vararg value... for expansion of llvm.va_start.
1656 static const unsigned GPArgRegs[] = {
1657 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1658 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1660 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1662 static const unsigned FPArgRegs[] = {
1663 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1666 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1668 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1670 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1673 // Make room for NumGPArgRegs and NumFPArgRegs.
1674 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1675 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1677 FuncInfo->setVarArgsStackOffset(
1678 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1679 CCInfo.getNextStackOffset(), true));
1681 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1682 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1684 // The fixed integer arguments of a variadic function are stored to the
1685 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1686 // the result of va_next.
1687 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1688 // Get an existing live-in vreg, or add a new one.
1689 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1691 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1693 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1694 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1695 MachinePointerInfo(), false, false, 0);
1696 MemOps.push_back(Store);
1697 // Increment the address by four for the next argument to store
1698 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1699 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1702 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1704 // The double arguments are stored to the VarArgsFrameIndex
1706 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1707 // Get an existing live-in vreg, or add a new one.
1708 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1710 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1712 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1713 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1714 MachinePointerInfo(), false, false, 0);
1715 MemOps.push_back(Store);
1716 // Increment the address by eight for the next argument to store
1717 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1719 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1723 if (!MemOps.empty())
1724 Chain = DAG.getNode(ISD::TokenFactor, dl,
1725 MVT::Other, &MemOps[0], MemOps.size());
1731 PPCTargetLowering::LowerFormalArguments_Darwin(
1733 CallingConv::ID CallConv, bool isVarArg,
1734 const SmallVectorImpl<ISD::InputArg>
1736 DebugLoc dl, SelectionDAG &DAG,
1737 SmallVectorImpl<SDValue> &InVals) const {
1738 // TODO: add description of PPC stack frame format, or at least some docs.
1740 MachineFunction &MF = DAG.getMachineFunction();
1741 MachineFrameInfo *MFI = MF.getFrameInfo();
1742 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1744 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1745 bool isPPC64 = PtrVT == MVT::i64;
1746 // Potential tail calls could cause overwriting of argument stack slots.
1747 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1748 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1750 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1751 // Area that is at least reserved in caller of this function.
1752 unsigned MinReservedArea = ArgOffset;
1754 static const unsigned GPR_32[] = { // 32-bit registers.
1755 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1756 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1758 static const unsigned GPR_64[] = { // 64-bit registers.
1759 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1760 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1763 static const unsigned *FPR = GetFPR();
1765 static const unsigned VR[] = {
1766 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1767 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1770 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1771 const unsigned Num_FPR_Regs = 13;
1772 const unsigned Num_VR_Regs = array_lengthof( VR);
1774 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1776 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1778 // In 32-bit non-varargs functions, the stack space for vectors is after the
1779 // stack space for non-vectors. We do not use this space unless we have
1780 // too many vectors to fit in registers, something that only occurs in
1781 // constructed examples:), but we have to walk the arglist to figure
1782 // that out...for the pathological case, compute VecArgOffset as the
1783 // start of the vector parameter area. Computing VecArgOffset is the
1784 // entire point of the following loop.
1785 unsigned VecArgOffset = ArgOffset;
1786 if (!isVarArg && !isPPC64) {
1787 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1789 EVT ObjectVT = Ins[ArgNo].VT;
1790 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1791 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1793 if (Flags.isByVal()) {
1794 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1795 ObjSize = Flags.getByValSize();
1797 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1798 VecArgOffset += ArgSize;
1802 switch(ObjectVT.getSimpleVT().SimpleTy) {
1803 default: llvm_unreachable("Unhandled argument type!");
1806 VecArgOffset += isPPC64 ? 8 : 4;
1808 case MVT::i64: // PPC64
1816 // Nothing to do, we're only looking at Nonvector args here.
1821 // We've found where the vector parameter area in memory is. Skip the
1822 // first 12 parameters; these don't use that memory.
1823 VecArgOffset = ((VecArgOffset+15)/16)*16;
1824 VecArgOffset += 12*16;
1826 // Add DAG nodes to load the arguments or copy them out of registers. On
1827 // entry to a function on PPC, the arguments start after the linkage area,
1828 // although the first ones are often in registers.
1830 SmallVector<SDValue, 8> MemOps;
1831 unsigned nAltivecParamsAtEnd = 0;
1832 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1834 bool needsLoad = false;
1835 EVT ObjectVT = Ins[ArgNo].VT;
1836 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1837 unsigned ArgSize = ObjSize;
1838 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1840 unsigned CurArgOffset = ArgOffset;
1842 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1843 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1844 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1845 if (isVarArg || isPPC64) {
1846 MinReservedArea = ((MinReservedArea+15)/16)*16;
1847 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1850 } else nAltivecParamsAtEnd++;
1852 // Calculate min reserved area.
1853 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1857 // FIXME the codegen can be much improved in some cases.
1858 // We do not have to keep everything in memory.
1859 if (Flags.isByVal()) {
1860 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1861 ObjSize = Flags.getByValSize();
1862 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1863 // Objects of size 1 and 2 are right justified, everything else is
1864 // left justified. This means the memory address is adjusted forwards.
1865 if (ObjSize==1 || ObjSize==2) {
1866 CurArgOffset = CurArgOffset + (4 - ObjSize);
1868 // The value of the object is its address.
1869 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
1870 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1871 InVals.push_back(FIN);
1872 if (ObjSize==1 || ObjSize==2) {
1873 if (GPR_idx != Num_GPR_Regs) {
1874 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1875 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1876 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1877 MachinePointerInfo(),
1878 ObjSize==1 ? MVT::i8 : MVT::i16,
1880 MemOps.push_back(Store);
1884 ArgOffset += PtrByteSize;
1888 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1889 // Store whatever pieces of the object are in registers
1890 // to memory. ArgVal will be address of the beginning of
1892 if (GPR_idx != Num_GPR_Regs) {
1893 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1894 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
1895 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1896 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1897 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1898 MachinePointerInfo(),
1900 MemOps.push_back(Store);
1902 ArgOffset += PtrByteSize;
1904 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1911 switch (ObjectVT.getSimpleVT().SimpleTy) {
1912 default: llvm_unreachable("Unhandled argument type!");
1915 if (GPR_idx != Num_GPR_Regs) {
1916 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1917 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1921 ArgSize = PtrByteSize;
1923 // All int arguments reserve stack space in the Darwin ABI.
1924 ArgOffset += PtrByteSize;
1928 case MVT::i64: // PPC64
1929 if (GPR_idx != Num_GPR_Regs) {
1930 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1931 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1933 if (ObjectVT == MVT::i32) {
1934 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1935 // value to MVT::i64 and then truncate to the correct register size.
1937 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1938 DAG.getValueType(ObjectVT));
1939 else if (Flags.isZExt())
1940 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1941 DAG.getValueType(ObjectVT));
1943 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1949 ArgSize = PtrByteSize;
1951 // All int arguments reserve stack space in the Darwin ABI.
1957 // Every 4 bytes of argument space consumes one of the GPRs available for
1958 // argument passing.
1959 if (GPR_idx != Num_GPR_Regs) {
1961 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1964 if (FPR_idx != Num_FPR_Regs) {
1967 if (ObjectVT == MVT::f32)
1968 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
1970 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
1972 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1978 // All FP arguments reserve stack space in the Darwin ABI.
1979 ArgOffset += isPPC64 ? 8 : ObjSize;
1985 // Note that vector arguments in registers don't reserve stack space,
1986 // except in varargs functions.
1987 if (VR_idx != Num_VR_Regs) {
1988 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
1989 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1991 while ((ArgOffset % 16) != 0) {
1992 ArgOffset += PtrByteSize;
1993 if (GPR_idx != Num_GPR_Regs)
1997 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2001 if (!isVarArg && !isPPC64) {
2002 // Vectors go after all the nonvectors.
2003 CurArgOffset = VecArgOffset;
2006 // Vectors are aligned.
2007 ArgOffset = ((ArgOffset+15)/16)*16;
2008 CurArgOffset = ArgOffset;
2016 // We need to load the argument to a virtual register if we determined above
2017 // that we ran out of physical registers of the appropriate type.
2019 int FI = MFI->CreateFixedObject(ObjSize,
2020 CurArgOffset + (ArgSize - ObjSize),
2022 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2023 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2027 InVals.push_back(ArgVal);
2030 // Set the size that is at least reserved in caller of this function. Tail
2031 // call optimized function's reserved stack space needs to be aligned so that
2032 // taking the difference between two stack areas will result in an aligned
2034 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2035 // Add the Altivec parameters at the end, if needed.
2036 if (nAltivecParamsAtEnd) {
2037 MinReservedArea = ((MinReservedArea+15)/16)*16;
2038 MinReservedArea += 16*nAltivecParamsAtEnd;
2041 std::max(MinReservedArea,
2042 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2043 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2044 getStackAlignment();
2045 unsigned AlignMask = TargetAlign-1;
2046 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2047 FI->setMinReservedArea(MinReservedArea);
2049 // If the function takes variable number of arguments, make a frame index for
2050 // the start of the first vararg value... for expansion of llvm.va_start.
2052 int Depth = ArgOffset;
2054 FuncInfo->setVarArgsFrameIndex(
2055 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2057 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2059 // If this function is vararg, store any remaining integer argument regs
2060 // to their spots on the stack so that they may be loaded by deferencing the
2061 // result of va_next.
2062 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2066 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2068 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2070 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2071 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2072 MachinePointerInfo(), false, false, 0);
2073 MemOps.push_back(Store);
2074 // Increment the address by four for the next argument to store
2075 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2076 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2080 if (!MemOps.empty())
2081 Chain = DAG.getNode(ISD::TokenFactor, dl,
2082 MVT::Other, &MemOps[0], MemOps.size());
2087 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2088 /// linkage area for the Darwin ABI.
2090 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2094 const SmallVectorImpl<ISD::OutputArg>
2096 const SmallVectorImpl<SDValue> &OutVals,
2097 unsigned &nAltivecParamsAtEnd) {
2098 // Count how many bytes are to be pushed on the stack, including the linkage
2099 // area, and parameter passing area. We start with 24/48 bytes, which is
2100 // prereserved space for [SP][CR][LR][3 x unused].
2101 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2102 unsigned NumOps = Outs.size();
2103 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2105 // Add up all the space actually used.
2106 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2107 // they all go in registers, but we must reserve stack space for them for
2108 // possible use by the caller. In varargs or 64-bit calls, parameters are
2109 // assigned stack space in order, with padding so Altivec parameters are
2111 nAltivecParamsAtEnd = 0;
2112 for (unsigned i = 0; i != NumOps; ++i) {
2113 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2114 EVT ArgVT = Outs[i].VT;
2115 // Varargs Altivec parameters are padded to a 16 byte boundary.
2116 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2117 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2118 if (!isVarArg && !isPPC64) {
2119 // Non-varargs Altivec parameters go after all the non-Altivec
2120 // parameters; handle those later so we know how much padding we need.
2121 nAltivecParamsAtEnd++;
2124 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2125 NumBytes = ((NumBytes+15)/16)*16;
2127 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2130 // Allow for Altivec parameters at the end, if needed.
2131 if (nAltivecParamsAtEnd) {
2132 NumBytes = ((NumBytes+15)/16)*16;
2133 NumBytes += 16*nAltivecParamsAtEnd;
2136 // The prolog code of the callee may store up to 8 GPR argument registers to
2137 // the stack, allowing va_start to index over them in memory if its varargs.
2138 // Because we cannot tell if this is needed on the caller side, we have to
2139 // conservatively assume that it is needed. As such, make sure we have at
2140 // least enough stack space for the caller to store the 8 GPRs.
2141 NumBytes = std::max(NumBytes,
2142 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2144 // Tail call needs the stack to be aligned.
2145 if (CC==CallingConv::Fast && GuaranteedTailCallOpt) {
2146 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2147 getStackAlignment();
2148 unsigned AlignMask = TargetAlign-1;
2149 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2155 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2156 /// adjusted to accomodate the arguments for the tailcall.
2157 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2158 unsigned ParamSize) {
2160 if (!isTailCall) return 0;
2162 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2163 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2164 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2165 // Remember only if the new adjustement is bigger.
2166 if (SPDiff < FI->getTailCallSPDelta())
2167 FI->setTailCallSPDelta(SPDiff);
2172 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2173 /// for tail call optimization. Targets which want to do tail call
2174 /// optimization should implement this function.
2176 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2177 CallingConv::ID CalleeCC,
2179 const SmallVectorImpl<ISD::InputArg> &Ins,
2180 SelectionDAG& DAG) const {
2181 if (!GuaranteedTailCallOpt)
2184 // Variable argument functions are not supported.
2188 MachineFunction &MF = DAG.getMachineFunction();
2189 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2190 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2191 // Functions containing by val parameters are not supported.
2192 for (unsigned i = 0; i != Ins.size(); i++) {
2193 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2194 if (Flags.isByVal()) return false;
2197 // Non PIC/GOT tail calls are supported.
2198 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2201 // At the moment we can only do local tail calls (in same module, hidden
2202 // or protected) if we are generating PIC.
2203 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2204 return G->getGlobal()->hasHiddenVisibility()
2205 || G->getGlobal()->hasProtectedVisibility();
2211 /// isCallCompatibleAddress - Return the immediate to use if the specified
2212 /// 32-bit value is representable in the immediate field of a BxA instruction.
2213 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2214 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2217 int Addr = C->getZExtValue();
2218 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2219 (Addr << 6 >> 6) != Addr)
2220 return 0; // Top 6 bits have to be sext of immediate.
2222 return DAG.getConstant((int)C->getZExtValue() >> 2,
2223 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2228 struct TailCallArgumentInfo {
2233 TailCallArgumentInfo() : FrameIdx(0) {}
2238 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2240 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2242 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2243 SmallVector<SDValue, 8> &MemOpChains,
2245 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2246 SDValue Arg = TailCallArgs[i].Arg;
2247 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2248 int FI = TailCallArgs[i].FrameIdx;
2249 // Store relative to framepointer.
2250 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2251 MachinePointerInfo::getFixedStack(FI),
2256 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2257 /// the appropriate stack slot for the tail call optimized function call.
2258 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2259 MachineFunction &MF,
2268 // Calculate the new stack slot for the return address.
2269 int SlotSize = isPPC64 ? 8 : 4;
2270 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2272 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2273 NewRetAddrLoc, true);
2274 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2275 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2276 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2277 MachinePointerInfo::getFixedStack(NewRetAddr),
2280 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2281 // slot as the FP is never overwritten.
2284 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2285 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2287 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2288 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2289 MachinePointerInfo::getFixedStack(NewFPIdx),
2296 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2297 /// the position of the argument.
2299 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2300 SDValue Arg, int SPDiff, unsigned ArgOffset,
2301 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2302 int Offset = ArgOffset + SPDiff;
2303 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2304 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2305 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2306 SDValue FIN = DAG.getFrameIndex(FI, VT);
2307 TailCallArgumentInfo Info;
2309 Info.FrameIdxOp = FIN;
2311 TailCallArguments.push_back(Info);
2314 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2315 /// stack slot. Returns the chain as result and the loaded frame pointers in
2316 /// LROpOut/FPOpout. Used when tail calling.
2317 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2323 DebugLoc dl) const {
2325 // Load the LR and FP stack slot for later adjusting.
2326 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2327 LROpOut = getReturnAddrFrameIndex(DAG);
2328 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2330 Chain = SDValue(LROpOut.getNode(), 1);
2332 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2333 // slot as the FP is never overwritten.
2335 FPOpOut = getFramePointerFrameIndex(DAG);
2336 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2338 Chain = SDValue(FPOpOut.getNode(), 1);
2344 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2345 /// by "Src" to address "Dst" of size "Size". Alignment information is
2346 /// specified by the specific parameter attribute. The copy will be passed as
2347 /// a byval function parameter.
2348 /// Sometimes what we are copying is the end of a larger object, the part that
2349 /// does not fit in registers.
2351 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2352 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2354 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2355 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2356 false, false, MachinePointerInfo(0),
2357 MachinePointerInfo(0));
2360 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2363 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2364 SDValue Arg, SDValue PtrOff, int SPDiff,
2365 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2366 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2367 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2369 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2374 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2376 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2377 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2378 DAG.getConstant(ArgOffset, PtrVT));
2380 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2381 MachinePointerInfo(), false, false, 0));
2382 // Calculate and remember argument location.
2383 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2388 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2389 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2390 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2391 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2392 MachineFunction &MF = DAG.getMachineFunction();
2394 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2395 // might overwrite each other in case of tail call optimization.
2396 SmallVector<SDValue, 8> MemOpChains2;
2397 // Do not flag preceeding copytoreg stuff together with the following stuff.
2399 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2401 if (!MemOpChains2.empty())
2402 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2403 &MemOpChains2[0], MemOpChains2.size());
2405 // Store the return address to the appropriate stack slot.
2406 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2407 isPPC64, isDarwinABI, dl);
2409 // Emit callseq_end just before tailcall node.
2410 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2411 DAG.getIntPtrConstant(0, true), InFlag);
2412 InFlag = Chain.getValue(1);
2416 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2417 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2418 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2419 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2420 const PPCSubtarget &PPCSubTarget) {
2422 bool isPPC64 = PPCSubTarget.isPPC64();
2423 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2425 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2426 NodeTys.push_back(MVT::Other); // Returns a chain
2427 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2429 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2431 bool needIndirectCall = true;
2432 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2433 // If this is an absolute destination address, use the munged value.
2434 Callee = SDValue(Dest, 0);
2435 needIndirectCall = false;
2438 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2439 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2440 // Use indirect calls for ALL functions calls in JIT mode, since the
2441 // far-call stubs may be outside relocation limits for a BL instruction.
2442 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2443 unsigned OpFlags = 0;
2444 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2445 PPCSubTarget.getDarwinVers() < 9 &&
2446 (G->getGlobal()->isDeclaration() ||
2447 G->getGlobal()->isWeakForLinker())) {
2448 // PC-relative references to external symbols should go through $stub,
2449 // unless we're building with the leopard linker or later, which
2450 // automatically synthesizes these stubs.
2451 OpFlags = PPCII::MO_DARWIN_STUB;
2454 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2455 // every direct call is) turn it into a TargetGlobalAddress /
2456 // TargetExternalSymbol node so that legalize doesn't hack it.
2457 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2458 Callee.getValueType(),
2460 needIndirectCall = false;
2464 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2465 unsigned char OpFlags = 0;
2467 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2468 PPCSubTarget.getDarwinVers() < 9) {
2469 // PC-relative references to external symbols should go through $stub,
2470 // unless we're building with the leopard linker or later, which
2471 // automatically synthesizes these stubs.
2472 OpFlags = PPCII::MO_DARWIN_STUB;
2475 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2477 needIndirectCall = false;
2480 if (needIndirectCall) {
2481 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2482 // to do the call, we can't use PPCISD::CALL.
2483 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2485 if (isSVR4ABI && isPPC64) {
2486 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2487 // entry point, but to the function descriptor (the function entry point
2488 // address is part of the function descriptor though).
2489 // The function descriptor is a three doubleword structure with the
2490 // following fields: function entry point, TOC base address and
2491 // environment pointer.
2492 // Thus for a call through a function pointer, the following actions need
2494 // 1. Save the TOC of the caller in the TOC save area of its stack
2495 // frame (this is done in LowerCall_Darwin()).
2496 // 2. Load the address of the function entry point from the function
2498 // 3. Load the TOC of the callee from the function descriptor into r2.
2499 // 4. Load the environment pointer from the function descriptor into
2501 // 5. Branch to the function entry point address.
2502 // 6. On return of the callee, the TOC of the caller needs to be
2503 // restored (this is done in FinishCall()).
2505 // All those operations are flagged together to ensure that no other
2506 // operations can be scheduled in between. E.g. without flagging the
2507 // operations together, a TOC access in the caller could be scheduled
2508 // between the load of the callee TOC and the branch to the callee, which
2509 // results in the TOC access going through the TOC of the callee instead
2510 // of going through the TOC of the caller, which leads to incorrect code.
2512 // Load the address of the function entry point from the function
2514 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2515 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2516 InFlag.getNode() ? 3 : 2);
2517 Chain = LoadFuncPtr.getValue(1);
2518 InFlag = LoadFuncPtr.getValue(2);
2520 // Load environment pointer into r11.
2521 // Offset of the environment pointer within the function descriptor.
2522 SDValue PtrOff = DAG.getIntPtrConstant(16);
2524 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2525 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2527 Chain = LoadEnvPtr.getValue(1);
2528 InFlag = LoadEnvPtr.getValue(2);
2530 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2532 Chain = EnvVal.getValue(0);
2533 InFlag = EnvVal.getValue(1);
2535 // Load TOC of the callee into r2. We are using a target-specific load
2536 // with r2 hard coded, because the result of a target-independent load
2537 // would never go directly into r2, since r2 is a reserved register (which
2538 // prevents the register allocator from allocating it), resulting in an
2539 // additional register being allocated and an unnecessary move instruction
2541 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2542 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2544 Chain = LoadTOCPtr.getValue(0);
2545 InFlag = LoadTOCPtr.getValue(1);
2547 MTCTROps[0] = Chain;
2548 MTCTROps[1] = LoadFuncPtr;
2549 MTCTROps[2] = InFlag;
2552 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2553 2 + (InFlag.getNode() != 0));
2554 InFlag = Chain.getValue(1);
2557 NodeTys.push_back(MVT::Other);
2558 NodeTys.push_back(MVT::Glue);
2559 Ops.push_back(Chain);
2560 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2562 // Add CTR register as callee so a bctr can be emitted later.
2564 Ops.push_back(DAG.getRegister(PPC::CTR, PtrVT));
2567 // If this is a direct call, pass the chain and the callee.
2568 if (Callee.getNode()) {
2569 Ops.push_back(Chain);
2570 Ops.push_back(Callee);
2572 // If this is a tail call add stack pointer delta.
2574 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2576 // Add argument registers to the end of the list so that they are known live
2578 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2579 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2580 RegsToPass[i].second.getValueType()));
2586 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2587 CallingConv::ID CallConv, bool isVarArg,
2588 const SmallVectorImpl<ISD::InputArg> &Ins,
2589 DebugLoc dl, SelectionDAG &DAG,
2590 SmallVectorImpl<SDValue> &InVals) const {
2592 SmallVector<CCValAssign, 16> RVLocs;
2593 CCState CCRetInfo(CallConv, isVarArg, getTargetMachine(),
2594 RVLocs, *DAG.getContext());
2595 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2597 // Copy all of the result registers out of their specified physreg.
2598 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2599 CCValAssign &VA = RVLocs[i];
2600 EVT VT = VA.getValVT();
2601 assert(VA.isRegLoc() && "Can only return in registers!");
2602 Chain = DAG.getCopyFromReg(Chain, dl,
2603 VA.getLocReg(), VT, InFlag).getValue(1);
2604 InVals.push_back(Chain.getValue(0));
2605 InFlag = Chain.getValue(2);
2612 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2613 bool isTailCall, bool isVarArg,
2615 SmallVector<std::pair<unsigned, SDValue>, 8>
2617 SDValue InFlag, SDValue Chain,
2619 int SPDiff, unsigned NumBytes,
2620 const SmallVectorImpl<ISD::InputArg> &Ins,
2621 SmallVectorImpl<SDValue> &InVals) const {
2622 std::vector<EVT> NodeTys;
2623 SmallVector<SDValue, 8> Ops;
2624 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2625 isTailCall, RegsToPass, Ops, NodeTys,
2628 // When performing tail call optimization the callee pops its arguments off
2629 // the stack. Account for this here so these bytes can be pushed back on in
2630 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2631 int BytesCalleePops =
2632 (CallConv==CallingConv::Fast && GuaranteedTailCallOpt) ? NumBytes : 0;
2634 if (InFlag.getNode())
2635 Ops.push_back(InFlag);
2639 // If this is the first return lowered for this function, add the regs
2640 // to the liveout set for the function.
2641 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2642 SmallVector<CCValAssign, 16> RVLocs;
2643 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2645 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2646 for (unsigned i = 0; i != RVLocs.size(); ++i)
2647 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2650 assert(((Callee.getOpcode() == ISD::Register &&
2651 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2652 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2653 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2654 isa<ConstantSDNode>(Callee)) &&
2655 "Expecting an global address, external symbol, absolute value or register");
2657 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2660 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2661 InFlag = Chain.getValue(1);
2663 // Add a NOP immediately after the branch instruction when using the 64-bit
2664 // SVR4 ABI. At link time, if caller and callee are in a different module and
2665 // thus have a different TOC, the call will be replaced with a call to a stub
2666 // function which saves the current TOC, loads the TOC of the callee and
2667 // branches to the callee. The NOP will be replaced with a load instruction
2668 // which restores the TOC of the caller from the TOC save slot of the current
2669 // stack frame. If caller and callee belong to the same module (and have the
2670 // same TOC), the NOP will remain unchanged.
2671 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2672 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2673 if (CallOpc == PPCISD::BCTRL_SVR4) {
2674 // This is a call through a function pointer.
2675 // Restore the caller TOC from the save area into R2.
2676 // See PrepareCall() for more information about calls through function
2677 // pointers in the 64-bit SVR4 ABI.
2678 // We are using a target-specific load with r2 hard coded, because the
2679 // result of a target-independent load would never go directly into r2,
2680 // since r2 is a reserved register (which prevents the register allocator
2681 // from allocating it), resulting in an additional register being
2682 // allocated and an unnecessary move instruction being generated.
2683 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2684 InFlag = Chain.getValue(1);
2686 // Otherwise insert NOP.
2687 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Glue, InFlag);
2691 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2692 DAG.getIntPtrConstant(BytesCalleePops, true),
2695 InFlag = Chain.getValue(1);
2697 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2698 Ins, dl, DAG, InVals);
2702 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2703 CallingConv::ID CallConv, bool isVarArg,
2705 const SmallVectorImpl<ISD::OutputArg> &Outs,
2706 const SmallVectorImpl<SDValue> &OutVals,
2707 const SmallVectorImpl<ISD::InputArg> &Ins,
2708 DebugLoc dl, SelectionDAG &DAG,
2709 SmallVectorImpl<SDValue> &InVals) const {
2711 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2714 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2715 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2716 isTailCall, Outs, OutVals, Ins,
2719 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2720 isTailCall, Outs, OutVals, Ins,
2725 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2726 CallingConv::ID CallConv, bool isVarArg,
2728 const SmallVectorImpl<ISD::OutputArg> &Outs,
2729 const SmallVectorImpl<SDValue> &OutVals,
2730 const SmallVectorImpl<ISD::InputArg> &Ins,
2731 DebugLoc dl, SelectionDAG &DAG,
2732 SmallVectorImpl<SDValue> &InVals) const {
2733 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2734 // of the 32-bit SVR4 ABI stack frame layout.
2736 assert((CallConv == CallingConv::C ||
2737 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2739 unsigned PtrByteSize = 4;
2741 MachineFunction &MF = DAG.getMachineFunction();
2743 // Mark this function as potentially containing a function that contains a
2744 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2745 // and restoring the callers stack pointer in this functions epilog. This is
2746 // done because by tail calling the called function might overwrite the value
2747 // in this function's (MF) stack pointer stack slot 0(SP).
2748 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2749 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2751 // Count how many bytes are to be pushed on the stack, including the linkage
2752 // area, parameter list area and the part of the local variable space which
2753 // contains copies of aggregates which are passed by value.
2755 // Assign locations to all of the outgoing arguments.
2756 SmallVector<CCValAssign, 16> ArgLocs;
2757 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
2758 ArgLocs, *DAG.getContext());
2760 // Reserve space for the linkage area on the stack.
2761 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2764 // Handle fixed and variable vector arguments differently.
2765 // Fixed vector arguments go into registers as long as registers are
2766 // available. Variable vector arguments always go into memory.
2767 unsigned NumArgs = Outs.size();
2769 for (unsigned i = 0; i != NumArgs; ++i) {
2770 MVT ArgVT = Outs[i].VT;
2771 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2774 if (Outs[i].IsFixed) {
2775 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2778 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2784 errs() << "Call operand #" << i << " has unhandled type "
2785 << EVT(ArgVT).getEVTString() << "\n";
2787 llvm_unreachable(0);
2791 // All arguments are treated the same.
2792 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2795 // Assign locations to all of the outgoing aggregate by value arguments.
2796 SmallVector<CCValAssign, 16> ByValArgLocs;
2797 CCState CCByValInfo(CallConv, isVarArg, getTargetMachine(), ByValArgLocs,
2800 // Reserve stack space for the allocations in CCInfo.
2801 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2803 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2805 // Size of the linkage area, parameter list area and the part of the local
2806 // space variable where copies of aggregates which are passed by value are
2808 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2810 // Calculate by how many bytes the stack has to be adjusted in case of tail
2811 // call optimization.
2812 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2814 // Adjust the stack pointer for the new arguments...
2815 // These operations are automatically eliminated by the prolog/epilog pass
2816 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2817 SDValue CallSeqStart = Chain;
2819 // Load the return address and frame pointer so it can be moved somewhere else
2822 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2825 // Set up a copy of the stack pointer for use loading and storing any
2826 // arguments that may not fit in the registers available for argument
2828 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2830 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2831 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2832 SmallVector<SDValue, 8> MemOpChains;
2834 // Walk the register/memloc assignments, inserting copies/loads.
2835 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2838 CCValAssign &VA = ArgLocs[i];
2839 SDValue Arg = OutVals[i];
2840 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2842 if (Flags.isByVal()) {
2843 // Argument is an aggregate which is passed by value, thus we need to
2844 // create a copy of it in the local variable space of the current stack
2845 // frame (which is the stack frame of the caller) and pass the address of
2846 // this copy to the callee.
2847 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2848 CCValAssign &ByValVA = ByValArgLocs[j++];
2849 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2851 // Memory reserved in the local variable space of the callers stack frame.
2852 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2854 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2855 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2857 // Create a copy of the argument in the local area of the current
2859 SDValue MemcpyCall =
2860 CreateCopyOfByValArgument(Arg, PtrOff,
2861 CallSeqStart.getNode()->getOperand(0),
2864 // This must go outside the CALLSEQ_START..END.
2865 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2866 CallSeqStart.getNode()->getOperand(1));
2867 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2868 NewCallSeqStart.getNode());
2869 Chain = CallSeqStart = NewCallSeqStart;
2871 // Pass the address of the aggregate copy on the stack either in a
2872 // physical register or in the parameter list area of the current stack
2873 // frame to the callee.
2877 if (VA.isRegLoc()) {
2878 // Put argument in a physical register.
2879 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2881 // Put argument in the parameter list area of the current stack frame.
2882 assert(VA.isMemLoc());
2883 unsigned LocMemOffset = VA.getLocMemOffset();
2886 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2887 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2889 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2890 MachinePointerInfo(),
2893 // Calculate and remember argument location.
2894 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
2900 if (!MemOpChains.empty())
2901 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2902 &MemOpChains[0], MemOpChains.size());
2904 // Build a sequence of copy-to-reg nodes chained together with token chain
2905 // and flag operands which copy the outgoing args into the appropriate regs.
2907 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2908 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2909 RegsToPass[i].second, InFlag);
2910 InFlag = Chain.getValue(1);
2913 // Set CR6 to true if this is a vararg call.
2915 SDValue SetCR(DAG.getMachineNode(PPC::CRSET, dl, MVT::i32), 0);
2916 Chain = DAG.getCopyToReg(Chain, dl, PPC::CR1EQ, SetCR, InFlag);
2917 InFlag = Chain.getValue(1);
2921 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
2922 false, TailCallArguments);
2924 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
2925 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
2930 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
2931 CallingConv::ID CallConv, bool isVarArg,
2933 const SmallVectorImpl<ISD::OutputArg> &Outs,
2934 const SmallVectorImpl<SDValue> &OutVals,
2935 const SmallVectorImpl<ISD::InputArg> &Ins,
2936 DebugLoc dl, SelectionDAG &DAG,
2937 SmallVectorImpl<SDValue> &InVals) const {
2939 unsigned NumOps = Outs.size();
2941 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2942 bool isPPC64 = PtrVT == MVT::i64;
2943 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2945 MachineFunction &MF = DAG.getMachineFunction();
2947 // Mark this function as potentially containing a function that contains a
2948 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2949 // and restoring the callers stack pointer in this functions epilog. This is
2950 // done because by tail calling the called function might overwrite the value
2951 // in this function's (MF) stack pointer stack slot 0(SP).
2952 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2953 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2955 unsigned nAltivecParamsAtEnd = 0;
2957 // Count how many bytes are to be pushed on the stack, including the linkage
2958 // area, and parameter passing area. We start with 24/48 bytes, which is
2959 // prereserved space for [SP][CR][LR][3 x unused].
2961 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
2963 nAltivecParamsAtEnd);
2965 // Calculate by how many bytes the stack has to be adjusted in case of tail
2966 // call optimization.
2967 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2969 // To protect arguments on the stack from being clobbered in a tail call,
2970 // force all the loads to happen before doing any other lowering.
2972 Chain = DAG.getStackArgumentTokenFactor(Chain);
2974 // Adjust the stack pointer for the new arguments...
2975 // These operations are automatically eliminated by the prolog/epilog pass
2976 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2977 SDValue CallSeqStart = Chain;
2979 // Load the return address and frame pointer so it can be move somewhere else
2982 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
2985 // Set up a copy of the stack pointer for use loading and storing any
2986 // arguments that may not fit in the registers available for argument
2990 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2992 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2994 // Figure out which arguments are going to go in registers, and which in
2995 // memory. Also, if this is a vararg function, floating point operations
2996 // must be stored to our stack, and loaded into integer regs as well, if
2997 // any integer regs are available for argument passing.
2998 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
2999 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3001 static const unsigned GPR_32[] = { // 32-bit registers.
3002 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3003 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3005 static const unsigned GPR_64[] = { // 64-bit registers.
3006 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3007 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3009 static const unsigned *FPR = GetFPR();
3011 static const unsigned VR[] = {
3012 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3013 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3015 const unsigned NumGPRs = array_lengthof(GPR_32);
3016 const unsigned NumFPRs = 13;
3017 const unsigned NumVRs = array_lengthof(VR);
3019 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
3021 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3022 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3024 SmallVector<SDValue, 8> MemOpChains;
3025 for (unsigned i = 0; i != NumOps; ++i) {
3026 SDValue Arg = OutVals[i];
3027 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3029 // PtrOff will be used to store the current argument to the stack if a
3030 // register cannot be found for it.
3033 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3035 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3037 // On PPC64, promote integers to 64-bit values.
3038 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3039 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3040 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3041 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3044 // FIXME memcpy is used way more than necessary. Correctness first.
3045 if (Flags.isByVal()) {
3046 unsigned Size = Flags.getByValSize();
3047 if (Size==1 || Size==2) {
3048 // Very small objects are passed right-justified.
3049 // Everything else is passed left-justified.
3050 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3051 if (GPR_idx != NumGPRs) {
3052 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3053 MachinePointerInfo(), VT,
3055 MemOpChains.push_back(Load.getValue(1));
3056 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3058 ArgOffset += PtrByteSize;
3060 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3061 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3062 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3063 CallSeqStart.getNode()->getOperand(0),
3065 // This must go outside the CALLSEQ_START..END.
3066 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3067 CallSeqStart.getNode()->getOperand(1));
3068 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3069 NewCallSeqStart.getNode());
3070 Chain = CallSeqStart = NewCallSeqStart;
3071 ArgOffset += PtrByteSize;
3075 // Copy entire object into memory. There are cases where gcc-generated
3076 // code assumes it is there, even if it could be put entirely into
3077 // registers. (This is not what the doc says.)
3078 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3079 CallSeqStart.getNode()->getOperand(0),
3081 // This must go outside the CALLSEQ_START..END.
3082 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3083 CallSeqStart.getNode()->getOperand(1));
3084 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3085 Chain = CallSeqStart = NewCallSeqStart;
3086 // And copy the pieces of it that fit into registers.
3087 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3088 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3089 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3090 if (GPR_idx != NumGPRs) {
3091 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3092 MachinePointerInfo(),
3094 MemOpChains.push_back(Load.getValue(1));
3095 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3096 ArgOffset += PtrByteSize;
3098 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3105 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3106 default: llvm_unreachable("Unexpected ValueType for argument!");
3109 if (GPR_idx != NumGPRs) {
3110 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3112 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3113 isPPC64, isTailCall, false, MemOpChains,
3114 TailCallArguments, dl);
3116 ArgOffset += PtrByteSize;
3120 if (FPR_idx != NumFPRs) {
3121 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3124 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3125 MachinePointerInfo(), false, false, 0);
3126 MemOpChains.push_back(Store);
3128 // Float varargs are always shadowed in available integer registers
3129 if (GPR_idx != NumGPRs) {
3130 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3131 MachinePointerInfo(), false, false, 0);
3132 MemOpChains.push_back(Load.getValue(1));
3133 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3135 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3136 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3137 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3138 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3139 MachinePointerInfo(),
3141 MemOpChains.push_back(Load.getValue(1));
3142 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3145 // If we have any FPRs remaining, we may also have GPRs remaining.
3146 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3148 if (GPR_idx != NumGPRs)
3150 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3151 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3155 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3156 isPPC64, isTailCall, false, MemOpChains,
3157 TailCallArguments, dl);
3162 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3169 // These go aligned on the stack, or in the corresponding R registers
3170 // when within range. The Darwin PPC ABI doc claims they also go in
3171 // V registers; in fact gcc does this only for arguments that are
3172 // prototyped, not for those that match the ... We do it for all
3173 // arguments, seems to work.
3174 while (ArgOffset % 16 !=0) {
3175 ArgOffset += PtrByteSize;
3176 if (GPR_idx != NumGPRs)
3179 // We could elide this store in the case where the object fits
3180 // entirely in R registers. Maybe later.
3181 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3182 DAG.getConstant(ArgOffset, PtrVT));
3183 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3184 MachinePointerInfo(), false, false, 0);
3185 MemOpChains.push_back(Store);
3186 if (VR_idx != NumVRs) {
3187 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3188 MachinePointerInfo(),
3190 MemOpChains.push_back(Load.getValue(1));
3191 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3194 for (unsigned i=0; i<16; i+=PtrByteSize) {
3195 if (GPR_idx == NumGPRs)
3197 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3198 DAG.getConstant(i, PtrVT));
3199 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3201 MemOpChains.push_back(Load.getValue(1));
3202 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3207 // Non-varargs Altivec params generally go in registers, but have
3208 // stack space allocated at the end.
3209 if (VR_idx != NumVRs) {
3210 // Doesn't have GPR space allocated.
3211 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3212 } else if (nAltivecParamsAtEnd==0) {
3213 // We are emitting Altivec params in order.
3214 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3215 isPPC64, isTailCall, true, MemOpChains,
3216 TailCallArguments, dl);
3222 // If all Altivec parameters fit in registers, as they usually do,
3223 // they get stack space following the non-Altivec parameters. We
3224 // don't track this here because nobody below needs it.
3225 // If there are more Altivec parameters than fit in registers emit
3227 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3229 // Offset is aligned; skip 1st 12 params which go in V registers.
3230 ArgOffset = ((ArgOffset+15)/16)*16;
3232 for (unsigned i = 0; i != NumOps; ++i) {
3233 SDValue Arg = OutVals[i];
3234 EVT ArgType = Outs[i].VT;
3235 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3236 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3239 // We are emitting Altivec params in order.
3240 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3241 isPPC64, isTailCall, true, MemOpChains,
3242 TailCallArguments, dl);
3249 if (!MemOpChains.empty())
3250 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3251 &MemOpChains[0], MemOpChains.size());
3253 // Check if this is an indirect call (MTCTR/BCTRL).
3254 // See PrepareCall() for more information about calls through function
3255 // pointers in the 64-bit SVR4 ABI.
3256 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3257 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3258 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3259 !isBLACompatibleAddress(Callee, DAG)) {
3260 // Load r2 into a virtual register and store it to the TOC save area.
3261 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3262 // TOC save area offset.
3263 SDValue PtrOff = DAG.getIntPtrConstant(40);
3264 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3265 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3269 // On Darwin, R12 must contain the address of an indirect callee. This does
3270 // not mean the MTCTR instruction must use R12; it's easier to model this as
3271 // an extra parameter, so do that.
3273 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3274 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3275 !isBLACompatibleAddress(Callee, DAG))
3276 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3277 PPC::R12), Callee));
3279 // Build a sequence of copy-to-reg nodes chained together with token chain
3280 // and flag operands which copy the outgoing args into the appropriate regs.
3282 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3283 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3284 RegsToPass[i].second, InFlag);
3285 InFlag = Chain.getValue(1);
3289 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3290 FPOp, true, TailCallArguments);
3292 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3293 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3298 PPCTargetLowering::LowerReturn(SDValue Chain,
3299 CallingConv::ID CallConv, bool isVarArg,
3300 const SmallVectorImpl<ISD::OutputArg> &Outs,
3301 const SmallVectorImpl<SDValue> &OutVals,
3302 DebugLoc dl, SelectionDAG &DAG) const {
3304 SmallVector<CCValAssign, 16> RVLocs;
3305 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
3306 RVLocs, *DAG.getContext());
3307 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3309 // If this is the first return lowered for this function, add the regs to the
3310 // liveout set for the function.
3311 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3312 for (unsigned i = 0; i != RVLocs.size(); ++i)
3313 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3318 // Copy the result values into the output registers.
3319 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3320 CCValAssign &VA = RVLocs[i];
3321 assert(VA.isRegLoc() && "Can only return in registers!");
3322 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3324 Flag = Chain.getValue(1);
3328 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3330 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3333 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3334 const PPCSubtarget &Subtarget) const {
3335 // When we pop the dynamic allocation we need to restore the SP link.
3336 DebugLoc dl = Op.getDebugLoc();
3338 // Get the corect type for pointers.
3339 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3341 // Construct the stack pointer operand.
3342 bool isPPC64 = Subtarget.isPPC64();
3343 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3344 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3346 // Get the operands for the STACKRESTORE.
3347 SDValue Chain = Op.getOperand(0);
3348 SDValue SaveSP = Op.getOperand(1);
3350 // Load the old link SP.
3351 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3352 MachinePointerInfo(),
3355 // Restore the stack pointer.
3356 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3358 // Store the old link SP.
3359 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3366 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3367 MachineFunction &MF = DAG.getMachineFunction();
3368 bool isPPC64 = PPCSubTarget.isPPC64();
3369 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3370 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3372 // Get current frame pointer save index. The users of this index will be
3373 // primarily DYNALLOC instructions.
3374 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3375 int RASI = FI->getReturnAddrSaveIndex();
3377 // If the frame pointer save index hasn't been defined yet.
3379 // Find out what the fix offset of the frame pointer save area.
3380 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3381 // Allocate the frame index for frame pointer save area.
3382 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3384 FI->setReturnAddrSaveIndex(RASI);
3386 return DAG.getFrameIndex(RASI, PtrVT);
3390 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3391 MachineFunction &MF = DAG.getMachineFunction();
3392 bool isPPC64 = PPCSubTarget.isPPC64();
3393 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3394 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3396 // Get current frame pointer save index. The users of this index will be
3397 // primarily DYNALLOC instructions.
3398 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3399 int FPSI = FI->getFramePointerSaveIndex();
3401 // If the frame pointer save index hasn't been defined yet.
3403 // Find out what the fix offset of the frame pointer save area.
3404 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3407 // Allocate the frame index for frame pointer save area.
3408 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3410 FI->setFramePointerSaveIndex(FPSI);
3412 return DAG.getFrameIndex(FPSI, PtrVT);
3415 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3417 const PPCSubtarget &Subtarget) const {
3419 SDValue Chain = Op.getOperand(0);
3420 SDValue Size = Op.getOperand(1);
3421 DebugLoc dl = Op.getDebugLoc();
3423 // Get the corect type for pointers.
3424 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3426 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3427 DAG.getConstant(0, PtrVT), Size);
3428 // Construct a node for the frame pointer save index.
3429 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3430 // Build a DYNALLOC node.
3431 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3432 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3433 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3436 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3438 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3439 // Not FP? Not a fsel.
3440 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3441 !Op.getOperand(2).getValueType().isFloatingPoint())
3444 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3446 // Cannot handle SETEQ/SETNE.
3447 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3449 EVT ResVT = Op.getValueType();
3450 EVT CmpVT = Op.getOperand(0).getValueType();
3451 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3452 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3453 DebugLoc dl = Op.getDebugLoc();
3455 // If the RHS of the comparison is a 0.0, we don't need to do the
3456 // subtraction at all.
3457 if (isFloatingPointZero(RHS))
3459 default: break; // SETUO etc aren't handled by fsel.
3462 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3465 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3466 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3467 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3470 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3473 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3474 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3475 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3476 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3481 default: break; // SETUO etc aren't handled by fsel.
3484 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3485 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3486 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3487 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3490 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3491 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3492 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3493 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3496 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3497 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3498 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3499 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3502 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3503 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3504 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3505 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3510 // FIXME: Split this code up when LegalizeDAGTypes lands.
3511 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3512 DebugLoc dl) const {
3513 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3514 SDValue Src = Op.getOperand(0);
3515 if (Src.getValueType() == MVT::f32)
3516 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3519 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3520 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3522 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3527 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3531 // Convert the FP value to an int value through memory.
3532 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3534 // Emit a store to the stack slot.
3535 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3536 MachinePointerInfo(), false, false, 0);
3538 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3540 if (Op.getValueType() == MVT::i32)
3541 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3542 DAG.getConstant(4, FIPtr.getValueType()));
3543 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3547 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3548 SelectionDAG &DAG) const {
3549 DebugLoc dl = Op.getDebugLoc();
3550 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3551 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3554 if (Op.getOperand(0).getValueType() == MVT::i64) {
3555 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3556 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3557 if (Op.getValueType() == MVT::f32)
3558 FP = DAG.getNode(ISD::FP_ROUND, dl,
3559 MVT::f32, FP, DAG.getIntPtrConstant(0));
3563 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3564 "Unhandled SINT_TO_FP type in custom expander!");
3565 // Since we only generate this in 64-bit mode, we can take advantage of
3566 // 64-bit registers. In particular, sign extend the input value into the
3567 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3568 // then lfd it and fcfid it.
3569 MachineFunction &MF = DAG.getMachineFunction();
3570 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3571 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3572 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3573 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3575 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3578 // STD the extended value into the stack slot.
3579 MachineMemOperand *MMO =
3580 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3581 MachineMemOperand::MOStore, 8, 8);
3582 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3584 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3585 Ops, 4, MVT::i64, MMO);
3586 // Load the value as a double.
3587 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3590 // FCFID it and return it.
3591 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3592 if (Op.getValueType() == MVT::f32)
3593 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3597 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3598 SelectionDAG &DAG) const {
3599 DebugLoc dl = Op.getDebugLoc();
3601 The rounding mode is in bits 30:31 of FPSR, and has the following
3608 FLT_ROUNDS, on the other hand, expects the following:
3615 To perform the conversion, we do:
3616 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3619 MachineFunction &MF = DAG.getMachineFunction();
3620 EVT VT = Op.getValueType();
3621 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3622 std::vector<EVT> NodeTys;
3623 SDValue MFFSreg, InFlag;
3625 // Save FP Control Word to register
3626 NodeTys.push_back(MVT::f64); // return register
3627 NodeTys.push_back(MVT::Glue); // unused in this context
3628 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3630 // Save FP register to stack slot
3631 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3632 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3633 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3634 StackSlot, MachinePointerInfo(), false, false,0);
3636 // Load FP Control Word from low 32 bits of stack slot.
3637 SDValue Four = DAG.getConstant(4, PtrVT);
3638 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3639 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3642 // Transform as necessary
3644 DAG.getNode(ISD::AND, dl, MVT::i32,
3645 CWD, DAG.getConstant(3, MVT::i32));
3647 DAG.getNode(ISD::SRL, dl, MVT::i32,
3648 DAG.getNode(ISD::AND, dl, MVT::i32,
3649 DAG.getNode(ISD::XOR, dl, MVT::i32,
3650 CWD, DAG.getConstant(3, MVT::i32)),
3651 DAG.getConstant(3, MVT::i32)),
3652 DAG.getConstant(1, MVT::i32));
3655 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3657 return DAG.getNode((VT.getSizeInBits() < 16 ?
3658 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3661 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3662 EVT VT = Op.getValueType();
3663 unsigned BitWidth = VT.getSizeInBits();
3664 DebugLoc dl = Op.getDebugLoc();
3665 assert(Op.getNumOperands() == 3 &&
3666 VT == Op.getOperand(1).getValueType() &&
3669 // Expand into a bunch of logical ops. Note that these ops
3670 // depend on the PPC behavior for oversized shift amounts.
3671 SDValue Lo = Op.getOperand(0);
3672 SDValue Hi = Op.getOperand(1);
3673 SDValue Amt = Op.getOperand(2);
3674 EVT AmtVT = Amt.getValueType();
3676 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3677 DAG.getConstant(BitWidth, AmtVT), Amt);
3678 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3679 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3680 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3681 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3682 DAG.getConstant(-BitWidth, AmtVT));
3683 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3684 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3685 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3686 SDValue OutOps[] = { OutLo, OutHi };
3687 return DAG.getMergeValues(OutOps, 2, dl);
3690 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3691 EVT VT = Op.getValueType();
3692 DebugLoc dl = Op.getDebugLoc();
3693 unsigned BitWidth = VT.getSizeInBits();
3694 assert(Op.getNumOperands() == 3 &&
3695 VT == Op.getOperand(1).getValueType() &&
3698 // Expand into a bunch of logical ops. Note that these ops
3699 // depend on the PPC behavior for oversized shift amounts.
3700 SDValue Lo = Op.getOperand(0);
3701 SDValue Hi = Op.getOperand(1);
3702 SDValue Amt = Op.getOperand(2);
3703 EVT AmtVT = Amt.getValueType();
3705 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3706 DAG.getConstant(BitWidth, AmtVT), Amt);
3707 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3708 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3709 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3710 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3711 DAG.getConstant(-BitWidth, AmtVT));
3712 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3713 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3714 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3715 SDValue OutOps[] = { OutLo, OutHi };
3716 return DAG.getMergeValues(OutOps, 2, dl);
3719 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3720 DebugLoc dl = Op.getDebugLoc();
3721 EVT VT = Op.getValueType();
3722 unsigned BitWidth = VT.getSizeInBits();
3723 assert(Op.getNumOperands() == 3 &&
3724 VT == Op.getOperand(1).getValueType() &&
3727 // Expand into a bunch of logical ops, followed by a select_cc.
3728 SDValue Lo = Op.getOperand(0);
3729 SDValue Hi = Op.getOperand(1);
3730 SDValue Amt = Op.getOperand(2);
3731 EVT AmtVT = Amt.getValueType();
3733 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3734 DAG.getConstant(BitWidth, AmtVT), Amt);
3735 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3736 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3737 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3738 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3739 DAG.getConstant(-BitWidth, AmtVT));
3740 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3741 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3742 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3743 Tmp4, Tmp6, ISD::SETLE);
3744 SDValue OutOps[] = { OutLo, OutHi };
3745 return DAG.getMergeValues(OutOps, 2, dl);
3748 //===----------------------------------------------------------------------===//
3749 // Vector related lowering.
3752 /// BuildSplatI - Build a canonical splati of Val with an element size of
3753 /// SplatSize. Cast the result to VT.
3754 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3755 SelectionDAG &DAG, DebugLoc dl) {
3756 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3758 static const EVT VTys[] = { // canonical VT to use for each size.
3759 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3762 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3764 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3768 EVT CanonicalVT = VTys[SplatSize-1];
3770 // Build a canonical splat for this value.
3771 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3772 SmallVector<SDValue, 8> Ops;
3773 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3774 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3775 &Ops[0], Ops.size());
3776 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3779 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3780 /// specified intrinsic ID.
3781 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3782 SelectionDAG &DAG, DebugLoc dl,
3783 EVT DestVT = MVT::Other) {
3784 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3785 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3786 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3789 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3790 /// specified intrinsic ID.
3791 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3792 SDValue Op2, SelectionDAG &DAG,
3793 DebugLoc dl, EVT DestVT = MVT::Other) {
3794 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3795 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3796 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3800 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3801 /// amount. The result has the specified value type.
3802 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3803 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3804 // Force LHS/RHS to be the right type.
3805 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3806 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3809 for (unsigned i = 0; i != 16; ++i)
3811 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3812 return DAG.getNode(ISD::BITCAST, dl, VT, T);
3815 // If this is a case we can't handle, return null and let the default
3816 // expansion code take care of it. If we CAN select this case, and if it
3817 // selects to a single instruction, return Op. Otherwise, if we can codegen
3818 // this case more efficiently than a constant pool load, lower it to the
3819 // sequence of ops that should be used.
3820 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3821 SelectionDAG &DAG) const {
3822 DebugLoc dl = Op.getDebugLoc();
3823 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3824 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3826 // Check if this is a splat of a constant value.
3827 APInt APSplatBits, APSplatUndef;
3828 unsigned SplatBitSize;
3830 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3831 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3834 unsigned SplatBits = APSplatBits.getZExtValue();
3835 unsigned SplatUndef = APSplatUndef.getZExtValue();
3836 unsigned SplatSize = SplatBitSize / 8;
3838 // First, handle single instruction cases.
3841 if (SplatBits == 0) {
3842 // Canonicalize all zero vectors to be v4i32.
3843 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3844 SDValue Z = DAG.getConstant(0, MVT::i32);
3845 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3846 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
3851 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3852 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3854 if (SextVal >= -16 && SextVal <= 15)
3855 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3858 // Two instruction sequences.
3860 // If this value is in the range [-32,30] and is even, use:
3861 // tmp = VSPLTI[bhw], result = add tmp, tmp
3862 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3863 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3864 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3865 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3868 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3869 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3871 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3872 // Make -1 and vspltisw -1:
3873 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3875 // Make the VSLW intrinsic, computing 0x8000_0000.
3876 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3879 // xor by OnesV to invert it.
3880 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3881 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3884 // Check to see if this is a wide variety of vsplti*, binop self cases.
3885 static const signed char SplatCsts[] = {
3886 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3887 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3890 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3891 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3892 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3893 int i = SplatCsts[idx];
3895 // Figure out what shift amount will be used by altivec if shifted by i in
3897 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3899 // vsplti + shl self.
3900 if (SextVal == (i << (int)TypeShiftAmt)) {
3901 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3902 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3903 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3904 Intrinsic::ppc_altivec_vslw
3906 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3907 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3910 // vsplti + srl self.
3911 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3912 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3913 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3914 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3915 Intrinsic::ppc_altivec_vsrw
3917 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3918 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3921 // vsplti + sra self.
3922 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3923 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3924 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3925 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3926 Intrinsic::ppc_altivec_vsraw
3928 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3929 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3932 // vsplti + rol self.
3933 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3934 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3935 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3936 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3937 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3938 Intrinsic::ppc_altivec_vrlw
3940 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3941 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3944 // t = vsplti c, result = vsldoi t, t, 1
3945 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
3946 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3947 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
3949 // t = vsplti c, result = vsldoi t, t, 2
3950 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
3951 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3952 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
3954 // t = vsplti c, result = vsldoi t, t, 3
3955 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
3956 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3957 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
3961 // Three instruction sequences.
3963 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3964 if (SextVal >= 0 && SextVal <= 31) {
3965 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
3966 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3967 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
3968 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
3970 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3971 if (SextVal >= -31 && SextVal <= 0) {
3972 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
3973 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3974 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
3975 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
3981 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3982 /// the specified operations to build the shuffle.
3983 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3984 SDValue RHS, SelectionDAG &DAG,
3986 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3987 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3988 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3991 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4003 if (OpNum == OP_COPY) {
4004 if (LHSID == (1*9+2)*9+3) return LHS;
4005 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4009 SDValue OpLHS, OpRHS;
4010 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4011 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4015 default: llvm_unreachable("Unknown i32 permute!");
4017 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4018 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4019 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4020 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4023 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4024 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4025 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4026 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4029 for (unsigned i = 0; i != 16; ++i)
4030 ShufIdxs[i] = (i&3)+0;
4033 for (unsigned i = 0; i != 16; ++i)
4034 ShufIdxs[i] = (i&3)+4;
4037 for (unsigned i = 0; i != 16; ++i)
4038 ShufIdxs[i] = (i&3)+8;
4041 for (unsigned i = 0; i != 16; ++i)
4042 ShufIdxs[i] = (i&3)+12;
4045 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4047 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4049 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4051 EVT VT = OpLHS.getValueType();
4052 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4053 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4054 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4055 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4058 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4059 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4060 /// return the code it can be lowered into. Worst case, it can always be
4061 /// lowered into a vperm.
4062 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4063 SelectionDAG &DAG) const {
4064 DebugLoc dl = Op.getDebugLoc();
4065 SDValue V1 = Op.getOperand(0);
4066 SDValue V2 = Op.getOperand(1);
4067 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4068 EVT VT = Op.getValueType();
4070 // Cases that are handled by instructions that take permute immediates
4071 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4072 // selected by the instruction selector.
4073 if (V2.getOpcode() == ISD::UNDEF) {
4074 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4075 PPC::isSplatShuffleMask(SVOp, 2) ||
4076 PPC::isSplatShuffleMask(SVOp, 4) ||
4077 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4078 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4079 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4080 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4081 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4082 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4083 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4084 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4085 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4090 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4091 // and produce a fixed permutation. If any of these match, do not lower to
4093 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4094 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4095 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4096 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4097 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4098 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4099 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4100 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4101 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4104 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4105 // perfect shuffle table to emit an optimal matching sequence.
4106 SmallVector<int, 16> PermMask;
4107 SVOp->getMask(PermMask);
4109 unsigned PFIndexes[4];
4110 bool isFourElementShuffle = true;
4111 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4112 unsigned EltNo = 8; // Start out undef.
4113 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4114 if (PermMask[i*4+j] < 0)
4115 continue; // Undef, ignore it.
4117 unsigned ByteSource = PermMask[i*4+j];
4118 if ((ByteSource & 3) != j) {
4119 isFourElementShuffle = false;
4124 EltNo = ByteSource/4;
4125 } else if (EltNo != ByteSource/4) {
4126 isFourElementShuffle = false;
4130 PFIndexes[i] = EltNo;
4133 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4134 // perfect shuffle vector to determine if it is cost effective to do this as
4135 // discrete instructions, or whether we should use a vperm.
4136 if (isFourElementShuffle) {
4137 // Compute the index in the perfect shuffle table.
4138 unsigned PFTableIndex =
4139 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4141 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4142 unsigned Cost = (PFEntry >> 30);
4144 // Determining when to avoid vperm is tricky. Many things affect the cost
4145 // of vperm, particularly how many times the perm mask needs to be computed.
4146 // For example, if the perm mask can be hoisted out of a loop or is already
4147 // used (perhaps because there are multiple permutes with the same shuffle
4148 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4149 // the loop requires an extra register.
4151 // As a compromise, we only emit discrete instructions if the shuffle can be
4152 // generated in 3 or fewer operations. When we have loop information
4153 // available, if this block is within a loop, we should avoid using vperm
4154 // for 3-operation perms and use a constant pool load instead.
4156 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4159 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4160 // vector that will get spilled to the constant pool.
4161 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4163 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4164 // that it is in input element units, not in bytes. Convert now.
4165 EVT EltVT = V1.getValueType().getVectorElementType();
4166 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4168 SmallVector<SDValue, 16> ResultMask;
4169 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4170 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4172 for (unsigned j = 0; j != BytesPerElement; ++j)
4173 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4177 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4178 &ResultMask[0], ResultMask.size());
4179 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4182 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4183 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4184 /// information about the intrinsic.
4185 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4187 unsigned IntrinsicID =
4188 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4191 switch (IntrinsicID) {
4192 default: return false;
4193 // Comparison predicates.
4194 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4195 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4196 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4197 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4198 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4199 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4200 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4201 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4202 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4203 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4204 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4205 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4206 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4208 // Normal Comparisons.
4209 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4210 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4211 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4212 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4213 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4214 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4215 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4216 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4217 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4218 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4219 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4220 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4221 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4226 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4227 /// lower, do it, otherwise return null.
4228 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4229 SelectionDAG &DAG) const {
4230 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4231 // opcode number of the comparison.
4232 DebugLoc dl = Op.getDebugLoc();
4235 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4236 return SDValue(); // Don't custom lower most intrinsics.
4238 // If this is a non-dot comparison, make the VCMP node and we are done.
4240 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4241 Op.getOperand(1), Op.getOperand(2),
4242 DAG.getConstant(CompareOpc, MVT::i32));
4243 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4246 // Create the PPCISD altivec 'dot' comparison node.
4248 Op.getOperand(2), // LHS
4249 Op.getOperand(3), // RHS
4250 DAG.getConstant(CompareOpc, MVT::i32)
4252 std::vector<EVT> VTs;
4253 VTs.push_back(Op.getOperand(2).getValueType());
4254 VTs.push_back(MVT::Glue);
4255 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4257 // Now that we have the comparison, emit a copy from the CR to a GPR.
4258 // This is flagged to the above dot comparison.
4259 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4260 DAG.getRegister(PPC::CR6, MVT::i32),
4261 CompNode.getValue(1));
4263 // Unpack the result based on how the target uses it.
4264 unsigned BitNo; // Bit # of CR6.
4265 bool InvertBit; // Invert result?
4266 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4267 default: // Can't happen, don't crash on invalid number though.
4268 case 0: // Return the value of the EQ bit of CR6.
4269 BitNo = 0; InvertBit = false;
4271 case 1: // Return the inverted value of the EQ bit of CR6.
4272 BitNo = 0; InvertBit = true;
4274 case 2: // Return the value of the LT bit of CR6.
4275 BitNo = 2; InvertBit = false;
4277 case 3: // Return the inverted value of the LT bit of CR6.
4278 BitNo = 2; InvertBit = true;
4282 // Shift the bit into the low position.
4283 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4284 DAG.getConstant(8-(3-BitNo), MVT::i32));
4286 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4287 DAG.getConstant(1, MVT::i32));
4289 // If we are supposed to, toggle the bit.
4291 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4292 DAG.getConstant(1, MVT::i32));
4296 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4297 SelectionDAG &DAG) const {
4298 DebugLoc dl = Op.getDebugLoc();
4299 // Create a stack slot that is 16-byte aligned.
4300 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4301 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4302 EVT PtrVT = getPointerTy();
4303 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4305 // Store the input value into Value#0 of the stack slot.
4306 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4307 Op.getOperand(0), FIdx, MachinePointerInfo(),
4310 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4314 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4315 DebugLoc dl = Op.getDebugLoc();
4316 if (Op.getValueType() == MVT::v4i32) {
4317 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4319 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4320 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4322 SDValue RHSSwap = // = vrlw RHS, 16
4323 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4325 // Shrinkify inputs to v8i16.
4326 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4327 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4328 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4330 // Low parts multiplied together, generating 32-bit results (we ignore the
4332 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4333 LHS, RHS, DAG, dl, MVT::v4i32);
4335 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4336 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4337 // Shift the high parts up 16 bits.
4338 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4340 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4341 } else if (Op.getValueType() == MVT::v8i16) {
4342 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4344 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4346 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4347 LHS, RHS, Zero, DAG, dl);
4348 } else if (Op.getValueType() == MVT::v16i8) {
4349 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4351 // Multiply the even 8-bit parts, producing 16-bit sums.
4352 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4353 LHS, RHS, DAG, dl, MVT::v8i16);
4354 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4356 // Multiply the odd 8-bit parts, producing 16-bit sums.
4357 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4358 LHS, RHS, DAG, dl, MVT::v8i16);
4359 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4361 // Merge the results together.
4363 for (unsigned i = 0; i != 8; ++i) {
4365 Ops[i*2+1] = 2*i+1+16;
4367 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4369 llvm_unreachable("Unknown mul to lower!");
4373 /// LowerOperation - Provide custom lowering hooks for some operations.
4375 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4376 switch (Op.getOpcode()) {
4377 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4378 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4379 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4380 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4381 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4382 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4383 case ISD::SETCC: return LowerSETCC(Op, DAG);
4384 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4386 return LowerVASTART(Op, DAG, PPCSubTarget);
4389 return LowerVAARG(Op, DAG, PPCSubTarget);
4391 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4392 case ISD::DYNAMIC_STACKALLOC:
4393 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4395 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4396 case ISD::FP_TO_UINT:
4397 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4399 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4400 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4402 // Lower 64-bit shifts.
4403 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4404 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4405 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4407 // Vector-related lowering.
4408 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4409 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4410 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4411 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4412 case ISD::MUL: return LowerMUL(Op, DAG);
4414 // Frame & Return address.
4415 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4416 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4421 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4422 SmallVectorImpl<SDValue>&Results,
4423 SelectionDAG &DAG) const {
4424 DebugLoc dl = N->getDebugLoc();
4425 switch (N->getOpcode()) {
4427 assert(false && "Do not know how to custom type legalize this operation!");
4429 case ISD::FP_ROUND_INREG: {
4430 assert(N->getValueType(0) == MVT::ppcf128);
4431 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4432 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4433 MVT::f64, N->getOperand(0),
4434 DAG.getIntPtrConstant(0));
4435 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4436 MVT::f64, N->getOperand(0),
4437 DAG.getIntPtrConstant(1));
4439 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4440 // of the long double, and puts FPSCR back the way it was. We do not
4441 // actually model FPSCR.
4442 std::vector<EVT> NodeTys;
4443 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4445 NodeTys.push_back(MVT::f64); // Return register
4446 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4447 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4448 MFFSreg = Result.getValue(0);
4449 InFlag = Result.getValue(1);
4452 NodeTys.push_back(MVT::Glue); // Returns a flag
4453 Ops[0] = DAG.getConstant(31, MVT::i32);
4455 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4456 InFlag = Result.getValue(0);
4459 NodeTys.push_back(MVT::Glue); // Returns a flag
4460 Ops[0] = DAG.getConstant(30, MVT::i32);
4462 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4463 InFlag = Result.getValue(0);
4466 NodeTys.push_back(MVT::f64); // result of add
4467 NodeTys.push_back(MVT::Glue); // Returns a flag
4471 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4472 FPreg = Result.getValue(0);
4473 InFlag = Result.getValue(1);
4476 NodeTys.push_back(MVT::f64);
4477 Ops[0] = DAG.getConstant(1, MVT::i32);
4481 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4482 FPreg = Result.getValue(0);
4484 // We know the low half is about to be thrown away, so just use something
4486 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4490 case ISD::FP_TO_SINT:
4491 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4497 //===----------------------------------------------------------------------===//
4498 // Other Lowering Code
4499 //===----------------------------------------------------------------------===//
4502 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4503 bool is64bit, unsigned BinOpcode) const {
4504 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4505 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4507 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4508 MachineFunction *F = BB->getParent();
4509 MachineFunction::iterator It = BB;
4512 unsigned dest = MI->getOperand(0).getReg();
4513 unsigned ptrA = MI->getOperand(1).getReg();
4514 unsigned ptrB = MI->getOperand(2).getReg();
4515 unsigned incr = MI->getOperand(3).getReg();
4516 DebugLoc dl = MI->getDebugLoc();
4518 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4519 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4520 F->insert(It, loopMBB);
4521 F->insert(It, exitMBB);
4522 exitMBB->splice(exitMBB->begin(), BB,
4523 llvm::next(MachineBasicBlock::iterator(MI)),
4525 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4527 MachineRegisterInfo &RegInfo = F->getRegInfo();
4528 unsigned TmpReg = (!BinOpcode) ? incr :
4529 RegInfo.createVirtualRegister(
4530 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4531 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4535 // fallthrough --> loopMBB
4536 BB->addSuccessor(loopMBB);
4539 // l[wd]arx dest, ptr
4540 // add r0, dest, incr
4541 // st[wd]cx. r0, ptr
4543 // fallthrough --> exitMBB
4545 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4546 .addReg(ptrA).addReg(ptrB);
4548 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4549 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4550 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4551 BuildMI(BB, dl, TII->get(PPC::BCC))
4552 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4553 BB->addSuccessor(loopMBB);
4554 BB->addSuccessor(exitMBB);
4563 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4564 MachineBasicBlock *BB,
4565 bool is8bit, // operation
4566 unsigned BinOpcode) const {
4567 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4568 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4569 // In 64 bit mode we have to use 64 bits for addresses, even though the
4570 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4571 // registers without caring whether they're 32 or 64, but here we're
4572 // doing actual arithmetic on the addresses.
4573 bool is64bit = PPCSubTarget.isPPC64();
4575 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4576 MachineFunction *F = BB->getParent();
4577 MachineFunction::iterator It = BB;
4580 unsigned dest = MI->getOperand(0).getReg();
4581 unsigned ptrA = MI->getOperand(1).getReg();
4582 unsigned ptrB = MI->getOperand(2).getReg();
4583 unsigned incr = MI->getOperand(3).getReg();
4584 DebugLoc dl = MI->getDebugLoc();
4586 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4587 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4588 F->insert(It, loopMBB);
4589 F->insert(It, exitMBB);
4590 exitMBB->splice(exitMBB->begin(), BB,
4591 llvm::next(MachineBasicBlock::iterator(MI)),
4593 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4595 MachineRegisterInfo &RegInfo = F->getRegInfo();
4596 const TargetRegisterClass *RC =
4597 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4598 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4599 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4600 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4601 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4602 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4603 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4604 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4605 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4606 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4607 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4608 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4609 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4611 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4615 // fallthrough --> loopMBB
4616 BB->addSuccessor(loopMBB);
4618 // The 4-byte load must be aligned, while a char or short may be
4619 // anywhere in the word. Hence all this nasty bookkeeping code.
4620 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4621 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4622 // xori shift, shift1, 24 [16]
4623 // rlwinm ptr, ptr1, 0, 0, 29
4624 // slw incr2, incr, shift
4625 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4626 // slw mask, mask2, shift
4628 // lwarx tmpDest, ptr
4629 // add tmp, tmpDest, incr2
4630 // andc tmp2, tmpDest, mask
4631 // and tmp3, tmp, mask
4632 // or tmp4, tmp3, tmp2
4635 // fallthrough --> exitMBB
4636 // srw dest, tmpDest, shift
4638 if (ptrA!=PPC::R0) {
4639 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4640 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4641 .addReg(ptrA).addReg(ptrB);
4645 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4646 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4647 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4648 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4650 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4651 .addReg(Ptr1Reg).addImm(0).addImm(61);
4653 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4654 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4655 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4656 .addReg(incr).addReg(ShiftReg);
4658 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4660 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4661 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4663 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4664 .addReg(Mask2Reg).addReg(ShiftReg);
4667 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4668 .addReg(PPC::R0).addReg(PtrReg);
4670 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4671 .addReg(Incr2Reg).addReg(TmpDestReg);
4672 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4673 .addReg(TmpDestReg).addReg(MaskReg);
4674 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4675 .addReg(TmpReg).addReg(MaskReg);
4676 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4677 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4678 BuildMI(BB, dl, TII->get(PPC::STWCX))
4679 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4680 BuildMI(BB, dl, TII->get(PPC::BCC))
4681 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4682 BB->addSuccessor(loopMBB);
4683 BB->addSuccessor(exitMBB);
4688 BuildMI(BB, dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4693 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4694 MachineBasicBlock *BB) const {
4695 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4697 // To "insert" these instructions we actually have to insert their
4698 // control-flow patterns.
4699 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4700 MachineFunction::iterator It = BB;
4703 MachineFunction *F = BB->getParent();
4705 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4706 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4707 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4708 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4709 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4711 // The incoming instruction knows the destination vreg to set, the
4712 // condition code register to branch on, the true/false values to
4713 // select between, and a branch opcode to use.
4718 // cmpTY ccX, r1, r2
4720 // fallthrough --> copy0MBB
4721 MachineBasicBlock *thisMBB = BB;
4722 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4723 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4724 unsigned SelectPred = MI->getOperand(4).getImm();
4725 DebugLoc dl = MI->getDebugLoc();
4726 F->insert(It, copy0MBB);
4727 F->insert(It, sinkMBB);
4729 // Transfer the remainder of BB and its successor edges to sinkMBB.
4730 sinkMBB->splice(sinkMBB->begin(), BB,
4731 llvm::next(MachineBasicBlock::iterator(MI)),
4733 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4735 // Next, add the true and fallthrough blocks as its successors.
4736 BB->addSuccessor(copy0MBB);
4737 BB->addSuccessor(sinkMBB);
4739 BuildMI(BB, dl, TII->get(PPC::BCC))
4740 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4743 // %FalseValue = ...
4744 // # fallthrough to sinkMBB
4747 // Update machine-CFG edges
4748 BB->addSuccessor(sinkMBB);
4751 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4754 BuildMI(*BB, BB->begin(), dl,
4755 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4756 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4757 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4759 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4760 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4761 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4762 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4763 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4764 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4765 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4766 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4768 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4769 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4770 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4771 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4772 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4773 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4774 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4775 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4777 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4778 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4779 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4780 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4781 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4782 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4783 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4784 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4786 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4787 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4788 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4789 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4790 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4791 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4792 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4793 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4795 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4796 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4797 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4798 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4799 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4800 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4801 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4802 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4804 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4805 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4806 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4807 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4808 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4809 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4810 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4811 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4813 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4814 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4815 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4816 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4817 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4818 BB = EmitAtomicBinary(MI, BB, false, 0);
4819 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4820 BB = EmitAtomicBinary(MI, BB, true, 0);
4822 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4823 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4824 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4826 unsigned dest = MI->getOperand(0).getReg();
4827 unsigned ptrA = MI->getOperand(1).getReg();
4828 unsigned ptrB = MI->getOperand(2).getReg();
4829 unsigned oldval = MI->getOperand(3).getReg();
4830 unsigned newval = MI->getOperand(4).getReg();
4831 DebugLoc dl = MI->getDebugLoc();
4833 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4834 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4835 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4836 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4837 F->insert(It, loop1MBB);
4838 F->insert(It, loop2MBB);
4839 F->insert(It, midMBB);
4840 F->insert(It, exitMBB);
4841 exitMBB->splice(exitMBB->begin(), BB,
4842 llvm::next(MachineBasicBlock::iterator(MI)),
4844 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4848 // fallthrough --> loopMBB
4849 BB->addSuccessor(loop1MBB);
4852 // l[wd]arx dest, ptr
4853 // cmp[wd] dest, oldval
4856 // st[wd]cx. newval, ptr
4860 // st[wd]cx. dest, ptr
4863 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4864 .addReg(ptrA).addReg(ptrB);
4865 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4866 .addReg(oldval).addReg(dest);
4867 BuildMI(BB, dl, TII->get(PPC::BCC))
4868 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4869 BB->addSuccessor(loop2MBB);
4870 BB->addSuccessor(midMBB);
4873 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4874 .addReg(newval).addReg(ptrA).addReg(ptrB);
4875 BuildMI(BB, dl, TII->get(PPC::BCC))
4876 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4877 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4878 BB->addSuccessor(loop1MBB);
4879 BB->addSuccessor(exitMBB);
4882 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4883 .addReg(dest).addReg(ptrA).addReg(ptrB);
4884 BB->addSuccessor(exitMBB);
4889 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4890 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4891 // We must use 64-bit registers for addresses when targeting 64-bit,
4892 // since we're actually doing arithmetic on them. Other registers
4894 bool is64bit = PPCSubTarget.isPPC64();
4895 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4897 unsigned dest = MI->getOperand(0).getReg();
4898 unsigned ptrA = MI->getOperand(1).getReg();
4899 unsigned ptrB = MI->getOperand(2).getReg();
4900 unsigned oldval = MI->getOperand(3).getReg();
4901 unsigned newval = MI->getOperand(4).getReg();
4902 DebugLoc dl = MI->getDebugLoc();
4904 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4905 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4906 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4907 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4908 F->insert(It, loop1MBB);
4909 F->insert(It, loop2MBB);
4910 F->insert(It, midMBB);
4911 F->insert(It, exitMBB);
4912 exitMBB->splice(exitMBB->begin(), BB,
4913 llvm::next(MachineBasicBlock::iterator(MI)),
4915 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4917 MachineRegisterInfo &RegInfo = F->getRegInfo();
4918 const TargetRegisterClass *RC =
4919 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4920 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4921 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4922 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4923 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4924 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4925 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4926 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4927 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4928 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4929 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4930 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4931 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4932 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4933 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4935 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4938 // fallthrough --> loopMBB
4939 BB->addSuccessor(loop1MBB);
4941 // The 4-byte load must be aligned, while a char or short may be
4942 // anywhere in the word. Hence all this nasty bookkeeping code.
4943 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4944 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4945 // xori shift, shift1, 24 [16]
4946 // rlwinm ptr, ptr1, 0, 0, 29
4947 // slw newval2, newval, shift
4948 // slw oldval2, oldval,shift
4949 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4950 // slw mask, mask2, shift
4951 // and newval3, newval2, mask
4952 // and oldval3, oldval2, mask
4954 // lwarx tmpDest, ptr
4955 // and tmp, tmpDest, mask
4956 // cmpw tmp, oldval3
4959 // andc tmp2, tmpDest, mask
4960 // or tmp4, tmp2, newval3
4965 // stwcx. tmpDest, ptr
4967 // srw dest, tmpDest, shift
4968 if (ptrA!=PPC::R0) {
4969 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4970 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4971 .addReg(ptrA).addReg(ptrB);
4975 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4976 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4977 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4978 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4980 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4981 .addReg(Ptr1Reg).addImm(0).addImm(61);
4983 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4984 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4985 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
4986 .addReg(newval).addReg(ShiftReg);
4987 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
4988 .addReg(oldval).addReg(ShiftReg);
4990 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4992 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4993 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
4994 .addReg(Mask3Reg).addImm(65535);
4996 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4997 .addReg(Mask2Reg).addReg(ShiftReg);
4998 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
4999 .addReg(NewVal2Reg).addReg(MaskReg);
5000 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5001 .addReg(OldVal2Reg).addReg(MaskReg);
5004 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5005 .addReg(PPC::R0).addReg(PtrReg);
5006 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5007 .addReg(TmpDestReg).addReg(MaskReg);
5008 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5009 .addReg(TmpReg).addReg(OldVal3Reg);
5010 BuildMI(BB, dl, TII->get(PPC::BCC))
5011 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5012 BB->addSuccessor(loop2MBB);
5013 BB->addSuccessor(midMBB);
5016 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5017 .addReg(TmpDestReg).addReg(MaskReg);
5018 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5019 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5020 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5021 .addReg(PPC::R0).addReg(PtrReg);
5022 BuildMI(BB, dl, TII->get(PPC::BCC))
5023 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5024 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5025 BB->addSuccessor(loop1MBB);
5026 BB->addSuccessor(exitMBB);
5029 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5030 .addReg(PPC::R0).addReg(PtrReg);
5031 BB->addSuccessor(exitMBB);
5036 BuildMI(BB, dl, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
5038 llvm_unreachable("Unexpected instr type to insert");
5041 MI->eraseFromParent(); // The pseudo instruction is gone now.
5045 //===----------------------------------------------------------------------===//
5046 // Target Optimization Hooks
5047 //===----------------------------------------------------------------------===//
5049 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5050 DAGCombinerInfo &DCI) const {
5051 const TargetMachine &TM = getTargetMachine();
5052 SelectionDAG &DAG = DCI.DAG;
5053 DebugLoc dl = N->getDebugLoc();
5054 switch (N->getOpcode()) {
5057 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5058 if (C->isNullValue()) // 0 << V -> 0.
5059 return N->getOperand(0);
5063 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5064 if (C->isNullValue()) // 0 >>u V -> 0.
5065 return N->getOperand(0);
5069 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5070 if (C->isNullValue() || // 0 >>s V -> 0.
5071 C->isAllOnesValue()) // -1 >>s V -> -1.
5072 return N->getOperand(0);
5076 case ISD::SINT_TO_FP:
5077 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5078 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5079 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5080 // We allow the src/dst to be either f32/f64, but the intermediate
5081 // type must be i64.
5082 if (N->getOperand(0).getValueType() == MVT::i64 &&
5083 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5084 SDValue Val = N->getOperand(0).getOperand(0);
5085 if (Val.getValueType() == MVT::f32) {
5086 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5087 DCI.AddToWorklist(Val.getNode());
5090 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5091 DCI.AddToWorklist(Val.getNode());
5092 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5093 DCI.AddToWorklist(Val.getNode());
5094 if (N->getValueType(0) == MVT::f32) {
5095 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5096 DAG.getIntPtrConstant(0));
5097 DCI.AddToWorklist(Val.getNode());
5100 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5101 // If the intermediate type is i32, we can avoid the load/store here
5108 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5109 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5110 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5111 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5112 N->getOperand(1).getValueType() == MVT::i32 &&
5113 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5114 SDValue Val = N->getOperand(1).getOperand(0);
5115 if (Val.getValueType() == MVT::f32) {
5116 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5117 DCI.AddToWorklist(Val.getNode());
5119 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5120 DCI.AddToWorklist(Val.getNode());
5122 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5123 N->getOperand(2), N->getOperand(3));
5124 DCI.AddToWorklist(Val.getNode());
5128 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5129 if (cast<StoreSDNode>(N)->isUnindexed() &&
5130 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5131 N->getOperand(1).getNode()->hasOneUse() &&
5132 (N->getOperand(1).getValueType() == MVT::i32 ||
5133 N->getOperand(1).getValueType() == MVT::i16)) {
5134 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5135 // Do an any-extend to 32-bits if this is a half-word input.
5136 if (BSwapOp.getValueType() == MVT::i16)
5137 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5140 N->getOperand(0), BSwapOp, N->getOperand(2),
5141 DAG.getValueType(N->getOperand(1).getValueType())
5144 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5145 Ops, array_lengthof(Ops),
5146 cast<StoreSDNode>(N)->getMemoryVT(),
5147 cast<StoreSDNode>(N)->getMemOperand());
5151 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5152 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5153 N->getOperand(0).hasOneUse() &&
5154 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5155 SDValue Load = N->getOperand(0);
5156 LoadSDNode *LD = cast<LoadSDNode>(Load);
5157 // Create the byte-swapping load.
5159 LD->getChain(), // Chain
5160 LD->getBasePtr(), // Ptr
5161 DAG.getValueType(N->getValueType(0)) // VT
5164 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5165 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5166 LD->getMemoryVT(), LD->getMemOperand());
5168 // If this is an i16 load, insert the truncate.
5169 SDValue ResVal = BSLoad;
5170 if (N->getValueType(0) == MVT::i16)
5171 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5173 // First, combine the bswap away. This makes the value produced by the
5175 DCI.CombineTo(N, ResVal);
5177 // Next, combine the load away, we give it a bogus result value but a real
5178 // chain result. The result value is dead because the bswap is dead.
5179 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5181 // Return N so it doesn't get rechecked!
5182 return SDValue(N, 0);
5186 case PPCISD::VCMP: {
5187 // If a VCMPo node already exists with exactly the same operands as this
5188 // node, use its result instead of this node (VCMPo computes both a CR6 and
5189 // a normal output).
5191 if (!N->getOperand(0).hasOneUse() &&
5192 !N->getOperand(1).hasOneUse() &&
5193 !N->getOperand(2).hasOneUse()) {
5195 // Scan all of the users of the LHS, looking for VCMPo's that match.
5196 SDNode *VCMPoNode = 0;
5198 SDNode *LHSN = N->getOperand(0).getNode();
5199 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5201 if (UI->getOpcode() == PPCISD::VCMPo &&
5202 UI->getOperand(1) == N->getOperand(1) &&
5203 UI->getOperand(2) == N->getOperand(2) &&
5204 UI->getOperand(0) == N->getOperand(0)) {
5209 // If there is no VCMPo node, or if the flag value has a single use, don't
5211 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5214 // Look at the (necessarily single) use of the flag value. If it has a
5215 // chain, this transformation is more complex. Note that multiple things
5216 // could use the value result, which we should ignore.
5217 SDNode *FlagUser = 0;
5218 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5219 FlagUser == 0; ++UI) {
5220 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5222 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5223 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5230 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5231 // give up for right now.
5232 if (FlagUser->getOpcode() == PPCISD::MFCR)
5233 return SDValue(VCMPoNode, 0);
5238 // If this is a branch on an altivec predicate comparison, lower this so
5239 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5240 // lowering is done pre-legalize, because the legalizer lowers the predicate
5241 // compare down to code that is difficult to reassemble.
5242 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5243 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5247 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5248 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5249 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5250 assert(isDot && "Can't compare against a vector result!");
5252 // If this is a comparison against something other than 0/1, then we know
5253 // that the condition is never/always true.
5254 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5255 if (Val != 0 && Val != 1) {
5256 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5257 return N->getOperand(0);
5258 // Always !=, turn it into an unconditional branch.
5259 return DAG.getNode(ISD::BR, dl, MVT::Other,
5260 N->getOperand(0), N->getOperand(4));
5263 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5265 // Create the PPCISD altivec 'dot' comparison node.
5266 std::vector<EVT> VTs;
5268 LHS.getOperand(2), // LHS of compare
5269 LHS.getOperand(3), // RHS of compare
5270 DAG.getConstant(CompareOpc, MVT::i32)
5272 VTs.push_back(LHS.getOperand(2).getValueType());
5273 VTs.push_back(MVT::Glue);
5274 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5276 // Unpack the result based on how the target uses it.
5277 PPC::Predicate CompOpc;
5278 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5279 default: // Can't happen, don't crash on invalid number though.
5280 case 0: // Branch on the value of the EQ bit of CR6.
5281 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5283 case 1: // Branch on the inverted value of the EQ bit of CR6.
5284 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5286 case 2: // Branch on the value of the LT bit of CR6.
5287 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5289 case 3: // Branch on the inverted value of the LT bit of CR6.
5290 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5294 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5295 DAG.getConstant(CompOpc, MVT::i32),
5296 DAG.getRegister(PPC::CR6, MVT::i32),
5297 N->getOperand(4), CompNode.getValue(1));
5306 //===----------------------------------------------------------------------===//
5307 // Inline Assembly Support
5308 //===----------------------------------------------------------------------===//
5310 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5314 const SelectionDAG &DAG,
5315 unsigned Depth) const {
5316 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5317 switch (Op.getOpcode()) {
5319 case PPCISD::LBRX: {
5320 // lhbrx is known to have the top bits cleared out.
5321 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5322 KnownZero = 0xFFFF0000;
5325 case ISD::INTRINSIC_WO_CHAIN: {
5326 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5328 case Intrinsic::ppc_altivec_vcmpbfp_p:
5329 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5330 case Intrinsic::ppc_altivec_vcmpequb_p:
5331 case Intrinsic::ppc_altivec_vcmpequh_p:
5332 case Intrinsic::ppc_altivec_vcmpequw_p:
5333 case Intrinsic::ppc_altivec_vcmpgefp_p:
5334 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5335 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5336 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5337 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5338 case Intrinsic::ppc_altivec_vcmpgtub_p:
5339 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5340 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5341 KnownZero = ~1U; // All bits but the low one are known to be zero.
5349 /// getConstraintType - Given a constraint, return the type of
5350 /// constraint it is for this target.
5351 PPCTargetLowering::ConstraintType
5352 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5353 if (Constraint.size() == 1) {
5354 switch (Constraint[0]) {
5361 return C_RegisterClass;
5364 return TargetLowering::getConstraintType(Constraint);
5367 /// Examine constraint type and operand type and determine a weight value.
5368 /// This object must already have been set up with the operand type
5369 /// and the current alternative constraint selected.
5370 TargetLowering::ConstraintWeight
5371 PPCTargetLowering::getSingleConstraintMatchWeight(
5372 AsmOperandInfo &info, const char *constraint) const {
5373 ConstraintWeight weight = CW_Invalid;
5374 Value *CallOperandVal = info.CallOperandVal;
5375 // If we don't have a value, we can't do a match,
5376 // but allow it at the lowest weight.
5377 if (CallOperandVal == NULL)
5379 const Type *type = CallOperandVal->getType();
5380 // Look at the constraint type.
5381 switch (*constraint) {
5383 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5386 if (type->isIntegerTy())
5387 weight = CW_Register;
5390 if (type->isFloatTy())
5391 weight = CW_Register;
5394 if (type->isDoubleTy())
5395 weight = CW_Register;
5398 if (type->isVectorTy())
5399 weight = CW_Register;
5402 weight = CW_Register;
5408 std::pair<unsigned, const TargetRegisterClass*>
5409 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5411 if (Constraint.size() == 1) {
5412 // GCC RS6000 Constraint Letters
5413 switch (Constraint[0]) {
5416 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5417 return std::make_pair(0U, PPC::G8RCRegisterClass);
5418 return std::make_pair(0U, PPC::GPRCRegisterClass);
5421 return std::make_pair(0U, PPC::F4RCRegisterClass);
5422 else if (VT == MVT::f64)
5423 return std::make_pair(0U, PPC::F8RCRegisterClass);
5426 return std::make_pair(0U, PPC::VRRCRegisterClass);
5428 return std::make_pair(0U, PPC::CRRCRegisterClass);
5432 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5436 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5437 /// vector. If it is invalid, don't add anything to Ops.
5438 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
5439 std::vector<SDValue>&Ops,
5440 SelectionDAG &DAG) const {
5441 SDValue Result(0,0);
5452 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5453 if (!CST) return; // Must be an immediate to match.
5454 unsigned Value = CST->getZExtValue();
5456 default: llvm_unreachable("Unknown constraint letter!");
5457 case 'I': // "I" is a signed 16-bit constant.
5458 if ((short)Value == (int)Value)
5459 Result = DAG.getTargetConstant(Value, Op.getValueType());
5461 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5462 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5463 if ((short)Value == 0)
5464 Result = DAG.getTargetConstant(Value, Op.getValueType());
5466 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5467 if ((Value >> 16) == 0)
5468 Result = DAG.getTargetConstant(Value, Op.getValueType());
5470 case 'M': // "M" is a constant that is greater than 31.
5472 Result = DAG.getTargetConstant(Value, Op.getValueType());
5474 case 'N': // "N" is a positive constant that is an exact power of two.
5475 if ((int)Value > 0 && isPowerOf2_32(Value))
5476 Result = DAG.getTargetConstant(Value, Op.getValueType());
5478 case 'O': // "O" is the constant zero.
5480 Result = DAG.getTargetConstant(Value, Op.getValueType());
5482 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5483 if ((short)-Value == (int)-Value)
5484 Result = DAG.getTargetConstant(Value, Op.getValueType());
5491 if (Result.getNode()) {
5492 Ops.push_back(Result);
5496 // Handle standard constraint letters.
5497 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
5500 // isLegalAddressingMode - Return true if the addressing mode represented
5501 // by AM is legal for this target, for a load/store of the specified type.
5502 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5503 const Type *Ty) const {
5504 // FIXME: PPC does not allow r+i addressing modes for vectors!
5506 // PPC allows a sign-extended 16-bit immediate field.
5507 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5510 // No global is ever allowed as a base.
5514 // PPC only support r+r,
5516 case 0: // "r+i" or just "i", depending on HasBaseReg.
5519 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5521 // Otherwise we have r+r or r+i.
5524 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5526 // Allow 2*r as r+r.
5529 // No other scales are supported.
5536 /// isLegalAddressImmediate - Return true if the integer value can be used
5537 /// as the offset of the target addressing mode for load / store of the
5539 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
5540 // PPC allows a sign-extended 16-bit immediate field.
5541 return (V > -(1 << 16) && V < (1 << 16)-1);
5544 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5548 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5549 SelectionDAG &DAG) const {
5550 MachineFunction &MF = DAG.getMachineFunction();
5551 MachineFrameInfo *MFI = MF.getFrameInfo();
5552 MFI->setReturnAddressIsTaken(true);
5554 DebugLoc dl = Op.getDebugLoc();
5555 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5557 // Make sure the function does not optimize away the store of the RA to
5559 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5560 FuncInfo->setLRStoreRequired();
5561 bool isPPC64 = PPCSubTarget.isPPC64();
5562 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5565 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5568 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5569 isPPC64? MVT::i64 : MVT::i32);
5570 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5571 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5573 MachinePointerInfo(), false, false, 0);
5576 // Just load the return address off the stack.
5577 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5578 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5579 RetAddrFI, MachinePointerInfo(), false, false, 0);
5582 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5583 SelectionDAG &DAG) const {
5584 DebugLoc dl = Op.getDebugLoc();
5585 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5587 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5588 bool isPPC64 = PtrVT == MVT::i64;
5590 MachineFunction &MF = DAG.getMachineFunction();
5591 MachineFrameInfo *MFI = MF.getFrameInfo();
5592 MFI->setFrameAddressIsTaken(true);
5593 bool is31 = (DisableFramePointerElim(MF) || MFI->hasVarSizedObjects()) &&
5594 MFI->getStackSize() &&
5595 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5596 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5597 (is31 ? PPC::R31 : PPC::R1);
5598 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5601 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5602 FrameAddr, MachinePointerInfo(), false, false, 0);
5607 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5608 // The PowerPC target isn't yet aware of offsets.
5612 /// getOptimalMemOpType - Returns the target specific optimal type for load
5613 /// and store operations as a result of memset, memcpy, and memmove
5614 /// lowering. If DstAlign is zero that means it's safe to destination
5615 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5616 /// means there isn't a need to check it against alignment requirement,
5617 /// probably because the source does not need to be loaded. If
5618 /// 'NonScalarIntSafe' is true, that means it's safe to return a
5619 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5620 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5621 /// constant so it does not need to be loaded.
5622 /// It returns EVT::Other if the type should be determined using generic
5623 /// target-independent logic.
5624 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5625 unsigned DstAlign, unsigned SrcAlign,
5626 bool NonScalarIntSafe,
5628 MachineFunction &MF) const {
5629 if (this->PPCSubTarget.isPPC64()) {