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 setMinFunctionAlignment(2);
398 if (PPCSubTarget.isDarwin())
399 setPrefFunctionAlignment(4);
401 computeRegisterProperties();
404 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
405 /// function arguments in the caller parameter area.
406 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
407 const TargetMachine &TM = getTargetMachine();
408 // Darwin passes everything on 4 byte boundary.
409 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
415 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
418 case PPCISD::FSEL: return "PPCISD::FSEL";
419 case PPCISD::FCFID: return "PPCISD::FCFID";
420 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
421 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
422 case PPCISD::STFIWX: return "PPCISD::STFIWX";
423 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
424 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
425 case PPCISD::VPERM: return "PPCISD::VPERM";
426 case PPCISD::Hi: return "PPCISD::Hi";
427 case PPCISD::Lo: return "PPCISD::Lo";
428 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
429 case PPCISD::TOC_RESTORE: return "PPCISD::TOC_RESTORE";
430 case PPCISD::LOAD: return "PPCISD::LOAD";
431 case PPCISD::LOAD_TOC: return "PPCISD::LOAD_TOC";
432 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
433 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
434 case PPCISD::SRL: return "PPCISD::SRL";
435 case PPCISD::SRA: return "PPCISD::SRA";
436 case PPCISD::SHL: return "PPCISD::SHL";
437 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
438 case PPCISD::STD_32: return "PPCISD::STD_32";
439 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
440 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
441 case PPCISD::NOP: return "PPCISD::NOP";
442 case PPCISD::MTCTR: return "PPCISD::MTCTR";
443 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
444 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
445 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
446 case PPCISD::MFCR: return "PPCISD::MFCR";
447 case PPCISD::VCMP: return "PPCISD::VCMP";
448 case PPCISD::VCMPo: return "PPCISD::VCMPo";
449 case PPCISD::LBRX: return "PPCISD::LBRX";
450 case PPCISD::STBRX: return "PPCISD::STBRX";
451 case PPCISD::LARX: return "PPCISD::LARX";
452 case PPCISD::STCX: return "PPCISD::STCX";
453 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
454 case PPCISD::MFFS: return "PPCISD::MFFS";
455 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
456 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
457 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
458 case PPCISD::MTFSF: return "PPCISD::MTFSF";
459 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
463 MVT::SimpleValueType PPCTargetLowering::getSetCCResultType(EVT VT) const {
467 //===----------------------------------------------------------------------===//
468 // Node matching predicates, for use by the tblgen matching code.
469 //===----------------------------------------------------------------------===//
471 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
472 static bool isFloatingPointZero(SDValue Op) {
473 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
474 return CFP->getValueAPF().isZero();
475 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
476 // Maybe this has already been legalized into the constant pool?
477 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
478 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
479 return CFP->getValueAPF().isZero();
484 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
485 /// true if Op is undef or if it matches the specified value.
486 static bool isConstantOrUndef(int Op, int Val) {
487 return Op < 0 || Op == Val;
490 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
491 /// VPKUHUM instruction.
492 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
494 for (unsigned i = 0; i != 16; ++i)
495 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
498 for (unsigned i = 0; i != 8; ++i)
499 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
500 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
506 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
507 /// VPKUWUM instruction.
508 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
510 for (unsigned i = 0; i != 16; i += 2)
511 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
512 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
515 for (unsigned i = 0; i != 8; i += 2)
516 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
517 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
518 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
519 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
525 /// isVMerge - Common function, used to match vmrg* shuffles.
527 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
528 unsigned LHSStart, unsigned RHSStart) {
529 assert(N->getValueType(0) == MVT::v16i8 &&
530 "PPC only supports shuffles by bytes!");
531 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
532 "Unsupported merge size!");
534 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
535 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
536 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
537 LHSStart+j+i*UnitSize) ||
538 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
539 RHSStart+j+i*UnitSize))
545 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
546 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
547 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
550 return isVMerge(N, UnitSize, 8, 24);
551 return isVMerge(N, UnitSize, 8, 8);
554 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
555 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
556 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
559 return isVMerge(N, UnitSize, 0, 16);
560 return isVMerge(N, UnitSize, 0, 0);
564 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
565 /// amount, otherwise return -1.
566 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
567 assert(N->getValueType(0) == MVT::v16i8 &&
568 "PPC only supports shuffles by bytes!");
570 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
572 // Find the first non-undef value in the shuffle mask.
574 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
577 if (i == 16) return -1; // all undef.
579 // Otherwise, check to see if the rest of the elements are consecutively
580 // numbered from this value.
581 unsigned ShiftAmt = SVOp->getMaskElt(i);
582 if (ShiftAmt < i) return -1;
586 // Check the rest of the elements to see if they are consecutive.
587 for (++i; i != 16; ++i)
588 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
591 // Check the rest of the elements to see if they are consecutive.
592 for (++i; i != 16; ++i)
593 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
599 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
600 /// specifies a splat of a single element that is suitable for input to
601 /// VSPLTB/VSPLTH/VSPLTW.
602 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
603 assert(N->getValueType(0) == MVT::v16i8 &&
604 (EltSize == 1 || EltSize == 2 || EltSize == 4));
606 // This is a splat operation if each element of the permute is the same, and
607 // if the value doesn't reference the second vector.
608 unsigned ElementBase = N->getMaskElt(0);
610 // FIXME: Handle UNDEF elements too!
611 if (ElementBase >= 16)
614 // Check that the indices are consecutive, in the case of a multi-byte element
615 // splatted with a v16i8 mask.
616 for (unsigned i = 1; i != EltSize; ++i)
617 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
620 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
621 if (N->getMaskElt(i) < 0) continue;
622 for (unsigned j = 0; j != EltSize; ++j)
623 if (N->getMaskElt(i+j) != N->getMaskElt(j))
629 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
631 bool PPC::isAllNegativeZeroVector(SDNode *N) {
632 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
634 APInt APVal, APUndef;
638 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
639 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
640 return CFP->getValueAPF().isNegZero();
645 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
646 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
647 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
648 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
649 assert(isSplatShuffleMask(SVOp, EltSize));
650 return SVOp->getMaskElt(0) / EltSize;
653 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
654 /// by using a vspltis[bhw] instruction of the specified element size, return
655 /// the constant being splatted. The ByteSize field indicates the number of
656 /// bytes of each element [124] -> [bhw].
657 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
660 // If ByteSize of the splat is bigger than the element size of the
661 // build_vector, then we have a case where we are checking for a splat where
662 // multiple elements of the buildvector are folded together into a single
663 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
664 unsigned EltSize = 16/N->getNumOperands();
665 if (EltSize < ByteSize) {
666 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
667 SDValue UniquedVals[4];
668 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
670 // See if all of the elements in the buildvector agree across.
671 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
672 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
673 // If the element isn't a constant, bail fully out.
674 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
677 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
678 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
679 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
680 return SDValue(); // no match.
683 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
684 // either constant or undef values that are identical for each chunk. See
685 // if these chunks can form into a larger vspltis*.
687 // Check to see if all of the leading entries are either 0 or -1. If
688 // neither, then this won't fit into the immediate field.
689 bool LeadingZero = true;
690 bool LeadingOnes = true;
691 for (unsigned i = 0; i != Multiple-1; ++i) {
692 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
694 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
695 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
697 // Finally, check the least significant entry.
699 if (UniquedVals[Multiple-1].getNode() == 0)
700 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
701 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
703 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
706 if (UniquedVals[Multiple-1].getNode() == 0)
707 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
708 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
709 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
710 return DAG.getTargetConstant(Val, MVT::i32);
716 // Check to see if this buildvec has a single non-undef value in its elements.
717 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
718 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
719 if (OpVal.getNode() == 0)
720 OpVal = N->getOperand(i);
721 else if (OpVal != N->getOperand(i))
725 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
727 unsigned ValSizeInBytes = EltSize;
729 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
730 Value = CN->getZExtValue();
731 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
732 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
733 Value = FloatToBits(CN->getValueAPF().convertToFloat());
736 // If the splat value is larger than the element value, then we can never do
737 // this splat. The only case that we could fit the replicated bits into our
738 // immediate field for would be zero, and we prefer to use vxor for it.
739 if (ValSizeInBytes < ByteSize) return SDValue();
741 // If the element value is larger than the splat value, cut it in half and
742 // check to see if the two halves are equal. Continue doing this until we
743 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
744 while (ValSizeInBytes > ByteSize) {
745 ValSizeInBytes >>= 1;
747 // If the top half equals the bottom half, we're still ok.
748 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
749 (Value & ((1 << (8*ValSizeInBytes))-1)))
753 // Properly sign extend the value.
754 int ShAmt = (4-ByteSize)*8;
755 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
757 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
758 if (MaskVal == 0) return SDValue();
760 // Finally, if this value fits in a 5 bit sext field, return it
761 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
762 return DAG.getTargetConstant(MaskVal, MVT::i32);
766 //===----------------------------------------------------------------------===//
767 // Addressing Mode Selection
768 //===----------------------------------------------------------------------===//
770 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
771 /// or 64-bit immediate, and if the value can be accurately represented as a
772 /// sign extension from a 16-bit value. If so, this returns true and the
774 static bool isIntS16Immediate(SDNode *N, short &Imm) {
775 if (N->getOpcode() != ISD::Constant)
778 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
779 if (N->getValueType(0) == MVT::i32)
780 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
782 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
784 static bool isIntS16Immediate(SDValue Op, short &Imm) {
785 return isIntS16Immediate(Op.getNode(), Imm);
789 /// SelectAddressRegReg - Given the specified addressed, check to see if it
790 /// can be represented as an indexed [r+r] operation. Returns false if it
791 /// can be more efficiently represented with [r+imm].
792 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
794 SelectionDAG &DAG) const {
796 if (N.getOpcode() == ISD::ADD) {
797 if (isIntS16Immediate(N.getOperand(1), imm))
799 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
802 Base = N.getOperand(0);
803 Index = N.getOperand(1);
805 } else if (N.getOpcode() == ISD::OR) {
806 if (isIntS16Immediate(N.getOperand(1), imm))
807 return false; // r+i can fold it if we can.
809 // If this is an or of disjoint bitfields, we can codegen this as an add
810 // (for better address arithmetic) if the LHS and RHS of the OR are provably
812 APInt LHSKnownZero, LHSKnownOne;
813 APInt RHSKnownZero, RHSKnownOne;
814 DAG.ComputeMaskedBits(N.getOperand(0),
815 APInt::getAllOnesValue(N.getOperand(0)
816 .getValueSizeInBits()),
817 LHSKnownZero, LHSKnownOne);
819 if (LHSKnownZero.getBoolValue()) {
820 DAG.ComputeMaskedBits(N.getOperand(1),
821 APInt::getAllOnesValue(N.getOperand(1)
822 .getValueSizeInBits()),
823 RHSKnownZero, RHSKnownOne);
824 // If all of the bits are known zero on the LHS or RHS, the add won't
826 if (~(LHSKnownZero | RHSKnownZero) == 0) {
827 Base = N.getOperand(0);
828 Index = N.getOperand(1);
837 /// Returns true if the address N can be represented by a base register plus
838 /// a signed 16-bit displacement [r+imm], and if it is not better
839 /// represented as reg+reg.
840 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
842 SelectionDAG &DAG) const {
843 // FIXME dl should come from parent load or store, not from address
844 DebugLoc dl = N.getDebugLoc();
845 // If this can be more profitably realized as r+r, fail.
846 if (SelectAddressRegReg(N, Disp, Base, DAG))
849 if (N.getOpcode() == ISD::ADD) {
851 if (isIntS16Immediate(N.getOperand(1), imm)) {
852 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
853 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
854 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
856 Base = N.getOperand(0);
858 return true; // [r+i]
859 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
860 // Match LOAD (ADD (X, Lo(G))).
861 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
862 && "Cannot handle constant offsets yet!");
863 Disp = N.getOperand(1).getOperand(0); // The global address.
864 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
865 Disp.getOpcode() == ISD::TargetConstantPool ||
866 Disp.getOpcode() == ISD::TargetJumpTable);
867 Base = N.getOperand(0);
868 return true; // [&g+r]
870 } else if (N.getOpcode() == ISD::OR) {
872 if (isIntS16Immediate(N.getOperand(1), imm)) {
873 // If this is an or of disjoint bitfields, we can codegen this as an add
874 // (for better address arithmetic) if the LHS and RHS of the OR are
875 // provably disjoint.
876 APInt LHSKnownZero, LHSKnownOne;
877 DAG.ComputeMaskedBits(N.getOperand(0),
878 APInt::getAllOnesValue(N.getOperand(0)
879 .getValueSizeInBits()),
880 LHSKnownZero, LHSKnownOne);
882 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
883 // If all of the bits are known zero on the LHS or RHS, the add won't
885 Base = N.getOperand(0);
886 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
890 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
891 // Loading from a constant address.
893 // If this address fits entirely in a 16-bit sext immediate field, codegen
896 if (isIntS16Immediate(CN, Imm)) {
897 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
898 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
899 CN->getValueType(0));
903 // Handle 32-bit sext immediates with LIS + addr mode.
904 if (CN->getValueType(0) == MVT::i32 ||
905 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
906 int Addr = (int)CN->getZExtValue();
908 // Otherwise, break this down into an LIS + disp.
909 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
911 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
912 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
913 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
918 Disp = DAG.getTargetConstant(0, getPointerTy());
919 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
920 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
923 return true; // [r+0]
926 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
927 /// represented as an indexed [r+r] operation.
928 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
930 SelectionDAG &DAG) const {
931 // Check to see if we can easily represent this as an [r+r] address. This
932 // will fail if it thinks that the address is more profitably represented as
933 // reg+imm, e.g. where imm = 0.
934 if (SelectAddressRegReg(N, Base, Index, DAG))
937 // If the operand is an addition, always emit this as [r+r], since this is
938 // better (for code size, and execution, as the memop does the add for free)
939 // than emitting an explicit add.
940 if (N.getOpcode() == ISD::ADD) {
941 Base = N.getOperand(0);
942 Index = N.getOperand(1);
946 // Otherwise, do it the hard way, using R0 as the base register.
947 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
953 /// SelectAddressRegImmShift - Returns true if the address N can be
954 /// represented by a base register plus a signed 14-bit displacement
955 /// [r+imm*4]. Suitable for use by STD and friends.
956 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
958 SelectionDAG &DAG) const {
959 // FIXME dl should come from the parent load or store, not the address
960 DebugLoc dl = N.getDebugLoc();
961 // If this can be more profitably realized as r+r, fail.
962 if (SelectAddressRegReg(N, Disp, Base, DAG))
965 if (N.getOpcode() == ISD::ADD) {
967 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
968 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
969 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
970 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
972 Base = N.getOperand(0);
974 return true; // [r+i]
975 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
976 // Match LOAD (ADD (X, Lo(G))).
977 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
978 && "Cannot handle constant offsets yet!");
979 Disp = N.getOperand(1).getOperand(0); // The global address.
980 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
981 Disp.getOpcode() == ISD::TargetConstantPool ||
982 Disp.getOpcode() == ISD::TargetJumpTable);
983 Base = N.getOperand(0);
984 return true; // [&g+r]
986 } else if (N.getOpcode() == ISD::OR) {
988 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
989 // If this is an or of disjoint bitfields, we can codegen this as an add
990 // (for better address arithmetic) if the LHS and RHS of the OR are
991 // provably disjoint.
992 APInt LHSKnownZero, LHSKnownOne;
993 DAG.ComputeMaskedBits(N.getOperand(0),
994 APInt::getAllOnesValue(N.getOperand(0)
995 .getValueSizeInBits()),
996 LHSKnownZero, LHSKnownOne);
997 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
998 // If all of the bits are known zero on the LHS or RHS, the add won't
1000 Base = N.getOperand(0);
1001 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1005 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1006 // Loading from a constant address. Verify low two bits are clear.
1007 if ((CN->getZExtValue() & 3) == 0) {
1008 // If this address fits entirely in a 14-bit sext immediate field, codegen
1011 if (isIntS16Immediate(CN, Imm)) {
1012 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1013 Base = DAG.getRegister(PPCSubTarget.isPPC64() ? PPC::X0 : PPC::R0,
1014 CN->getValueType(0));
1018 // Fold the low-part of 32-bit absolute addresses into addr mode.
1019 if (CN->getValueType(0) == MVT::i32 ||
1020 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1021 int Addr = (int)CN->getZExtValue();
1023 // Otherwise, break this down into an LIS + disp.
1024 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1025 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1026 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1027 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1033 Disp = DAG.getTargetConstant(0, getPointerTy());
1034 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1035 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1038 return true; // [r+0]
1042 /// getPreIndexedAddressParts - returns true by value, base pointer and
1043 /// offset pointer and addressing mode by reference if the node's address
1044 /// can be legally represented as pre-indexed load / store address.
1045 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1047 ISD::MemIndexedMode &AM,
1048 SelectionDAG &DAG) const {
1049 // Disabled by default for now.
1050 if (!EnablePPCPreinc) return false;
1054 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1055 Ptr = LD->getBasePtr();
1056 VT = LD->getMemoryVT();
1058 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1059 Ptr = ST->getBasePtr();
1060 VT = ST->getMemoryVT();
1064 // PowerPC doesn't have preinc load/store instructions for vectors.
1068 // TODO: Check reg+reg first.
1070 // LDU/STU use reg+imm*4, others use reg+imm.
1071 if (VT != MVT::i64) {
1073 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1077 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1081 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1082 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1083 // sext i32 to i64 when addr mode is r+i.
1084 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1085 LD->getExtensionType() == ISD::SEXTLOAD &&
1086 isa<ConstantSDNode>(Offset))
1094 //===----------------------------------------------------------------------===//
1095 // LowerOperation implementation
1096 //===----------------------------------------------------------------------===//
1098 /// GetLabelAccessInfo - Return true if we should reference labels using a
1099 /// PICBase, set the HiOpFlags and LoOpFlags to the target MO flags.
1100 static bool GetLabelAccessInfo(const TargetMachine &TM, unsigned &HiOpFlags,
1101 unsigned &LoOpFlags, const GlobalValue *GV = 0) {
1102 HiOpFlags = PPCII::MO_HA16;
1103 LoOpFlags = PPCII::MO_LO16;
1105 // Don't use the pic base if not in PIC relocation model. Or if we are on a
1106 // non-darwin platform. We don't support PIC on other platforms yet.
1107 bool isPIC = TM.getRelocationModel() == Reloc::PIC_ &&
1108 TM.getSubtarget<PPCSubtarget>().isDarwin();
1110 HiOpFlags |= PPCII::MO_PIC_FLAG;
1111 LoOpFlags |= PPCII::MO_PIC_FLAG;
1114 // If this is a reference to a global value that requires a non-lazy-ptr, make
1115 // sure that instruction lowering adds it.
1116 if (GV && TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM)) {
1117 HiOpFlags |= PPCII::MO_NLP_FLAG;
1118 LoOpFlags |= PPCII::MO_NLP_FLAG;
1120 if (GV->hasHiddenVisibility()) {
1121 HiOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1122 LoOpFlags |= PPCII::MO_NLP_HIDDEN_FLAG;
1129 static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC,
1130 SelectionDAG &DAG) {
1131 EVT PtrVT = HiPart.getValueType();
1132 SDValue Zero = DAG.getConstant(0, PtrVT);
1133 DebugLoc DL = HiPart.getDebugLoc();
1135 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero);
1136 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero);
1138 // With PIC, the first instruction is actually "GR+hi(&G)".
1140 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1141 DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi);
1143 // Generate non-pic code that has direct accesses to the constant pool.
1144 // The address of the global is just (hi(&g)+lo(&g)).
1145 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1148 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1149 SelectionDAG &DAG) const {
1150 EVT PtrVT = Op.getValueType();
1151 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1152 const Constant *C = CP->getConstVal();
1154 unsigned MOHiFlag, MOLoFlag;
1155 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1157 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOHiFlag);
1159 DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment(), 0, MOLoFlag);
1160 return LowerLabelRef(CPIHi, CPILo, isPIC, DAG);
1163 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
1164 EVT PtrVT = Op.getValueType();
1165 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1167 unsigned MOHiFlag, MOLoFlag;
1168 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1169 SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag);
1170 SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag);
1171 return LowerLabelRef(JTIHi, JTILo, isPIC, DAG);
1174 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op,
1175 SelectionDAG &DAG) const {
1176 EVT PtrVT = Op.getValueType();
1178 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1180 unsigned MOHiFlag, MOLoFlag;
1181 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag);
1182 SDValue TgtBAHi = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOHiFlag);
1183 SDValue TgtBALo = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true, MOLoFlag);
1184 return LowerLabelRef(TgtBAHi, TgtBALo, isPIC, DAG);
1187 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1188 SelectionDAG &DAG) const {
1189 EVT PtrVT = Op.getValueType();
1190 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1191 DebugLoc DL = GSDN->getDebugLoc();
1192 const GlobalValue *GV = GSDN->getGlobal();
1194 // 64-bit SVR4 ABI code is always position-independent.
1195 // The actual address of the GlobalValue is stored in the TOC.
1196 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1197 SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset());
1198 return DAG.getNode(PPCISD::TOC_ENTRY, DL, MVT::i64, GA,
1199 DAG.getRegister(PPC::X2, MVT::i64));
1202 unsigned MOHiFlag, MOLoFlag;
1203 bool isPIC = GetLabelAccessInfo(DAG.getTarget(), MOHiFlag, MOLoFlag, GV);
1206 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag);
1208 DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag);
1210 SDValue Ptr = LowerLabelRef(GAHi, GALo, isPIC, DAG);
1212 // If the global reference is actually to a non-lazy-pointer, we have to do an
1213 // extra load to get the address of the global.
1214 if (MOHiFlag & PPCII::MO_NLP_FLAG)
1215 Ptr = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo(),
1220 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1221 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1222 DebugLoc dl = Op.getDebugLoc();
1224 // If we're comparing for equality to zero, expose the fact that this is
1225 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1226 // fold the new nodes.
1227 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1228 if (C->isNullValue() && CC == ISD::SETEQ) {
1229 EVT VT = Op.getOperand(0).getValueType();
1230 SDValue Zext = Op.getOperand(0);
1231 if (VT.bitsLT(MVT::i32)) {
1233 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1235 unsigned Log2b = Log2_32(VT.getSizeInBits());
1236 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1237 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1238 DAG.getConstant(Log2b, MVT::i32));
1239 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1241 // Leave comparisons against 0 and -1 alone for now, since they're usually
1242 // optimized. FIXME: revisit this when we can custom lower all setcc
1244 if (C->isAllOnesValue() || C->isNullValue())
1248 // If we have an integer seteq/setne, turn it into a compare against zero
1249 // by xor'ing the rhs with the lhs, which is faster than setting a
1250 // condition register, reading it back out, and masking the correct bit. The
1251 // normal approach here uses sub to do this instead of xor. Using xor exposes
1252 // the result to other bit-twiddling opportunities.
1253 EVT LHSVT = Op.getOperand(0).getValueType();
1254 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1255 EVT VT = Op.getValueType();
1256 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1258 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1263 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1264 const PPCSubtarget &Subtarget) const {
1266 llvm_unreachable("VAARG not yet implemented for the SVR4 ABI!");
1267 return SDValue(); // Not reached
1270 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op,
1271 SelectionDAG &DAG) const {
1272 SDValue Chain = Op.getOperand(0);
1273 SDValue Trmp = Op.getOperand(1); // trampoline
1274 SDValue FPtr = Op.getOperand(2); // nested function
1275 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1276 DebugLoc dl = Op.getDebugLoc();
1278 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1279 bool isPPC64 = (PtrVT == MVT::i64);
1280 const Type *IntPtrTy =
1281 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1284 TargetLowering::ArgListTy Args;
1285 TargetLowering::ArgListEntry Entry;
1287 Entry.Ty = IntPtrTy;
1288 Entry.Node = Trmp; Args.push_back(Entry);
1290 // TrampSize == (isPPC64 ? 48 : 40);
1291 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1292 isPPC64 ? MVT::i64 : MVT::i32);
1293 Args.push_back(Entry);
1295 Entry.Node = FPtr; Args.push_back(Entry);
1296 Entry.Node = Nest; Args.push_back(Entry);
1298 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1299 std::pair<SDValue, SDValue> CallResult =
1300 LowerCallTo(Chain, Op.getValueType().getTypeForEVT(*DAG.getContext()),
1301 false, false, false, false, 0, CallingConv::C, false,
1302 /*isReturnValueUsed=*/true,
1303 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1307 { CallResult.first, CallResult.second };
1309 return DAG.getMergeValues(Ops, 2, dl);
1312 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1313 const PPCSubtarget &Subtarget) const {
1314 MachineFunction &MF = DAG.getMachineFunction();
1315 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1317 DebugLoc dl = Op.getDebugLoc();
1319 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1320 // vastart just stores the address of the VarArgsFrameIndex slot into the
1321 // memory location argument.
1322 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1323 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1324 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1325 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
1326 MachinePointerInfo(SV),
1330 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1331 // We suppose the given va_list is already allocated.
1334 // char gpr; /* index into the array of 8 GPRs
1335 // * stored in the register save area
1336 // * gpr=0 corresponds to r3,
1337 // * gpr=1 to r4, etc.
1339 // char fpr; /* index into the array of 8 FPRs
1340 // * stored in the register save area
1341 // * fpr=0 corresponds to f1,
1342 // * fpr=1 to f2, etc.
1344 // char *overflow_arg_area;
1345 // /* location on stack that holds
1346 // * the next overflow argument
1348 // char *reg_save_area;
1349 // /* where r3:r10 and f1:f8 (if saved)
1355 SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), MVT::i32);
1356 SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), MVT::i32);
1359 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1361 SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(),
1363 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1366 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1367 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1369 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1370 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1372 uint64_t FPROffset = 1;
1373 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1375 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1377 // Store first byte : number of int regs
1378 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1380 MachinePointerInfo(SV),
1381 MVT::i8, false, false, 0);
1382 uint64_t nextOffset = FPROffset;
1383 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1386 // Store second byte : number of float regs
1387 SDValue secondStore =
1388 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr,
1389 MachinePointerInfo(SV, nextOffset), MVT::i8,
1391 nextOffset += StackOffset;
1392 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1394 // Store second word : arguments given on stack
1395 SDValue thirdStore =
1396 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr,
1397 MachinePointerInfo(SV, nextOffset),
1399 nextOffset += FrameOffset;
1400 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1402 // Store third word : arguments given in registers
1403 return DAG.getStore(thirdStore, dl, FR, nextPtr,
1404 MachinePointerInfo(SV, nextOffset),
1409 #include "PPCGenCallingConv.inc"
1411 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1412 CCValAssign::LocInfo &LocInfo,
1413 ISD::ArgFlagsTy &ArgFlags,
1418 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1420 CCValAssign::LocInfo &LocInfo,
1421 ISD::ArgFlagsTy &ArgFlags,
1423 static const unsigned ArgRegs[] = {
1424 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1425 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1427 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1429 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1431 // Skip one register if the first unallocated register has an even register
1432 // number and there are still argument registers available which have not been
1433 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1434 // need to skip a register if RegNum is odd.
1435 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1436 State.AllocateReg(ArgRegs[RegNum]);
1439 // Always return false here, as this function only makes sure that the first
1440 // unallocated register has an odd register number and does not actually
1441 // allocate a register for the current argument.
1445 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1447 CCValAssign::LocInfo &LocInfo,
1448 ISD::ArgFlagsTy &ArgFlags,
1450 static const unsigned ArgRegs[] = {
1451 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1455 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1457 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1459 // If there is only one Floating-point register left we need to put both f64
1460 // values of a split ppc_fp128 value on the stack.
1461 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1462 State.AllocateReg(ArgRegs[RegNum]);
1465 // Always return false here, as this function only makes sure that the two f64
1466 // values a ppc_fp128 value is split into are both passed in registers or both
1467 // passed on the stack and does not actually allocate a register for the
1468 // current argument.
1472 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1474 static const unsigned *GetFPR() {
1475 static const unsigned FPR[] = {
1476 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1477 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1483 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1485 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1486 unsigned PtrByteSize) {
1487 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1488 if (Flags.isByVal())
1489 ArgSize = Flags.getByValSize();
1490 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1496 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1497 CallingConv::ID CallConv, bool isVarArg,
1498 const SmallVectorImpl<ISD::InputArg>
1500 DebugLoc dl, SelectionDAG &DAG,
1501 SmallVectorImpl<SDValue> &InVals)
1503 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1504 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1507 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1513 PPCTargetLowering::LowerFormalArguments_SVR4(
1515 CallingConv::ID CallConv, bool isVarArg,
1516 const SmallVectorImpl<ISD::InputArg>
1518 DebugLoc dl, SelectionDAG &DAG,
1519 SmallVectorImpl<SDValue> &InVals) const {
1521 // 32-bit SVR4 ABI Stack Frame Layout:
1522 // +-----------------------------------+
1523 // +--> | Back chain |
1524 // | +-----------------------------------+
1525 // | | Floating-point register save area |
1526 // | +-----------------------------------+
1527 // | | General register save area |
1528 // | +-----------------------------------+
1529 // | | CR save word |
1530 // | +-----------------------------------+
1531 // | | VRSAVE save word |
1532 // | +-----------------------------------+
1533 // | | Alignment padding |
1534 // | +-----------------------------------+
1535 // | | Vector register save area |
1536 // | +-----------------------------------+
1537 // | | Local variable space |
1538 // | +-----------------------------------+
1539 // | | Parameter list area |
1540 // | +-----------------------------------+
1541 // | | LR save word |
1542 // | +-----------------------------------+
1543 // SP--> +--- | Back chain |
1544 // +-----------------------------------+
1547 // System V Application Binary Interface PowerPC Processor Supplement
1548 // AltiVec Technology Programming Interface Manual
1550 MachineFunction &MF = DAG.getMachineFunction();
1551 MachineFrameInfo *MFI = MF.getFrameInfo();
1552 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1554 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1555 // Potential tail calls could cause overwriting of argument stack slots.
1556 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1557 unsigned PtrByteSize = 4;
1559 // Assign locations to all of the incoming arguments.
1560 SmallVector<CCValAssign, 16> ArgLocs;
1561 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1562 getTargetMachine(), ArgLocs, *DAG.getContext());
1564 // Reserve space for the linkage area on the stack.
1565 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
1567 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1569 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1570 CCValAssign &VA = ArgLocs[i];
1572 // Arguments stored in registers.
1573 if (VA.isRegLoc()) {
1574 TargetRegisterClass *RC;
1575 EVT ValVT = VA.getValVT();
1577 switch (ValVT.getSimpleVT().SimpleTy) {
1579 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1581 RC = PPC::GPRCRegisterClass;
1584 RC = PPC::F4RCRegisterClass;
1587 RC = PPC::F8RCRegisterClass;
1593 RC = PPC::VRRCRegisterClass;
1597 // Transform the arguments stored in physical registers into virtual ones.
1598 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1599 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1601 InVals.push_back(ArgValue);
1603 // Argument stored in memory.
1604 assert(VA.isMemLoc());
1606 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1607 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1610 // Create load nodes to retrieve arguments from the stack.
1611 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1612 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
1613 MachinePointerInfo(),
1618 // Assign locations to all of the incoming aggregate by value arguments.
1619 // Aggregates passed by value are stored in the local variable space of the
1620 // caller's stack frame, right above the parameter list area.
1621 SmallVector<CCValAssign, 16> ByValArgLocs;
1622 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1623 getTargetMachine(), ByValArgLocs, *DAG.getContext());
1625 // Reserve stack space for the allocations in CCInfo.
1626 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1628 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1630 // Area that is at least reserved in the caller of this function.
1631 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1633 // Set the size that is at least reserved in caller of this function. Tail
1634 // call optimized function's reserved stack space needs to be aligned so that
1635 // taking the difference between two stack areas will result in an aligned
1637 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1640 std::max(MinReservedArea,
1641 PPCFrameLowering::getMinCallFrameSize(false, false));
1643 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
1644 getStackAlignment();
1645 unsigned AlignMask = TargetAlign-1;
1646 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1648 FI->setMinReservedArea(MinReservedArea);
1650 SmallVector<SDValue, 8> MemOps;
1652 // If the function takes variable number of arguments, make a frame index for
1653 // the start of the first vararg value... for expansion of llvm.va_start.
1655 static const unsigned GPArgRegs[] = {
1656 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1657 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1659 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1661 static const unsigned FPArgRegs[] = {
1662 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1665 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1667 FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs,
1669 FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs,
1672 // Make room for NumGPArgRegs and NumFPArgRegs.
1673 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1674 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1676 FuncInfo->setVarArgsStackOffset(
1677 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1678 CCInfo.getNextStackOffset(), true));
1680 FuncInfo->setVarArgsFrameIndex(MFI->CreateStackObject(Depth, 8, false));
1681 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1683 // The fixed integer arguments of a variadic function are stored to the
1684 // VarArgsFrameIndex on the stack so that they may be loaded by deferencing
1685 // the result of va_next.
1686 for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1687 // Get an existing live-in vreg, or add a new one.
1688 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]);
1690 VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1692 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1693 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1694 MachinePointerInfo(), false, false, 0);
1695 MemOps.push_back(Store);
1696 // Increment the address by four for the next argument to store
1697 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1698 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1701 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1703 // The double arguments are stored to the VarArgsFrameIndex
1705 for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1706 // Get an existing live-in vreg, or add a new one.
1707 unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]);
1709 VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1711 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1712 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1713 MachinePointerInfo(), false, false, 0);
1714 MemOps.push_back(Store);
1715 // Increment the address by eight for the next argument to store
1716 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1718 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1722 if (!MemOps.empty())
1723 Chain = DAG.getNode(ISD::TokenFactor, dl,
1724 MVT::Other, &MemOps[0], MemOps.size());
1730 PPCTargetLowering::LowerFormalArguments_Darwin(
1732 CallingConv::ID CallConv, bool isVarArg,
1733 const SmallVectorImpl<ISD::InputArg>
1735 DebugLoc dl, SelectionDAG &DAG,
1736 SmallVectorImpl<SDValue> &InVals) const {
1737 // TODO: add description of PPC stack frame format, or at least some docs.
1739 MachineFunction &MF = DAG.getMachineFunction();
1740 MachineFrameInfo *MFI = MF.getFrameInfo();
1741 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
1743 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1744 bool isPPC64 = PtrVT == MVT::i64;
1745 // Potential tail calls could cause overwriting of argument stack slots.
1746 bool isImmutable = !(GuaranteedTailCallOpt && (CallConv==CallingConv::Fast));
1747 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1749 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
1750 // Area that is at least reserved in caller of this function.
1751 unsigned MinReservedArea = ArgOffset;
1753 static const unsigned GPR_32[] = { // 32-bit registers.
1754 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1755 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1757 static const unsigned GPR_64[] = { // 64-bit registers.
1758 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1759 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1762 static const unsigned *FPR = GetFPR();
1764 static const unsigned VR[] = {
1765 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1766 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1769 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1770 const unsigned Num_FPR_Regs = 13;
1771 const unsigned Num_VR_Regs = array_lengthof( VR);
1773 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1775 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1777 // In 32-bit non-varargs functions, the stack space for vectors is after the
1778 // stack space for non-vectors. We do not use this space unless we have
1779 // too many vectors to fit in registers, something that only occurs in
1780 // constructed examples:), but we have to walk the arglist to figure
1781 // that out...for the pathological case, compute VecArgOffset as the
1782 // start of the vector parameter area. Computing VecArgOffset is the
1783 // entire point of the following loop.
1784 unsigned VecArgOffset = ArgOffset;
1785 if (!isVarArg && !isPPC64) {
1786 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1788 EVT ObjectVT = Ins[ArgNo].VT;
1789 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1790 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1792 if (Flags.isByVal()) {
1793 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1794 ObjSize = Flags.getByValSize();
1796 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1797 VecArgOffset += ArgSize;
1801 switch(ObjectVT.getSimpleVT().SimpleTy) {
1802 default: llvm_unreachable("Unhandled argument type!");
1805 VecArgOffset += isPPC64 ? 8 : 4;
1807 case MVT::i64: // PPC64
1815 // Nothing to do, we're only looking at Nonvector args here.
1820 // We've found where the vector parameter area in memory is. Skip the
1821 // first 12 parameters; these don't use that memory.
1822 VecArgOffset = ((VecArgOffset+15)/16)*16;
1823 VecArgOffset += 12*16;
1825 // Add DAG nodes to load the arguments or copy them out of registers. On
1826 // entry to a function on PPC, the arguments start after the linkage area,
1827 // although the first ones are often in registers.
1829 SmallVector<SDValue, 8> MemOps;
1830 unsigned nAltivecParamsAtEnd = 0;
1831 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1833 bool needsLoad = false;
1834 EVT ObjectVT = Ins[ArgNo].VT;
1835 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1836 unsigned ArgSize = ObjSize;
1837 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1839 unsigned CurArgOffset = ArgOffset;
1841 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1842 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1843 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1844 if (isVarArg || isPPC64) {
1845 MinReservedArea = ((MinReservedArea+15)/16)*16;
1846 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1849 } else nAltivecParamsAtEnd++;
1851 // Calculate min reserved area.
1852 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1856 // FIXME the codegen can be much improved in some cases.
1857 // We do not have to keep everything in memory.
1858 if (Flags.isByVal()) {
1859 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1860 ObjSize = Flags.getByValSize();
1861 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1862 // Objects of size 1 and 2 are right justified, everything else is
1863 // left justified. This means the memory address is adjusted forwards.
1864 if (ObjSize==1 || ObjSize==2) {
1865 CurArgOffset = CurArgOffset + (4 - ObjSize);
1867 // The value of the object is its address.
1868 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true);
1869 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1870 InVals.push_back(FIN);
1871 if (ObjSize==1 || ObjSize==2) {
1872 if (GPR_idx != Num_GPR_Regs) {
1875 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1877 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1878 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1879 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1880 MachinePointerInfo(),
1881 ObjSize==1 ? MVT::i8 : MVT::i16,
1883 MemOps.push_back(Store);
1887 ArgOffset += PtrByteSize;
1891 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1892 // Store whatever pieces of the object are in registers
1893 // to memory. ArgVal will be address of the beginning of
1895 if (GPR_idx != Num_GPR_Regs) {
1898 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1900 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1901 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true);
1902 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1903 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1904 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
1905 MachinePointerInfo(),
1907 MemOps.push_back(Store);
1909 ArgOffset += PtrByteSize;
1911 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1918 switch (ObjectVT.getSimpleVT().SimpleTy) {
1919 default: llvm_unreachable("Unhandled argument type!");
1922 if (GPR_idx != Num_GPR_Regs) {
1923 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1924 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1928 ArgSize = PtrByteSize;
1930 // All int arguments reserve stack space in the Darwin ABI.
1931 ArgOffset += PtrByteSize;
1935 case MVT::i64: // PPC64
1936 if (GPR_idx != Num_GPR_Regs) {
1937 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1938 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1940 if (ObjectVT == MVT::i32) {
1941 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1942 // value to MVT::i64 and then truncate to the correct register size.
1944 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1945 DAG.getValueType(ObjectVT));
1946 else if (Flags.isZExt())
1947 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1948 DAG.getValueType(ObjectVT));
1950 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1956 ArgSize = PtrByteSize;
1958 // All int arguments reserve stack space in the Darwin ABI.
1964 // Every 4 bytes of argument space consumes one of the GPRs available for
1965 // argument passing.
1966 if (GPR_idx != Num_GPR_Regs) {
1968 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1971 if (FPR_idx != Num_FPR_Regs) {
1974 if (ObjectVT == MVT::f32)
1975 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
1977 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
1979 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1985 // All FP arguments reserve stack space in the Darwin ABI.
1986 ArgOffset += isPPC64 ? 8 : ObjSize;
1992 // Note that vector arguments in registers don't reserve stack space,
1993 // except in varargs functions.
1994 if (VR_idx != Num_VR_Regs) {
1995 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
1996 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1998 while ((ArgOffset % 16) != 0) {
1999 ArgOffset += PtrByteSize;
2000 if (GPR_idx != Num_GPR_Regs)
2004 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2008 if (!isVarArg && !isPPC64) {
2009 // Vectors go after all the nonvectors.
2010 CurArgOffset = VecArgOffset;
2013 // Vectors are aligned.
2014 ArgOffset = ((ArgOffset+15)/16)*16;
2015 CurArgOffset = ArgOffset;
2023 // We need to load the argument to a virtual register if we determined above
2024 // that we ran out of physical registers of the appropriate type.
2026 int FI = MFI->CreateFixedObject(ObjSize,
2027 CurArgOffset + (ArgSize - ObjSize),
2029 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2030 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
2034 InVals.push_back(ArgVal);
2037 // Set the size that is at least reserved in caller of this function. Tail
2038 // call optimized function's reserved stack space needs to be aligned so that
2039 // taking the difference between two stack areas will result in an aligned
2041 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2042 // Add the Altivec parameters at the end, if needed.
2043 if (nAltivecParamsAtEnd) {
2044 MinReservedArea = ((MinReservedArea+15)/16)*16;
2045 MinReservedArea += 16*nAltivecParamsAtEnd;
2048 std::max(MinReservedArea,
2049 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2050 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2051 getStackAlignment();
2052 unsigned AlignMask = TargetAlign-1;
2053 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2054 FI->setMinReservedArea(MinReservedArea);
2056 // If the function takes variable number of arguments, make a frame index for
2057 // the start of the first vararg value... for expansion of llvm.va_start.
2059 int Depth = ArgOffset;
2061 FuncInfo->setVarArgsFrameIndex(
2062 MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2064 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2066 // If this function is vararg, store any remaining integer argument regs
2067 // to their spots on the stack so that they may be loaded by deferencing the
2068 // result of va_next.
2069 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2073 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2075 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2077 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2078 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
2079 MachinePointerInfo(), false, false, 0);
2080 MemOps.push_back(Store);
2081 // Increment the address by four for the next argument to store
2082 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2083 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2087 if (!MemOps.empty())
2088 Chain = DAG.getNode(ISD::TokenFactor, dl,
2089 MVT::Other, &MemOps[0], MemOps.size());
2094 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2095 /// linkage area for the Darwin ABI.
2097 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2101 const SmallVectorImpl<ISD::OutputArg>
2103 const SmallVectorImpl<SDValue> &OutVals,
2104 unsigned &nAltivecParamsAtEnd) {
2105 // Count how many bytes are to be pushed on the stack, including the linkage
2106 // area, and parameter passing area. We start with 24/48 bytes, which is
2107 // prereserved space for [SP][CR][LR][3 x unused].
2108 unsigned NumBytes = PPCFrameLowering::getLinkageSize(isPPC64, true);
2109 unsigned NumOps = Outs.size();
2110 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2112 // Add up all the space actually used.
2113 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2114 // they all go in registers, but we must reserve stack space for them for
2115 // possible use by the caller. In varargs or 64-bit calls, parameters are
2116 // assigned stack space in order, with padding so Altivec parameters are
2118 nAltivecParamsAtEnd = 0;
2119 for (unsigned i = 0; i != NumOps; ++i) {
2120 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2121 EVT ArgVT = Outs[i].VT;
2122 // Varargs Altivec parameters are padded to a 16 byte boundary.
2123 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2124 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2125 if (!isVarArg && !isPPC64) {
2126 // Non-varargs Altivec parameters go after all the non-Altivec
2127 // parameters; handle those later so we know how much padding we need.
2128 nAltivecParamsAtEnd++;
2131 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2132 NumBytes = ((NumBytes+15)/16)*16;
2134 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2137 // Allow for Altivec parameters at the end, if needed.
2138 if (nAltivecParamsAtEnd) {
2139 NumBytes = ((NumBytes+15)/16)*16;
2140 NumBytes += 16*nAltivecParamsAtEnd;
2143 // The prolog code of the callee may store up to 8 GPR argument registers to
2144 // the stack, allowing va_start to index over them in memory if its varargs.
2145 // Because we cannot tell if this is needed on the caller side, we have to
2146 // conservatively assume that it is needed. As such, make sure we have at
2147 // least enough stack space for the caller to store the 8 GPRs.
2148 NumBytes = std::max(NumBytes,
2149 PPCFrameLowering::getMinCallFrameSize(isPPC64, true));
2151 // Tail call needs the stack to be aligned.
2152 if (CC==CallingConv::Fast && GuaranteedTailCallOpt) {
2153 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameLowering()->
2154 getStackAlignment();
2155 unsigned AlignMask = TargetAlign-1;
2156 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2162 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2163 /// adjusted to accommodate the arguments for the tailcall.
2164 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall,
2165 unsigned ParamSize) {
2167 if (!isTailCall) return 0;
2169 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2170 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2171 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2172 // Remember only if the new adjustement is bigger.
2173 if (SPDiff < FI->getTailCallSPDelta())
2174 FI->setTailCallSPDelta(SPDiff);
2179 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2180 /// for tail call optimization. Targets which want to do tail call
2181 /// optimization should implement this function.
2183 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2184 CallingConv::ID CalleeCC,
2186 const SmallVectorImpl<ISD::InputArg> &Ins,
2187 SelectionDAG& DAG) const {
2188 if (!GuaranteedTailCallOpt)
2191 // Variable argument functions are not supported.
2195 MachineFunction &MF = DAG.getMachineFunction();
2196 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2197 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2198 // Functions containing by val parameters are not supported.
2199 for (unsigned i = 0; i != Ins.size(); i++) {
2200 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2201 if (Flags.isByVal()) return false;
2204 // Non PIC/GOT tail calls are supported.
2205 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2208 // At the moment we can only do local tail calls (in same module, hidden
2209 // or protected) if we are generating PIC.
2210 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2211 return G->getGlobal()->hasHiddenVisibility()
2212 || G->getGlobal()->hasProtectedVisibility();
2218 /// isCallCompatibleAddress - Return the immediate to use if the specified
2219 /// 32-bit value is representable in the immediate field of a BxA instruction.
2220 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2221 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2224 int Addr = C->getZExtValue();
2225 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2226 (Addr << 6 >> 6) != Addr)
2227 return 0; // Top 6 bits have to be sext of immediate.
2229 return DAG.getConstant((int)C->getZExtValue() >> 2,
2230 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2235 struct TailCallArgumentInfo {
2240 TailCallArgumentInfo() : FrameIdx(0) {}
2245 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2247 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2249 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2250 SmallVector<SDValue, 8> &MemOpChains,
2252 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2253 SDValue Arg = TailCallArgs[i].Arg;
2254 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2255 int FI = TailCallArgs[i].FrameIdx;
2256 // Store relative to framepointer.
2257 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2258 MachinePointerInfo::getFixedStack(FI),
2263 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2264 /// the appropriate stack slot for the tail call optimized function call.
2265 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2266 MachineFunction &MF,
2275 // Calculate the new stack slot for the return address.
2276 int SlotSize = isPPC64 ? 8 : 4;
2277 int NewRetAddrLoc = SPDiff + PPCFrameLowering::getReturnSaveOffset(isPPC64,
2279 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2280 NewRetAddrLoc, true);
2281 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2282 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2283 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2284 MachinePointerInfo::getFixedStack(NewRetAddr),
2287 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2288 // slot as the FP is never overwritten.
2291 SPDiff + PPCFrameLowering::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2292 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2294 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2295 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2296 MachinePointerInfo::getFixedStack(NewFPIdx),
2303 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2304 /// the position of the argument.
2306 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2307 SDValue Arg, int SPDiff, unsigned ArgOffset,
2308 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2309 int Offset = ArgOffset + SPDiff;
2310 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2311 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2312 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2313 SDValue FIN = DAG.getFrameIndex(FI, VT);
2314 TailCallArgumentInfo Info;
2316 Info.FrameIdxOp = FIN;
2318 TailCallArguments.push_back(Info);
2321 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2322 /// stack slot. Returns the chain as result and the loaded frame pointers in
2323 /// LROpOut/FPOpout. Used when tail calling.
2324 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2330 DebugLoc dl) const {
2332 // Load the LR and FP stack slot for later adjusting.
2333 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2334 LROpOut = getReturnAddrFrameIndex(DAG);
2335 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo(),
2337 Chain = SDValue(LROpOut.getNode(), 1);
2339 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2340 // slot as the FP is never overwritten.
2342 FPOpOut = getFramePointerFrameIndex(DAG);
2343 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, MachinePointerInfo(),
2345 Chain = SDValue(FPOpOut.getNode(), 1);
2351 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2352 /// by "Src" to address "Dst" of size "Size". Alignment information is
2353 /// specified by the specific parameter attribute. The copy will be passed as
2354 /// a byval function parameter.
2355 /// Sometimes what we are copying is the end of a larger object, the part that
2356 /// does not fit in registers.
2358 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2359 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2361 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2362 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2363 false, false, MachinePointerInfo(0),
2364 MachinePointerInfo(0));
2367 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2370 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2371 SDValue Arg, SDValue PtrOff, int SPDiff,
2372 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2373 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2374 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments,
2376 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2381 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2383 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2384 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2385 DAG.getConstant(ArgOffset, PtrVT));
2387 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2388 MachinePointerInfo(), false, false, 0));
2389 // Calculate and remember argument location.
2390 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2395 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2396 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2397 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2398 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2399 MachineFunction &MF = DAG.getMachineFunction();
2401 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2402 // might overwrite each other in case of tail call optimization.
2403 SmallVector<SDValue, 8> MemOpChains2;
2404 // Do not flag preceding copytoreg stuff together with the following stuff.
2406 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2408 if (!MemOpChains2.empty())
2409 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2410 &MemOpChains2[0], MemOpChains2.size());
2412 // Store the return address to the appropriate stack slot.
2413 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2414 isPPC64, isDarwinABI, dl);
2416 // Emit callseq_end just before tailcall node.
2417 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2418 DAG.getIntPtrConstant(0, true), InFlag);
2419 InFlag = Chain.getValue(1);
2423 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2424 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2425 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2426 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2427 const PPCSubtarget &PPCSubTarget) {
2429 bool isPPC64 = PPCSubTarget.isPPC64();
2430 bool isSVR4ABI = PPCSubTarget.isSVR4ABI();
2432 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2433 NodeTys.push_back(MVT::Other); // Returns a chain
2434 NodeTys.push_back(MVT::Glue); // Returns a flag for retval copy to use.
2436 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2438 bool needIndirectCall = true;
2439 if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) {
2440 // If this is an absolute destination address, use the munged value.
2441 Callee = SDValue(Dest, 0);
2442 needIndirectCall = false;
2445 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2446 // XXX Work around for http://llvm.org/bugs/show_bug.cgi?id=5201
2447 // Use indirect calls for ALL functions calls in JIT mode, since the
2448 // far-call stubs may be outside relocation limits for a BL instruction.
2449 if (!DAG.getTarget().getSubtarget<PPCSubtarget>().isJITCodeModel()) {
2450 unsigned OpFlags = 0;
2451 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2452 (!PPCSubTarget.getTargetTriple().isMacOSX() ||
2453 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5)) &&
2454 (G->getGlobal()->isDeclaration() ||
2455 G->getGlobal()->isWeakForLinker())) {
2456 // PC-relative references to external symbols should go through $stub,
2457 // unless we're building with the leopard linker or later, which
2458 // automatically synthesizes these stubs.
2459 OpFlags = PPCII::MO_DARWIN_STUB;
2462 // If the callee is a GlobalAddress/ExternalSymbol node (quite common,
2463 // every direct call is) turn it into a TargetGlobalAddress /
2464 // TargetExternalSymbol node so that legalize doesn't hack it.
2465 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl,
2466 Callee.getValueType(),
2468 needIndirectCall = false;
2472 if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2473 unsigned char OpFlags = 0;
2475 if (DAG.getTarget().getRelocationModel() != Reloc::Static &&
2476 (!PPCSubTarget.getTargetTriple().isMacOSX() ||
2477 PPCSubTarget.getTargetTriple().isMacOSXVersionLT(10, 5))) {
2478 // PC-relative references to external symbols should go through $stub,
2479 // unless we're building with the leopard linker or later, which
2480 // automatically synthesizes these stubs.
2481 OpFlags = PPCII::MO_DARWIN_STUB;
2484 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType(),
2486 needIndirectCall = false;
2489 if (needIndirectCall) {
2490 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2491 // to do the call, we can't use PPCISD::CALL.
2492 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2494 if (isSVR4ABI && isPPC64) {
2495 // Function pointers in the 64-bit SVR4 ABI do not point to the function
2496 // entry point, but to the function descriptor (the function entry point
2497 // address is part of the function descriptor though).
2498 // The function descriptor is a three doubleword structure with the
2499 // following fields: function entry point, TOC base address and
2500 // environment pointer.
2501 // Thus for a call through a function pointer, the following actions need
2503 // 1. Save the TOC of the caller in the TOC save area of its stack
2504 // frame (this is done in LowerCall_Darwin()).
2505 // 2. Load the address of the function entry point from the function
2507 // 3. Load the TOC of the callee from the function descriptor into r2.
2508 // 4. Load the environment pointer from the function descriptor into
2510 // 5. Branch to the function entry point address.
2511 // 6. On return of the callee, the TOC of the caller needs to be
2512 // restored (this is done in FinishCall()).
2514 // All those operations are flagged together to ensure that no other
2515 // operations can be scheduled in between. E.g. without flagging the
2516 // operations together, a TOC access in the caller could be scheduled
2517 // between the load of the callee TOC and the branch to the callee, which
2518 // results in the TOC access going through the TOC of the callee instead
2519 // of going through the TOC of the caller, which leads to incorrect code.
2521 // Load the address of the function entry point from the function
2523 SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Other, MVT::Glue);
2524 SDValue LoadFuncPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, MTCTROps,
2525 InFlag.getNode() ? 3 : 2);
2526 Chain = LoadFuncPtr.getValue(1);
2527 InFlag = LoadFuncPtr.getValue(2);
2529 // Load environment pointer into r11.
2530 // Offset of the environment pointer within the function descriptor.
2531 SDValue PtrOff = DAG.getIntPtrConstant(16);
2533 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, MVT::i64, Callee, PtrOff);
2534 SDValue LoadEnvPtr = DAG.getNode(PPCISD::LOAD, dl, VTs, Chain, AddPtr,
2536 Chain = LoadEnvPtr.getValue(1);
2537 InFlag = LoadEnvPtr.getValue(2);
2539 SDValue EnvVal = DAG.getCopyToReg(Chain, dl, PPC::X11, LoadEnvPtr,
2541 Chain = EnvVal.getValue(0);
2542 InFlag = EnvVal.getValue(1);
2544 // Load TOC of the callee into r2. We are using a target-specific load
2545 // with r2 hard coded, because the result of a target-independent load
2546 // would never go directly into r2, since r2 is a reserved register (which
2547 // prevents the register allocator from allocating it), resulting in an
2548 // additional register being allocated and an unnecessary move instruction
2550 VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2551 SDValue LoadTOCPtr = DAG.getNode(PPCISD::LOAD_TOC, dl, VTs, Chain,
2553 Chain = LoadTOCPtr.getValue(0);
2554 InFlag = LoadTOCPtr.getValue(1);
2556 MTCTROps[0] = Chain;
2557 MTCTROps[1] = LoadFuncPtr;
2558 MTCTROps[2] = InFlag;
2561 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2562 2 + (InFlag.getNode() != 0));
2563 InFlag = Chain.getValue(1);
2566 NodeTys.push_back(MVT::Other);
2567 NodeTys.push_back(MVT::Glue);
2568 Ops.push_back(Chain);
2569 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2571 // Add CTR register as callee so a bctr can be emitted later.
2573 Ops.push_back(DAG.getRegister(isPPC64 ? PPC::CTR8 : PPC::CTR, PtrVT));
2576 // If this is a direct call, pass the chain and the callee.
2577 if (Callee.getNode()) {
2578 Ops.push_back(Chain);
2579 Ops.push_back(Callee);
2581 // If this is a tail call add stack pointer delta.
2583 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2585 // Add argument registers to the end of the list so that they are known live
2587 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2588 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2589 RegsToPass[i].second.getValueType()));
2595 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2596 CallingConv::ID CallConv, bool isVarArg,
2597 const SmallVectorImpl<ISD::InputArg> &Ins,
2598 DebugLoc dl, SelectionDAG &DAG,
2599 SmallVectorImpl<SDValue> &InVals) const {
2601 SmallVector<CCValAssign, 16> RVLocs;
2602 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2603 getTargetMachine(), RVLocs, *DAG.getContext());
2604 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2606 // Copy all of the result registers out of their specified physreg.
2607 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2608 CCValAssign &VA = RVLocs[i];
2609 EVT VT = VA.getValVT();
2610 assert(VA.isRegLoc() && "Can only return in registers!");
2611 Chain = DAG.getCopyFromReg(Chain, dl,
2612 VA.getLocReg(), VT, InFlag).getValue(1);
2613 InVals.push_back(Chain.getValue(0));
2614 InFlag = Chain.getValue(2);
2621 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2622 bool isTailCall, bool isVarArg,
2624 SmallVector<std::pair<unsigned, SDValue>, 8>
2626 SDValue InFlag, SDValue Chain,
2628 int SPDiff, unsigned NumBytes,
2629 const SmallVectorImpl<ISD::InputArg> &Ins,
2630 SmallVectorImpl<SDValue> &InVals) const {
2631 std::vector<EVT> NodeTys;
2632 SmallVector<SDValue, 8> Ops;
2633 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2634 isTailCall, RegsToPass, Ops, NodeTys,
2637 // When performing tail call optimization the callee pops its arguments off
2638 // the stack. Account for this here so these bytes can be pushed back on in
2639 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2640 int BytesCalleePops =
2641 (CallConv==CallingConv::Fast && GuaranteedTailCallOpt) ? NumBytes : 0;
2643 if (InFlag.getNode())
2644 Ops.push_back(InFlag);
2648 // If this is the first return lowered for this function, add the regs
2649 // to the liveout set for the function.
2650 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2651 SmallVector<CCValAssign, 16> RVLocs;
2652 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2653 getTargetMachine(), RVLocs, *DAG.getContext());
2654 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2655 for (unsigned i = 0; i != RVLocs.size(); ++i)
2656 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2659 assert(((Callee.getOpcode() == ISD::Register &&
2660 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2661 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2662 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2663 isa<ConstantSDNode>(Callee)) &&
2664 "Expecting an global address, external symbol, absolute value or register");
2666 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2669 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2670 InFlag = Chain.getValue(1);
2672 // Add a NOP immediately after the branch instruction when using the 64-bit
2673 // SVR4 ABI. At link time, if caller and callee are in a different module and
2674 // thus have a different TOC, the call will be replaced with a call to a stub
2675 // function which saves the current TOC, loads the TOC of the callee and
2676 // branches to the callee. The NOP will be replaced with a load instruction
2677 // which restores the TOC of the caller from the TOC save slot of the current
2678 // stack frame. If caller and callee belong to the same module (and have the
2679 // same TOC), the NOP will remain unchanged.
2680 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2681 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
2682 if (CallOpc == PPCISD::BCTRL_SVR4) {
2683 // This is a call through a function pointer.
2684 // Restore the caller TOC from the save area into R2.
2685 // See PrepareCall() for more information about calls through function
2686 // pointers in the 64-bit SVR4 ABI.
2687 // We are using a target-specific load with r2 hard coded, because the
2688 // result of a target-independent load would never go directly into r2,
2689 // since r2 is a reserved register (which prevents the register allocator
2690 // from allocating it), resulting in an additional register being
2691 // allocated and an unnecessary move instruction being generated.
2692 Chain = DAG.getNode(PPCISD::TOC_RESTORE, dl, VTs, Chain, InFlag);
2693 InFlag = Chain.getValue(1);
2695 // Otherwise insert NOP.
2696 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Glue, InFlag);
2700 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2701 DAG.getIntPtrConstant(BytesCalleePops, true),
2704 InFlag = Chain.getValue(1);
2706 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2707 Ins, dl, DAG, InVals);
2711 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2712 CallingConv::ID CallConv, bool isVarArg,
2714 const SmallVectorImpl<ISD::OutputArg> &Outs,
2715 const SmallVectorImpl<SDValue> &OutVals,
2716 const SmallVectorImpl<ISD::InputArg> &Ins,
2717 DebugLoc dl, SelectionDAG &DAG,
2718 SmallVectorImpl<SDValue> &InVals) const {
2720 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
2723 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64())
2724 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2725 isTailCall, Outs, OutVals, Ins,
2728 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2729 isTailCall, Outs, OutVals, Ins,
2734 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2735 CallingConv::ID CallConv, bool isVarArg,
2737 const SmallVectorImpl<ISD::OutputArg> &Outs,
2738 const SmallVectorImpl<SDValue> &OutVals,
2739 const SmallVectorImpl<ISD::InputArg> &Ins,
2740 DebugLoc dl, SelectionDAG &DAG,
2741 SmallVectorImpl<SDValue> &InVals) const {
2742 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2743 // of the 32-bit SVR4 ABI stack frame layout.
2745 assert((CallConv == CallingConv::C ||
2746 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2748 unsigned PtrByteSize = 4;
2750 MachineFunction &MF = DAG.getMachineFunction();
2752 // Mark this function as potentially containing a function that contains a
2753 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2754 // and restoring the callers stack pointer in this functions epilog. This is
2755 // done because by tail calling the called function might overwrite the value
2756 // in this function's (MF) stack pointer stack slot 0(SP).
2757 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2758 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2760 // Count how many bytes are to be pushed on the stack, including the linkage
2761 // area, parameter list area and the part of the local variable space which
2762 // contains copies of aggregates which are passed by value.
2764 // Assign locations to all of the outgoing arguments.
2765 SmallVector<CCValAssign, 16> ArgLocs;
2766 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2767 getTargetMachine(), ArgLocs, *DAG.getContext());
2769 // Reserve space for the linkage area on the stack.
2770 CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false), PtrByteSize);
2773 // Handle fixed and variable vector arguments differently.
2774 // Fixed vector arguments go into registers as long as registers are
2775 // available. Variable vector arguments always go into memory.
2776 unsigned NumArgs = Outs.size();
2778 for (unsigned i = 0; i != NumArgs; ++i) {
2779 MVT ArgVT = Outs[i].VT;
2780 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2783 if (Outs[i].IsFixed) {
2784 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2787 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2793 errs() << "Call operand #" << i << " has unhandled type "
2794 << EVT(ArgVT).getEVTString() << "\n";
2796 llvm_unreachable(0);
2800 // All arguments are treated the same.
2801 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2804 // Assign locations to all of the outgoing aggregate by value arguments.
2805 SmallVector<CCValAssign, 16> ByValArgLocs;
2806 CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2807 getTargetMachine(), ByValArgLocs, *DAG.getContext());
2809 // Reserve stack space for the allocations in CCInfo.
2810 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2812 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2814 // Size of the linkage area, parameter list area and the part of the local
2815 // space variable where copies of aggregates which are passed by value are
2817 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2819 // Calculate by how many bytes the stack has to be adjusted in case of tail
2820 // call optimization.
2821 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2823 // Adjust the stack pointer for the new arguments...
2824 // These operations are automatically eliminated by the prolog/epilog pass
2825 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2826 SDValue CallSeqStart = Chain;
2828 // Load the return address and frame pointer so it can be moved somewhere else
2831 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2834 // Set up a copy of the stack pointer for use loading and storing any
2835 // arguments that may not fit in the registers available for argument
2837 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2839 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2840 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2841 SmallVector<SDValue, 8> MemOpChains;
2843 // Walk the register/memloc assignments, inserting copies/loads.
2844 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2847 CCValAssign &VA = ArgLocs[i];
2848 SDValue Arg = OutVals[i];
2849 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2851 if (Flags.isByVal()) {
2852 // Argument is an aggregate which is passed by value, thus we need to
2853 // create a copy of it in the local variable space of the current stack
2854 // frame (which is the stack frame of the caller) and pass the address of
2855 // this copy to the callee.
2856 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2857 CCValAssign &ByValVA = ByValArgLocs[j++];
2858 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2860 // Memory reserved in the local variable space of the callers stack frame.
2861 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2863 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2864 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2866 // Create a copy of the argument in the local area of the current
2868 SDValue MemcpyCall =
2869 CreateCopyOfByValArgument(Arg, PtrOff,
2870 CallSeqStart.getNode()->getOperand(0),
2873 // This must go outside the CALLSEQ_START..END.
2874 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2875 CallSeqStart.getNode()->getOperand(1));
2876 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2877 NewCallSeqStart.getNode());
2878 Chain = CallSeqStart = NewCallSeqStart;
2880 // Pass the address of the aggregate copy on the stack either in a
2881 // physical register or in the parameter list area of the current stack
2882 // frame to the callee.
2886 if (VA.isRegLoc()) {
2887 // Put argument in a physical register.
2888 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2890 // Put argument in the parameter list area of the current stack frame.
2891 assert(VA.isMemLoc());
2892 unsigned LocMemOffset = VA.getLocMemOffset();
2895 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2896 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2898 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2899 MachinePointerInfo(),
2902 // Calculate and remember argument location.
2903 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
2909 if (!MemOpChains.empty())
2910 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2911 &MemOpChains[0], MemOpChains.size());
2913 // Set CR6 to true if this is a vararg call.
2915 SDValue SetCR(DAG.getMachineNode(PPC::CRSET, dl, MVT::i32), 0);
2916 RegsToPass.push_back(std::make_pair(unsigned(PPC::CR1EQ), SetCR));
2919 // Build a sequence of copy-to-reg nodes chained together with token chain
2920 // and flag operands which copy the outgoing args into the appropriate regs.
2922 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2923 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2924 RegsToPass[i].second, InFlag);
2925 InFlag = Chain.getValue(1);
2929 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
2930 false, TailCallArguments);
2932 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
2933 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
2938 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
2939 CallingConv::ID CallConv, bool isVarArg,
2941 const SmallVectorImpl<ISD::OutputArg> &Outs,
2942 const SmallVectorImpl<SDValue> &OutVals,
2943 const SmallVectorImpl<ISD::InputArg> &Ins,
2944 DebugLoc dl, SelectionDAG &DAG,
2945 SmallVectorImpl<SDValue> &InVals) const {
2947 unsigned NumOps = Outs.size();
2949 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2950 bool isPPC64 = PtrVT == MVT::i64;
2951 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2953 MachineFunction &MF = DAG.getMachineFunction();
2955 // Mark this function as potentially containing a function that contains a
2956 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2957 // and restoring the callers stack pointer in this functions epilog. This is
2958 // done because by tail calling the called function might overwrite the value
2959 // in this function's (MF) stack pointer stack slot 0(SP).
2960 if (GuaranteedTailCallOpt && CallConv==CallingConv::Fast)
2961 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2963 unsigned nAltivecParamsAtEnd = 0;
2965 // Count how many bytes are to be pushed on the stack, including the linkage
2966 // area, and parameter passing area. We start with 24/48 bytes, which is
2967 // prereserved space for [SP][CR][LR][3 x unused].
2969 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
2971 nAltivecParamsAtEnd);
2973 // Calculate by how many bytes the stack has to be adjusted in case of tail
2974 // call optimization.
2975 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2977 // To protect arguments on the stack from being clobbered in a tail call,
2978 // force all the loads to happen before doing any other lowering.
2980 Chain = DAG.getStackArgumentTokenFactor(Chain);
2982 // Adjust the stack pointer for the new arguments...
2983 // These operations are automatically eliminated by the prolog/epilog pass
2984 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2985 SDValue CallSeqStart = Chain;
2987 // Load the return address and frame pointer so it can be move somewhere else
2990 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
2993 // Set up a copy of the stack pointer for use loading and storing any
2994 // arguments that may not fit in the registers available for argument
2998 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
3000 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
3002 // Figure out which arguments are going to go in registers, and which in
3003 // memory. Also, if this is a vararg function, floating point operations
3004 // must be stored to our stack, and loaded into integer regs as well, if
3005 // any integer regs are available for argument passing.
3006 unsigned ArgOffset = PPCFrameLowering::getLinkageSize(isPPC64, true);
3007 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
3009 static const unsigned GPR_32[] = { // 32-bit registers.
3010 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
3011 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
3013 static const unsigned GPR_64[] = { // 64-bit registers.
3014 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
3015 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
3017 static const unsigned *FPR = GetFPR();
3019 static const unsigned VR[] = {
3020 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
3021 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
3023 const unsigned NumGPRs = array_lengthof(GPR_32);
3024 const unsigned NumFPRs = 13;
3025 const unsigned NumVRs = array_lengthof(VR);
3027 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
3029 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3030 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
3032 SmallVector<SDValue, 8> MemOpChains;
3033 for (unsigned i = 0; i != NumOps; ++i) {
3034 SDValue Arg = OutVals[i];
3035 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3037 // PtrOff will be used to store the current argument to the stack if a
3038 // register cannot be found for it.
3041 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
3043 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3045 // On PPC64, promote integers to 64-bit values.
3046 if (isPPC64 && Arg.getValueType() == MVT::i32) {
3047 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
3048 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
3049 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
3052 // FIXME memcpy is used way more than necessary. Correctness first.
3053 if (Flags.isByVal()) {
3054 unsigned Size = Flags.getByValSize();
3055 if (Size==1 || Size==2) {
3056 // Very small objects are passed right-justified.
3057 // Everything else is passed left-justified.
3058 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
3059 if (GPR_idx != NumGPRs) {
3060 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
3061 MachinePointerInfo(), VT,
3063 MemOpChains.push_back(Load.getValue(1));
3064 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3066 ArgOffset += PtrByteSize;
3068 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
3069 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
3070 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
3071 CallSeqStart.getNode()->getOperand(0),
3073 // This must go outside the CALLSEQ_START..END.
3074 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3075 CallSeqStart.getNode()->getOperand(1));
3076 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
3077 NewCallSeqStart.getNode());
3078 Chain = CallSeqStart = NewCallSeqStart;
3079 ArgOffset += PtrByteSize;
3083 // Copy entire object into memory. There are cases where gcc-generated
3084 // code assumes it is there, even if it could be put entirely into
3085 // registers. (This is not what the doc says.)
3086 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
3087 CallSeqStart.getNode()->getOperand(0),
3089 // This must go outside the CALLSEQ_START..END.
3090 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
3091 CallSeqStart.getNode()->getOperand(1));
3092 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
3093 Chain = CallSeqStart = NewCallSeqStart;
3094 // And copy the pieces of it that fit into registers.
3095 for (unsigned j=0; j<Size; j+=PtrByteSize) {
3096 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
3097 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
3098 if (GPR_idx != NumGPRs) {
3099 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
3100 MachinePointerInfo(),
3102 MemOpChains.push_back(Load.getValue(1));
3103 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3104 ArgOffset += PtrByteSize;
3106 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
3113 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
3114 default: llvm_unreachable("Unexpected ValueType for argument!");
3117 if (GPR_idx != NumGPRs) {
3118 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
3120 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3121 isPPC64, isTailCall, false, MemOpChains,
3122 TailCallArguments, dl);
3124 ArgOffset += PtrByteSize;
3128 if (FPR_idx != NumFPRs) {
3129 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3132 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3133 MachinePointerInfo(), false, false, 0);
3134 MemOpChains.push_back(Store);
3136 // Float varargs are always shadowed in available integer registers
3137 if (GPR_idx != NumGPRs) {
3138 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3139 MachinePointerInfo(), false, false, 0);
3140 MemOpChains.push_back(Load.getValue(1));
3141 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3143 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3144 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3145 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3146 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
3147 MachinePointerInfo(),
3149 MemOpChains.push_back(Load.getValue(1));
3150 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3153 // If we have any FPRs remaining, we may also have GPRs remaining.
3154 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3156 if (GPR_idx != NumGPRs)
3158 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3159 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3163 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3164 isPPC64, isTailCall, false, MemOpChains,
3165 TailCallArguments, dl);
3170 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3177 // These go aligned on the stack, or in the corresponding R registers
3178 // when within range. The Darwin PPC ABI doc claims they also go in
3179 // V registers; in fact gcc does this only for arguments that are
3180 // prototyped, not for those that match the ... We do it for all
3181 // arguments, seems to work.
3182 while (ArgOffset % 16 !=0) {
3183 ArgOffset += PtrByteSize;
3184 if (GPR_idx != NumGPRs)
3187 // We could elide this store in the case where the object fits
3188 // entirely in R registers. Maybe later.
3189 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3190 DAG.getConstant(ArgOffset, PtrVT));
3191 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff,
3192 MachinePointerInfo(), false, false, 0);
3193 MemOpChains.push_back(Store);
3194 if (VR_idx != NumVRs) {
3195 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff,
3196 MachinePointerInfo(),
3198 MemOpChains.push_back(Load.getValue(1));
3199 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3202 for (unsigned i=0; i<16; i+=PtrByteSize) {
3203 if (GPR_idx == NumGPRs)
3205 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3206 DAG.getConstant(i, PtrVT));
3207 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo(),
3209 MemOpChains.push_back(Load.getValue(1));
3210 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3215 // Non-varargs Altivec params generally go in registers, but have
3216 // stack space allocated at the end.
3217 if (VR_idx != NumVRs) {
3218 // Doesn't have GPR space allocated.
3219 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3220 } else if (nAltivecParamsAtEnd==0) {
3221 // We are emitting Altivec params in order.
3222 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3223 isPPC64, isTailCall, true, MemOpChains,
3224 TailCallArguments, dl);
3230 // If all Altivec parameters fit in registers, as they usually do,
3231 // they get stack space following the non-Altivec parameters. We
3232 // don't track this here because nobody below needs it.
3233 // If there are more Altivec parameters than fit in registers emit
3235 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3237 // Offset is aligned; skip 1st 12 params which go in V registers.
3238 ArgOffset = ((ArgOffset+15)/16)*16;
3240 for (unsigned i = 0; i != NumOps; ++i) {
3241 SDValue Arg = OutVals[i];
3242 EVT ArgType = Outs[i].VT;
3243 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3244 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3247 // We are emitting Altivec params in order.
3248 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3249 isPPC64, isTailCall, true, MemOpChains,
3250 TailCallArguments, dl);
3257 if (!MemOpChains.empty())
3258 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3259 &MemOpChains[0], MemOpChains.size());
3261 // Check if this is an indirect call (MTCTR/BCTRL).
3262 // See PrepareCall() for more information about calls through function
3263 // pointers in the 64-bit SVR4 ABI.
3264 if (!isTailCall && isPPC64 && PPCSubTarget.isSVR4ABI() &&
3265 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3266 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3267 !isBLACompatibleAddress(Callee, DAG)) {
3268 // Load r2 into a virtual register and store it to the TOC save area.
3269 SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64);
3270 // TOC save area offset.
3271 SDValue PtrOff = DAG.getIntPtrConstant(40);
3272 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
3273 Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, MachinePointerInfo(),
3277 // On Darwin, R12 must contain the address of an indirect callee. This does
3278 // not mean the MTCTR instruction must use R12; it's easier to model this as
3279 // an extra parameter, so do that.
3281 !dyn_cast<GlobalAddressSDNode>(Callee) &&
3282 !dyn_cast<ExternalSymbolSDNode>(Callee) &&
3283 !isBLACompatibleAddress(Callee, DAG))
3284 RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 :
3285 PPC::R12), Callee));
3287 // Build a sequence of copy-to-reg nodes chained together with token chain
3288 // and flag operands which copy the outgoing args into the appropriate regs.
3290 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3291 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3292 RegsToPass[i].second, InFlag);
3293 InFlag = Chain.getValue(1);
3297 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3298 FPOp, true, TailCallArguments);
3300 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3301 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3306 PPCTargetLowering::LowerReturn(SDValue Chain,
3307 CallingConv::ID CallConv, bool isVarArg,
3308 const SmallVectorImpl<ISD::OutputArg> &Outs,
3309 const SmallVectorImpl<SDValue> &OutVals,
3310 DebugLoc dl, SelectionDAG &DAG) const {
3312 SmallVector<CCValAssign, 16> RVLocs;
3313 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
3314 getTargetMachine(), RVLocs, *DAG.getContext());
3315 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3317 // If this is the first return lowered for this function, add the regs to the
3318 // liveout set for the function.
3319 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3320 for (unsigned i = 0; i != RVLocs.size(); ++i)
3321 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3326 // Copy the result values into the output registers.
3327 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3328 CCValAssign &VA = RVLocs[i];
3329 assert(VA.isRegLoc() && "Can only return in registers!");
3330 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3332 Flag = Chain.getValue(1);
3336 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3338 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3341 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3342 const PPCSubtarget &Subtarget) const {
3343 // When we pop the dynamic allocation we need to restore the SP link.
3344 DebugLoc dl = Op.getDebugLoc();
3346 // Get the corect type for pointers.
3347 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3349 // Construct the stack pointer operand.
3350 bool isPPC64 = Subtarget.isPPC64();
3351 unsigned SP = isPPC64 ? PPC::X1 : PPC::R1;
3352 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3354 // Get the operands for the STACKRESTORE.
3355 SDValue Chain = Op.getOperand(0);
3356 SDValue SaveSP = Op.getOperand(1);
3358 // Load the old link SP.
3359 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr,
3360 MachinePointerInfo(),
3363 // Restore the stack pointer.
3364 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3366 // Store the old link SP.
3367 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo(),
3374 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3375 MachineFunction &MF = DAG.getMachineFunction();
3376 bool isPPC64 = PPCSubTarget.isPPC64();
3377 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3378 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3380 // Get current frame pointer save index. The users of this index will be
3381 // primarily DYNALLOC instructions.
3382 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3383 int RASI = FI->getReturnAddrSaveIndex();
3385 // If the frame pointer save index hasn't been defined yet.
3387 // Find out what the fix offset of the frame pointer save area.
3388 int LROffset = PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI);
3389 // Allocate the frame index for frame pointer save area.
3390 RASI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, LROffset, true);
3392 FI->setReturnAddrSaveIndex(RASI);
3394 return DAG.getFrameIndex(RASI, PtrVT);
3398 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3399 MachineFunction &MF = DAG.getMachineFunction();
3400 bool isPPC64 = PPCSubTarget.isPPC64();
3401 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3402 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3404 // Get current frame pointer save index. The users of this index will be
3405 // primarily DYNALLOC instructions.
3406 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3407 int FPSI = FI->getFramePointerSaveIndex();
3409 // If the frame pointer save index hasn't been defined yet.
3411 // Find out what the fix offset of the frame pointer save area.
3412 int FPOffset = PPCFrameLowering::getFramePointerSaveOffset(isPPC64,
3415 // Allocate the frame index for frame pointer save area.
3416 FPSI = MF.getFrameInfo()->CreateFixedObject(isPPC64? 8 : 4, FPOffset, true);
3418 FI->setFramePointerSaveIndex(FPSI);
3420 return DAG.getFrameIndex(FPSI, PtrVT);
3423 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3425 const PPCSubtarget &Subtarget) const {
3427 SDValue Chain = Op.getOperand(0);
3428 SDValue Size = Op.getOperand(1);
3429 DebugLoc dl = Op.getDebugLoc();
3431 // Get the corect type for pointers.
3432 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3434 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3435 DAG.getConstant(0, PtrVT), Size);
3436 // Construct a node for the frame pointer save index.
3437 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3438 // Build a DYNALLOC node.
3439 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3440 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3441 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3444 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3446 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3447 // Not FP? Not a fsel.
3448 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3449 !Op.getOperand(2).getValueType().isFloatingPoint())
3452 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3454 // Cannot handle SETEQ/SETNE.
3455 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3457 EVT ResVT = Op.getValueType();
3458 EVT CmpVT = Op.getOperand(0).getValueType();
3459 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3460 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3461 DebugLoc dl = Op.getDebugLoc();
3463 // If the RHS of the comparison is a 0.0, we don't need to do the
3464 // subtraction at all.
3465 if (isFloatingPointZero(RHS))
3467 default: break; // SETUO etc aren't handled by fsel.
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, LHS, TV, FV);
3478 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3481 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3482 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3483 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3484 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3489 default: break; // SETUO etc aren't handled by fsel.
3492 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3493 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3494 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3495 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3498 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3499 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3500 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3501 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3504 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3505 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3506 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3507 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3510 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3511 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3512 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3513 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3518 // FIXME: Split this code up when LegalizeDAGTypes lands.
3519 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3520 DebugLoc dl) const {
3521 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3522 SDValue Src = Op.getOperand(0);
3523 if (Src.getValueType() == MVT::f32)
3524 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3527 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3528 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3530 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3535 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3539 // Convert the FP value to an int value through memory.
3540 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3542 // Emit a store to the stack slot.
3543 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr,
3544 MachinePointerInfo(), false, false, 0);
3546 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3548 if (Op.getValueType() == MVT::i32)
3549 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3550 DAG.getConstant(4, FIPtr.getValueType()));
3551 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, MachinePointerInfo(),
3555 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op,
3556 SelectionDAG &DAG) const {
3557 DebugLoc dl = Op.getDebugLoc();
3558 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3559 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3562 if (Op.getOperand(0).getValueType() == MVT::i64) {
3563 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op.getOperand(0));
3564 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3565 if (Op.getValueType() == MVT::f32)
3566 FP = DAG.getNode(ISD::FP_ROUND, dl,
3567 MVT::f32, FP, DAG.getIntPtrConstant(0));
3571 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3572 "Unhandled SINT_TO_FP type in custom expander!");
3573 // Since we only generate this in 64-bit mode, we can take advantage of
3574 // 64-bit registers. In particular, sign extend the input value into the
3575 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3576 // then lfd it and fcfid it.
3577 MachineFunction &MF = DAG.getMachineFunction();
3578 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3579 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3580 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3581 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3583 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3586 // STD the extended value into the stack slot.
3587 MachineMemOperand *MMO =
3588 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FrameIdx),
3589 MachineMemOperand::MOStore, 8, 8);
3590 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3592 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3593 Ops, 4, MVT::i64, MMO);
3594 // Load the value as a double.
3595 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, MachinePointerInfo(),
3598 // FCFID it and return it.
3599 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3600 if (Op.getValueType() == MVT::f32)
3601 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3605 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3606 SelectionDAG &DAG) const {
3607 DebugLoc dl = Op.getDebugLoc();
3609 The rounding mode is in bits 30:31 of FPSR, and has the following
3616 FLT_ROUNDS, on the other hand, expects the following:
3623 To perform the conversion, we do:
3624 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3627 MachineFunction &MF = DAG.getMachineFunction();
3628 EVT VT = Op.getValueType();
3629 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3630 std::vector<EVT> NodeTys;
3631 SDValue MFFSreg, InFlag;
3633 // Save FP Control Word to register
3634 NodeTys.push_back(MVT::f64); // return register
3635 NodeTys.push_back(MVT::Glue); // unused in this context
3636 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3638 // Save FP register to stack slot
3639 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3640 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3641 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3642 StackSlot, MachinePointerInfo(), false, false,0);
3644 // Load FP Control Word from low 32 bits of stack slot.
3645 SDValue Four = DAG.getConstant(4, PtrVT);
3646 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3647 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, MachinePointerInfo(),
3650 // Transform as necessary
3652 DAG.getNode(ISD::AND, dl, MVT::i32,
3653 CWD, DAG.getConstant(3, MVT::i32));
3655 DAG.getNode(ISD::SRL, dl, MVT::i32,
3656 DAG.getNode(ISD::AND, dl, MVT::i32,
3657 DAG.getNode(ISD::XOR, dl, MVT::i32,
3658 CWD, DAG.getConstant(3, MVT::i32)),
3659 DAG.getConstant(3, MVT::i32)),
3660 DAG.getConstant(1, MVT::i32));
3663 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3665 return DAG.getNode((VT.getSizeInBits() < 16 ?
3666 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3669 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3670 EVT VT = Op.getValueType();
3671 unsigned BitWidth = VT.getSizeInBits();
3672 DebugLoc dl = Op.getDebugLoc();
3673 assert(Op.getNumOperands() == 3 &&
3674 VT == Op.getOperand(1).getValueType() &&
3677 // Expand into a bunch of logical ops. Note that these ops
3678 // depend on the PPC behavior for oversized shift amounts.
3679 SDValue Lo = Op.getOperand(0);
3680 SDValue Hi = Op.getOperand(1);
3681 SDValue Amt = Op.getOperand(2);
3682 EVT AmtVT = Amt.getValueType();
3684 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3685 DAG.getConstant(BitWidth, AmtVT), Amt);
3686 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3687 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3688 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3689 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3690 DAG.getConstant(-BitWidth, AmtVT));
3691 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3692 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3693 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3694 SDValue OutOps[] = { OutLo, OutHi };
3695 return DAG.getMergeValues(OutOps, 2, dl);
3698 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const {
3699 EVT VT = Op.getValueType();
3700 DebugLoc dl = Op.getDebugLoc();
3701 unsigned BitWidth = VT.getSizeInBits();
3702 assert(Op.getNumOperands() == 3 &&
3703 VT == Op.getOperand(1).getValueType() &&
3706 // Expand into a bunch of logical ops. Note that these ops
3707 // depend on the PPC behavior for oversized shift amounts.
3708 SDValue Lo = Op.getOperand(0);
3709 SDValue Hi = Op.getOperand(1);
3710 SDValue Amt = Op.getOperand(2);
3711 EVT AmtVT = Amt.getValueType();
3713 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3714 DAG.getConstant(BitWidth, AmtVT), Amt);
3715 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3716 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3717 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3718 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3719 DAG.getConstant(-BitWidth, AmtVT));
3720 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3721 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3722 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3723 SDValue OutOps[] = { OutLo, OutHi };
3724 return DAG.getMergeValues(OutOps, 2, dl);
3727 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const {
3728 DebugLoc dl = Op.getDebugLoc();
3729 EVT VT = Op.getValueType();
3730 unsigned BitWidth = VT.getSizeInBits();
3731 assert(Op.getNumOperands() == 3 &&
3732 VT == Op.getOperand(1).getValueType() &&
3735 // Expand into a bunch of logical ops, followed by a select_cc.
3736 SDValue Lo = Op.getOperand(0);
3737 SDValue Hi = Op.getOperand(1);
3738 SDValue Amt = Op.getOperand(2);
3739 EVT AmtVT = Amt.getValueType();
3741 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3742 DAG.getConstant(BitWidth, AmtVT), Amt);
3743 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3744 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3745 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3746 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3747 DAG.getConstant(-BitWidth, AmtVT));
3748 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3749 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3750 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3751 Tmp4, Tmp6, ISD::SETLE);
3752 SDValue OutOps[] = { OutLo, OutHi };
3753 return DAG.getMergeValues(OutOps, 2, dl);
3756 //===----------------------------------------------------------------------===//
3757 // Vector related lowering.
3760 /// BuildSplatI - Build a canonical splati of Val with an element size of
3761 /// SplatSize. Cast the result to VT.
3762 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3763 SelectionDAG &DAG, DebugLoc dl) {
3764 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3766 static const EVT VTys[] = { // canonical VT to use for each size.
3767 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3770 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3772 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3776 EVT CanonicalVT = VTys[SplatSize-1];
3778 // Build a canonical splat for this value.
3779 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3780 SmallVector<SDValue, 8> Ops;
3781 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3782 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3783 &Ops[0], Ops.size());
3784 return DAG.getNode(ISD::BITCAST, dl, ReqVT, Res);
3787 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3788 /// specified intrinsic ID.
3789 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3790 SelectionDAG &DAG, DebugLoc dl,
3791 EVT DestVT = MVT::Other) {
3792 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3793 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3794 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3797 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3798 /// specified intrinsic ID.
3799 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3800 SDValue Op2, SelectionDAG &DAG,
3801 DebugLoc dl, EVT DestVT = MVT::Other) {
3802 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3803 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3804 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3808 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3809 /// amount. The result has the specified value type.
3810 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3811 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3812 // Force LHS/RHS to be the right type.
3813 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS);
3814 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS);
3817 for (unsigned i = 0; i != 16; ++i)
3819 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3820 return DAG.getNode(ISD::BITCAST, dl, VT, T);
3823 // If this is a case we can't handle, return null and let the default
3824 // expansion code take care of it. If we CAN select this case, and if it
3825 // selects to a single instruction, return Op. Otherwise, if we can codegen
3826 // this case more efficiently than a constant pool load, lower it to the
3827 // sequence of ops that should be used.
3828 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3829 SelectionDAG &DAG) const {
3830 DebugLoc dl = Op.getDebugLoc();
3831 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3832 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3834 // Check if this is a splat of a constant value.
3835 APInt APSplatBits, APSplatUndef;
3836 unsigned SplatBitSize;
3838 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3839 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3842 unsigned SplatBits = APSplatBits.getZExtValue();
3843 unsigned SplatUndef = APSplatUndef.getZExtValue();
3844 unsigned SplatSize = SplatBitSize / 8;
3846 // First, handle single instruction cases.
3849 if (SplatBits == 0) {
3850 // Canonicalize all zero vectors to be v4i32.
3851 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3852 SDValue Z = DAG.getConstant(0, MVT::i32);
3853 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3854 Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z);
3859 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3860 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3862 if (SextVal >= -16 && SextVal <= 15)
3863 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3866 // Two instruction sequences.
3868 // If this value is in the range [-32,30] and is even, use:
3869 // tmp = VSPLTI[bhw], result = add tmp, tmp
3870 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3871 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3872 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3873 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3876 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3877 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3879 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3880 // Make -1 and vspltisw -1:
3881 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3883 // Make the VSLW intrinsic, computing 0x8000_0000.
3884 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3887 // xor by OnesV to invert it.
3888 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3889 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3892 // Check to see if this is a wide variety of vsplti*, binop self cases.
3893 static const signed char SplatCsts[] = {
3894 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3895 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3898 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3899 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3900 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3901 int i = SplatCsts[idx];
3903 // Figure out what shift amount will be used by altivec if shifted by i in
3905 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3907 // vsplti + shl self.
3908 if (SextVal == (i << (int)TypeShiftAmt)) {
3909 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3910 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3911 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3912 Intrinsic::ppc_altivec_vslw
3914 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3915 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3918 // vsplti + srl self.
3919 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3920 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3921 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3922 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3923 Intrinsic::ppc_altivec_vsrw
3925 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3926 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3929 // vsplti + sra self.
3930 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3931 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3932 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3933 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3934 Intrinsic::ppc_altivec_vsraw
3936 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3937 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3940 // vsplti + rol self.
3941 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3942 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3943 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3944 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3945 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3946 Intrinsic::ppc_altivec_vrlw
3948 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3949 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
3952 // t = vsplti c, result = vsldoi t, t, 1
3953 if (SextVal == ((i << 8) | (i < 0 ? 0xFF : 0))) {
3954 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3955 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
3957 // t = vsplti c, result = vsldoi t, t, 2
3958 if (SextVal == ((i << 16) | (i < 0 ? 0xFFFF : 0))) {
3959 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3960 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
3962 // t = vsplti c, result = vsldoi t, t, 3
3963 if (SextVal == ((i << 24) | (i < 0 ? 0xFFFFFF : 0))) {
3964 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3965 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
3969 // Three instruction sequences.
3971 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3972 if (SextVal >= 0 && SextVal <= 31) {
3973 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
3974 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3975 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
3976 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
3978 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3979 if (SextVal >= -31 && SextVal <= 0) {
3980 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
3981 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3982 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
3983 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), LHS);
3989 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3990 /// the specified operations to build the shuffle.
3991 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3992 SDValue RHS, SelectionDAG &DAG,
3994 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3995 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3996 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3999 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4011 if (OpNum == OP_COPY) {
4012 if (LHSID == (1*9+2)*9+3) return LHS;
4013 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4017 SDValue OpLHS, OpRHS;
4018 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4019 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4023 default: llvm_unreachable("Unknown i32 permute!");
4025 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
4026 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
4027 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
4028 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
4031 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
4032 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
4033 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
4034 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4037 for (unsigned i = 0; i != 16; ++i)
4038 ShufIdxs[i] = (i&3)+0;
4041 for (unsigned i = 0; i != 16; ++i)
4042 ShufIdxs[i] = (i&3)+4;
4045 for (unsigned i = 0; i != 16; ++i)
4046 ShufIdxs[i] = (i&3)+8;
4049 for (unsigned i = 0; i != 16; ++i)
4050 ShufIdxs[i] = (i&3)+12;
4053 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4055 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4057 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4059 EVT VT = OpLHS.getValueType();
4060 OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS);
4061 OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS);
4062 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4063 return DAG.getNode(ISD::BITCAST, dl, VT, T);
4066 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4067 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4068 /// return the code it can be lowered into. Worst case, it can always be
4069 /// lowered into a vperm.
4070 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4071 SelectionDAG &DAG) const {
4072 DebugLoc dl = Op.getDebugLoc();
4073 SDValue V1 = Op.getOperand(0);
4074 SDValue V2 = Op.getOperand(1);
4075 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4076 EVT VT = Op.getValueType();
4078 // Cases that are handled by instructions that take permute immediates
4079 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4080 // selected by the instruction selector.
4081 if (V2.getOpcode() == ISD::UNDEF) {
4082 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4083 PPC::isSplatShuffleMask(SVOp, 2) ||
4084 PPC::isSplatShuffleMask(SVOp, 4) ||
4085 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4086 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4087 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4088 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4089 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4090 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4091 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4092 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4093 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4098 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4099 // and produce a fixed permutation. If any of these match, do not lower to
4101 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4102 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4103 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4104 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4105 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4106 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4107 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4108 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4109 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4112 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4113 // perfect shuffle table to emit an optimal matching sequence.
4114 SmallVector<int, 16> PermMask;
4115 SVOp->getMask(PermMask);
4117 unsigned PFIndexes[4];
4118 bool isFourElementShuffle = true;
4119 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4120 unsigned EltNo = 8; // Start out undef.
4121 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4122 if (PermMask[i*4+j] < 0)
4123 continue; // Undef, ignore it.
4125 unsigned ByteSource = PermMask[i*4+j];
4126 if ((ByteSource & 3) != j) {
4127 isFourElementShuffle = false;
4132 EltNo = ByteSource/4;
4133 } else if (EltNo != ByteSource/4) {
4134 isFourElementShuffle = false;
4138 PFIndexes[i] = EltNo;
4141 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4142 // perfect shuffle vector to determine if it is cost effective to do this as
4143 // discrete instructions, or whether we should use a vperm.
4144 if (isFourElementShuffle) {
4145 // Compute the index in the perfect shuffle table.
4146 unsigned PFTableIndex =
4147 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4149 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4150 unsigned Cost = (PFEntry >> 30);
4152 // Determining when to avoid vperm is tricky. Many things affect the cost
4153 // of vperm, particularly how many times the perm mask needs to be computed.
4154 // For example, if the perm mask can be hoisted out of a loop or is already
4155 // used (perhaps because there are multiple permutes with the same shuffle
4156 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4157 // the loop requires an extra register.
4159 // As a compromise, we only emit discrete instructions if the shuffle can be
4160 // generated in 3 or fewer operations. When we have loop information
4161 // available, if this block is within a loop, we should avoid using vperm
4162 // for 3-operation perms and use a constant pool load instead.
4164 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4167 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4168 // vector that will get spilled to the constant pool.
4169 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4171 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4172 // that it is in input element units, not in bytes. Convert now.
4173 EVT EltVT = V1.getValueType().getVectorElementType();
4174 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4176 SmallVector<SDValue, 16> ResultMask;
4177 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4178 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4180 for (unsigned j = 0; j != BytesPerElement; ++j)
4181 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4185 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4186 &ResultMask[0], ResultMask.size());
4187 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4190 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4191 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4192 /// information about the intrinsic.
4193 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4195 unsigned IntrinsicID =
4196 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4199 switch (IntrinsicID) {
4200 default: return false;
4201 // Comparison predicates.
4202 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4203 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4204 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4205 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4206 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4207 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4208 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4209 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4210 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4211 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4212 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4213 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4214 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4216 // Normal Comparisons.
4217 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4218 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4219 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4220 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4221 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4222 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4223 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4224 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4225 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4226 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4227 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4228 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4229 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4234 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4235 /// lower, do it, otherwise return null.
4236 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4237 SelectionDAG &DAG) const {
4238 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4239 // opcode number of the comparison.
4240 DebugLoc dl = Op.getDebugLoc();
4243 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4244 return SDValue(); // Don't custom lower most intrinsics.
4246 // If this is a non-dot comparison, make the VCMP node and we are done.
4248 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4249 Op.getOperand(1), Op.getOperand(2),
4250 DAG.getConstant(CompareOpc, MVT::i32));
4251 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp);
4254 // Create the PPCISD altivec 'dot' comparison node.
4256 Op.getOperand(2), // LHS
4257 Op.getOperand(3), // RHS
4258 DAG.getConstant(CompareOpc, MVT::i32)
4260 std::vector<EVT> VTs;
4261 VTs.push_back(Op.getOperand(2).getValueType());
4262 VTs.push_back(MVT::Glue);
4263 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4265 // Now that we have the comparison, emit a copy from the CR to a GPR.
4266 // This is flagged to the above dot comparison.
4267 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4268 DAG.getRegister(PPC::CR6, MVT::i32),
4269 CompNode.getValue(1));
4271 // Unpack the result based on how the target uses it.
4272 unsigned BitNo; // Bit # of CR6.
4273 bool InvertBit; // Invert result?
4274 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4275 default: // Can't happen, don't crash on invalid number though.
4276 case 0: // Return the value of the EQ bit of CR6.
4277 BitNo = 0; InvertBit = false;
4279 case 1: // Return the inverted value of the EQ bit of CR6.
4280 BitNo = 0; InvertBit = true;
4282 case 2: // Return the value of the LT bit of CR6.
4283 BitNo = 2; InvertBit = false;
4285 case 3: // Return the inverted value of the LT bit of CR6.
4286 BitNo = 2; InvertBit = true;
4290 // Shift the bit into the low position.
4291 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4292 DAG.getConstant(8-(3-BitNo), MVT::i32));
4294 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4295 DAG.getConstant(1, MVT::i32));
4297 // If we are supposed to, toggle the bit.
4299 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4300 DAG.getConstant(1, MVT::i32));
4304 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4305 SelectionDAG &DAG) const {
4306 DebugLoc dl = Op.getDebugLoc();
4307 // Create a stack slot that is 16-byte aligned.
4308 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4309 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4310 EVT PtrVT = getPointerTy();
4311 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4313 // Store the input value into Value#0 of the stack slot.
4314 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4315 Op.getOperand(0), FIdx, MachinePointerInfo(),
4318 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo(),
4322 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
4323 DebugLoc dl = Op.getDebugLoc();
4324 if (Op.getValueType() == MVT::v4i32) {
4325 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4327 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4328 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4330 SDValue RHSSwap = // = vrlw RHS, 16
4331 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4333 // Shrinkify inputs to v8i16.
4334 LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS);
4335 RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS);
4336 RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap);
4338 // Low parts multiplied together, generating 32-bit results (we ignore the
4340 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4341 LHS, RHS, DAG, dl, MVT::v4i32);
4343 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4344 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4345 // Shift the high parts up 16 bits.
4346 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4348 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4349 } else if (Op.getValueType() == MVT::v8i16) {
4350 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4352 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4354 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4355 LHS, RHS, Zero, DAG, dl);
4356 } else if (Op.getValueType() == MVT::v16i8) {
4357 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4359 // Multiply the even 8-bit parts, producing 16-bit sums.
4360 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4361 LHS, RHS, DAG, dl, MVT::v8i16);
4362 EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts);
4364 // Multiply the odd 8-bit parts, producing 16-bit sums.
4365 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4366 LHS, RHS, DAG, dl, MVT::v8i16);
4367 OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts);
4369 // Merge the results together.
4371 for (unsigned i = 0; i != 8; ++i) {
4373 Ops[i*2+1] = 2*i+1+16;
4375 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4377 llvm_unreachable("Unknown mul to lower!");
4381 /// LowerOperation - Provide custom lowering hooks for some operations.
4383 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4384 switch (Op.getOpcode()) {
4385 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4386 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4387 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4388 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4389 case ISD::GlobalTLSAddress: llvm_unreachable("TLS not implemented for PPC");
4390 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4391 case ISD::SETCC: return LowerSETCC(Op, DAG);
4392 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4394 return LowerVASTART(Op, DAG, PPCSubTarget);
4397 return LowerVAARG(Op, DAG, PPCSubTarget);
4399 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4400 case ISD::DYNAMIC_STACKALLOC:
4401 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4403 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4404 case ISD::FP_TO_UINT:
4405 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4407 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4408 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4410 // Lower 64-bit shifts.
4411 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4412 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4413 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4415 // Vector-related lowering.
4416 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4417 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4418 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4419 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4420 case ISD::MUL: return LowerMUL(Op, DAG);
4422 // Frame & Return address.
4423 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4424 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4429 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4430 SmallVectorImpl<SDValue>&Results,
4431 SelectionDAG &DAG) const {
4432 DebugLoc dl = N->getDebugLoc();
4433 switch (N->getOpcode()) {
4435 assert(false && "Do not know how to custom type legalize this operation!");
4437 case ISD::FP_ROUND_INREG: {
4438 assert(N->getValueType(0) == MVT::ppcf128);
4439 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4440 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4441 MVT::f64, N->getOperand(0),
4442 DAG.getIntPtrConstant(0));
4443 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4444 MVT::f64, N->getOperand(0),
4445 DAG.getIntPtrConstant(1));
4447 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4448 // of the long double, and puts FPSCR back the way it was. We do not
4449 // actually model FPSCR.
4450 std::vector<EVT> NodeTys;
4451 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4453 NodeTys.push_back(MVT::f64); // Return register
4454 NodeTys.push_back(MVT::Glue); // Returns a flag for later insns
4455 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4456 MFFSreg = Result.getValue(0);
4457 InFlag = Result.getValue(1);
4460 NodeTys.push_back(MVT::Glue); // Returns a flag
4461 Ops[0] = DAG.getConstant(31, MVT::i32);
4463 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4464 InFlag = Result.getValue(0);
4467 NodeTys.push_back(MVT::Glue); // Returns a flag
4468 Ops[0] = DAG.getConstant(30, MVT::i32);
4470 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4471 InFlag = Result.getValue(0);
4474 NodeTys.push_back(MVT::f64); // result of add
4475 NodeTys.push_back(MVT::Glue); // Returns a flag
4479 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4480 FPreg = Result.getValue(0);
4481 InFlag = Result.getValue(1);
4484 NodeTys.push_back(MVT::f64);
4485 Ops[0] = DAG.getConstant(1, MVT::i32);
4489 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4490 FPreg = Result.getValue(0);
4492 // We know the low half is about to be thrown away, so just use something
4494 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4498 case ISD::FP_TO_SINT:
4499 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4505 //===----------------------------------------------------------------------===//
4506 // Other Lowering Code
4507 //===----------------------------------------------------------------------===//
4510 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4511 bool is64bit, unsigned BinOpcode) const {
4512 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4513 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4515 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4516 MachineFunction *F = BB->getParent();
4517 MachineFunction::iterator It = BB;
4520 unsigned dest = MI->getOperand(0).getReg();
4521 unsigned ptrA = MI->getOperand(1).getReg();
4522 unsigned ptrB = MI->getOperand(2).getReg();
4523 unsigned incr = MI->getOperand(3).getReg();
4524 DebugLoc dl = MI->getDebugLoc();
4526 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4527 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4528 F->insert(It, loopMBB);
4529 F->insert(It, exitMBB);
4530 exitMBB->splice(exitMBB->begin(), BB,
4531 llvm::next(MachineBasicBlock::iterator(MI)),
4533 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4535 MachineRegisterInfo &RegInfo = F->getRegInfo();
4536 unsigned TmpReg = (!BinOpcode) ? incr :
4537 RegInfo.createVirtualRegister(
4538 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4539 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4543 // fallthrough --> loopMBB
4544 BB->addSuccessor(loopMBB);
4547 // l[wd]arx dest, ptr
4548 // add r0, dest, incr
4549 // st[wd]cx. r0, ptr
4551 // fallthrough --> exitMBB
4553 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4554 .addReg(ptrA).addReg(ptrB);
4556 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4557 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4558 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4559 BuildMI(BB, dl, TII->get(PPC::BCC))
4560 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4561 BB->addSuccessor(loopMBB);
4562 BB->addSuccessor(exitMBB);
4571 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4572 MachineBasicBlock *BB,
4573 bool is8bit, // operation
4574 unsigned BinOpcode) const {
4575 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4576 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4577 // In 64 bit mode we have to use 64 bits for addresses, even though the
4578 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4579 // registers without caring whether they're 32 or 64, but here we're
4580 // doing actual arithmetic on the addresses.
4581 bool is64bit = PPCSubTarget.isPPC64();
4582 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4584 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4585 MachineFunction *F = BB->getParent();
4586 MachineFunction::iterator It = BB;
4589 unsigned dest = MI->getOperand(0).getReg();
4590 unsigned ptrA = MI->getOperand(1).getReg();
4591 unsigned ptrB = MI->getOperand(2).getReg();
4592 unsigned incr = MI->getOperand(3).getReg();
4593 DebugLoc dl = MI->getDebugLoc();
4595 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4596 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4597 F->insert(It, loopMBB);
4598 F->insert(It, exitMBB);
4599 exitMBB->splice(exitMBB->begin(), BB,
4600 llvm::next(MachineBasicBlock::iterator(MI)),
4602 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4604 MachineRegisterInfo &RegInfo = F->getRegInfo();
4605 const TargetRegisterClass *RC =
4606 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4607 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4608 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4609 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4610 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4611 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4612 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4613 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4614 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4615 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4616 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4617 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4618 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4620 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4624 // fallthrough --> loopMBB
4625 BB->addSuccessor(loopMBB);
4627 // The 4-byte load must be aligned, while a char or short may be
4628 // anywhere in the word. Hence all this nasty bookkeeping code.
4629 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4630 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4631 // xori shift, shift1, 24 [16]
4632 // rlwinm ptr, ptr1, 0, 0, 29
4633 // slw incr2, incr, shift
4634 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4635 // slw mask, mask2, shift
4637 // lwarx tmpDest, ptr
4638 // add tmp, tmpDest, incr2
4639 // andc tmp2, tmpDest, mask
4640 // and tmp3, tmp, mask
4641 // or tmp4, tmp3, tmp2
4644 // fallthrough --> exitMBB
4645 // srw dest, tmpDest, shift
4646 if (ptrA != ZeroReg) {
4647 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4648 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4649 .addReg(ptrA).addReg(ptrB);
4653 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4654 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4655 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4656 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4658 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4659 .addReg(Ptr1Reg).addImm(0).addImm(61);
4661 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4662 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4663 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4664 .addReg(incr).addReg(ShiftReg);
4666 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4668 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4669 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4671 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4672 .addReg(Mask2Reg).addReg(ShiftReg);
4675 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4676 .addReg(ZeroReg).addReg(PtrReg);
4678 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4679 .addReg(Incr2Reg).addReg(TmpDestReg);
4680 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4681 .addReg(TmpDestReg).addReg(MaskReg);
4682 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4683 .addReg(TmpReg).addReg(MaskReg);
4684 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4685 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4686 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4687 .addReg(Tmp4Reg).addReg(ZeroReg).addReg(PtrReg);
4688 BuildMI(BB, dl, TII->get(PPC::BCC))
4689 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4690 BB->addSuccessor(loopMBB);
4691 BB->addSuccessor(exitMBB);
4696 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg)
4702 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4703 MachineBasicBlock *BB) const {
4704 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4706 // To "insert" these instructions we actually have to insert their
4707 // control-flow patterns.
4708 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4709 MachineFunction::iterator It = BB;
4712 MachineFunction *F = BB->getParent();
4714 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4715 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4716 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4717 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4718 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4720 // The incoming instruction knows the destination vreg to set, the
4721 // condition code register to branch on, the true/false values to
4722 // select between, and a branch opcode to use.
4727 // cmpTY ccX, r1, r2
4729 // fallthrough --> copy0MBB
4730 MachineBasicBlock *thisMBB = BB;
4731 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4732 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4733 unsigned SelectPred = MI->getOperand(4).getImm();
4734 DebugLoc dl = MI->getDebugLoc();
4735 F->insert(It, copy0MBB);
4736 F->insert(It, sinkMBB);
4738 // Transfer the remainder of BB and its successor edges to sinkMBB.
4739 sinkMBB->splice(sinkMBB->begin(), BB,
4740 llvm::next(MachineBasicBlock::iterator(MI)),
4742 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
4744 // Next, add the true and fallthrough blocks as its successors.
4745 BB->addSuccessor(copy0MBB);
4746 BB->addSuccessor(sinkMBB);
4748 BuildMI(BB, dl, TII->get(PPC::BCC))
4749 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4752 // %FalseValue = ...
4753 // # fallthrough to sinkMBB
4756 // Update machine-CFG edges
4757 BB->addSuccessor(sinkMBB);
4760 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4763 BuildMI(*BB, BB->begin(), dl,
4764 TII->get(PPC::PHI), MI->getOperand(0).getReg())
4765 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4766 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4768 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4769 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4770 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4771 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4772 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4773 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4774 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4775 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4777 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4778 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4779 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4780 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4781 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4782 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4783 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4784 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4786 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4787 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4788 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4789 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4790 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4791 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4792 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4793 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4795 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4796 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4797 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4798 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4799 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4800 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4801 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4802 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4804 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4805 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4806 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4807 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4808 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4809 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4810 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4811 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4813 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4814 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4815 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4816 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4817 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4818 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4819 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4820 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4822 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4823 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4824 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4825 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4826 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4827 BB = EmitAtomicBinary(MI, BB, false, 0);
4828 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4829 BB = EmitAtomicBinary(MI, BB, true, 0);
4831 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4832 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4833 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4835 unsigned dest = MI->getOperand(0).getReg();
4836 unsigned ptrA = MI->getOperand(1).getReg();
4837 unsigned ptrB = MI->getOperand(2).getReg();
4838 unsigned oldval = MI->getOperand(3).getReg();
4839 unsigned newval = MI->getOperand(4).getReg();
4840 DebugLoc dl = MI->getDebugLoc();
4842 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4843 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4844 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4845 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4846 F->insert(It, loop1MBB);
4847 F->insert(It, loop2MBB);
4848 F->insert(It, midMBB);
4849 F->insert(It, exitMBB);
4850 exitMBB->splice(exitMBB->begin(), BB,
4851 llvm::next(MachineBasicBlock::iterator(MI)),
4853 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4857 // fallthrough --> loopMBB
4858 BB->addSuccessor(loop1MBB);
4861 // l[wd]arx dest, ptr
4862 // cmp[wd] dest, oldval
4865 // st[wd]cx. newval, ptr
4869 // st[wd]cx. dest, ptr
4872 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4873 .addReg(ptrA).addReg(ptrB);
4874 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4875 .addReg(oldval).addReg(dest);
4876 BuildMI(BB, dl, TII->get(PPC::BCC))
4877 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4878 BB->addSuccessor(loop2MBB);
4879 BB->addSuccessor(midMBB);
4882 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4883 .addReg(newval).addReg(ptrA).addReg(ptrB);
4884 BuildMI(BB, dl, TII->get(PPC::BCC))
4885 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4886 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4887 BB->addSuccessor(loop1MBB);
4888 BB->addSuccessor(exitMBB);
4891 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4892 .addReg(dest).addReg(ptrA).addReg(ptrB);
4893 BB->addSuccessor(exitMBB);
4898 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4899 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4900 // We must use 64-bit registers for addresses when targeting 64-bit,
4901 // since we're actually doing arithmetic on them. Other registers
4903 bool is64bit = PPCSubTarget.isPPC64();
4904 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4906 unsigned dest = MI->getOperand(0).getReg();
4907 unsigned ptrA = MI->getOperand(1).getReg();
4908 unsigned ptrB = MI->getOperand(2).getReg();
4909 unsigned oldval = MI->getOperand(3).getReg();
4910 unsigned newval = MI->getOperand(4).getReg();
4911 DebugLoc dl = MI->getDebugLoc();
4913 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4914 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4915 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4916 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4917 F->insert(It, loop1MBB);
4918 F->insert(It, loop2MBB);
4919 F->insert(It, midMBB);
4920 F->insert(It, exitMBB);
4921 exitMBB->splice(exitMBB->begin(), BB,
4922 llvm::next(MachineBasicBlock::iterator(MI)),
4924 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
4926 MachineRegisterInfo &RegInfo = F->getRegInfo();
4927 const TargetRegisterClass *RC =
4928 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4929 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4930 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4931 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4932 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4933 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4934 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4935 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4936 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4937 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4938 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4939 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4940 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4941 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4942 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4944 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4945 unsigned ZeroReg = is64bit ? PPC::X0 : PPC::R0;
4948 // fallthrough --> loopMBB
4949 BB->addSuccessor(loop1MBB);
4951 // The 4-byte load must be aligned, while a char or short may be
4952 // anywhere in the word. Hence all this nasty bookkeeping code.
4953 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4954 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4955 // xori shift, shift1, 24 [16]
4956 // rlwinm ptr, ptr1, 0, 0, 29
4957 // slw newval2, newval, shift
4958 // slw oldval2, oldval,shift
4959 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4960 // slw mask, mask2, shift
4961 // and newval3, newval2, mask
4962 // and oldval3, oldval2, mask
4964 // lwarx tmpDest, ptr
4965 // and tmp, tmpDest, mask
4966 // cmpw tmp, oldval3
4969 // andc tmp2, tmpDest, mask
4970 // or tmp4, tmp2, newval3
4975 // stwcx. tmpDest, ptr
4977 // srw dest, tmpDest, shift
4978 if (ptrA != ZeroReg) {
4979 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4980 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4981 .addReg(ptrA).addReg(ptrB);
4985 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4986 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4987 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4988 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4990 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4991 .addReg(Ptr1Reg).addImm(0).addImm(61);
4993 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4994 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4995 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
4996 .addReg(newval).addReg(ShiftReg);
4997 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
4998 .addReg(oldval).addReg(ShiftReg);
5000 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
5002 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
5003 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
5004 .addReg(Mask3Reg).addImm(65535);
5006 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
5007 .addReg(Mask2Reg).addReg(ShiftReg);
5008 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
5009 .addReg(NewVal2Reg).addReg(MaskReg);
5010 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
5011 .addReg(OldVal2Reg).addReg(MaskReg);
5014 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
5015 .addReg(ZeroReg).addReg(PtrReg);
5016 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
5017 .addReg(TmpDestReg).addReg(MaskReg);
5018 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
5019 .addReg(TmpReg).addReg(OldVal3Reg);
5020 BuildMI(BB, dl, TII->get(PPC::BCC))
5021 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
5022 BB->addSuccessor(loop2MBB);
5023 BB->addSuccessor(midMBB);
5026 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
5027 .addReg(TmpDestReg).addReg(MaskReg);
5028 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
5029 .addReg(Tmp2Reg).addReg(NewVal3Reg);
5030 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
5031 .addReg(ZeroReg).addReg(PtrReg);
5032 BuildMI(BB, dl, TII->get(PPC::BCC))
5033 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
5034 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
5035 BB->addSuccessor(loop1MBB);
5036 BB->addSuccessor(exitMBB);
5039 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5040 .addReg(ZeroReg).addReg(PtrReg);
5041 BB->addSuccessor(exitMBB);
5046 BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW),dest).addReg(TmpReg)
5049 llvm_unreachable("Unexpected instr type to insert");
5052 MI->eraseFromParent(); // The pseudo instruction is gone now.
5056 //===----------------------------------------------------------------------===//
5057 // Target Optimization Hooks
5058 //===----------------------------------------------------------------------===//
5060 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5061 DAGCombinerInfo &DCI) const {
5062 const TargetMachine &TM = getTargetMachine();
5063 SelectionDAG &DAG = DCI.DAG;
5064 DebugLoc dl = N->getDebugLoc();
5065 switch (N->getOpcode()) {
5068 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5069 if (C->isNullValue()) // 0 << V -> 0.
5070 return N->getOperand(0);
5074 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5075 if (C->isNullValue()) // 0 >>u V -> 0.
5076 return N->getOperand(0);
5080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5081 if (C->isNullValue() || // 0 >>s V -> 0.
5082 C->isAllOnesValue()) // -1 >>s V -> -1.
5083 return N->getOperand(0);
5087 case ISD::SINT_TO_FP:
5088 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5089 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5090 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5091 // We allow the src/dst to be either f32/f64, but the intermediate
5092 // type must be i64.
5093 if (N->getOperand(0).getValueType() == MVT::i64 &&
5094 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5095 SDValue Val = N->getOperand(0).getOperand(0);
5096 if (Val.getValueType() == MVT::f32) {
5097 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5098 DCI.AddToWorklist(Val.getNode());
5101 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5102 DCI.AddToWorklist(Val.getNode());
5103 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5104 DCI.AddToWorklist(Val.getNode());
5105 if (N->getValueType(0) == MVT::f32) {
5106 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5107 DAG.getIntPtrConstant(0));
5108 DCI.AddToWorklist(Val.getNode());
5111 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5112 // If the intermediate type is i32, we can avoid the load/store here
5119 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5120 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5121 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5122 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5123 N->getOperand(1).getValueType() == MVT::i32 &&
5124 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5125 SDValue Val = N->getOperand(1).getOperand(0);
5126 if (Val.getValueType() == MVT::f32) {
5127 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5128 DCI.AddToWorklist(Val.getNode());
5130 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5131 DCI.AddToWorklist(Val.getNode());
5133 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5134 N->getOperand(2), N->getOperand(3));
5135 DCI.AddToWorklist(Val.getNode());
5139 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5140 if (cast<StoreSDNode>(N)->isUnindexed() &&
5141 N->getOperand(1).getOpcode() == ISD::BSWAP &&
5142 N->getOperand(1).getNode()->hasOneUse() &&
5143 (N->getOperand(1).getValueType() == MVT::i32 ||
5144 N->getOperand(1).getValueType() == MVT::i16)) {
5145 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5146 // Do an any-extend to 32-bits if this is a half-word input.
5147 if (BSwapOp.getValueType() == MVT::i16)
5148 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5151 N->getOperand(0), BSwapOp, N->getOperand(2),
5152 DAG.getValueType(N->getOperand(1).getValueType())
5155 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
5156 Ops, array_lengthof(Ops),
5157 cast<StoreSDNode>(N)->getMemoryVT(),
5158 cast<StoreSDNode>(N)->getMemOperand());
5162 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5163 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5164 N->getOperand(0).hasOneUse() &&
5165 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5166 SDValue Load = N->getOperand(0);
5167 LoadSDNode *LD = cast<LoadSDNode>(Load);
5168 // Create the byte-swapping load.
5170 LD->getChain(), // Chain
5171 LD->getBasePtr(), // Ptr
5172 DAG.getValueType(N->getValueType(0)) // VT
5175 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
5176 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
5177 LD->getMemoryVT(), LD->getMemOperand());
5179 // If this is an i16 load, insert the truncate.
5180 SDValue ResVal = BSLoad;
5181 if (N->getValueType(0) == MVT::i16)
5182 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5184 // First, combine the bswap away. This makes the value produced by the
5186 DCI.CombineTo(N, ResVal);
5188 // Next, combine the load away, we give it a bogus result value but a real
5189 // chain result. The result value is dead because the bswap is dead.
5190 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5192 // Return N so it doesn't get rechecked!
5193 return SDValue(N, 0);
5197 case PPCISD::VCMP: {
5198 // If a VCMPo node already exists with exactly the same operands as this
5199 // node, use its result instead of this node (VCMPo computes both a CR6 and
5200 // a normal output).
5202 if (!N->getOperand(0).hasOneUse() &&
5203 !N->getOperand(1).hasOneUse() &&
5204 !N->getOperand(2).hasOneUse()) {
5206 // Scan all of the users of the LHS, looking for VCMPo's that match.
5207 SDNode *VCMPoNode = 0;
5209 SDNode *LHSN = N->getOperand(0).getNode();
5210 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5212 if (UI->getOpcode() == PPCISD::VCMPo &&
5213 UI->getOperand(1) == N->getOperand(1) &&
5214 UI->getOperand(2) == N->getOperand(2) &&
5215 UI->getOperand(0) == N->getOperand(0)) {
5220 // If there is no VCMPo node, or if the flag value has a single use, don't
5222 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5225 // Look at the (necessarily single) use of the flag value. If it has a
5226 // chain, this transformation is more complex. Note that multiple things
5227 // could use the value result, which we should ignore.
5228 SDNode *FlagUser = 0;
5229 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5230 FlagUser == 0; ++UI) {
5231 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5233 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5234 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5241 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5242 // give up for right now.
5243 if (FlagUser->getOpcode() == PPCISD::MFCR)
5244 return SDValue(VCMPoNode, 0);
5249 // If this is a branch on an altivec predicate comparison, lower this so
5250 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5251 // lowering is done pre-legalize, because the legalizer lowers the predicate
5252 // compare down to code that is difficult to reassemble.
5253 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5254 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5258 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5259 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5260 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5261 assert(isDot && "Can't compare against a vector result!");
5263 // If this is a comparison against something other than 0/1, then we know
5264 // that the condition is never/always true.
5265 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5266 if (Val != 0 && Val != 1) {
5267 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5268 return N->getOperand(0);
5269 // Always !=, turn it into an unconditional branch.
5270 return DAG.getNode(ISD::BR, dl, MVT::Other,
5271 N->getOperand(0), N->getOperand(4));
5274 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5276 // Create the PPCISD altivec 'dot' comparison node.
5277 std::vector<EVT> VTs;
5279 LHS.getOperand(2), // LHS of compare
5280 LHS.getOperand(3), // RHS of compare
5281 DAG.getConstant(CompareOpc, MVT::i32)
5283 VTs.push_back(LHS.getOperand(2).getValueType());
5284 VTs.push_back(MVT::Glue);
5285 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5287 // Unpack the result based on how the target uses it.
5288 PPC::Predicate CompOpc;
5289 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5290 default: // Can't happen, don't crash on invalid number though.
5291 case 0: // Branch on the value of the EQ bit of CR6.
5292 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5294 case 1: // Branch on the inverted value of the EQ bit of CR6.
5295 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5297 case 2: // Branch on the value of the LT bit of CR6.
5298 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5300 case 3: // Branch on the inverted value of the LT bit of CR6.
5301 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5305 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5306 DAG.getConstant(CompOpc, MVT::i32),
5307 DAG.getRegister(PPC::CR6, MVT::i32),
5308 N->getOperand(4), CompNode.getValue(1));
5317 //===----------------------------------------------------------------------===//
5318 // Inline Assembly Support
5319 //===----------------------------------------------------------------------===//
5321 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5325 const SelectionDAG &DAG,
5326 unsigned Depth) const {
5327 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5328 switch (Op.getOpcode()) {
5330 case PPCISD::LBRX: {
5331 // lhbrx is known to have the top bits cleared out.
5332 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5333 KnownZero = 0xFFFF0000;
5336 case ISD::INTRINSIC_WO_CHAIN: {
5337 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5339 case Intrinsic::ppc_altivec_vcmpbfp_p:
5340 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5341 case Intrinsic::ppc_altivec_vcmpequb_p:
5342 case Intrinsic::ppc_altivec_vcmpequh_p:
5343 case Intrinsic::ppc_altivec_vcmpequw_p:
5344 case Intrinsic::ppc_altivec_vcmpgefp_p:
5345 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5346 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5347 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5348 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5349 case Intrinsic::ppc_altivec_vcmpgtub_p:
5350 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5351 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5352 KnownZero = ~1U; // All bits but the low one are known to be zero.
5360 /// getConstraintType - Given a constraint, return the type of
5361 /// constraint it is for this target.
5362 PPCTargetLowering::ConstraintType
5363 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5364 if (Constraint.size() == 1) {
5365 switch (Constraint[0]) {
5372 return C_RegisterClass;
5375 return TargetLowering::getConstraintType(Constraint);
5378 /// Examine constraint type and operand type and determine a weight value.
5379 /// This object must already have been set up with the operand type
5380 /// and the current alternative constraint selected.
5381 TargetLowering::ConstraintWeight
5382 PPCTargetLowering::getSingleConstraintMatchWeight(
5383 AsmOperandInfo &info, const char *constraint) const {
5384 ConstraintWeight weight = CW_Invalid;
5385 Value *CallOperandVal = info.CallOperandVal;
5386 // If we don't have a value, we can't do a match,
5387 // but allow it at the lowest weight.
5388 if (CallOperandVal == NULL)
5390 const Type *type = CallOperandVal->getType();
5391 // Look at the constraint type.
5392 switch (*constraint) {
5394 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
5397 if (type->isIntegerTy())
5398 weight = CW_Register;
5401 if (type->isFloatTy())
5402 weight = CW_Register;
5405 if (type->isDoubleTy())
5406 weight = CW_Register;
5409 if (type->isVectorTy())
5410 weight = CW_Register;
5413 weight = CW_Register;
5419 std::pair<unsigned, const TargetRegisterClass*>
5420 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5422 if (Constraint.size() == 1) {
5423 // GCC RS6000 Constraint Letters
5424 switch (Constraint[0]) {
5427 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5428 return std::make_pair(0U, PPC::G8RCRegisterClass);
5429 return std::make_pair(0U, PPC::GPRCRegisterClass);
5432 return std::make_pair(0U, PPC::F4RCRegisterClass);
5433 else if (VT == MVT::f64)
5434 return std::make_pair(0U, PPC::F8RCRegisterClass);
5437 return std::make_pair(0U, PPC::VRRCRegisterClass);
5439 return std::make_pair(0U, PPC::CRRCRegisterClass);
5443 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5447 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5448 /// vector. If it is invalid, don't add anything to Ops.
5449 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5450 std::string &Constraint,
5451 std::vector<SDValue>&Ops,
5452 SelectionDAG &DAG) const {
5453 SDValue Result(0,0);
5455 // Only support length 1 constraints.
5456 if (Constraint.length() > 1) return;
5458 char Letter = Constraint[0];
5469 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5470 if (!CST) return; // Must be an immediate to match.
5471 unsigned Value = CST->getZExtValue();
5473 default: llvm_unreachable("Unknown constraint letter!");
5474 case 'I': // "I" is a signed 16-bit constant.
5475 if ((short)Value == (int)Value)
5476 Result = DAG.getTargetConstant(Value, Op.getValueType());
5478 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5479 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5480 if ((short)Value == 0)
5481 Result = DAG.getTargetConstant(Value, Op.getValueType());
5483 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5484 if ((Value >> 16) == 0)
5485 Result = DAG.getTargetConstant(Value, Op.getValueType());
5487 case 'M': // "M" is a constant that is greater than 31.
5489 Result = DAG.getTargetConstant(Value, Op.getValueType());
5491 case 'N': // "N" is a positive constant that is an exact power of two.
5492 if ((int)Value > 0 && isPowerOf2_32(Value))
5493 Result = DAG.getTargetConstant(Value, Op.getValueType());
5495 case 'O': // "O" is the constant zero.
5497 Result = DAG.getTargetConstant(Value, Op.getValueType());
5499 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5500 if ((short)-Value == (int)-Value)
5501 Result = DAG.getTargetConstant(Value, Op.getValueType());
5508 if (Result.getNode()) {
5509 Ops.push_back(Result);
5513 // Handle standard constraint letters.
5514 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5517 // isLegalAddressingMode - Return true if the addressing mode represented
5518 // by AM is legal for this target, for a load/store of the specified type.
5519 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5520 const Type *Ty) const {
5521 // FIXME: PPC does not allow r+i addressing modes for vectors!
5523 // PPC allows a sign-extended 16-bit immediate field.
5524 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5527 // No global is ever allowed as a base.
5531 // PPC only support r+r,
5533 case 0: // "r+i" or just "i", depending on HasBaseReg.
5536 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5538 // Otherwise we have r+r or r+i.
5541 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5543 // Allow 2*r as r+r.
5546 // No other scales are supported.
5553 /// isLegalAddressImmediate - Return true if the integer value can be used
5554 /// as the offset of the target addressing mode for load / store of the
5556 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
5557 // PPC allows a sign-extended 16-bit immediate field.
5558 return (V > -(1 << 16) && V < (1 << 16)-1);
5561 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5565 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op,
5566 SelectionDAG &DAG) const {
5567 MachineFunction &MF = DAG.getMachineFunction();
5568 MachineFrameInfo *MFI = MF.getFrameInfo();
5569 MFI->setReturnAddressIsTaken(true);
5571 DebugLoc dl = Op.getDebugLoc();
5572 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5574 // Make sure the function does not optimize away the store of the RA to
5576 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5577 FuncInfo->setLRStoreRequired();
5578 bool isPPC64 = PPCSubTarget.isPPC64();
5579 bool isDarwinABI = PPCSubTarget.isDarwinABI();
5582 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
5585 DAG.getConstant(PPCFrameLowering::getReturnSaveOffset(isPPC64, isDarwinABI),
5586 isPPC64? MVT::i64 : MVT::i32);
5587 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5588 DAG.getNode(ISD::ADD, dl, getPointerTy(),
5590 MachinePointerInfo(), false, false, 0);
5593 // Just load the return address off the stack.
5594 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5595 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
5596 RetAddrFI, MachinePointerInfo(), false, false, 0);
5599 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op,
5600 SelectionDAG &DAG) const {
5601 DebugLoc dl = Op.getDebugLoc();
5602 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
5604 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5605 bool isPPC64 = PtrVT == MVT::i64;
5607 MachineFunction &MF = DAG.getMachineFunction();
5608 MachineFrameInfo *MFI = MF.getFrameInfo();
5609 MFI->setFrameAddressIsTaken(true);
5610 bool is31 = (DisableFramePointerElim(MF) || MFI->hasVarSizedObjects()) &&
5611 MFI->getStackSize() &&
5612 !MF.getFunction()->hasFnAttr(Attribute::Naked);
5613 unsigned FrameReg = isPPC64 ? (is31 ? PPC::X31 : PPC::X1) :
5614 (is31 ? PPC::R31 : PPC::R1);
5615 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg,
5618 FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(),
5619 FrameAddr, MachinePointerInfo(), false, false, 0);
5624 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5625 // The PowerPC target isn't yet aware of offsets.
5629 /// getOptimalMemOpType - Returns the target specific optimal type for load
5630 /// and store operations as a result of memset, memcpy, and memmove
5631 /// lowering. If DstAlign is zero that means it's safe to destination
5632 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
5633 /// means there isn't a need to check it against alignment requirement,
5634 /// probably because the source does not need to be loaded. If
5635 /// 'NonScalarIntSafe' is true, that means it's safe to return a
5636 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
5637 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
5638 /// constant so it does not need to be loaded.
5639 /// It returns EVT::Other if the type should be determined using generic
5640 /// target-independent logic.
5641 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size,
5642 unsigned DstAlign, unsigned SrcAlign,
5643 bool NonScalarIntSafe,
5645 MachineFunction &MF) const {
5646 if (this->PPCSubTarget.isPPC64()) {