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 "PPCPredicates.h"
17 #include "PPCTargetMachine.h"
18 #include "PPCPerfectShuffle.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/CallingConv.h"
29 #include "llvm/Constants.h"
30 #include "llvm/Function.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Target/TargetLoweringObjectFile.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, EVT &ValVT, EVT &LocVT,
42 CCValAssign::LocInfo &LocInfo,
43 ISD::ArgFlagsTy &ArgFlags,
45 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, EVT &ValVT,
47 CCValAssign::LocInfo &LocInfo,
48 ISD::ArgFlagsTy &ArgFlags,
50 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, EVT &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();
63 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 // Set up the register classes.
77 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
78 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
79 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
81 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
82 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
83 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
85 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
87 // PowerPC has pre-inc load and store's.
88 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
89 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
90 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
91 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
92 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
93 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
94 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
95 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
96 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
97 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
99 // This is used in the ppcf128->int sequence. Note it has different semantics
100 // from FP_ROUND: that rounds to nearest, this rounds to zero.
101 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
103 // PowerPC has no SREM/UREM instructions
104 setOperationAction(ISD::SREM, MVT::i32, Expand);
105 setOperationAction(ISD::UREM, MVT::i32, Expand);
106 setOperationAction(ISD::SREM, MVT::i64, Expand);
107 setOperationAction(ISD::UREM, MVT::i64, Expand);
109 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
110 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
111 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
112 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
113 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
114 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
115 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
116 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
117 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
119 // We don't support sin/cos/sqrt/fmod/pow
120 setOperationAction(ISD::FSIN , MVT::f64, Expand);
121 setOperationAction(ISD::FCOS , MVT::f64, Expand);
122 setOperationAction(ISD::FREM , MVT::f64, Expand);
123 setOperationAction(ISD::FPOW , MVT::f64, Expand);
124 setOperationAction(ISD::FSIN , MVT::f32, Expand);
125 setOperationAction(ISD::FCOS , MVT::f32, Expand);
126 setOperationAction(ISD::FREM , MVT::f32, Expand);
127 setOperationAction(ISD::FPOW , MVT::f32, Expand);
129 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
131 // If we're enabling GP optimizations, use hardware square root
132 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
133 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
134 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
137 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
138 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
140 // PowerPC does not have BSWAP, CTPOP or CTTZ
141 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
142 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
143 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
144 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
145 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
146 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
148 // PowerPC does not have ROTR
149 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
150 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
152 // PowerPC does not have Select
153 setOperationAction(ISD::SELECT, MVT::i32, Expand);
154 setOperationAction(ISD::SELECT, MVT::i64, Expand);
155 setOperationAction(ISD::SELECT, MVT::f32, Expand);
156 setOperationAction(ISD::SELECT, MVT::f64, Expand);
158 // PowerPC wants to turn select_cc of FP into fsel when possible.
159 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
160 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
162 // PowerPC wants to optimize integer setcc a bit
163 setOperationAction(ISD::SETCC, MVT::i32, Custom);
165 // PowerPC does not have BRCOND which requires SetCC
166 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
168 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
170 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
171 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
173 // PowerPC does not have [U|S]INT_TO_FP
174 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
175 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
177 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
178 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
179 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
180 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
182 // We cannot sextinreg(i1). Expand to shifts.
183 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
185 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
186 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
187 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
188 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
191 // We want to legalize GlobalAddress and ConstantPool nodes into the
192 // appropriate instructions to materialize the address.
193 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
194 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
195 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
196 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
197 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
198 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
199 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
200 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
201 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
202 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
205 setOperationAction(ISD::TRAP, MVT::Other, Legal);
207 // TRAMPOLINE is custom lowered.
208 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
210 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
211 setOperationAction(ISD::VASTART , MVT::Other, Custom);
213 // VAARG is custom lowered with the 32-bit SVR4 ABI.
214 if ( TM.getSubtarget<PPCSubtarget>().isSVR4ABI()
215 && !TM.getSubtarget<PPCSubtarget>().isPPC64())
216 setOperationAction(ISD::VAARG, MVT::Other, Custom);
218 setOperationAction(ISD::VAARG, MVT::Other, Expand);
220 // Use the default implementation.
221 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
222 setOperationAction(ISD::VAEND , MVT::Other, Expand);
223 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
224 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
225 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
226 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
228 // We want to custom lower some of our intrinsics.
229 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
231 // Comparisons that require checking two conditions.
232 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
233 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
234 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
235 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
236 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
237 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
238 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
239 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
240 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
241 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
242 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
243 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
245 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
246 // They also have instructions for converting between i64 and fp.
247 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
248 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
249 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
250 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
251 // This is just the low 32 bits of a (signed) fp->i64 conversion.
252 // We cannot do this with Promote because i64 is not a legal type.
253 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
255 // FIXME: disable this lowered code. This generates 64-bit register values,
256 // and we don't model the fact that the top part is clobbered by calls. We
257 // need to flag these together so that the value isn't live across a call.
258 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
260 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
261 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
264 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
265 // 64-bit PowerPC implementations can support i64 types directly
266 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
267 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
268 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
269 // 64-bit PowerPC wants to expand i128 shifts itself.
270 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
271 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
272 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
274 // 32-bit PowerPC wants to expand i64 shifts itself.
275 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
276 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
277 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
280 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
281 // First set operation action for all vector types to expand. Then we
282 // will selectively turn on ones that can be effectively codegen'd.
283 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
284 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
285 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
287 // add/sub are legal for all supported vector VT's.
288 setOperationAction(ISD::ADD , VT, Legal);
289 setOperationAction(ISD::SUB , VT, Legal);
291 // We promote all shuffles to v16i8.
292 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
293 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
295 // We promote all non-typed operations to v4i32.
296 setOperationAction(ISD::AND , VT, Promote);
297 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
298 setOperationAction(ISD::OR , VT, Promote);
299 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
300 setOperationAction(ISD::XOR , VT, Promote);
301 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
302 setOperationAction(ISD::LOAD , VT, Promote);
303 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
304 setOperationAction(ISD::SELECT, VT, Promote);
305 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
306 setOperationAction(ISD::STORE, VT, Promote);
307 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
309 // No other operations are legal.
310 setOperationAction(ISD::MUL , VT, Expand);
311 setOperationAction(ISD::SDIV, VT, Expand);
312 setOperationAction(ISD::SREM, VT, Expand);
313 setOperationAction(ISD::UDIV, VT, Expand);
314 setOperationAction(ISD::UREM, VT, Expand);
315 setOperationAction(ISD::FDIV, VT, Expand);
316 setOperationAction(ISD::FNEG, VT, Expand);
317 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
318 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
319 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
320 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
321 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
322 setOperationAction(ISD::UDIVREM, VT, Expand);
323 setOperationAction(ISD::SDIVREM, VT, Expand);
324 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
325 setOperationAction(ISD::FPOW, VT, Expand);
326 setOperationAction(ISD::CTPOP, VT, Expand);
327 setOperationAction(ISD::CTLZ, VT, Expand);
328 setOperationAction(ISD::CTTZ, VT, Expand);
331 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
332 // with merges, splats, etc.
333 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
335 setOperationAction(ISD::AND , MVT::v4i32, Legal);
336 setOperationAction(ISD::OR , MVT::v4i32, Legal);
337 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
338 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
339 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
340 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
342 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
343 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
344 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
345 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
347 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
348 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
349 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
350 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
352 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
353 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
355 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
356 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
357 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
358 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
361 setShiftAmountType(MVT::i32);
362 setBooleanContents(ZeroOrOneBooleanContent);
364 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
365 setStackPointerRegisterToSaveRestore(PPC::X1);
366 setExceptionPointerRegister(PPC::X3);
367 setExceptionSelectorRegister(PPC::X4);
369 setStackPointerRegisterToSaveRestore(PPC::R1);
370 setExceptionPointerRegister(PPC::R3);
371 setExceptionSelectorRegister(PPC::R4);
374 // We have target-specific dag combine patterns for the following nodes:
375 setTargetDAGCombine(ISD::SINT_TO_FP);
376 setTargetDAGCombine(ISD::STORE);
377 setTargetDAGCombine(ISD::BR_CC);
378 setTargetDAGCombine(ISD::BSWAP);
380 // Darwin long double math library functions have $LDBL128 appended.
381 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
382 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
383 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
384 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
385 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
386 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
387 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
388 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
389 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
390 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
391 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
394 computeRegisterProperties();
397 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
398 /// function arguments in the caller parameter area.
399 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
400 TargetMachine &TM = getTargetMachine();
401 // Darwin passes everything on 4 byte boundary.
402 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
408 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
411 case PPCISD::FSEL: return "PPCISD::FSEL";
412 case PPCISD::FCFID: return "PPCISD::FCFID";
413 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
414 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
415 case PPCISD::STFIWX: return "PPCISD::STFIWX";
416 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
417 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
418 case PPCISD::VPERM: return "PPCISD::VPERM";
419 case PPCISD::Hi: return "PPCISD::Hi";
420 case PPCISD::Lo: return "PPCISD::Lo";
421 case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY";
422 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
423 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
424 case PPCISD::SRL: return "PPCISD::SRL";
425 case PPCISD::SRA: return "PPCISD::SRA";
426 case PPCISD::SHL: return "PPCISD::SHL";
427 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
428 case PPCISD::STD_32: return "PPCISD::STD_32";
429 case PPCISD::CALL_SVR4: return "PPCISD::CALL_SVR4";
430 case PPCISD::CALL_Darwin: return "PPCISD::CALL_Darwin";
431 case PPCISD::NOP: return "PPCISD::NOP";
432 case PPCISD::MTCTR: return "PPCISD::MTCTR";
433 case PPCISD::BCTRL_Darwin: return "PPCISD::BCTRL_Darwin";
434 case PPCISD::BCTRL_SVR4: return "PPCISD::BCTRL_SVR4";
435 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
436 case PPCISD::MFCR: return "PPCISD::MFCR";
437 case PPCISD::VCMP: return "PPCISD::VCMP";
438 case PPCISD::VCMPo: return "PPCISD::VCMPo";
439 case PPCISD::LBRX: return "PPCISD::LBRX";
440 case PPCISD::STBRX: return "PPCISD::STBRX";
441 case PPCISD::LARX: return "PPCISD::LARX";
442 case PPCISD::STCX: return "PPCISD::STCX";
443 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
444 case PPCISD::MFFS: return "PPCISD::MFFS";
445 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
446 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
447 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
448 case PPCISD::MTFSF: return "PPCISD::MTFSF";
449 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
453 MVT::SimpleValueType PPCTargetLowering::getSetCCResultType(EVT VT) const {
457 /// getFunctionAlignment - Return the Log2 alignment of this function.
458 unsigned PPCTargetLowering::getFunctionAlignment(const Function *F) const {
459 if (getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin())
460 return F->hasFnAttr(Attribute::OptimizeForSize) ? 2 : 4;
465 //===----------------------------------------------------------------------===//
466 // Node matching predicates, for use by the tblgen matching code.
467 //===----------------------------------------------------------------------===//
469 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
470 static bool isFloatingPointZero(SDValue Op) {
471 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
472 return CFP->getValueAPF().isZero();
473 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
474 // Maybe this has already been legalized into the constant pool?
475 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
476 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
477 return CFP->getValueAPF().isZero();
482 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
483 /// true if Op is undef or if it matches the specified value.
484 static bool isConstantOrUndef(int Op, int Val) {
485 return Op < 0 || Op == Val;
488 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
489 /// VPKUHUM instruction.
490 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
492 for (unsigned i = 0; i != 16; ++i)
493 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
496 for (unsigned i = 0; i != 8; ++i)
497 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
498 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
504 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
505 /// VPKUWUM instruction.
506 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
508 for (unsigned i = 0; i != 16; i += 2)
509 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
510 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
513 for (unsigned i = 0; i != 8; i += 2)
514 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
515 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
516 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
517 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
523 /// isVMerge - Common function, used to match vmrg* shuffles.
525 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
526 unsigned LHSStart, unsigned RHSStart) {
527 assert(N->getValueType(0) == MVT::v16i8 &&
528 "PPC only supports shuffles by bytes!");
529 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
530 "Unsupported merge size!");
532 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
533 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
534 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
535 LHSStart+j+i*UnitSize) ||
536 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
537 RHSStart+j+i*UnitSize))
543 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
544 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
545 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
548 return isVMerge(N, UnitSize, 8, 24);
549 return isVMerge(N, UnitSize, 8, 8);
552 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
553 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
554 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
557 return isVMerge(N, UnitSize, 0, 16);
558 return isVMerge(N, UnitSize, 0, 0);
562 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
563 /// amount, otherwise return -1.
564 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
565 assert(N->getValueType(0) == MVT::v16i8 &&
566 "PPC only supports shuffles by bytes!");
568 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
570 // Find the first non-undef value in the shuffle mask.
572 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
575 if (i == 16) return -1; // all undef.
577 // Otherwise, check to see if the rest of the elements are consecutively
578 // numbered from this value.
579 unsigned ShiftAmt = SVOp->getMaskElt(i);
580 if (ShiftAmt < i) return -1;
584 // Check the rest of the elements to see if they are consecutive.
585 for (++i; i != 16; ++i)
586 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
589 // Check the rest of the elements to see if they are consecutive.
590 for (++i; i != 16; ++i)
591 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
597 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
598 /// specifies a splat of a single element that is suitable for input to
599 /// VSPLTB/VSPLTH/VSPLTW.
600 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
601 assert(N->getValueType(0) == MVT::v16i8 &&
602 (EltSize == 1 || EltSize == 2 || EltSize == 4));
604 // This is a splat operation if each element of the permute is the same, and
605 // if the value doesn't reference the second vector.
606 unsigned ElementBase = N->getMaskElt(0);
608 // FIXME: Handle UNDEF elements too!
609 if (ElementBase >= 16)
612 // Check that the indices are consecutive, in the case of a multi-byte element
613 // splatted with a v16i8 mask.
614 for (unsigned i = 1; i != EltSize; ++i)
615 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
618 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
619 if (N->getMaskElt(i) < 0) continue;
620 for (unsigned j = 0; j != EltSize; ++j)
621 if (N->getMaskElt(i+j) != N->getMaskElt(j))
627 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
629 bool PPC::isAllNegativeZeroVector(SDNode *N) {
630 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
632 APInt APVal, APUndef;
636 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32, true))
637 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
638 return CFP->getValueAPF().isNegZero();
643 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
644 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
645 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
646 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
647 assert(isSplatShuffleMask(SVOp, EltSize));
648 return SVOp->getMaskElt(0) / EltSize;
651 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
652 /// by using a vspltis[bhw] instruction of the specified element size, return
653 /// the constant being splatted. The ByteSize field indicates the number of
654 /// bytes of each element [124] -> [bhw].
655 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
658 // If ByteSize of the splat is bigger than the element size of the
659 // build_vector, then we have a case where we are checking for a splat where
660 // multiple elements of the buildvector are folded together into a single
661 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
662 unsigned EltSize = 16/N->getNumOperands();
663 if (EltSize < ByteSize) {
664 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
665 SDValue UniquedVals[4];
666 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
668 // See if all of the elements in the buildvector agree across.
669 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
670 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
671 // If the element isn't a constant, bail fully out.
672 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
675 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
676 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
677 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
678 return SDValue(); // no match.
681 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
682 // either constant or undef values that are identical for each chunk. See
683 // if these chunks can form into a larger vspltis*.
685 // Check to see if all of the leading entries are either 0 or -1. If
686 // neither, then this won't fit into the immediate field.
687 bool LeadingZero = true;
688 bool LeadingOnes = true;
689 for (unsigned i = 0; i != Multiple-1; ++i) {
690 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
692 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
693 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
695 // Finally, check the least significant entry.
697 if (UniquedVals[Multiple-1].getNode() == 0)
698 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
699 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
701 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
704 if (UniquedVals[Multiple-1].getNode() == 0)
705 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
706 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
707 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
708 return DAG.getTargetConstant(Val, MVT::i32);
714 // Check to see if this buildvec has a single non-undef value in its elements.
715 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
716 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
717 if (OpVal.getNode() == 0)
718 OpVal = N->getOperand(i);
719 else if (OpVal != N->getOperand(i))
723 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
725 unsigned ValSizeInBytes = EltSize;
727 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
728 Value = CN->getZExtValue();
729 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
730 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
731 Value = FloatToBits(CN->getValueAPF().convertToFloat());
734 // If the splat value is larger than the element value, then we can never do
735 // this splat. The only case that we could fit the replicated bits into our
736 // immediate field for would be zero, and we prefer to use vxor for it.
737 if (ValSizeInBytes < ByteSize) return SDValue();
739 // If the element value is larger than the splat value, cut it in half and
740 // check to see if the two halves are equal. Continue doing this until we
741 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
742 while (ValSizeInBytes > ByteSize) {
743 ValSizeInBytes >>= 1;
745 // If the top half equals the bottom half, we're still ok.
746 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
747 (Value & ((1 << (8*ValSizeInBytes))-1)))
751 // Properly sign extend the value.
752 int ShAmt = (4-ByteSize)*8;
753 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
755 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
756 if (MaskVal == 0) return SDValue();
758 // Finally, if this value fits in a 5 bit sext field, return it
759 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
760 return DAG.getTargetConstant(MaskVal, MVT::i32);
764 //===----------------------------------------------------------------------===//
765 // Addressing Mode Selection
766 //===----------------------------------------------------------------------===//
768 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
769 /// or 64-bit immediate, and if the value can be accurately represented as a
770 /// sign extension from a 16-bit value. If so, this returns true and the
772 static bool isIntS16Immediate(SDNode *N, short &Imm) {
773 if (N->getOpcode() != ISD::Constant)
776 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
777 if (N->getValueType(0) == MVT::i32)
778 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
780 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
782 static bool isIntS16Immediate(SDValue Op, short &Imm) {
783 return isIntS16Immediate(Op.getNode(), Imm);
787 /// SelectAddressRegReg - Given the specified addressed, check to see if it
788 /// can be represented as an indexed [r+r] operation. Returns false if it
789 /// can be more efficiently represented with [r+imm].
790 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
792 SelectionDAG &DAG) const {
794 if (N.getOpcode() == ISD::ADD) {
795 if (isIntS16Immediate(N.getOperand(1), imm))
797 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
800 Base = N.getOperand(0);
801 Index = N.getOperand(1);
803 } else if (N.getOpcode() == ISD::OR) {
804 if (isIntS16Immediate(N.getOperand(1), imm))
805 return false; // r+i can fold it if we can.
807 // If this is an or of disjoint bitfields, we can codegen this as an add
808 // (for better address arithmetic) if the LHS and RHS of the OR are provably
810 APInt LHSKnownZero, LHSKnownOne;
811 APInt RHSKnownZero, RHSKnownOne;
812 DAG.ComputeMaskedBits(N.getOperand(0),
813 APInt::getAllOnesValue(N.getOperand(0)
814 .getValueSizeInBits()),
815 LHSKnownZero, LHSKnownOne);
817 if (LHSKnownZero.getBoolValue()) {
818 DAG.ComputeMaskedBits(N.getOperand(1),
819 APInt::getAllOnesValue(N.getOperand(1)
820 .getValueSizeInBits()),
821 RHSKnownZero, RHSKnownOne);
822 // If all of the bits are known zero on the LHS or RHS, the add won't
824 if (~(LHSKnownZero | RHSKnownZero) == 0) {
825 Base = N.getOperand(0);
826 Index = N.getOperand(1);
835 /// Returns true if the address N can be represented by a base register plus
836 /// a signed 16-bit displacement [r+imm], and if it is not better
837 /// represented as reg+reg.
838 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
840 SelectionDAG &DAG) const {
841 // FIXME dl should come from parent load or store, not from address
842 DebugLoc dl = N.getDebugLoc();
843 // If this can be more profitably realized as r+r, fail.
844 if (SelectAddressRegReg(N, Disp, Base, DAG))
847 if (N.getOpcode() == ISD::ADD) {
849 if (isIntS16Immediate(N.getOperand(1), imm)) {
850 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
851 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
852 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
854 Base = N.getOperand(0);
856 return true; // [r+i]
857 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
858 // Match LOAD (ADD (X, Lo(G))).
859 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
860 && "Cannot handle constant offsets yet!");
861 Disp = N.getOperand(1).getOperand(0); // The global address.
862 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
863 Disp.getOpcode() == ISD::TargetConstantPool ||
864 Disp.getOpcode() == ISD::TargetJumpTable);
865 Base = N.getOperand(0);
866 return true; // [&g+r]
868 } else if (N.getOpcode() == ISD::OR) {
870 if (isIntS16Immediate(N.getOperand(1), imm)) {
871 // If this is an or of disjoint bitfields, we can codegen this as an add
872 // (for better address arithmetic) if the LHS and RHS of the OR are
873 // provably disjoint.
874 APInt LHSKnownZero, LHSKnownOne;
875 DAG.ComputeMaskedBits(N.getOperand(0),
876 APInt::getAllOnesValue(N.getOperand(0)
877 .getValueSizeInBits()),
878 LHSKnownZero, LHSKnownOne);
880 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
881 // If all of the bits are known zero on the LHS or RHS, the add won't
883 Base = N.getOperand(0);
884 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
888 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
889 // Loading from a constant address.
891 // If this address fits entirely in a 16-bit sext immediate field, codegen
894 if (isIntS16Immediate(CN, Imm)) {
895 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
896 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
900 // Handle 32-bit sext immediates with LIS + addr mode.
901 if (CN->getValueType(0) == MVT::i32 ||
902 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
903 int Addr = (int)CN->getZExtValue();
905 // Otherwise, break this down into an LIS + disp.
906 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
908 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
909 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
910 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0);
915 Disp = DAG.getTargetConstant(0, getPointerTy());
916 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
917 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
920 return true; // [r+0]
923 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
924 /// represented as an indexed [r+r] operation.
925 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
927 SelectionDAG &DAG) const {
928 // Check to see if we can easily represent this as an [r+r] address. This
929 // will fail if it thinks that the address is more profitably represented as
930 // reg+imm, e.g. where imm = 0.
931 if (SelectAddressRegReg(N, Base, Index, DAG))
934 // If the operand is an addition, always emit this as [r+r], since this is
935 // better (for code size, and execution, as the memop does the add for free)
936 // than emitting an explicit add.
937 if (N.getOpcode() == ISD::ADD) {
938 Base = N.getOperand(0);
939 Index = N.getOperand(1);
943 // Otherwise, do it the hard way, using R0 as the base register.
944 Base = DAG.getRegister(PPC::R0, N.getValueType());
949 /// SelectAddressRegImmShift - Returns true if the address N can be
950 /// represented by a base register plus a signed 14-bit displacement
951 /// [r+imm*4]. Suitable for use by STD and friends.
952 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
954 SelectionDAG &DAG) const {
955 // FIXME dl should come from the parent load or store, not the address
956 DebugLoc dl = N.getDebugLoc();
957 // If this can be more profitably realized as r+r, fail.
958 if (SelectAddressRegReg(N, Disp, Base, DAG))
961 if (N.getOpcode() == ISD::ADD) {
963 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
964 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
965 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
966 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
968 Base = N.getOperand(0);
970 return true; // [r+i]
971 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
972 // Match LOAD (ADD (X, Lo(G))).
973 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
974 && "Cannot handle constant offsets yet!");
975 Disp = N.getOperand(1).getOperand(0); // The global address.
976 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
977 Disp.getOpcode() == ISD::TargetConstantPool ||
978 Disp.getOpcode() == ISD::TargetJumpTable);
979 Base = N.getOperand(0);
980 return true; // [&g+r]
982 } else if (N.getOpcode() == ISD::OR) {
984 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
985 // If this is an or of disjoint bitfields, we can codegen this as an add
986 // (for better address arithmetic) if the LHS and RHS of the OR are
987 // provably disjoint.
988 APInt LHSKnownZero, LHSKnownOne;
989 DAG.ComputeMaskedBits(N.getOperand(0),
990 APInt::getAllOnesValue(N.getOperand(0)
991 .getValueSizeInBits()),
992 LHSKnownZero, LHSKnownOne);
993 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
994 // If all of the bits are known zero on the LHS or RHS, the add won't
996 Base = N.getOperand(0);
997 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
1001 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
1002 // Loading from a constant address. Verify low two bits are clear.
1003 if ((CN->getZExtValue() & 3) == 0) {
1004 // If this address fits entirely in a 14-bit sext immediate field, codegen
1007 if (isIntS16Immediate(CN, Imm)) {
1008 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1009 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
1013 // Fold the low-part of 32-bit absolute addresses into addr mode.
1014 if (CN->getValueType(0) == MVT::i32 ||
1015 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1016 int Addr = (int)CN->getZExtValue();
1018 // Otherwise, break this down into an LIS + disp.
1019 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1020 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1021 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1022 Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base),0);
1028 Disp = DAG.getTargetConstant(0, getPointerTy());
1029 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1030 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1033 return true; // [r+0]
1037 /// getPreIndexedAddressParts - returns true by value, base pointer and
1038 /// offset pointer and addressing mode by reference if the node's address
1039 /// can be legally represented as pre-indexed load / store address.
1040 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1042 ISD::MemIndexedMode &AM,
1043 SelectionDAG &DAG) const {
1044 // Disabled by default for now.
1045 if (!EnablePPCPreinc) return false;
1049 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1050 Ptr = LD->getBasePtr();
1051 VT = LD->getMemoryVT();
1053 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1055 Ptr = ST->getBasePtr();
1056 VT = ST->getMemoryVT();
1060 // PowerPC doesn't have preinc load/store instructions for vectors.
1064 // TODO: Check reg+reg first.
1066 // LDU/STU use reg+imm*4, others use reg+imm.
1067 if (VT != MVT::i64) {
1069 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1073 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1077 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1078 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1079 // sext i32 to i64 when addr mode is r+i.
1080 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1081 LD->getExtensionType() == ISD::SEXTLOAD &&
1082 isa<ConstantSDNode>(Offset))
1090 //===----------------------------------------------------------------------===//
1091 // LowerOperation implementation
1092 //===----------------------------------------------------------------------===//
1094 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1095 SelectionDAG &DAG) {
1096 EVT PtrVT = Op.getValueType();
1097 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1098 Constant *C = CP->getConstVal();
1099 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1100 SDValue Zero = DAG.getConstant(0, PtrVT);
1101 // FIXME there isn't really any debug info here
1102 DebugLoc dl = Op.getDebugLoc();
1104 const TargetMachine &TM = DAG.getTarget();
1106 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, CPI, Zero);
1107 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, CPI, Zero);
1109 // If this is a non-darwin platform, we don't support non-static relo models
1111 if (TM.getRelocationModel() == Reloc::Static ||
1112 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1113 // Generate non-pic code that has direct accesses to the constant pool.
1114 // The address of the global is just (hi(&g)+lo(&g)).
1115 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1118 if (TM.getRelocationModel() == Reloc::PIC_) {
1119 // With PIC, the first instruction is actually "GR+hi(&G)".
1120 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1121 DAG.getNode(PPCISD::GlobalBaseReg,
1122 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1125 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1129 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1130 EVT PtrVT = Op.getValueType();
1131 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1132 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1133 SDValue Zero = DAG.getConstant(0, PtrVT);
1134 // FIXME there isn't really any debug loc here
1135 DebugLoc dl = Op.getDebugLoc();
1137 const TargetMachine &TM = DAG.getTarget();
1139 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, JTI, Zero);
1140 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, JTI, Zero);
1142 // If this is a non-darwin platform, we don't support non-static relo models
1144 if (TM.getRelocationModel() == Reloc::Static ||
1145 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1146 // Generate non-pic code that has direct accesses to the constant pool.
1147 // The address of the global is just (hi(&g)+lo(&g)).
1148 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1151 if (TM.getRelocationModel() == Reloc::PIC_) {
1152 // With PIC, the first instruction is actually "GR+hi(&G)".
1153 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1154 DAG.getNode(PPCISD::GlobalBaseReg,
1155 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1158 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1162 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1163 SelectionDAG &DAG) {
1164 llvm_unreachable("TLS not implemented for PPC.");
1165 return SDValue(); // Not reached
1168 SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
1169 EVT PtrVT = Op.getValueType();
1170 DebugLoc DL = Op.getDebugLoc();
1172 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1173 SDValue TgtBA = DAG.getBlockAddress(BA, PtrVT, /*isTarget=*/true);
1174 SDValue Zero = DAG.getConstant(0, PtrVT);
1175 SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, TgtBA, Zero);
1176 SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, TgtBA, Zero);
1178 // If this is a non-darwin platform, we don't support non-static relo models
1180 const TargetMachine &TM = DAG.getTarget();
1181 if (TM.getRelocationModel() == Reloc::Static ||
1182 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1183 // Generate non-pic code that has direct accesses to globals.
1184 // The address of the global is just (hi(&g)+lo(&g)).
1185 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1188 if (TM.getRelocationModel() == Reloc::PIC_) {
1189 // With PIC, the first instruction is actually "GR+hi(&G)".
1190 Hi = DAG.getNode(ISD::ADD, DL, PtrVT,
1191 DAG.getNode(PPCISD::GlobalBaseReg,
1192 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1195 return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1198 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1199 SelectionDAG &DAG) {
1200 EVT PtrVT = Op.getValueType();
1201 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1202 GlobalValue *GV = GSDN->getGlobal();
1203 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1204 SDValue Zero = DAG.getConstant(0, PtrVT);
1205 // FIXME there isn't really any debug info here
1206 DebugLoc dl = GSDN->getDebugLoc();
1208 const TargetMachine &TM = DAG.getTarget();
1210 // 64-bit SVR4 ABI code is always position-independent.
1211 // The actual address of the GlobalValue is stored in the TOC.
1212 if (PPCSubTarget.isSVR4ABI() && PPCSubTarget.isPPC64()) {
1213 return DAG.getNode(PPCISD::TOC_ENTRY, dl, MVT::i64, GA,
1214 DAG.getRegister(PPC::X2, MVT::i64));
1217 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, GA, Zero);
1218 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, GA, Zero);
1220 // If this is a non-darwin platform, we don't support non-static relo models
1222 if (TM.getRelocationModel() == Reloc::Static ||
1223 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1224 // Generate non-pic code that has direct accesses to globals.
1225 // The address of the global is just (hi(&g)+lo(&g)).
1226 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1229 if (TM.getRelocationModel() == Reloc::PIC_) {
1230 // With PIC, the first instruction is actually "GR+hi(&G)".
1231 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1232 DAG.getNode(PPCISD::GlobalBaseReg,
1233 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1236 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1238 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV, TM))
1241 // If the global is weak or external, we have to go through the lazy
1243 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Lo, NULL, 0);
1246 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1247 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1248 DebugLoc dl = Op.getDebugLoc();
1250 // If we're comparing for equality to zero, expose the fact that this is
1251 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1252 // fold the new nodes.
1253 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1254 if (C->isNullValue() && CC == ISD::SETEQ) {
1255 EVT VT = Op.getOperand(0).getValueType();
1256 SDValue Zext = Op.getOperand(0);
1257 if (VT.bitsLT(MVT::i32)) {
1259 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1261 unsigned Log2b = Log2_32(VT.getSizeInBits());
1262 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1263 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1264 DAG.getConstant(Log2b, MVT::i32));
1265 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1267 // Leave comparisons against 0 and -1 alone for now, since they're usually
1268 // optimized. FIXME: revisit this when we can custom lower all setcc
1270 if (C->isAllOnesValue() || C->isNullValue())
1274 // If we have an integer seteq/setne, turn it into a compare against zero
1275 // by xor'ing the rhs with the lhs, which is faster than setting a
1276 // condition register, reading it back out, and masking the correct bit. The
1277 // normal approach here uses sub to do this instead of xor. Using xor exposes
1278 // the result to other bit-twiddling opportunities.
1279 EVT LHSVT = Op.getOperand(0).getValueType();
1280 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1281 EVT VT = Op.getValueType();
1282 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1284 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1289 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1290 int VarArgsFrameIndex,
1291 int VarArgsStackOffset,
1292 unsigned VarArgsNumGPR,
1293 unsigned VarArgsNumFPR,
1294 const PPCSubtarget &Subtarget) {
1296 llvm_unreachable("VAARG not yet implemented for the SVR4 ABI!");
1297 return SDValue(); // Not reached
1300 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
1301 SDValue Chain = Op.getOperand(0);
1302 SDValue Trmp = Op.getOperand(1); // trampoline
1303 SDValue FPtr = Op.getOperand(2); // nested function
1304 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1305 DebugLoc dl = Op.getDebugLoc();
1307 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1308 bool isPPC64 = (PtrVT == MVT::i64);
1309 const Type *IntPtrTy =
1310 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType(
1313 TargetLowering::ArgListTy Args;
1314 TargetLowering::ArgListEntry Entry;
1316 Entry.Ty = IntPtrTy;
1317 Entry.Node = Trmp; Args.push_back(Entry);
1319 // TrampSize == (isPPC64 ? 48 : 40);
1320 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1321 isPPC64 ? MVT::i64 : MVT::i32);
1322 Args.push_back(Entry);
1324 Entry.Node = FPtr; Args.push_back(Entry);
1325 Entry.Node = Nest; Args.push_back(Entry);
1327 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1328 std::pair<SDValue, SDValue> CallResult =
1329 LowerCallTo(Chain, Op.getValueType().getTypeForEVT(*DAG.getContext()),
1330 false, false, false, false, 0, CallingConv::C, false,
1331 /*isReturnValueUsed=*/true,
1332 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1336 { CallResult.first, CallResult.second };
1338 return DAG.getMergeValues(Ops, 2, dl);
1341 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1342 int VarArgsFrameIndex,
1343 int VarArgsStackOffset,
1344 unsigned VarArgsNumGPR,
1345 unsigned VarArgsNumFPR,
1346 const PPCSubtarget &Subtarget) {
1347 DebugLoc dl = Op.getDebugLoc();
1349 if (Subtarget.isDarwinABI() || Subtarget.isPPC64()) {
1350 // vastart just stores the address of the VarArgsFrameIndex slot into the
1351 // memory location argument.
1352 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1353 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1354 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1355 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1358 // For the 32-bit SVR4 ABI we follow the layout of the va_list struct.
1359 // We suppose the given va_list is already allocated.
1362 // char gpr; /* index into the array of 8 GPRs
1363 // * stored in the register save area
1364 // * gpr=0 corresponds to r3,
1365 // * gpr=1 to r4, etc.
1367 // char fpr; /* index into the array of 8 FPRs
1368 // * stored in the register save area
1369 // * fpr=0 corresponds to f1,
1370 // * fpr=1 to f2, etc.
1372 // char *overflow_arg_area;
1373 // /* location on stack that holds
1374 // * the next overflow argument
1376 // char *reg_save_area;
1377 // /* where r3:r10 and f1:f8 (if saved)
1383 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i32);
1384 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i32);
1387 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1389 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1390 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1392 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1393 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1395 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1396 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1398 uint64_t FPROffset = 1;
1399 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1401 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1403 // Store first byte : number of int regs
1404 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1405 Op.getOperand(1), SV, 0, MVT::i8);
1406 uint64_t nextOffset = FPROffset;
1407 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1410 // Store second byte : number of float regs
1411 SDValue secondStore =
1412 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr, SV, nextOffset, MVT::i8);
1413 nextOffset += StackOffset;
1414 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1416 // Store second word : arguments given on stack
1417 SDValue thirdStore =
1418 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr, SV, nextOffset);
1419 nextOffset += FrameOffset;
1420 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1422 // Store third word : arguments given in registers
1423 return DAG.getStore(thirdStore, dl, FR, nextPtr, SV, nextOffset);
1427 #include "PPCGenCallingConv.inc"
1429 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, EVT &ValVT, EVT &LocVT,
1430 CCValAssign::LocInfo &LocInfo,
1431 ISD::ArgFlagsTy &ArgFlags,
1436 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, EVT &ValVT,
1438 CCValAssign::LocInfo &LocInfo,
1439 ISD::ArgFlagsTy &ArgFlags,
1441 static const unsigned ArgRegs[] = {
1442 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1443 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1445 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1447 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1449 // Skip one register if the first unallocated register has an even register
1450 // number and there are still argument registers available which have not been
1451 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1452 // need to skip a register if RegNum is odd.
1453 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1454 State.AllocateReg(ArgRegs[RegNum]);
1457 // Always return false here, as this function only makes sure that the first
1458 // unallocated register has an odd register number and does not actually
1459 // allocate a register for the current argument.
1463 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, EVT &ValVT,
1465 CCValAssign::LocInfo &LocInfo,
1466 ISD::ArgFlagsTy &ArgFlags,
1468 static const unsigned ArgRegs[] = {
1469 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1473 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1475 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1477 // If there is only one Floating-point register left we need to put both f64
1478 // values of a split ppc_fp128 value on the stack.
1479 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1480 State.AllocateReg(ArgRegs[RegNum]);
1483 // Always return false here, as this function only makes sure that the two f64
1484 // values a ppc_fp128 value is split into are both passed in registers or both
1485 // passed on the stack and does not actually allocate a register for the
1486 // current argument.
1490 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1492 static const unsigned *GetFPR() {
1493 static const unsigned FPR[] = {
1494 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1495 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1501 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1503 static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags,
1504 unsigned PtrByteSize) {
1505 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1506 if (Flags.isByVal())
1507 ArgSize = Flags.getByValSize();
1508 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1514 PPCTargetLowering::LowerFormalArguments(SDValue Chain,
1515 CallingConv::ID CallConv, bool isVarArg,
1516 const SmallVectorImpl<ISD::InputArg>
1518 DebugLoc dl, SelectionDAG &DAG,
1519 SmallVectorImpl<SDValue> &InVals) {
1520 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
1521 return LowerFormalArguments_SVR4(Chain, CallConv, isVarArg, Ins,
1524 return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins,
1530 PPCTargetLowering::LowerFormalArguments_SVR4(
1532 CallingConv::ID CallConv, bool isVarArg,
1533 const SmallVectorImpl<ISD::InputArg>
1535 DebugLoc dl, SelectionDAG &DAG,
1536 SmallVectorImpl<SDValue> &InVals) {
1538 // 32-bit SVR4 ABI Stack Frame Layout:
1539 // +-----------------------------------+
1540 // +--> | Back chain |
1541 // | +-----------------------------------+
1542 // | | Floating-point register save area |
1543 // | +-----------------------------------+
1544 // | | General register save area |
1545 // | +-----------------------------------+
1546 // | | CR save word |
1547 // | +-----------------------------------+
1548 // | | VRSAVE save word |
1549 // | +-----------------------------------+
1550 // | | Alignment padding |
1551 // | +-----------------------------------+
1552 // | | Vector register save area |
1553 // | +-----------------------------------+
1554 // | | Local variable space |
1555 // | +-----------------------------------+
1556 // | | Parameter list area |
1557 // | +-----------------------------------+
1558 // | | LR save word |
1559 // | +-----------------------------------+
1560 // SP--> +--- | Back chain |
1561 // +-----------------------------------+
1564 // System V Application Binary Interface PowerPC Processor Supplement
1565 // AltiVec Technology Programming Interface Manual
1567 MachineFunction &MF = DAG.getMachineFunction();
1568 MachineFrameInfo *MFI = MF.getFrameInfo();
1570 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1571 // Potential tail calls could cause overwriting of argument stack slots.
1572 bool isImmutable = !(PerformTailCallOpt && (CallConv==CallingConv::Fast));
1573 unsigned PtrByteSize = 4;
1575 // Assign locations to all of the incoming arguments.
1576 SmallVector<CCValAssign, 16> ArgLocs;
1577 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1580 // Reserve space for the linkage area on the stack.
1581 CCInfo.AllocateStack(PPCFrameInfo::getLinkageSize(false, false), PtrByteSize);
1583 CCInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4);
1585 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1586 CCValAssign &VA = ArgLocs[i];
1588 // Arguments stored in registers.
1589 if (VA.isRegLoc()) {
1590 TargetRegisterClass *RC;
1591 EVT ValVT = VA.getValVT();
1593 switch (ValVT.getSimpleVT().SimpleTy) {
1595 llvm_unreachable("ValVT not supported by formal arguments Lowering");
1597 RC = PPC::GPRCRegisterClass;
1600 RC = PPC::F4RCRegisterClass;
1603 RC = PPC::F8RCRegisterClass;
1609 RC = PPC::VRRCRegisterClass;
1613 // Transform the arguments stored in physical registers into virtual ones.
1614 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1615 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, ValVT);
1617 InVals.push_back(ArgValue);
1619 // Argument stored in memory.
1620 assert(VA.isMemLoc());
1622 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1623 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1624 isImmutable, false);
1626 // Create load nodes to retrieve arguments from the stack.
1627 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1628 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
1632 // Assign locations to all of the incoming aggregate by value arguments.
1633 // Aggregates passed by value are stored in the local variable space of the
1634 // caller's stack frame, right above the parameter list area.
1635 SmallVector<CCValAssign, 16> ByValArgLocs;
1636 CCState CCByValInfo(CallConv, isVarArg, getTargetMachine(),
1637 ByValArgLocs, *DAG.getContext());
1639 // Reserve stack space for the allocations in CCInfo.
1640 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1642 CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC_SVR4_ByVal);
1644 // Area that is at least reserved in the caller of this function.
1645 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1647 // Set the size that is at least reserved in caller of this function. Tail
1648 // call optimized function's reserved stack space needs to be aligned so that
1649 // taking the difference between two stack areas will result in an aligned
1651 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1654 std::max(MinReservedArea,
1655 PPCFrameInfo::getMinCallFrameSize(false, false));
1657 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1658 getStackAlignment();
1659 unsigned AlignMask = TargetAlign-1;
1660 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1662 FI->setMinReservedArea(MinReservedArea);
1664 SmallVector<SDValue, 8> MemOps;
1666 // If the function takes variable number of arguments, make a frame index for
1667 // the start of the first vararg value... for expansion of llvm.va_start.
1669 static const unsigned GPArgRegs[] = {
1670 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1671 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1673 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1675 static const unsigned FPArgRegs[] = {
1676 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1679 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1681 VarArgsNumGPR = CCInfo.getFirstUnallocated(GPArgRegs, NumGPArgRegs);
1682 VarArgsNumFPR = CCInfo.getFirstUnallocated(FPArgRegs, NumFPArgRegs);
1684 // Make room for NumGPArgRegs and NumFPArgRegs.
1685 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1686 NumFPArgRegs * EVT(MVT::f64).getSizeInBits()/8;
1688 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1689 CCInfo.getNextStackOffset(),
1692 VarArgsFrameIndex = MFI->CreateStackObject(Depth, 8, false);
1693 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1695 // The fixed integer arguments of a variadic function are
1696 // stored to the VarArgsFrameIndex on the stack.
1697 unsigned GPRIndex = 0;
1698 for (; GPRIndex != VarArgsNumGPR; ++GPRIndex) {
1699 SDValue Val = DAG.getRegister(GPArgRegs[GPRIndex], PtrVT);
1700 SDValue Store = DAG.getStore(Chain, dl, Val, FIN, NULL, 0);
1701 MemOps.push_back(Store);
1702 // Increment the address by four for the next argument to store
1703 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1704 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1707 // If this function is vararg, store any remaining integer argument regs
1708 // to their spots on the stack so that they may be loaded by deferencing the
1709 // result of va_next.
1710 for (; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1711 unsigned VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1713 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1714 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1715 MemOps.push_back(Store);
1716 // Increment the address by four for the next argument to store
1717 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1718 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1721 // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6
1724 // The double arguments are stored to the VarArgsFrameIndex
1726 unsigned FPRIndex = 0;
1727 for (FPRIndex = 0; FPRIndex != VarArgsNumFPR; ++FPRIndex) {
1728 SDValue Val = DAG.getRegister(FPArgRegs[FPRIndex], MVT::f64);
1729 SDValue Store = DAG.getStore(Chain, dl, Val, FIN, NULL, 0);
1730 MemOps.push_back(Store);
1731 // Increment the address by eight for the next argument to store
1732 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1734 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1737 for (; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1738 unsigned VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1740 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64);
1741 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1742 MemOps.push_back(Store);
1743 // Increment the address by eight for the next argument to store
1744 SDValue PtrOff = DAG.getConstant(EVT(MVT::f64).getSizeInBits()/8,
1746 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1750 if (!MemOps.empty())
1751 Chain = DAG.getNode(ISD::TokenFactor, dl,
1752 MVT::Other, &MemOps[0], MemOps.size());
1758 PPCTargetLowering::LowerFormalArguments_Darwin(
1760 CallingConv::ID CallConv, bool isVarArg,
1761 const SmallVectorImpl<ISD::InputArg>
1763 DebugLoc dl, SelectionDAG &DAG,
1764 SmallVectorImpl<SDValue> &InVals) {
1765 // TODO: add description of PPC stack frame format, or at least some docs.
1767 MachineFunction &MF = DAG.getMachineFunction();
1768 MachineFrameInfo *MFI = MF.getFrameInfo();
1770 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1771 bool isPPC64 = PtrVT == MVT::i64;
1772 // Potential tail calls could cause overwriting of argument stack slots.
1773 bool isImmutable = !(PerformTailCallOpt && (CallConv==CallingConv::Fast));
1774 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1776 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, true);
1777 // Area that is at least reserved in caller of this function.
1778 unsigned MinReservedArea = ArgOffset;
1780 static const unsigned GPR_32[] = { // 32-bit registers.
1781 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1782 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1784 static const unsigned GPR_64[] = { // 64-bit registers.
1785 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1786 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1789 static const unsigned *FPR = GetFPR();
1791 static const unsigned VR[] = {
1792 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1793 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1796 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1797 const unsigned Num_FPR_Regs = 13;
1798 const unsigned Num_VR_Regs = array_lengthof( VR);
1800 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1802 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1804 // In 32-bit non-varargs functions, the stack space for vectors is after the
1805 // stack space for non-vectors. We do not use this space unless we have
1806 // too many vectors to fit in registers, something that only occurs in
1807 // constructed examples:), but we have to walk the arglist to figure
1808 // that out...for the pathological case, compute VecArgOffset as the
1809 // start of the vector parameter area. Computing VecArgOffset is the
1810 // entire point of the following loop.
1811 unsigned VecArgOffset = ArgOffset;
1812 if (!isVarArg && !isPPC64) {
1813 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e;
1815 EVT ObjectVT = Ins[ArgNo].VT;
1816 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1817 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1819 if (Flags.isByVal()) {
1820 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1821 ObjSize = Flags.getByValSize();
1823 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1824 VecArgOffset += ArgSize;
1828 switch(ObjectVT.getSimpleVT().SimpleTy) {
1829 default: llvm_unreachable("Unhandled argument type!");
1832 VecArgOffset += isPPC64 ? 8 : 4;
1834 case MVT::i64: // PPC64
1842 // Nothing to do, we're only looking at Nonvector args here.
1847 // We've found where the vector parameter area in memory is. Skip the
1848 // first 12 parameters; these don't use that memory.
1849 VecArgOffset = ((VecArgOffset+15)/16)*16;
1850 VecArgOffset += 12*16;
1852 // Add DAG nodes to load the arguments or copy them out of registers. On
1853 // entry to a function on PPC, the arguments start after the linkage area,
1854 // although the first ones are often in registers.
1856 SmallVector<SDValue, 8> MemOps;
1857 unsigned nAltivecParamsAtEnd = 0;
1858 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1860 bool needsLoad = false;
1861 EVT ObjectVT = Ins[ArgNo].VT;
1862 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1863 unsigned ArgSize = ObjSize;
1864 ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
1866 unsigned CurArgOffset = ArgOffset;
1868 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1869 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1870 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1871 if (isVarArg || isPPC64) {
1872 MinReservedArea = ((MinReservedArea+15)/16)*16;
1873 MinReservedArea += CalculateStackSlotSize(ObjectVT,
1876 } else nAltivecParamsAtEnd++;
1878 // Calculate min reserved area.
1879 MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT,
1883 // FIXME the codegen can be much improved in some cases.
1884 // We do not have to keep everything in memory.
1885 if (Flags.isByVal()) {
1886 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1887 ObjSize = Flags.getByValSize();
1888 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1889 // Objects of size 1 and 2 are right justified, everything else is
1890 // left justified. This means the memory address is adjusted forwards.
1891 if (ObjSize==1 || ObjSize==2) {
1892 CurArgOffset = CurArgOffset + (4 - ObjSize);
1894 // The value of the object is its address.
1895 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset, true, false);
1896 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1897 InVals.push_back(FIN);
1898 if (ObjSize==1 || ObjSize==2) {
1899 if (GPR_idx != Num_GPR_Regs) {
1900 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1901 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1902 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1903 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1904 MemOps.push_back(Store);
1908 ArgOffset += PtrByteSize;
1912 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1913 // Store whatever pieces of the object are in registers
1914 // to memory. ArgVal will be address of the beginning of
1916 if (GPR_idx != Num_GPR_Regs) {
1917 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1918 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset, true, false);
1919 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1920 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
1921 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1922 MemOps.push_back(Store);
1924 ArgOffset += PtrByteSize;
1926 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1933 switch (ObjectVT.getSimpleVT().SimpleTy) {
1934 default: llvm_unreachable("Unhandled argument type!");
1937 if (GPR_idx != Num_GPR_Regs) {
1938 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1939 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
1943 ArgSize = PtrByteSize;
1945 // All int arguments reserve stack space in the Darwin ABI.
1946 ArgOffset += PtrByteSize;
1950 case MVT::i64: // PPC64
1951 if (GPR_idx != Num_GPR_Regs) {
1952 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1953 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1955 if (ObjectVT == MVT::i32) {
1956 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1957 // value to MVT::i64 and then truncate to the correct register size.
1959 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1960 DAG.getValueType(ObjectVT));
1961 else if (Flags.isZExt())
1962 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1963 DAG.getValueType(ObjectVT));
1965 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1971 ArgSize = PtrByteSize;
1973 // All int arguments reserve stack space in the Darwin ABI.
1979 // Every 4 bytes of argument space consumes one of the GPRs available for
1980 // argument passing.
1981 if (GPR_idx != Num_GPR_Regs) {
1983 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1986 if (FPR_idx != Num_FPR_Regs) {
1989 if (ObjectVT == MVT::f32)
1990 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
1992 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
1994 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2000 // All FP arguments reserve stack space in the Darwin ABI.
2001 ArgOffset += isPPC64 ? 8 : ObjSize;
2007 // Note that vector arguments in registers don't reserve stack space,
2008 // except in varargs functions.
2009 if (VR_idx != Num_VR_Regs) {
2010 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
2011 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
2013 while ((ArgOffset % 16) != 0) {
2014 ArgOffset += PtrByteSize;
2015 if (GPR_idx != Num_GPR_Regs)
2019 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64?
2023 if (!isVarArg && !isPPC64) {
2024 // Vectors go after all the nonvectors.
2025 CurArgOffset = VecArgOffset;
2028 // Vectors are aligned.
2029 ArgOffset = ((ArgOffset+15)/16)*16;
2030 CurArgOffset = ArgOffset;
2038 // We need to load the argument to a virtual register if we determined above
2039 // that we ran out of physical registers of the appropriate type.
2041 int FI = MFI->CreateFixedObject(ObjSize,
2042 CurArgOffset + (ArgSize - ObjSize),
2043 isImmutable, false);
2044 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2045 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, NULL, 0);
2048 InVals.push_back(ArgVal);
2051 // Set the size that is at least reserved in caller of this function. Tail
2052 // call optimized function's reserved stack space needs to be aligned so that
2053 // taking the difference between two stack areas will result in an aligned
2055 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2056 // Add the Altivec parameters at the end, if needed.
2057 if (nAltivecParamsAtEnd) {
2058 MinReservedArea = ((MinReservedArea+15)/16)*16;
2059 MinReservedArea += 16*nAltivecParamsAtEnd;
2062 std::max(MinReservedArea,
2063 PPCFrameInfo::getMinCallFrameSize(isPPC64, true));
2064 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
2065 getStackAlignment();
2066 unsigned AlignMask = TargetAlign-1;
2067 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2068 FI->setMinReservedArea(MinReservedArea);
2070 // If the function takes variable number of arguments, make a frame index for
2071 // the start of the first vararg value... for expansion of llvm.va_start.
2073 int Depth = ArgOffset;
2075 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2076 Depth, true, false);
2077 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
2079 // If this function is vararg, store any remaining integer argument regs
2080 // to their spots on the stack so that they may be loaded by deferencing the
2081 // result of va_next.
2082 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2086 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2088 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2090 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT);
2091 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
2092 MemOps.push_back(Store);
2093 // Increment the address by four for the next argument to store
2094 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2095 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2099 if (!MemOps.empty())
2100 Chain = DAG.getNode(ISD::TokenFactor, dl,
2101 MVT::Other, &MemOps[0], MemOps.size());
2106 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2107 /// linkage area for the Darwin ABI.
2109 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2113 const SmallVectorImpl<ISD::OutputArg>
2115 unsigned &nAltivecParamsAtEnd) {
2116 // Count how many bytes are to be pushed on the stack, including the linkage
2117 // area, and parameter passing area. We start with 24/48 bytes, which is
2118 // prereserved space for [SP][CR][LR][3 x unused].
2119 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, true);
2120 unsigned NumOps = Outs.size();
2121 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2123 // Add up all the space actually used.
2124 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2125 // they all go in registers, but we must reserve stack space for them for
2126 // possible use by the caller. In varargs or 64-bit calls, parameters are
2127 // assigned stack space in order, with padding so Altivec parameters are
2129 nAltivecParamsAtEnd = 0;
2130 for (unsigned i = 0; i != NumOps; ++i) {
2131 SDValue Arg = Outs[i].Val;
2132 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2133 EVT ArgVT = Arg.getValueType();
2134 // Varargs Altivec parameters are padded to a 16 byte boundary.
2135 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2136 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2137 if (!isVarArg && !isPPC64) {
2138 // Non-varargs Altivec parameters go after all the non-Altivec
2139 // parameters; handle those later so we know how much padding we need.
2140 nAltivecParamsAtEnd++;
2143 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2144 NumBytes = ((NumBytes+15)/16)*16;
2146 NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
2149 // Allow for Altivec parameters at the end, if needed.
2150 if (nAltivecParamsAtEnd) {
2151 NumBytes = ((NumBytes+15)/16)*16;
2152 NumBytes += 16*nAltivecParamsAtEnd;
2155 // The prolog code of the callee may store up to 8 GPR argument registers to
2156 // the stack, allowing va_start to index over them in memory if its varargs.
2157 // Because we cannot tell if this is needed on the caller side, we have to
2158 // conservatively assume that it is needed. As such, make sure we have at
2159 // least enough stack space for the caller to store the 8 GPRs.
2160 NumBytes = std::max(NumBytes,
2161 PPCFrameInfo::getMinCallFrameSize(isPPC64, true));
2163 // Tail call needs the stack to be aligned.
2164 if (CC==CallingConv::Fast && PerformTailCallOpt) {
2165 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
2166 getStackAlignment();
2167 unsigned AlignMask = TargetAlign-1;
2168 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2174 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2175 /// adjusted to accomodate the arguments for the tailcall.
2176 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
2177 unsigned ParamSize) {
2179 if (!IsTailCall) return 0;
2181 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2182 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2183 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2184 // Remember only if the new adjustement is bigger.
2185 if (SPDiff < FI->getTailCallSPDelta())
2186 FI->setTailCallSPDelta(SPDiff);
2191 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2192 /// for tail call optimization. Targets which want to do tail call
2193 /// optimization should implement this function.
2195 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2196 CallingConv::ID CalleeCC,
2198 const SmallVectorImpl<ISD::InputArg> &Ins,
2199 SelectionDAG& DAG) const {
2200 // Variable argument functions are not supported.
2204 MachineFunction &MF = DAG.getMachineFunction();
2205 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2206 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2207 // Functions containing by val parameters are not supported.
2208 for (unsigned i = 0; i != Ins.size(); i++) {
2209 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2210 if (Flags.isByVal()) return false;
2213 // Non PIC/GOT tail calls are supported.
2214 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2217 // At the moment we can only do local tail calls (in same module, hidden
2218 // or protected) if we are generating PIC.
2219 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2220 return G->getGlobal()->hasHiddenVisibility()
2221 || G->getGlobal()->hasProtectedVisibility();
2227 /// isCallCompatibleAddress - Return the immediate to use if the specified
2228 /// 32-bit value is representable in the immediate field of a BxA instruction.
2229 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2230 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2233 int Addr = C->getZExtValue();
2234 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2235 (Addr << 6 >> 6) != Addr)
2236 return 0; // Top 6 bits have to be sext of immediate.
2238 return DAG.getConstant((int)C->getZExtValue() >> 2,
2239 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2244 struct TailCallArgumentInfo {
2249 TailCallArgumentInfo() : FrameIdx(0) {}
2254 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2256 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2258 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2259 SmallVector<SDValue, 8> &MemOpChains,
2261 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2262 SDValue Arg = TailCallArgs[i].Arg;
2263 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2264 int FI = TailCallArgs[i].FrameIdx;
2265 // Store relative to framepointer.
2266 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2267 PseudoSourceValue::getFixedStack(FI),
2272 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2273 /// the appropriate stack slot for the tail call optimized function call.
2274 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2275 MachineFunction &MF,
2284 // Calculate the new stack slot for the return address.
2285 int SlotSize = isPPC64 ? 8 : 4;
2286 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
2288 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2291 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2292 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2293 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2294 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
2296 // When using the 32/64-bit SVR4 ABI there is no need to move the FP stack
2297 // slot as the FP is never overwritten.
2300 SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64, isDarwinABI);
2301 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc,
2303 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2304 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2305 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
2311 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2312 /// the position of the argument.
2314 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2315 SDValue Arg, int SPDiff, unsigned ArgOffset,
2316 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2317 int Offset = ArgOffset + SPDiff;
2318 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2319 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true,false);
2320 EVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2321 SDValue FIN = DAG.getFrameIndex(FI, VT);
2322 TailCallArgumentInfo Info;
2324 Info.FrameIdxOp = FIN;
2326 TailCallArguments.push_back(Info);
2329 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2330 /// stack slot. Returns the chain as result and the loaded frame pointers in
2331 /// LROpOut/FPOpout. Used when tail calling.
2332 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2340 // Load the LR and FP stack slot for later adjusting.
2341 EVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2342 LROpOut = getReturnAddrFrameIndex(DAG);
2343 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, NULL, 0);
2344 Chain = SDValue(LROpOut.getNode(), 1);
2346 // When using the 32/64-bit SVR4 ABI there is no need to load the FP stack
2347 // slot as the FP is never overwritten.
2349 FPOpOut = getFramePointerFrameIndex(DAG);
2350 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, NULL, 0);
2351 Chain = SDValue(FPOpOut.getNode(), 1);
2357 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2358 /// by "Src" to address "Dst" of size "Size". Alignment information is
2359 /// specified by the specific parameter attribute. The copy will be passed as
2360 /// a byval function parameter.
2361 /// Sometimes what we are copying is the end of a larger object, the part that
2362 /// does not fit in registers.
2364 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2365 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2367 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2368 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2369 false, NULL, 0, NULL, 0);
2372 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2375 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2376 SDValue Arg, SDValue PtrOff, int SPDiff,
2377 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2378 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2379 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments,
2381 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2386 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2388 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2389 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2390 DAG.getConstant(ArgOffset, PtrVT));
2392 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0));
2393 // Calculate and remember argument location.
2394 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2399 void PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain,
2400 DebugLoc dl, bool isPPC64, int SPDiff, unsigned NumBytes,
2401 SDValue LROp, SDValue FPOp, bool isDarwinABI,
2402 SmallVector<TailCallArgumentInfo, 8> &TailCallArguments) {
2403 MachineFunction &MF = DAG.getMachineFunction();
2405 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2406 // might overwrite each other in case of tail call optimization.
2407 SmallVector<SDValue, 8> MemOpChains2;
2408 // Do not flag preceeding copytoreg stuff together with the following stuff.
2410 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2412 if (!MemOpChains2.empty())
2413 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2414 &MemOpChains2[0], MemOpChains2.size());
2416 // Store the return address to the appropriate stack slot.
2417 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2418 isPPC64, isDarwinABI, dl);
2420 // Emit callseq_end just before tailcall node.
2421 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2422 DAG.getIntPtrConstant(0, true), InFlag);
2423 InFlag = Chain.getValue(1);
2427 unsigned PrepareCall(SelectionDAG &DAG, SDValue &Callee, SDValue &InFlag,
2428 SDValue &Chain, DebugLoc dl, int SPDiff, bool isTailCall,
2429 SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass,
2430 SmallVector<SDValue, 8> &Ops, std::vector<EVT> &NodeTys,
2432 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2433 NodeTys.push_back(MVT::Other); // Returns a chain
2434 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2436 unsigned CallOpc = isSVR4ABI ? PPCISD::CALL_SVR4 : PPCISD::CALL_Darwin;
2438 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2439 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2440 // node so that legalize doesn't hack it.
2441 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2442 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2443 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2444 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2445 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2446 // If this is an absolute destination address, use the munged value.
2447 Callee = SDValue(Dest, 0);
2449 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2450 // to do the call, we can't use PPCISD::CALL.
2451 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2452 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2453 2 + (InFlag.getNode() != 0));
2454 InFlag = Chain.getValue(1);
2457 NodeTys.push_back(MVT::Other);
2458 NodeTys.push_back(MVT::Flag);
2459 Ops.push_back(Chain);
2460 CallOpc = isSVR4ABI ? PPCISD::BCTRL_SVR4 : PPCISD::BCTRL_Darwin;
2462 // Add CTR register as callee so a bctr can be emitted later.
2464 Ops.push_back(DAG.getRegister(PPC::CTR, PtrVT));
2467 // If this is a direct call, pass the chain and the callee.
2468 if (Callee.getNode()) {
2469 Ops.push_back(Chain);
2470 Ops.push_back(Callee);
2472 // If this is a tail call add stack pointer delta.
2474 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2476 // Add argument registers to the end of the list so that they are known live
2478 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2479 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2480 RegsToPass[i].second.getValueType()));
2486 PPCTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2487 CallingConv::ID CallConv, bool isVarArg,
2488 const SmallVectorImpl<ISD::InputArg> &Ins,
2489 DebugLoc dl, SelectionDAG &DAG,
2490 SmallVectorImpl<SDValue> &InVals) {
2492 SmallVector<CCValAssign, 16> RVLocs;
2493 CCState CCRetInfo(CallConv, isVarArg, getTargetMachine(),
2494 RVLocs, *DAG.getContext());
2495 CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2497 // Copy all of the result registers out of their specified physreg.
2498 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2499 CCValAssign &VA = RVLocs[i];
2500 EVT VT = VA.getValVT();
2501 assert(VA.isRegLoc() && "Can only return in registers!");
2502 Chain = DAG.getCopyFromReg(Chain, dl,
2503 VA.getLocReg(), VT, InFlag).getValue(1);
2504 InVals.push_back(Chain.getValue(0));
2505 InFlag = Chain.getValue(2);
2512 PPCTargetLowering::FinishCall(CallingConv::ID CallConv, DebugLoc dl,
2513 bool isTailCall, bool isVarArg,
2515 SmallVector<std::pair<unsigned, SDValue>, 8>
2517 SDValue InFlag, SDValue Chain,
2519 int SPDiff, unsigned NumBytes,
2520 const SmallVectorImpl<ISD::InputArg> &Ins,
2521 SmallVectorImpl<SDValue> &InVals) {
2522 std::vector<EVT> NodeTys;
2523 SmallVector<SDValue, 8> Ops;
2524 unsigned CallOpc = PrepareCall(DAG, Callee, InFlag, Chain, dl, SPDiff,
2525 isTailCall, RegsToPass, Ops, NodeTys,
2526 PPCSubTarget.isSVR4ABI());
2528 // When performing tail call optimization the callee pops its arguments off
2529 // the stack. Account for this here so these bytes can be pushed back on in
2530 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2531 int BytesCalleePops =
2532 (CallConv==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2534 if (InFlag.getNode())
2535 Ops.push_back(InFlag);
2539 // If this is the first return lowered for this function, add the regs
2540 // to the liveout set for the function.
2541 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2542 SmallVector<CCValAssign, 16> RVLocs;
2543 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2545 CCInfo.AnalyzeCallResult(Ins, RetCC_PPC);
2546 for (unsigned i = 0; i != RVLocs.size(); ++i)
2547 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2550 assert(((Callee.getOpcode() == ISD::Register &&
2551 cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||
2552 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2553 Callee.getOpcode() == ISD::TargetGlobalAddress ||
2554 isa<ConstantSDNode>(Callee)) &&
2555 "Expecting an global address, external symbol, absolute value or register");
2557 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Ops[0], Ops.size());
2560 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2561 InFlag = Chain.getValue(1);
2563 // Add a NOP immediately after the branch instruction when using the 64-bit
2564 // SVR4 ABI. At link time, if caller and callee are in a different module and
2565 // thus have a different TOC, the call will be replaced with a call to a stub
2566 // function which saves the current TOC, loads the TOC of the callee and
2567 // branches to the callee. The NOP will be replaced with a load instruction
2568 // which restores the TOC of the caller from the TOC save slot of the current
2569 // stack frame. If caller and callee belong to the same module (and have the
2570 // same TOC), the NOP will remain unchanged.
2571 if (!isTailCall && PPCSubTarget.isSVR4ABI()&& PPCSubTarget.isPPC64()) {
2573 InFlag = DAG.getNode(PPCISD::NOP, dl, MVT::Flag, InFlag);
2576 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2577 DAG.getIntPtrConstant(BytesCalleePops, true),
2580 InFlag = Chain.getValue(1);
2582 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2583 Ins, dl, DAG, InVals);
2587 PPCTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2588 CallingConv::ID CallConv, bool isVarArg,
2590 const SmallVectorImpl<ISD::OutputArg> &Outs,
2591 const SmallVectorImpl<ISD::InputArg> &Ins,
2592 DebugLoc dl, SelectionDAG &DAG,
2593 SmallVectorImpl<SDValue> &InVals) {
2594 if (PPCSubTarget.isSVR4ABI() && !PPCSubTarget.isPPC64()) {
2595 return LowerCall_SVR4(Chain, Callee, CallConv, isVarArg,
2596 isTailCall, Outs, Ins,
2599 return LowerCall_Darwin(Chain, Callee, CallConv, isVarArg,
2600 isTailCall, Outs, Ins,
2606 PPCTargetLowering::LowerCall_SVR4(SDValue Chain, SDValue Callee,
2607 CallingConv::ID CallConv, bool isVarArg,
2609 const SmallVectorImpl<ISD::OutputArg> &Outs,
2610 const SmallVectorImpl<ISD::InputArg> &Ins,
2611 DebugLoc dl, SelectionDAG &DAG,
2612 SmallVectorImpl<SDValue> &InVals) {
2613 // See PPCTargetLowering::LowerFormalArguments_SVR4() for a description
2614 // of the 32-bit SVR4 ABI stack frame layout.
2616 assert((!isTailCall ||
2617 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
2618 "IsEligibleForTailCallOptimization missed a case!");
2620 assert((CallConv == CallingConv::C ||
2621 CallConv == CallingConv::Fast) && "Unknown calling convention!");
2623 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2624 unsigned PtrByteSize = 4;
2626 MachineFunction &MF = DAG.getMachineFunction();
2628 // Mark this function as potentially containing a function that contains a
2629 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2630 // and restoring the callers stack pointer in this functions epilog. This is
2631 // done because by tail calling the called function might overwrite the value
2632 // in this function's (MF) stack pointer stack slot 0(SP).
2633 if (PerformTailCallOpt && CallConv==CallingConv::Fast)
2634 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2636 // Count how many bytes are to be pushed on the stack, including the linkage
2637 // area, parameter list area and the part of the local variable space which
2638 // contains copies of aggregates which are passed by value.
2640 // Assign locations to all of the outgoing arguments.
2641 SmallVector<CCValAssign, 16> ArgLocs;
2642 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
2643 ArgLocs, *DAG.getContext());
2645 // Reserve space for the linkage area on the stack.
2646 CCInfo.AllocateStack(PPCFrameInfo::getLinkageSize(false, false), PtrByteSize);
2649 // Handle fixed and variable vector arguments differently.
2650 // Fixed vector arguments go into registers as long as registers are
2651 // available. Variable vector arguments always go into memory.
2652 unsigned NumArgs = Outs.size();
2654 for (unsigned i = 0; i != NumArgs; ++i) {
2655 EVT ArgVT = Outs[i].Val.getValueType();
2656 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2659 if (Outs[i].IsFixed) {
2660 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2663 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2669 errs() << "Call operand #" << i << " has unhandled type "
2670 << ArgVT.getEVTString() << "\n";
2672 llvm_unreachable(0);
2676 // All arguments are treated the same.
2677 CCInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4);
2680 // Assign locations to all of the outgoing aggregate by value arguments.
2681 SmallVector<CCValAssign, 16> ByValArgLocs;
2682 CCState CCByValInfo(CallConv, isVarArg, getTargetMachine(), ByValArgLocs,
2685 // Reserve stack space for the allocations in CCInfo.
2686 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2688 CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC_SVR4_ByVal);
2690 // Size of the linkage area, parameter list area and the part of the local
2691 // space variable where copies of aggregates which are passed by value are
2693 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2695 // Calculate by how many bytes the stack has to be adjusted in case of tail
2696 // call optimization.
2697 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2699 // Adjust the stack pointer for the new arguments...
2700 // These operations are automatically eliminated by the prolog/epilog pass
2701 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2702 SDValue CallSeqStart = Chain;
2704 // Load the return address and frame pointer so it can be moved somewhere else
2707 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2710 // Set up a copy of the stack pointer for use loading and storing any
2711 // arguments that may not fit in the registers available for argument
2713 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2715 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2716 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2717 SmallVector<SDValue, 8> MemOpChains;
2719 // Walk the register/memloc assignments, inserting copies/loads.
2720 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2723 CCValAssign &VA = ArgLocs[i];
2724 SDValue Arg = Outs[i].Val;
2725 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2727 if (Flags.isByVal()) {
2728 // Argument is an aggregate which is passed by value, thus we need to
2729 // create a copy of it in the local variable space of the current stack
2730 // frame (which is the stack frame of the caller) and pass the address of
2731 // this copy to the callee.
2732 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2733 CCValAssign &ByValVA = ByValArgLocs[j++];
2734 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2736 // Memory reserved in the local variable space of the callers stack frame.
2737 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2739 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2740 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2742 // Create a copy of the argument in the local area of the current
2744 SDValue MemcpyCall =
2745 CreateCopyOfByValArgument(Arg, PtrOff,
2746 CallSeqStart.getNode()->getOperand(0),
2749 // This must go outside the CALLSEQ_START..END.
2750 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2751 CallSeqStart.getNode()->getOperand(1));
2752 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2753 NewCallSeqStart.getNode());
2754 Chain = CallSeqStart = NewCallSeqStart;
2756 // Pass the address of the aggregate copy on the stack either in a
2757 // physical register or in the parameter list area of the current stack
2758 // frame to the callee.
2762 if (VA.isRegLoc()) {
2763 // Put argument in a physical register.
2764 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2766 // Put argument in the parameter list area of the current stack frame.
2767 assert(VA.isMemLoc());
2768 unsigned LocMemOffset = VA.getLocMemOffset();
2771 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2772 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2774 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2775 PseudoSourceValue::getStack(), LocMemOffset));
2777 // Calculate and remember argument location.
2778 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
2784 if (!MemOpChains.empty())
2785 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2786 &MemOpChains[0], MemOpChains.size());
2788 // Build a sequence of copy-to-reg nodes chained together with token chain
2789 // and flag operands which copy the outgoing args into the appropriate regs.
2791 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2792 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2793 RegsToPass[i].second, InFlag);
2794 InFlag = Chain.getValue(1);
2797 // Set CR6 to true if this is a vararg call.
2799 SDValue SetCR(DAG.getMachineNode(PPC::CRSET, dl, MVT::i32), 0);
2800 Chain = DAG.getCopyToReg(Chain, dl, PPC::CR1EQ, SetCR, InFlag);
2801 InFlag = Chain.getValue(1);
2805 PrepareTailCall(DAG, InFlag, Chain, dl, false, SPDiff, NumBytes, LROp, FPOp,
2806 false, TailCallArguments);
2809 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
2810 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
2815 PPCTargetLowering::LowerCall_Darwin(SDValue Chain, SDValue Callee,
2816 CallingConv::ID CallConv, bool isVarArg,
2818 const SmallVectorImpl<ISD::OutputArg> &Outs,
2819 const SmallVectorImpl<ISD::InputArg> &Ins,
2820 DebugLoc dl, SelectionDAG &DAG,
2821 SmallVectorImpl<SDValue> &InVals) {
2823 unsigned NumOps = Outs.size();
2825 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2826 bool isPPC64 = PtrVT == MVT::i64;
2827 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2829 MachineFunction &MF = DAG.getMachineFunction();
2831 // Mark this function as potentially containing a function that contains a
2832 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2833 // and restoring the callers stack pointer in this functions epilog. This is
2834 // done because by tail calling the called function might overwrite the value
2835 // in this function's (MF) stack pointer stack slot 0(SP).
2836 if (PerformTailCallOpt && CallConv==CallingConv::Fast)
2837 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2839 unsigned nAltivecParamsAtEnd = 0;
2841 // Count how many bytes are to be pushed on the stack, including the linkage
2842 // area, and parameter passing area. We start with 24/48 bytes, which is
2843 // prereserved space for [SP][CR][LR][3 x unused].
2845 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isVarArg, CallConv,
2847 nAltivecParamsAtEnd);
2849 // Calculate by how many bytes the stack has to be adjusted in case of tail
2850 // call optimization.
2851 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2853 // To protect arguments on the stack from being clobbered in a tail call,
2854 // force all the loads to happen before doing any other lowering.
2856 Chain = DAG.getStackArgumentTokenFactor(Chain);
2858 // Adjust the stack pointer for the new arguments...
2859 // These operations are automatically eliminated by the prolog/epilog pass
2860 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2861 SDValue CallSeqStart = Chain;
2863 // Load the return address and frame pointer so it can be move somewhere else
2866 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
2869 // Set up a copy of the stack pointer for use loading and storing any
2870 // arguments that may not fit in the registers available for argument
2874 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2876 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2878 // Figure out which arguments are going to go in registers, and which in
2879 // memory. Also, if this is a vararg function, floating point operations
2880 // must be stored to our stack, and loaded into integer regs as well, if
2881 // any integer regs are available for argument passing.
2882 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, true);
2883 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2885 static const unsigned GPR_32[] = { // 32-bit registers.
2886 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2887 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2889 static const unsigned GPR_64[] = { // 64-bit registers.
2890 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2891 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2893 static const unsigned *FPR = GetFPR();
2895 static const unsigned VR[] = {
2896 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2897 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2899 const unsigned NumGPRs = array_lengthof(GPR_32);
2900 const unsigned NumFPRs = 13;
2901 const unsigned NumVRs = array_lengthof(VR);
2903 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2905 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2906 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2908 SmallVector<SDValue, 8> MemOpChains;
2909 for (unsigned i = 0; i != NumOps; ++i) {
2910 SDValue Arg = Outs[i].Val;
2911 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2913 // PtrOff will be used to store the current argument to the stack if a
2914 // register cannot be found for it.
2917 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2919 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
2921 // On PPC64, promote integers to 64-bit values.
2922 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2923 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2924 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2925 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
2928 // FIXME memcpy is used way more than necessary. Correctness first.
2929 if (Flags.isByVal()) {
2930 unsigned Size = Flags.getByValSize();
2931 if (Size==1 || Size==2) {
2932 // Very small objects are passed right-justified.
2933 // Everything else is passed left-justified.
2934 EVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2935 if (GPR_idx != NumGPRs) {
2936 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
2938 MemOpChains.push_back(Load.getValue(1));
2939 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2941 ArgOffset += PtrByteSize;
2943 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2944 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
2945 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2946 CallSeqStart.getNode()->getOperand(0),
2948 // This must go outside the CALLSEQ_START..END.
2949 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2950 CallSeqStart.getNode()->getOperand(1));
2951 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2952 NewCallSeqStart.getNode());
2953 Chain = CallSeqStart = NewCallSeqStart;
2954 ArgOffset += PtrByteSize;
2958 // Copy entire object into memory. There are cases where gcc-generated
2959 // code assumes it is there, even if it could be put entirely into
2960 // registers. (This is not what the doc says.)
2961 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2962 CallSeqStart.getNode()->getOperand(0),
2964 // This must go outside the CALLSEQ_START..END.
2965 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2966 CallSeqStart.getNode()->getOperand(1));
2967 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2968 Chain = CallSeqStart = NewCallSeqStart;
2969 // And copy the pieces of it that fit into registers.
2970 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2971 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2972 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
2973 if (GPR_idx != NumGPRs) {
2974 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, NULL, 0);
2975 MemOpChains.push_back(Load.getValue(1));
2976 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2977 ArgOffset += PtrByteSize;
2979 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2986 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
2987 default: llvm_unreachable("Unexpected ValueType for argument!");
2990 if (GPR_idx != NumGPRs) {
2991 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2993 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2994 isPPC64, isTailCall, false, MemOpChains,
2995 TailCallArguments, dl);
2997 ArgOffset += PtrByteSize;
3001 if (FPR_idx != NumFPRs) {
3002 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
3005 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
3006 MemOpChains.push_back(Store);
3008 // Float varargs are always shadowed in available integer registers
3009 if (GPR_idx != NumGPRs) {
3010 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
3011 MemOpChains.push_back(Load.getValue(1));
3012 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3014 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
3015 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
3016 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
3017 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
3018 MemOpChains.push_back(Load.getValue(1));
3019 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3022 // If we have any FPRs remaining, we may also have GPRs remaining.
3023 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
3025 if (GPR_idx != NumGPRs)
3027 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3028 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3032 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3033 isPPC64, isTailCall, false, MemOpChains,
3034 TailCallArguments, dl);
3039 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3046 // These go aligned on the stack, or in the corresponding R registers
3047 // when within range. The Darwin PPC ABI doc claims they also go in
3048 // V registers; in fact gcc does this only for arguments that are
3049 // prototyped, not for those that match the ... We do it for all
3050 // arguments, seems to work.
3051 while (ArgOffset % 16 !=0) {
3052 ArgOffset += PtrByteSize;
3053 if (GPR_idx != NumGPRs)
3056 // We could elide this store in the case where the object fits
3057 // entirely in R registers. Maybe later.
3058 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3059 DAG.getConstant(ArgOffset, PtrVT));
3060 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
3061 MemOpChains.push_back(Store);
3062 if (VR_idx != NumVRs) {
3063 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, NULL, 0);
3064 MemOpChains.push_back(Load.getValue(1));
3065 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3068 for (unsigned i=0; i<16; i+=PtrByteSize) {
3069 if (GPR_idx == NumGPRs)
3071 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3072 DAG.getConstant(i, PtrVT));
3073 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, NULL, 0);
3074 MemOpChains.push_back(Load.getValue(1));
3075 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3080 // Non-varargs Altivec params generally go in registers, but have
3081 // stack space allocated at the end.
3082 if (VR_idx != NumVRs) {
3083 // Doesn't have GPR space allocated.
3084 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3085 } else if (nAltivecParamsAtEnd==0) {
3086 // We are emitting Altivec params in order.
3087 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3088 isPPC64, isTailCall, true, MemOpChains,
3089 TailCallArguments, dl);
3095 // If all Altivec parameters fit in registers, as they usually do,
3096 // they get stack space following the non-Altivec parameters. We
3097 // don't track this here because nobody below needs it.
3098 // If there are more Altivec parameters than fit in registers emit
3100 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3102 // Offset is aligned; skip 1st 12 params which go in V registers.
3103 ArgOffset = ((ArgOffset+15)/16)*16;
3105 for (unsigned i = 0; i != NumOps; ++i) {
3106 SDValue Arg = Outs[i].Val;
3107 EVT ArgType = Arg.getValueType();
3108 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3109 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3112 // We are emitting Altivec params in order.
3113 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3114 isPPC64, isTailCall, true, MemOpChains,
3115 TailCallArguments, dl);
3122 if (!MemOpChains.empty())
3123 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3124 &MemOpChains[0], MemOpChains.size());
3126 // Build a sequence of copy-to-reg nodes chained together with token chain
3127 // and flag operands which copy the outgoing args into the appropriate regs.
3129 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3130 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3131 RegsToPass[i].second, InFlag);
3132 InFlag = Chain.getValue(1);
3136 PrepareTailCall(DAG, InFlag, Chain, dl, isPPC64, SPDiff, NumBytes, LROp,
3137 FPOp, true, TailCallArguments);
3140 return FinishCall(CallConv, dl, isTailCall, isVarArg, DAG,
3141 RegsToPass, InFlag, Chain, Callee, SPDiff, NumBytes,
3146 PPCTargetLowering::LowerReturn(SDValue Chain,
3147 CallingConv::ID CallConv, bool isVarArg,
3148 const SmallVectorImpl<ISD::OutputArg> &Outs,
3149 DebugLoc dl, SelectionDAG &DAG) {
3151 SmallVector<CCValAssign, 16> RVLocs;
3152 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
3153 RVLocs, *DAG.getContext());
3154 CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
3156 // If this is the first return lowered for this function, add the regs to the
3157 // liveout set for the function.
3158 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3159 for (unsigned i = 0; i != RVLocs.size(); ++i)
3160 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3165 // Copy the result values into the output registers.
3166 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3167 CCValAssign &VA = RVLocs[i];
3168 assert(VA.isRegLoc() && "Can only return in registers!");
3169 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3171 Flag = Chain.getValue(1);
3175 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3177 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3180 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3181 const PPCSubtarget &Subtarget) {
3182 // When we pop the dynamic allocation we need to restore the SP link.
3183 DebugLoc dl = Op.getDebugLoc();
3185 // Get the corect type for pointers.
3186 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3188 // Construct the stack pointer operand.
3189 bool IsPPC64 = Subtarget.isPPC64();
3190 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
3191 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3193 // Get the operands for the STACKRESTORE.
3194 SDValue Chain = Op.getOperand(0);
3195 SDValue SaveSP = Op.getOperand(1);
3197 // Load the old link SP.
3198 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr, NULL, 0);
3200 // Restore the stack pointer.
3201 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3203 // Store the old link SP.
3204 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, NULL, 0);
3210 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3211 MachineFunction &MF = DAG.getMachineFunction();
3212 bool IsPPC64 = PPCSubTarget.isPPC64();
3213 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3214 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3216 // Get current frame pointer save index. The users of this index will be
3217 // primarily DYNALLOC instructions.
3218 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3219 int RASI = FI->getReturnAddrSaveIndex();
3221 // If the frame pointer save index hasn't been defined yet.
3223 // Find out what the fix offset of the frame pointer save area.
3224 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isDarwinABI);
3225 // Allocate the frame index for frame pointer save area.
3226 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset,
3229 FI->setReturnAddrSaveIndex(RASI);
3231 return DAG.getFrameIndex(RASI, PtrVT);
3235 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3236 MachineFunction &MF = DAG.getMachineFunction();
3237 bool IsPPC64 = PPCSubTarget.isPPC64();
3238 bool isDarwinABI = PPCSubTarget.isDarwinABI();
3239 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3241 // Get current frame pointer save index. The users of this index will be
3242 // primarily DYNALLOC instructions.
3243 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3244 int FPSI = FI->getFramePointerSaveIndex();
3246 // If the frame pointer save index hasn't been defined yet.
3248 // Find out what the fix offset of the frame pointer save area.
3249 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64,
3252 // Allocate the frame index for frame pointer save area.
3253 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset,
3256 FI->setFramePointerSaveIndex(FPSI);
3258 return DAG.getFrameIndex(FPSI, PtrVT);
3261 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3263 const PPCSubtarget &Subtarget) {
3265 SDValue Chain = Op.getOperand(0);
3266 SDValue Size = Op.getOperand(1);
3267 DebugLoc dl = Op.getDebugLoc();
3269 // Get the corect type for pointers.
3270 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3272 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3273 DAG.getConstant(0, PtrVT), Size);
3274 // Construct a node for the frame pointer save index.
3275 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3276 // Build a DYNALLOC node.
3277 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3278 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3279 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3282 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3284 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
3285 // Not FP? Not a fsel.
3286 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3287 !Op.getOperand(2).getValueType().isFloatingPoint())
3290 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3292 // Cannot handle SETEQ/SETNE.
3293 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3295 EVT ResVT = Op.getValueType();
3296 EVT CmpVT = Op.getOperand(0).getValueType();
3297 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3298 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3299 DebugLoc dl = Op.getDebugLoc();
3301 // If the RHS of the comparison is a 0.0, we don't need to do the
3302 // subtraction at all.
3303 if (isFloatingPointZero(RHS))
3305 default: break; // SETUO etc aren't handled by fsel.
3308 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3311 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3312 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3313 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3316 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3319 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3320 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3321 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3322 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3327 default: break; // SETUO etc aren't handled by fsel.
3330 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3331 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3332 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3333 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3336 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3337 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3338 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3339 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3342 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3343 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3344 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3345 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3348 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3349 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3350 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3351 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3356 // FIXME: Split this code up when LegalizeDAGTypes lands.
3357 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3359 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3360 SDValue Src = Op.getOperand(0);
3361 if (Src.getValueType() == MVT::f32)
3362 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3365 switch (Op.getValueType().getSimpleVT().SimpleTy) {
3366 default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!");
3368 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3373 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3377 // Convert the FP value to an int value through memory.
3378 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3380 // Emit a store to the stack slot.
3381 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr, NULL, 0);
3383 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3385 if (Op.getValueType() == MVT::i32)
3386 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3387 DAG.getConstant(4, FIPtr.getValueType()));
3388 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, NULL, 0);
3391 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3392 DebugLoc dl = Op.getDebugLoc();
3393 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3394 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3397 if (Op.getOperand(0).getValueType() == MVT::i64) {
3398 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, dl,
3399 MVT::f64, Op.getOperand(0));
3400 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3401 if (Op.getValueType() == MVT::f32)
3402 FP = DAG.getNode(ISD::FP_ROUND, dl,
3403 MVT::f32, FP, DAG.getIntPtrConstant(0));
3407 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3408 "Unhandled SINT_TO_FP type in custom expander!");
3409 // Since we only generate this in 64-bit mode, we can take advantage of
3410 // 64-bit registers. In particular, sign extend the input value into the
3411 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3412 // then lfd it and fcfid it.
3413 MachineFunction &MF = DAG.getMachineFunction();
3414 MachineFrameInfo *FrameInfo = MF.getFrameInfo();
3415 int FrameIdx = FrameInfo->CreateStackObject(8, 8, false);
3416 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3417 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3419 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3422 // STD the extended value into the stack slot.
3423 MachineMemOperand *MMO =
3424 MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FrameIdx),
3425 MachineMemOperand::MOStore, 0, 8, 8);
3426 SDValue Ops[] = { DAG.getEntryNode(), Ext64, FIdx };
3428 DAG.getMemIntrinsicNode(PPCISD::STD_32, dl, DAG.getVTList(MVT::Other),
3429 Ops, 4, MVT::i64, MMO);
3430 // Load the value as a double.
3431 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, NULL, 0);
3433 // FCFID it and return it.
3434 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3435 if (Op.getValueType() == MVT::f32)
3436 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3440 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
3441 DebugLoc dl = Op.getDebugLoc();
3443 The rounding mode is in bits 30:31 of FPSR, and has the following
3450 FLT_ROUNDS, on the other hand, expects the following:
3457 To perform the conversion, we do:
3458 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3461 MachineFunction &MF = DAG.getMachineFunction();
3462 EVT VT = Op.getValueType();
3463 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3464 std::vector<EVT> NodeTys;
3465 SDValue MFFSreg, InFlag;
3467 // Save FP Control Word to register
3468 NodeTys.push_back(MVT::f64); // return register
3469 NodeTys.push_back(MVT::Flag); // unused in this context
3470 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3472 // Save FP register to stack slot
3473 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
3474 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3475 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3476 StackSlot, NULL, 0);
3478 // Load FP Control Word from low 32 bits of stack slot.
3479 SDValue Four = DAG.getConstant(4, PtrVT);
3480 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3481 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, NULL, 0);
3483 // Transform as necessary
3485 DAG.getNode(ISD::AND, dl, MVT::i32,
3486 CWD, DAG.getConstant(3, MVT::i32));
3488 DAG.getNode(ISD::SRL, dl, MVT::i32,
3489 DAG.getNode(ISD::AND, dl, MVT::i32,
3490 DAG.getNode(ISD::XOR, dl, MVT::i32,
3491 CWD, DAG.getConstant(3, MVT::i32)),
3492 DAG.getConstant(3, MVT::i32)),
3493 DAG.getConstant(1, MVT::i32));
3496 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3498 return DAG.getNode((VT.getSizeInBits() < 16 ?
3499 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3502 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3503 EVT VT = Op.getValueType();
3504 unsigned BitWidth = VT.getSizeInBits();
3505 DebugLoc dl = Op.getDebugLoc();
3506 assert(Op.getNumOperands() == 3 &&
3507 VT == Op.getOperand(1).getValueType() &&
3510 // Expand into a bunch of logical ops. Note that these ops
3511 // depend on the PPC behavior for oversized shift amounts.
3512 SDValue Lo = Op.getOperand(0);
3513 SDValue Hi = Op.getOperand(1);
3514 SDValue Amt = Op.getOperand(2);
3515 EVT AmtVT = Amt.getValueType();
3517 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3518 DAG.getConstant(BitWidth, AmtVT), Amt);
3519 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3520 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3521 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3522 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3523 DAG.getConstant(-BitWidth, AmtVT));
3524 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3525 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3526 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3527 SDValue OutOps[] = { OutLo, OutHi };
3528 return DAG.getMergeValues(OutOps, 2, dl);
3531 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3532 EVT VT = Op.getValueType();
3533 DebugLoc dl = Op.getDebugLoc();
3534 unsigned BitWidth = VT.getSizeInBits();
3535 assert(Op.getNumOperands() == 3 &&
3536 VT == Op.getOperand(1).getValueType() &&
3539 // Expand into a bunch of logical ops. Note that these ops
3540 // depend on the PPC behavior for oversized shift amounts.
3541 SDValue Lo = Op.getOperand(0);
3542 SDValue Hi = Op.getOperand(1);
3543 SDValue Amt = Op.getOperand(2);
3544 EVT AmtVT = Amt.getValueType();
3546 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3547 DAG.getConstant(BitWidth, AmtVT), Amt);
3548 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3549 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3550 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3551 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3552 DAG.getConstant(-BitWidth, AmtVT));
3553 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3554 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3555 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3556 SDValue OutOps[] = { OutLo, OutHi };
3557 return DAG.getMergeValues(OutOps, 2, dl);
3560 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3561 DebugLoc dl = Op.getDebugLoc();
3562 EVT VT = Op.getValueType();
3563 unsigned BitWidth = VT.getSizeInBits();
3564 assert(Op.getNumOperands() == 3 &&
3565 VT == Op.getOperand(1).getValueType() &&
3568 // Expand into a bunch of logical ops, followed by a select_cc.
3569 SDValue Lo = Op.getOperand(0);
3570 SDValue Hi = Op.getOperand(1);
3571 SDValue Amt = Op.getOperand(2);
3572 EVT AmtVT = Amt.getValueType();
3574 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3575 DAG.getConstant(BitWidth, AmtVT), Amt);
3576 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3577 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3578 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3579 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3580 DAG.getConstant(-BitWidth, AmtVT));
3581 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3582 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3583 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3584 Tmp4, Tmp6, ISD::SETLE);
3585 SDValue OutOps[] = { OutLo, OutHi };
3586 return DAG.getMergeValues(OutOps, 2, dl);
3589 //===----------------------------------------------------------------------===//
3590 // Vector related lowering.
3593 /// BuildSplatI - Build a canonical splati of Val with an element size of
3594 /// SplatSize. Cast the result to VT.
3595 static SDValue BuildSplatI(int Val, unsigned SplatSize, EVT VT,
3596 SelectionDAG &DAG, DebugLoc dl) {
3597 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3599 static const EVT VTys[] = { // canonical VT to use for each size.
3600 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3603 EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3605 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3609 EVT CanonicalVT = VTys[SplatSize-1];
3611 // Build a canonical splat for this value.
3612 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3613 SmallVector<SDValue, 8> Ops;
3614 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3615 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3616 &Ops[0], Ops.size());
3617 return DAG.getNode(ISD::BIT_CONVERT, dl, ReqVT, Res);
3620 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3621 /// specified intrinsic ID.
3622 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3623 SelectionDAG &DAG, DebugLoc dl,
3624 EVT DestVT = MVT::Other) {
3625 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3626 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3627 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3630 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3631 /// specified intrinsic ID.
3632 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3633 SDValue Op2, SelectionDAG &DAG,
3634 DebugLoc dl, EVT DestVT = MVT::Other) {
3635 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3636 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3637 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3641 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3642 /// amount. The result has the specified value type.
3643 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3644 EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3645 // Force LHS/RHS to be the right type.
3646 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, LHS);
3647 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, RHS);
3650 for (unsigned i = 0; i != 16; ++i)
3652 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3653 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3656 // If this is a case we can't handle, return null and let the default
3657 // expansion code take care of it. If we CAN select this case, and if it
3658 // selects to a single instruction, return Op. Otherwise, if we can codegen
3659 // this case more efficiently than a constant pool load, lower it to the
3660 // sequence of ops that should be used.
3661 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3662 DebugLoc dl = Op.getDebugLoc();
3663 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3664 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3666 // Check if this is a splat of a constant value.
3667 APInt APSplatBits, APSplatUndef;
3668 unsigned SplatBitSize;
3670 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3671 HasAnyUndefs, 0, true) || SplatBitSize > 32)
3674 unsigned SplatBits = APSplatBits.getZExtValue();
3675 unsigned SplatUndef = APSplatUndef.getZExtValue();
3676 unsigned SplatSize = SplatBitSize / 8;
3678 // First, handle single instruction cases.
3681 if (SplatBits == 0) {
3682 // Canonicalize all zero vectors to be v4i32.
3683 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3684 SDValue Z = DAG.getConstant(0, MVT::i32);
3685 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3686 Op = DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Z);
3691 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3692 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3694 if (SextVal >= -16 && SextVal <= 15)
3695 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3698 // Two instruction sequences.
3700 // If this value is in the range [-32,30] and is even, use:
3701 // tmp = VSPLTI[bhw], result = add tmp, tmp
3702 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3703 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3704 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3705 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3708 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3709 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3711 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3712 // Make -1 and vspltisw -1:
3713 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3715 // Make the VSLW intrinsic, computing 0x8000_0000.
3716 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3719 // xor by OnesV to invert it.
3720 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3721 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3724 // Check to see if this is a wide variety of vsplti*, binop self cases.
3725 static const signed char SplatCsts[] = {
3726 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3727 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3730 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3731 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3732 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3733 int i = SplatCsts[idx];
3735 // Figure out what shift amount will be used by altivec if shifted by i in
3737 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3739 // vsplti + shl self.
3740 if (SextVal == (i << (int)TypeShiftAmt)) {
3741 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3742 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3743 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3744 Intrinsic::ppc_altivec_vslw
3746 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3747 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3750 // vsplti + srl self.
3751 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3752 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3753 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3754 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3755 Intrinsic::ppc_altivec_vsrw
3757 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3758 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3761 // vsplti + sra self.
3762 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3763 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3764 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3765 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3766 Intrinsic::ppc_altivec_vsraw
3768 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3769 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3772 // vsplti + rol self.
3773 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3774 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3775 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3776 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3777 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3778 Intrinsic::ppc_altivec_vrlw
3780 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3781 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3784 // t = vsplti c, result = vsldoi t, t, 1
3785 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3786 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3787 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
3789 // t = vsplti c, result = vsldoi t, t, 2
3790 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3791 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3792 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
3794 // t = vsplti c, result = vsldoi t, t, 3
3795 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3796 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3797 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
3801 // Three instruction sequences.
3803 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3804 if (SextVal >= 0 && SextVal <= 31) {
3805 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
3806 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3807 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
3808 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3810 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3811 if (SextVal >= -31 && SextVal <= 0) {
3812 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
3813 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3814 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
3815 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3821 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3822 /// the specified operations to build the shuffle.
3823 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3824 SDValue RHS, SelectionDAG &DAG,
3826 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3827 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3828 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3831 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3843 if (OpNum == OP_COPY) {
3844 if (LHSID == (1*9+2)*9+3) return LHS;
3845 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3849 SDValue OpLHS, OpRHS;
3850 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3851 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3855 default: llvm_unreachable("Unknown i32 permute!");
3857 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3858 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3859 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3860 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3863 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3864 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3865 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3866 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3869 for (unsigned i = 0; i != 16; ++i)
3870 ShufIdxs[i] = (i&3)+0;
3873 for (unsigned i = 0; i != 16; ++i)
3874 ShufIdxs[i] = (i&3)+4;
3877 for (unsigned i = 0; i != 16; ++i)
3878 ShufIdxs[i] = (i&3)+8;
3881 for (unsigned i = 0; i != 16; ++i)
3882 ShufIdxs[i] = (i&3)+12;
3885 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
3887 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
3889 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
3891 EVT VT = OpLHS.getValueType();
3892 OpLHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpLHS);
3893 OpRHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpRHS);
3894 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
3895 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3898 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3899 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3900 /// return the code it can be lowered into. Worst case, it can always be
3901 /// lowered into a vperm.
3902 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
3903 SelectionDAG &DAG) {
3904 DebugLoc dl = Op.getDebugLoc();
3905 SDValue V1 = Op.getOperand(0);
3906 SDValue V2 = Op.getOperand(1);
3907 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
3908 EVT VT = Op.getValueType();
3910 // Cases that are handled by instructions that take permute immediates
3911 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3912 // selected by the instruction selector.
3913 if (V2.getOpcode() == ISD::UNDEF) {
3914 if (PPC::isSplatShuffleMask(SVOp, 1) ||
3915 PPC::isSplatShuffleMask(SVOp, 2) ||
3916 PPC::isSplatShuffleMask(SVOp, 4) ||
3917 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
3918 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
3919 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
3920 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
3921 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
3922 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
3923 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
3924 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
3925 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
3930 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3931 // and produce a fixed permutation. If any of these match, do not lower to
3933 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
3934 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
3935 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
3936 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
3937 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
3938 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
3939 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
3940 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
3941 PPC::isVMRGHShuffleMask(SVOp, 4, false))
3944 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3945 // perfect shuffle table to emit an optimal matching sequence.
3946 SmallVector<int, 16> PermMask;
3947 SVOp->getMask(PermMask);
3949 unsigned PFIndexes[4];
3950 bool isFourElementShuffle = true;
3951 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3952 unsigned EltNo = 8; // Start out undef.
3953 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3954 if (PermMask[i*4+j] < 0)
3955 continue; // Undef, ignore it.
3957 unsigned ByteSource = PermMask[i*4+j];
3958 if ((ByteSource & 3) != j) {
3959 isFourElementShuffle = false;
3964 EltNo = ByteSource/4;
3965 } else if (EltNo != ByteSource/4) {
3966 isFourElementShuffle = false;
3970 PFIndexes[i] = EltNo;
3973 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3974 // perfect shuffle vector to determine if it is cost effective to do this as
3975 // discrete instructions, or whether we should use a vperm.
3976 if (isFourElementShuffle) {
3977 // Compute the index in the perfect shuffle table.
3978 unsigned PFTableIndex =
3979 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3981 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3982 unsigned Cost = (PFEntry >> 30);
3984 // Determining when to avoid vperm is tricky. Many things affect the cost
3985 // of vperm, particularly how many times the perm mask needs to be computed.
3986 // For example, if the perm mask can be hoisted out of a loop or is already
3987 // used (perhaps because there are multiple permutes with the same shuffle
3988 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3989 // the loop requires an extra register.
3991 // As a compromise, we only emit discrete instructions if the shuffle can be
3992 // generated in 3 or fewer operations. When we have loop information
3993 // available, if this block is within a loop, we should avoid using vperm
3994 // for 3-operation perms and use a constant pool load instead.
3996 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
3999 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4000 // vector that will get spilled to the constant pool.
4001 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4003 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4004 // that it is in input element units, not in bytes. Convert now.
4005 EVT EltVT = V1.getValueType().getVectorElementType();
4006 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4008 SmallVector<SDValue, 16> ResultMask;
4009 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4010 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4012 for (unsigned j = 0; j != BytesPerElement; ++j)
4013 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4017 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4018 &ResultMask[0], ResultMask.size());
4019 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4022 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4023 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4024 /// information about the intrinsic.
4025 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4027 unsigned IntrinsicID =
4028 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4031 switch (IntrinsicID) {
4032 default: return false;
4033 // Comparison predicates.
4034 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4035 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4036 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4037 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4038 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4039 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4040 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4041 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4042 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4043 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4044 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4045 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4046 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4048 // Normal Comparisons.
4049 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4050 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4051 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4052 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4053 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4054 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4055 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4056 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4057 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4058 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4059 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4060 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4061 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4066 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4067 /// lower, do it, otherwise return null.
4068 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4069 SelectionDAG &DAG) {
4070 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4071 // opcode number of the comparison.
4072 DebugLoc dl = Op.getDebugLoc();
4075 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4076 return SDValue(); // Don't custom lower most intrinsics.
4078 // If this is a non-dot comparison, make the VCMP node and we are done.
4080 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4081 Op.getOperand(1), Op.getOperand(2),
4082 DAG.getConstant(CompareOpc, MVT::i32));
4083 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Tmp);
4086 // Create the PPCISD altivec 'dot' comparison node.
4088 Op.getOperand(2), // LHS
4089 Op.getOperand(3), // RHS
4090 DAG.getConstant(CompareOpc, MVT::i32)
4092 std::vector<EVT> VTs;
4093 VTs.push_back(Op.getOperand(2).getValueType());
4094 VTs.push_back(MVT::Flag);
4095 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4097 // Now that we have the comparison, emit a copy from the CR to a GPR.
4098 // This is flagged to the above dot comparison.
4099 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4100 DAG.getRegister(PPC::CR6, MVT::i32),
4101 CompNode.getValue(1));
4103 // Unpack the result based on how the target uses it.
4104 unsigned BitNo; // Bit # of CR6.
4105 bool InvertBit; // Invert result?
4106 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4107 default: // Can't happen, don't crash on invalid number though.
4108 case 0: // Return the value of the EQ bit of CR6.
4109 BitNo = 0; InvertBit = false;
4111 case 1: // Return the inverted value of the EQ bit of CR6.
4112 BitNo = 0; InvertBit = true;
4114 case 2: // Return the value of the LT bit of CR6.
4115 BitNo = 2; InvertBit = false;
4117 case 3: // Return the inverted value of the LT bit of CR6.
4118 BitNo = 2; InvertBit = true;
4122 // Shift the bit into the low position.
4123 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4124 DAG.getConstant(8-(3-BitNo), MVT::i32));
4126 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4127 DAG.getConstant(1, MVT::i32));
4129 // If we are supposed to, toggle the bit.
4131 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4132 DAG.getConstant(1, MVT::i32));
4136 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4137 SelectionDAG &DAG) {
4138 DebugLoc dl = Op.getDebugLoc();
4139 // Create a stack slot that is 16-byte aligned.
4140 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4141 int FrameIdx = FrameInfo->CreateStackObject(16, 16, false);
4142 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4143 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4145 // Store the input value into Value#0 of the stack slot.
4146 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4147 Op.getOperand(0), FIdx, NULL, 0);
4149 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, NULL, 0);
4152 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
4153 DebugLoc dl = Op.getDebugLoc();
4154 if (Op.getValueType() == MVT::v4i32) {
4155 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4157 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4158 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4160 SDValue RHSSwap = // = vrlw RHS, 16
4161 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4163 // Shrinkify inputs to v8i16.
4164 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, LHS);
4165 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHS);
4166 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHSSwap);
4168 // Low parts multiplied together, generating 32-bit results (we ignore the
4170 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4171 LHS, RHS, DAG, dl, MVT::v4i32);
4173 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4174 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4175 // Shift the high parts up 16 bits.
4176 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4178 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4179 } else if (Op.getValueType() == MVT::v8i16) {
4180 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4182 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4184 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4185 LHS, RHS, Zero, DAG, dl);
4186 } else if (Op.getValueType() == MVT::v16i8) {
4187 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4189 // Multiply the even 8-bit parts, producing 16-bit sums.
4190 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4191 LHS, RHS, DAG, dl, MVT::v8i16);
4192 EvenParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, EvenParts);
4194 // Multiply the odd 8-bit parts, producing 16-bit sums.
4195 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4196 LHS, RHS, DAG, dl, MVT::v8i16);
4197 OddParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OddParts);
4199 // Merge the results together.
4201 for (unsigned i = 0; i != 8; ++i) {
4203 Ops[i*2+1] = 2*i+1+16;
4205 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4207 llvm_unreachable("Unknown mul to lower!");
4211 /// LowerOperation - Provide custom lowering hooks for some operations.
4213 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
4214 switch (Op.getOpcode()) {
4215 default: llvm_unreachable("Wasn't expecting to be able to lower this!");
4216 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4217 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
4218 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4219 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4220 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4221 case ISD::SETCC: return LowerSETCC(Op, DAG);
4222 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4224 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
4225 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
4228 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
4229 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
4231 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4232 case ISD::DYNAMIC_STACKALLOC:
4233 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4235 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4236 case ISD::FP_TO_UINT:
4237 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4239 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4240 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4242 // Lower 64-bit shifts.
4243 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4244 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4245 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4247 // Vector-related lowering.
4248 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4249 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4250 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4251 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4252 case ISD::MUL: return LowerMUL(Op, DAG);
4254 // Frame & Return address.
4255 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4256 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4261 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4262 SmallVectorImpl<SDValue>&Results,
4263 SelectionDAG &DAG) {
4264 DebugLoc dl = N->getDebugLoc();
4265 switch (N->getOpcode()) {
4267 assert(false && "Do not know how to custom type legalize this operation!");
4269 case ISD::FP_ROUND_INREG: {
4270 assert(N->getValueType(0) == MVT::ppcf128);
4271 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4272 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4273 MVT::f64, N->getOperand(0),
4274 DAG.getIntPtrConstant(0));
4275 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4276 MVT::f64, N->getOperand(0),
4277 DAG.getIntPtrConstant(1));
4279 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4280 // of the long double, and puts FPSCR back the way it was. We do not
4281 // actually model FPSCR.
4282 std::vector<EVT> NodeTys;
4283 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4285 NodeTys.push_back(MVT::f64); // Return register
4286 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
4287 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4288 MFFSreg = Result.getValue(0);
4289 InFlag = Result.getValue(1);
4292 NodeTys.push_back(MVT::Flag); // Returns a flag
4293 Ops[0] = DAG.getConstant(31, MVT::i32);
4295 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4296 InFlag = Result.getValue(0);
4299 NodeTys.push_back(MVT::Flag); // Returns a flag
4300 Ops[0] = DAG.getConstant(30, MVT::i32);
4302 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4303 InFlag = Result.getValue(0);
4306 NodeTys.push_back(MVT::f64); // result of add
4307 NodeTys.push_back(MVT::Flag); // Returns a flag
4311 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4312 FPreg = Result.getValue(0);
4313 InFlag = Result.getValue(1);
4316 NodeTys.push_back(MVT::f64);
4317 Ops[0] = DAG.getConstant(1, MVT::i32);
4321 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4322 FPreg = Result.getValue(0);
4324 // We know the low half is about to be thrown away, so just use something
4326 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4330 case ISD::FP_TO_SINT:
4331 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4337 //===----------------------------------------------------------------------===//
4338 // Other Lowering Code
4339 //===----------------------------------------------------------------------===//
4342 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4343 bool is64bit, unsigned BinOpcode) const {
4344 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4345 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4347 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4348 MachineFunction *F = BB->getParent();
4349 MachineFunction::iterator It = BB;
4352 unsigned dest = MI->getOperand(0).getReg();
4353 unsigned ptrA = MI->getOperand(1).getReg();
4354 unsigned ptrB = MI->getOperand(2).getReg();
4355 unsigned incr = MI->getOperand(3).getReg();
4356 DebugLoc dl = MI->getDebugLoc();
4358 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4359 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4360 F->insert(It, loopMBB);
4361 F->insert(It, exitMBB);
4362 exitMBB->transferSuccessors(BB);
4364 MachineRegisterInfo &RegInfo = F->getRegInfo();
4365 unsigned TmpReg = (!BinOpcode) ? incr :
4366 RegInfo.createVirtualRegister(
4367 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4368 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4372 // fallthrough --> loopMBB
4373 BB->addSuccessor(loopMBB);
4376 // l[wd]arx dest, ptr
4377 // add r0, dest, incr
4378 // st[wd]cx. r0, ptr
4380 // fallthrough --> exitMBB
4382 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4383 .addReg(ptrA).addReg(ptrB);
4385 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4386 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4387 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4388 BuildMI(BB, dl, TII->get(PPC::BCC))
4389 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4390 BB->addSuccessor(loopMBB);
4391 BB->addSuccessor(exitMBB);
4400 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4401 MachineBasicBlock *BB,
4402 bool is8bit, // operation
4403 unsigned BinOpcode) const {
4404 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4405 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4406 // In 64 bit mode we have to use 64 bits for addresses, even though the
4407 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4408 // registers without caring whether they're 32 or 64, but here we're
4409 // doing actual arithmetic on the addresses.
4410 bool is64bit = PPCSubTarget.isPPC64();
4412 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4413 MachineFunction *F = BB->getParent();
4414 MachineFunction::iterator It = BB;
4417 unsigned dest = MI->getOperand(0).getReg();
4418 unsigned ptrA = MI->getOperand(1).getReg();
4419 unsigned ptrB = MI->getOperand(2).getReg();
4420 unsigned incr = MI->getOperand(3).getReg();
4421 DebugLoc dl = MI->getDebugLoc();
4423 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4424 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4425 F->insert(It, loopMBB);
4426 F->insert(It, exitMBB);
4427 exitMBB->transferSuccessors(BB);
4429 MachineRegisterInfo &RegInfo = F->getRegInfo();
4430 const TargetRegisterClass *RC =
4431 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4432 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4433 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4434 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4435 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4436 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4437 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4438 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4439 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4440 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4441 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4442 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4443 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4445 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4449 // fallthrough --> loopMBB
4450 BB->addSuccessor(loopMBB);
4452 // The 4-byte load must be aligned, while a char or short may be
4453 // anywhere in the word. Hence all this nasty bookkeeping code.
4454 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4455 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4456 // xori shift, shift1, 24 [16]
4457 // rlwinm ptr, ptr1, 0, 0, 29
4458 // slw incr2, incr, shift
4459 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4460 // slw mask, mask2, shift
4462 // lwarx tmpDest, ptr
4463 // add tmp, tmpDest, incr2
4464 // andc tmp2, tmpDest, mask
4465 // and tmp3, tmp, mask
4466 // or tmp4, tmp3, tmp2
4469 // fallthrough --> exitMBB
4470 // srw dest, tmpDest, shift
4472 if (ptrA!=PPC::R0) {
4473 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4474 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4475 .addReg(ptrA).addReg(ptrB);
4479 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4480 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4481 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4482 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4484 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4485 .addReg(Ptr1Reg).addImm(0).addImm(61);
4487 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4488 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4489 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4490 .addReg(incr).addReg(ShiftReg);
4492 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4494 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4495 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4497 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4498 .addReg(Mask2Reg).addReg(ShiftReg);
4501 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4502 .addReg(PPC::R0).addReg(PtrReg);
4504 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4505 .addReg(Incr2Reg).addReg(TmpDestReg);
4506 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4507 .addReg(TmpDestReg).addReg(MaskReg);
4508 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4509 .addReg(TmpReg).addReg(MaskReg);
4510 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4511 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4512 BuildMI(BB, dl, TII->get(PPC::STWCX))
4513 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4514 BuildMI(BB, dl, TII->get(PPC::BCC))
4515 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4516 BB->addSuccessor(loopMBB);
4517 BB->addSuccessor(exitMBB);
4522 BuildMI(BB, dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4527 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4528 MachineBasicBlock *BB,
4529 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
4530 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4532 // To "insert" these instructions we actually have to insert their
4533 // control-flow patterns.
4534 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4535 MachineFunction::iterator It = BB;
4538 MachineFunction *F = BB->getParent();
4540 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4541 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4542 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4543 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4544 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4546 // The incoming instruction knows the destination vreg to set, the
4547 // condition code register to branch on, the true/false values to
4548 // select between, and a branch opcode to use.
4553 // cmpTY ccX, r1, r2
4555 // fallthrough --> copy0MBB
4556 MachineBasicBlock *thisMBB = BB;
4557 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4558 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4559 unsigned SelectPred = MI->getOperand(4).getImm();
4560 DebugLoc dl = MI->getDebugLoc();
4561 BuildMI(BB, dl, TII->get(PPC::BCC))
4562 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4563 F->insert(It, copy0MBB);
4564 F->insert(It, sinkMBB);
4565 // Update machine-CFG edges by first adding all successors of the current
4566 // block to the new block which will contain the Phi node for the select.
4567 // Also inform sdisel of the edge changes.
4568 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
4569 E = BB->succ_end(); I != E; ++I) {
4570 EM->insert(std::make_pair(*I, sinkMBB));
4571 sinkMBB->addSuccessor(*I);
4573 // Next, remove all successors of the current block, and add the true
4574 // and fallthrough blocks as its successors.
4575 while (!BB->succ_empty())
4576 BB->removeSuccessor(BB->succ_begin());
4577 // Next, add the true and fallthrough blocks as its successors.
4578 BB->addSuccessor(copy0MBB);
4579 BB->addSuccessor(sinkMBB);
4582 // %FalseValue = ...
4583 // # fallthrough to sinkMBB
4586 // Update machine-CFG edges
4587 BB->addSuccessor(sinkMBB);
4590 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4593 BuildMI(BB, dl, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4594 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4595 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4597 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4598 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4599 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4600 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4601 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4602 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4603 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4604 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4606 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4607 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4608 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4609 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4610 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4611 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4612 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4613 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4615 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4616 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4617 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4618 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4619 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4620 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4621 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4622 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4624 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4625 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4626 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4627 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4628 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4629 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4630 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4631 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4633 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4634 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4635 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4636 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4637 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4638 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4639 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4640 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4642 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4643 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4644 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4645 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4646 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4647 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4648 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4649 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4651 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4652 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4653 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4654 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4655 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4656 BB = EmitAtomicBinary(MI, BB, false, 0);
4657 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4658 BB = EmitAtomicBinary(MI, BB, true, 0);
4660 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4661 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4662 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4664 unsigned dest = MI->getOperand(0).getReg();
4665 unsigned ptrA = MI->getOperand(1).getReg();
4666 unsigned ptrB = MI->getOperand(2).getReg();
4667 unsigned oldval = MI->getOperand(3).getReg();
4668 unsigned newval = MI->getOperand(4).getReg();
4669 DebugLoc dl = MI->getDebugLoc();
4671 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4672 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4673 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4674 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4675 F->insert(It, loop1MBB);
4676 F->insert(It, loop2MBB);
4677 F->insert(It, midMBB);
4678 F->insert(It, exitMBB);
4679 exitMBB->transferSuccessors(BB);
4683 // fallthrough --> loopMBB
4684 BB->addSuccessor(loop1MBB);
4687 // l[wd]arx dest, ptr
4688 // cmp[wd] dest, oldval
4691 // st[wd]cx. newval, ptr
4695 // st[wd]cx. dest, ptr
4698 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4699 .addReg(ptrA).addReg(ptrB);
4700 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4701 .addReg(oldval).addReg(dest);
4702 BuildMI(BB, dl, TII->get(PPC::BCC))
4703 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4704 BB->addSuccessor(loop2MBB);
4705 BB->addSuccessor(midMBB);
4708 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4709 .addReg(newval).addReg(ptrA).addReg(ptrB);
4710 BuildMI(BB, dl, TII->get(PPC::BCC))
4711 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4712 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4713 BB->addSuccessor(loop1MBB);
4714 BB->addSuccessor(exitMBB);
4717 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4718 .addReg(dest).addReg(ptrA).addReg(ptrB);
4719 BB->addSuccessor(exitMBB);
4724 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4725 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4726 // We must use 64-bit registers for addresses when targeting 64-bit,
4727 // since we're actually doing arithmetic on them. Other registers
4729 bool is64bit = PPCSubTarget.isPPC64();
4730 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4732 unsigned dest = MI->getOperand(0).getReg();
4733 unsigned ptrA = MI->getOperand(1).getReg();
4734 unsigned ptrB = MI->getOperand(2).getReg();
4735 unsigned oldval = MI->getOperand(3).getReg();
4736 unsigned newval = MI->getOperand(4).getReg();
4737 DebugLoc dl = MI->getDebugLoc();
4739 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4740 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4741 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4742 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4743 F->insert(It, loop1MBB);
4744 F->insert(It, loop2MBB);
4745 F->insert(It, midMBB);
4746 F->insert(It, exitMBB);
4747 exitMBB->transferSuccessors(BB);
4749 MachineRegisterInfo &RegInfo = F->getRegInfo();
4750 const TargetRegisterClass *RC =
4751 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4752 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4753 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4754 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4755 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4756 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4757 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4758 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4759 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4760 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4761 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4762 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4763 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4764 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4765 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4767 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4770 // fallthrough --> loopMBB
4771 BB->addSuccessor(loop1MBB);
4773 // The 4-byte load must be aligned, while a char or short may be
4774 // anywhere in the word. Hence all this nasty bookkeeping code.
4775 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4776 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4777 // xori shift, shift1, 24 [16]
4778 // rlwinm ptr, ptr1, 0, 0, 29
4779 // slw newval2, newval, shift
4780 // slw oldval2, oldval,shift
4781 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4782 // slw mask, mask2, shift
4783 // and newval3, newval2, mask
4784 // and oldval3, oldval2, mask
4786 // lwarx tmpDest, ptr
4787 // and tmp, tmpDest, mask
4788 // cmpw tmp, oldval3
4791 // andc tmp2, tmpDest, mask
4792 // or tmp4, tmp2, newval3
4797 // stwcx. tmpDest, ptr
4799 // srw dest, tmpDest, shift
4800 if (ptrA!=PPC::R0) {
4801 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4802 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4803 .addReg(ptrA).addReg(ptrB);
4807 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4808 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4809 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4810 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4812 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4813 .addReg(Ptr1Reg).addImm(0).addImm(61);
4815 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4816 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4817 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
4818 .addReg(newval).addReg(ShiftReg);
4819 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
4820 .addReg(oldval).addReg(ShiftReg);
4822 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4824 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4825 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
4826 .addReg(Mask3Reg).addImm(65535);
4828 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4829 .addReg(Mask2Reg).addReg(ShiftReg);
4830 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
4831 .addReg(NewVal2Reg).addReg(MaskReg);
4832 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
4833 .addReg(OldVal2Reg).addReg(MaskReg);
4836 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4837 .addReg(PPC::R0).addReg(PtrReg);
4838 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
4839 .addReg(TmpDestReg).addReg(MaskReg);
4840 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
4841 .addReg(TmpReg).addReg(OldVal3Reg);
4842 BuildMI(BB, dl, TII->get(PPC::BCC))
4843 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4844 BB->addSuccessor(loop2MBB);
4845 BB->addSuccessor(midMBB);
4848 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
4849 .addReg(TmpDestReg).addReg(MaskReg);
4850 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
4851 .addReg(Tmp2Reg).addReg(NewVal3Reg);
4852 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4853 .addReg(PPC::R0).addReg(PtrReg);
4854 BuildMI(BB, dl, TII->get(PPC::BCC))
4855 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4856 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4857 BB->addSuccessor(loop1MBB);
4858 BB->addSuccessor(exitMBB);
4861 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
4862 .addReg(PPC::R0).addReg(PtrReg);
4863 BB->addSuccessor(exitMBB);
4868 BuildMI(BB, dl, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
4870 llvm_unreachable("Unexpected instr type to insert");
4873 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
4877 //===----------------------------------------------------------------------===//
4878 // Target Optimization Hooks
4879 //===----------------------------------------------------------------------===//
4881 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
4882 DAGCombinerInfo &DCI) const {
4883 TargetMachine &TM = getTargetMachine();
4884 SelectionDAG &DAG = DCI.DAG;
4885 DebugLoc dl = N->getDebugLoc();
4886 switch (N->getOpcode()) {
4889 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4890 if (C->getZExtValue() == 0) // 0 << V -> 0.
4891 return N->getOperand(0);
4895 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4896 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
4897 return N->getOperand(0);
4901 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4902 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
4903 C->isAllOnesValue()) // -1 >>s V -> -1.
4904 return N->getOperand(0);
4908 case ISD::SINT_TO_FP:
4909 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
4910 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
4911 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
4912 // We allow the src/dst to be either f32/f64, but the intermediate
4913 // type must be i64.
4914 if (N->getOperand(0).getValueType() == MVT::i64 &&
4915 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
4916 SDValue Val = N->getOperand(0).getOperand(0);
4917 if (Val.getValueType() == MVT::f32) {
4918 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
4919 DCI.AddToWorklist(Val.getNode());
4922 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
4923 DCI.AddToWorklist(Val.getNode());
4924 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
4925 DCI.AddToWorklist(Val.getNode());
4926 if (N->getValueType(0) == MVT::f32) {
4927 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
4928 DAG.getIntPtrConstant(0));
4929 DCI.AddToWorklist(Val.getNode());
4932 } else if (N->getOperand(0).getValueType() == MVT::i32) {
4933 // If the intermediate type is i32, we can avoid the load/store here
4940 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
4941 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
4942 !cast<StoreSDNode>(N)->isTruncatingStore() &&
4943 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
4944 N->getOperand(1).getValueType() == MVT::i32 &&
4945 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
4946 SDValue Val = N->getOperand(1).getOperand(0);
4947 if (Val.getValueType() == MVT::f32) {
4948 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
4949 DCI.AddToWorklist(Val.getNode());
4951 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
4952 DCI.AddToWorklist(Val.getNode());
4954 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
4955 N->getOperand(2), N->getOperand(3));
4956 DCI.AddToWorklist(Val.getNode());
4960 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
4961 if (cast<StoreSDNode>(N)->isUnindexed() &&
4962 N->getOperand(1).getOpcode() == ISD::BSWAP &&
4963 N->getOperand(1).getNode()->hasOneUse() &&
4964 (N->getOperand(1).getValueType() == MVT::i32 ||
4965 N->getOperand(1).getValueType() == MVT::i16)) {
4966 SDValue BSwapOp = N->getOperand(1).getOperand(0);
4967 // Do an any-extend to 32-bits if this is a half-word input.
4968 if (BSwapOp.getValueType() == MVT::i16)
4969 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
4972 N->getOperand(0), BSwapOp, N->getOperand(2),
4973 DAG.getValueType(N->getOperand(1).getValueType())
4976 DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
4977 Ops, array_lengthof(Ops),
4978 cast<StoreSDNode>(N)->getMemoryVT(),
4979 cast<StoreSDNode>(N)->getMemOperand());
4983 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
4984 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
4985 N->getOperand(0).hasOneUse() &&
4986 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
4987 SDValue Load = N->getOperand(0);
4988 LoadSDNode *LD = cast<LoadSDNode>(Load);
4989 // Create the byte-swapping load.
4991 LD->getChain(), // Chain
4992 LD->getBasePtr(), // Ptr
4993 DAG.getValueType(N->getValueType(0)) // VT
4996 DAG.getMemIntrinsicNode(PPCISD::LBRX, dl,
4997 DAG.getVTList(MVT::i32, MVT::Other), Ops, 3,
4998 LD->getMemoryVT(), LD->getMemOperand());
5000 // If this is an i16 load, insert the truncate.
5001 SDValue ResVal = BSLoad;
5002 if (N->getValueType(0) == MVT::i16)
5003 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5005 // First, combine the bswap away. This makes the value produced by the
5007 DCI.CombineTo(N, ResVal);
5009 // Next, combine the load away, we give it a bogus result value but a real
5010 // chain result. The result value is dead because the bswap is dead.
5011 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5013 // Return N so it doesn't get rechecked!
5014 return SDValue(N, 0);
5018 case PPCISD::VCMP: {
5019 // If a VCMPo node already exists with exactly the same operands as this
5020 // node, use its result instead of this node (VCMPo computes both a CR6 and
5021 // a normal output).
5023 if (!N->getOperand(0).hasOneUse() &&
5024 !N->getOperand(1).hasOneUse() &&
5025 !N->getOperand(2).hasOneUse()) {
5027 // Scan all of the users of the LHS, looking for VCMPo's that match.
5028 SDNode *VCMPoNode = 0;
5030 SDNode *LHSN = N->getOperand(0).getNode();
5031 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5033 if (UI->getOpcode() == PPCISD::VCMPo &&
5034 UI->getOperand(1) == N->getOperand(1) &&
5035 UI->getOperand(2) == N->getOperand(2) &&
5036 UI->getOperand(0) == N->getOperand(0)) {
5041 // If there is no VCMPo node, or if the flag value has a single use, don't
5043 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5046 // Look at the (necessarily single) use of the flag value. If it has a
5047 // chain, this transformation is more complex. Note that multiple things
5048 // could use the value result, which we should ignore.
5049 SDNode *FlagUser = 0;
5050 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5051 FlagUser == 0; ++UI) {
5052 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5054 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5055 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5062 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5063 // give up for right now.
5064 if (FlagUser->getOpcode() == PPCISD::MFCR)
5065 return SDValue(VCMPoNode, 0);
5070 // If this is a branch on an altivec predicate comparison, lower this so
5071 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5072 // lowering is done pre-legalize, because the legalizer lowers the predicate
5073 // compare down to code that is difficult to reassemble.
5074 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5075 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5079 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5080 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5081 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5082 assert(isDot && "Can't compare against a vector result!");
5084 // If this is a comparison against something other than 0/1, then we know
5085 // that the condition is never/always true.
5086 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5087 if (Val != 0 && Val != 1) {
5088 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5089 return N->getOperand(0);
5090 // Always !=, turn it into an unconditional branch.
5091 return DAG.getNode(ISD::BR, dl, MVT::Other,
5092 N->getOperand(0), N->getOperand(4));
5095 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5097 // Create the PPCISD altivec 'dot' comparison node.
5098 std::vector<EVT> VTs;
5100 LHS.getOperand(2), // LHS of compare
5101 LHS.getOperand(3), // RHS of compare
5102 DAG.getConstant(CompareOpc, MVT::i32)
5104 VTs.push_back(LHS.getOperand(2).getValueType());
5105 VTs.push_back(MVT::Flag);
5106 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5108 // Unpack the result based on how the target uses it.
5109 PPC::Predicate CompOpc;
5110 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5111 default: // Can't happen, don't crash on invalid number though.
5112 case 0: // Branch on the value of the EQ bit of CR6.
5113 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5115 case 1: // Branch on the inverted value of the EQ bit of CR6.
5116 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5118 case 2: // Branch on the value of the LT bit of CR6.
5119 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5121 case 3: // Branch on the inverted value of the LT bit of CR6.
5122 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5126 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5127 DAG.getConstant(CompOpc, MVT::i32),
5128 DAG.getRegister(PPC::CR6, MVT::i32),
5129 N->getOperand(4), CompNode.getValue(1));
5138 //===----------------------------------------------------------------------===//
5139 // Inline Assembly Support
5140 //===----------------------------------------------------------------------===//
5142 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5146 const SelectionDAG &DAG,
5147 unsigned Depth) const {
5148 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5149 switch (Op.getOpcode()) {
5151 case PPCISD::LBRX: {
5152 // lhbrx is known to have the top bits cleared out.
5153 if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16)
5154 KnownZero = 0xFFFF0000;
5157 case ISD::INTRINSIC_WO_CHAIN: {
5158 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5160 case Intrinsic::ppc_altivec_vcmpbfp_p:
5161 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5162 case Intrinsic::ppc_altivec_vcmpequb_p:
5163 case Intrinsic::ppc_altivec_vcmpequh_p:
5164 case Intrinsic::ppc_altivec_vcmpequw_p:
5165 case Intrinsic::ppc_altivec_vcmpgefp_p:
5166 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5167 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5168 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5169 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5170 case Intrinsic::ppc_altivec_vcmpgtub_p:
5171 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5172 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5173 KnownZero = ~1U; // All bits but the low one are known to be zero.
5181 /// getConstraintType - Given a constraint, return the type of
5182 /// constraint it is for this target.
5183 PPCTargetLowering::ConstraintType
5184 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5185 if (Constraint.size() == 1) {
5186 switch (Constraint[0]) {
5193 return C_RegisterClass;
5196 return TargetLowering::getConstraintType(Constraint);
5199 std::pair<unsigned, const TargetRegisterClass*>
5200 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5202 if (Constraint.size() == 1) {
5203 // GCC RS6000 Constraint Letters
5204 switch (Constraint[0]) {
5207 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5208 return std::make_pair(0U, PPC::G8RCRegisterClass);
5209 return std::make_pair(0U, PPC::GPRCRegisterClass);
5212 return std::make_pair(0U, PPC::F4RCRegisterClass);
5213 else if (VT == MVT::f64)
5214 return std::make_pair(0U, PPC::F8RCRegisterClass);
5217 return std::make_pair(0U, PPC::VRRCRegisterClass);
5219 return std::make_pair(0U, PPC::CRRCRegisterClass);
5223 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5227 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5228 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is true
5229 /// it means one of the asm constraint of the inline asm instruction being
5230 /// processed is 'm'.
5231 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
5233 std::vector<SDValue>&Ops,
5234 SelectionDAG &DAG) const {
5235 SDValue Result(0,0);
5246 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5247 if (!CST) return; // Must be an immediate to match.
5248 unsigned Value = CST->getZExtValue();
5250 default: llvm_unreachable("Unknown constraint letter!");
5251 case 'I': // "I" is a signed 16-bit constant.
5252 if ((short)Value == (int)Value)
5253 Result = DAG.getTargetConstant(Value, Op.getValueType());
5255 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5256 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5257 if ((short)Value == 0)
5258 Result = DAG.getTargetConstant(Value, Op.getValueType());
5260 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5261 if ((Value >> 16) == 0)
5262 Result = DAG.getTargetConstant(Value, Op.getValueType());
5264 case 'M': // "M" is a constant that is greater than 31.
5266 Result = DAG.getTargetConstant(Value, Op.getValueType());
5268 case 'N': // "N" is a positive constant that is an exact power of two.
5269 if ((int)Value > 0 && isPowerOf2_32(Value))
5270 Result = DAG.getTargetConstant(Value, Op.getValueType());
5272 case 'O': // "O" is the constant zero.
5274 Result = DAG.getTargetConstant(Value, Op.getValueType());
5276 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5277 if ((short)-Value == (int)-Value)
5278 Result = DAG.getTargetConstant(Value, Op.getValueType());
5285 if (Result.getNode()) {
5286 Ops.push_back(Result);
5290 // Handle standard constraint letters.
5291 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, hasMemory, Ops, DAG);
5294 // isLegalAddressingMode - Return true if the addressing mode represented
5295 // by AM is legal for this target, for a load/store of the specified type.
5296 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5297 const Type *Ty) const {
5298 // FIXME: PPC does not allow r+i addressing modes for vectors!
5300 // PPC allows a sign-extended 16-bit immediate field.
5301 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5304 // No global is ever allowed as a base.
5308 // PPC only support r+r,
5310 case 0: // "r+i" or just "i", depending on HasBaseReg.
5313 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5315 // Otherwise we have r+r or r+i.
5318 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5320 // Allow 2*r as r+r.
5323 // No other scales are supported.
5330 /// isLegalAddressImmediate - Return true if the integer value can be used
5331 /// as the offset of the target addressing mode for load / store of the
5333 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
5334 // PPC allows a sign-extended 16-bit immediate field.
5335 return (V > -(1 << 16) && V < (1 << 16)-1);
5338 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5342 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5343 DebugLoc dl = Op.getDebugLoc();
5344 // Depths > 0 not supported yet!
5345 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5348 MachineFunction &MF = DAG.getMachineFunction();
5349 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5351 // Just load the return address off the stack.
5352 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5354 // Make sure the function really does not optimize away the store of the RA
5356 FuncInfo->setLRStoreRequired();
5357 return DAG.getLoad(getPointerTy(), dl,
5358 DAG.getEntryNode(), RetAddrFI, NULL, 0);
5361 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5362 DebugLoc dl = Op.getDebugLoc();
5363 // Depths > 0 not supported yet!
5364 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5367 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5368 bool isPPC64 = PtrVT == MVT::i64;
5370 MachineFunction &MF = DAG.getMachineFunction();
5371 MachineFrameInfo *MFI = MF.getFrameInfo();
5372 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
5373 && MFI->getStackSize();
5376 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::X31 : PPC::X1,
5379 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::R31 : PPC::R1,
5384 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5385 // The PowerPC target isn't yet aware of offsets.
5389 EVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
5390 bool isSrcConst, bool isSrcStr,
5391 SelectionDAG &DAG) const {
5392 if (this->PPCSubTarget.isPPC64()) {