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/Support/CommandLine.h"
35 #include "llvm/DerivedTypes.h"
38 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
39 CCValAssign::LocInfo &LocInfo,
40 ISD::ArgFlagsTy &ArgFlags,
42 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
44 CCValAssign::LocInfo &LocInfo,
45 ISD::ArgFlagsTy &ArgFlags,
47 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
49 CCValAssign::LocInfo &LocInfo,
50 ISD::ArgFlagsTy &ArgFlags,
53 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
54 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
57 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
58 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
62 // Use _setjmp/_longjmp instead of setjmp/longjmp.
63 setUseUnderscoreSetJmp(true);
64 setUseUnderscoreLongJmp(true);
66 // Set up the register classes.
67 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
68 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
69 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
71 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
72 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
73 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Expand);
75 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
77 // PowerPC has pre-inc load and store's.
78 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
79 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
80 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
81 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
82 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
83 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
84 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
85 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
86 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
87 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
89 // This is used in the ppcf128->int sequence. Note it has different semantics
90 // from FP_ROUND: that rounds to nearest, this rounds to zero.
91 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
93 // PowerPC has no SREM/UREM instructions
94 setOperationAction(ISD::SREM, MVT::i32, Expand);
95 setOperationAction(ISD::UREM, MVT::i32, Expand);
96 setOperationAction(ISD::SREM, MVT::i64, Expand);
97 setOperationAction(ISD::UREM, MVT::i64, Expand);
99 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
100 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
101 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
102 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
103 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
104 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
105 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
106 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
107 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
109 // We don't support sin/cos/sqrt/fmod/pow
110 setOperationAction(ISD::FSIN , MVT::f64, Expand);
111 setOperationAction(ISD::FCOS , MVT::f64, Expand);
112 setOperationAction(ISD::FREM , MVT::f64, Expand);
113 setOperationAction(ISD::FPOW , MVT::f64, Expand);
114 setOperationAction(ISD::FSIN , MVT::f32, Expand);
115 setOperationAction(ISD::FCOS , MVT::f32, Expand);
116 setOperationAction(ISD::FREM , MVT::f32, Expand);
117 setOperationAction(ISD::FPOW , MVT::f32, Expand);
119 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
121 // If we're enabling GP optimizations, use hardware square root
122 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
123 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
124 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
127 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
128 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
130 // PowerPC does not have BSWAP, CTPOP or CTTZ
131 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
132 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
133 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
134 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
135 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
136 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
138 // PowerPC does not have ROTR
139 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
140 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
142 // PowerPC does not have Select
143 setOperationAction(ISD::SELECT, MVT::i32, Expand);
144 setOperationAction(ISD::SELECT, MVT::i64, Expand);
145 setOperationAction(ISD::SELECT, MVT::f32, Expand);
146 setOperationAction(ISD::SELECT, MVT::f64, Expand);
148 // PowerPC wants to turn select_cc of FP into fsel when possible.
149 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
150 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
152 // PowerPC wants to optimize integer setcc a bit
153 setOperationAction(ISD::SETCC, MVT::i32, Custom);
155 // PowerPC does not have BRCOND which requires SetCC
156 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
158 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
160 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
161 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
163 // PowerPC does not have [U|S]INT_TO_FP
164 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
165 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
167 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
168 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
169 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
170 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
172 // We cannot sextinreg(i1). Expand to shifts.
173 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
175 // Support label based line numbers.
176 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
177 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
179 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
180 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
181 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
182 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
185 // We want to legalize GlobalAddress and ConstantPool nodes into the
186 // appropriate instructions to materialize the address.
187 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
188 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
189 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
190 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
191 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
192 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
193 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
194 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
196 // RET must be custom lowered, to meet ABI requirements.
197 setOperationAction(ISD::RET , MVT::Other, Custom);
200 setOperationAction(ISD::TRAP, MVT::Other, Legal);
202 // TRAMPOLINE is custom lowered.
203 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
205 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
206 setOperationAction(ISD::VASTART , MVT::Other, Custom);
208 // VAARG is custom lowered with ELF 32 ABI
209 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
210 setOperationAction(ISD::VAARG, MVT::Other, Custom);
212 setOperationAction(ISD::VAARG, MVT::Other, Expand);
214 // Use the default implementation.
215 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
216 setOperationAction(ISD::VAEND , MVT::Other, Expand);
217 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
218 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
219 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
220 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
222 // We want to custom lower some of our intrinsics.
223 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
225 // Comparisons that require checking two conditions.
226 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
227 setCondCodeAction(ISD::SETULT, MVT::f64, Expand);
228 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
229 setCondCodeAction(ISD::SETUGT, MVT::f64, Expand);
230 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
231 setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand);
232 setCondCodeAction(ISD::SETOGE, MVT::f32, Expand);
233 setCondCodeAction(ISD::SETOGE, MVT::f64, Expand);
234 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
235 setCondCodeAction(ISD::SETOLE, MVT::f64, Expand);
236 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
237 setCondCodeAction(ISD::SETONE, MVT::f64, Expand);
239 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
240 // They also have instructions for converting between i64 and fp.
241 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
242 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
243 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
244 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
245 // This is just the low 32 bits of a (signed) fp->i64 conversion.
246 // We cannot do this with Promote because i64 is not a legal type.
247 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
249 // FIXME: disable this lowered code. This generates 64-bit register values,
250 // and we don't model the fact that the top part is clobbered by calls. We
251 // need to flag these together so that the value isn't live across a call.
252 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
254 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
255 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
258 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
259 // 64-bit PowerPC implementations can support i64 types directly
260 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
261 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
262 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
263 // 64-bit PowerPC wants to expand i128 shifts itself.
264 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
265 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
266 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
268 // 32-bit PowerPC wants to expand i64 shifts itself.
269 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
270 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
271 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
274 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
275 // First set operation action for all vector types to expand. Then we
276 // will selectively turn on ones that can be effectively codegen'd.
277 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
278 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
279 MVT VT = (MVT::SimpleValueType)i;
281 // add/sub are legal for all supported vector VT's.
282 setOperationAction(ISD::ADD , VT, Legal);
283 setOperationAction(ISD::SUB , VT, Legal);
285 // We promote all shuffles to v16i8.
286 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
287 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
289 // We promote all non-typed operations to v4i32.
290 setOperationAction(ISD::AND , VT, Promote);
291 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
292 setOperationAction(ISD::OR , VT, Promote);
293 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
294 setOperationAction(ISD::XOR , VT, Promote);
295 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
296 setOperationAction(ISD::LOAD , VT, Promote);
297 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
298 setOperationAction(ISD::SELECT, VT, Promote);
299 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
300 setOperationAction(ISD::STORE, VT, Promote);
301 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
303 // No other operations are legal.
304 setOperationAction(ISD::MUL , VT, Expand);
305 setOperationAction(ISD::SDIV, VT, Expand);
306 setOperationAction(ISD::SREM, VT, Expand);
307 setOperationAction(ISD::UDIV, VT, Expand);
308 setOperationAction(ISD::UREM, VT, Expand);
309 setOperationAction(ISD::FDIV, VT, Expand);
310 setOperationAction(ISD::FNEG, VT, Expand);
311 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
312 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
313 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
314 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
315 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
316 setOperationAction(ISD::UDIVREM, VT, Expand);
317 setOperationAction(ISD::SDIVREM, VT, Expand);
318 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
319 setOperationAction(ISD::FPOW, VT, Expand);
320 setOperationAction(ISD::CTPOP, VT, Expand);
321 setOperationAction(ISD::CTLZ, VT, Expand);
322 setOperationAction(ISD::CTTZ, VT, Expand);
325 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
326 // with merges, splats, etc.
327 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
329 setOperationAction(ISD::AND , MVT::v4i32, Legal);
330 setOperationAction(ISD::OR , MVT::v4i32, Legal);
331 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
332 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
333 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
334 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
336 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
337 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
338 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
339 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
341 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
342 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
343 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
344 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
346 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
347 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
349 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
350 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
351 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
352 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
355 setShiftAmountType(MVT::i32);
356 setBooleanContents(ZeroOrOneBooleanContent);
358 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
359 setStackPointerRegisterToSaveRestore(PPC::X1);
360 setExceptionPointerRegister(PPC::X3);
361 setExceptionSelectorRegister(PPC::X4);
363 setStackPointerRegisterToSaveRestore(PPC::R1);
364 setExceptionPointerRegister(PPC::R3);
365 setExceptionSelectorRegister(PPC::R4);
368 // We have target-specific dag combine patterns for the following nodes:
369 setTargetDAGCombine(ISD::SINT_TO_FP);
370 setTargetDAGCombine(ISD::STORE);
371 setTargetDAGCombine(ISD::BR_CC);
372 setTargetDAGCombine(ISD::BSWAP);
374 // Darwin long double math library functions have $LDBL128 appended.
375 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
376 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
377 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
378 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
379 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
380 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
381 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
382 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
383 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
384 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
385 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
388 computeRegisterProperties();
391 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
392 /// function arguments in the caller parameter area.
393 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
394 TargetMachine &TM = getTargetMachine();
395 // Darwin passes everything on 4 byte boundary.
396 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
402 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
405 case PPCISD::FSEL: return "PPCISD::FSEL";
406 case PPCISD::FCFID: return "PPCISD::FCFID";
407 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
408 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
409 case PPCISD::STFIWX: return "PPCISD::STFIWX";
410 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
411 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
412 case PPCISD::VPERM: return "PPCISD::VPERM";
413 case PPCISD::Hi: return "PPCISD::Hi";
414 case PPCISD::Lo: return "PPCISD::Lo";
415 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
416 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
417 case PPCISD::SRL: return "PPCISD::SRL";
418 case PPCISD::SRA: return "PPCISD::SRA";
419 case PPCISD::SHL: return "PPCISD::SHL";
420 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
421 case PPCISD::STD_32: return "PPCISD::STD_32";
422 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
423 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
424 case PPCISD::MTCTR: return "PPCISD::MTCTR";
425 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
426 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
427 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
428 case PPCISD::MFCR: return "PPCISD::MFCR";
429 case PPCISD::VCMP: return "PPCISD::VCMP";
430 case PPCISD::VCMPo: return "PPCISD::VCMPo";
431 case PPCISD::LBRX: return "PPCISD::LBRX";
432 case PPCISD::STBRX: return "PPCISD::STBRX";
433 case PPCISD::LARX: return "PPCISD::LARX";
434 case PPCISD::STCX: return "PPCISD::STCX";
435 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
436 case PPCISD::MFFS: return "PPCISD::MFFS";
437 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
438 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
439 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
440 case PPCISD::MTFSF: return "PPCISD::MTFSF";
441 case PPCISD::TAILCALL: return "PPCISD::TAILCALL";
442 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
446 MVT PPCTargetLowering::getSetCCResultType(MVT VT) const {
450 /// getFunctionAlignment - Return the Log2 alignment of this function.
451 unsigned PPCTargetLowering::getFunctionAlignment(const Function *F) const {
452 if (getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin())
453 return F->hasFnAttr(Attribute::OptimizeForSize) ? 2 : 4;
458 //===----------------------------------------------------------------------===//
459 // Node matching predicates, for use by the tblgen matching code.
460 //===----------------------------------------------------------------------===//
462 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
463 static bool isFloatingPointZero(SDValue Op) {
464 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
465 return CFP->getValueAPF().isZero();
466 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
467 // Maybe this has already been legalized into the constant pool?
468 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
469 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
470 return CFP->getValueAPF().isZero();
475 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
476 /// true if Op is undef or if it matches the specified value.
477 static bool isConstantOrUndef(int Op, int Val) {
478 return Op < 0 || Op == Val;
481 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
482 /// VPKUHUM instruction.
483 bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
485 for (unsigned i = 0; i != 16; ++i)
486 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
489 for (unsigned i = 0; i != 8; ++i)
490 if (!isConstantOrUndef(N->getMaskElt(i), i*2+1) ||
491 !isConstantOrUndef(N->getMaskElt(i+8), i*2+1))
497 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
498 /// VPKUWUM instruction.
499 bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary) {
501 for (unsigned i = 0; i != 16; i += 2)
502 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
503 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
506 for (unsigned i = 0; i != 8; i += 2)
507 if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
508 !isConstantOrUndef(N->getMaskElt(i+1), i*2+3) ||
509 !isConstantOrUndef(N->getMaskElt(i+8), i*2+2) ||
510 !isConstantOrUndef(N->getMaskElt(i+9), i*2+3))
516 /// isVMerge - Common function, used to match vmrg* shuffles.
518 static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize,
519 unsigned LHSStart, unsigned RHSStart) {
520 assert(N->getValueType(0) == MVT::v16i8 &&
521 "PPC only supports shuffles by bytes!");
522 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
523 "Unsupported merge size!");
525 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
526 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
527 if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j),
528 LHSStart+j+i*UnitSize) ||
529 !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j),
530 RHSStart+j+i*UnitSize))
536 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
537 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
538 bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
541 return isVMerge(N, UnitSize, 8, 24);
542 return isVMerge(N, UnitSize, 8, 8);
545 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
546 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
547 bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
550 return isVMerge(N, UnitSize, 0, 16);
551 return isVMerge(N, UnitSize, 0, 0);
555 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
556 /// amount, otherwise return -1.
557 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
558 assert(N->getValueType(0) == MVT::v16i8 &&
559 "PPC only supports shuffles by bytes!");
561 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
563 // Find the first non-undef value in the shuffle mask.
565 for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i)
568 if (i == 16) return -1; // all undef.
570 // Otherwise, check to see if the rest of the elements are consecutively
571 // numbered from this value.
572 unsigned ShiftAmt = SVOp->getMaskElt(i);
573 if (ShiftAmt < i) return -1;
577 // Check the rest of the elements to see if they are consecutive.
578 for (++i; i != 16; ++i)
579 if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
582 // Check the rest of the elements to see if they are consecutive.
583 for (++i; i != 16; ++i)
584 if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
590 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
591 /// specifies a splat of a single element that is suitable for input to
592 /// VSPLTB/VSPLTH/VSPLTW.
593 bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) {
594 assert(N->getValueType(0) == MVT::v16i8 &&
595 (EltSize == 1 || EltSize == 2 || EltSize == 4));
597 // This is a splat operation if each element of the permute is the same, and
598 // if the value doesn't reference the second vector.
599 unsigned ElementBase = N->getMaskElt(0);
601 // FIXME: Handle UNDEF elements too!
602 if (ElementBase >= 16)
605 // Check that the indices are consecutive, in the case of a multi-byte element
606 // splatted with a v16i8 mask.
607 for (unsigned i = 1; i != EltSize; ++i)
608 if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase))
611 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
612 if (N->getMaskElt(i) < 0) continue;
613 for (unsigned j = 0; j != EltSize; ++j)
614 if (N->getMaskElt(i+j) != N->getMaskElt(j))
620 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
622 bool PPC::isAllNegativeZeroVector(SDNode *N) {
623 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
625 APInt APVal, APUndef;
629 if (BV->isConstantSplat(APVal, APUndef, BitSize, HasAnyUndefs, 32))
630 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
631 return CFP->getValueAPF().isNegZero();
636 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
637 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
638 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
639 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
640 assert(isSplatShuffleMask(SVOp, EltSize));
641 return SVOp->getMaskElt(0) / EltSize;
644 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
645 /// by using a vspltis[bhw] instruction of the specified element size, return
646 /// the constant being splatted. The ByteSize field indicates the number of
647 /// bytes of each element [124] -> [bhw].
648 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
651 // If ByteSize of the splat is bigger than the element size of the
652 // build_vector, then we have a case where we are checking for a splat where
653 // multiple elements of the buildvector are folded together into a single
654 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
655 unsigned EltSize = 16/N->getNumOperands();
656 if (EltSize < ByteSize) {
657 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
658 SDValue UniquedVals[4];
659 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
661 // See if all of the elements in the buildvector agree across.
662 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
663 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
664 // If the element isn't a constant, bail fully out.
665 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
668 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
669 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
670 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
671 return SDValue(); // no match.
674 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
675 // either constant or undef values that are identical for each chunk. See
676 // if these chunks can form into a larger vspltis*.
678 // Check to see if all of the leading entries are either 0 or -1. If
679 // neither, then this won't fit into the immediate field.
680 bool LeadingZero = true;
681 bool LeadingOnes = true;
682 for (unsigned i = 0; i != Multiple-1; ++i) {
683 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
685 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
686 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
688 // Finally, check the least significant entry.
690 if (UniquedVals[Multiple-1].getNode() == 0)
691 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
692 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
694 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
697 if (UniquedVals[Multiple-1].getNode() == 0)
698 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
699 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue();
700 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
701 return DAG.getTargetConstant(Val, MVT::i32);
707 // Check to see if this buildvec has a single non-undef value in its elements.
708 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
709 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
710 if (OpVal.getNode() == 0)
711 OpVal = N->getOperand(i);
712 else if (OpVal != N->getOperand(i))
716 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
718 unsigned ValSizeInBytes = EltSize;
720 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
721 Value = CN->getZExtValue();
722 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
723 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
724 Value = FloatToBits(CN->getValueAPF().convertToFloat());
727 // If the splat value is larger than the element value, then we can never do
728 // this splat. The only case that we could fit the replicated bits into our
729 // immediate field for would be zero, and we prefer to use vxor for it.
730 if (ValSizeInBytes < ByteSize) return SDValue();
732 // If the element value is larger than the splat value, cut it in half and
733 // check to see if the two halves are equal. Continue doing this until we
734 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
735 while (ValSizeInBytes > ByteSize) {
736 ValSizeInBytes >>= 1;
738 // If the top half equals the bottom half, we're still ok.
739 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
740 (Value & ((1 << (8*ValSizeInBytes))-1)))
744 // Properly sign extend the value.
745 int ShAmt = (4-ByteSize)*8;
746 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
748 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
749 if (MaskVal == 0) return SDValue();
751 // Finally, if this value fits in a 5 bit sext field, return it
752 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
753 return DAG.getTargetConstant(MaskVal, MVT::i32);
757 //===----------------------------------------------------------------------===//
758 // Addressing Mode Selection
759 //===----------------------------------------------------------------------===//
761 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
762 /// or 64-bit immediate, and if the value can be accurately represented as a
763 /// sign extension from a 16-bit value. If so, this returns true and the
765 static bool isIntS16Immediate(SDNode *N, short &Imm) {
766 if (N->getOpcode() != ISD::Constant)
769 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
770 if (N->getValueType(0) == MVT::i32)
771 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
773 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
775 static bool isIntS16Immediate(SDValue Op, short &Imm) {
776 return isIntS16Immediate(Op.getNode(), Imm);
780 /// SelectAddressRegReg - Given the specified addressed, check to see if it
781 /// can be represented as an indexed [r+r] operation. Returns false if it
782 /// can be more efficiently represented with [r+imm].
783 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
785 SelectionDAG &DAG) const {
787 if (N.getOpcode() == ISD::ADD) {
788 if (isIntS16Immediate(N.getOperand(1), imm))
790 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
793 Base = N.getOperand(0);
794 Index = N.getOperand(1);
796 } else if (N.getOpcode() == ISD::OR) {
797 if (isIntS16Immediate(N.getOperand(1), imm))
798 return false; // r+i can fold it if we can.
800 // If this is an or of disjoint bitfields, we can codegen this as an add
801 // (for better address arithmetic) if the LHS and RHS of the OR are provably
803 APInt LHSKnownZero, LHSKnownOne;
804 APInt RHSKnownZero, RHSKnownOne;
805 DAG.ComputeMaskedBits(N.getOperand(0),
806 APInt::getAllOnesValue(N.getOperand(0)
807 .getValueSizeInBits()),
808 LHSKnownZero, LHSKnownOne);
810 if (LHSKnownZero.getBoolValue()) {
811 DAG.ComputeMaskedBits(N.getOperand(1),
812 APInt::getAllOnesValue(N.getOperand(1)
813 .getValueSizeInBits()),
814 RHSKnownZero, RHSKnownOne);
815 // If all of the bits are known zero on the LHS or RHS, the add won't
817 if (~(LHSKnownZero | RHSKnownZero) == 0) {
818 Base = N.getOperand(0);
819 Index = N.getOperand(1);
828 /// Returns true if the address N can be represented by a base register plus
829 /// a signed 16-bit displacement [r+imm], and if it is not better
830 /// represented as reg+reg.
831 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
833 SelectionDAG &DAG) const {
834 // FIXME dl should come from parent load or store, not from address
835 DebugLoc dl = N.getDebugLoc();
836 // If this can be more profitably realized as r+r, fail.
837 if (SelectAddressRegReg(N, Disp, Base, DAG))
840 if (N.getOpcode() == ISD::ADD) {
842 if (isIntS16Immediate(N.getOperand(1), imm)) {
843 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
844 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
845 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
847 Base = N.getOperand(0);
849 return true; // [r+i]
850 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
851 // Match LOAD (ADD (X, Lo(G))).
852 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
853 && "Cannot handle constant offsets yet!");
854 Disp = N.getOperand(1).getOperand(0); // The global address.
855 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
856 Disp.getOpcode() == ISD::TargetConstantPool ||
857 Disp.getOpcode() == ISD::TargetJumpTable);
858 Base = N.getOperand(0);
859 return true; // [&g+r]
861 } else if (N.getOpcode() == ISD::OR) {
863 if (isIntS16Immediate(N.getOperand(1), imm)) {
864 // If this is an or of disjoint bitfields, we can codegen this as an add
865 // (for better address arithmetic) if the LHS and RHS of the OR are
866 // provably disjoint.
867 APInt LHSKnownZero, LHSKnownOne;
868 DAG.ComputeMaskedBits(N.getOperand(0),
869 APInt::getAllOnesValue(N.getOperand(0)
870 .getValueSizeInBits()),
871 LHSKnownZero, LHSKnownOne);
873 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
874 // If all of the bits are known zero on the LHS or RHS, the add won't
876 Base = N.getOperand(0);
877 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
881 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
882 // Loading from a constant address.
884 // If this address fits entirely in a 16-bit sext immediate field, codegen
887 if (isIntS16Immediate(CN, Imm)) {
888 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
889 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
893 // Handle 32-bit sext immediates with LIS + addr mode.
894 if (CN->getValueType(0) == MVT::i32 ||
895 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
896 int Addr = (int)CN->getZExtValue();
898 // Otherwise, break this down into an LIS + disp.
899 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
901 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
902 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
903 Base = SDValue(DAG.getTargetNode(Opc, dl, CN->getValueType(0), Base), 0);
908 Disp = DAG.getTargetConstant(0, getPointerTy());
909 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
910 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
913 return true; // [r+0]
916 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
917 /// represented as an indexed [r+r] operation.
918 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
920 SelectionDAG &DAG) const {
921 // Check to see if we can easily represent this as an [r+r] address. This
922 // will fail if it thinks that the address is more profitably represented as
923 // reg+imm, e.g. where imm = 0.
924 if (SelectAddressRegReg(N, Base, Index, DAG))
927 // If the operand is an addition, always emit this as [r+r], since this is
928 // better (for code size, and execution, as the memop does the add for free)
929 // than emitting an explicit add.
930 if (N.getOpcode() == ISD::ADD) {
931 Base = N.getOperand(0);
932 Index = N.getOperand(1);
936 // Otherwise, do it the hard way, using R0 as the base register.
937 Base = DAG.getRegister(PPC::R0, N.getValueType());
942 /// SelectAddressRegImmShift - Returns true if the address N can be
943 /// represented by a base register plus a signed 14-bit displacement
944 /// [r+imm*4]. Suitable for use by STD and friends.
945 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
947 SelectionDAG &DAG) const {
948 // FIXME dl should come from the parent load or store, not the address
949 DebugLoc dl = N.getDebugLoc();
950 // If this can be more profitably realized as r+r, fail.
951 if (SelectAddressRegReg(N, Disp, Base, DAG))
954 if (N.getOpcode() == ISD::ADD) {
956 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
957 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
958 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
959 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
961 Base = N.getOperand(0);
963 return true; // [r+i]
964 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
965 // Match LOAD (ADD (X, Lo(G))).
966 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
967 && "Cannot handle constant offsets yet!");
968 Disp = N.getOperand(1).getOperand(0); // The global address.
969 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
970 Disp.getOpcode() == ISD::TargetConstantPool ||
971 Disp.getOpcode() == ISD::TargetJumpTable);
972 Base = N.getOperand(0);
973 return true; // [&g+r]
975 } else if (N.getOpcode() == ISD::OR) {
977 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
978 // If this is an or of disjoint bitfields, we can codegen this as an add
979 // (for better address arithmetic) if the LHS and RHS of the OR are
980 // provably disjoint.
981 APInt LHSKnownZero, LHSKnownOne;
982 DAG.ComputeMaskedBits(N.getOperand(0),
983 APInt::getAllOnesValue(N.getOperand(0)
984 .getValueSizeInBits()),
985 LHSKnownZero, LHSKnownOne);
986 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
987 // If all of the bits are known zero on the LHS or RHS, the add won't
989 Base = N.getOperand(0);
990 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
994 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
995 // Loading from a constant address. Verify low two bits are clear.
996 if ((CN->getZExtValue() & 3) == 0) {
997 // If this address fits entirely in a 14-bit sext immediate field, codegen
1000 if (isIntS16Immediate(CN, Imm)) {
1001 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
1002 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
1006 // Fold the low-part of 32-bit absolute addresses into addr mode.
1007 if (CN->getValueType(0) == MVT::i32 ||
1008 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
1009 int Addr = (int)CN->getZExtValue();
1011 // Otherwise, break this down into an LIS + disp.
1012 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
1013 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
1014 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
1015 Base = SDValue(DAG.getTargetNode(Opc, dl, CN->getValueType(0), Base),0);
1021 Disp = DAG.getTargetConstant(0, getPointerTy());
1022 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1023 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1026 return true; // [r+0]
1030 /// getPreIndexedAddressParts - returns true by value, base pointer and
1031 /// offset pointer and addressing mode by reference if the node's address
1032 /// can be legally represented as pre-indexed load / store address.
1033 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1035 ISD::MemIndexedMode &AM,
1036 SelectionDAG &DAG) const {
1037 // Disabled by default for now.
1038 if (!EnablePPCPreinc) return false;
1042 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1043 Ptr = LD->getBasePtr();
1044 VT = LD->getMemoryVT();
1046 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1048 Ptr = ST->getBasePtr();
1049 VT = ST->getMemoryVT();
1053 // PowerPC doesn't have preinc load/store instructions for vectors.
1057 // TODO: Check reg+reg first.
1059 // LDU/STU use reg+imm*4, others use reg+imm.
1060 if (VT != MVT::i64) {
1062 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1066 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1070 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1071 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1072 // sext i32 to i64 when addr mode is r+i.
1073 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1074 LD->getExtensionType() == ISD::SEXTLOAD &&
1075 isa<ConstantSDNode>(Offset))
1083 //===----------------------------------------------------------------------===//
1084 // LowerOperation implementation
1085 //===----------------------------------------------------------------------===//
1087 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1088 SelectionDAG &DAG) {
1089 MVT PtrVT = Op.getValueType();
1090 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1091 Constant *C = CP->getConstVal();
1092 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1093 SDValue Zero = DAG.getConstant(0, PtrVT);
1094 // FIXME there isn't really any debug info here
1095 DebugLoc dl = Op.getDebugLoc();
1097 const TargetMachine &TM = DAG.getTarget();
1099 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, CPI, Zero);
1100 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, CPI, Zero);
1102 // If this is a non-darwin platform, we don't support non-static relo models
1104 if (TM.getRelocationModel() == Reloc::Static ||
1105 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1106 // Generate non-pic code that has direct accesses to the constant pool.
1107 // The address of the global is just (hi(&g)+lo(&g)).
1108 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1111 if (TM.getRelocationModel() == Reloc::PIC_) {
1112 // With PIC, the first instruction is actually "GR+hi(&G)".
1113 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1114 DAG.getNode(PPCISD::GlobalBaseReg,
1115 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1118 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1122 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1123 MVT PtrVT = Op.getValueType();
1124 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1125 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1126 SDValue Zero = DAG.getConstant(0, PtrVT);
1127 // FIXME there isn't really any debug loc here
1128 DebugLoc dl = Op.getDebugLoc();
1130 const TargetMachine &TM = DAG.getTarget();
1132 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, JTI, Zero);
1133 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, JTI, Zero);
1135 // If this is a non-darwin platform, we don't support non-static relo models
1137 if (TM.getRelocationModel() == Reloc::Static ||
1138 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1139 // Generate non-pic code that has direct accesses to the constant pool.
1140 // The address of the global is just (hi(&g)+lo(&g)).
1141 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1144 if (TM.getRelocationModel() == Reloc::PIC_) {
1145 // With PIC, the first instruction is actually "GR+hi(&G)".
1146 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1147 DAG.getNode(PPCISD::GlobalBaseReg,
1148 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1151 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1155 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1156 SelectionDAG &DAG) {
1157 assert(0 && "TLS not implemented for PPC.");
1158 return SDValue(); // Not reached
1161 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1162 SelectionDAG &DAG) {
1163 MVT PtrVT = Op.getValueType();
1164 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1165 GlobalValue *GV = GSDN->getGlobal();
1166 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1167 SDValue Zero = DAG.getConstant(0, PtrVT);
1168 // FIXME there isn't really any debug info here
1169 DebugLoc dl = GSDN->getDebugLoc();
1171 const TargetMachine &TM = DAG.getTarget();
1173 SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, GA, Zero);
1174 SDValue Lo = DAG.getNode(PPCISD::Lo, dl, PtrVT, GA, Zero);
1176 // If this is a non-darwin platform, we don't support non-static relo models
1178 if (TM.getRelocationModel() == Reloc::Static ||
1179 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1180 // Generate non-pic code that has direct accesses to globals.
1181 // The address of the global is just (hi(&g)+lo(&g)).
1182 return DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1185 if (TM.getRelocationModel() == Reloc::PIC_) {
1186 // With PIC, the first instruction is actually "GR+hi(&G)".
1187 Hi = DAG.getNode(ISD::ADD, dl, PtrVT,
1188 DAG.getNode(PPCISD::GlobalBaseReg,
1189 DebugLoc::getUnknownLoc(), PtrVT), Hi);
1192 Lo = DAG.getNode(ISD::ADD, dl, PtrVT, Hi, Lo);
1194 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1197 // If the global is weak or external, we have to go through the lazy
1199 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Lo, NULL, 0);
1202 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1203 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1204 DebugLoc dl = Op.getDebugLoc();
1206 // If we're comparing for equality to zero, expose the fact that this is
1207 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1208 // fold the new nodes.
1209 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1210 if (C->isNullValue() && CC == ISD::SETEQ) {
1211 MVT VT = Op.getOperand(0).getValueType();
1212 SDValue Zext = Op.getOperand(0);
1213 if (VT.bitsLT(MVT::i32)) {
1215 Zext = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Op.getOperand(0));
1217 unsigned Log2b = Log2_32(VT.getSizeInBits());
1218 SDValue Clz = DAG.getNode(ISD::CTLZ, dl, VT, Zext);
1219 SDValue Scc = DAG.getNode(ISD::SRL, dl, VT, Clz,
1220 DAG.getConstant(Log2b, MVT::i32));
1221 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Scc);
1223 // Leave comparisons against 0 and -1 alone for now, since they're usually
1224 // optimized. FIXME: revisit this when we can custom lower all setcc
1226 if (C->isAllOnesValue() || C->isNullValue())
1230 // If we have an integer seteq/setne, turn it into a compare against zero
1231 // by xor'ing the rhs with the lhs, which is faster than setting a
1232 // condition register, reading it back out, and masking the correct bit. The
1233 // normal approach here uses sub to do this instead of xor. Using xor exposes
1234 // the result to other bit-twiddling opportunities.
1235 MVT LHSVT = Op.getOperand(0).getValueType();
1236 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1237 MVT VT = Op.getValueType();
1238 SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0),
1240 return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, LHSVT), CC);
1245 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1246 int VarArgsFrameIndex,
1247 int VarArgsStackOffset,
1248 unsigned VarArgsNumGPR,
1249 unsigned VarArgsNumFPR,
1250 const PPCSubtarget &Subtarget) {
1252 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1253 return SDValue(); // Not reached
1256 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
1257 SDValue Chain = Op.getOperand(0);
1258 SDValue Trmp = Op.getOperand(1); // trampoline
1259 SDValue FPtr = Op.getOperand(2); // nested function
1260 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1261 DebugLoc dl = Op.getDebugLoc();
1263 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1264 bool isPPC64 = (PtrVT == MVT::i64);
1265 const Type *IntPtrTy =
1266 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType();
1268 TargetLowering::ArgListTy Args;
1269 TargetLowering::ArgListEntry Entry;
1271 Entry.Ty = IntPtrTy;
1272 Entry.Node = Trmp; Args.push_back(Entry);
1274 // TrampSize == (isPPC64 ? 48 : 40);
1275 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1276 isPPC64 ? MVT::i64 : MVT::i32);
1277 Args.push_back(Entry);
1279 Entry.Node = FPtr; Args.push_back(Entry);
1280 Entry.Node = Nest; Args.push_back(Entry);
1282 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1283 std::pair<SDValue, SDValue> CallResult =
1284 LowerCallTo(Chain, Op.getValueType().getTypeForMVT(), false, false,
1285 false, false, 0, CallingConv::C, false,
1286 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1290 { CallResult.first, CallResult.second };
1292 return DAG.getMergeValues(Ops, 2, dl);
1295 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1296 int VarArgsFrameIndex,
1297 int VarArgsStackOffset,
1298 unsigned VarArgsNumGPR,
1299 unsigned VarArgsNumFPR,
1300 const PPCSubtarget &Subtarget) {
1301 DebugLoc dl = Op.getDebugLoc();
1303 if (Subtarget.isMachoABI()) {
1304 // vastart just stores the address of the VarArgsFrameIndex slot into the
1305 // memory location argument.
1306 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1307 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1308 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1309 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
1312 // For ELF 32 ABI we follow the layout of the va_list struct.
1313 // We suppose the given va_list is already allocated.
1316 // char gpr; /* index into the array of 8 GPRs
1317 // * stored in the register save area
1318 // * gpr=0 corresponds to r3,
1319 // * gpr=1 to r4, etc.
1321 // char fpr; /* index into the array of 8 FPRs
1322 // * stored in the register save area
1323 // * fpr=0 corresponds to f1,
1324 // * fpr=1 to f2, etc.
1326 // char *overflow_arg_area;
1327 // /* location on stack that holds
1328 // * the next overflow argument
1330 // char *reg_save_area;
1331 // /* where r3:r10 and f1:f8 (if saved)
1337 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i32);
1338 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i32);
1341 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1343 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1344 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1346 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1347 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1349 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1350 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1352 uint64_t FPROffset = 1;
1353 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1355 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1357 // Store first byte : number of int regs
1358 SDValue firstStore = DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR,
1359 Op.getOperand(1), SV, 0, MVT::i8);
1360 uint64_t nextOffset = FPROffset;
1361 SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1),
1364 // Store second byte : number of float regs
1365 SDValue secondStore =
1366 DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr, SV, nextOffset, MVT::i8);
1367 nextOffset += StackOffset;
1368 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset);
1370 // Store second word : arguments given on stack
1371 SDValue thirdStore =
1372 DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr, SV, nextOffset);
1373 nextOffset += FrameOffset;
1374 nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset);
1376 // Store third word : arguments given in registers
1377 return DAG.getStore(thirdStore, dl, FR, nextPtr, SV, nextOffset);
1381 #include "PPCGenCallingConv.inc"
1383 static bool CC_PPC_SVR4_Custom_Dummy(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
1384 CCValAssign::LocInfo &LocInfo,
1385 ISD::ArgFlagsTy &ArgFlags,
1390 static bool CC_PPC_SVR4_Custom_AlignArgRegs(unsigned &ValNo, MVT &ValVT,
1392 CCValAssign::LocInfo &LocInfo,
1393 ISD::ArgFlagsTy &ArgFlags,
1395 static const unsigned ArgRegs[] = {
1396 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1397 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1399 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1401 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1403 // Skip one register if the first unallocated register has an even register
1404 // number and there are still argument registers available which have not been
1405 // allocated yet. RegNum is actually an index into ArgRegs, which means we
1406 // need to skip a register if RegNum is odd.
1407 if (RegNum != NumArgRegs && RegNum % 2 == 1) {
1408 State.AllocateReg(ArgRegs[RegNum]);
1411 // Always return false here, as this function only makes sure that the first
1412 // unallocated register has an odd register number and does not actually
1413 // allocate a register for the current argument.
1417 static bool CC_PPC_SVR4_Custom_AlignFPArgRegs(unsigned &ValNo, MVT &ValVT,
1419 CCValAssign::LocInfo &LocInfo,
1420 ISD::ArgFlagsTy &ArgFlags,
1422 static const unsigned ArgRegs[] = {
1423 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1427 const unsigned NumArgRegs = array_lengthof(ArgRegs);
1429 unsigned RegNum = State.getFirstUnallocated(ArgRegs, NumArgRegs);
1431 // If there is only one Floating-point register left we need to put both f64
1432 // values of a split ppc_fp128 value on the stack.
1433 if (RegNum != NumArgRegs && ArgRegs[RegNum] == PPC::F8) {
1434 State.AllocateReg(ArgRegs[RegNum]);
1437 // Always return false here, as this function only makes sure that the two f64
1438 // values a ppc_fp128 value is split into are both passed in registers or both
1439 // passed on the stack and does not actually allocate a register for the
1440 // current argument.
1444 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1445 /// depending on which subtarget is selected.
1446 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1447 if (Subtarget.isMachoABI()) {
1448 static const unsigned FPR[] = {
1449 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1450 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1456 static const unsigned FPR[] = {
1457 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1463 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1465 static unsigned CalculateStackSlotSize(SDValue Arg, ISD::ArgFlagsTy Flags,
1466 unsigned PtrByteSize) {
1467 MVT ArgVT = Arg.getValueType();
1468 unsigned ArgSize = ArgVT.getSizeInBits()/8;
1469 if (Flags.isByVal())
1470 ArgSize = Flags.getByValSize();
1471 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1477 PPCTargetLowering::LowerFORMAL_ARGUMENTS_SVR4(SDValue Op,
1479 int &VarArgsFrameIndex,
1480 int &VarArgsStackOffset,
1481 unsigned &VarArgsNumGPR,
1482 unsigned &VarArgsNumFPR,
1483 const PPCSubtarget &Subtarget) {
1484 // SVR4 ABI Stack Frame Layout:
1485 // +-----------------------------------+
1486 // +--> | Back chain |
1487 // | +-----------------------------------+
1488 // | | Floating-point register save area |
1489 // | +-----------------------------------+
1490 // | | General register save area |
1491 // | +-----------------------------------+
1492 // | | CR save word |
1493 // | +-----------------------------------+
1494 // | | VRSAVE save word |
1495 // | +-----------------------------------+
1496 // | | Alignment padding |
1497 // | +-----------------------------------+
1498 // | | Vector register save area |
1499 // | +-----------------------------------+
1500 // | | Local variable space |
1501 // | +-----------------------------------+
1502 // | | Parameter list area |
1503 // | +-----------------------------------+
1504 // | | LR save word |
1505 // | +-----------------------------------+
1506 // SP--> +--- | Back chain |
1507 // +-----------------------------------+
1510 // System V Application Binary Interface PowerPC Processor Supplement
1511 // AltiVec Technology Programming Interface Manual
1513 MachineFunction &MF = DAG.getMachineFunction();
1514 MachineFrameInfo *MFI = MF.getFrameInfo();
1515 SmallVector<SDValue, 8> ArgValues;
1516 SDValue Root = Op.getOperand(0);
1517 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1518 DebugLoc dl = Op.getDebugLoc();
1520 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1521 // Potential tail calls could cause overwriting of argument stack slots.
1522 unsigned CC = MF.getFunction()->getCallingConv();
1523 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1524 unsigned PtrByteSize = 4;
1526 // Assign locations to all of the incoming arguments.
1527 SmallVector<CCValAssign, 16> ArgLocs;
1528 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
1530 // Reserve space for the linkage area on the stack.
1531 CCInfo.AllocateStack(PPCFrameInfo::getLinkageSize(false, false), PtrByteSize);
1533 CCInfo.AnalyzeFormalArguments(Op.getNode(), CC_PPC_SVR4);
1535 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1536 CCValAssign &VA = ArgLocs[i];
1538 // Arguments stored in registers.
1539 if (VA.isRegLoc()) {
1540 TargetRegisterClass *RC;
1541 MVT ValVT = VA.getValVT();
1543 switch (ValVT.getSimpleVT()) {
1545 assert(0 && "ValVT not supported by FORMAL_ARGUMENTS Lowering");
1547 RC = PPC::GPRCRegisterClass;
1550 RC = PPC::F4RCRegisterClass;
1553 RC = PPC::F8RCRegisterClass;
1559 RC = PPC::VRRCRegisterClass;
1563 // Transform the arguments stored in physical registers into virtual ones.
1564 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1565 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, ValVT);
1567 ArgValues.push_back(ArgValue);
1569 // Argument stored in memory.
1570 assert(VA.isMemLoc());
1572 unsigned ArgSize = VA.getLocVT().getSizeInBits() / 8;
1573 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
1576 // Create load nodes to retrieve arguments from the stack.
1577 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1578 ArgValues.push_back(DAG.getLoad(VA.getValVT(), dl, Root, FIN, NULL, 0));
1582 // Assign locations to all of the incoming aggregate by value arguments.
1583 // Aggregates passed by value are stored in the local variable space of the
1584 // caller's stack frame, right above the parameter list area.
1585 SmallVector<CCValAssign, 16> ByValArgLocs;
1586 CCState CCByValInfo(CC, isVarArg, getTargetMachine(), ByValArgLocs);
1588 // Reserve stack space for the allocations in CCInfo.
1589 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
1591 CCByValInfo.AnalyzeFormalArguments(Op.getNode(), CC_PPC_SVR4_ByVal);
1593 // Area that is at least reserved in the caller of this function.
1594 unsigned MinReservedArea = CCByValInfo.getNextStackOffset();
1596 // Set the size that is at least reserved in caller of this function. Tail
1597 // call optimized function's reserved stack space needs to be aligned so that
1598 // taking the difference between two stack areas will result in an aligned
1600 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1603 std::max(MinReservedArea,
1604 PPCFrameInfo::getMinCallFrameSize(false, false));
1606 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1607 getStackAlignment();
1608 unsigned AlignMask = TargetAlign-1;
1609 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1611 FI->setMinReservedArea(MinReservedArea);
1613 SmallVector<SDValue, 8> MemOps;
1615 // If the function takes variable number of arguments, make a frame index for
1616 // the start of the first vararg value... for expansion of llvm.va_start.
1618 static const unsigned GPArgRegs[] = {
1619 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1620 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1622 const unsigned NumGPArgRegs = array_lengthof(GPArgRegs);
1624 static const unsigned FPArgRegs[] = {
1625 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1628 const unsigned NumFPArgRegs = array_lengthof(FPArgRegs);
1630 VarArgsNumGPR = CCInfo.getFirstUnallocated(GPArgRegs, NumGPArgRegs);
1631 VarArgsNumFPR = CCInfo.getFirstUnallocated(FPArgRegs, NumFPArgRegs);
1633 // Make room for NumGPArgRegs and NumFPArgRegs.
1634 int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 +
1635 NumFPArgRegs * MVT(MVT::f64).getSizeInBits()/8;
1637 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1638 CCInfo.getNextStackOffset());
1640 VarArgsFrameIndex = MFI->CreateStackObject(Depth, 8);
1641 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1643 // The fixed integer arguments of a variadic function are
1644 // stored to the VarArgsFrameIndex on the stack.
1645 unsigned GPRIndex = 0;
1646 for (; GPRIndex != VarArgsNumGPR; ++GPRIndex) {
1647 SDValue Val = DAG.getRegister(GPArgRegs[GPRIndex], PtrVT);
1648 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
1649 MemOps.push_back(Store);
1650 // Increment the address by four for the next argument to store
1651 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1652 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1655 // If this function is vararg, store any remaining integer argument regs
1656 // to their spots on the stack so that they may be loaded by deferencing the
1657 // result of va_next.
1658 for (; GPRIndex != NumGPArgRegs; ++GPRIndex) {
1659 unsigned VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass);
1661 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1662 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1663 MemOps.push_back(Store);
1664 // Increment the address by four for the next argument to store
1665 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1666 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1669 // FIXME SVR4: We only need to save FP argument registers if CR bit 6 is
1672 // The double arguments are stored to the VarArgsFrameIndex
1674 unsigned FPRIndex = 0;
1675 for (FPRIndex = 0; FPRIndex != VarArgsNumFPR; ++FPRIndex) {
1676 SDValue Val = DAG.getRegister(FPArgRegs[FPRIndex], MVT::f64);
1677 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
1678 MemOps.push_back(Store);
1679 // Increment the address by eight for the next argument to store
1680 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1682 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1685 for (; FPRIndex != NumFPArgRegs; ++FPRIndex) {
1686 unsigned VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass);
1688 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::f64);
1689 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1690 MemOps.push_back(Store);
1691 // Increment the address by eight for the next argument to store
1692 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1694 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
1698 if (!MemOps.empty())
1699 Root = DAG.getNode(ISD::TokenFactor, dl,
1700 MVT::Other, &MemOps[0], MemOps.size());
1703 ArgValues.push_back(Root);
1705 // Return the new list of results.
1706 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1707 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1711 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op,
1713 int &VarArgsFrameIndex,
1714 int &VarArgsStackOffset,
1715 unsigned &VarArgsNumGPR,
1716 unsigned &VarArgsNumFPR,
1717 const PPCSubtarget &Subtarget) {
1718 // TODO: add description of PPC stack frame format, or at least some docs.
1720 MachineFunction &MF = DAG.getMachineFunction();
1721 MachineFrameInfo *MFI = MF.getFrameInfo();
1722 SmallVector<SDValue, 8> ArgValues;
1723 SDValue Root = Op.getOperand(0);
1724 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1725 DebugLoc dl = Op.getDebugLoc();
1727 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1728 bool isPPC64 = PtrVT == MVT::i64;
1729 bool isMachoABI = Subtarget.isMachoABI();
1730 bool isELF32_ABI = Subtarget.isELF32_ABI();
1731 // Potential tail calls could cause overwriting of argument stack slots.
1732 unsigned CC = MF.getFunction()->getCallingConv();
1733 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1734 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1736 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1737 // Area that is at least reserved in caller of this function.
1738 unsigned MinReservedArea = ArgOffset;
1740 static const unsigned GPR_32[] = { // 32-bit registers.
1741 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1742 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1744 static const unsigned GPR_64[] = { // 64-bit registers.
1745 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1746 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1749 static const unsigned *FPR = GetFPR(Subtarget);
1751 static const unsigned VR[] = {
1752 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1753 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1756 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1757 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1758 const unsigned Num_VR_Regs = array_lengthof( VR);
1760 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1762 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1764 // In 32-bit non-varargs functions, the stack space for vectors is after the
1765 // stack space for non-vectors. We do not use this space unless we have
1766 // too many vectors to fit in registers, something that only occurs in
1767 // constructed examples:), but we have to walk the arglist to figure
1768 // that out...for the pathological case, compute VecArgOffset as the
1769 // start of the vector parameter area. Computing VecArgOffset is the
1770 // entire point of the following loop.
1771 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1772 // to handle Elf here.
1773 unsigned VecArgOffset = ArgOffset;
1774 if (!isVarArg && !isPPC64) {
1775 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e;
1777 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1778 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1779 ISD::ArgFlagsTy Flags =
1780 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1782 if (Flags.isByVal()) {
1783 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1784 ObjSize = Flags.getByValSize();
1786 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1787 VecArgOffset += ArgSize;
1791 switch(ObjectVT.getSimpleVT()) {
1792 default: assert(0 && "Unhandled argument type!");
1795 VecArgOffset += isPPC64 ? 8 : 4;
1797 case MVT::i64: // PPC64
1805 // Nothing to do, we're only looking at Nonvector args here.
1810 // We've found where the vector parameter area in memory is. Skip the
1811 // first 12 parameters; these don't use that memory.
1812 VecArgOffset = ((VecArgOffset+15)/16)*16;
1813 VecArgOffset += 12*16;
1815 // Add DAG nodes to load the arguments or copy them out of registers. On
1816 // entry to a function on PPC, the arguments start after the linkage area,
1817 // although the first ones are often in registers.
1819 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1820 // represented with two words (long long or double) must be copied to an
1821 // even GPR_idx value or to an even ArgOffset value.
1823 SmallVector<SDValue, 8> MemOps;
1824 unsigned nAltivecParamsAtEnd = 0;
1825 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues() - 1;
1826 ArgNo != e; ++ArgNo) {
1828 bool needsLoad = false;
1829 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1830 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1831 unsigned ArgSize = ObjSize;
1832 ISD::ArgFlagsTy Flags =
1833 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1834 // See if next argument requires stack alignment in ELF
1835 bool Align = Flags.isSplit();
1837 unsigned CurArgOffset = ArgOffset;
1839 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1840 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1841 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1842 if (isVarArg || isPPC64) {
1843 MinReservedArea = ((MinReservedArea+15)/16)*16;
1844 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1847 } else nAltivecParamsAtEnd++;
1849 // Calculate min reserved area.
1850 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1854 // FIXME alignment for ELF may not be right
1855 // FIXME the codegen can be much improved in some cases.
1856 // We do not have to keep everything in memory.
1857 if (Flags.isByVal()) {
1858 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1859 ObjSize = Flags.getByValSize();
1860 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1861 // Double word align in ELF
1862 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1863 // Objects of size 1 and 2 are right justified, everything else is
1864 // left justified. This means the memory address is adjusted forwards.
1865 if (ObjSize==1 || ObjSize==2) {
1866 CurArgOffset = CurArgOffset + (4 - ObjSize);
1868 // The value of the object is its address.
1869 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1870 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1871 ArgValues.push_back(FIN);
1872 if (ObjSize==1 || ObjSize==2) {
1873 if (GPR_idx != Num_GPR_Regs) {
1874 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1875 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1876 SDValue Store = DAG.getTruncStore(Val.getValue(1), dl, Val, FIN,
1877 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1878 MemOps.push_back(Store);
1880 if (isMachoABI) ArgOffset += PtrByteSize;
1882 ArgOffset += PtrByteSize;
1886 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1887 // Store whatever pieces of the object are in registers
1888 // to memory. ArgVal will be address of the beginning of
1890 if (GPR_idx != Num_GPR_Regs) {
1891 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1892 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1893 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1894 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
1895 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
1896 MemOps.push_back(Store);
1898 if (isMachoABI) ArgOffset += PtrByteSize;
1900 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1907 switch (ObjectVT.getSimpleVT()) {
1908 default: assert(0 && "Unhandled argument type!");
1911 // Double word align in ELF
1912 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1914 if (GPR_idx != Num_GPR_Regs) {
1915 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
1916 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32);
1920 ArgSize = PtrByteSize;
1922 // Stack align in ELF
1923 if (needsLoad && Align && isELF32_ABI)
1924 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1925 // All int arguments reserve stack space in Macho ABI.
1926 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1930 case MVT::i64: // PPC64
1931 if (GPR_idx != Num_GPR_Regs) {
1932 unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
1933 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1935 if (ObjectVT == MVT::i32) {
1936 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1937 // value to MVT::i64 and then truncate to the correct register size.
1939 ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal,
1940 DAG.getValueType(ObjectVT));
1941 else if (Flags.isZExt())
1942 ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal,
1943 DAG.getValueType(ObjectVT));
1945 ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
1951 ArgSize = PtrByteSize;
1953 // All int arguments reserve stack space in Macho ABI.
1954 if (isMachoABI || needsLoad) ArgOffset += 8;
1959 // Every 4 bytes of argument space consumes one of the GPRs available for
1960 // argument passing.
1961 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1963 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1966 if (FPR_idx != Num_FPR_Regs) {
1969 if (ObjectVT == MVT::f32)
1970 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass);
1972 VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
1974 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, ObjectVT);
1980 // Stack align in ELF
1981 if (needsLoad && Align && isELF32_ABI)
1982 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1983 // All FP arguments reserve stack space in Macho ABI.
1984 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1990 // Note that vector arguments in registers don't reserve stack space,
1991 // except in varargs functions.
1992 if (VR_idx != Num_VR_Regs) {
1993 unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass);
1994 ArgVal = DAG.getCopyFromReg(Root, dl, VReg, ObjectVT);
1996 while ((ArgOffset % 16) != 0) {
1997 ArgOffset += PtrByteSize;
1998 if (GPR_idx != Num_GPR_Regs)
2002 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
2006 if (!isVarArg && !isPPC64) {
2007 // Vectors go after all the nonvectors.
2008 CurArgOffset = VecArgOffset;
2011 // Vectors are aligned.
2012 ArgOffset = ((ArgOffset+15)/16)*16;
2013 CurArgOffset = ArgOffset;
2021 // We need to load the argument to a virtual register if we determined above
2022 // that we ran out of physical registers of the appropriate type.
2024 int FI = MFI->CreateFixedObject(ObjSize,
2025 CurArgOffset + (ArgSize - ObjSize),
2027 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
2028 ArgVal = DAG.getLoad(ObjectVT, dl, Root, FIN, NULL, 0);
2031 ArgValues.push_back(ArgVal);
2034 // Set the size that is at least reserved in caller of this function. Tail
2035 // call optimized function's reserved stack space needs to be aligned so that
2036 // taking the difference between two stack areas will result in an aligned
2038 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2039 // Add the Altivec parameters at the end, if needed.
2040 if (nAltivecParamsAtEnd) {
2041 MinReservedArea = ((MinReservedArea+15)/16)*16;
2042 MinReservedArea += 16*nAltivecParamsAtEnd;
2045 std::max(MinReservedArea,
2046 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
2047 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
2048 getStackAlignment();
2049 unsigned AlignMask = TargetAlign-1;
2050 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
2051 FI->setMinReservedArea(MinReservedArea);
2053 // If the function takes variable number of arguments, make a frame index for
2054 // the start of the first vararg value... for expansion of llvm.va_start.
2059 VarArgsNumGPR = GPR_idx;
2060 VarArgsNumFPR = FPR_idx;
2062 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
2064 depth = -(Num_GPR_Regs * PtrVT.getSizeInBits()/8 +
2065 Num_FPR_Regs * MVT(MVT::f64).getSizeInBits()/8 +
2066 PtrVT.getSizeInBits()/8);
2068 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2075 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
2077 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
2079 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
2080 // stored to the VarArgsFrameIndex on the stack.
2082 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
2083 SDValue Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
2084 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
2085 MemOps.push_back(Store);
2086 // Increment the address by four for the next argument to store
2087 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2088 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2092 // If this function is vararg, store any remaining integer argument regs
2093 // to their spots on the stack so that they may be loaded by deferencing the
2094 // result of va_next.
2095 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
2099 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
2101 VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass);
2103 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, PtrVT);
2104 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
2105 MemOps.push_back(Store);
2106 // Increment the address by four for the next argument to store
2107 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
2108 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2111 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
2114 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
2115 SDValue Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
2116 SDValue Store = DAG.getStore(Root, dl, Val, FIN, NULL, 0);
2117 MemOps.push_back(Store);
2118 // Increment the address by eight for the next argument to store
2119 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
2121 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2124 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
2125 unsigned VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass);
2127 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::f64);
2128 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
2129 MemOps.push_back(Store);
2130 // Increment the address by eight for the next argument to store
2131 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
2133 FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff);
2138 if (!MemOps.empty())
2139 Root = DAG.getNode(ISD::TokenFactor, dl,
2140 MVT::Other, &MemOps[0], MemOps.size());
2142 ArgValues.push_back(Root);
2144 // Return the new list of results.
2145 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
2146 &ArgValues[0], ArgValues.size());
2149 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
2152 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
2157 CallSDNode *TheCall,
2158 unsigned &nAltivecParamsAtEnd) {
2159 // Count how many bytes are to be pushed on the stack, including the linkage
2160 // area, and parameter passing area. We start with 24/48 bytes, which is
2161 // prereserved space for [SP][CR][LR][3 x unused].
2162 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2163 unsigned NumOps = TheCall->getNumArgs();
2164 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2166 // Add up all the space actually used.
2167 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
2168 // they all go in registers, but we must reserve stack space for them for
2169 // possible use by the caller. In varargs or 64-bit calls, parameters are
2170 // assigned stack space in order, with padding so Altivec parameters are
2172 nAltivecParamsAtEnd = 0;
2173 for (unsigned i = 0; i != NumOps; ++i) {
2174 SDValue Arg = TheCall->getArg(i);
2175 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2176 MVT ArgVT = Arg.getValueType();
2177 // Varargs Altivec parameters are padded to a 16 byte boundary.
2178 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
2179 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
2180 if (!isVarArg && !isPPC64) {
2181 // Non-varargs Altivec parameters go after all the non-Altivec
2182 // parameters; handle those later so we know how much padding we need.
2183 nAltivecParamsAtEnd++;
2186 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
2187 NumBytes = ((NumBytes+15)/16)*16;
2189 NumBytes += CalculateStackSlotSize(Arg, Flags, PtrByteSize);
2192 // Allow for Altivec parameters at the end, if needed.
2193 if (nAltivecParamsAtEnd) {
2194 NumBytes = ((NumBytes+15)/16)*16;
2195 NumBytes += 16*nAltivecParamsAtEnd;
2198 // The prolog code of the callee may store up to 8 GPR argument registers to
2199 // the stack, allowing va_start to index over them in memory if its varargs.
2200 // Because we cannot tell if this is needed on the caller side, we have to
2201 // conservatively assume that it is needed. As such, make sure we have at
2202 // least enough stack space for the caller to store the 8 GPRs.
2203 NumBytes = std::max(NumBytes,
2204 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
2206 // Tail call needs the stack to be aligned.
2207 if (CC==CallingConv::Fast && PerformTailCallOpt) {
2208 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
2209 getStackAlignment();
2210 unsigned AlignMask = TargetAlign-1;
2211 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
2217 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
2218 /// adjusted to accomodate the arguments for the tailcall.
2219 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
2220 unsigned ParamSize) {
2222 if (!IsTailCall) return 0;
2224 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
2225 unsigned CallerMinReservedArea = FI->getMinReservedArea();
2226 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
2227 // Remember only if the new adjustement is bigger.
2228 if (SPDiff < FI->getTailCallSPDelta())
2229 FI->setTailCallSPDelta(SPDiff);
2234 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
2235 /// following the call is a return. A function is eligible if caller/callee
2236 /// calling conventions match, currently only fastcc supports tail calls, and
2237 /// the function CALL is immediatly followed by a RET.
2239 PPCTargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
2241 SelectionDAG& DAG) const {
2242 // Variable argument functions are not supported.
2243 if (!PerformTailCallOpt || TheCall->isVarArg())
2246 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
2247 MachineFunction &MF = DAG.getMachineFunction();
2248 unsigned CallerCC = MF.getFunction()->getCallingConv();
2249 unsigned CalleeCC = TheCall->getCallingConv();
2250 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
2251 // Functions containing by val parameters are not supported.
2252 for (unsigned i = 0; i != TheCall->getNumArgs(); i++) {
2253 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2254 if (Flags.isByVal()) return false;
2257 SDValue Callee = TheCall->getCallee();
2258 // Non PIC/GOT tail calls are supported.
2259 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
2262 // At the moment we can only do local tail calls (in same module, hidden
2263 // or protected) if we are generating PIC.
2264 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2265 return G->getGlobal()->hasHiddenVisibility()
2266 || G->getGlobal()->hasProtectedVisibility();
2273 /// isCallCompatibleAddress - Return the immediate to use if the specified
2274 /// 32-bit value is representable in the immediate field of a BxA instruction.
2275 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
2276 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2279 int Addr = C->getZExtValue();
2280 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
2281 (Addr << 6 >> 6) != Addr)
2282 return 0; // Top 6 bits have to be sext of immediate.
2284 return DAG.getConstant((int)C->getZExtValue() >> 2,
2285 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
2290 struct TailCallArgumentInfo {
2295 TailCallArgumentInfo() : FrameIdx(0) {}
2300 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
2302 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
2304 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
2305 SmallVector<SDValue, 8> &MemOpChains,
2307 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
2308 SDValue Arg = TailCallArgs[i].Arg;
2309 SDValue FIN = TailCallArgs[i].FrameIdxOp;
2310 int FI = TailCallArgs[i].FrameIdx;
2311 // Store relative to framepointer.
2312 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, FIN,
2313 PseudoSourceValue::getFixedStack(FI),
2318 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
2319 /// the appropriate stack slot for the tail call optimized function call.
2320 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2321 MachineFunction &MF,
2330 // Calculate the new stack slot for the return address.
2331 int SlotSize = isPPC64 ? 8 : 4;
2332 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
2334 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2336 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2337 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2338 Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx,
2339 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
2341 // When using the SVR4 ABI there is no need to move the FP stack slot
2342 // as the FP is never overwritten.
2345 SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64, isMachoABI);
2346 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc);
2347 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2348 Chain = DAG.getStore(Chain, dl, OldFP, NewFramePtrIdx,
2349 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
2355 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2356 /// the position of the argument.
2358 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2359 SDValue Arg, int SPDiff, unsigned ArgOffset,
2360 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2361 int Offset = ArgOffset + SPDiff;
2362 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2363 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
2364 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2365 SDValue FIN = DAG.getFrameIndex(FI, VT);
2366 TailCallArgumentInfo Info;
2368 Info.FrameIdxOp = FIN;
2370 TailCallArguments.push_back(Info);
2373 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2374 /// stack slot. Returns the chain as result and the loaded frame pointers in
2375 /// LROpOut/FPOpout. Used when tail calling.
2376 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2384 // Load the LR and FP stack slot for later adjusting.
2385 MVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2386 LROpOut = getReturnAddrFrameIndex(DAG);
2387 LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, NULL, 0);
2388 Chain = SDValue(LROpOut.getNode(), 1);
2390 // When using the SVR4 ABI there is no need to load the FP stack slot
2391 // as the FP is never overwritten.
2393 FPOpOut = getFramePointerFrameIndex(DAG);
2394 FPOpOut = DAG.getLoad(VT, dl, Chain, FPOpOut, NULL, 0);
2395 Chain = SDValue(FPOpOut.getNode(), 1);
2401 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2402 /// by "Src" to address "Dst" of size "Size". Alignment information is
2403 /// specified by the specific parameter attribute. The copy will be passed as
2404 /// a byval function parameter.
2405 /// Sometimes what we are copying is the end of a larger object, the part that
2406 /// does not fit in registers.
2408 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2409 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2411 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2412 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2413 false, NULL, 0, NULL, 0);
2416 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2419 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2420 SDValue Arg, SDValue PtrOff, int SPDiff,
2421 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2422 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2423 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments,
2425 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2430 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2432 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2433 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
2434 DAG.getConstant(ArgOffset, PtrVT));
2436 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0));
2437 // Calculate and remember argument location.
2438 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2442 SDValue PPCTargetLowering::LowerCALL_SVR4(SDValue Op, SelectionDAG &DAG,
2443 const PPCSubtarget &Subtarget,
2444 TargetMachine &TM) {
2445 // See PPCTargetLowering::LowerFORMAL_ARGUMENTS_SVR4() for a description
2446 // of the SVR4 ABI stack frame layout.
2447 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
2448 SDValue Chain = TheCall->getChain();
2449 bool isVarArg = TheCall->isVarArg();
2450 unsigned CC = TheCall->getCallingConv();
2451 assert((CC == CallingConv::C ||
2452 CC == CallingConv::Fast) && "Unknown calling convention!");
2453 bool isTailCall = TheCall->isTailCall()
2454 && CC == CallingConv::Fast && PerformTailCallOpt;
2455 SDValue Callee = TheCall->getCallee();
2456 DebugLoc dl = TheCall->getDebugLoc();
2458 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2459 unsigned PtrByteSize = 4;
2461 MachineFunction &MF = DAG.getMachineFunction();
2463 // Mark this function as potentially containing a function that contains a
2464 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2465 // and restoring the callers stack pointer in this functions epilog. This is
2466 // done because by tail calling the called function might overwrite the value
2467 // in this function's (MF) stack pointer stack slot 0(SP).
2468 if (PerformTailCallOpt && CC==CallingConv::Fast)
2469 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2471 // Count how many bytes are to be pushed on the stack, including the linkage
2472 // area, parameter list area and the part of the local variable space which
2473 // contains copies of aggregates which are passed by value.
2475 // Assign locations to all of the outgoing arguments.
2476 SmallVector<CCValAssign, 16> ArgLocs;
2477 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs);
2479 // Reserve space for the linkage area on the stack.
2480 CCInfo.AllocateStack(PPCFrameInfo::getLinkageSize(false, false), PtrByteSize);
2483 // Handle fixed and variable vector arguments differently.
2484 // Fixed vector arguments go into registers as long as registers are
2485 // available. Variable vector arguments always go into memory.
2486 unsigned NumArgs = TheCall->getNumArgs();
2487 unsigned NumFixedArgs = TheCall->getNumFixedArgs();
2489 for (unsigned i = 0; i != NumArgs; ++i) {
2490 MVT ArgVT = TheCall->getArg(i).getValueType();
2491 ISD::ArgFlagsTy ArgFlags = TheCall->getArgFlags(i);
2494 if (i < NumFixedArgs) {
2495 Result = CC_PPC_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags,
2498 Result = CC_PPC_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full,
2503 cerr << "Call operand #" << i << " has unhandled type "
2504 << ArgVT.getMVTString() << "\n";
2509 // All arguments are treated the same.
2510 CCInfo.AnalyzeCallOperands(TheCall, CC_PPC_SVR4);
2513 // Assign locations to all of the outgoing aggregate by value arguments.
2514 SmallVector<CCValAssign, 16> ByValArgLocs;
2515 CCState CCByValInfo(CC, isVarArg, getTargetMachine(), ByValArgLocs);
2517 // Reserve stack space for the allocations in CCInfo.
2518 CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
2520 CCByValInfo.AnalyzeCallOperands(TheCall, CC_PPC_SVR4_ByVal);
2522 // Size of the linkage area, parameter list area and the part of the local
2523 // space variable where copies of aggregates which are passed by value are
2525 unsigned NumBytes = CCByValInfo.getNextStackOffset();
2527 // Calculate by how many bytes the stack has to be adjusted in case of tail
2528 // call optimization.
2529 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2531 // Adjust the stack pointer for the new arguments...
2532 // These operations are automatically eliminated by the prolog/epilog pass
2533 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2534 SDValue CallSeqStart = Chain;
2536 // Load the return address and frame pointer so it can be moved somewhere else
2539 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, false,
2542 // Set up a copy of the stack pointer for use loading and storing any
2543 // arguments that may not fit in the registers available for argument
2545 SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2547 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2548 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2549 SmallVector<SDValue, 8> MemOpChains;
2551 // Walk the register/memloc assignments, inserting copies/loads.
2552 for (unsigned i = 0, j = 0, e = ArgLocs.size();
2555 CCValAssign &VA = ArgLocs[i];
2556 SDValue Arg = TheCall->getArg(i);
2557 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2559 if (Flags.isByVal()) {
2560 // Argument is an aggregate which is passed by value, thus we need to
2561 // create a copy of it in the local variable space of the current stack
2562 // frame (which is the stack frame of the caller) and pass the address of
2563 // this copy to the callee.
2564 assert((j < ByValArgLocs.size()) && "Index out of bounds!");
2565 CCValAssign &ByValVA = ByValArgLocs[j++];
2566 assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!");
2568 // Memory reserved in the local variable space of the callers stack frame.
2569 unsigned LocMemOffset = ByValVA.getLocMemOffset();
2571 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2572 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2574 // Create a copy of the argument in the local area of the current
2576 SDValue MemcpyCall =
2577 CreateCopyOfByValArgument(Arg, PtrOff,
2578 CallSeqStart.getNode()->getOperand(0),
2581 // This must go outside the CALLSEQ_START..END.
2582 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2583 CallSeqStart.getNode()->getOperand(1));
2584 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2585 NewCallSeqStart.getNode());
2586 Chain = CallSeqStart = NewCallSeqStart;
2588 // Pass the address of the aggregate copy on the stack either in a
2589 // physical register or in the parameter list area of the current stack
2590 // frame to the callee.
2594 if (VA.isRegLoc()) {
2595 // Put argument in a physical register.
2596 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2598 // Put argument in the parameter list area of the current stack frame.
2599 assert(VA.isMemLoc());
2600 unsigned LocMemOffset = VA.getLocMemOffset();
2603 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2604 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2606 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
2607 PseudoSourceValue::getStack(), LocMemOffset));
2609 // Calculate and remember argument location.
2610 CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset,
2616 if (!MemOpChains.empty())
2617 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2618 &MemOpChains[0], MemOpChains.size());
2620 // Build a sequence of copy-to-reg nodes chained together with token chain
2621 // and flag operands which copy the outgoing args into the appropriate regs.
2623 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2624 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2625 RegsToPass[i].second, InFlag);
2626 InFlag = Chain.getValue(1);
2629 // Set CR6 to true if this is a vararg call.
2631 SDValue SetCR(DAG.getTargetNode(PPC::CRSET, dl, MVT::i32), 0);
2632 Chain = DAG.getCopyToReg(Chain, dl, PPC::CR1EQ, SetCR, InFlag);
2633 InFlag = Chain.getValue(1);
2636 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2637 // might overwrite each other in case of tail call optimization.
2639 SmallVector<SDValue, 8> MemOpChains2;
2640 // Do not flag preceeding copytoreg stuff together with the following stuff.
2642 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2644 if (!MemOpChains2.empty())
2645 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2646 &MemOpChains2[0], MemOpChains2.size());
2648 // Store the return address to the appropriate stack slot.
2649 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2653 // Emit callseq_end just before tailcall node.
2655 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2656 DAG.getIntPtrConstant(0, true), InFlag);
2657 InFlag = Chain.getValue(1);
2660 std::vector<MVT> NodeTys;
2661 NodeTys.push_back(MVT::Other); // Returns a chain
2662 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2664 SmallVector<SDValue, 8> Ops;
2665 unsigned CallOpc = PPCISD::CALL_ELF;
2667 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2668 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2669 // node so that legalize doesn't hack it.
2670 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2671 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2672 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2673 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2674 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2675 // If this is an absolute destination address, use the munged value.
2676 Callee = SDValue(Dest, 0);
2678 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2679 // to do the call, we can't use PPCISD::CALL.
2680 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2681 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
2682 2 + (InFlag.getNode() != 0));
2683 InFlag = Chain.getValue(1);
2686 NodeTys.push_back(MVT::Other);
2687 NodeTys.push_back(MVT::Flag);
2688 Ops.push_back(Chain);
2689 CallOpc = PPCISD::BCTRL_ELF;
2691 // Add CTR register as callee so a bctr can be emitted later.
2693 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
2696 // If this is a direct call, pass the chain and the callee.
2697 if (Callee.getNode()) {
2698 Ops.push_back(Chain);
2699 Ops.push_back(Callee);
2701 // If this is a tail call add stack pointer delta.
2703 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2705 // Add argument registers to the end of the list so that they are known live
2707 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2708 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2709 RegsToPass[i].second.getValueType()));
2711 // When performing tail call optimization the callee pops its arguments off
2712 // the stack. Account for this here so these bytes can be pushed back on in
2713 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2714 int BytesCalleePops =
2715 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2717 if (InFlag.getNode())
2718 Ops.push_back(InFlag);
2722 assert(InFlag.getNode() &&
2723 "Flag must be set. Depend on flag being set in LowerRET");
2724 Chain = DAG.getNode(PPCISD::TAILCALL, dl,
2725 TheCall->getVTList(), &Ops[0], Ops.size());
2726 return SDValue(Chain.getNode(), Op.getResNo());
2729 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
2730 InFlag = Chain.getValue(1);
2732 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2733 DAG.getIntPtrConstant(BytesCalleePops, true),
2735 if (TheCall->getValueType(0) != MVT::Other)
2736 InFlag = Chain.getValue(1);
2738 SmallVector<SDValue, 16> ResultVals;
2739 SmallVector<CCValAssign, 16> RVLocs;
2740 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
2741 CCState CCRetInfo(CallerCC, isVarArg, TM, RVLocs);
2742 CCRetInfo.AnalyzeCallResult(TheCall, RetCC_PPC);
2744 // Copy all of the result registers out of their specified physreg.
2745 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2746 CCValAssign &VA = RVLocs[i];
2747 MVT VT = VA.getValVT();
2748 assert(VA.isRegLoc() && "Can only return in registers!");
2749 Chain = DAG.getCopyFromReg(Chain, dl,
2750 VA.getLocReg(), VT, InFlag).getValue(1);
2751 ResultVals.push_back(Chain.getValue(0));
2752 InFlag = Chain.getValue(2);
2755 // If the function returns void, just return the chain.
2759 // Otherwise, merge everything together with a MERGE_VALUES node.
2760 ResultVals.push_back(Chain);
2761 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
2762 &ResultVals[0], ResultVals.size());
2763 return Res.getValue(Op.getResNo());
2766 SDValue PPCTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG,
2767 const PPCSubtarget &Subtarget,
2768 TargetMachine &TM) {
2769 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
2770 SDValue Chain = TheCall->getChain();
2771 bool isVarArg = TheCall->isVarArg();
2772 unsigned CC = TheCall->getCallingConv();
2773 bool isTailCall = TheCall->isTailCall()
2774 && CC == CallingConv::Fast && PerformTailCallOpt;
2775 SDValue Callee = TheCall->getCallee();
2776 unsigned NumOps = TheCall->getNumArgs();
2777 DebugLoc dl = TheCall->getDebugLoc();
2779 bool isMachoABI = Subtarget.isMachoABI();
2780 bool isELF32_ABI = Subtarget.isELF32_ABI();
2782 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2783 bool isPPC64 = PtrVT == MVT::i64;
2784 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2786 MachineFunction &MF = DAG.getMachineFunction();
2788 // Mark this function as potentially containing a function that contains a
2789 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2790 // and restoring the callers stack pointer in this functions epilog. This is
2791 // done because by tail calling the called function might overwrite the value
2792 // in this function's (MF) stack pointer stack slot 0(SP).
2793 if (PerformTailCallOpt && CC==CallingConv::Fast)
2794 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2796 unsigned nAltivecParamsAtEnd = 0;
2798 // Count how many bytes are to be pushed on the stack, including the linkage
2799 // area, and parameter passing area. We start with 24/48 bytes, which is
2800 // prereserved space for [SP][CR][LR][3 x unused].
2802 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isMachoABI, isVarArg, CC,
2803 TheCall, nAltivecParamsAtEnd);
2805 // Calculate by how many bytes the stack has to be adjusted in case of tail
2806 // call optimization.
2807 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2809 // Adjust the stack pointer for the new arguments...
2810 // These operations are automatically eliminated by the prolog/epilog pass
2811 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2812 SDValue CallSeqStart = Chain;
2814 // Load the return address and frame pointer so it can be move somewhere else
2817 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, true,
2820 // Set up a copy of the stack pointer for use loading and storing any
2821 // arguments that may not fit in the registers available for argument
2825 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2827 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2829 // Figure out which arguments are going to go in registers, and which in
2830 // memory. Also, if this is a vararg function, floating point operations
2831 // must be stored to our stack, and loaded into integer regs as well, if
2832 // any integer regs are available for argument passing.
2833 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2834 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2836 static const unsigned GPR_32[] = { // 32-bit registers.
2837 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2838 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2840 static const unsigned GPR_64[] = { // 64-bit registers.
2841 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2842 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2844 static const unsigned *FPR = GetFPR(Subtarget);
2846 static const unsigned VR[] = {
2847 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2848 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2850 const unsigned NumGPRs = array_lengthof(GPR_32);
2851 const unsigned NumFPRs = isMachoABI ? 13 : 8;
2852 const unsigned NumVRs = array_lengthof(VR);
2854 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2856 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
2857 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2859 SmallVector<SDValue, 8> MemOpChains;
2860 for (unsigned i = 0; i != NumOps; ++i) {
2862 SDValue Arg = TheCall->getArg(i);
2863 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2864 // See if next argument requires stack alignment in ELF
2865 bool Align = Flags.isSplit();
2867 // PtrOff will be used to store the current argument to the stack if a
2868 // register cannot be found for it.
2871 // Stack align in ELF 32
2872 if (isELF32_ABI && Align)
2873 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
2874 StackPtr.getValueType());
2876 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2878 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
2880 // On PPC64, promote integers to 64-bit values.
2881 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2882 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2883 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2884 Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg);
2887 // FIXME Elf untested, what are alignment rules?
2888 // FIXME memcpy is used way more than necessary. Correctness first.
2889 if (Flags.isByVal()) {
2890 unsigned Size = Flags.getByValSize();
2891 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2892 if (Size==1 || Size==2) {
2893 // Very small objects are passed right-justified.
2894 // Everything else is passed left-justified.
2895 MVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2896 if (GPR_idx != NumGPRs) {
2897 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
2899 MemOpChains.push_back(Load.getValue(1));
2900 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2902 ArgOffset += PtrByteSize;
2904 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2905 SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const);
2906 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2907 CallSeqStart.getNode()->getOperand(0),
2909 // This must go outside the CALLSEQ_START..END.
2910 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2911 CallSeqStart.getNode()->getOperand(1));
2912 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2913 NewCallSeqStart.getNode());
2914 Chain = CallSeqStart = NewCallSeqStart;
2915 ArgOffset += PtrByteSize;
2919 // Copy entire object into memory. There are cases where gcc-generated
2920 // code assumes it is there, even if it could be put entirely into
2921 // registers. (This is not what the doc says.)
2922 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2923 CallSeqStart.getNode()->getOperand(0),
2925 // This must go outside the CALLSEQ_START..END.
2926 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2927 CallSeqStart.getNode()->getOperand(1));
2928 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2929 Chain = CallSeqStart = NewCallSeqStart;
2930 // And copy the pieces of it that fit into registers.
2931 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2932 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2933 SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
2934 if (GPR_idx != NumGPRs) {
2935 SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg, NULL, 0);
2936 MemOpChains.push_back(Load.getValue(1));
2937 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2939 ArgOffset += PtrByteSize;
2941 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2948 switch (Arg.getValueType().getSimpleVT()) {
2949 default: assert(0 && "Unexpected ValueType for argument!");
2952 // Double word align in ELF
2953 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2954 if (GPR_idx != NumGPRs) {
2955 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2957 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2958 isPPC64, isTailCall, false, MemOpChains,
2959 TailCallArguments, dl);
2962 if (inMem || isMachoABI) {
2963 // Stack align in ELF
2964 if (isELF32_ABI && Align)
2965 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2967 ArgOffset += PtrByteSize;
2972 if (FPR_idx != NumFPRs) {
2973 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
2976 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
2977 MemOpChains.push_back(Store);
2979 // Float varargs are always shadowed in available integer registers
2980 if (GPR_idx != NumGPRs) {
2981 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
2982 MemOpChains.push_back(Load.getValue(1));
2983 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2986 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
2987 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
2988 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
2989 SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff, NULL, 0);
2990 MemOpChains.push_back(Load.getValue(1));
2991 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2995 // If we have any FPRs remaining, we may also have GPRs remaining.
2996 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
2999 if (GPR_idx != NumGPRs)
3001 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
3002 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
3007 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3008 isPPC64, isTailCall, false, MemOpChains,
3009 TailCallArguments, dl);
3012 if (inMem || isMachoABI) {
3013 // Stack align in ELF
3014 if (isELF32_ABI && Align)
3015 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
3019 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
3027 // These go aligned on the stack, or in the corresponding R registers
3028 // when within range. The Darwin PPC ABI doc claims they also go in
3029 // V registers; in fact gcc does this only for arguments that are
3030 // prototyped, not for those that match the ... We do it for all
3031 // arguments, seems to work.
3032 while (ArgOffset % 16 !=0) {
3033 ArgOffset += PtrByteSize;
3034 if (GPR_idx != NumGPRs)
3037 // We could elide this store in the case where the object fits
3038 // entirely in R registers. Maybe later.
3039 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr,
3040 DAG.getConstant(ArgOffset, PtrVT));
3041 SDValue Store = DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0);
3042 MemOpChains.push_back(Store);
3043 if (VR_idx != NumVRs) {
3044 SDValue Load = DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, NULL, 0);
3045 MemOpChains.push_back(Load.getValue(1));
3046 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
3049 for (unsigned i=0; i<16; i+=PtrByteSize) {
3050 if (GPR_idx == NumGPRs)
3052 SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff,
3053 DAG.getConstant(i, PtrVT));
3054 SDValue Load = DAG.getLoad(PtrVT, dl, Store, Ix, NULL, 0);
3055 MemOpChains.push_back(Load.getValue(1));
3056 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
3061 // Non-varargs Altivec params generally go in registers, but have
3062 // stack space allocated at the end.
3063 if (VR_idx != NumVRs) {
3064 // Doesn't have GPR space allocated.
3065 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
3066 } else if (nAltivecParamsAtEnd==0) {
3067 // We are emitting Altivec params in order.
3068 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3069 isPPC64, isTailCall, true, MemOpChains,
3070 TailCallArguments, dl);
3076 // If all Altivec parameters fit in registers, as they usually do,
3077 // they get stack space following the non-Altivec parameters. We
3078 // don't track this here because nobody below needs it.
3079 // If there are more Altivec parameters than fit in registers emit
3081 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
3083 // Offset is aligned; skip 1st 12 params which go in V registers.
3084 ArgOffset = ((ArgOffset+15)/16)*16;
3086 for (unsigned i = 0; i != NumOps; ++i) {
3087 SDValue Arg = TheCall->getArg(i);
3088 MVT ArgType = Arg.getValueType();
3089 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
3090 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
3093 // We are emitting Altivec params in order.
3094 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
3095 isPPC64, isTailCall, true, MemOpChains,
3096 TailCallArguments, dl);
3103 if (!MemOpChains.empty())
3104 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3105 &MemOpChains[0], MemOpChains.size());
3107 // Build a sequence of copy-to-reg nodes chained together with token chain
3108 // and flag operands which copy the outgoing args into the appropriate regs.
3110 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3111 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3112 RegsToPass[i].second, InFlag);
3113 InFlag = Chain.getValue(1);
3116 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
3117 // might overwrite each other in case of tail call optimization.
3119 SmallVector<SDValue, 8> MemOpChains2;
3120 // Do not flag preceeding copytoreg stuff together with the following stuff.
3122 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
3124 if (!MemOpChains2.empty())
3125 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
3126 &MemOpChains2[0], MemOpChains2.size());
3128 // Store the return address to the appropriate stack slot.
3129 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
3130 isPPC64, isMachoABI, dl);
3133 // Emit callseq_end just before tailcall node.
3135 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3136 DAG.getIntPtrConstant(0, true), InFlag);
3137 InFlag = Chain.getValue(1);
3140 std::vector<MVT> NodeTys;
3141 NodeTys.push_back(MVT::Other); // Returns a chain
3142 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
3144 SmallVector<SDValue, 8> Ops;
3145 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
3147 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
3148 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
3149 // node so that legalize doesn't hack it.
3150 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
3151 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
3152 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
3153 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
3154 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
3155 // If this is an absolute destination address, use the munged value.
3156 Callee = SDValue(Dest, 0);
3158 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
3159 // to do the call, we can't use PPCISD::CALL.
3160 SDValue MTCTROps[] = {Chain, Callee, InFlag};
3161 Chain = DAG.getNode(PPCISD::MTCTR, dl, NodeTys, MTCTROps,
3162 2 + (InFlag.getNode() != 0));
3163 InFlag = Chain.getValue(1);
3165 // Copy the callee address into R12/X12 on darwin.
3167 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
3168 Chain = DAG.getCopyToReg(Chain, dl, Reg, Callee, InFlag);
3169 InFlag = Chain.getValue(1);
3173 NodeTys.push_back(MVT::Other);
3174 NodeTys.push_back(MVT::Flag);
3175 Ops.push_back(Chain);
3176 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
3178 // Add CTR register as callee so a bctr can be emitted later.
3180 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
3183 // If this is a direct call, pass the chain and the callee.
3184 if (Callee.getNode()) {
3185 Ops.push_back(Chain);
3186 Ops.push_back(Callee);
3188 // If this is a tail call add stack pointer delta.
3190 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
3192 // Add argument registers to the end of the list so that they are known live
3194 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3195 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3196 RegsToPass[i].second.getValueType()));
3198 // When performing tail call optimization the callee pops its arguments off
3199 // the stack. Account for this here so these bytes can be pushed back on in
3200 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
3201 int BytesCalleePops =
3202 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
3204 if (InFlag.getNode())
3205 Ops.push_back(InFlag);
3209 assert(InFlag.getNode() &&
3210 "Flag must be set. Depend on flag being set in LowerRET");
3211 Chain = DAG.getNode(PPCISD::TAILCALL, dl,
3212 TheCall->getVTList(), &Ops[0], Ops.size());
3213 return SDValue(Chain.getNode(), Op.getResNo());
3216 Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
3217 InFlag = Chain.getValue(1);
3219 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
3220 DAG.getIntPtrConstant(BytesCalleePops, true),
3222 if (TheCall->getValueType(0) != MVT::Other)
3223 InFlag = Chain.getValue(1);
3225 SmallVector<SDValue, 16> ResultVals;
3226 SmallVector<CCValAssign, 16> RVLocs;
3227 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
3228 CCState CCInfo(CallerCC, isVarArg, TM, RVLocs);
3229 CCInfo.AnalyzeCallResult(TheCall, RetCC_PPC);
3231 // Copy all of the result registers out of their specified physreg.
3232 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3233 CCValAssign &VA = RVLocs[i];
3234 MVT VT = VA.getValVT();
3235 assert(VA.isRegLoc() && "Can only return in registers!");
3236 Chain = DAG.getCopyFromReg(Chain, dl,
3237 VA.getLocReg(), VT, InFlag).getValue(1);
3238 ResultVals.push_back(Chain.getValue(0));
3239 InFlag = Chain.getValue(2);
3242 // If the function returns void, just return the chain.
3246 // Otherwise, merge everything together with a MERGE_VALUES node.
3247 ResultVals.push_back(Chain);
3248 SDValue Res = DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
3249 &ResultVals[0], ResultVals.size());
3250 return Res.getValue(Op.getResNo());
3253 SDValue PPCTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG,
3254 TargetMachine &TM) {
3255 SmallVector<CCValAssign, 16> RVLocs;
3256 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
3257 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
3258 DebugLoc dl = Op.getDebugLoc();
3259 CCState CCInfo(CC, isVarArg, TM, RVLocs);
3260 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_PPC);
3262 // If this is the first return lowered for this function, add the regs to the
3263 // liveout set for the function.
3264 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
3265 for (unsigned i = 0; i != RVLocs.size(); ++i)
3266 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
3269 SDValue Chain = Op.getOperand(0);
3271 Chain = GetPossiblePreceedingTailCall(Chain, PPCISD::TAILCALL);
3272 if (Chain.getOpcode() == PPCISD::TAILCALL) {
3273 SDValue TailCall = Chain;
3274 SDValue TargetAddress = TailCall.getOperand(1);
3275 SDValue StackAdjustment = TailCall.getOperand(2);
3277 assert(((TargetAddress.getOpcode() == ISD::Register &&
3278 cast<RegisterSDNode>(TargetAddress)->getReg() == PPC::CTR) ||
3279 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
3280 TargetAddress.getOpcode() == ISD::TargetGlobalAddress ||
3281 isa<ConstantSDNode>(TargetAddress)) &&
3282 "Expecting an global address, external symbol, absolute value or register");
3284 assert(StackAdjustment.getOpcode() == ISD::Constant &&
3285 "Expecting a const value");
3287 SmallVector<SDValue,8> Operands;
3288 Operands.push_back(Chain.getOperand(0));
3289 Operands.push_back(TargetAddress);
3290 Operands.push_back(StackAdjustment);
3291 // Copy registers used by the call. Last operand is a flag so it is not
3293 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
3294 Operands.push_back(Chain.getOperand(i));
3296 return DAG.getNode(PPCISD::TC_RETURN, dl, MVT::Other, &Operands[0],
3302 // Copy the result values into the output registers.
3303 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3304 CCValAssign &VA = RVLocs[i];
3305 assert(VA.isRegLoc() && "Can only return in registers!");
3306 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
3307 Op.getOperand(i*2+1), Flag);
3308 Flag = Chain.getValue(1);
3312 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
3314 return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, Chain);
3317 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
3318 const PPCSubtarget &Subtarget) {
3319 // When we pop the dynamic allocation we need to restore the SP link.
3320 DebugLoc dl = Op.getDebugLoc();
3322 // Get the corect type for pointers.
3323 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3325 // Construct the stack pointer operand.
3326 bool IsPPC64 = Subtarget.isPPC64();
3327 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
3328 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
3330 // Get the operands for the STACKRESTORE.
3331 SDValue Chain = Op.getOperand(0);
3332 SDValue SaveSP = Op.getOperand(1);
3334 // Load the old link SP.
3335 SDValue LoadLinkSP = DAG.getLoad(PtrVT, dl, Chain, StackPtr, NULL, 0);
3337 // Restore the stack pointer.
3338 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP);
3340 // Store the old link SP.
3341 return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, NULL, 0);
3347 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
3348 MachineFunction &MF = DAG.getMachineFunction();
3349 bool IsPPC64 = PPCSubTarget.isPPC64();
3350 bool isMachoABI = PPCSubTarget.isMachoABI();
3351 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3353 // Get current frame pointer save index. The users of this index will be
3354 // primarily DYNALLOC instructions.
3355 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3356 int RASI = FI->getReturnAddrSaveIndex();
3358 // If the frame pointer save index hasn't been defined yet.
3360 // Find out what the fix offset of the frame pointer save area.
3361 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isMachoABI);
3362 // Allocate the frame index for frame pointer save area.
3363 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset);
3365 FI->setReturnAddrSaveIndex(RASI);
3367 return DAG.getFrameIndex(RASI, PtrVT);
3371 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
3372 MachineFunction &MF = DAG.getMachineFunction();
3373 bool IsPPC64 = PPCSubTarget.isPPC64();
3374 bool isMachoABI = PPCSubTarget.isMachoABI();
3375 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3377 // Get current frame pointer save index. The users of this index will be
3378 // primarily DYNALLOC instructions.
3379 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
3380 int FPSI = FI->getFramePointerSaveIndex();
3382 // If the frame pointer save index hasn't been defined yet.
3384 // Find out what the fix offset of the frame pointer save area.
3385 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
3387 // Allocate the frame index for frame pointer save area.
3388 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
3390 FI->setFramePointerSaveIndex(FPSI);
3392 return DAG.getFrameIndex(FPSI, PtrVT);
3395 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
3397 const PPCSubtarget &Subtarget) {
3399 SDValue Chain = Op.getOperand(0);
3400 SDValue Size = Op.getOperand(1);
3401 DebugLoc dl = Op.getDebugLoc();
3403 // Get the corect type for pointers.
3404 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3406 SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT,
3407 DAG.getConstant(0, PtrVT), Size);
3408 // Construct a node for the frame pointer save index.
3409 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
3410 // Build a DYNALLOC node.
3411 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
3412 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
3413 return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops, 3);
3416 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
3418 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
3419 // Not FP? Not a fsel.
3420 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
3421 !Op.getOperand(2).getValueType().isFloatingPoint())
3424 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3426 // Cannot handle SETEQ/SETNE.
3427 if (CC == ISD::SETEQ || CC == ISD::SETNE) return Op;
3429 MVT ResVT = Op.getValueType();
3430 MVT CmpVT = Op.getOperand(0).getValueType();
3431 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3432 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
3433 DebugLoc dl = Op.getDebugLoc();
3435 // If the RHS of the comparison is a 0.0, we don't need to do the
3436 // subtraction at all.
3437 if (isFloatingPointZero(RHS))
3439 default: break; // SETUO etc aren't handled by fsel.
3442 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3445 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3446 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3447 return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV);
3450 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
3453 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
3454 LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS);
3455 return DAG.getNode(PPCISD::FSEL, dl, ResVT,
3456 DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV);
3461 default: break; // SETUO etc aren't handled by fsel.
3464 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3465 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3466 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3467 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3470 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS);
3471 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3472 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3473 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3476 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3477 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3478 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3479 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV);
3482 Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS);
3483 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
3484 Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp);
3485 return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV);
3490 // FIXME: Split this code up when LegalizeDAGTypes lands.
3491 SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
3493 assert(Op.getOperand(0).getValueType().isFloatingPoint());
3494 SDValue Src = Op.getOperand(0);
3495 if (Src.getValueType() == MVT::f32)
3496 Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src);
3499 switch (Op.getValueType().getSimpleVT()) {
3500 default: assert(0 && "Unhandled FP_TO_INT type in custom expander!");
3502 Tmp = DAG.getNode(Op.getOpcode()==ISD::FP_TO_SINT ? PPCISD::FCTIWZ :
3507 Tmp = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Src);
3511 // Convert the FP value to an int value through memory.
3512 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
3514 // Emit a store to the stack slot.
3515 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Tmp, FIPtr, NULL, 0);
3517 // Result is a load from the stack slot. If loading 4 bytes, make sure to
3519 if (Op.getValueType() == MVT::i32)
3520 FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr,
3521 DAG.getConstant(4, FIPtr.getValueType()));
3522 return DAG.getLoad(Op.getValueType(), dl, Chain, FIPtr, NULL, 0);
3525 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3526 DebugLoc dl = Op.getDebugLoc();
3527 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
3528 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
3531 if (Op.getOperand(0).getValueType() == MVT::i64) {
3532 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, dl,
3533 MVT::f64, Op.getOperand(0));
3534 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Bits);
3535 if (Op.getValueType() == MVT::f32)
3536 FP = DAG.getNode(ISD::FP_ROUND, dl,
3537 MVT::f32, FP, DAG.getIntPtrConstant(0));
3541 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
3542 "Unhandled SINT_TO_FP type in custom expander!");
3543 // Since we only generate this in 64-bit mode, we can take advantage of
3544 // 64-bit registers. In particular, sign extend the input value into the
3545 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
3546 // then lfd it and fcfid it.
3547 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3548 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
3549 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3550 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3552 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, dl, MVT::i32,
3555 // STD the extended value into the stack slot.
3556 MachineMemOperand MO(PseudoSourceValue::getFixedStack(FrameIdx),
3557 MachineMemOperand::MOStore, 0, 8, 8);
3558 SDValue Store = DAG.getNode(PPCISD::STD_32, dl, MVT::Other,
3559 DAG.getEntryNode(), Ext64, FIdx,
3560 DAG.getMemOperand(MO));
3561 // Load the value as a double.
3562 SDValue Ld = DAG.getLoad(MVT::f64, dl, Store, FIdx, NULL, 0);
3564 // FCFID it and return it.
3565 SDValue FP = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Ld);
3566 if (Op.getValueType() == MVT::f32)
3567 FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, DAG.getIntPtrConstant(0));
3571 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
3572 DebugLoc dl = Op.getDebugLoc();
3574 The rounding mode is in bits 30:31 of FPSR, and has the following
3581 FLT_ROUNDS, on the other hand, expects the following:
3588 To perform the conversion, we do:
3589 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
3592 MachineFunction &MF = DAG.getMachineFunction();
3593 MVT VT = Op.getValueType();
3594 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3595 std::vector<MVT> NodeTys;
3596 SDValue MFFSreg, InFlag;
3598 // Save FP Control Word to register
3599 NodeTys.push_back(MVT::f64); // return register
3600 NodeTys.push_back(MVT::Flag); // unused in this context
3601 SDValue Chain = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
3603 // Save FP register to stack slot
3604 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3605 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3606 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Chain,
3607 StackSlot, NULL, 0);
3609 // Load FP Control Word from low 32 bits of stack slot.
3610 SDValue Four = DAG.getConstant(4, PtrVT);
3611 SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four);
3612 SDValue CWD = DAG.getLoad(MVT::i32, dl, Store, Addr, NULL, 0);
3614 // Transform as necessary
3616 DAG.getNode(ISD::AND, dl, MVT::i32,
3617 CWD, DAG.getConstant(3, MVT::i32));
3619 DAG.getNode(ISD::SRL, dl, MVT::i32,
3620 DAG.getNode(ISD::AND, dl, MVT::i32,
3621 DAG.getNode(ISD::XOR, dl, MVT::i32,
3622 CWD, DAG.getConstant(3, MVT::i32)),
3623 DAG.getConstant(3, MVT::i32)),
3624 DAG.getConstant(1, MVT::i32));
3627 DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2);
3629 return DAG.getNode((VT.getSizeInBits() < 16 ?
3630 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
3633 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3634 MVT VT = Op.getValueType();
3635 unsigned BitWidth = VT.getSizeInBits();
3636 DebugLoc dl = Op.getDebugLoc();
3637 assert(Op.getNumOperands() == 3 &&
3638 VT == Op.getOperand(1).getValueType() &&
3641 // Expand into a bunch of logical ops. Note that these ops
3642 // depend on the PPC behavior for oversized shift amounts.
3643 SDValue Lo = Op.getOperand(0);
3644 SDValue Hi = Op.getOperand(1);
3645 SDValue Amt = Op.getOperand(2);
3646 MVT AmtVT = Amt.getValueType();
3648 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3649 DAG.getConstant(BitWidth, AmtVT), Amt);
3650 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt);
3651 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1);
3652 SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3);
3653 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3654 DAG.getConstant(-BitWidth, AmtVT));
3655 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5);
3656 SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3657 SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt);
3658 SDValue OutOps[] = { OutLo, OutHi };
3659 return DAG.getMergeValues(OutOps, 2, dl);
3662 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3663 MVT VT = Op.getValueType();
3664 DebugLoc dl = Op.getDebugLoc();
3665 unsigned BitWidth = VT.getSizeInBits();
3666 assert(Op.getNumOperands() == 3 &&
3667 VT == Op.getOperand(1).getValueType() &&
3670 // Expand into a bunch of logical ops. Note that these ops
3671 // depend on the PPC behavior for oversized shift amounts.
3672 SDValue Lo = Op.getOperand(0);
3673 SDValue Hi = Op.getOperand(1);
3674 SDValue Amt = Op.getOperand(2);
3675 MVT AmtVT = Amt.getValueType();
3677 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3678 DAG.getConstant(BitWidth, AmtVT), Amt);
3679 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3680 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3681 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3682 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3683 DAG.getConstant(-BitWidth, AmtVT));
3684 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5);
3685 SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6);
3686 SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt);
3687 SDValue OutOps[] = { OutLo, OutHi };
3688 return DAG.getMergeValues(OutOps, 2, dl);
3691 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3692 DebugLoc dl = Op.getDebugLoc();
3693 MVT VT = Op.getValueType();
3694 unsigned BitWidth = VT.getSizeInBits();
3695 assert(Op.getNumOperands() == 3 &&
3696 VT == Op.getOperand(1).getValueType() &&
3699 // Expand into a bunch of logical ops, followed by a select_cc.
3700 SDValue Lo = Op.getOperand(0);
3701 SDValue Hi = Op.getOperand(1);
3702 SDValue Amt = Op.getOperand(2);
3703 MVT AmtVT = Amt.getValueType();
3705 SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT,
3706 DAG.getConstant(BitWidth, AmtVT), Amt);
3707 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt);
3708 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1);
3709 SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3);
3710 SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt,
3711 DAG.getConstant(-BitWidth, AmtVT));
3712 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5);
3713 SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt);
3714 SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, AmtVT),
3715 Tmp4, Tmp6, ISD::SETLE);
3716 SDValue OutOps[] = { OutLo, OutHi };
3717 return DAG.getMergeValues(OutOps, 2, dl);
3720 //===----------------------------------------------------------------------===//
3721 // Vector related lowering.
3724 /// BuildSplatI - Build a canonical splati of Val with an element size of
3725 /// SplatSize. Cast the result to VT.
3726 static SDValue BuildSplatI(int Val, unsigned SplatSize, MVT VT,
3727 SelectionDAG &DAG, DebugLoc dl) {
3728 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3730 static const MVT VTys[] = { // canonical VT to use for each size.
3731 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3734 MVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3736 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3740 MVT CanonicalVT = VTys[SplatSize-1];
3742 // Build a canonical splat for this value.
3743 SDValue Elt = DAG.getConstant(Val, MVT::i32);
3744 SmallVector<SDValue, 8> Ops;
3745 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3746 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT,
3747 &Ops[0], Ops.size());
3748 return DAG.getNode(ISD::BIT_CONVERT, dl, ReqVT, Res);
3751 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3752 /// specified intrinsic ID.
3753 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3754 SelectionDAG &DAG, DebugLoc dl,
3755 MVT DestVT = MVT::Other) {
3756 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3757 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3758 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3761 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3762 /// specified intrinsic ID.
3763 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3764 SDValue Op2, SelectionDAG &DAG,
3765 DebugLoc dl, MVT DestVT = MVT::Other) {
3766 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3767 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
3768 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3772 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3773 /// amount. The result has the specified value type.
3774 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3775 MVT VT, SelectionDAG &DAG, DebugLoc dl) {
3776 // Force LHS/RHS to be the right type.
3777 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, LHS);
3778 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, RHS);
3781 for (unsigned i = 0; i != 16; ++i)
3783 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops);
3784 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
3787 // If this is a case we can't handle, return null and let the default
3788 // expansion code take care of it. If we CAN select this case, and if it
3789 // selects to a single instruction, return Op. Otherwise, if we can codegen
3790 // this case more efficiently than a constant pool load, lower it to the
3791 // sequence of ops that should be used.
3792 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3793 DebugLoc dl = Op.getDebugLoc();
3794 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3795 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
3797 // Check if this is a splat of a constant value.
3798 APInt APSplatBits, APSplatUndef;
3799 unsigned SplatBitSize;
3801 if (! BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
3802 HasAnyUndefs) || SplatBitSize > 32)
3805 unsigned SplatBits = APSplatBits.getZExtValue();
3806 unsigned SplatUndef = APSplatUndef.getZExtValue();
3807 unsigned SplatSize = SplatBitSize / 8;
3809 // First, handle single instruction cases.
3812 if (SplatBits == 0) {
3813 // Canonicalize all zero vectors to be v4i32.
3814 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3815 SDValue Z = DAG.getConstant(0, MVT::i32);
3816 Z = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Z, Z, Z, Z);
3817 Op = DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Z);
3822 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3823 int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >>
3825 if (SextVal >= -16 && SextVal <= 15)
3826 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG, dl);
3829 // Two instruction sequences.
3831 // If this value is in the range [-32,30] and is even, use:
3832 // tmp = VSPLTI[bhw], result = add tmp, tmp
3833 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3834 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG, dl);
3835 Res = DAG.getNode(ISD::ADD, dl, Res.getValueType(), Res, Res);
3836 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3839 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3840 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3842 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3843 // Make -1 and vspltisw -1:
3844 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG, dl);
3846 // Make the VSLW intrinsic, computing 0x8000_0000.
3847 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3850 // xor by OnesV to invert it.
3851 Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV);
3852 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3855 // Check to see if this is a wide variety of vsplti*, binop self cases.
3856 static const signed char SplatCsts[] = {
3857 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3858 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3861 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3862 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3863 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3864 int i = SplatCsts[idx];
3866 // Figure out what shift amount will be used by altivec if shifted by i in
3868 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3870 // vsplti + shl self.
3871 if (SextVal == (i << (int)TypeShiftAmt)) {
3872 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3873 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3874 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3875 Intrinsic::ppc_altivec_vslw
3877 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3878 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3881 // vsplti + srl self.
3882 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3883 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3884 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3885 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3886 Intrinsic::ppc_altivec_vsrw
3888 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3889 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3892 // vsplti + sra self.
3893 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3894 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3895 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3896 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3897 Intrinsic::ppc_altivec_vsraw
3899 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3900 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3903 // vsplti + rol self.
3904 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3905 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3906 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG, dl);
3907 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3908 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3909 Intrinsic::ppc_altivec_vrlw
3911 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl);
3912 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Res);
3915 // t = vsplti c, result = vsldoi t, t, 1
3916 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3917 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3918 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG, dl);
3920 // t = vsplti c, result = vsldoi t, t, 2
3921 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3922 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3923 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG, dl);
3925 // t = vsplti c, result = vsldoi t, t, 3
3926 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3927 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG, dl);
3928 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG, dl);
3932 // Three instruction sequences.
3934 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3935 if (SextVal >= 0 && SextVal <= 31) {
3936 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG, dl);
3937 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3938 LHS = DAG.getNode(ISD::SUB, dl, LHS.getValueType(), LHS, RHS);
3939 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3941 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3942 if (SextVal >= -31 && SextVal <= 0) {
3943 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG, dl);
3944 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG, dl);
3945 LHS = DAG.getNode(ISD::ADD, dl, LHS.getValueType(), LHS, RHS);
3946 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), LHS);
3952 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3953 /// the specified operations to build the shuffle.
3954 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3955 SDValue RHS, SelectionDAG &DAG,
3957 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3958 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3959 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3962 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3974 if (OpNum == OP_COPY) {
3975 if (LHSID == (1*9+2)*9+3) return LHS;
3976 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3980 SDValue OpLHS, OpRHS;
3981 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
3982 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
3986 default: assert(0 && "Unknown i32 permute!");
3988 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3989 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3990 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3991 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3994 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3995 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3996 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3997 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
4000 for (unsigned i = 0; i != 16; ++i)
4001 ShufIdxs[i] = (i&3)+0;
4004 for (unsigned i = 0; i != 16; ++i)
4005 ShufIdxs[i] = (i&3)+4;
4008 for (unsigned i = 0; i != 16; ++i)
4009 ShufIdxs[i] = (i&3)+8;
4012 for (unsigned i = 0; i != 16; ++i)
4013 ShufIdxs[i] = (i&3)+12;
4016 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl);
4018 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl);
4020 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl);
4022 MVT VT = OpLHS.getValueType();
4023 OpLHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpLHS);
4024 OpRHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OpRHS);
4025 SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs);
4026 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, T);
4029 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
4030 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
4031 /// return the code it can be lowered into. Worst case, it can always be
4032 /// lowered into a vperm.
4033 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4034 SelectionDAG &DAG) {
4035 DebugLoc dl = Op.getDebugLoc();
4036 SDValue V1 = Op.getOperand(0);
4037 SDValue V2 = Op.getOperand(1);
4038 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4039 MVT VT = Op.getValueType();
4041 // Cases that are handled by instructions that take permute immediates
4042 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
4043 // selected by the instruction selector.
4044 if (V2.getOpcode() == ISD::UNDEF) {
4045 if (PPC::isSplatShuffleMask(SVOp, 1) ||
4046 PPC::isSplatShuffleMask(SVOp, 2) ||
4047 PPC::isSplatShuffleMask(SVOp, 4) ||
4048 PPC::isVPKUWUMShuffleMask(SVOp, true) ||
4049 PPC::isVPKUHUMShuffleMask(SVOp, true) ||
4050 PPC::isVSLDOIShuffleMask(SVOp, true) != -1 ||
4051 PPC::isVMRGLShuffleMask(SVOp, 1, true) ||
4052 PPC::isVMRGLShuffleMask(SVOp, 2, true) ||
4053 PPC::isVMRGLShuffleMask(SVOp, 4, true) ||
4054 PPC::isVMRGHShuffleMask(SVOp, 1, true) ||
4055 PPC::isVMRGHShuffleMask(SVOp, 2, true) ||
4056 PPC::isVMRGHShuffleMask(SVOp, 4, true)) {
4061 // Altivec has a variety of "shuffle immediates" that take two vector inputs
4062 // and produce a fixed permutation. If any of these match, do not lower to
4064 if (PPC::isVPKUWUMShuffleMask(SVOp, false) ||
4065 PPC::isVPKUHUMShuffleMask(SVOp, false) ||
4066 PPC::isVSLDOIShuffleMask(SVOp, false) != -1 ||
4067 PPC::isVMRGLShuffleMask(SVOp, 1, false) ||
4068 PPC::isVMRGLShuffleMask(SVOp, 2, false) ||
4069 PPC::isVMRGLShuffleMask(SVOp, 4, false) ||
4070 PPC::isVMRGHShuffleMask(SVOp, 1, false) ||
4071 PPC::isVMRGHShuffleMask(SVOp, 2, false) ||
4072 PPC::isVMRGHShuffleMask(SVOp, 4, false))
4075 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
4076 // perfect shuffle table to emit an optimal matching sequence.
4077 SmallVector<int, 16> PermMask;
4078 SVOp->getMask(PermMask);
4080 unsigned PFIndexes[4];
4081 bool isFourElementShuffle = true;
4082 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
4083 unsigned EltNo = 8; // Start out undef.
4084 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
4085 if (PermMask[i*4+j] < 0)
4086 continue; // Undef, ignore it.
4088 unsigned ByteSource = PermMask[i*4+j];
4089 if ((ByteSource & 3) != j) {
4090 isFourElementShuffle = false;
4095 EltNo = ByteSource/4;
4096 } else if (EltNo != ByteSource/4) {
4097 isFourElementShuffle = false;
4101 PFIndexes[i] = EltNo;
4104 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
4105 // perfect shuffle vector to determine if it is cost effective to do this as
4106 // discrete instructions, or whether we should use a vperm.
4107 if (isFourElementShuffle) {
4108 // Compute the index in the perfect shuffle table.
4109 unsigned PFTableIndex =
4110 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4112 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4113 unsigned Cost = (PFEntry >> 30);
4115 // Determining when to avoid vperm is tricky. Many things affect the cost
4116 // of vperm, particularly how many times the perm mask needs to be computed.
4117 // For example, if the perm mask can be hoisted out of a loop or is already
4118 // used (perhaps because there are multiple permutes with the same shuffle
4119 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
4120 // the loop requires an extra register.
4122 // As a compromise, we only emit discrete instructions if the shuffle can be
4123 // generated in 3 or fewer operations. When we have loop information
4124 // available, if this block is within a loop, we should avoid using vperm
4125 // for 3-operation perms and use a constant pool load instead.
4127 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4130 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
4131 // vector that will get spilled to the constant pool.
4132 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
4134 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
4135 // that it is in input element units, not in bytes. Convert now.
4136 MVT EltVT = V1.getValueType().getVectorElementType();
4137 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
4139 SmallVector<SDValue, 16> ResultMask;
4140 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
4141 unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i];
4143 for (unsigned j = 0; j != BytesPerElement; ++j)
4144 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
4148 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
4149 &ResultMask[0], ResultMask.size());
4150 return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), V1, V2, VPermMask);
4153 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
4154 /// altivec comparison. If it is, return true and fill in Opc/isDot with
4155 /// information about the intrinsic.
4156 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
4158 unsigned IntrinsicID =
4159 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
4162 switch (IntrinsicID) {
4163 default: return false;
4164 // Comparison predicates.
4165 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
4166 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
4167 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
4168 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
4169 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
4170 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
4171 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
4172 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
4173 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
4174 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
4175 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
4176 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
4177 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
4179 // Normal Comparisons.
4180 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
4181 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
4182 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
4183 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
4184 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
4185 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
4186 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
4187 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
4188 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
4189 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
4190 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
4191 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
4192 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
4197 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
4198 /// lower, do it, otherwise return null.
4199 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
4200 SelectionDAG &DAG) {
4201 // If this is a lowered altivec predicate compare, CompareOpc is set to the
4202 // opcode number of the comparison.
4203 DebugLoc dl = Op.getDebugLoc();
4206 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
4207 return SDValue(); // Don't custom lower most intrinsics.
4209 // If this is a non-dot comparison, make the VCMP node and we are done.
4211 SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(),
4212 Op.getOperand(1), Op.getOperand(2),
4213 DAG.getConstant(CompareOpc, MVT::i32));
4214 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Tmp);
4217 // Create the PPCISD altivec 'dot' comparison node.
4219 Op.getOperand(2), // LHS
4220 Op.getOperand(3), // RHS
4221 DAG.getConstant(CompareOpc, MVT::i32)
4223 std::vector<MVT> VTs;
4224 VTs.push_back(Op.getOperand(2).getValueType());
4225 VTs.push_back(MVT::Flag);
4226 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
4228 // Now that we have the comparison, emit a copy from the CR to a GPR.
4229 // This is flagged to the above dot comparison.
4230 SDValue Flags = DAG.getNode(PPCISD::MFCR, dl, MVT::i32,
4231 DAG.getRegister(PPC::CR6, MVT::i32),
4232 CompNode.getValue(1));
4234 // Unpack the result based on how the target uses it.
4235 unsigned BitNo; // Bit # of CR6.
4236 bool InvertBit; // Invert result?
4237 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
4238 default: // Can't happen, don't crash on invalid number though.
4239 case 0: // Return the value of the EQ bit of CR6.
4240 BitNo = 0; InvertBit = false;
4242 case 1: // Return the inverted value of the EQ bit of CR6.
4243 BitNo = 0; InvertBit = true;
4245 case 2: // Return the value of the LT bit of CR6.
4246 BitNo = 2; InvertBit = false;
4248 case 3: // Return the inverted value of the LT bit of CR6.
4249 BitNo = 2; InvertBit = true;
4253 // Shift the bit into the low position.
4254 Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags,
4255 DAG.getConstant(8-(3-BitNo), MVT::i32));
4257 Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags,
4258 DAG.getConstant(1, MVT::i32));
4260 // If we are supposed to, toggle the bit.
4262 Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags,
4263 DAG.getConstant(1, MVT::i32));
4267 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
4268 SelectionDAG &DAG) {
4269 DebugLoc dl = Op.getDebugLoc();
4270 // Create a stack slot that is 16-byte aligned.
4271 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
4272 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
4273 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4274 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
4276 // Store the input value into Value#0 of the stack slot.
4277 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl,
4278 Op.getOperand(0), FIdx, NULL, 0);
4280 return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, NULL, 0);
4283 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
4284 DebugLoc dl = Op.getDebugLoc();
4285 if (Op.getValueType() == MVT::v4i32) {
4286 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4288 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG, dl);
4289 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG, dl);//+16 as shift amt.
4291 SDValue RHSSwap = // = vrlw RHS, 16
4292 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl);
4294 // Shrinkify inputs to v8i16.
4295 LHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, LHS);
4296 RHS = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHS);
4297 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, RHSSwap);
4299 // Low parts multiplied together, generating 32-bit results (we ignore the
4301 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
4302 LHS, RHS, DAG, dl, MVT::v4i32);
4304 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
4305 LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32);
4306 // Shift the high parts up 16 bits.
4307 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd,
4309 return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd);
4310 } else if (Op.getValueType() == MVT::v8i16) {
4311 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4313 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG, dl);
4315 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
4316 LHS, RHS, Zero, DAG, dl);
4317 } else if (Op.getValueType() == MVT::v16i8) {
4318 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
4320 // Multiply the even 8-bit parts, producing 16-bit sums.
4321 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
4322 LHS, RHS, DAG, dl, MVT::v8i16);
4323 EvenParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, EvenParts);
4325 // Multiply the odd 8-bit parts, producing 16-bit sums.
4326 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
4327 LHS, RHS, DAG, dl, MVT::v8i16);
4328 OddParts = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, OddParts);
4330 // Merge the results together.
4332 for (unsigned i = 0; i != 8; ++i) {
4334 Ops[i*2+1] = 2*i+1+16;
4336 return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops);
4338 assert(0 && "Unknown mul to lower!");
4343 /// LowerOperation - Provide custom lowering hooks for some operations.
4345 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
4346 switch (Op.getOpcode()) {
4347 default: assert(0 && "Wasn't expecting to be able to lower this!");
4348 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
4349 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
4350 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4351 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
4352 case ISD::SETCC: return LowerSETCC(Op, DAG);
4353 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
4355 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
4356 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
4359 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
4360 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
4362 case ISD::FORMAL_ARGUMENTS:
4363 if (PPCSubTarget.isELF32_ABI()) {
4364 return LowerFORMAL_ARGUMENTS_SVR4(Op, DAG, VarArgsFrameIndex,
4365 VarArgsStackOffset, VarArgsNumGPR,
4366 VarArgsNumFPR, PPCSubTarget);
4368 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
4369 VarArgsStackOffset, VarArgsNumGPR,
4370 VarArgsNumFPR, PPCSubTarget);
4374 if (PPCSubTarget.isELF32_ABI()) {
4375 return LowerCALL_SVR4(Op, DAG, PPCSubTarget, getTargetMachine());
4377 return LowerCALL(Op, DAG, PPCSubTarget, getTargetMachine());
4380 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
4381 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
4382 case ISD::DYNAMIC_STACKALLOC:
4383 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
4385 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
4386 case ISD::FP_TO_UINT:
4387 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG,
4389 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
4390 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
4392 // Lower 64-bit shifts.
4393 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
4394 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
4395 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
4397 // Vector-related lowering.
4398 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
4399 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4400 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
4401 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
4402 case ISD::MUL: return LowerMUL(Op, DAG);
4404 // Frame & Return address.
4405 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
4406 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
4411 void PPCTargetLowering::ReplaceNodeResults(SDNode *N,
4412 SmallVectorImpl<SDValue>&Results,
4413 SelectionDAG &DAG) {
4414 DebugLoc dl = N->getDebugLoc();
4415 switch (N->getOpcode()) {
4417 assert(false && "Do not know how to custom type legalize this operation!");
4419 case ISD::FP_ROUND_INREG: {
4420 assert(N->getValueType(0) == MVT::ppcf128);
4421 assert(N->getOperand(0).getValueType() == MVT::ppcf128);
4422 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4423 MVT::f64, N->getOperand(0),
4424 DAG.getIntPtrConstant(0));
4425 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
4426 MVT::f64, N->getOperand(0),
4427 DAG.getIntPtrConstant(1));
4429 // This sequence changes FPSCR to do round-to-zero, adds the two halves
4430 // of the long double, and puts FPSCR back the way it was. We do not
4431 // actually model FPSCR.
4432 std::vector<MVT> NodeTys;
4433 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
4435 NodeTys.push_back(MVT::f64); // Return register
4436 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
4437 Result = DAG.getNode(PPCISD::MFFS, dl, NodeTys, &InFlag, 0);
4438 MFFSreg = Result.getValue(0);
4439 InFlag = Result.getValue(1);
4442 NodeTys.push_back(MVT::Flag); // Returns a flag
4443 Ops[0] = DAG.getConstant(31, MVT::i32);
4445 Result = DAG.getNode(PPCISD::MTFSB1, dl, NodeTys, Ops, 2);
4446 InFlag = Result.getValue(0);
4449 NodeTys.push_back(MVT::Flag); // Returns a flag
4450 Ops[0] = DAG.getConstant(30, MVT::i32);
4452 Result = DAG.getNode(PPCISD::MTFSB0, dl, NodeTys, Ops, 2);
4453 InFlag = Result.getValue(0);
4456 NodeTys.push_back(MVT::f64); // result of add
4457 NodeTys.push_back(MVT::Flag); // Returns a flag
4461 Result = DAG.getNode(PPCISD::FADDRTZ, dl, NodeTys, Ops, 3);
4462 FPreg = Result.getValue(0);
4463 InFlag = Result.getValue(1);
4466 NodeTys.push_back(MVT::f64);
4467 Ops[0] = DAG.getConstant(1, MVT::i32);
4471 Result = DAG.getNode(PPCISD::MTFSF, dl, NodeTys, Ops, 4);
4472 FPreg = Result.getValue(0);
4474 // We know the low half is about to be thrown away, so just use something
4476 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::ppcf128,
4480 case ISD::FP_TO_SINT:
4481 Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl));
4487 //===----------------------------------------------------------------------===//
4488 // Other Lowering Code
4489 //===----------------------------------------------------------------------===//
4492 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
4493 bool is64bit, unsigned BinOpcode) const {
4494 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4495 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4497 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4498 MachineFunction *F = BB->getParent();
4499 MachineFunction::iterator It = BB;
4502 unsigned dest = MI->getOperand(0).getReg();
4503 unsigned ptrA = MI->getOperand(1).getReg();
4504 unsigned ptrB = MI->getOperand(2).getReg();
4505 unsigned incr = MI->getOperand(3).getReg();
4506 DebugLoc dl = MI->getDebugLoc();
4508 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4509 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4510 F->insert(It, loopMBB);
4511 F->insert(It, exitMBB);
4512 exitMBB->transferSuccessors(BB);
4514 MachineRegisterInfo &RegInfo = F->getRegInfo();
4515 unsigned TmpReg = (!BinOpcode) ? incr :
4516 RegInfo.createVirtualRegister(
4517 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4518 (const TargetRegisterClass *) &PPC::GPRCRegClass);
4522 // fallthrough --> loopMBB
4523 BB->addSuccessor(loopMBB);
4526 // l[wd]arx dest, ptr
4527 // add r0, dest, incr
4528 // st[wd]cx. r0, ptr
4530 // fallthrough --> exitMBB
4532 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4533 .addReg(ptrA).addReg(ptrB);
4535 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
4536 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4537 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
4538 BuildMI(BB, dl, TII->get(PPC::BCC))
4539 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4540 BB->addSuccessor(loopMBB);
4541 BB->addSuccessor(exitMBB);
4550 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
4551 MachineBasicBlock *BB,
4552 bool is8bit, // operation
4553 unsigned BinOpcode) const {
4554 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
4555 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4556 // In 64 bit mode we have to use 64 bits for addresses, even though the
4557 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
4558 // registers without caring whether they're 32 or 64, but here we're
4559 // doing actual arithmetic on the addresses.
4560 bool is64bit = PPCSubTarget.isPPC64();
4562 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4563 MachineFunction *F = BB->getParent();
4564 MachineFunction::iterator It = BB;
4567 unsigned dest = MI->getOperand(0).getReg();
4568 unsigned ptrA = MI->getOperand(1).getReg();
4569 unsigned ptrB = MI->getOperand(2).getReg();
4570 unsigned incr = MI->getOperand(3).getReg();
4571 DebugLoc dl = MI->getDebugLoc();
4573 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
4574 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4575 F->insert(It, loopMBB);
4576 F->insert(It, exitMBB);
4577 exitMBB->transferSuccessors(BB);
4579 MachineRegisterInfo &RegInfo = F->getRegInfo();
4580 const TargetRegisterClass *RC =
4581 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4582 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4583 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4584 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4585 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4586 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4587 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4588 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4589 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4590 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4591 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4592 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4593 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4595 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4599 // fallthrough --> loopMBB
4600 BB->addSuccessor(loopMBB);
4602 // The 4-byte load must be aligned, while a char or short may be
4603 // anywhere in the word. Hence all this nasty bookkeeping code.
4604 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4605 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4606 // xori shift, shift1, 24 [16]
4607 // rlwinm ptr, ptr1, 0, 0, 29
4608 // slw incr2, incr, shift
4609 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4610 // slw mask, mask2, shift
4612 // lwarx tmpDest, ptr
4613 // add tmp, tmpDest, incr2
4614 // andc tmp2, tmpDest, mask
4615 // and tmp3, tmp, mask
4616 // or tmp4, tmp3, tmp2
4619 // fallthrough --> exitMBB
4620 // srw dest, tmpDest, shift
4622 if (ptrA!=PPC::R0) {
4623 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4624 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4625 .addReg(ptrA).addReg(ptrB);
4629 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4630 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4631 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4632 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4634 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4635 .addReg(Ptr1Reg).addImm(0).addImm(61);
4637 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4638 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4639 BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg)
4640 .addReg(incr).addReg(ShiftReg);
4642 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4644 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4645 BuildMI(BB, dl, TII->get(PPC::ORI),Mask2Reg).addReg(Mask3Reg).addImm(65535);
4647 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4648 .addReg(Mask2Reg).addReg(ShiftReg);
4651 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4652 .addReg(PPC::R0).addReg(PtrReg);
4654 BuildMI(BB, dl, TII->get(BinOpcode), TmpReg)
4655 .addReg(Incr2Reg).addReg(TmpDestReg);
4656 BuildMI(BB, dl, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4657 .addReg(TmpDestReg).addReg(MaskReg);
4658 BuildMI(BB, dl, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4659 .addReg(TmpReg).addReg(MaskReg);
4660 BuildMI(BB, dl, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4661 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4662 BuildMI(BB, dl, TII->get(PPC::STWCX))
4663 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4664 BuildMI(BB, dl, TII->get(PPC::BCC))
4665 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4666 BB->addSuccessor(loopMBB);
4667 BB->addSuccessor(exitMBB);
4672 BuildMI(BB, dl, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4677 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4678 MachineBasicBlock *BB) const {
4679 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4681 // To "insert" these instructions we actually have to insert their
4682 // control-flow patterns.
4683 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4684 MachineFunction::iterator It = BB;
4687 MachineFunction *F = BB->getParent();
4689 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4690 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4691 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4692 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4693 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4695 // The incoming instruction knows the destination vreg to set, the
4696 // condition code register to branch on, the true/false values to
4697 // select between, and a branch opcode to use.
4702 // cmpTY ccX, r1, r2
4704 // fallthrough --> copy0MBB
4705 MachineBasicBlock *thisMBB = BB;
4706 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4707 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4708 unsigned SelectPred = MI->getOperand(4).getImm();
4709 DebugLoc dl = MI->getDebugLoc();
4710 BuildMI(BB, dl, TII->get(PPC::BCC))
4711 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4712 F->insert(It, copy0MBB);
4713 F->insert(It, sinkMBB);
4714 // Update machine-CFG edges by transferring all successors of the current
4715 // block to the new block which will contain the Phi node for the select.
4716 sinkMBB->transferSuccessors(BB);
4717 // Next, add the true and fallthrough blocks as its successors.
4718 BB->addSuccessor(copy0MBB);
4719 BB->addSuccessor(sinkMBB);
4722 // %FalseValue = ...
4723 // # fallthrough to sinkMBB
4726 // Update machine-CFG edges
4727 BB->addSuccessor(sinkMBB);
4730 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4733 BuildMI(BB, dl, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4734 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4735 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4737 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4738 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4739 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4740 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4741 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4742 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4743 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4744 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4746 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4747 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4748 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4749 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4750 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4751 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4752 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4753 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4755 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4756 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4757 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4758 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4759 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4760 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4761 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4762 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4764 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4765 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4766 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4767 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4768 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4769 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4770 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4771 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4773 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4774 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4775 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4776 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4777 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4778 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4779 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4780 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4782 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4783 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4784 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4785 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4786 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4787 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4788 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4789 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4791 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4792 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4793 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4794 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4795 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4796 BB = EmitAtomicBinary(MI, BB, false, 0);
4797 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4798 BB = EmitAtomicBinary(MI, BB, true, 0);
4800 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4801 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4802 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4804 unsigned dest = MI->getOperand(0).getReg();
4805 unsigned ptrA = MI->getOperand(1).getReg();
4806 unsigned ptrB = MI->getOperand(2).getReg();
4807 unsigned oldval = MI->getOperand(3).getReg();
4808 unsigned newval = MI->getOperand(4).getReg();
4809 DebugLoc dl = MI->getDebugLoc();
4811 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4812 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4813 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4814 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4815 F->insert(It, loop1MBB);
4816 F->insert(It, loop2MBB);
4817 F->insert(It, midMBB);
4818 F->insert(It, exitMBB);
4819 exitMBB->transferSuccessors(BB);
4823 // fallthrough --> loopMBB
4824 BB->addSuccessor(loop1MBB);
4827 // l[wd]arx dest, ptr
4828 // cmp[wd] dest, oldval
4831 // st[wd]cx. newval, ptr
4835 // st[wd]cx. dest, ptr
4838 BuildMI(BB, dl, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4839 .addReg(ptrA).addReg(ptrB);
4840 BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4841 .addReg(oldval).addReg(dest);
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(is64bit ? PPC::STDCX : PPC::STWCX))
4849 .addReg(newval).addReg(ptrA).addReg(ptrB);
4850 BuildMI(BB, dl, TII->get(PPC::BCC))
4851 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4852 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4853 BB->addSuccessor(loop1MBB);
4854 BB->addSuccessor(exitMBB);
4857 BuildMI(BB, dl, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4858 .addReg(dest).addReg(ptrA).addReg(ptrB);
4859 BB->addSuccessor(exitMBB);
4864 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4865 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4866 // We must use 64-bit registers for addresses when targeting 64-bit,
4867 // since we're actually doing arithmetic on them. Other registers
4869 bool is64bit = PPCSubTarget.isPPC64();
4870 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4872 unsigned dest = MI->getOperand(0).getReg();
4873 unsigned ptrA = MI->getOperand(1).getReg();
4874 unsigned ptrB = MI->getOperand(2).getReg();
4875 unsigned oldval = MI->getOperand(3).getReg();
4876 unsigned newval = MI->getOperand(4).getReg();
4877 DebugLoc dl = MI->getDebugLoc();
4879 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4880 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4881 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4882 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4883 F->insert(It, loop1MBB);
4884 F->insert(It, loop2MBB);
4885 F->insert(It, midMBB);
4886 F->insert(It, exitMBB);
4887 exitMBB->transferSuccessors(BB);
4889 MachineRegisterInfo &RegInfo = F->getRegInfo();
4890 const TargetRegisterClass *RC =
4891 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4892 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4893 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4894 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4895 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4896 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4897 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4898 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4899 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4900 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4901 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4902 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4903 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4904 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4905 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4907 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4910 // fallthrough --> loopMBB
4911 BB->addSuccessor(loop1MBB);
4913 // The 4-byte load must be aligned, while a char or short may be
4914 // anywhere in the word. Hence all this nasty bookkeeping code.
4915 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4916 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4917 // xori shift, shift1, 24 [16]
4918 // rlwinm ptr, ptr1, 0, 0, 29
4919 // slw newval2, newval, shift
4920 // slw oldval2, oldval,shift
4921 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4922 // slw mask, mask2, shift
4923 // and newval3, newval2, mask
4924 // and oldval3, oldval2, mask
4926 // lwarx tmpDest, ptr
4927 // and tmp, tmpDest, mask
4928 // cmpw tmp, oldval3
4931 // andc tmp2, tmpDest, mask
4932 // or tmp4, tmp2, newval3
4937 // stwcx. tmpDest, ptr
4939 // srw dest, tmpDest, shift
4940 if (ptrA!=PPC::R0) {
4941 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4942 BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4943 .addReg(ptrA).addReg(ptrB);
4947 BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4948 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4949 BuildMI(BB, dl, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4950 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4952 BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg)
4953 .addReg(Ptr1Reg).addImm(0).addImm(61);
4955 BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg)
4956 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4957 BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg)
4958 .addReg(newval).addReg(ShiftReg);
4959 BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg)
4960 .addReg(oldval).addReg(ShiftReg);
4962 BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255);
4964 BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0);
4965 BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg)
4966 .addReg(Mask3Reg).addImm(65535);
4968 BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg)
4969 .addReg(Mask2Reg).addReg(ShiftReg);
4970 BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg)
4971 .addReg(NewVal2Reg).addReg(MaskReg);
4972 BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg)
4973 .addReg(OldVal2Reg).addReg(MaskReg);
4976 BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg)
4977 .addReg(PPC::R0).addReg(PtrReg);
4978 BuildMI(BB, dl, TII->get(PPC::AND),TmpReg)
4979 .addReg(TmpDestReg).addReg(MaskReg);
4980 BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0)
4981 .addReg(TmpReg).addReg(OldVal3Reg);
4982 BuildMI(BB, dl, TII->get(PPC::BCC))
4983 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4984 BB->addSuccessor(loop2MBB);
4985 BB->addSuccessor(midMBB);
4988 BuildMI(BB, dl, TII->get(PPC::ANDC),Tmp2Reg)
4989 .addReg(TmpDestReg).addReg(MaskReg);
4990 BuildMI(BB, dl, TII->get(PPC::OR),Tmp4Reg)
4991 .addReg(Tmp2Reg).addReg(NewVal3Reg);
4992 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4993 .addReg(PPC::R0).addReg(PtrReg);
4994 BuildMI(BB, dl, TII->get(PPC::BCC))
4995 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4996 BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB);
4997 BB->addSuccessor(loop1MBB);
4998 BB->addSuccessor(exitMBB);
5001 BuildMI(BB, dl, TII->get(PPC::STWCX)).addReg(TmpDestReg)
5002 .addReg(PPC::R0).addReg(PtrReg);
5003 BB->addSuccessor(exitMBB);
5008 BuildMI(BB, dl, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
5010 assert(0 && "Unexpected instr type to insert");
5013 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
5017 //===----------------------------------------------------------------------===//
5018 // Target Optimization Hooks
5019 //===----------------------------------------------------------------------===//
5021 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
5022 DAGCombinerInfo &DCI) const {
5023 TargetMachine &TM = getTargetMachine();
5024 SelectionDAG &DAG = DCI.DAG;
5025 DebugLoc dl = N->getDebugLoc();
5026 switch (N->getOpcode()) {
5029 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5030 if (C->getZExtValue() == 0) // 0 << V -> 0.
5031 return N->getOperand(0);
5035 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5036 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
5037 return N->getOperand(0);
5041 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
5042 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
5043 C->isAllOnesValue()) // -1 >>s V -> -1.
5044 return N->getOperand(0);
5048 case ISD::SINT_TO_FP:
5049 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
5050 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
5051 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
5052 // We allow the src/dst to be either f32/f64, but the intermediate
5053 // type must be i64.
5054 if (N->getOperand(0).getValueType() == MVT::i64 &&
5055 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
5056 SDValue Val = N->getOperand(0).getOperand(0);
5057 if (Val.getValueType() == MVT::f32) {
5058 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5059 DCI.AddToWorklist(Val.getNode());
5062 Val = DAG.getNode(PPCISD::FCTIDZ, dl, MVT::f64, Val);
5063 DCI.AddToWorklist(Val.getNode());
5064 Val = DAG.getNode(PPCISD::FCFID, dl, MVT::f64, Val);
5065 DCI.AddToWorklist(Val.getNode());
5066 if (N->getValueType(0) == MVT::f32) {
5067 Val = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Val,
5068 DAG.getIntPtrConstant(0));
5069 DCI.AddToWorklist(Val.getNode());
5072 } else if (N->getOperand(0).getValueType() == MVT::i32) {
5073 // If the intermediate type is i32, we can avoid the load/store here
5080 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
5081 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
5082 !cast<StoreSDNode>(N)->isTruncatingStore() &&
5083 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
5084 N->getOperand(1).getValueType() == MVT::i32 &&
5085 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
5086 SDValue Val = N->getOperand(1).getOperand(0);
5087 if (Val.getValueType() == MVT::f32) {
5088 Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val);
5089 DCI.AddToWorklist(Val.getNode());
5091 Val = DAG.getNode(PPCISD::FCTIWZ, dl, MVT::f64, Val);
5092 DCI.AddToWorklist(Val.getNode());
5094 Val = DAG.getNode(PPCISD::STFIWX, dl, MVT::Other, N->getOperand(0), Val,
5095 N->getOperand(2), N->getOperand(3));
5096 DCI.AddToWorklist(Val.getNode());
5100 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
5101 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
5102 N->getOperand(1).getNode()->hasOneUse() &&
5103 (N->getOperand(1).getValueType() == MVT::i32 ||
5104 N->getOperand(1).getValueType() == MVT::i16)) {
5105 SDValue BSwapOp = N->getOperand(1).getOperand(0);
5106 // Do an any-extend to 32-bits if this is a half-word input.
5107 if (BSwapOp.getValueType() == MVT::i16)
5108 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
5110 return DAG.getNode(PPCISD::STBRX, dl, MVT::Other, N->getOperand(0),
5111 BSwapOp, N->getOperand(2), N->getOperand(3),
5112 DAG.getValueType(N->getOperand(1).getValueType()));
5116 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
5117 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
5118 N->getOperand(0).hasOneUse() &&
5119 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
5120 SDValue Load = N->getOperand(0);
5121 LoadSDNode *LD = cast<LoadSDNode>(Load);
5122 // Create the byte-swapping load.
5123 std::vector<MVT> VTs;
5124 VTs.push_back(MVT::i32);
5125 VTs.push_back(MVT::Other);
5126 SDValue MO = DAG.getMemOperand(LD->getMemOperand());
5128 LD->getChain(), // Chain
5129 LD->getBasePtr(), // Ptr
5131 DAG.getValueType(N->getValueType(0)) // VT
5133 SDValue BSLoad = DAG.getNode(PPCISD::LBRX, dl, VTs, Ops, 4);
5135 // If this is an i16 load, insert the truncate.
5136 SDValue ResVal = BSLoad;
5137 if (N->getValueType(0) == MVT::i16)
5138 ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad);
5140 // First, combine the bswap away. This makes the value produced by the
5142 DCI.CombineTo(N, ResVal);
5144 // Next, combine the load away, we give it a bogus result value but a real
5145 // chain result. The result value is dead because the bswap is dead.
5146 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
5148 // Return N so it doesn't get rechecked!
5149 return SDValue(N, 0);
5153 case PPCISD::VCMP: {
5154 // If a VCMPo node already exists with exactly the same operands as this
5155 // node, use its result instead of this node (VCMPo computes both a CR6 and
5156 // a normal output).
5158 if (!N->getOperand(0).hasOneUse() &&
5159 !N->getOperand(1).hasOneUse() &&
5160 !N->getOperand(2).hasOneUse()) {
5162 // Scan all of the users of the LHS, looking for VCMPo's that match.
5163 SDNode *VCMPoNode = 0;
5165 SDNode *LHSN = N->getOperand(0).getNode();
5166 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
5168 if (UI->getOpcode() == PPCISD::VCMPo &&
5169 UI->getOperand(1) == N->getOperand(1) &&
5170 UI->getOperand(2) == N->getOperand(2) &&
5171 UI->getOperand(0) == N->getOperand(0)) {
5176 // If there is no VCMPo node, or if the flag value has a single use, don't
5178 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
5181 // Look at the (necessarily single) use of the flag value. If it has a
5182 // chain, this transformation is more complex. Note that multiple things
5183 // could use the value result, which we should ignore.
5184 SDNode *FlagUser = 0;
5185 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
5186 FlagUser == 0; ++UI) {
5187 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
5189 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
5190 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
5197 // If the user is a MFCR instruction, we know this is safe. Otherwise we
5198 // give up for right now.
5199 if (FlagUser->getOpcode() == PPCISD::MFCR)
5200 return SDValue(VCMPoNode, 0);
5205 // If this is a branch on an altivec predicate comparison, lower this so
5206 // that we don't have to do a MFCR: instead, branch directly on CR6. This
5207 // lowering is done pre-legalize, because the legalizer lowers the predicate
5208 // compare down to code that is difficult to reassemble.
5209 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
5210 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
5214 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
5215 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
5216 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
5217 assert(isDot && "Can't compare against a vector result!");
5219 // If this is a comparison against something other than 0/1, then we know
5220 // that the condition is never/always true.
5221 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
5222 if (Val != 0 && Val != 1) {
5223 if (CC == ISD::SETEQ) // Cond never true, remove branch.
5224 return N->getOperand(0);
5225 // Always !=, turn it into an unconditional branch.
5226 return DAG.getNode(ISD::BR, dl, MVT::Other,
5227 N->getOperand(0), N->getOperand(4));
5230 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
5232 // Create the PPCISD altivec 'dot' comparison node.
5233 std::vector<MVT> VTs;
5235 LHS.getOperand(2), // LHS of compare
5236 LHS.getOperand(3), // RHS of compare
5237 DAG.getConstant(CompareOpc, MVT::i32)
5239 VTs.push_back(LHS.getOperand(2).getValueType());
5240 VTs.push_back(MVT::Flag);
5241 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops, 3);
5243 // Unpack the result based on how the target uses it.
5244 PPC::Predicate CompOpc;
5245 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
5246 default: // Can't happen, don't crash on invalid number though.
5247 case 0: // Branch on the value of the EQ bit of CR6.
5248 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
5250 case 1: // Branch on the inverted value of the EQ bit of CR6.
5251 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
5253 case 2: // Branch on the value of the LT bit of CR6.
5254 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
5256 case 3: // Branch on the inverted value of the LT bit of CR6.
5257 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
5261 return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0),
5262 DAG.getConstant(CompOpc, MVT::i32),
5263 DAG.getRegister(PPC::CR6, MVT::i32),
5264 N->getOperand(4), CompNode.getValue(1));
5273 //===----------------------------------------------------------------------===//
5274 // Inline Assembly Support
5275 //===----------------------------------------------------------------------===//
5277 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
5281 const SelectionDAG &DAG,
5282 unsigned Depth) const {
5283 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
5284 switch (Op.getOpcode()) {
5286 case PPCISD::LBRX: {
5287 // lhbrx is known to have the top bits cleared out.
5288 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
5289 KnownZero = 0xFFFF0000;
5292 case ISD::INTRINSIC_WO_CHAIN: {
5293 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
5295 case Intrinsic::ppc_altivec_vcmpbfp_p:
5296 case Intrinsic::ppc_altivec_vcmpeqfp_p:
5297 case Intrinsic::ppc_altivec_vcmpequb_p:
5298 case Intrinsic::ppc_altivec_vcmpequh_p:
5299 case Intrinsic::ppc_altivec_vcmpequw_p:
5300 case Intrinsic::ppc_altivec_vcmpgefp_p:
5301 case Intrinsic::ppc_altivec_vcmpgtfp_p:
5302 case Intrinsic::ppc_altivec_vcmpgtsb_p:
5303 case Intrinsic::ppc_altivec_vcmpgtsh_p:
5304 case Intrinsic::ppc_altivec_vcmpgtsw_p:
5305 case Intrinsic::ppc_altivec_vcmpgtub_p:
5306 case Intrinsic::ppc_altivec_vcmpgtuh_p:
5307 case Intrinsic::ppc_altivec_vcmpgtuw_p:
5308 KnownZero = ~1U; // All bits but the low one are known to be zero.
5316 /// getConstraintType - Given a constraint, return the type of
5317 /// constraint it is for this target.
5318 PPCTargetLowering::ConstraintType
5319 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
5320 if (Constraint.size() == 1) {
5321 switch (Constraint[0]) {
5328 return C_RegisterClass;
5331 return TargetLowering::getConstraintType(Constraint);
5334 std::pair<unsigned, const TargetRegisterClass*>
5335 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
5337 if (Constraint.size() == 1) {
5338 // GCC RS6000 Constraint Letters
5339 switch (Constraint[0]) {
5342 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
5343 return std::make_pair(0U, PPC::G8RCRegisterClass);
5344 return std::make_pair(0U, PPC::GPRCRegisterClass);
5347 return std::make_pair(0U, PPC::F4RCRegisterClass);
5348 else if (VT == MVT::f64)
5349 return std::make_pair(0U, PPC::F8RCRegisterClass);
5352 return std::make_pair(0U, PPC::VRRCRegisterClass);
5354 return std::make_pair(0U, PPC::CRRCRegisterClass);
5358 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5362 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
5363 /// vector. If it is invalid, don't add anything to Ops. If hasMemory is true
5364 /// it means one of the asm constraint of the inline asm instruction being
5365 /// processed is 'm'.
5366 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
5368 std::vector<SDValue>&Ops,
5369 SelectionDAG &DAG) const {
5370 SDValue Result(0,0);
5381 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
5382 if (!CST) return; // Must be an immediate to match.
5383 unsigned Value = CST->getZExtValue();
5385 default: assert(0 && "Unknown constraint letter!");
5386 case 'I': // "I" is a signed 16-bit constant.
5387 if ((short)Value == (int)Value)
5388 Result = DAG.getTargetConstant(Value, Op.getValueType());
5390 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
5391 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
5392 if ((short)Value == 0)
5393 Result = DAG.getTargetConstant(Value, Op.getValueType());
5395 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
5396 if ((Value >> 16) == 0)
5397 Result = DAG.getTargetConstant(Value, Op.getValueType());
5399 case 'M': // "M" is a constant that is greater than 31.
5401 Result = DAG.getTargetConstant(Value, Op.getValueType());
5403 case 'N': // "N" is a positive constant that is an exact power of two.
5404 if ((int)Value > 0 && isPowerOf2_32(Value))
5405 Result = DAG.getTargetConstant(Value, Op.getValueType());
5407 case 'O': // "O" is the constant zero.
5409 Result = DAG.getTargetConstant(Value, Op.getValueType());
5411 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
5412 if ((short)-Value == (int)-Value)
5413 Result = DAG.getTargetConstant(Value, Op.getValueType());
5420 if (Result.getNode()) {
5421 Ops.push_back(Result);
5425 // Handle standard constraint letters.
5426 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, hasMemory, Ops, DAG);
5429 // isLegalAddressingMode - Return true if the addressing mode represented
5430 // by AM is legal for this target, for a load/store of the specified type.
5431 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
5432 const Type *Ty) const {
5433 // FIXME: PPC does not allow r+i addressing modes for vectors!
5435 // PPC allows a sign-extended 16-bit immediate field.
5436 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
5439 // No global is ever allowed as a base.
5443 // PPC only support r+r,
5445 case 0: // "r+i" or just "i", depending on HasBaseReg.
5448 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
5450 // Otherwise we have r+r or r+i.
5453 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
5455 // Allow 2*r as r+r.
5458 // No other scales are supported.
5465 /// isLegalAddressImmediate - Return true if the integer value can be used
5466 /// as the offset of the target addressing mode for load / store of the
5468 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
5469 // PPC allows a sign-extended 16-bit immediate field.
5470 return (V > -(1 << 16) && V < (1 << 16)-1);
5473 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
5477 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
5478 DebugLoc dl = Op.getDebugLoc();
5479 // Depths > 0 not supported yet!
5480 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5483 MachineFunction &MF = DAG.getMachineFunction();
5484 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
5486 // Just load the return address off the stack.
5487 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
5489 // Make sure the function really does not optimize away the store of the RA
5491 FuncInfo->setLRStoreRequired();
5492 return DAG.getLoad(getPointerTy(), dl,
5493 DAG.getEntryNode(), RetAddrFI, NULL, 0);
5496 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
5497 DebugLoc dl = Op.getDebugLoc();
5498 // Depths > 0 not supported yet!
5499 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
5502 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
5503 bool isPPC64 = PtrVT == MVT::i64;
5505 MachineFunction &MF = DAG.getMachineFunction();
5506 MachineFrameInfo *MFI = MF.getFrameInfo();
5507 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
5508 && MFI->getStackSize();
5511 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::X31 : PPC::X1,
5514 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, is31 ? PPC::R31 : PPC::R1,
5519 PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
5520 // The PowerPC target isn't yet aware of offsets.
5524 MVT PPCTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
5525 bool isSrcConst, bool isSrcStr,
5526 SelectionDAG &DAG) const {
5527 if (this->PPCSubTarget.isPPC64()) {