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/ParameterAttributes.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
38 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
39 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
42 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
43 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
47 // Use _setjmp/_longjmp instead of setjmp/longjmp.
48 setUseUnderscoreSetJmp(true);
49 setUseUnderscoreLongJmp(true);
51 // Set up the register classes.
52 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
53 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
54 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
56 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
57 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
58 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Expand);
60 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
62 // PowerPC has pre-inc load and store's.
63 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
64 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
65 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
66 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
67 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
68 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
69 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
70 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
71 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
72 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
74 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
75 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
76 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
77 // This is used in the ppcf128->int sequence. Note it has different semantics
78 // from FP_ROUND: that rounds to nearest, this rounds to zero.
79 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
81 // PowerPC has no SREM/UREM instructions
82 setOperationAction(ISD::SREM, MVT::i32, Expand);
83 setOperationAction(ISD::UREM, MVT::i32, Expand);
84 setOperationAction(ISD::SREM, MVT::i64, Expand);
85 setOperationAction(ISD::UREM, MVT::i64, Expand);
87 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
88 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
89 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
90 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
91 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
92 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
93 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
94 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
95 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
97 // We don't support sin/cos/sqrt/fmod/pow
98 setOperationAction(ISD::FSIN , MVT::f64, Expand);
99 setOperationAction(ISD::FCOS , MVT::f64, Expand);
100 setOperationAction(ISD::FREM , MVT::f64, Expand);
101 setOperationAction(ISD::FPOW , MVT::f64, Expand);
102 setOperationAction(ISD::FLOG , MVT::f64, Expand);
103 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
104 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
105 setOperationAction(ISD::FEXP ,MVT::f64, Expand);
106 setOperationAction(ISD::FEXP2 ,MVT::f64, Expand);
107 setOperationAction(ISD::FSIN , MVT::f32, Expand);
108 setOperationAction(ISD::FCOS , MVT::f32, Expand);
109 setOperationAction(ISD::FREM , MVT::f32, Expand);
110 setOperationAction(ISD::FPOW , MVT::f32, Expand);
111 setOperationAction(ISD::FLOG , MVT::f32, Expand);
112 setOperationAction(ISD::FLOG2 ,MVT::f32, Expand);
113 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
114 setOperationAction(ISD::FEXP ,MVT::f32, Expand);
115 setOperationAction(ISD::FEXP2 ,MVT::f32, Expand);
117 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
119 // If we're enabling GP optimizations, use hardware square root
120 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
121 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
122 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
125 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
126 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
128 // PowerPC does not have BSWAP, CTPOP or CTTZ
129 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
130 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
131 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
132 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
133 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
134 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
136 // PowerPC does not have ROTR
137 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
138 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
140 // PowerPC does not have Select
141 setOperationAction(ISD::SELECT, MVT::i32, Expand);
142 setOperationAction(ISD::SELECT, MVT::i64, Expand);
143 setOperationAction(ISD::SELECT, MVT::f32, Expand);
144 setOperationAction(ISD::SELECT, MVT::f64, Expand);
146 // PowerPC wants to turn select_cc of FP into fsel when possible.
147 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
148 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
150 // PowerPC wants to optimize integer setcc a bit
151 setOperationAction(ISD::SETCC, MVT::i32, Custom);
153 // PowerPC does not have BRCOND which requires SetCC
154 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
156 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
158 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
159 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
161 // PowerPC does not have [U|S]INT_TO_FP
162 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
163 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
165 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
166 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
167 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
168 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
170 // We cannot sextinreg(i1). Expand to shifts.
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
173 // Support label based line numbers.
174 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
175 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
177 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
178 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
179 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
180 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
183 // We want to legalize GlobalAddress and ConstantPool nodes into the
184 // appropriate instructions to materialize the address.
185 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
186 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
187 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
188 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
189 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
190 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
191 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
192 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
194 // RET must be custom lowered, to meet ABI requirements.
195 setOperationAction(ISD::RET , MVT::Other, Custom);
198 setOperationAction(ISD::TRAP, MVT::Other, Legal);
200 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
201 setOperationAction(ISD::VASTART , MVT::Other, Custom);
203 // VAARG is custom lowered with ELF 32 ABI
204 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
205 setOperationAction(ISD::VAARG, MVT::Other, Custom);
207 setOperationAction(ISD::VAARG, MVT::Other, Expand);
209 // Use the default implementation.
210 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
211 setOperationAction(ISD::VAEND , MVT::Other, Expand);
212 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
213 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
214 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
215 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
217 // We want to custom lower some of our intrinsics.
218 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
220 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
221 // They also have instructions for converting between i64 and fp.
222 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
223 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
224 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
225 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
226 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
228 // FIXME: disable this lowered code. This generates 64-bit register values,
229 // and we don't model the fact that the top part is clobbered by calls. We
230 // need to flag these together so that the value isn't live across a call.
231 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
233 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
234 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
236 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
237 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
240 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
241 // 64-bit PowerPC implementations can support i64 types directly
242 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
243 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
244 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
245 // 64-bit PowerPC wants to expand i128 shifts itself.
246 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
247 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
248 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
250 // 32-bit PowerPC wants to expand i64 shifts itself.
251 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
252 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
253 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
256 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
257 // First set operation action for all vector types to expand. Then we
258 // will selectively turn on ones that can be effectively codegen'd.
259 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
260 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
261 MVT VT = (MVT::SimpleValueType)i;
263 // add/sub are legal for all supported vector VT's.
264 setOperationAction(ISD::ADD , VT, Legal);
265 setOperationAction(ISD::SUB , VT, Legal);
267 // We promote all shuffles to v16i8.
268 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
269 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
271 // We promote all non-typed operations to v4i32.
272 setOperationAction(ISD::AND , VT, Promote);
273 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
274 setOperationAction(ISD::OR , VT, Promote);
275 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
276 setOperationAction(ISD::XOR , VT, Promote);
277 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
278 setOperationAction(ISD::LOAD , VT, Promote);
279 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
280 setOperationAction(ISD::SELECT, VT, Promote);
281 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
282 setOperationAction(ISD::STORE, VT, Promote);
283 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
285 // No other operations are legal.
286 setOperationAction(ISD::MUL , VT, Expand);
287 setOperationAction(ISD::SDIV, VT, Expand);
288 setOperationAction(ISD::SREM, VT, Expand);
289 setOperationAction(ISD::UDIV, VT, Expand);
290 setOperationAction(ISD::UREM, VT, Expand);
291 setOperationAction(ISD::FDIV, VT, Expand);
292 setOperationAction(ISD::FNEG, VT, Expand);
293 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
294 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
295 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
296 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
297 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
298 setOperationAction(ISD::UDIVREM, VT, Expand);
299 setOperationAction(ISD::SDIVREM, VT, Expand);
300 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
301 setOperationAction(ISD::FPOW, VT, Expand);
302 setOperationAction(ISD::CTPOP, VT, Expand);
303 setOperationAction(ISD::CTLZ, VT, Expand);
304 setOperationAction(ISD::CTTZ, VT, Expand);
307 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
308 // with merges, splats, etc.
309 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
311 setOperationAction(ISD::AND , MVT::v4i32, Legal);
312 setOperationAction(ISD::OR , MVT::v4i32, Legal);
313 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
314 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
315 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
316 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
318 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
319 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
320 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
321 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
323 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
324 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
325 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
326 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
328 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
329 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
331 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
332 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
333 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
334 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
337 setShiftAmountType(MVT::i32);
338 setSetCCResultContents(ZeroOrOneSetCCResult);
340 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
341 setStackPointerRegisterToSaveRestore(PPC::X1);
342 setExceptionPointerRegister(PPC::X3);
343 setExceptionSelectorRegister(PPC::X4);
345 setStackPointerRegisterToSaveRestore(PPC::R1);
346 setExceptionPointerRegister(PPC::R3);
347 setExceptionSelectorRegister(PPC::R4);
350 // We have target-specific dag combine patterns for the following nodes:
351 setTargetDAGCombine(ISD::SINT_TO_FP);
352 setTargetDAGCombine(ISD::STORE);
353 setTargetDAGCombine(ISD::BR_CC);
354 setTargetDAGCombine(ISD::BSWAP);
356 // Darwin long double math library functions have $LDBL128 appended.
357 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
358 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
359 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
360 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
361 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
362 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
363 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
364 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
365 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
366 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
367 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
370 computeRegisterProperties();
373 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
374 /// function arguments in the caller parameter area.
375 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
376 TargetMachine &TM = getTargetMachine();
377 // Darwin passes everything on 4 byte boundary.
378 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
384 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
387 case PPCISD::FSEL: return "PPCISD::FSEL";
388 case PPCISD::FCFID: return "PPCISD::FCFID";
389 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
390 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
391 case PPCISD::STFIWX: return "PPCISD::STFIWX";
392 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
393 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
394 case PPCISD::VPERM: return "PPCISD::VPERM";
395 case PPCISD::Hi: return "PPCISD::Hi";
396 case PPCISD::Lo: return "PPCISD::Lo";
397 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
398 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
399 case PPCISD::SRL: return "PPCISD::SRL";
400 case PPCISD::SRA: return "PPCISD::SRA";
401 case PPCISD::SHL: return "PPCISD::SHL";
402 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
403 case PPCISD::STD_32: return "PPCISD::STD_32";
404 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
405 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
406 case PPCISD::MTCTR: return "PPCISD::MTCTR";
407 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
408 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
409 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
410 case PPCISD::MFCR: return "PPCISD::MFCR";
411 case PPCISD::VCMP: return "PPCISD::VCMP";
412 case PPCISD::VCMPo: return "PPCISD::VCMPo";
413 case PPCISD::LBRX: return "PPCISD::LBRX";
414 case PPCISD::STBRX: return "PPCISD::STBRX";
415 case PPCISD::LARX: return "PPCISD::LARX";
416 case PPCISD::STCX: return "PPCISD::STCX";
417 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
418 case PPCISD::MFFS: return "PPCISD::MFFS";
419 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
420 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
421 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
422 case PPCISD::MTFSF: return "PPCISD::MTFSF";
423 case PPCISD::TAILCALL: return "PPCISD::TAILCALL";
424 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
429 MVT PPCTargetLowering::getSetCCResultType(const SDValue &) const {
434 //===----------------------------------------------------------------------===//
435 // Node matching predicates, for use by the tblgen matching code.
436 //===----------------------------------------------------------------------===//
438 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
439 static bool isFloatingPointZero(SDValue Op) {
440 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
441 return CFP->getValueAPF().isZero();
442 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
443 // Maybe this has already been legalized into the constant pool?
444 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
445 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
446 return CFP->getValueAPF().isZero();
451 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
452 /// true if Op is undef or if it matches the specified value.
453 static bool isConstantOrUndef(SDValue Op, unsigned Val) {
454 return Op.getOpcode() == ISD::UNDEF ||
455 cast<ConstantSDNode>(Op)->getZExtValue() == Val;
458 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
459 /// VPKUHUM instruction.
460 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
462 for (unsigned i = 0; i != 16; ++i)
463 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
466 for (unsigned i = 0; i != 8; ++i)
467 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
468 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
474 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
475 /// VPKUWUM instruction.
476 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
478 for (unsigned i = 0; i != 16; i += 2)
479 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
480 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
483 for (unsigned i = 0; i != 8; i += 2)
484 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
485 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
486 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
487 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
493 /// isVMerge - Common function, used to match vmrg* shuffles.
495 static bool isVMerge(SDNode *N, unsigned UnitSize,
496 unsigned LHSStart, unsigned RHSStart) {
497 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
498 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
499 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
500 "Unsupported merge size!");
502 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
503 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
504 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
505 LHSStart+j+i*UnitSize) ||
506 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
507 RHSStart+j+i*UnitSize))
513 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
514 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
515 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
517 return isVMerge(N, UnitSize, 8, 24);
518 return isVMerge(N, UnitSize, 8, 8);
521 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
522 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
523 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
525 return isVMerge(N, UnitSize, 0, 16);
526 return isVMerge(N, UnitSize, 0, 0);
530 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
531 /// amount, otherwise return -1.
532 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
533 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
534 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
535 // Find the first non-undef value in the shuffle mask.
537 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
540 if (i == 16) return -1; // all undef.
542 // Otherwise, check to see if the rest of the elements are consequtively
543 // numbered from this value.
544 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getZExtValue();
545 if (ShiftAmt < i) return -1;
549 // Check the rest of the elements to see if they are consequtive.
550 for (++i; i != 16; ++i)
551 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
554 // Check the rest of the elements to see if they are consequtive.
555 for (++i; i != 16; ++i)
556 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
563 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
564 /// specifies a splat of a single element that is suitable for input to
565 /// VSPLTB/VSPLTH/VSPLTW.
566 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
567 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
568 N->getNumOperands() == 16 &&
569 (EltSize == 1 || EltSize == 2 || EltSize == 4));
571 // This is a splat operation if each element of the permute is the same, and
572 // if the value doesn't reference the second vector.
573 unsigned ElementBase = 0;
574 SDValue Elt = N->getOperand(0);
575 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
576 ElementBase = EltV->getZExtValue();
578 return false; // FIXME: Handle UNDEF elements too!
580 if (cast<ConstantSDNode>(Elt)->getZExtValue() >= 16)
583 // Check that they are consequtive.
584 for (unsigned i = 1; i != EltSize; ++i) {
585 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
586 cast<ConstantSDNode>(N->getOperand(i))->getZExtValue() != i+ElementBase)
590 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
591 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
592 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
593 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
594 "Invalid VECTOR_SHUFFLE mask!");
595 for (unsigned j = 0; j != EltSize; ++j)
596 if (N->getOperand(i+j) != N->getOperand(j))
603 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
605 bool PPC::isAllNegativeZeroVector(SDNode *N) {
606 assert(N->getOpcode() == ISD::BUILD_VECTOR);
607 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
608 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
609 return CFP->getValueAPF().isNegZero();
613 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
614 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
615 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
616 assert(isSplatShuffleMask(N, EltSize));
617 return cast<ConstantSDNode>(N->getOperand(0))->getZExtValue() / EltSize;
620 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
621 /// by using a vspltis[bhw] instruction of the specified element size, return
622 /// the constant being splatted. The ByteSize field indicates the number of
623 /// bytes of each element [124] -> [bhw].
624 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
627 // If ByteSize of the splat is bigger than the element size of the
628 // build_vector, then we have a case where we are checking for a splat where
629 // multiple elements of the buildvector are folded together into a single
630 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
631 unsigned EltSize = 16/N->getNumOperands();
632 if (EltSize < ByteSize) {
633 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
634 SDValue UniquedVals[4];
635 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
637 // See if all of the elements in the buildvector agree across.
638 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
639 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
640 // If the element isn't a constant, bail fully out.
641 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
644 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
645 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
646 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
647 return SDValue(); // no match.
650 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
651 // either constant or undef values that are identical for each chunk. See
652 // if these chunks can form into a larger vspltis*.
654 // Check to see if all of the leading entries are either 0 or -1. If
655 // neither, then this won't fit into the immediate field.
656 bool LeadingZero = true;
657 bool LeadingOnes = true;
658 for (unsigned i = 0; i != Multiple-1; ++i) {
659 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
661 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
662 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
664 // Finally, check the least significant entry.
666 if (UniquedVals[Multiple-1].getNode() == 0)
667 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
668 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
670 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
673 if (UniquedVals[Multiple-1].getNode() == 0)
674 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
675 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
676 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
677 return DAG.getTargetConstant(Val, MVT::i32);
683 // Check to see if this buildvec has a single non-undef value in its elements.
684 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
685 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
686 if (OpVal.getNode() == 0)
687 OpVal = N->getOperand(i);
688 else if (OpVal != N->getOperand(i))
692 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
694 unsigned ValSizeInBytes = 0;
696 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
697 Value = CN->getZExtValue();
698 ValSizeInBytes = CN->getValueType(0).getSizeInBits()/8;
699 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
700 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
701 Value = FloatToBits(CN->getValueAPF().convertToFloat());
705 // If the splat value is larger than the element value, then we can never do
706 // this splat. The only case that we could fit the replicated bits into our
707 // immediate field for would be zero, and we prefer to use vxor for it.
708 if (ValSizeInBytes < ByteSize) return SDValue();
710 // If the element value is larger than the splat value, cut it in half and
711 // check to see if the two halves are equal. Continue doing this until we
712 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
713 while (ValSizeInBytes > ByteSize) {
714 ValSizeInBytes >>= 1;
716 // If the top half equals the bottom half, we're still ok.
717 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
718 (Value & ((1 << (8*ValSizeInBytes))-1)))
722 // Properly sign extend the value.
723 int ShAmt = (4-ByteSize)*8;
724 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
726 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
727 if (MaskVal == 0) return SDValue();
729 // Finally, if this value fits in a 5 bit sext field, return it
730 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
731 return DAG.getTargetConstant(MaskVal, MVT::i32);
735 //===----------------------------------------------------------------------===//
736 // Addressing Mode Selection
737 //===----------------------------------------------------------------------===//
739 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
740 /// or 64-bit immediate, and if the value can be accurately represented as a
741 /// sign extension from a 16-bit value. If so, this returns true and the
743 static bool isIntS16Immediate(SDNode *N, short &Imm) {
744 if (N->getOpcode() != ISD::Constant)
747 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
748 if (N->getValueType(0) == MVT::i32)
749 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
751 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
753 static bool isIntS16Immediate(SDValue Op, short &Imm) {
754 return isIntS16Immediate(Op.getNode(), Imm);
758 /// SelectAddressRegReg - Given the specified addressed, check to see if it
759 /// can be represented as an indexed [r+r] operation. Returns false if it
760 /// can be more efficiently represented with [r+imm].
761 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
765 if (N.getOpcode() == ISD::ADD) {
766 if (isIntS16Immediate(N.getOperand(1), imm))
768 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
771 Base = N.getOperand(0);
772 Index = N.getOperand(1);
774 } else if (N.getOpcode() == ISD::OR) {
775 if (isIntS16Immediate(N.getOperand(1), imm))
776 return false; // r+i can fold it if we can.
778 // If this is an or of disjoint bitfields, we can codegen this as an add
779 // (for better address arithmetic) if the LHS and RHS of the OR are provably
781 APInt LHSKnownZero, LHSKnownOne;
782 APInt RHSKnownZero, RHSKnownOne;
783 DAG.ComputeMaskedBits(N.getOperand(0),
784 APInt::getAllOnesValue(N.getOperand(0)
785 .getValueSizeInBits()),
786 LHSKnownZero, LHSKnownOne);
788 if (LHSKnownZero.getBoolValue()) {
789 DAG.ComputeMaskedBits(N.getOperand(1),
790 APInt::getAllOnesValue(N.getOperand(1)
791 .getValueSizeInBits()),
792 RHSKnownZero, RHSKnownOne);
793 // If all of the bits are known zero on the LHS or RHS, the add won't
795 if (~(LHSKnownZero | RHSKnownZero) == 0) {
796 Base = N.getOperand(0);
797 Index = N.getOperand(1);
806 /// Returns true if the address N can be represented by a base register plus
807 /// a signed 16-bit displacement [r+imm], and if it is not better
808 /// represented as reg+reg.
809 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
810 SDValue &Base, SelectionDAG &DAG){
811 // If this can be more profitably realized as r+r, fail.
812 if (SelectAddressRegReg(N, Disp, Base, DAG))
815 if (N.getOpcode() == ISD::ADD) {
817 if (isIntS16Immediate(N.getOperand(1), imm)) {
818 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
819 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
820 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
822 Base = N.getOperand(0);
824 return true; // [r+i]
825 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
826 // Match LOAD (ADD (X, Lo(G))).
827 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
828 && "Cannot handle constant offsets yet!");
829 Disp = N.getOperand(1).getOperand(0); // The global address.
830 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
831 Disp.getOpcode() == ISD::TargetConstantPool ||
832 Disp.getOpcode() == ISD::TargetJumpTable);
833 Base = N.getOperand(0);
834 return true; // [&g+r]
836 } else if (N.getOpcode() == ISD::OR) {
838 if (isIntS16Immediate(N.getOperand(1), imm)) {
839 // If this is an or of disjoint bitfields, we can codegen this as an add
840 // (for better address arithmetic) if the LHS and RHS of the OR are
841 // provably disjoint.
842 APInt LHSKnownZero, LHSKnownOne;
843 DAG.ComputeMaskedBits(N.getOperand(0),
844 APInt::getAllOnesValue(N.getOperand(0)
845 .getValueSizeInBits()),
846 LHSKnownZero, LHSKnownOne);
848 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
849 // If all of the bits are known zero on the LHS or RHS, the add won't
851 Base = N.getOperand(0);
852 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
856 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
857 // Loading from a constant address.
859 // If this address fits entirely in a 16-bit sext immediate field, codegen
862 if (isIntS16Immediate(CN, Imm)) {
863 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
864 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
868 // Handle 32-bit sext immediates with LIS + addr mode.
869 if (CN->getValueType(0) == MVT::i32 ||
870 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
871 int Addr = (int)CN->getZExtValue();
873 // Otherwise, break this down into an LIS + disp.
874 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
876 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
877 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
878 Base = SDValue(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
883 Disp = DAG.getTargetConstant(0, getPointerTy());
884 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
885 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
888 return true; // [r+0]
891 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
892 /// represented as an indexed [r+r] operation.
893 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
896 // Check to see if we can easily represent this as an [r+r] address. This
897 // will fail if it thinks that the address is more profitably represented as
898 // reg+imm, e.g. where imm = 0.
899 if (SelectAddressRegReg(N, Base, Index, DAG))
902 // If the operand is an addition, always emit this as [r+r], since this is
903 // better (for code size, and execution, as the memop does the add for free)
904 // than emitting an explicit add.
905 if (N.getOpcode() == ISD::ADD) {
906 Base = N.getOperand(0);
907 Index = N.getOperand(1);
911 // Otherwise, do it the hard way, using R0 as the base register.
912 Base = DAG.getRegister(PPC::R0, N.getValueType());
917 /// SelectAddressRegImmShift - Returns true if the address N can be
918 /// represented by a base register plus a signed 14-bit displacement
919 /// [r+imm*4]. Suitable for use by STD and friends.
920 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
923 // If this can be more profitably realized as r+r, fail.
924 if (SelectAddressRegReg(N, Disp, Base, DAG))
927 if (N.getOpcode() == ISD::ADD) {
929 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
930 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
931 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
932 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
934 Base = N.getOperand(0);
936 return true; // [r+i]
937 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
938 // Match LOAD (ADD (X, Lo(G))).
939 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
940 && "Cannot handle constant offsets yet!");
941 Disp = N.getOperand(1).getOperand(0); // The global address.
942 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
943 Disp.getOpcode() == ISD::TargetConstantPool ||
944 Disp.getOpcode() == ISD::TargetJumpTable);
945 Base = N.getOperand(0);
946 return true; // [&g+r]
948 } else if (N.getOpcode() == ISD::OR) {
950 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
951 // If this is an or of disjoint bitfields, we can codegen this as an add
952 // (for better address arithmetic) if the LHS and RHS of the OR are
953 // provably disjoint.
954 APInt LHSKnownZero, LHSKnownOne;
955 DAG.ComputeMaskedBits(N.getOperand(0),
956 APInt::getAllOnesValue(N.getOperand(0)
957 .getValueSizeInBits()),
958 LHSKnownZero, LHSKnownOne);
959 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
960 // If all of the bits are known zero on the LHS or RHS, the add won't
962 Base = N.getOperand(0);
963 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
967 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
968 // Loading from a constant address. Verify low two bits are clear.
969 if ((CN->getZExtValue() & 3) == 0) {
970 // If this address fits entirely in a 14-bit sext immediate field, codegen
973 if (isIntS16Immediate(CN, Imm)) {
974 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
975 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
979 // Fold the low-part of 32-bit absolute addresses into addr mode.
980 if (CN->getValueType(0) == MVT::i32 ||
981 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
982 int Addr = (int)CN->getZExtValue();
984 // Otherwise, break this down into an LIS + disp.
985 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
987 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
988 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
989 Base = SDValue(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
995 Disp = DAG.getTargetConstant(0, getPointerTy());
996 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
997 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1000 return true; // [r+0]
1004 /// getPreIndexedAddressParts - returns true by value, base pointer and
1005 /// offset pointer and addressing mode by reference if the node's address
1006 /// can be legally represented as pre-indexed load / store address.
1007 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1009 ISD::MemIndexedMode &AM,
1010 SelectionDAG &DAG) {
1011 // Disabled by default for now.
1012 if (!EnablePPCPreinc) return false;
1016 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1017 Ptr = LD->getBasePtr();
1018 VT = LD->getMemoryVT();
1020 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1022 Ptr = ST->getBasePtr();
1023 VT = ST->getMemoryVT();
1027 // PowerPC doesn't have preinc load/store instructions for vectors.
1031 // TODO: Check reg+reg first.
1033 // LDU/STU use reg+imm*4, others use reg+imm.
1034 if (VT != MVT::i64) {
1036 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1040 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1044 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1045 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1046 // sext i32 to i64 when addr mode is r+i.
1047 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1048 LD->getExtensionType() == ISD::SEXTLOAD &&
1049 isa<ConstantSDNode>(Offset))
1057 //===----------------------------------------------------------------------===//
1058 // LowerOperation implementation
1059 //===----------------------------------------------------------------------===//
1061 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1062 SelectionDAG &DAG) {
1063 MVT PtrVT = Op.getValueType();
1064 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1065 Constant *C = CP->getConstVal();
1066 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1067 SDValue Zero = DAG.getConstant(0, PtrVT);
1069 const TargetMachine &TM = DAG.getTarget();
1071 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1072 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1074 // If this is a non-darwin platform, we don't support non-static relo models
1076 if (TM.getRelocationModel() == Reloc::Static ||
1077 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1078 // Generate non-pic code that has direct accesses to the constant pool.
1079 // The address of the global is just (hi(&g)+lo(&g)).
1080 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1083 if (TM.getRelocationModel() == Reloc::PIC_) {
1084 // With PIC, the first instruction is actually "GR+hi(&G)".
1085 Hi = DAG.getNode(ISD::ADD, PtrVT,
1086 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1089 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1093 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1094 MVT PtrVT = Op.getValueType();
1095 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1096 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1097 SDValue Zero = DAG.getConstant(0, PtrVT);
1099 const TargetMachine &TM = DAG.getTarget();
1101 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1102 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1104 // If this is a non-darwin platform, we don't support non-static relo models
1106 if (TM.getRelocationModel() == Reloc::Static ||
1107 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1108 // Generate non-pic code that has direct accesses to the constant pool.
1109 // The address of the global is just (hi(&g)+lo(&g)).
1110 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1113 if (TM.getRelocationModel() == Reloc::PIC_) {
1114 // With PIC, the first instruction is actually "GR+hi(&G)".
1115 Hi = DAG.getNode(ISD::ADD, PtrVT,
1116 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1119 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1123 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1124 SelectionDAG &DAG) {
1125 assert(0 && "TLS not implemented for PPC.");
1126 return SDValue(); // Not reached
1129 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1130 SelectionDAG &DAG) {
1131 MVT PtrVT = Op.getValueType();
1132 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1133 GlobalValue *GV = GSDN->getGlobal();
1134 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1135 // If it's a debug information descriptor, don't mess with it.
1136 if (DAG.isVerifiedDebugInfoDesc(Op))
1138 SDValue Zero = DAG.getConstant(0, PtrVT);
1140 const TargetMachine &TM = DAG.getTarget();
1142 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1143 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1145 // If this is a non-darwin platform, we don't support non-static relo models
1147 if (TM.getRelocationModel() == Reloc::Static ||
1148 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1149 // Generate non-pic code that has direct accesses to globals.
1150 // The address of the global is just (hi(&g)+lo(&g)).
1151 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1154 if (TM.getRelocationModel() == Reloc::PIC_) {
1155 // With PIC, the first instruction is actually "GR+hi(&G)".
1156 Hi = DAG.getNode(ISD::ADD, PtrVT,
1157 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1160 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1162 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1165 // If the global is weak or external, we have to go through the lazy
1167 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1170 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1171 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1173 // If we're comparing for equality to zero, expose the fact that this is
1174 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1175 // fold the new nodes.
1176 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1177 if (C->isNullValue() && CC == ISD::SETEQ) {
1178 MVT VT = Op.getOperand(0).getValueType();
1179 SDValue Zext = Op.getOperand(0);
1180 if (VT.bitsLT(MVT::i32)) {
1182 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1184 unsigned Log2b = Log2_32(VT.getSizeInBits());
1185 SDValue Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1186 SDValue Scc = DAG.getNode(ISD::SRL, VT, Clz,
1187 DAG.getConstant(Log2b, MVT::i32));
1188 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1190 // Leave comparisons against 0 and -1 alone for now, since they're usually
1191 // optimized. FIXME: revisit this when we can custom lower all setcc
1193 if (C->isAllOnesValue() || C->isNullValue())
1197 // If we have an integer seteq/setne, turn it into a compare against zero
1198 // by xor'ing the rhs with the lhs, which is faster than setting a
1199 // condition register, reading it back out, and masking the correct bit. The
1200 // normal approach here uses sub to do this instead of xor. Using xor exposes
1201 // the result to other bit-twiddling opportunities.
1202 MVT LHSVT = Op.getOperand(0).getValueType();
1203 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1204 MVT VT = Op.getValueType();
1205 SDValue Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1207 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1212 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1213 int VarArgsFrameIndex,
1214 int VarArgsStackOffset,
1215 unsigned VarArgsNumGPR,
1216 unsigned VarArgsNumFPR,
1217 const PPCSubtarget &Subtarget) {
1219 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1220 return SDValue(); // Not reached
1223 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1224 int VarArgsFrameIndex,
1225 int VarArgsStackOffset,
1226 unsigned VarArgsNumGPR,
1227 unsigned VarArgsNumFPR,
1228 const PPCSubtarget &Subtarget) {
1230 if (Subtarget.isMachoABI()) {
1231 // vastart just stores the address of the VarArgsFrameIndex slot into the
1232 // memory location argument.
1233 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1234 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1235 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1236 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1239 // For ELF 32 ABI we follow the layout of the va_list struct.
1240 // We suppose the given va_list is already allocated.
1243 // char gpr; /* index into the array of 8 GPRs
1244 // * stored in the register save area
1245 // * gpr=0 corresponds to r3,
1246 // * gpr=1 to r4, etc.
1248 // char fpr; /* index into the array of 8 FPRs
1249 // * stored in the register save area
1250 // * fpr=0 corresponds to f1,
1251 // * fpr=1 to f2, etc.
1253 // char *overflow_arg_area;
1254 // /* location on stack that holds
1255 // * the next overflow argument
1257 // char *reg_save_area;
1258 // /* where r3:r10 and f1:f8 (if saved)
1264 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1265 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1268 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1270 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1271 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1273 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1274 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1276 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1277 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1279 uint64_t FPROffset = 1;
1280 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1282 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1284 // Store first byte : number of int regs
1285 SDValue firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1286 Op.getOperand(1), SV, 0);
1287 uint64_t nextOffset = FPROffset;
1288 SDValue nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1291 // Store second byte : number of float regs
1292 SDValue secondStore =
1293 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1294 nextOffset += StackOffset;
1295 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1297 // Store second word : arguments given on stack
1298 SDValue thirdStore =
1299 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1300 nextOffset += FrameOffset;
1301 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1303 // Store third word : arguments given in registers
1304 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1308 #include "PPCGenCallingConv.inc"
1310 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1311 /// depending on which subtarget is selected.
1312 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1313 if (Subtarget.isMachoABI()) {
1314 static const unsigned FPR[] = {
1315 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1316 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1322 static const unsigned FPR[] = {
1323 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1329 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1331 static unsigned CalculateStackSlotSize(SDValue Arg, SDValue Flag,
1332 bool isVarArg, unsigned PtrByteSize) {
1333 MVT ArgVT = Arg.getValueType();
1334 ISD::ArgFlagsTy Flags = cast<ARG_FLAGSSDNode>(Flag)->getArgFlags();
1335 unsigned ArgSize =ArgVT.getSizeInBits()/8;
1336 if (Flags.isByVal())
1337 ArgSize = Flags.getByValSize();
1338 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1344 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op,
1346 int &VarArgsFrameIndex,
1347 int &VarArgsStackOffset,
1348 unsigned &VarArgsNumGPR,
1349 unsigned &VarArgsNumFPR,
1350 const PPCSubtarget &Subtarget) {
1351 // TODO: add description of PPC stack frame format, or at least some docs.
1353 MachineFunction &MF = DAG.getMachineFunction();
1354 MachineFrameInfo *MFI = MF.getFrameInfo();
1355 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1356 SmallVector<SDValue, 8> ArgValues;
1357 SDValue Root = Op.getOperand(0);
1358 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1360 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1361 bool isPPC64 = PtrVT == MVT::i64;
1362 bool isMachoABI = Subtarget.isMachoABI();
1363 bool isELF32_ABI = Subtarget.isELF32_ABI();
1364 // Potential tail calls could cause overwriting of argument stack slots.
1365 unsigned CC = MF.getFunction()->getCallingConv();
1366 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1367 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1369 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1370 // Area that is at least reserved in caller of this function.
1371 unsigned MinReservedArea = ArgOffset;
1373 static const unsigned GPR_32[] = { // 32-bit registers.
1374 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1375 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1377 static const unsigned GPR_64[] = { // 64-bit registers.
1378 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1379 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1382 static const unsigned *FPR = GetFPR(Subtarget);
1384 static const unsigned VR[] = {
1385 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1386 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1389 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1390 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1391 const unsigned Num_VR_Regs = array_lengthof( VR);
1393 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1395 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1397 // In 32-bit non-varargs functions, the stack space for vectors is after the
1398 // stack space for non-vectors. We do not use this space unless we have
1399 // too many vectors to fit in registers, something that only occurs in
1400 // constructed examples:), but we have to walk the arglist to figure
1401 // that out...for the pathological case, compute VecArgOffset as the
1402 // start of the vector parameter area. Computing VecArgOffset is the
1403 // entire point of the following loop.
1404 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1405 // to handle Elf here.
1406 unsigned VecArgOffset = ArgOffset;
1407 if (!isVarArg && !isPPC64) {
1408 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e;
1410 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1411 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1412 ISD::ArgFlagsTy Flags =
1413 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1415 if (Flags.isByVal()) {
1416 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1417 ObjSize = Flags.getByValSize();
1419 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1420 VecArgOffset += ArgSize;
1424 switch(ObjectVT.getSimpleVT()) {
1425 default: assert(0 && "Unhandled argument type!");
1428 VecArgOffset += isPPC64 ? 8 : 4;
1430 case MVT::i64: // PPC64
1438 // Nothing to do, we're only looking at Nonvector args here.
1443 // We've found where the vector parameter area in memory is. Skip the
1444 // first 12 parameters; these don't use that memory.
1445 VecArgOffset = ((VecArgOffset+15)/16)*16;
1446 VecArgOffset += 12*16;
1448 // Add DAG nodes to load the arguments or copy them out of registers. On
1449 // entry to a function on PPC, the arguments start after the linkage area,
1450 // although the first ones are often in registers.
1452 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1453 // represented with two words (long long or double) must be copied to an
1454 // even GPR_idx value or to an even ArgOffset value.
1456 SmallVector<SDValue, 8> MemOps;
1457 unsigned nAltivecParamsAtEnd = 0;
1458 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues() - 1;
1459 ArgNo != e; ++ArgNo) {
1461 bool needsLoad = false;
1462 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1463 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1464 unsigned ArgSize = ObjSize;
1465 ISD::ArgFlagsTy Flags =
1466 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1467 // See if next argument requires stack alignment in ELF
1468 bool Align = Flags.isSplit();
1470 unsigned CurArgOffset = ArgOffset;
1472 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1473 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1474 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1475 if (isVarArg || isPPC64) {
1476 MinReservedArea = ((MinReservedArea+15)/16)*16;
1477 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1478 Op.getOperand(ArgNo+3),
1481 } else nAltivecParamsAtEnd++;
1483 // Calculate min reserved area.
1484 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1485 Op.getOperand(ArgNo+3),
1489 // FIXME alignment for ELF may not be right
1490 // FIXME the codegen can be much improved in some cases.
1491 // We do not have to keep everything in memory.
1492 if (Flags.isByVal()) {
1493 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1494 ObjSize = Flags.getByValSize();
1495 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1496 // Double word align in ELF
1497 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1498 // Objects of size 1 and 2 are right justified, everything else is
1499 // left justified. This means the memory address is adjusted forwards.
1500 if (ObjSize==1 || ObjSize==2) {
1501 CurArgOffset = CurArgOffset + (4 - ObjSize);
1503 // The value of the object is its address.
1504 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1505 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1506 ArgValues.push_back(FIN);
1507 if (ObjSize==1 || ObjSize==2) {
1508 if (GPR_idx != Num_GPR_Regs) {
1509 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1510 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1511 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1512 SDValue Store = DAG.getTruncStore(Val.getValue(1), Val, FIN,
1513 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1514 MemOps.push_back(Store);
1516 if (isMachoABI) ArgOffset += PtrByteSize;
1518 ArgOffset += PtrByteSize;
1522 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1523 // Store whatever pieces of the object are in registers
1524 // to memory. ArgVal will be address of the beginning of
1526 if (GPR_idx != Num_GPR_Regs) {
1527 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1528 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1529 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1530 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1531 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1532 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1533 MemOps.push_back(Store);
1535 if (isMachoABI) ArgOffset += PtrByteSize;
1537 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1544 switch (ObjectVT.getSimpleVT()) {
1545 default: assert(0 && "Unhandled argument type!");
1548 // Double word align in ELF
1549 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1551 if (GPR_idx != Num_GPR_Regs) {
1552 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1553 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1554 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1558 ArgSize = PtrByteSize;
1560 // Stack align in ELF
1561 if (needsLoad && Align && isELF32_ABI)
1562 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1563 // All int arguments reserve stack space in Macho ABI.
1564 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1568 case MVT::i64: // PPC64
1569 if (GPR_idx != Num_GPR_Regs) {
1570 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1571 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1572 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1574 if (ObjectVT == MVT::i32) {
1575 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1576 // value to MVT::i64 and then truncate to the correct register size.
1578 ArgVal = DAG.getNode(ISD::AssertSext, MVT::i64, ArgVal,
1579 DAG.getValueType(ObjectVT));
1580 else if (Flags.isZExt())
1581 ArgVal = DAG.getNode(ISD::AssertZext, MVT::i64, ArgVal,
1582 DAG.getValueType(ObjectVT));
1584 ArgVal = DAG.getNode(ISD::TRUNCATE, MVT::i32, ArgVal);
1590 ArgSize = PtrByteSize;
1592 // All int arguments reserve stack space in Macho ABI.
1593 if (isMachoABI || needsLoad) ArgOffset += 8;
1598 // Every 4 bytes of argument space consumes one of the GPRs available for
1599 // argument passing.
1600 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1602 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1605 if (FPR_idx != Num_FPR_Regs) {
1607 if (ObjectVT == MVT::f32)
1608 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1610 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1611 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1612 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1618 // Stack align in ELF
1619 if (needsLoad && Align && isELF32_ABI)
1620 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1621 // All FP arguments reserve stack space in Macho ABI.
1622 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1628 // Note that vector arguments in registers don't reserve stack space,
1629 // except in varargs functions.
1630 if (VR_idx != Num_VR_Regs) {
1631 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1632 RegInfo.addLiveIn(VR[VR_idx], VReg);
1633 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1635 while ((ArgOffset % 16) != 0) {
1636 ArgOffset += PtrByteSize;
1637 if (GPR_idx != Num_GPR_Regs)
1641 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1645 if (!isVarArg && !isPPC64) {
1646 // Vectors go after all the nonvectors.
1647 CurArgOffset = VecArgOffset;
1650 // Vectors are aligned.
1651 ArgOffset = ((ArgOffset+15)/16)*16;
1652 CurArgOffset = ArgOffset;
1660 // We need to load the argument to a virtual register if we determined above
1661 // that we ran out of physical registers of the appropriate type.
1663 int FI = MFI->CreateFixedObject(ObjSize,
1664 CurArgOffset + (ArgSize - ObjSize),
1666 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1667 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1670 ArgValues.push_back(ArgVal);
1673 // Set the size that is at least reserved in caller of this function. Tail
1674 // call optimized function's reserved stack space needs to be aligned so that
1675 // taking the difference between two stack areas will result in an aligned
1677 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1678 // Add the Altivec parameters at the end, if needed.
1679 if (nAltivecParamsAtEnd) {
1680 MinReservedArea = ((MinReservedArea+15)/16)*16;
1681 MinReservedArea += 16*nAltivecParamsAtEnd;
1684 std::max(MinReservedArea,
1685 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1686 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1687 getStackAlignment();
1688 unsigned AlignMask = TargetAlign-1;
1689 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1690 FI->setMinReservedArea(MinReservedArea);
1692 // If the function takes variable number of arguments, make a frame index for
1693 // the start of the first vararg value... for expansion of llvm.va_start.
1698 VarArgsNumGPR = GPR_idx;
1699 VarArgsNumFPR = FPR_idx;
1701 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1703 depth = -(Num_GPR_Regs * PtrVT.getSizeInBits()/8 +
1704 Num_FPR_Regs * MVT(MVT::f64).getSizeInBits()/8 +
1705 PtrVT.getSizeInBits()/8);
1707 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1714 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1716 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1718 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1719 // stored to the VarArgsFrameIndex on the stack.
1721 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1722 SDValue Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1723 SDValue Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1724 MemOps.push_back(Store);
1725 // Increment the address by four for the next argument to store
1726 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1727 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1731 // If this function is vararg, store any remaining integer argument regs
1732 // to their spots on the stack so that they may be loaded by deferencing the
1733 // result of va_next.
1734 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1737 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1739 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1741 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1742 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1743 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1744 MemOps.push_back(Store);
1745 // Increment the address by four for the next argument to store
1746 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1747 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1750 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1753 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1754 SDValue Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1755 SDValue Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1756 MemOps.push_back(Store);
1757 // Increment the address by eight for the next argument to store
1758 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1760 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1763 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1765 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1767 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1768 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1769 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1770 MemOps.push_back(Store);
1771 // Increment the address by eight for the next argument to store
1772 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1774 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1779 if (!MemOps.empty())
1780 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1782 ArgValues.push_back(Root);
1784 // Return the new list of results.
1785 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1789 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
1792 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
1798 unsigned &nAltivecParamsAtEnd) {
1799 // Count how many bytes are to be pushed on the stack, including the linkage
1800 // area, and parameter passing area. We start with 24/48 bytes, which is
1801 // prereserved space for [SP][CR][LR][3 x unused].
1802 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1803 unsigned NumOps = (Call.getNumOperands() - 5) / 2;
1804 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1806 // Add up all the space actually used.
1807 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1808 // they all go in registers, but we must reserve stack space for them for
1809 // possible use by the caller. In varargs or 64-bit calls, parameters are
1810 // assigned stack space in order, with padding so Altivec parameters are
1812 nAltivecParamsAtEnd = 0;
1813 for (unsigned i = 0; i != NumOps; ++i) {
1814 SDValue Arg = Call.getOperand(5+2*i);
1815 SDValue Flag = Call.getOperand(5+2*i+1);
1816 MVT ArgVT = Arg.getValueType();
1817 // Varargs Altivec parameters are padded to a 16 byte boundary.
1818 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1819 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1820 if (!isVarArg && !isPPC64) {
1821 // Non-varargs Altivec parameters go after all the non-Altivec
1822 // parameters; handle those later so we know how much padding we need.
1823 nAltivecParamsAtEnd++;
1826 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1827 NumBytes = ((NumBytes+15)/16)*16;
1829 NumBytes += CalculateStackSlotSize(Arg, Flag, isVarArg, PtrByteSize);
1832 // Allow for Altivec parameters at the end, if needed.
1833 if (nAltivecParamsAtEnd) {
1834 NumBytes = ((NumBytes+15)/16)*16;
1835 NumBytes += 16*nAltivecParamsAtEnd;
1838 // The prolog code of the callee may store up to 8 GPR argument registers to
1839 // the stack, allowing va_start to index over them in memory if its varargs.
1840 // Because we cannot tell if this is needed on the caller side, we have to
1841 // conservatively assume that it is needed. As such, make sure we have at
1842 // least enough stack space for the caller to store the 8 GPRs.
1843 NumBytes = std::max(NumBytes,
1844 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1846 // Tail call needs the stack to be aligned.
1847 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1848 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1849 getStackAlignment();
1850 unsigned AlignMask = TargetAlign-1;
1851 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
1857 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
1858 /// adjusted to accomodate the arguments for the tailcall.
1859 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
1860 unsigned ParamSize) {
1862 if (!IsTailCall) return 0;
1864 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
1865 unsigned CallerMinReservedArea = FI->getMinReservedArea();
1866 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
1867 // Remember only if the new adjustement is bigger.
1868 if (SPDiff < FI->getTailCallSPDelta())
1869 FI->setTailCallSPDelta(SPDiff);
1874 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1875 /// following the call is a return. A function is eligible if caller/callee
1876 /// calling conventions match, currently only fastcc supports tail calls, and
1877 /// the function CALL is immediatly followed by a RET.
1879 PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Call,
1881 SelectionDAG& DAG) const {
1882 // Variable argument functions are not supported.
1883 if (!PerformTailCallOpt ||
1884 cast<ConstantSDNode>(Call.getOperand(2))->getZExtValue() != 0)
1887 if (CheckTailCallReturnConstraints(Call, Ret)) {
1888 MachineFunction &MF = DAG.getMachineFunction();
1889 unsigned CallerCC = MF.getFunction()->getCallingConv();
1890 unsigned CalleeCC= cast<ConstantSDNode>(Call.getOperand(1))->getZExtValue();
1891 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1892 // Functions containing by val parameters are not supported.
1893 for (unsigned i = 0; i != ((Call.getNumOperands()-5)/2); i++) {
1894 ISD::ArgFlagsTy Flags = cast<ARG_FLAGSSDNode>(Call.getOperand(5+2*i+1))
1896 if (Flags.isByVal()) return false;
1899 SDValue Callee = Call.getOperand(4);
1900 // Non PIC/GOT tail calls are supported.
1901 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1904 // At the moment we can only do local tail calls (in same module, hidden
1905 // or protected) if we are generating PIC.
1906 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1907 return G->getGlobal()->hasHiddenVisibility()
1908 || G->getGlobal()->hasProtectedVisibility();
1915 /// isCallCompatibleAddress - Return the immediate to use if the specified
1916 /// 32-bit value is representable in the immediate field of a BxA instruction.
1917 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
1918 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1921 int Addr = C->getZExtValue();
1922 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1923 (Addr << 6 >> 6) != Addr)
1924 return 0; // Top 6 bits have to be sext of immediate.
1926 return DAG.getConstant((int)C->getZExtValue() >> 2,
1927 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
1932 struct TailCallArgumentInfo {
1937 TailCallArgumentInfo() : FrameIdx(0) {}
1942 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
1944 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
1946 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
1947 SmallVector<SDValue, 8> &MemOpChains) {
1948 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
1949 SDValue Arg = TailCallArgs[i].Arg;
1950 SDValue FIN = TailCallArgs[i].FrameIdxOp;
1951 int FI = TailCallArgs[i].FrameIdx;
1952 // Store relative to framepointer.
1953 MemOpChains.push_back(DAG.getStore(Chain, Arg, FIN,
1954 PseudoSourceValue::getFixedStack(FI),
1959 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
1960 /// the appropriate stack slot for the tail call optimized function call.
1961 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
1962 MachineFunction &MF,
1970 // Calculate the new stack slot for the return address.
1971 int SlotSize = isPPC64 ? 8 : 4;
1972 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
1974 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
1976 int NewFPLoc = SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64,
1978 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc);
1980 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
1981 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
1982 Chain = DAG.getStore(Chain, OldRetAddr, NewRetAddrFrIdx,
1983 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
1984 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
1985 Chain = DAG.getStore(Chain, OldFP, NewFramePtrIdx,
1986 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
1991 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
1992 /// the position of the argument.
1994 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
1995 SDValue Arg, int SPDiff, unsigned ArgOffset,
1996 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
1997 int Offset = ArgOffset + SPDiff;
1998 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
1999 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
2000 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2001 SDValue FIN = DAG.getFrameIndex(FI, VT);
2002 TailCallArgumentInfo Info;
2004 Info.FrameIdxOp = FIN;
2006 TailCallArguments.push_back(Info);
2009 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2010 /// stack slot. Returns the chain as result and the loaded frame pointers in
2011 /// LROpOut/FPOpout. Used when tail calling.
2012 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2018 // Load the LR and FP stack slot for later adjusting.
2019 MVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2020 LROpOut = getReturnAddrFrameIndex(DAG);
2021 LROpOut = DAG.getLoad(VT, Chain, LROpOut, NULL, 0);
2022 Chain = SDValue(LROpOut.getNode(), 1);
2023 FPOpOut = getFramePointerFrameIndex(DAG);
2024 FPOpOut = DAG.getLoad(VT, Chain, FPOpOut, NULL, 0);
2025 Chain = SDValue(FPOpOut.getNode(), 1);
2030 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2031 /// by "Src" to address "Dst" of size "Size". Alignment information is
2032 /// specified by the specific parameter attribute. The copy will be passed as
2033 /// a byval function parameter.
2034 /// Sometimes what we are copying is the end of a larger object, the part that
2035 /// does not fit in registers.
2037 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2038 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2040 SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
2041 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(), false,
2045 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2048 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2049 SDValue Arg, SDValue PtrOff, int SPDiff,
2050 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2051 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2052 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2053 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2058 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2060 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2061 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2062 DAG.getConstant(ArgOffset, PtrVT));
2064 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
2065 // Calculate and remember argument location.
2066 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2070 SDValue PPCTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG,
2071 const PPCSubtarget &Subtarget,
2072 TargetMachine &TM) {
2073 SDValue Chain = Op.getOperand(0);
2074 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
2075 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2076 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue() != 0
2077 && CC == CallingConv::Fast && PerformTailCallOpt;
2078 SDValue Callee = Op.getOperand(4);
2079 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
2081 bool isMachoABI = Subtarget.isMachoABI();
2082 bool isELF32_ABI = Subtarget.isELF32_ABI();
2084 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2085 bool isPPC64 = PtrVT == MVT::i64;
2086 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2088 MachineFunction &MF = DAG.getMachineFunction();
2090 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
2091 // SelectExpr to use to put the arguments in the appropriate registers.
2092 std::vector<SDValue> args_to_use;
2094 // Mark this function as potentially containing a function that contains a
2095 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2096 // and restoring the callers stack pointer in this functions epilog. This is
2097 // done because by tail calling the called function might overwrite the value
2098 // in this function's (MF) stack pointer stack slot 0(SP).
2099 if (PerformTailCallOpt && CC==CallingConv::Fast)
2100 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2102 unsigned nAltivecParamsAtEnd = 0;
2104 // Count how many bytes are to be pushed on the stack, including the linkage
2105 // area, and parameter passing area. We start with 24/48 bytes, which is
2106 // prereserved space for [SP][CR][LR][3 x unused].
2108 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isMachoABI, isVarArg, CC,
2109 Op, nAltivecParamsAtEnd);
2111 // Calculate by how many bytes the stack has to be adjusted in case of tail
2112 // call optimization.
2113 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2115 // Adjust the stack pointer for the new arguments...
2116 // These operations are automatically eliminated by the prolog/epilog pass
2117 Chain = DAG.getCALLSEQ_START(Chain,
2118 DAG.getConstant(NumBytes, PtrVT));
2119 SDValue CallSeqStart = Chain;
2121 // Load the return address and frame pointer so it can be move somewhere else
2124 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp);
2126 // Set up a copy of the stack pointer for use loading and storing any
2127 // arguments that may not fit in the registers available for argument
2131 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2133 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2135 // Figure out which arguments are going to go in registers, and which in
2136 // memory. Also, if this is a vararg function, floating point operations
2137 // must be stored to our stack, and loaded into integer regs as well, if
2138 // any integer regs are available for argument passing.
2139 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2140 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2142 static const unsigned GPR_32[] = { // 32-bit registers.
2143 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2144 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2146 static const unsigned GPR_64[] = { // 64-bit registers.
2147 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2148 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2150 static const unsigned *FPR = GetFPR(Subtarget);
2152 static const unsigned VR[] = {
2153 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2154 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2156 const unsigned NumGPRs = array_lengthof(GPR_32);
2157 const unsigned NumFPRs = isMachoABI ? 13 : 8;
2158 const unsigned NumVRs = array_lengthof( VR);
2160 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2162 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
2163 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2165 SmallVector<SDValue, 8> MemOpChains;
2166 for (unsigned i = 0; i != NumOps; ++i) {
2168 SDValue Arg = Op.getOperand(5+2*i);
2169 ISD::ArgFlagsTy Flags =
2170 cast<ARG_FLAGSSDNode>(Op.getOperand(5+2*i+1))->getArgFlags();
2171 // See if next argument requires stack alignment in ELF
2172 bool Align = Flags.isSplit();
2174 // PtrOff will be used to store the current argument to the stack if a
2175 // register cannot be found for it.
2178 // Stack align in ELF 32
2179 if (isELF32_ABI && Align)
2180 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
2181 StackPtr.getValueType());
2183 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2185 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
2187 // On PPC64, promote integers to 64-bit values.
2188 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2189 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2190 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2191 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
2194 // FIXME Elf untested, what are alignment rules?
2195 // FIXME memcpy is used way more than necessary. Correctness first.
2196 if (Flags.isByVal()) {
2197 unsigned Size = Flags.getByValSize();
2198 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2199 if (Size==1 || Size==2) {
2200 // Very small objects are passed right-justified.
2201 // Everything else is passed left-justified.
2202 MVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2203 if (GPR_idx != NumGPRs) {
2204 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
2206 MemOpChains.push_back(Load.getValue(1));
2207 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2209 ArgOffset += PtrByteSize;
2211 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2212 SDValue AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
2213 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2214 CallSeqStart.getNode()->getOperand(0),
2216 // This must go outside the CALLSEQ_START..END.
2217 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2218 CallSeqStart.getNode()->getOperand(1));
2219 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2220 NewCallSeqStart.getNode());
2221 Chain = CallSeqStart = NewCallSeqStart;
2222 ArgOffset += PtrByteSize;
2226 // Copy entire object into memory. There are cases where gcc-generated
2227 // code assumes it is there, even if it could be put entirely into
2228 // registers. (This is not what the doc says.)
2229 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2230 CallSeqStart.getNode()->getOperand(0),
2232 // This must go outside the CALLSEQ_START..END.
2233 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2234 CallSeqStart.getNode()->getOperand(1));
2235 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2236 Chain = CallSeqStart = NewCallSeqStart;
2237 // And copy the pieces of it that fit into registers.
2238 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2239 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2240 SDValue AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
2241 if (GPR_idx != NumGPRs) {
2242 SDValue Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
2243 MemOpChains.push_back(Load.getValue(1));
2244 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2246 ArgOffset += PtrByteSize;
2248 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2255 switch (Arg.getValueType().getSimpleVT()) {
2256 default: assert(0 && "Unexpected ValueType for argument!");
2259 // Double word align in ELF
2260 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2261 if (GPR_idx != NumGPRs) {
2262 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2264 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2265 isPPC64, isTailCall, false, MemOpChains,
2269 if (inMem || isMachoABI) {
2270 // Stack align in ELF
2271 if (isELF32_ABI && Align)
2272 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2274 ArgOffset += PtrByteSize;
2279 if (FPR_idx != NumFPRs) {
2280 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
2283 SDValue Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2284 MemOpChains.push_back(Store);
2286 // Float varargs are always shadowed in available integer registers
2287 if (GPR_idx != NumGPRs) {
2288 SDValue Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
2289 MemOpChains.push_back(Load.getValue(1));
2290 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2293 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
2294 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
2295 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
2296 SDValue Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
2297 MemOpChains.push_back(Load.getValue(1));
2298 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2302 // If we have any FPRs remaining, we may also have GPRs remaining.
2303 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
2306 if (GPR_idx != NumGPRs)
2308 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
2309 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
2314 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2315 isPPC64, isTailCall, false, MemOpChains,
2319 if (inMem || isMachoABI) {
2320 // Stack align in ELF
2321 if (isELF32_ABI && Align)
2322 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2326 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2334 // These go aligned on the stack, or in the corresponding R registers
2335 // when within range. The Darwin PPC ABI doc claims they also go in
2336 // V registers; in fact gcc does this only for arguments that are
2337 // prototyped, not for those that match the ... We do it for all
2338 // arguments, seems to work.
2339 while (ArgOffset % 16 !=0) {
2340 ArgOffset += PtrByteSize;
2341 if (GPR_idx != NumGPRs)
2344 // We could elide this store in the case where the object fits
2345 // entirely in R registers. Maybe later.
2346 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2347 DAG.getConstant(ArgOffset, PtrVT));
2348 SDValue Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2349 MemOpChains.push_back(Store);
2350 if (VR_idx != NumVRs) {
2351 SDValue Load = DAG.getLoad(MVT::v4f32, Store, PtrOff, NULL, 0);
2352 MemOpChains.push_back(Load.getValue(1));
2353 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2356 for (unsigned i=0; i<16; i+=PtrByteSize) {
2357 if (GPR_idx == NumGPRs)
2359 SDValue Ix = DAG.getNode(ISD::ADD, PtrVT, PtrOff,
2360 DAG.getConstant(i, PtrVT));
2361 SDValue Load = DAG.getLoad(PtrVT, Store, Ix, NULL, 0);
2362 MemOpChains.push_back(Load.getValue(1));
2363 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2368 // Non-varargs Altivec params generally go in registers, but have
2369 // stack space allocated at the end.
2370 if (VR_idx != NumVRs) {
2371 // Doesn't have GPR space allocated.
2372 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2373 } else if (nAltivecParamsAtEnd==0) {
2374 // We are emitting Altivec params in order.
2375 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2376 isPPC64, isTailCall, true, MemOpChains,
2383 // If all Altivec parameters fit in registers, as they usually do,
2384 // they get stack space following the non-Altivec parameters. We
2385 // don't track this here because nobody below needs it.
2386 // If there are more Altivec parameters than fit in registers emit
2388 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2390 // Offset is aligned; skip 1st 12 params which go in V registers.
2391 ArgOffset = ((ArgOffset+15)/16)*16;
2393 for (unsigned i = 0; i != NumOps; ++i) {
2394 SDValue Arg = Op.getOperand(5+2*i);
2395 MVT ArgType = Arg.getValueType();
2396 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2397 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2400 // We are emitting Altivec params in order.
2401 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2402 isPPC64, isTailCall, true, MemOpChains,
2410 if (!MemOpChains.empty())
2411 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2412 &MemOpChains[0], MemOpChains.size());
2414 // Build a sequence of copy-to-reg nodes chained together with token chain
2415 // and flag operands which copy the outgoing args into the appropriate regs.
2417 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2418 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
2420 InFlag = Chain.getValue(1);
2423 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2424 if (isVarArg && isELF32_ABI) {
2425 SDValue SetCR(DAG.getTargetNode(PPC::CRSET, MVT::i32), 0);
2426 Chain = DAG.getCopyToReg(Chain, PPC::CR1EQ, SetCR, InFlag);
2427 InFlag = Chain.getValue(1);
2430 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2431 // might overwrite each other in case of tail call optimization.
2433 SmallVector<SDValue, 8> MemOpChains2;
2434 // Do not flag preceeding copytoreg stuff together with the following stuff.
2436 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2438 if (!MemOpChains2.empty())
2439 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2440 &MemOpChains2[0], MemOpChains2.size());
2442 // Store the return address to the appropriate stack slot.
2443 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2444 isPPC64, isMachoABI);
2447 // Emit callseq_end just before tailcall node.
2449 SmallVector<SDValue, 8> CallSeqOps;
2450 SDVTList CallSeqNodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2451 CallSeqOps.push_back(Chain);
2452 CallSeqOps.push_back(DAG.getIntPtrConstant(NumBytes));
2453 CallSeqOps.push_back(DAG.getIntPtrConstant(0));
2454 if (InFlag.getNode())
2455 CallSeqOps.push_back(InFlag);
2456 Chain = DAG.getNode(ISD::CALLSEQ_END, CallSeqNodeTys, &CallSeqOps[0],
2458 InFlag = Chain.getValue(1);
2461 std::vector<MVT> NodeTys;
2462 NodeTys.push_back(MVT::Other); // Returns a chain
2463 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2465 SmallVector<SDValue, 8> Ops;
2466 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2468 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2469 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2470 // node so that legalize doesn't hack it.
2471 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2472 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2473 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2474 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2475 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2476 // If this is an absolute destination address, use the munged value.
2477 Callee = SDValue(Dest, 0);
2479 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2480 // to do the call, we can't use PPCISD::CALL.
2481 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2482 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps,
2483 2 + (InFlag.getNode() != 0));
2484 InFlag = Chain.getValue(1);
2486 // Copy the callee address into R12/X12 on darwin.
2488 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2489 Chain = DAG.getCopyToReg(Chain, Reg, Callee, InFlag);
2490 InFlag = Chain.getValue(1);
2494 NodeTys.push_back(MVT::Other);
2495 NodeTys.push_back(MVT::Flag);
2496 Ops.push_back(Chain);
2497 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2499 // Add CTR register as callee so a bctr can be emitted later.
2501 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
2504 // If this is a direct call, pass the chain and the callee.
2505 if (Callee.getNode()) {
2506 Ops.push_back(Chain);
2507 Ops.push_back(Callee);
2509 // If this is a tail call add stack pointer delta.
2511 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2513 // Add argument registers to the end of the list so that they are known live
2515 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2516 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2517 RegsToPass[i].second.getValueType()));
2519 // When performing tail call optimization the callee pops its arguments off
2520 // the stack. Account for this here so these bytes can be pushed back on in
2521 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2522 int BytesCalleePops =
2523 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2525 if (InFlag.getNode())
2526 Ops.push_back(InFlag);
2530 assert(InFlag.getNode() &&
2531 "Flag must be set. Depend on flag being set in LowerRET");
2532 Chain = DAG.getNode(PPCISD::TAILCALL,
2533 Op.getNode()->getVTList(), &Ops[0], Ops.size());
2534 return SDValue(Chain.getNode(), Op.getResNo());
2537 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
2538 InFlag = Chain.getValue(1);
2540 Chain = DAG.getCALLSEQ_END(Chain,
2541 DAG.getConstant(NumBytes, PtrVT),
2542 DAG.getConstant(BytesCalleePops, PtrVT),
2544 if (Op.getNode()->getValueType(0) != MVT::Other)
2545 InFlag = Chain.getValue(1);
2547 SmallVector<SDValue, 16> ResultVals;
2548 SmallVector<CCValAssign, 16> RVLocs;
2549 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
2550 CCState CCInfo(CallerCC, isVarArg, TM, RVLocs);
2551 CCInfo.AnalyzeCallResult(Op.getNode(), RetCC_PPC);
2553 // Copy all of the result registers out of their specified physreg.
2554 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2555 CCValAssign &VA = RVLocs[i];
2556 MVT VT = VA.getValVT();
2557 assert(VA.isRegLoc() && "Can only return in registers!");
2558 Chain = DAG.getCopyFromReg(Chain, VA.getLocReg(), VT, InFlag).getValue(1);
2559 ResultVals.push_back(Chain.getValue(0));
2560 InFlag = Chain.getValue(2);
2563 // If the function returns void, just return the chain.
2567 // Otherwise, merge everything together with a MERGE_VALUES node.
2568 ResultVals.push_back(Chain);
2569 SDValue Res = DAG.getMergeValues(Op.getNode()->getVTList(), &ResultVals[0],
2571 return Res.getValue(Op.getResNo());
2574 SDValue PPCTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG,
2575 TargetMachine &TM) {
2576 SmallVector<CCValAssign, 16> RVLocs;
2577 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2578 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2579 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2580 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_PPC);
2582 // If this is the first return lowered for this function, add the regs to the
2583 // liveout set for the function.
2584 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2585 for (unsigned i = 0; i != RVLocs.size(); ++i)
2586 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2589 SDValue Chain = Op.getOperand(0);
2591 Chain = GetPossiblePreceedingTailCall(Chain, PPCISD::TAILCALL);
2592 if (Chain.getOpcode() == PPCISD::TAILCALL) {
2593 SDValue TailCall = Chain;
2594 SDValue TargetAddress = TailCall.getOperand(1);
2595 SDValue StackAdjustment = TailCall.getOperand(2);
2597 assert(((TargetAddress.getOpcode() == ISD::Register &&
2598 cast<RegisterSDNode>(TargetAddress)->getReg() == PPC::CTR) ||
2599 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
2600 TargetAddress.getOpcode() == ISD::TargetGlobalAddress ||
2601 isa<ConstantSDNode>(TargetAddress)) &&
2602 "Expecting an global address, external symbol, absolute value or register");
2604 assert(StackAdjustment.getOpcode() == ISD::Constant &&
2605 "Expecting a const value");
2607 SmallVector<SDValue,8> Operands;
2608 Operands.push_back(Chain.getOperand(0));
2609 Operands.push_back(TargetAddress);
2610 Operands.push_back(StackAdjustment);
2611 // Copy registers used by the call. Last operand is a flag so it is not
2613 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
2614 Operands.push_back(Chain.getOperand(i));
2616 return DAG.getNode(PPCISD::TC_RETURN, MVT::Other, &Operands[0],
2622 // Copy the result values into the output registers.
2623 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2624 CCValAssign &VA = RVLocs[i];
2625 assert(VA.isRegLoc() && "Can only return in registers!");
2626 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2627 Flag = Chain.getValue(1);
2631 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2633 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2636 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
2637 const PPCSubtarget &Subtarget) {
2638 // When we pop the dynamic allocation we need to restore the SP link.
2640 // Get the corect type for pointers.
2641 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2643 // Construct the stack pointer operand.
2644 bool IsPPC64 = Subtarget.isPPC64();
2645 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2646 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
2648 // Get the operands for the STACKRESTORE.
2649 SDValue Chain = Op.getOperand(0);
2650 SDValue SaveSP = Op.getOperand(1);
2652 // Load the old link SP.
2653 SDValue LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2655 // Restore the stack pointer.
2656 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2658 // Store the old link SP.
2659 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2665 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
2666 MachineFunction &MF = DAG.getMachineFunction();
2667 bool IsPPC64 = PPCSubTarget.isPPC64();
2668 bool isMachoABI = PPCSubTarget.isMachoABI();
2669 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2671 // Get current frame pointer save index. The users of this index will be
2672 // primarily DYNALLOC instructions.
2673 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2674 int RASI = FI->getReturnAddrSaveIndex();
2676 // If the frame pointer save index hasn't been defined yet.
2678 // Find out what the fix offset of the frame pointer save area.
2679 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isMachoABI);
2680 // Allocate the frame index for frame pointer save area.
2681 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset);
2683 FI->setReturnAddrSaveIndex(RASI);
2685 return DAG.getFrameIndex(RASI, PtrVT);
2689 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
2690 MachineFunction &MF = DAG.getMachineFunction();
2691 bool IsPPC64 = PPCSubTarget.isPPC64();
2692 bool isMachoABI = PPCSubTarget.isMachoABI();
2693 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2695 // Get current frame pointer save index. The users of this index will be
2696 // primarily DYNALLOC instructions.
2697 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2698 int FPSI = FI->getFramePointerSaveIndex();
2700 // If the frame pointer save index hasn't been defined yet.
2702 // Find out what the fix offset of the frame pointer save area.
2703 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2705 // Allocate the frame index for frame pointer save area.
2706 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2708 FI->setFramePointerSaveIndex(FPSI);
2710 return DAG.getFrameIndex(FPSI, PtrVT);
2713 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
2715 const PPCSubtarget &Subtarget) {
2717 SDValue Chain = Op.getOperand(0);
2718 SDValue Size = Op.getOperand(1);
2720 // Get the corect type for pointers.
2721 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2723 SDValue NegSize = DAG.getNode(ISD::SUB, PtrVT,
2724 DAG.getConstant(0, PtrVT), Size);
2725 // Construct a node for the frame pointer save index.
2726 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
2727 // Build a DYNALLOC node.
2728 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
2729 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2730 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2733 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2735 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
2736 // Not FP? Not a fsel.
2737 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
2738 !Op.getOperand(2).getValueType().isFloatingPoint())
2741 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2743 // Cannot handle SETEQ/SETNE.
2744 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDValue();
2746 MVT ResVT = Op.getValueType();
2747 MVT CmpVT = Op.getOperand(0).getValueType();
2748 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2749 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
2751 // If the RHS of the comparison is a 0.0, we don't need to do the
2752 // subtraction at all.
2753 if (isFloatingPointZero(RHS))
2755 default: break; // SETUO etc aren't handled by fsel.
2758 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2761 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2762 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2763 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2766 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2769 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2770 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2771 return DAG.getNode(PPCISD::FSEL, ResVT,
2772 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2777 default: break; // SETUO etc aren't handled by fsel.
2780 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2781 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2782 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2783 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2786 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2787 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2788 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2789 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2792 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2793 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2794 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2795 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2798 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2799 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2800 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2801 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2806 // FIXME: Split this code up when LegalizeDAGTypes lands.
2807 SDValue PPCTargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
2808 assert(Op.getOperand(0).getValueType().isFloatingPoint());
2809 SDValue Src = Op.getOperand(0);
2810 if (Src.getValueType() == MVT::f32)
2811 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2814 switch (Op.getValueType().getSimpleVT()) {
2815 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2817 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2820 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2824 // Convert the FP value to an int value through memory.
2825 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
2827 // Emit a store to the stack slot.
2828 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2830 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2832 if (Op.getValueType() == MVT::i32)
2833 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2834 DAG.getConstant(4, FIPtr.getValueType()));
2835 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2838 SDValue PPCTargetLowering::LowerFP_ROUND_INREG(SDValue Op,
2839 SelectionDAG &DAG) {
2840 assert(Op.getValueType() == MVT::ppcf128);
2841 SDNode *Node = Op.getNode();
2842 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2843 assert(Node->getOperand(0).getNode()->getOpcode() == ISD::BUILD_PAIR);
2844 SDValue Lo = Node->getOperand(0).getNode()->getOperand(0);
2845 SDValue Hi = Node->getOperand(0).getNode()->getOperand(1);
2847 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2848 // of the long double, and puts FPSCR back the way it was. We do not
2849 // actually model FPSCR.
2850 std::vector<MVT> NodeTys;
2851 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
2853 NodeTys.push_back(MVT::f64); // Return register
2854 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2855 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2856 MFFSreg = Result.getValue(0);
2857 InFlag = Result.getValue(1);
2860 NodeTys.push_back(MVT::Flag); // Returns a flag
2861 Ops[0] = DAG.getConstant(31, MVT::i32);
2863 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2864 InFlag = Result.getValue(0);
2867 NodeTys.push_back(MVT::Flag); // Returns a flag
2868 Ops[0] = DAG.getConstant(30, MVT::i32);
2870 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2871 InFlag = Result.getValue(0);
2874 NodeTys.push_back(MVT::f64); // result of add
2875 NodeTys.push_back(MVT::Flag); // Returns a flag
2879 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2880 FPreg = Result.getValue(0);
2881 InFlag = Result.getValue(1);
2884 NodeTys.push_back(MVT::f64);
2885 Ops[0] = DAG.getConstant(1, MVT::i32);
2889 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2890 FPreg = Result.getValue(0);
2892 // We know the low half is about to be thrown away, so just use something
2894 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2897 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2898 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2899 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2902 if (Op.getOperand(0).getValueType() == MVT::i64) {
2903 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2904 SDValue FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2905 if (Op.getValueType() == MVT::f32)
2906 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2910 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2911 "Unhandled SINT_TO_FP type in custom expander!");
2912 // Since we only generate this in 64-bit mode, we can take advantage of
2913 // 64-bit registers. In particular, sign extend the input value into the
2914 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2915 // then lfd it and fcfid it.
2916 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2917 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2918 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2919 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2921 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2924 // STD the extended value into the stack slot.
2925 MachineMemOperand MO(PseudoSourceValue::getFixedStack(FrameIdx),
2926 MachineMemOperand::MOStore, 0, 8, 8);
2927 SDValue Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2928 DAG.getEntryNode(), Ext64, FIdx,
2929 DAG.getMemOperand(MO));
2930 // Load the value as a double.
2931 SDValue Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2933 // FCFID it and return it.
2934 SDValue FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2935 if (Op.getValueType() == MVT::f32)
2936 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2940 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
2942 The rounding mode is in bits 30:31 of FPSR, and has the following
2949 FLT_ROUNDS, on the other hand, expects the following:
2956 To perform the conversion, we do:
2957 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2960 MachineFunction &MF = DAG.getMachineFunction();
2961 MVT VT = Op.getValueType();
2962 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2963 std::vector<MVT> NodeTys;
2964 SDValue MFFSreg, InFlag;
2966 // Save FP Control Word to register
2967 NodeTys.push_back(MVT::f64); // return register
2968 NodeTys.push_back(MVT::Flag); // unused in this context
2969 SDValue Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2971 // Save FP register to stack slot
2972 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2973 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2974 SDValue Store = DAG.getStore(DAG.getEntryNode(), Chain,
2975 StackSlot, NULL, 0);
2977 // Load FP Control Word from low 32 bits of stack slot.
2978 SDValue Four = DAG.getConstant(4, PtrVT);
2979 SDValue Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2980 SDValue CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2982 // Transform as necessary
2984 DAG.getNode(ISD::AND, MVT::i32,
2985 CWD, DAG.getConstant(3, MVT::i32));
2987 DAG.getNode(ISD::SRL, MVT::i32,
2988 DAG.getNode(ISD::AND, MVT::i32,
2989 DAG.getNode(ISD::XOR, MVT::i32,
2990 CWD, DAG.getConstant(3, MVT::i32)),
2991 DAG.getConstant(3, MVT::i32)),
2992 DAG.getConstant(1, MVT::i8));
2995 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2997 return DAG.getNode((VT.getSizeInBits() < 16 ?
2998 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
3001 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3002 MVT VT = Op.getValueType();
3003 unsigned BitWidth = VT.getSizeInBits();
3004 assert(Op.getNumOperands() == 3 &&
3005 VT == Op.getOperand(1).getValueType() &&
3008 // Expand into a bunch of logical ops. Note that these ops
3009 // depend on the PPC behavior for oversized shift amounts.
3010 SDValue Lo = Op.getOperand(0);
3011 SDValue Hi = Op.getOperand(1);
3012 SDValue Amt = Op.getOperand(2);
3013 MVT AmtVT = Amt.getValueType();
3015 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3016 DAG.getConstant(BitWidth, AmtVT), Amt);
3017 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
3018 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
3019 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3020 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3021 DAG.getConstant(-BitWidth, AmtVT));
3022 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
3023 SDValue OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
3024 SDValue OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
3025 SDValue OutOps[] = { OutLo, OutHi };
3026 return DAG.getMergeValues(OutOps, 2);
3029 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3030 MVT VT = Op.getValueType();
3031 unsigned BitWidth = VT.getSizeInBits();
3032 assert(Op.getNumOperands() == 3 &&
3033 VT == Op.getOperand(1).getValueType() &&
3036 // Expand into a bunch of logical ops. Note that these ops
3037 // depend on the PPC behavior for oversized shift amounts.
3038 SDValue Lo = Op.getOperand(0);
3039 SDValue Hi = Op.getOperand(1);
3040 SDValue Amt = Op.getOperand(2);
3041 MVT AmtVT = Amt.getValueType();
3043 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3044 DAG.getConstant(BitWidth, AmtVT), Amt);
3045 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
3046 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
3047 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3048 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3049 DAG.getConstant(-BitWidth, AmtVT));
3050 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
3051 SDValue OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
3052 SDValue OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
3053 SDValue OutOps[] = { OutLo, OutHi };
3054 return DAG.getMergeValues(OutOps, 2);
3057 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3058 MVT VT = Op.getValueType();
3059 unsigned BitWidth = VT.getSizeInBits();
3060 assert(Op.getNumOperands() == 3 &&
3061 VT == Op.getOperand(1).getValueType() &&
3064 // Expand into a bunch of logical ops, followed by a select_cc.
3065 SDValue Lo = Op.getOperand(0);
3066 SDValue Hi = Op.getOperand(1);
3067 SDValue Amt = Op.getOperand(2);
3068 MVT AmtVT = Amt.getValueType();
3070 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3071 DAG.getConstant(BitWidth, AmtVT), Amt);
3072 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
3073 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
3074 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3075 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3076 DAG.getConstant(-BitWidth, AmtVT));
3077 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
3078 SDValue OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
3079 SDValue OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
3080 Tmp4, Tmp6, ISD::SETLE);
3081 SDValue OutOps[] = { OutLo, OutHi };
3082 return DAG.getMergeValues(OutOps, 2);
3085 //===----------------------------------------------------------------------===//
3086 // Vector related lowering.
3089 // If this is a vector of constants or undefs, get the bits. A bit in
3090 // UndefBits is set if the corresponding element of the vector is an
3091 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
3092 // zero. Return true if this is not an array of constants, false if it is.
3094 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
3095 uint64_t UndefBits[2]) {
3096 // Start with zero'd results.
3097 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
3099 unsigned EltBitSize = BV->getOperand(0).getValueType().getSizeInBits();
3100 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
3101 SDValue OpVal = BV->getOperand(i);
3103 unsigned PartNo = i >= e/2; // In the upper 128 bits?
3104 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
3106 uint64_t EltBits = 0;
3107 if (OpVal.getOpcode() == ISD::UNDEF) {
3108 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
3109 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
3111 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
3112 EltBits = CN->getZExtValue() & (~0U >> (32-EltBitSize));
3113 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
3114 assert(CN->getValueType(0) == MVT::f32 &&
3115 "Only one legal FP vector type!");
3116 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
3118 // Nonconstant element.
3122 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
3125 //printf("%llx %llx %llx %llx\n",
3126 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
3130 // If this is a splat (repetition) of a value across the whole vector, return
3131 // the smallest size that splats it. For example, "0x01010101010101..." is a
3132 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
3133 // SplatSize = 1 byte.
3134 static bool isConstantSplat(const uint64_t Bits128[2],
3135 const uint64_t Undef128[2],
3136 unsigned &SplatBits, unsigned &SplatUndef,
3137 unsigned &SplatSize) {
3139 // Don't let undefs prevent splats from matching. See if the top 64-bits are
3140 // the same as the lower 64-bits, ignoring undefs.
3141 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
3142 return false; // Can't be a splat if two pieces don't match.
3144 uint64_t Bits64 = Bits128[0] | Bits128[1];
3145 uint64_t Undef64 = Undef128[0] & Undef128[1];
3147 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
3149 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
3150 return false; // Can't be a splat if two pieces don't match.
3152 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
3153 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
3155 // If the top 16-bits are different than the lower 16-bits, ignoring
3156 // undefs, we have an i32 splat.
3157 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
3159 SplatUndef = Undef32;
3164 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
3165 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
3167 // If the top 8-bits are different than the lower 8-bits, ignoring
3168 // undefs, we have an i16 splat.
3169 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
3171 SplatUndef = Undef16;
3176 // Otherwise, we have an 8-bit splat.
3177 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
3178 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
3183 /// BuildSplatI - Build a canonical splati of Val with an element size of
3184 /// SplatSize. Cast the result to VT.
3185 static SDValue BuildSplatI(int Val, unsigned SplatSize, MVT VT,
3186 SelectionDAG &DAG) {
3187 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3189 static const MVT VTys[] = { // canonical VT to use for each size.
3190 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3193 MVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3195 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3199 MVT CanonicalVT = VTys[SplatSize-1];
3201 // Build a canonical splat for this value.
3202 SDValue Elt = DAG.getConstant(Val, CanonicalVT.getVectorElementType());
3203 SmallVector<SDValue, 8> Ops;
3204 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3205 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
3206 &Ops[0], Ops.size());
3207 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
3210 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3211 /// specified intrinsic ID.
3212 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3214 MVT DestVT = MVT::Other) {
3215 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3216 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
3217 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3220 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3221 /// specified intrinsic ID.
3222 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3223 SDValue Op2, SelectionDAG &DAG,
3224 MVT DestVT = MVT::Other) {
3225 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3226 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
3227 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3231 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3232 /// amount. The result has the specified value type.
3233 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3234 MVT VT, SelectionDAG &DAG) {
3235 // Force LHS/RHS to be the right type.
3236 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
3237 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
3240 for (unsigned i = 0; i != 16; ++i)
3241 Ops[i] = DAG.getConstant(i+Amt, MVT::i8);
3242 SDValue T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
3243 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
3244 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
3247 // If this is a case we can't handle, return null and let the default
3248 // expansion code take care of it. If we CAN select this case, and if it
3249 // selects to a single instruction, return Op. Otherwise, if we can codegen
3250 // this case more efficiently than a constant pool load, lower it to the
3251 // sequence of ops that should be used.
3252 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3253 SelectionDAG &DAG) {
3254 // If this is a vector of constants or undefs, get the bits. A bit in
3255 // UndefBits is set if the corresponding element of the vector is an
3256 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
3258 uint64_t VectorBits[2];
3259 uint64_t UndefBits[2];
3260 if (GetConstantBuildVectorBits(Op.getNode(), VectorBits, UndefBits))
3261 return SDValue(); // Not a constant vector.
3263 // If this is a splat (repetition) of a value across the whole vector, return
3264 // the smallest size that splats it. For example, "0x01010101010101..." is a
3265 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
3266 // SplatSize = 1 byte.
3267 unsigned SplatBits, SplatUndef, SplatSize;
3268 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
3269 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
3271 // First, handle single instruction cases.
3274 if (SplatBits == 0) {
3275 // Canonicalize all zero vectors to be v4i32.
3276 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3277 SDValue Z = DAG.getConstant(0, MVT::i32);
3278 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
3279 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
3284 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3285 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
3286 if (SextVal >= -16 && SextVal <= 15)
3287 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
3290 // Two instruction sequences.
3292 // If this value is in the range [-32,30] and is even, use:
3293 // tmp = VSPLTI[bhw], result = add tmp, tmp
3294 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3295 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG);
3296 Res = DAG.getNode(ISD::ADD, Res.getValueType(), Res, Res);
3297 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3300 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3301 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3303 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3304 // Make -1 and vspltisw -1:
3305 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
3307 // Make the VSLW intrinsic, computing 0x8000_0000.
3308 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3311 // xor by OnesV to invert it.
3312 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
3313 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3316 // Check to see if this is a wide variety of vsplti*, binop self cases.
3317 unsigned SplatBitSize = SplatSize*8;
3318 static const signed char SplatCsts[] = {
3319 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3320 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3323 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3324 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3325 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3326 int i = SplatCsts[idx];
3328 // Figure out what shift amount will be used by altivec if shifted by i in
3330 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3332 // vsplti + shl self.
3333 if (SextVal == (i << (int)TypeShiftAmt)) {
3334 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3335 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3336 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3337 Intrinsic::ppc_altivec_vslw
3339 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3340 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3343 // vsplti + srl self.
3344 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3345 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3346 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3347 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3348 Intrinsic::ppc_altivec_vsrw
3350 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3351 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3354 // vsplti + sra self.
3355 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3356 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3357 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3358 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3359 Intrinsic::ppc_altivec_vsraw
3361 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3362 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3365 // vsplti + rol self.
3366 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3367 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3368 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3369 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3370 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3371 Intrinsic::ppc_altivec_vrlw
3373 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3374 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3377 // t = vsplti c, result = vsldoi t, t, 1
3378 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3379 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3380 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
3382 // t = vsplti c, result = vsldoi t, t, 2
3383 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3384 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3385 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
3387 // t = vsplti c, result = vsldoi t, t, 3
3388 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3389 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3390 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
3394 // Three instruction sequences.
3396 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3397 if (SextVal >= 0 && SextVal <= 31) {
3398 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
3399 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
3400 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
3401 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
3403 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3404 if (SextVal >= -31 && SextVal <= 0) {
3405 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
3406 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
3407 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
3408 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
3415 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3416 /// the specified operations to build the shuffle.
3417 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3418 SDValue RHS, SelectionDAG &DAG) {
3419 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3420 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3421 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3424 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3436 if (OpNum == OP_COPY) {
3437 if (LHSID == (1*9+2)*9+3) return LHS;
3438 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3442 SDValue OpLHS, OpRHS;
3443 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
3444 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
3446 unsigned ShufIdxs[16];
3448 default: assert(0 && "Unknown i32 permute!");
3450 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3451 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3452 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3453 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3456 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3457 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3458 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3459 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3462 for (unsigned i = 0; i != 16; ++i)
3463 ShufIdxs[i] = (i&3)+0;
3466 for (unsigned i = 0; i != 16; ++i)
3467 ShufIdxs[i] = (i&3)+4;
3470 for (unsigned i = 0; i != 16; ++i)
3471 ShufIdxs[i] = (i&3)+8;
3474 for (unsigned i = 0; i != 16; ++i)
3475 ShufIdxs[i] = (i&3)+12;
3478 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
3480 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
3482 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
3485 for (unsigned i = 0; i != 16; ++i)
3486 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i8);
3488 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
3489 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3492 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3493 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3494 /// return the code it can be lowered into. Worst case, it can always be
3495 /// lowered into a vperm.
3496 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
3497 SelectionDAG &DAG) {
3498 SDValue V1 = Op.getOperand(0);
3499 SDValue V2 = Op.getOperand(1);
3500 SDValue PermMask = Op.getOperand(2);
3502 // Cases that are handled by instructions that take permute immediates
3503 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3504 // selected by the instruction selector.
3505 if (V2.getOpcode() == ISD::UNDEF) {
3506 if (PPC::isSplatShuffleMask(PermMask.getNode(), 1) ||
3507 PPC::isSplatShuffleMask(PermMask.getNode(), 2) ||
3508 PPC::isSplatShuffleMask(PermMask.getNode(), 4) ||
3509 PPC::isVPKUWUMShuffleMask(PermMask.getNode(), true) ||
3510 PPC::isVPKUHUMShuffleMask(PermMask.getNode(), true) ||
3511 PPC::isVSLDOIShuffleMask(PermMask.getNode(), true) != -1 ||
3512 PPC::isVMRGLShuffleMask(PermMask.getNode(), 1, true) ||
3513 PPC::isVMRGLShuffleMask(PermMask.getNode(), 2, true) ||
3514 PPC::isVMRGLShuffleMask(PermMask.getNode(), 4, true) ||
3515 PPC::isVMRGHShuffleMask(PermMask.getNode(), 1, true) ||
3516 PPC::isVMRGHShuffleMask(PermMask.getNode(), 2, true) ||
3517 PPC::isVMRGHShuffleMask(PermMask.getNode(), 4, true)) {
3522 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3523 // and produce a fixed permutation. If any of these match, do not lower to
3525 if (PPC::isVPKUWUMShuffleMask(PermMask.getNode(), false) ||
3526 PPC::isVPKUHUMShuffleMask(PermMask.getNode(), false) ||
3527 PPC::isVSLDOIShuffleMask(PermMask.getNode(), false) != -1 ||
3528 PPC::isVMRGLShuffleMask(PermMask.getNode(), 1, false) ||
3529 PPC::isVMRGLShuffleMask(PermMask.getNode(), 2, false) ||
3530 PPC::isVMRGLShuffleMask(PermMask.getNode(), 4, false) ||
3531 PPC::isVMRGHShuffleMask(PermMask.getNode(), 1, false) ||
3532 PPC::isVMRGHShuffleMask(PermMask.getNode(), 2, false) ||
3533 PPC::isVMRGHShuffleMask(PermMask.getNode(), 4, false))
3536 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3537 // perfect shuffle table to emit an optimal matching sequence.
3538 unsigned PFIndexes[4];
3539 bool isFourElementShuffle = true;
3540 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3541 unsigned EltNo = 8; // Start out undef.
3542 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3543 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
3544 continue; // Undef, ignore it.
3546 unsigned ByteSource =
3547 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getZExtValue();
3548 if ((ByteSource & 3) != j) {
3549 isFourElementShuffle = false;
3554 EltNo = ByteSource/4;
3555 } else if (EltNo != ByteSource/4) {
3556 isFourElementShuffle = false;
3560 PFIndexes[i] = EltNo;
3563 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3564 // perfect shuffle vector to determine if it is cost effective to do this as
3565 // discrete instructions, or whether we should use a vperm.
3566 if (isFourElementShuffle) {
3567 // Compute the index in the perfect shuffle table.
3568 unsigned PFTableIndex =
3569 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3571 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3572 unsigned Cost = (PFEntry >> 30);
3574 // Determining when to avoid vperm is tricky. Many things affect the cost
3575 // of vperm, particularly how many times the perm mask needs to be computed.
3576 // For example, if the perm mask can be hoisted out of a loop or is already
3577 // used (perhaps because there are multiple permutes with the same shuffle
3578 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3579 // the loop requires an extra register.
3581 // As a compromise, we only emit discrete instructions if the shuffle can be
3582 // generated in 3 or fewer operations. When we have loop information
3583 // available, if this block is within a loop, we should avoid using vperm
3584 // for 3-operation perms and use a constant pool load instead.
3586 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3589 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3590 // vector that will get spilled to the constant pool.
3591 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3593 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3594 // that it is in input element units, not in bytes. Convert now.
3595 MVT EltVT = V1.getValueType().getVectorElementType();
3596 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
3598 SmallVector<SDValue, 16> ResultMask;
3599 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3601 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3604 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getZExtValue();
3606 for (unsigned j = 0; j != BytesPerElement; ++j)
3607 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3611 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3612 &ResultMask[0], ResultMask.size());
3613 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3616 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3617 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3618 /// information about the intrinsic.
3619 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
3621 unsigned IntrinsicID =
3622 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
3625 switch (IntrinsicID) {
3626 default: return false;
3627 // Comparison predicates.
3628 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3629 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3630 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3631 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3632 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3633 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3634 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3635 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3636 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3637 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3638 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3639 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3640 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3642 // Normal Comparisons.
3643 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3644 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3645 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3646 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3647 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3648 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3649 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3650 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3651 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3652 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3653 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3654 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3655 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3660 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3661 /// lower, do it, otherwise return null.
3662 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
3663 SelectionDAG &DAG) {
3664 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3665 // opcode number of the comparison.
3668 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3669 return SDValue(); // Don't custom lower most intrinsics.
3671 // If this is a non-dot comparison, make the VCMP node and we are done.
3673 SDValue Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3674 Op.getOperand(1), Op.getOperand(2),
3675 DAG.getConstant(CompareOpc, MVT::i32));
3676 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3679 // Create the PPCISD altivec 'dot' comparison node.
3681 Op.getOperand(2), // LHS
3682 Op.getOperand(3), // RHS
3683 DAG.getConstant(CompareOpc, MVT::i32)
3685 std::vector<MVT> VTs;
3686 VTs.push_back(Op.getOperand(2).getValueType());
3687 VTs.push_back(MVT::Flag);
3688 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3690 // Now that we have the comparison, emit a copy from the CR to a GPR.
3691 // This is flagged to the above dot comparison.
3692 SDValue Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3693 DAG.getRegister(PPC::CR6, MVT::i32),
3694 CompNode.getValue(1));
3696 // Unpack the result based on how the target uses it.
3697 unsigned BitNo; // Bit # of CR6.
3698 bool InvertBit; // Invert result?
3699 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
3700 default: // Can't happen, don't crash on invalid number though.
3701 case 0: // Return the value of the EQ bit of CR6.
3702 BitNo = 0; InvertBit = false;
3704 case 1: // Return the inverted value of the EQ bit of CR6.
3705 BitNo = 0; InvertBit = true;
3707 case 2: // Return the value of the LT bit of CR6.
3708 BitNo = 2; InvertBit = false;
3710 case 3: // Return the inverted value of the LT bit of CR6.
3711 BitNo = 2; InvertBit = true;
3715 // Shift the bit into the low position.
3716 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3717 DAG.getConstant(8-(3-BitNo), MVT::i32));
3719 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3720 DAG.getConstant(1, MVT::i32));
3722 // If we are supposed to, toggle the bit.
3724 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3725 DAG.getConstant(1, MVT::i32));
3729 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
3730 SelectionDAG &DAG) {
3731 // Create a stack slot that is 16-byte aligned.
3732 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3733 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3734 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3735 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3737 // Store the input value into Value#0 of the stack slot.
3738 SDValue Store = DAG.getStore(DAG.getEntryNode(),
3739 Op.getOperand(0), FIdx, NULL, 0);
3741 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3744 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
3745 if (Op.getValueType() == MVT::v4i32) {
3746 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3748 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3749 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3751 SDValue RHSSwap = // = vrlw RHS, 16
3752 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3754 // Shrinkify inputs to v8i16.
3755 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3756 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3757 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3759 // Low parts multiplied together, generating 32-bit results (we ignore the
3761 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3762 LHS, RHS, DAG, MVT::v4i32);
3764 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3765 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3766 // Shift the high parts up 16 bits.
3767 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3768 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3769 } else if (Op.getValueType() == MVT::v8i16) {
3770 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3772 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3774 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3775 LHS, RHS, Zero, DAG);
3776 } else if (Op.getValueType() == MVT::v16i8) {
3777 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3779 // Multiply the even 8-bit parts, producing 16-bit sums.
3780 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3781 LHS, RHS, DAG, MVT::v8i16);
3782 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3784 // Multiply the odd 8-bit parts, producing 16-bit sums.
3785 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3786 LHS, RHS, DAG, MVT::v8i16);
3787 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3789 // Merge the results together.
3791 for (unsigned i = 0; i != 8; ++i) {
3792 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3793 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3795 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3796 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3798 assert(0 && "Unknown mul to lower!");
3803 /// LowerOperation - Provide custom lowering hooks for some operations.
3805 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3806 switch (Op.getOpcode()) {
3807 default: assert(0 && "Wasn't expecting to be able to lower this!");
3808 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3809 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3810 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3811 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3812 case ISD::SETCC: return LowerSETCC(Op, DAG);
3814 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3815 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3818 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3819 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3821 case ISD::FORMAL_ARGUMENTS:
3822 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3823 VarArgsStackOffset, VarArgsNumGPR,
3824 VarArgsNumFPR, PPCSubTarget);
3826 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3827 getTargetMachine());
3828 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3829 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3830 case ISD::DYNAMIC_STACKALLOC:
3831 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3833 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3834 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3835 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3836 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3837 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3839 // Lower 64-bit shifts.
3840 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3841 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3842 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3844 // Vector-related lowering.
3845 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3846 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3847 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3848 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3849 case ISD::MUL: return LowerMUL(Op, DAG);
3851 // Frame & Return address.
3852 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3853 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3858 SDNode *PPCTargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
3859 switch (N->getOpcode()) {
3860 default: assert(0 && "Wasn't expecting to be able to lower this!");
3861 case ISD::FP_TO_SINT: {
3862 SDValue Res = LowerFP_TO_SINT(SDValue(N, 0), DAG);
3863 // Use MERGE_VALUES to drop the chain result value and get a node with one
3864 // result. This requires turning off getMergeValues simplification, since
3865 // otherwise it will give us Res back.
3866 return DAG.getMergeValues(&Res, 1, false).getNode();
3872 //===----------------------------------------------------------------------===//
3873 // Other Lowering Code
3874 //===----------------------------------------------------------------------===//
3877 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3878 bool is64bit, unsigned BinOpcode) {
3879 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3880 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3882 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3883 MachineFunction *F = BB->getParent();
3884 MachineFunction::iterator It = BB;
3887 unsigned dest = MI->getOperand(0).getReg();
3888 unsigned ptrA = MI->getOperand(1).getReg();
3889 unsigned ptrB = MI->getOperand(2).getReg();
3890 unsigned incr = MI->getOperand(3).getReg();
3892 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3893 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3894 F->insert(It, loopMBB);
3895 F->insert(It, exitMBB);
3896 exitMBB->transferSuccessors(BB);
3898 MachineRegisterInfo &RegInfo = F->getRegInfo();
3899 unsigned TmpReg = (!BinOpcode) ? incr :
3900 RegInfo.createVirtualRegister(
3901 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3902 (const TargetRegisterClass *) &PPC::GPRCRegClass);
3906 // fallthrough --> loopMBB
3907 BB->addSuccessor(loopMBB);
3910 // l[wd]arx dest, ptr
3911 // add r0, dest, incr
3912 // st[wd]cx. r0, ptr
3914 // fallthrough --> exitMBB
3916 BuildMI(BB, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
3917 .addReg(ptrA).addReg(ptrB);
3919 BuildMI(BB, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
3920 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
3921 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
3922 BuildMI(BB, TII->get(PPC::BCC))
3923 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
3924 BB->addSuccessor(loopMBB);
3925 BB->addSuccessor(exitMBB);
3934 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
3935 MachineBasicBlock *BB,
3936 bool is8bit, // operation
3937 unsigned BinOpcode) {
3938 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3939 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3940 // In 64 bit mode we have to use 64 bits for addresses, even though the
3941 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
3942 // registers without caring whether they're 32 or 64, but here we're
3943 // doing actual arithmetic on the addresses.
3944 bool is64bit = PPCSubTarget.isPPC64();
3946 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3947 MachineFunction *F = BB->getParent();
3948 MachineFunction::iterator It = BB;
3951 unsigned dest = MI->getOperand(0).getReg();
3952 unsigned ptrA = MI->getOperand(1).getReg();
3953 unsigned ptrB = MI->getOperand(2).getReg();
3954 unsigned incr = MI->getOperand(3).getReg();
3956 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3957 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3958 F->insert(It, loopMBB);
3959 F->insert(It, exitMBB);
3960 exitMBB->transferSuccessors(BB);
3962 MachineRegisterInfo &RegInfo = F->getRegInfo();
3963 const TargetRegisterClass *RC =
3964 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3965 (const TargetRegisterClass *) &PPC::GPRCRegClass;
3966 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
3967 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
3968 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
3969 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
3970 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
3971 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
3972 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
3973 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
3974 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
3975 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
3976 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
3978 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
3982 // fallthrough --> loopMBB
3983 BB->addSuccessor(loopMBB);
3985 // The 4-byte load must be aligned, while a char or short may be
3986 // anywhere in the word. Hence all this nasty bookkeeping code.
3987 // add ptr1, ptrA, ptrB [copy if ptrA==0]
3988 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
3989 // xori shift, shift1, 24 [16]
3990 // rlwinm ptr, ptr1, 0, 0, 29
3991 // slw incr2, incr, shift
3992 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
3993 // slw mask, mask2, shift
3995 // lwarx tmpDest, ptr
3996 // add tmp, tmpDest, incr2
3997 // andc tmp2, tmpDest, mask
3998 // and tmp3, tmp, mask
3999 // or tmp4, tmp3, tmp2
4002 // fallthrough --> exitMBB
4003 // srw dest, tmpDest, shift
4005 if (ptrA!=PPC::R0) {
4006 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4007 BuildMI(BB, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4008 .addReg(ptrA).addReg(ptrB);
4012 BuildMI(BB, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4013 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4014 BuildMI(BB, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4015 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4017 BuildMI(BB, TII->get(PPC::RLDICR), PtrReg)
4018 .addReg(Ptr1Reg).addImm(0).addImm(61);
4020 BuildMI(BB, TII->get(PPC::RLWINM), PtrReg)
4021 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4022 BuildMI(BB, TII->get(PPC::SLW), Incr2Reg)
4023 .addReg(incr).addReg(ShiftReg);
4025 BuildMI(BB, TII->get(PPC::LI), Mask2Reg).addImm(255);
4027 BuildMI(BB, TII->get(PPC::LI), Mask3Reg).addImm(0);
4028 BuildMI(BB, TII->get(PPC::ORI), Mask2Reg).addReg(Mask3Reg).addImm(65535);
4030 BuildMI(BB, TII->get(PPC::SLW), MaskReg)
4031 .addReg(Mask2Reg).addReg(ShiftReg);
4034 BuildMI(BB, TII->get(PPC::LWARX), TmpDestReg)
4035 .addReg(PPC::R0).addReg(PtrReg);
4037 BuildMI(BB, TII->get(BinOpcode), TmpReg)
4038 .addReg(Incr2Reg).addReg(TmpDestReg);
4039 BuildMI(BB, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4040 .addReg(TmpDestReg).addReg(MaskReg);
4041 BuildMI(BB, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4042 .addReg(TmpReg).addReg(MaskReg);
4043 BuildMI(BB, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4044 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4045 BuildMI(BB, TII->get(PPC::STWCX))
4046 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4047 BuildMI(BB, TII->get(PPC::BCC))
4048 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4049 BB->addSuccessor(loopMBB);
4050 BB->addSuccessor(exitMBB);
4055 BuildMI(BB, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4060 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4061 MachineBasicBlock *BB) {
4062 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4064 // To "insert" these instructions we actually have to insert their
4065 // control-flow patterns.
4066 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4067 MachineFunction::iterator It = BB;
4070 MachineFunction *F = BB->getParent();
4072 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4073 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4074 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4075 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4076 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4078 // The incoming instruction knows the destination vreg to set, the
4079 // condition code register to branch on, the true/false values to
4080 // select between, and a branch opcode to use.
4085 // cmpTY ccX, r1, r2
4087 // fallthrough --> copy0MBB
4088 MachineBasicBlock *thisMBB = BB;
4089 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4090 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4091 unsigned SelectPred = MI->getOperand(4).getImm();
4092 BuildMI(BB, TII->get(PPC::BCC))
4093 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4094 F->insert(It, copy0MBB);
4095 F->insert(It, sinkMBB);
4096 // Update machine-CFG edges by transferring all successors of the current
4097 // block to the new block which will contain the Phi node for the select.
4098 sinkMBB->transferSuccessors(BB);
4099 // Next, add the true and fallthrough blocks as its successors.
4100 BB->addSuccessor(copy0MBB);
4101 BB->addSuccessor(sinkMBB);
4104 // %FalseValue = ...
4105 // # fallthrough to sinkMBB
4108 // Update machine-CFG edges
4109 BB->addSuccessor(sinkMBB);
4112 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4115 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4116 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4117 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4119 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4120 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4121 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4122 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4123 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4124 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4125 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4126 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4128 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4129 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4130 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4131 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4132 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4133 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4134 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4135 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4137 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4138 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4139 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4140 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4141 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4142 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4143 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4144 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4146 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4147 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4148 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4149 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4150 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4151 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4152 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4153 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4155 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4156 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4157 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4158 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4159 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4160 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4161 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4162 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4164 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4165 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4166 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4167 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4168 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4169 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4170 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4171 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4173 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4174 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4175 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4176 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4177 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4178 BB = EmitAtomicBinary(MI, BB, false, 0);
4179 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4180 BB = EmitAtomicBinary(MI, BB, true, 0);
4182 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4183 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4184 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4186 unsigned dest = MI->getOperand(0).getReg();
4187 unsigned ptrA = MI->getOperand(1).getReg();
4188 unsigned ptrB = MI->getOperand(2).getReg();
4189 unsigned oldval = MI->getOperand(3).getReg();
4190 unsigned newval = MI->getOperand(4).getReg();
4192 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4193 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4194 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4195 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4196 F->insert(It, loop1MBB);
4197 F->insert(It, loop2MBB);
4198 F->insert(It, midMBB);
4199 F->insert(It, exitMBB);
4200 exitMBB->transferSuccessors(BB);
4204 // fallthrough --> loopMBB
4205 BB->addSuccessor(loop1MBB);
4208 // l[wd]arx dest, ptr
4209 // cmp[wd] dest, oldval
4212 // st[wd]cx. newval, ptr
4216 // st[wd]cx. dest, ptr
4219 BuildMI(BB, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4220 .addReg(ptrA).addReg(ptrB);
4221 BuildMI(BB, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4222 .addReg(oldval).addReg(dest);
4223 BuildMI(BB, TII->get(PPC::BCC))
4224 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4225 BB->addSuccessor(loop2MBB);
4226 BB->addSuccessor(midMBB);
4229 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4230 .addReg(newval).addReg(ptrA).addReg(ptrB);
4231 BuildMI(BB, TII->get(PPC::BCC))
4232 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4233 BuildMI(BB, TII->get(PPC::B)).addMBB(exitMBB);
4234 BB->addSuccessor(loop1MBB);
4235 BB->addSuccessor(exitMBB);
4238 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4239 .addReg(dest).addReg(ptrA).addReg(ptrB);
4240 BB->addSuccessor(exitMBB);
4245 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4246 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4247 // We must use 64-bit registers for addresses when targeting 64-bit,
4248 // since we're actually doing arithmetic on them. Other registers
4250 bool is64bit = PPCSubTarget.isPPC64();
4251 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4253 unsigned dest = MI->getOperand(0).getReg();
4254 unsigned ptrA = MI->getOperand(1).getReg();
4255 unsigned ptrB = MI->getOperand(2).getReg();
4256 unsigned oldval = MI->getOperand(3).getReg();
4257 unsigned newval = MI->getOperand(4).getReg();
4259 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4260 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4261 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4262 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4263 F->insert(It, loop1MBB);
4264 F->insert(It, loop2MBB);
4265 F->insert(It, midMBB);
4266 F->insert(It, exitMBB);
4267 exitMBB->transferSuccessors(BB);
4269 MachineRegisterInfo &RegInfo = F->getRegInfo();
4270 const TargetRegisterClass *RC =
4271 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4272 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4273 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4274 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4275 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4276 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4277 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4278 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4279 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4280 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4281 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4282 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4283 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4284 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4285 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4287 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4290 // fallthrough --> loopMBB
4291 BB->addSuccessor(loop1MBB);
4293 // The 4-byte load must be aligned, while a char or short may be
4294 // anywhere in the word. Hence all this nasty bookkeeping code.
4295 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4296 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4297 // xori shift, shift1, 24 [16]
4298 // rlwinm ptr, ptr1, 0, 0, 29
4299 // slw newval2, newval, shift
4300 // slw oldval2, oldval,shift
4301 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4302 // slw mask, mask2, shift
4303 // and newval3, newval2, mask
4304 // and oldval3, oldval2, mask
4306 // lwarx tmpDest, ptr
4307 // and tmp, tmpDest, mask
4308 // cmpw tmp, oldval3
4311 // andc tmp2, tmpDest, mask
4312 // or tmp4, tmp2, newval3
4317 // stwcx. tmpDest, ptr
4319 // srw dest, tmpDest, shift
4320 if (ptrA!=PPC::R0) {
4321 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4322 BuildMI(BB, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4323 .addReg(ptrA).addReg(ptrB);
4327 BuildMI(BB, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4328 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4329 BuildMI(BB, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4330 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4332 BuildMI(BB, TII->get(PPC::RLDICR), PtrReg)
4333 .addReg(Ptr1Reg).addImm(0).addImm(61);
4335 BuildMI(BB, TII->get(PPC::RLWINM), PtrReg)
4336 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4337 BuildMI(BB, TII->get(PPC::SLW), NewVal2Reg)
4338 .addReg(newval).addReg(ShiftReg);
4339 BuildMI(BB, TII->get(PPC::SLW), OldVal2Reg)
4340 .addReg(oldval).addReg(ShiftReg);
4342 BuildMI(BB, TII->get(PPC::LI), Mask2Reg).addImm(255);
4344 BuildMI(BB, TII->get(PPC::LI), Mask3Reg).addImm(0);
4345 BuildMI(BB, TII->get(PPC::ORI), Mask2Reg).addReg(Mask3Reg).addImm(65535);
4347 BuildMI(BB, TII->get(PPC::SLW), MaskReg)
4348 .addReg(Mask2Reg).addReg(ShiftReg);
4349 BuildMI(BB, TII->get(PPC::AND), NewVal3Reg)
4350 .addReg(NewVal2Reg).addReg(MaskReg);
4351 BuildMI(BB, TII->get(PPC::AND), OldVal3Reg)
4352 .addReg(OldVal2Reg).addReg(MaskReg);
4355 BuildMI(BB, TII->get(PPC::LWARX), TmpDestReg)
4356 .addReg(PPC::R0).addReg(PtrReg);
4357 BuildMI(BB, TII->get(PPC::AND),TmpReg).addReg(TmpDestReg).addReg(MaskReg);
4358 BuildMI(BB, TII->get(PPC::CMPW), PPC::CR0)
4359 .addReg(TmpReg).addReg(OldVal3Reg);
4360 BuildMI(BB, TII->get(PPC::BCC))
4361 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4362 BB->addSuccessor(loop2MBB);
4363 BB->addSuccessor(midMBB);
4366 BuildMI(BB, TII->get(PPC::ANDC),Tmp2Reg).addReg(TmpDestReg).addReg(MaskReg);
4367 BuildMI(BB, TII->get(PPC::OR),Tmp4Reg).addReg(Tmp2Reg).addReg(NewVal3Reg);
4368 BuildMI(BB, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4369 .addReg(PPC::R0).addReg(PtrReg);
4370 BuildMI(BB, TII->get(PPC::BCC))
4371 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4372 BuildMI(BB, TII->get(PPC::B)).addMBB(exitMBB);
4373 BB->addSuccessor(loop1MBB);
4374 BB->addSuccessor(exitMBB);
4377 BuildMI(BB, TII->get(PPC::STWCX)).addReg(TmpDestReg)
4378 .addReg(PPC::R0).addReg(PtrReg);
4379 BB->addSuccessor(exitMBB);
4384 BuildMI(BB, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
4386 assert(0 && "Unexpected instr type to insert");
4389 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
4393 //===----------------------------------------------------------------------===//
4394 // Target Optimization Hooks
4395 //===----------------------------------------------------------------------===//
4397 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
4398 DAGCombinerInfo &DCI) const {
4399 TargetMachine &TM = getTargetMachine();
4400 SelectionDAG &DAG = DCI.DAG;
4401 switch (N->getOpcode()) {
4404 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4405 if (C->getZExtValue() == 0) // 0 << V -> 0.
4406 return N->getOperand(0);
4410 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4411 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
4412 return N->getOperand(0);
4416 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4417 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
4418 C->isAllOnesValue()) // -1 >>s V -> -1.
4419 return N->getOperand(0);
4423 case ISD::SINT_TO_FP:
4424 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
4425 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
4426 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
4427 // We allow the src/dst to be either f32/f64, but the intermediate
4428 // type must be i64.
4429 if (N->getOperand(0).getValueType() == MVT::i64 &&
4430 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
4431 SDValue Val = N->getOperand(0).getOperand(0);
4432 if (Val.getValueType() == MVT::f32) {
4433 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
4434 DCI.AddToWorklist(Val.getNode());
4437 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
4438 DCI.AddToWorklist(Val.getNode());
4439 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
4440 DCI.AddToWorklist(Val.getNode());
4441 if (N->getValueType(0) == MVT::f32) {
4442 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
4443 DAG.getIntPtrConstant(0));
4444 DCI.AddToWorklist(Val.getNode());
4447 } else if (N->getOperand(0).getValueType() == MVT::i32) {
4448 // If the intermediate type is i32, we can avoid the load/store here
4455 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
4456 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
4457 !cast<StoreSDNode>(N)->isTruncatingStore() &&
4458 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
4459 N->getOperand(1).getValueType() == MVT::i32 &&
4460 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
4461 SDValue Val = N->getOperand(1).getOperand(0);
4462 if (Val.getValueType() == MVT::f32) {
4463 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
4464 DCI.AddToWorklist(Val.getNode());
4466 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
4467 DCI.AddToWorklist(Val.getNode());
4469 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
4470 N->getOperand(2), N->getOperand(3));
4471 DCI.AddToWorklist(Val.getNode());
4475 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
4476 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
4477 N->getOperand(1).getNode()->hasOneUse() &&
4478 (N->getOperand(1).getValueType() == MVT::i32 ||
4479 N->getOperand(1).getValueType() == MVT::i16)) {
4480 SDValue BSwapOp = N->getOperand(1).getOperand(0);
4481 // Do an any-extend to 32-bits if this is a half-word input.
4482 if (BSwapOp.getValueType() == MVT::i16)
4483 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
4485 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
4486 N->getOperand(2), N->getOperand(3),
4487 DAG.getValueType(N->getOperand(1).getValueType()));
4491 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
4492 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
4493 N->getOperand(0).hasOneUse() &&
4494 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
4495 SDValue Load = N->getOperand(0);
4496 LoadSDNode *LD = cast<LoadSDNode>(Load);
4497 // Create the byte-swapping load.
4498 std::vector<MVT> VTs;
4499 VTs.push_back(MVT::i32);
4500 VTs.push_back(MVT::Other);
4501 SDValue MO = DAG.getMemOperand(LD->getMemOperand());
4503 LD->getChain(), // Chain
4504 LD->getBasePtr(), // Ptr
4506 DAG.getValueType(N->getValueType(0)) // VT
4508 SDValue BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
4510 // If this is an i16 load, insert the truncate.
4511 SDValue ResVal = BSLoad;
4512 if (N->getValueType(0) == MVT::i16)
4513 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
4515 // First, combine the bswap away. This makes the value produced by the
4517 DCI.CombineTo(N, ResVal);
4519 // Next, combine the load away, we give it a bogus result value but a real
4520 // chain result. The result value is dead because the bswap is dead.
4521 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
4523 // Return N so it doesn't get rechecked!
4524 return SDValue(N, 0);
4528 case PPCISD::VCMP: {
4529 // If a VCMPo node already exists with exactly the same operands as this
4530 // node, use its result instead of this node (VCMPo computes both a CR6 and
4531 // a normal output).
4533 if (!N->getOperand(0).hasOneUse() &&
4534 !N->getOperand(1).hasOneUse() &&
4535 !N->getOperand(2).hasOneUse()) {
4537 // Scan all of the users of the LHS, looking for VCMPo's that match.
4538 SDNode *VCMPoNode = 0;
4540 SDNode *LHSN = N->getOperand(0).getNode();
4541 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
4543 if (UI->getOpcode() == PPCISD::VCMPo &&
4544 UI->getOperand(1) == N->getOperand(1) &&
4545 UI->getOperand(2) == N->getOperand(2) &&
4546 UI->getOperand(0) == N->getOperand(0)) {
4551 // If there is no VCMPo node, or if the flag value has a single use, don't
4553 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
4556 // Look at the (necessarily single) use of the flag value. If it has a
4557 // chain, this transformation is more complex. Note that multiple things
4558 // could use the value result, which we should ignore.
4559 SDNode *FlagUser = 0;
4560 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
4561 FlagUser == 0; ++UI) {
4562 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
4564 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
4565 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
4572 // If the user is a MFCR instruction, we know this is safe. Otherwise we
4573 // give up for right now.
4574 if (FlagUser->getOpcode() == PPCISD::MFCR)
4575 return SDValue(VCMPoNode, 0);
4580 // If this is a branch on an altivec predicate comparison, lower this so
4581 // that we don't have to do a MFCR: instead, branch directly on CR6. This
4582 // lowering is done pre-legalize, because the legalizer lowers the predicate
4583 // compare down to code that is difficult to reassemble.
4584 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
4585 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
4589 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
4590 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
4591 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
4592 assert(isDot && "Can't compare against a vector result!");
4594 // If this is a comparison against something other than 0/1, then we know
4595 // that the condition is never/always true.
4596 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
4597 if (Val != 0 && Val != 1) {
4598 if (CC == ISD::SETEQ) // Cond never true, remove branch.
4599 return N->getOperand(0);
4600 // Always !=, turn it into an unconditional branch.
4601 return DAG.getNode(ISD::BR, MVT::Other,
4602 N->getOperand(0), N->getOperand(4));
4605 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
4607 // Create the PPCISD altivec 'dot' comparison node.
4608 std::vector<MVT> VTs;
4610 LHS.getOperand(2), // LHS of compare
4611 LHS.getOperand(3), // RHS of compare
4612 DAG.getConstant(CompareOpc, MVT::i32)
4614 VTs.push_back(LHS.getOperand(2).getValueType());
4615 VTs.push_back(MVT::Flag);
4616 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
4618 // Unpack the result based on how the target uses it.
4619 PPC::Predicate CompOpc;
4620 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
4621 default: // Can't happen, don't crash on invalid number though.
4622 case 0: // Branch on the value of the EQ bit of CR6.
4623 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
4625 case 1: // Branch on the inverted value of the EQ bit of CR6.
4626 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
4628 case 2: // Branch on the value of the LT bit of CR6.
4629 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
4631 case 3: // Branch on the inverted value of the LT bit of CR6.
4632 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
4636 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
4637 DAG.getConstant(CompOpc, MVT::i32),
4638 DAG.getRegister(PPC::CR6, MVT::i32),
4639 N->getOperand(4), CompNode.getValue(1));
4648 //===----------------------------------------------------------------------===//
4649 // Inline Assembly Support
4650 //===----------------------------------------------------------------------===//
4652 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4656 const SelectionDAG &DAG,
4657 unsigned Depth) const {
4658 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4659 switch (Op.getOpcode()) {
4661 case PPCISD::LBRX: {
4662 // lhbrx is known to have the top bits cleared out.
4663 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
4664 KnownZero = 0xFFFF0000;
4667 case ISD::INTRINSIC_WO_CHAIN: {
4668 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
4670 case Intrinsic::ppc_altivec_vcmpbfp_p:
4671 case Intrinsic::ppc_altivec_vcmpeqfp_p:
4672 case Intrinsic::ppc_altivec_vcmpequb_p:
4673 case Intrinsic::ppc_altivec_vcmpequh_p:
4674 case Intrinsic::ppc_altivec_vcmpequw_p:
4675 case Intrinsic::ppc_altivec_vcmpgefp_p:
4676 case Intrinsic::ppc_altivec_vcmpgtfp_p:
4677 case Intrinsic::ppc_altivec_vcmpgtsb_p:
4678 case Intrinsic::ppc_altivec_vcmpgtsh_p:
4679 case Intrinsic::ppc_altivec_vcmpgtsw_p:
4680 case Intrinsic::ppc_altivec_vcmpgtub_p:
4681 case Intrinsic::ppc_altivec_vcmpgtuh_p:
4682 case Intrinsic::ppc_altivec_vcmpgtuw_p:
4683 KnownZero = ~1U; // All bits but the low one are known to be zero.
4691 /// getConstraintType - Given a constraint, return the type of
4692 /// constraint it is for this target.
4693 PPCTargetLowering::ConstraintType
4694 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
4695 if (Constraint.size() == 1) {
4696 switch (Constraint[0]) {
4703 return C_RegisterClass;
4706 return TargetLowering::getConstraintType(Constraint);
4709 std::pair<unsigned, const TargetRegisterClass*>
4710 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4712 if (Constraint.size() == 1) {
4713 // GCC RS6000 Constraint Letters
4714 switch (Constraint[0]) {
4717 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
4718 return std::make_pair(0U, PPC::G8RCRegisterClass);
4719 return std::make_pair(0U, PPC::GPRCRegisterClass);
4722 return std::make_pair(0U, PPC::F4RCRegisterClass);
4723 else if (VT == MVT::f64)
4724 return std::make_pair(0U, PPC::F8RCRegisterClass);
4727 return std::make_pair(0U, PPC::VRRCRegisterClass);
4729 return std::make_pair(0U, PPC::CRRCRegisterClass);
4733 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4737 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4738 /// vector. If it is invalid, don't add anything to Ops.
4739 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
4740 std::vector<SDValue>&Ops,
4741 SelectionDAG &DAG) const {
4742 SDValue Result(0,0);
4753 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
4754 if (!CST) return; // Must be an immediate to match.
4755 unsigned Value = CST->getZExtValue();
4757 default: assert(0 && "Unknown constraint letter!");
4758 case 'I': // "I" is a signed 16-bit constant.
4759 if ((short)Value == (int)Value)
4760 Result = DAG.getTargetConstant(Value, Op.getValueType());
4762 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
4763 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
4764 if ((short)Value == 0)
4765 Result = DAG.getTargetConstant(Value, Op.getValueType());
4767 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
4768 if ((Value >> 16) == 0)
4769 Result = DAG.getTargetConstant(Value, Op.getValueType());
4771 case 'M': // "M" is a constant that is greater than 31.
4773 Result = DAG.getTargetConstant(Value, Op.getValueType());
4775 case 'N': // "N" is a positive constant that is an exact power of two.
4776 if ((int)Value > 0 && isPowerOf2_32(Value))
4777 Result = DAG.getTargetConstant(Value, Op.getValueType());
4779 case 'O': // "O" is the constant zero.
4781 Result = DAG.getTargetConstant(Value, Op.getValueType());
4783 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
4784 if ((short)-Value == (int)-Value)
4785 Result = DAG.getTargetConstant(Value, Op.getValueType());
4792 if (Result.getNode()) {
4793 Ops.push_back(Result);
4797 // Handle standard constraint letters.
4798 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
4801 // isLegalAddressingMode - Return true if the addressing mode represented
4802 // by AM is legal for this target, for a load/store of the specified type.
4803 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4804 const Type *Ty) const {
4805 // FIXME: PPC does not allow r+i addressing modes for vectors!
4807 // PPC allows a sign-extended 16-bit immediate field.
4808 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
4811 // No global is ever allowed as a base.
4815 // PPC only support r+r,
4817 case 0: // "r+i" or just "i", depending on HasBaseReg.
4820 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
4822 // Otherwise we have r+r or r+i.
4825 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
4827 // Allow 2*r as r+r.
4830 // No other scales are supported.
4837 /// isLegalAddressImmediate - Return true if the integer value can be used
4838 /// as the offset of the target addressing mode for load / store of the
4840 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
4841 // PPC allows a sign-extended 16-bit immediate field.
4842 return (V > -(1 << 16) && V < (1 << 16)-1);
4845 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
4849 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
4850 // Depths > 0 not supported yet!
4851 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4854 MachineFunction &MF = DAG.getMachineFunction();
4855 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4857 // Just load the return address off the stack.
4858 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
4860 // Make sure the function really does not optimize away the store of the RA
4862 FuncInfo->setLRStoreRequired();
4863 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4866 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
4867 // Depths > 0 not supported yet!
4868 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4871 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4872 bool isPPC64 = PtrVT == MVT::i64;
4874 MachineFunction &MF = DAG.getMachineFunction();
4875 MachineFrameInfo *MFI = MF.getFrameInfo();
4876 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4877 && MFI->getStackSize();
4880 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
4883 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,