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/Analysis/ScalarEvolutionExpressions.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/PseudoSourceValue.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/Constants.h"
30 #include "llvm/Function.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Support/CommandLine.h"
37 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
38 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
41 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
42 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
46 // Use _setjmp/_longjmp instead of setjmp/longjmp.
47 setUseUnderscoreSetJmp(true);
48 setUseUnderscoreLongJmp(true);
50 // Set up the register classes.
51 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
52 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
53 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
55 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
56 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
57 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Expand);
59 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
61 // PowerPC has pre-inc load and store's.
62 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
63 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
64 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
65 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
66 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
67 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
68 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
69 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
70 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
71 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
73 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
74 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
75 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
76 // This is used in the ppcf128->int sequence. Note it has different semantics
77 // from FP_ROUND: that rounds to nearest, this rounds to zero.
78 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
80 // PowerPC has no intrinsics for these particular operations
81 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
82 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
83 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
84 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
86 // PowerPC has no SREM/UREM instructions
87 setOperationAction(ISD::SREM, MVT::i32, Expand);
88 setOperationAction(ISD::UREM, MVT::i32, Expand);
89 setOperationAction(ISD::SREM, MVT::i64, Expand);
90 setOperationAction(ISD::UREM, MVT::i64, Expand);
92 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
93 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
94 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
95 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
96 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
97 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
98 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
99 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
100 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
102 // We don't support sin/cos/sqrt/fmod/pow
103 setOperationAction(ISD::FSIN , MVT::f64, Expand);
104 setOperationAction(ISD::FCOS , MVT::f64, Expand);
105 setOperationAction(ISD::FREM , MVT::f64, Expand);
106 setOperationAction(ISD::FPOW , 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);
112 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
114 // If we're enabling GP optimizations, use hardware square root
115 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
116 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
117 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
120 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
121 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
123 // PowerPC does not have BSWAP, CTPOP or CTTZ
124 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
125 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
126 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
127 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
128 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
129 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
131 // PowerPC does not have ROTR
132 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
134 // PowerPC does not have Select
135 setOperationAction(ISD::SELECT, MVT::i32, Expand);
136 setOperationAction(ISD::SELECT, MVT::i64, Expand);
137 setOperationAction(ISD::SELECT, MVT::f32, Expand);
138 setOperationAction(ISD::SELECT, MVT::f64, Expand);
140 // PowerPC wants to turn select_cc of FP into fsel when possible.
141 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
142 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
144 // PowerPC wants to optimize integer setcc a bit
145 setOperationAction(ISD::SETCC, MVT::i32, Custom);
147 // PowerPC does not have BRCOND which requires SetCC
148 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
150 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
152 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
153 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
155 // PowerPC does not have [U|S]INT_TO_FP
156 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
157 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
159 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
160 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
161 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
162 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
164 // We cannot sextinreg(i1). Expand to shifts.
165 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
167 // Support label based line numbers.
168 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
169 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
171 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
172 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
173 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
174 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
177 // We want to legalize GlobalAddress and ConstantPool nodes into the
178 // appropriate instructions to materialize the address.
179 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
180 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
181 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
182 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
183 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
184 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
185 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
186 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
188 // RET must be custom lowered, to meet ABI requirements
189 setOperationAction(ISD::RET , MVT::Other, Custom);
191 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
192 setOperationAction(ISD::VASTART , MVT::Other, Custom);
194 // VAARG is custom lowered with ELF 32 ABI
195 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
196 setOperationAction(ISD::VAARG, MVT::Other, Custom);
198 setOperationAction(ISD::VAARG, MVT::Other, Expand);
200 // Use the default implementation.
201 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
202 setOperationAction(ISD::VAEND , MVT::Other, Expand);
203 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
204 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
205 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
206 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
208 // We want to custom lower some of our intrinsics.
209 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
211 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
212 // They also have instructions for converting between i64 and fp.
213 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
214 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
215 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
216 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
217 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
219 // FIXME: disable this lowered code. This generates 64-bit register values,
220 // and we don't model the fact that the top part is clobbered by calls. We
221 // need to flag these together so that the value isn't live across a call.
222 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
224 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
225 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
227 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
228 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
231 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
232 // 64-bit PowerPC implementations can support i64 types directly
233 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
234 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
235 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
236 // 64-bit PowerPC wants to expand i128 shifts itself.
237 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
238 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
239 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
241 // 32-bit PowerPC wants to expand i64 shifts itself.
242 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
243 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
244 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
247 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
248 // First set operation action for all vector types to expand. Then we
249 // will selectively turn on ones that can be effectively codegen'd.
250 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
251 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
252 // add/sub are legal for all supported vector VT's.
253 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
254 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
256 // We promote all shuffles to v16i8.
257 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
258 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
260 // We promote all non-typed operations to v4i32.
261 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
262 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
263 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
264 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
265 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
266 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
267 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
268 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
269 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
270 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
271 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
272 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
274 // No other operations are legal.
275 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
276 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
277 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
278 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
279 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
280 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
281 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
282 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
283 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
284 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
285 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
286 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
287 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
288 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
289 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
290 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
291 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
292 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
293 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
296 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
297 // with merges, splats, etc.
298 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
300 setOperationAction(ISD::AND , MVT::v4i32, Legal);
301 setOperationAction(ISD::OR , MVT::v4i32, Legal);
302 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
303 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
304 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
305 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
307 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
308 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
309 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
310 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
312 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
313 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
314 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
315 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
317 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
318 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
320 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
321 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
322 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
323 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
326 setShiftAmountType(MVT::i32);
327 setSetCCResultContents(ZeroOrOneSetCCResult);
329 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
330 setStackPointerRegisterToSaveRestore(PPC::X1);
331 setExceptionPointerRegister(PPC::X3);
332 setExceptionSelectorRegister(PPC::X4);
334 setStackPointerRegisterToSaveRestore(PPC::R1);
335 setExceptionPointerRegister(PPC::R3);
336 setExceptionSelectorRegister(PPC::R4);
339 // We have target-specific dag combine patterns for the following nodes:
340 setTargetDAGCombine(ISD::SINT_TO_FP);
341 setTargetDAGCombine(ISD::STORE);
342 setTargetDAGCombine(ISD::BR_CC);
343 setTargetDAGCombine(ISD::BSWAP);
345 // Darwin long double math library functions have $LDBL128 appended.
346 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
347 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
348 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
349 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
350 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
351 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
354 computeRegisterProperties();
357 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
358 /// function arguments in the caller parameter area.
359 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
360 TargetMachine &TM = getTargetMachine();
361 // Darwin passes everything on 4 byte boundary.
362 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
368 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
371 case PPCISD::FSEL: return "PPCISD::FSEL";
372 case PPCISD::FCFID: return "PPCISD::FCFID";
373 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
374 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
375 case PPCISD::STFIWX: return "PPCISD::STFIWX";
376 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
377 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
378 case PPCISD::VPERM: return "PPCISD::VPERM";
379 case PPCISD::Hi: return "PPCISD::Hi";
380 case PPCISD::Lo: return "PPCISD::Lo";
381 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
382 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
383 case PPCISD::SRL: return "PPCISD::SRL";
384 case PPCISD::SRA: return "PPCISD::SRA";
385 case PPCISD::SHL: return "PPCISD::SHL";
386 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
387 case PPCISD::STD_32: return "PPCISD::STD_32";
388 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
389 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
390 case PPCISD::MTCTR: return "PPCISD::MTCTR";
391 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
392 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
393 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
394 case PPCISD::MFCR: return "PPCISD::MFCR";
395 case PPCISD::VCMP: return "PPCISD::VCMP";
396 case PPCISD::VCMPo: return "PPCISD::VCMPo";
397 case PPCISD::LBRX: return "PPCISD::LBRX";
398 case PPCISD::STBRX: return "PPCISD::STBRX";
399 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
400 case PPCISD::MFFS: return "PPCISD::MFFS";
401 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
402 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
403 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
404 case PPCISD::MTFSF: return "PPCISD::MTFSF";
410 PPCTargetLowering::getSetCCResultType(const SDOperand &) const {
415 //===----------------------------------------------------------------------===//
416 // Node matching predicates, for use by the tblgen matching code.
417 //===----------------------------------------------------------------------===//
419 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
420 static bool isFloatingPointZero(SDOperand Op) {
421 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
422 return CFP->getValueAPF().isZero();
423 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
424 // Maybe this has already been legalized into the constant pool?
425 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
426 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
427 return CFP->getValueAPF().isZero();
432 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
433 /// true if Op is undef or if it matches the specified value.
434 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
435 return Op.getOpcode() == ISD::UNDEF ||
436 cast<ConstantSDNode>(Op)->getValue() == Val;
439 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
440 /// VPKUHUM instruction.
441 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
443 for (unsigned i = 0; i != 16; ++i)
444 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
447 for (unsigned i = 0; i != 8; ++i)
448 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
449 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
455 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
456 /// VPKUWUM instruction.
457 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
459 for (unsigned i = 0; i != 16; i += 2)
460 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
461 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
464 for (unsigned i = 0; i != 8; i += 2)
465 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
466 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
467 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
468 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
474 /// isVMerge - Common function, used to match vmrg* shuffles.
476 static bool isVMerge(SDNode *N, unsigned UnitSize,
477 unsigned LHSStart, unsigned RHSStart) {
478 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
479 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
480 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
481 "Unsupported merge size!");
483 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
484 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
485 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
486 LHSStart+j+i*UnitSize) ||
487 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
488 RHSStart+j+i*UnitSize))
494 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
495 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
496 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
498 return isVMerge(N, UnitSize, 8, 24);
499 return isVMerge(N, UnitSize, 8, 8);
502 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
503 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
504 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
506 return isVMerge(N, UnitSize, 0, 16);
507 return isVMerge(N, UnitSize, 0, 0);
511 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
512 /// amount, otherwise return -1.
513 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
514 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
515 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
516 // Find the first non-undef value in the shuffle mask.
518 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
521 if (i == 16) return -1; // all undef.
523 // Otherwise, check to see if the rest of the elements are consequtively
524 // numbered from this value.
525 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
526 if (ShiftAmt < i) return -1;
530 // Check the rest of the elements to see if they are consequtive.
531 for (++i; i != 16; ++i)
532 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
535 // Check the rest of the elements to see if they are consequtive.
536 for (++i; i != 16; ++i)
537 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
544 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
545 /// specifies a splat of a single element that is suitable for input to
546 /// VSPLTB/VSPLTH/VSPLTW.
547 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
548 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
549 N->getNumOperands() == 16 &&
550 (EltSize == 1 || EltSize == 2 || EltSize == 4));
552 // This is a splat operation if each element of the permute is the same, and
553 // if the value doesn't reference the second vector.
554 unsigned ElementBase = 0;
555 SDOperand Elt = N->getOperand(0);
556 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
557 ElementBase = EltV->getValue();
559 return false; // FIXME: Handle UNDEF elements too!
561 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
564 // Check that they are consequtive.
565 for (unsigned i = 1; i != EltSize; ++i) {
566 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
567 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
571 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
572 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
573 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
574 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
575 "Invalid VECTOR_SHUFFLE mask!");
576 for (unsigned j = 0; j != EltSize; ++j)
577 if (N->getOperand(i+j) != N->getOperand(j))
584 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
586 bool PPC::isAllNegativeZeroVector(SDNode *N) {
587 assert(N->getOpcode() == ISD::BUILD_VECTOR);
588 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
589 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
590 return CFP->getValueAPF().isNegZero();
594 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
595 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
596 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
597 assert(isSplatShuffleMask(N, EltSize));
598 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
601 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
602 /// by using a vspltis[bhw] instruction of the specified element size, return
603 /// the constant being splatted. The ByteSize field indicates the number of
604 /// bytes of each element [124] -> [bhw].
605 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
606 SDOperand OpVal(0, 0);
608 // If ByteSize of the splat is bigger than the element size of the
609 // build_vector, then we have a case where we are checking for a splat where
610 // multiple elements of the buildvector are folded together into a single
611 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
612 unsigned EltSize = 16/N->getNumOperands();
613 if (EltSize < ByteSize) {
614 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
615 SDOperand UniquedVals[4];
616 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
618 // See if all of the elements in the buildvector agree across.
619 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
620 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
621 // If the element isn't a constant, bail fully out.
622 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
625 if (UniquedVals[i&(Multiple-1)].Val == 0)
626 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
627 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
628 return SDOperand(); // no match.
631 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
632 // either constant or undef values that are identical for each chunk. See
633 // if these chunks can form into a larger vspltis*.
635 // Check to see if all of the leading entries are either 0 or -1. If
636 // neither, then this won't fit into the immediate field.
637 bool LeadingZero = true;
638 bool LeadingOnes = true;
639 for (unsigned i = 0; i != Multiple-1; ++i) {
640 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
642 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
643 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
645 // Finally, check the least significant entry.
647 if (UniquedVals[Multiple-1].Val == 0)
648 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
649 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
651 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
654 if (UniquedVals[Multiple-1].Val == 0)
655 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
656 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
657 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
658 return DAG.getTargetConstant(Val, MVT::i32);
664 // Check to see if this buildvec has a single non-undef value in its elements.
665 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
666 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
668 OpVal = N->getOperand(i);
669 else if (OpVal != N->getOperand(i))
673 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
675 unsigned ValSizeInBytes = 0;
677 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
678 Value = CN->getValue();
679 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
680 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
681 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
682 Value = FloatToBits(CN->getValueAPF().convertToFloat());
686 // If the splat value is larger than the element value, then we can never do
687 // this splat. The only case that we could fit the replicated bits into our
688 // immediate field for would be zero, and we prefer to use vxor for it.
689 if (ValSizeInBytes < ByteSize) return SDOperand();
691 // If the element value is larger than the splat value, cut it in half and
692 // check to see if the two halves are equal. Continue doing this until we
693 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
694 while (ValSizeInBytes > ByteSize) {
695 ValSizeInBytes >>= 1;
697 // If the top half equals the bottom half, we're still ok.
698 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
699 (Value & ((1 << (8*ValSizeInBytes))-1)))
703 // Properly sign extend the value.
704 int ShAmt = (4-ByteSize)*8;
705 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
707 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
708 if (MaskVal == 0) return SDOperand();
710 // Finally, if this value fits in a 5 bit sext field, return it
711 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
712 return DAG.getTargetConstant(MaskVal, MVT::i32);
716 //===----------------------------------------------------------------------===//
717 // Addressing Mode Selection
718 //===----------------------------------------------------------------------===//
720 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
721 /// or 64-bit immediate, and if the value can be accurately represented as a
722 /// sign extension from a 16-bit value. If so, this returns true and the
724 static bool isIntS16Immediate(SDNode *N, short &Imm) {
725 if (N->getOpcode() != ISD::Constant)
728 Imm = (short)cast<ConstantSDNode>(N)->getValue();
729 if (N->getValueType(0) == MVT::i32)
730 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
732 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
734 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
735 return isIntS16Immediate(Op.Val, Imm);
739 /// SelectAddressRegReg - Given the specified addressed, check to see if it
740 /// can be represented as an indexed [r+r] operation. Returns false if it
741 /// can be more efficiently represented with [r+imm].
742 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
746 if (N.getOpcode() == ISD::ADD) {
747 if (isIntS16Immediate(N.getOperand(1), imm))
749 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
752 Base = N.getOperand(0);
753 Index = N.getOperand(1);
755 } else if (N.getOpcode() == ISD::OR) {
756 if (isIntS16Immediate(N.getOperand(1), imm))
757 return false; // r+i can fold it if we can.
759 // If this is an or of disjoint bitfields, we can codegen this as an add
760 // (for better address arithmetic) if the LHS and RHS of the OR are provably
762 APInt LHSKnownZero, LHSKnownOne;
763 APInt RHSKnownZero, RHSKnownOne;
764 DAG.ComputeMaskedBits(N.getOperand(0),
765 APInt::getAllOnesValue(N.getOperand(0)
766 .getValueSizeInBits()),
767 LHSKnownZero, LHSKnownOne);
769 if (LHSKnownZero.getBoolValue()) {
770 DAG.ComputeMaskedBits(N.getOperand(1),
771 APInt::getAllOnesValue(N.getOperand(1)
772 .getValueSizeInBits()),
773 RHSKnownZero, RHSKnownOne);
774 // If all of the bits are known zero on the LHS or RHS, the add won't
776 if (~(LHSKnownZero | RHSKnownZero) == 0) {
777 Base = N.getOperand(0);
778 Index = N.getOperand(1);
787 /// Returns true if the address N can be represented by a base register plus
788 /// a signed 16-bit displacement [r+imm], and if it is not better
789 /// represented as reg+reg.
790 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
791 SDOperand &Base, SelectionDAG &DAG){
792 // If this can be more profitably realized as r+r, fail.
793 if (SelectAddressRegReg(N, Disp, Base, DAG))
796 if (N.getOpcode() == ISD::ADD) {
798 if (isIntS16Immediate(N.getOperand(1), imm)) {
799 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
800 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
801 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
803 Base = N.getOperand(0);
805 return true; // [r+i]
806 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
807 // Match LOAD (ADD (X, Lo(G))).
808 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
809 && "Cannot handle constant offsets yet!");
810 Disp = N.getOperand(1).getOperand(0); // The global address.
811 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
812 Disp.getOpcode() == ISD::TargetConstantPool ||
813 Disp.getOpcode() == ISD::TargetJumpTable);
814 Base = N.getOperand(0);
815 return true; // [&g+r]
817 } else if (N.getOpcode() == ISD::OR) {
819 if (isIntS16Immediate(N.getOperand(1), imm)) {
820 // If this is an or of disjoint bitfields, we can codegen this as an add
821 // (for better address arithmetic) if the LHS and RHS of the OR are
822 // provably disjoint.
823 APInt LHSKnownZero, LHSKnownOne;
824 DAG.ComputeMaskedBits(N.getOperand(0),
825 APInt::getAllOnesValue(32),
826 LHSKnownZero, LHSKnownOne);
827 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
828 // If all of the bits are known zero on the LHS or RHS, the add won't
830 Base = N.getOperand(0);
831 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
835 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
836 // Loading from a constant address.
838 // If this address fits entirely in a 16-bit sext immediate field, codegen
841 if (isIntS16Immediate(CN, Imm)) {
842 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
843 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
847 // Handle 32-bit sext immediates with LIS + addr mode.
848 if (CN->getValueType(0) == MVT::i32 ||
849 (int64_t)CN->getValue() == (int)CN->getValue()) {
850 int Addr = (int)CN->getValue();
852 // Otherwise, break this down into an LIS + disp.
853 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
855 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
856 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
857 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
862 Disp = DAG.getTargetConstant(0, getPointerTy());
863 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
864 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
867 return true; // [r+0]
870 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
871 /// represented as an indexed [r+r] operation.
872 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
875 // Check to see if we can easily represent this as an [r+r] address. This
876 // will fail if it thinks that the address is more profitably represented as
877 // reg+imm, e.g. where imm = 0.
878 if (SelectAddressRegReg(N, Base, Index, DAG))
881 // If the operand is an addition, always emit this as [r+r], since this is
882 // better (for code size, and execution, as the memop does the add for free)
883 // than emitting an explicit add.
884 if (N.getOpcode() == ISD::ADD) {
885 Base = N.getOperand(0);
886 Index = N.getOperand(1);
890 // Otherwise, do it the hard way, using R0 as the base register.
891 Base = DAG.getRegister(PPC::R0, N.getValueType());
896 /// SelectAddressRegImmShift - Returns true if the address N can be
897 /// represented by a base register plus a signed 14-bit displacement
898 /// [r+imm*4]. Suitable for use by STD and friends.
899 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
902 // If this can be more profitably realized as r+r, fail.
903 if (SelectAddressRegReg(N, Disp, Base, DAG))
906 if (N.getOpcode() == ISD::ADD) {
908 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
909 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
910 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
911 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
913 Base = N.getOperand(0);
915 return true; // [r+i]
916 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
917 // Match LOAD (ADD (X, Lo(G))).
918 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
919 && "Cannot handle constant offsets yet!");
920 Disp = N.getOperand(1).getOperand(0); // The global address.
921 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
922 Disp.getOpcode() == ISD::TargetConstantPool ||
923 Disp.getOpcode() == ISD::TargetJumpTable);
924 Base = N.getOperand(0);
925 return true; // [&g+r]
927 } else if (N.getOpcode() == ISD::OR) {
929 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
930 // If this is an or of disjoint bitfields, we can codegen this as an add
931 // (for better address arithmetic) if the LHS and RHS of the OR are
932 // provably disjoint.
933 APInt LHSKnownZero, LHSKnownOne;
934 DAG.ComputeMaskedBits(N.getOperand(0),
935 APInt::getAllOnesValue(32),
936 LHSKnownZero, LHSKnownOne);
937 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
938 // If all of the bits are known zero on the LHS or RHS, the add won't
940 Base = N.getOperand(0);
941 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
945 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
946 // Loading from a constant address. Verify low two bits are clear.
947 if ((CN->getValue() & 3) == 0) {
948 // If this address fits entirely in a 14-bit sext immediate field, codegen
951 if (isIntS16Immediate(CN, Imm)) {
952 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
953 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
957 // Fold the low-part of 32-bit absolute addresses into addr mode.
958 if (CN->getValueType(0) == MVT::i32 ||
959 (int64_t)CN->getValue() == (int)CN->getValue()) {
960 int Addr = (int)CN->getValue();
962 // Otherwise, break this down into an LIS + disp.
963 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
965 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
966 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
967 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
973 Disp = DAG.getTargetConstant(0, getPointerTy());
974 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
975 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
978 return true; // [r+0]
982 /// getPreIndexedAddressParts - returns true by value, base pointer and
983 /// offset pointer and addressing mode by reference if the node's address
984 /// can be legally represented as pre-indexed load / store address.
985 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
987 ISD::MemIndexedMode &AM,
989 // Disabled by default for now.
990 if (!EnablePPCPreinc) return false;
994 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
995 Ptr = LD->getBasePtr();
996 VT = LD->getMemoryVT();
998 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1000 Ptr = ST->getBasePtr();
1001 VT = ST->getMemoryVT();
1005 // PowerPC doesn't have preinc load/store instructions for vectors.
1006 if (MVT::isVector(VT))
1009 // TODO: Check reg+reg first.
1011 // LDU/STU use reg+imm*4, others use reg+imm.
1012 if (VT != MVT::i64) {
1014 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1018 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1022 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1023 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1024 // sext i32 to i64 when addr mode is r+i.
1025 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1026 LD->getExtensionType() == ISD::SEXTLOAD &&
1027 isa<ConstantSDNode>(Offset))
1035 //===----------------------------------------------------------------------===//
1036 // LowerOperation implementation
1037 //===----------------------------------------------------------------------===//
1039 SDOperand PPCTargetLowering::LowerConstantPool(SDOperand Op,
1040 SelectionDAG &DAG) {
1041 MVT::ValueType PtrVT = Op.getValueType();
1042 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1043 Constant *C = CP->getConstVal();
1044 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1045 SDOperand Zero = DAG.getConstant(0, PtrVT);
1047 const TargetMachine &TM = DAG.getTarget();
1049 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1050 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1052 // If this is a non-darwin platform, we don't support non-static relo models
1054 if (TM.getRelocationModel() == Reloc::Static ||
1055 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1056 // Generate non-pic code that has direct accesses to the constant pool.
1057 // The address of the global is just (hi(&g)+lo(&g)).
1058 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1061 if (TM.getRelocationModel() == Reloc::PIC_) {
1062 // With PIC, the first instruction is actually "GR+hi(&G)".
1063 Hi = DAG.getNode(ISD::ADD, PtrVT,
1064 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1067 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1071 SDOperand PPCTargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1072 MVT::ValueType PtrVT = Op.getValueType();
1073 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1074 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1075 SDOperand Zero = DAG.getConstant(0, PtrVT);
1077 const TargetMachine &TM = DAG.getTarget();
1079 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1080 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1082 // If this is a non-darwin platform, we don't support non-static relo models
1084 if (TM.getRelocationModel() == Reloc::Static ||
1085 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1086 // Generate non-pic code that has direct accesses to the constant pool.
1087 // The address of the global is just (hi(&g)+lo(&g)).
1088 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1091 if (TM.getRelocationModel() == Reloc::PIC_) {
1092 // With PIC, the first instruction is actually "GR+hi(&G)".
1093 Hi = DAG.getNode(ISD::ADD, PtrVT,
1094 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1097 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1101 SDOperand PPCTargetLowering::LowerGlobalTLSAddress(SDOperand Op,
1102 SelectionDAG &DAG) {
1103 assert(0 && "TLS not implemented for PPC.");
1106 SDOperand PPCTargetLowering::LowerGlobalAddress(SDOperand Op,
1107 SelectionDAG &DAG) {
1108 MVT::ValueType PtrVT = Op.getValueType();
1109 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1110 GlobalValue *GV = GSDN->getGlobal();
1111 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1112 // If it's a debug information descriptor, don't mess with it.
1113 if (DAG.isVerifiedDebugInfoDesc(Op))
1115 SDOperand Zero = DAG.getConstant(0, PtrVT);
1117 const TargetMachine &TM = DAG.getTarget();
1119 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1120 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1122 // If this is a non-darwin platform, we don't support non-static relo models
1124 if (TM.getRelocationModel() == Reloc::Static ||
1125 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1126 // Generate non-pic code that has direct accesses to globals.
1127 // The address of the global is just (hi(&g)+lo(&g)).
1128 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1131 if (TM.getRelocationModel() == Reloc::PIC_) {
1132 // With PIC, the first instruction is actually "GR+hi(&G)".
1133 Hi = DAG.getNode(ISD::ADD, PtrVT,
1134 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1137 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1139 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1142 // If the global is weak or external, we have to go through the lazy
1144 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1147 SDOperand PPCTargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1148 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1150 // If we're comparing for equality to zero, expose the fact that this is
1151 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1152 // fold the new nodes.
1153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1154 if (C->isNullValue() && CC == ISD::SETEQ) {
1155 MVT::ValueType VT = Op.getOperand(0).getValueType();
1156 SDOperand Zext = Op.getOperand(0);
1157 if (VT < MVT::i32) {
1159 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1161 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1162 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1163 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1164 DAG.getConstant(Log2b, MVT::i32));
1165 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1167 // Leave comparisons against 0 and -1 alone for now, since they're usually
1168 // optimized. FIXME: revisit this when we can custom lower all setcc
1170 if (C->isAllOnesValue() || C->isNullValue())
1174 // If we have an integer seteq/setne, turn it into a compare against zero
1175 // by xor'ing the rhs with the lhs, which is faster than setting a
1176 // condition register, reading it back out, and masking the correct bit. The
1177 // normal approach here uses sub to do this instead of xor. Using xor exposes
1178 // the result to other bit-twiddling opportunities.
1179 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1180 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1181 MVT::ValueType VT = Op.getValueType();
1182 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1184 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1189 SDOperand PPCTargetLowering::LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1190 int VarArgsFrameIndex,
1191 int VarArgsStackOffset,
1192 unsigned VarArgsNumGPR,
1193 unsigned VarArgsNumFPR,
1194 const PPCSubtarget &Subtarget) {
1196 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1199 SDOperand PPCTargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1200 int VarArgsFrameIndex,
1201 int VarArgsStackOffset,
1202 unsigned VarArgsNumGPR,
1203 unsigned VarArgsNumFPR,
1204 const PPCSubtarget &Subtarget) {
1206 if (Subtarget.isMachoABI()) {
1207 // vastart just stores the address of the VarArgsFrameIndex slot into the
1208 // memory location argument.
1209 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1210 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1211 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1212 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1215 // For ELF 32 ABI we follow the layout of the va_list struct.
1216 // We suppose the given va_list is already allocated.
1219 // char gpr; /* index into the array of 8 GPRs
1220 // * stored in the register save area
1221 // * gpr=0 corresponds to r3,
1222 // * gpr=1 to r4, etc.
1224 // char fpr; /* index into the array of 8 FPRs
1225 // * stored in the register save area
1226 // * fpr=0 corresponds to f1,
1227 // * fpr=1 to f2, etc.
1229 // char *overflow_arg_area;
1230 // /* location on stack that holds
1231 // * the next overflow argument
1233 // char *reg_save_area;
1234 // /* where r3:r10 and f1:f8 (if saved)
1240 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1241 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1244 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1246 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1247 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1249 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1250 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1252 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1253 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1255 uint64_t FPROffset = 1;
1256 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1258 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1260 // Store first byte : number of int regs
1261 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1262 Op.getOperand(1), SV, 0);
1263 uint64_t nextOffset = FPROffset;
1264 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1267 // Store second byte : number of float regs
1268 SDOperand secondStore =
1269 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1270 nextOffset += StackOffset;
1271 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1273 // Store second word : arguments given on stack
1274 SDOperand thirdStore =
1275 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1276 nextOffset += FrameOffset;
1277 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1279 // Store third word : arguments given in registers
1280 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1284 #include "PPCGenCallingConv.inc"
1286 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1287 /// depending on which subtarget is selected.
1288 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1289 if (Subtarget.isMachoABI()) {
1290 static const unsigned FPR[] = {
1291 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1292 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1298 static const unsigned FPR[] = {
1299 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1306 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op,
1308 int &VarArgsFrameIndex,
1309 int &VarArgsStackOffset,
1310 unsigned &VarArgsNumGPR,
1311 unsigned &VarArgsNumFPR,
1312 const PPCSubtarget &Subtarget) {
1313 // TODO: add description of PPC stack frame format, or at least some docs.
1315 MachineFunction &MF = DAG.getMachineFunction();
1316 MachineFrameInfo *MFI = MF.getFrameInfo();
1317 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1318 SmallVector<SDOperand, 8> ArgValues;
1319 SDOperand Root = Op.getOperand(0);
1320 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1322 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1323 bool isPPC64 = PtrVT == MVT::i64;
1324 bool isMachoABI = Subtarget.isMachoABI();
1325 bool isELF32_ABI = Subtarget.isELF32_ABI();
1326 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1328 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1330 static const unsigned GPR_32[] = { // 32-bit registers.
1331 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1332 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1334 static const unsigned GPR_64[] = { // 64-bit registers.
1335 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1336 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1339 static const unsigned *FPR = GetFPR(Subtarget);
1341 static const unsigned VR[] = {
1342 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1343 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1346 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1347 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1348 const unsigned Num_VR_Regs = array_lengthof( VR);
1350 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1352 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1354 // In 32-bit non-varargs functions, the stack space for vectors is after the
1355 // stack space for non-vectors. We do not use this space unless we have
1356 // too many vectors to fit in registers, something that only occurs in
1357 // constructed examples:), but we have to walk the arglist to figure
1358 // that out...for the pathological case, compute VecArgOffset as the
1359 // start of the vector parameter area. Computing VecArgOffset is the
1360 // entire point of the following loop.
1361 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1362 // to handle Elf here.
1363 unsigned VecArgOffset = ArgOffset;
1364 if (!isVarArg && !isPPC64) {
1365 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e;
1367 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1368 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1369 ISD::ArgFlagsTy Flags =
1370 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1372 if (Flags.isByVal()) {
1373 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1374 ObjSize = Flags.getByValSize();
1376 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1377 VecArgOffset += ArgSize;
1382 default: assert(0 && "Unhandled argument type!");
1385 VecArgOffset += isPPC64 ? 8 : 4;
1387 case MVT::i64: // PPC64
1395 // Nothing to do, we're only looking at Nonvector args here.
1400 // We've found where the vector parameter area in memory is. Skip the
1401 // first 12 parameters; these don't use that memory.
1402 VecArgOffset = ((VecArgOffset+15)/16)*16;
1403 VecArgOffset += 12*16;
1405 // Add DAG nodes to load the arguments or copy them out of registers. On
1406 // entry to a function on PPC, the arguments start after the linkage area,
1407 // although the first ones are often in registers.
1409 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1410 // represented with two words (long long or double) must be copied to an
1411 // even GPR_idx value or to an even ArgOffset value. TODO: implement this.
1413 SmallVector<SDOperand, 8> MemOps;
1415 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1417 bool needsLoad = false;
1418 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1419 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1420 unsigned ArgSize = ObjSize;
1421 ISD::ArgFlagsTy Flags =
1422 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1423 // See if next argument requires stack alignment in ELF
1424 bool Expand = false; // TODO: implement this.
1426 unsigned CurArgOffset = ArgOffset;
1428 // FIXME alignment for ELF may not be right
1429 // FIXME the codegen can be much improved in some cases.
1430 // We do not have to keep everything in memory.
1431 if (Flags.isByVal()) {
1432 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1433 ObjSize = Flags.getByValSize();
1434 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1435 // Double word align in ELF
1436 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1437 // Objects of size 1 and 2 are right justified, everything else is
1438 // left justified. This means the memory address is adjusted forwards.
1439 if (ObjSize==1 || ObjSize==2) {
1440 CurArgOffset = CurArgOffset + (4 - ObjSize);
1442 // The value of the object is its address.
1443 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1444 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1445 ArgValues.push_back(FIN);
1446 if (ObjSize==1 || ObjSize==2) {
1447 if (GPR_idx != Num_GPR_Regs) {
1448 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1449 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1450 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1451 SDOperand Store = DAG.getTruncStore(Val.getValue(1), Val, FIN,
1452 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1453 MemOps.push_back(Store);
1455 if (isMachoABI) ArgOffset += PtrByteSize;
1457 ArgOffset += PtrByteSize;
1461 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1462 // Store whatever pieces of the object are in registers
1463 // to memory. ArgVal will be address of the beginning of
1465 if (GPR_idx != Num_GPR_Regs) {
1466 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1467 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1468 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1469 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1470 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1471 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1472 MemOps.push_back(Store);
1474 if (isMachoABI) ArgOffset += PtrByteSize;
1476 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1484 default: assert(0 && "Unhandled argument type!");
1487 // Double word align in ELF
1488 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1490 if (GPR_idx != Num_GPR_Regs) {
1491 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1492 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1493 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1497 ArgSize = PtrByteSize;
1499 // Stack align in ELF
1500 if (needsLoad && Expand && isELF32_ABI)
1501 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1502 // All int arguments reserve stack space in Macho ABI.
1503 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1507 case MVT::i64: // PPC64
1508 if (GPR_idx != Num_GPR_Regs) {
1509 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1510 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1511 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1513 if (ObjectVT == MVT::i32) {
1514 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1515 // value to MVT::i64 and then truncate to the correct register size.
1517 ArgVal = DAG.getNode(ISD::AssertSext, MVT::i64, ArgVal,
1518 DAG.getValueType(ObjectVT));
1519 else if (Flags.isZExt())
1520 ArgVal = DAG.getNode(ISD::AssertZext, MVT::i64, ArgVal,
1521 DAG.getValueType(ObjectVT));
1523 ArgVal = DAG.getNode(ISD::TRUNCATE, MVT::i32, ArgVal);
1530 // All int arguments reserve stack space in Macho ABI.
1531 if (isMachoABI || needsLoad) ArgOffset += 8;
1536 // Every 4 bytes of argument space consumes one of the GPRs available for
1537 // argument passing.
1538 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1540 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1543 if (FPR_idx != Num_FPR_Regs) {
1545 if (ObjectVT == MVT::f32)
1546 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1548 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1549 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1550 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1556 // Stack align in ELF
1557 if (needsLoad && Expand && isELF32_ABI)
1558 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1559 // All FP arguments reserve stack space in Macho ABI.
1560 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1566 // Note that vector arguments in registers don't reserve stack space,
1567 // except in varargs functions.
1568 if (VR_idx != Num_VR_Regs) {
1569 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1570 RegInfo.addLiveIn(VR[VR_idx], VReg);
1571 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1573 while ((ArgOffset % 16) != 0) {
1574 ArgOffset += PtrByteSize;
1575 if (GPR_idx != Num_GPR_Regs)
1579 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1583 if (!isVarArg && !isPPC64) {
1584 // Vectors go after all the nonvectors.
1585 CurArgOffset = VecArgOffset;
1588 // Vectors are aligned.
1589 ArgOffset = ((ArgOffset+15)/16)*16;
1590 CurArgOffset = ArgOffset;
1598 // We need to load the argument to a virtual register if we determined above
1599 // that we ran out of physical registers of the appropriate type.
1601 int FI = MFI->CreateFixedObject(ObjSize,
1602 CurArgOffset + (ArgSize - ObjSize));
1603 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1604 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1607 ArgValues.push_back(ArgVal);
1610 // If the function takes variable number of arguments, make a frame index for
1611 // the start of the first vararg value... for expansion of llvm.va_start.
1616 VarArgsNumGPR = GPR_idx;
1617 VarArgsNumFPR = FPR_idx;
1619 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1621 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1622 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1623 MVT::getSizeInBits(PtrVT)/8);
1625 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1632 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1634 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1636 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1637 // stored to the VarArgsFrameIndex on the stack.
1639 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1640 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1641 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1642 MemOps.push_back(Store);
1643 // Increment the address by four for the next argument to store
1644 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1645 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1649 // If this function is vararg, store any remaining integer argument regs
1650 // to their spots on the stack so that they may be loaded by deferencing the
1651 // result of va_next.
1652 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1655 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1657 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1659 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1660 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1661 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1662 MemOps.push_back(Store);
1663 // Increment the address by four for the next argument to store
1664 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1665 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1668 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1671 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1672 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1673 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1674 MemOps.push_back(Store);
1675 // Increment the address by eight for the next argument to store
1676 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1678 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1681 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1683 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1685 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1686 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1687 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1688 MemOps.push_back(Store);
1689 // Increment the address by eight for the next argument to store
1690 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1692 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1697 if (!MemOps.empty())
1698 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1700 ArgValues.push_back(Root);
1702 // Return the new list of results.
1703 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1704 Op.Val->value_end());
1705 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1708 /// isCallCompatibleAddress - Return the immediate to use if the specified
1709 /// 32-bit value is representable in the immediate field of a BxA instruction.
1710 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1711 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1714 int Addr = C->getValue();
1715 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1716 (Addr << 6 >> 6) != Addr)
1717 return 0; // Top 6 bits have to be sext of immediate.
1719 return DAG.getConstant((int)C->getValue() >> 2,
1720 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1723 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1724 /// by "Src" to address "Dst" of size "Size". Alignment information is
1725 /// specified by the specific parameter attribute. The copy will be passed as
1726 /// a byval function parameter.
1727 /// Sometimes what we are copying is the end of a larger object, the part that
1728 /// does not fit in registers.
1730 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1731 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1733 SDOperand AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1734 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1735 SDOperand AlwaysInline = DAG.getConstant(0, MVT::i32);
1736 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1739 SDOperand PPCTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG,
1740 const PPCSubtarget &Subtarget,
1741 TargetMachine &TM) {
1742 SDOperand Chain = Op.getOperand(0);
1743 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1744 SDOperand Callee = Op.getOperand(4);
1745 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1747 bool isMachoABI = Subtarget.isMachoABI();
1748 bool isELF32_ABI = Subtarget.isELF32_ABI();
1750 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1751 bool isPPC64 = PtrVT == MVT::i64;
1752 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1754 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1755 // SelectExpr to use to put the arguments in the appropriate registers.
1756 std::vector<SDOperand> args_to_use;
1758 // Count how many bytes are to be pushed on the stack, including the linkage
1759 // area, and parameter passing area. We start with 24/48 bytes, which is
1760 // prereserved space for [SP][CR][LR][3 x unused].
1761 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1763 // Add up all the space actually used.
1764 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1765 // they all go in registers, but we must reserve stack space for them for
1766 // possible use by the caller. In varargs or 64-bit calls, parameters are
1767 // assigned stack space in order, with padding so Altivec parameters are
1769 unsigned nAltivecParamsAtEnd = 0;
1770 for (unsigned i = 0; i != NumOps; ++i) {
1771 SDOperand Arg = Op.getOperand(5+2*i);
1772 MVT::ValueType ArgVT = Arg.getValueType();
1773 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1774 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1775 if (!isVarArg && !isPPC64) {
1776 // Non-varargs Altivec parameters go after all the non-Altivec parameters;
1777 // do those last so we know how much padding we need.
1778 nAltivecParamsAtEnd++;
1781 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1782 NumBytes = ((NumBytes+15)/16)*16;
1785 ISD::ArgFlagsTy Flags =
1786 cast<ARG_FLAGSSDNode>(Op.getOperand(5+2*i+1))->getArgFlags();
1787 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1788 if (Flags.isByVal())
1789 ArgSize = Flags.getByValSize();
1790 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1791 NumBytes += ArgSize;
1793 // Allow for Altivec parameters at the end, if needed.
1794 if (nAltivecParamsAtEnd) {
1795 NumBytes = ((NumBytes+15)/16)*16;
1796 NumBytes += 16*nAltivecParamsAtEnd;
1799 // The prolog code of the callee may store up to 8 GPR argument registers to
1800 // the stack, allowing va_start to index over them in memory if its varargs.
1801 // Because we cannot tell if this is needed on the caller side, we have to
1802 // conservatively assume that it is needed. As such, make sure we have at
1803 // least enough stack space for the caller to store the 8 GPRs.
1804 NumBytes = std::max(NumBytes,
1805 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1807 // Adjust the stack pointer for the new arguments...
1808 // These operations are automatically eliminated by the prolog/epilog pass
1809 Chain = DAG.getCALLSEQ_START(Chain,
1810 DAG.getConstant(NumBytes, PtrVT));
1811 SDOperand CallSeqStart = Chain;
1813 // Set up a copy of the stack pointer for use loading and storing any
1814 // arguments that may not fit in the registers available for argument
1818 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1820 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1822 // Figure out which arguments are going to go in registers, and which in
1823 // memory. Also, if this is a vararg function, floating point operations
1824 // must be stored to our stack, and loaded into integer regs as well, if
1825 // any integer regs are available for argument passing.
1826 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1827 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1829 static const unsigned GPR_32[] = { // 32-bit registers.
1830 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1831 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1833 static const unsigned GPR_64[] = { // 64-bit registers.
1834 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1835 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1837 static const unsigned *FPR = GetFPR(Subtarget);
1839 static const unsigned VR[] = {
1840 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1841 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1843 const unsigned NumGPRs = array_lengthof(GPR_32);
1844 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1845 const unsigned NumVRs = array_lengthof( VR);
1847 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1849 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1850 SmallVector<SDOperand, 8> MemOpChains;
1851 for (unsigned i = 0; i != NumOps; ++i) {
1853 SDOperand Arg = Op.getOperand(5+2*i);
1854 ISD::ArgFlagsTy Flags =
1855 cast<ARG_FLAGSSDNode>(Op.getOperand(5+2*i+1))->getArgFlags();
1856 // See if next argument requires stack alignment in ELF
1857 bool Expand = false; // TODO: implement this.
1859 // PtrOff will be used to store the current argument to the stack if a
1860 // register cannot be found for it.
1863 // Stack align in ELF 32
1864 if (isELF32_ABI && Expand)
1865 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1866 StackPtr.getValueType());
1868 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1870 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1872 // On PPC64, promote integers to 64-bit values.
1873 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1874 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
1875 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1876 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1879 // FIXME Elf untested, what are alignment rules?
1880 // FIXME memcpy is used way more than necessary. Correctness first.
1881 if (Flags.isByVal()) {
1882 unsigned Size = Flags.getByValSize();
1883 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1884 if (Size==1 || Size==2) {
1885 // Very small objects are passed right-justified.
1886 // Everything else is passed left-justified.
1887 MVT::ValueType VT = (Size==1) ? MVT::i8 : MVT::i16;
1888 if (GPR_idx != NumGPRs) {
1889 SDOperand Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
1891 MemOpChains.push_back(Load.getValue(1));
1892 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1894 ArgOffset += PtrByteSize;
1896 SDOperand Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
1897 SDOperand AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
1898 SDOperand MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
1899 CallSeqStart.Val->getOperand(0),
1901 // This must go outside the CALLSEQ_START..END.
1902 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1903 CallSeqStart.Val->getOperand(1));
1904 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1905 Chain = CallSeqStart = NewCallSeqStart;
1906 ArgOffset += PtrByteSize;
1910 // Copy entire object into memory. There are cases where gcc-generated
1911 // code assumes it is there, even if it could be put entirely into
1912 // registers. (This is not what the doc says.)
1913 SDOperand MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
1914 CallSeqStart.Val->getOperand(0),
1916 // This must go outside the CALLSEQ_START..END.
1917 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1918 CallSeqStart.Val->getOperand(1));
1919 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1920 Chain = CallSeqStart = NewCallSeqStart;
1921 // And copy the pieces of it that fit into registers.
1922 for (unsigned j=0; j<Size; j+=PtrByteSize) {
1923 SDOperand Const = DAG.getConstant(j, PtrOff.getValueType());
1924 SDOperand AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
1925 if (GPR_idx != NumGPRs) {
1926 SDOperand Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
1927 MemOpChains.push_back(Load.getValue(1));
1928 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1930 ArgOffset += PtrByteSize;
1932 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
1939 switch (Arg.getValueType()) {
1940 default: assert(0 && "Unexpected ValueType for argument!");
1943 // Double word align in ELF
1944 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1945 if (GPR_idx != NumGPRs) {
1946 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1948 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1951 if (inMem || isMachoABI) {
1952 // Stack align in ELF
1953 if (isELF32_ABI && Expand)
1954 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1956 ArgOffset += PtrByteSize;
1961 if (FPR_idx != NumFPRs) {
1962 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1965 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1966 MemOpChains.push_back(Store);
1968 // Float varargs are always shadowed in available integer registers
1969 if (GPR_idx != NumGPRs) {
1970 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1971 MemOpChains.push_back(Load.getValue(1));
1972 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1975 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1976 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1977 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1978 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1979 MemOpChains.push_back(Load.getValue(1));
1980 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1984 // If we have any FPRs remaining, we may also have GPRs remaining.
1985 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1988 if (GPR_idx != NumGPRs)
1990 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1991 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1996 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1999 if (inMem || isMachoABI) {
2000 // Stack align in ELF
2001 if (isELF32_ABI && Expand)
2002 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2006 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2014 // These go aligned on the stack, or in the corresponding R registers
2015 // when within range. The Darwin PPC ABI doc claims they also go in
2016 // V registers; in fact gcc does this only for arguments that are
2017 // prototyped, not for those that match the ... We do it for all
2018 // arguments, seems to work.
2019 while (ArgOffset % 16 !=0) {
2020 ArgOffset += PtrByteSize;
2021 if (GPR_idx != NumGPRs)
2024 // We could elide this store in the case where the object fits
2025 // entirely in R registers. Maybe later.
2026 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2027 DAG.getConstant(ArgOffset, PtrVT));
2028 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2029 MemOpChains.push_back(Store);
2030 if (VR_idx != NumVRs) {
2031 SDOperand Load = DAG.getLoad(MVT::v4f32, Store, PtrOff, NULL, 0);
2032 MemOpChains.push_back(Load.getValue(1));
2033 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2036 for (unsigned i=0; i<16; i+=PtrByteSize) {
2037 if (GPR_idx == NumGPRs)
2039 SDOperand Ix = DAG.getNode(ISD::ADD, PtrVT, PtrOff,
2040 DAG.getConstant(i, PtrVT));
2041 SDOperand Load = DAG.getLoad(PtrVT, Store, Ix, NULL, 0);
2042 MemOpChains.push_back(Load.getValue(1));
2043 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2047 // Non-varargs Altivec params generally go in registers, but have
2048 // stack space allocated at the end.
2049 if (VR_idx != NumVRs) {
2050 // Doesn't have GPR space allocated.
2051 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2052 } else if (nAltivecParamsAtEnd==0) {
2053 // We are emitting Altivec params in order.
2054 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2055 DAG.getConstant(ArgOffset, PtrVT));
2056 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2057 MemOpChains.push_back(Store);
2063 // If all Altivec parameters fit in registers, as they usually do,
2064 // they get stack space following the non-Altivec parameters. We
2065 // don't track this here because nobody below needs it.
2066 // If there are more Altivec parameters than fit in registers emit
2068 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2070 // Offset is aligned; skip 1st 12 params which go in V registers.
2071 ArgOffset = ((ArgOffset+15)/16)*16;
2073 for (unsigned i = 0; i != NumOps; ++i) {
2074 SDOperand Arg = Op.getOperand(5+2*i);
2075 MVT::ValueType ArgType = Arg.getValueType();
2076 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2077 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2079 SDOperand PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2080 DAG.getConstant(ArgOffset, PtrVT));
2081 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2082 MemOpChains.push_back(Store);
2089 if (!MemOpChains.empty())
2090 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2091 &MemOpChains[0], MemOpChains.size());
2093 // Build a sequence of copy-to-reg nodes chained together with token chain
2094 // and flag operands which copy the outgoing args into the appropriate regs.
2096 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2097 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
2099 InFlag = Chain.getValue(1);
2102 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2103 if (isVarArg && isELF32_ABI) {
2104 SDOperand SetCR(DAG.getTargetNode(PPC::CRSET, MVT::i32), 0);
2105 Chain = DAG.getCopyToReg(Chain, PPC::CR1EQ, SetCR, InFlag);
2106 InFlag = Chain.getValue(1);
2109 std::vector<MVT::ValueType> NodeTys;
2110 NodeTys.push_back(MVT::Other); // Returns a chain
2111 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2113 SmallVector<SDOperand, 8> Ops;
2114 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2116 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2117 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2118 // node so that legalize doesn't hack it.
2119 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2120 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2121 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2122 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2123 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2124 // If this is an absolute destination address, use the munged value.
2125 Callee = SDOperand(Dest, 0);
2127 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2128 // to do the call, we can't use PPCISD::CALL.
2129 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
2130 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
2131 InFlag = Chain.getValue(1);
2133 // Copy the callee address into R12/X12 on darwin.
2135 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2136 Chain = DAG.getCopyToReg(Chain, Reg, Callee, InFlag);
2137 InFlag = Chain.getValue(1);
2141 NodeTys.push_back(MVT::Other);
2142 NodeTys.push_back(MVT::Flag);
2143 Ops.push_back(Chain);
2144 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2148 // If this is a direct call, pass the chain and the callee.
2150 Ops.push_back(Chain);
2151 Ops.push_back(Callee);
2154 // Add argument registers to the end of the list so that they are known live
2156 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2157 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2158 RegsToPass[i].second.getValueType()));
2161 Ops.push_back(InFlag);
2162 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
2163 InFlag = Chain.getValue(1);
2165 Chain = DAG.getCALLSEQ_END(Chain,
2166 DAG.getConstant(NumBytes, PtrVT),
2167 DAG.getConstant(0, PtrVT),
2169 if (Op.Val->getValueType(0) != MVT::Other)
2170 InFlag = Chain.getValue(1);
2172 SmallVector<SDOperand, 16> ResultVals;
2173 SmallVector<CCValAssign, 16> RVLocs;
2174 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2175 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2176 CCInfo.AnalyzeCallResult(Op.Val, RetCC_PPC);
2178 // Copy all of the result registers out of their specified physreg.
2179 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2180 CCValAssign &VA = RVLocs[i];
2181 MVT::ValueType VT = VA.getValVT();
2182 assert(VA.isRegLoc() && "Can only return in registers!");
2183 Chain = DAG.getCopyFromReg(Chain, VA.getLocReg(), VT, InFlag).getValue(1);
2184 ResultVals.push_back(Chain.getValue(0));
2185 InFlag = Chain.getValue(2);
2188 // If the function returns void, just return the chain.
2192 // Otherwise, merge everything together with a MERGE_VALUES node.
2193 ResultVals.push_back(Chain);
2194 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
2195 &ResultVals[0], ResultVals.size());
2196 return Res.getValue(Op.ResNo);
2199 SDOperand PPCTargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG,
2200 TargetMachine &TM) {
2201 SmallVector<CCValAssign, 16> RVLocs;
2202 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2203 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2204 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2205 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
2207 // If this is the first return lowered for this function, add the regs to the
2208 // liveout set for the function.
2209 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2210 for (unsigned i = 0; i != RVLocs.size(); ++i)
2211 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2214 SDOperand Chain = Op.getOperand(0);
2217 // Copy the result values into the output registers.
2218 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2219 CCValAssign &VA = RVLocs[i];
2220 assert(VA.isRegLoc() && "Can only return in registers!");
2221 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2222 Flag = Chain.getValue(1);
2226 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2228 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2231 SDOperand PPCTargetLowering::LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
2232 const PPCSubtarget &Subtarget) {
2233 // When we pop the dynamic allocation we need to restore the SP link.
2235 // Get the corect type for pointers.
2236 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2238 // Construct the stack pointer operand.
2239 bool IsPPC64 = Subtarget.isPPC64();
2240 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2241 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
2243 // Get the operands for the STACKRESTORE.
2244 SDOperand Chain = Op.getOperand(0);
2245 SDOperand SaveSP = Op.getOperand(1);
2247 // Load the old link SP.
2248 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2250 // Restore the stack pointer.
2251 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2253 // Store the old link SP.
2254 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2257 SDOperand PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
2259 const PPCSubtarget &Subtarget) {
2260 MachineFunction &MF = DAG.getMachineFunction();
2261 bool IsPPC64 = Subtarget.isPPC64();
2262 bool isMachoABI = Subtarget.isMachoABI();
2264 // Get current frame pointer save index. The users of this index will be
2265 // primarily DYNALLOC instructions.
2266 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2267 int FPSI = FI->getFramePointerSaveIndex();
2269 // If the frame pointer save index hasn't been defined yet.
2271 // Find out what the fix offset of the frame pointer save area.
2272 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2274 // Allocate the frame index for frame pointer save area.
2275 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2277 FI->setFramePointerSaveIndex(FPSI);
2281 SDOperand Chain = Op.getOperand(0);
2282 SDOperand Size = Op.getOperand(1);
2284 // Get the corect type for pointers.
2285 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2287 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
2288 DAG.getConstant(0, PtrVT), Size);
2289 // Construct a node for the frame pointer save index.
2290 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
2291 // Build a DYNALLOC node.
2292 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
2293 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2294 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2298 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2300 SDOperand PPCTargetLowering::LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2301 // Not FP? Not a fsel.
2302 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2303 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2306 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2308 // Cannot handle SETEQ/SETNE.
2309 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2311 MVT::ValueType ResVT = Op.getValueType();
2312 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2313 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2314 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2316 // If the RHS of the comparison is a 0.0, we don't need to do the
2317 // subtraction at all.
2318 if (isFloatingPointZero(RHS))
2320 default: break; // SETUO etc aren't handled by fsel.
2324 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2328 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2329 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2330 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2334 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2338 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2339 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2340 return DAG.getNode(PPCISD::FSEL, ResVT,
2341 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2346 default: break; // SETUO etc aren't handled by fsel.
2350 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2351 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2352 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2353 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2357 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2358 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2359 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2360 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2364 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2365 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2366 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2367 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2371 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2372 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2373 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2374 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2379 // FIXME: Split this code up when LegalizeDAGTypes lands.
2380 SDOperand PPCTargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2381 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2382 SDOperand Src = Op.getOperand(0);
2383 if (Src.getValueType() == MVT::f32)
2384 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2387 switch (Op.getValueType()) {
2388 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2390 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2393 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2397 // Convert the FP value to an int value through memory.
2398 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2400 // Emit a store to the stack slot.
2401 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2403 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2405 if (Op.getValueType() == MVT::i32)
2406 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2407 DAG.getConstant(4, FIPtr.getValueType()));
2408 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2411 SDOperand PPCTargetLowering::LowerFP_ROUND_INREG(SDOperand Op,
2412 SelectionDAG &DAG) {
2413 assert(Op.getValueType() == MVT::ppcf128);
2414 SDNode *Node = Op.Val;
2415 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2416 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2417 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2418 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2420 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2421 // of the long double, and puts FPSCR back the way it was. We do not
2422 // actually model FPSCR.
2423 std::vector<MVT::ValueType> NodeTys;
2424 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2426 NodeTys.push_back(MVT::f64); // Return register
2427 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2428 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2429 MFFSreg = Result.getValue(0);
2430 InFlag = Result.getValue(1);
2433 NodeTys.push_back(MVT::Flag); // Returns a flag
2434 Ops[0] = DAG.getConstant(31, MVT::i32);
2436 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2437 InFlag = Result.getValue(0);
2440 NodeTys.push_back(MVT::Flag); // Returns a flag
2441 Ops[0] = DAG.getConstant(30, MVT::i32);
2443 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2444 InFlag = Result.getValue(0);
2447 NodeTys.push_back(MVT::f64); // result of add
2448 NodeTys.push_back(MVT::Flag); // Returns a flag
2452 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2453 FPreg = Result.getValue(0);
2454 InFlag = Result.getValue(1);
2457 NodeTys.push_back(MVT::f64);
2458 Ops[0] = DAG.getConstant(1, MVT::i32);
2462 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2463 FPreg = Result.getValue(0);
2465 // We know the low half is about to be thrown away, so just use something
2467 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2470 SDOperand PPCTargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2471 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2472 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2475 if (Op.getOperand(0).getValueType() == MVT::i64) {
2476 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2477 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2478 if (Op.getValueType() == MVT::f32)
2479 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2483 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2484 "Unhandled SINT_TO_FP type in custom expander!");
2485 // Since we only generate this in 64-bit mode, we can take advantage of
2486 // 64-bit registers. In particular, sign extend the input value into the
2487 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2488 // then lfd it and fcfid it.
2489 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2490 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2491 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2492 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2494 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2497 // STD the extended value into the stack slot.
2498 MemOperand MO(PseudoSourceValue::getFixedStack(),
2499 MemOperand::MOStore, FrameIdx, 8, 8);
2500 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2501 DAG.getEntryNode(), Ext64, FIdx,
2502 DAG.getMemOperand(MO));
2503 // Load the value as a double.
2504 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2506 // FCFID it and return it.
2507 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2508 if (Op.getValueType() == MVT::f32)
2509 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2513 SDOperand PPCTargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2515 The rounding mode is in bits 30:31 of FPSR, and has the following
2522 FLT_ROUNDS, on the other hand, expects the following:
2529 To perform the conversion, we do:
2530 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2533 MachineFunction &MF = DAG.getMachineFunction();
2534 MVT::ValueType VT = Op.getValueType();
2535 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2536 std::vector<MVT::ValueType> NodeTys;
2537 SDOperand MFFSreg, InFlag;
2539 // Save FP Control Word to register
2540 NodeTys.push_back(MVT::f64); // return register
2541 NodeTys.push_back(MVT::Flag); // unused in this context
2542 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2544 // Save FP register to stack slot
2545 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2546 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2547 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2548 StackSlot, NULL, 0);
2550 // Load FP Control Word from low 32 bits of stack slot.
2551 SDOperand Four = DAG.getConstant(4, PtrVT);
2552 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2553 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2555 // Transform as necessary
2557 DAG.getNode(ISD::AND, MVT::i32,
2558 CWD, DAG.getConstant(3, MVT::i32));
2560 DAG.getNode(ISD::SRL, MVT::i32,
2561 DAG.getNode(ISD::AND, MVT::i32,
2562 DAG.getNode(ISD::XOR, MVT::i32,
2563 CWD, DAG.getConstant(3, MVT::i32)),
2564 DAG.getConstant(3, MVT::i32)),
2565 DAG.getConstant(1, MVT::i8));
2568 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2570 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2571 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2574 SDOperand PPCTargetLowering::LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2575 MVT::ValueType VT = Op.getValueType();
2576 unsigned BitWidth = MVT::getSizeInBits(VT);
2577 assert(Op.getNumOperands() == 3 &&
2578 VT == Op.getOperand(1).getValueType() &&
2581 // Expand into a bunch of logical ops. Note that these ops
2582 // depend on the PPC behavior for oversized shift amounts.
2583 SDOperand Lo = Op.getOperand(0);
2584 SDOperand Hi = Op.getOperand(1);
2585 SDOperand Amt = Op.getOperand(2);
2586 MVT::ValueType AmtVT = Amt.getValueType();
2588 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2589 DAG.getConstant(BitWidth, AmtVT), Amt);
2590 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
2591 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
2592 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2593 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2594 DAG.getConstant(-BitWidth, AmtVT));
2595 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
2596 SDOperand OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2597 SDOperand OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
2598 SDOperand OutOps[] = { OutLo, OutHi };
2599 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2603 SDOperand PPCTargetLowering::LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2604 MVT::ValueType VT = Op.getValueType();
2605 unsigned BitWidth = MVT::getSizeInBits(VT);
2606 assert(Op.getNumOperands() == 3 &&
2607 VT == Op.getOperand(1).getValueType() &&
2610 // Expand into a bunch of logical ops. Note that these ops
2611 // depend on the PPC behavior for oversized shift amounts.
2612 SDOperand Lo = Op.getOperand(0);
2613 SDOperand Hi = Op.getOperand(1);
2614 SDOperand Amt = Op.getOperand(2);
2615 MVT::ValueType AmtVT = Amt.getValueType();
2617 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2618 DAG.getConstant(BitWidth, AmtVT), Amt);
2619 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2620 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2621 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2622 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2623 DAG.getConstant(-BitWidth, AmtVT));
2624 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
2625 SDOperand OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2626 SDOperand OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
2627 SDOperand OutOps[] = { OutLo, OutHi };
2628 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2632 SDOperand PPCTargetLowering::LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2633 MVT::ValueType VT = Op.getValueType();
2634 unsigned BitWidth = MVT::getSizeInBits(VT);
2635 assert(Op.getNumOperands() == 3 &&
2636 VT == Op.getOperand(1).getValueType() &&
2639 // Expand into a bunch of logical ops, followed by a select_cc.
2640 SDOperand Lo = Op.getOperand(0);
2641 SDOperand Hi = Op.getOperand(1);
2642 SDOperand Amt = Op.getOperand(2);
2643 MVT::ValueType AmtVT = Amt.getValueType();
2645 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2646 DAG.getConstant(BitWidth, AmtVT), Amt);
2647 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2648 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2649 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2650 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2651 DAG.getConstant(-BitWidth, AmtVT));
2652 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
2653 SDOperand OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
2654 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
2655 Tmp4, Tmp6, ISD::SETLE);
2656 SDOperand OutOps[] = { OutLo, OutHi };
2657 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2661 //===----------------------------------------------------------------------===//
2662 // Vector related lowering.
2665 // If this is a vector of constants or undefs, get the bits. A bit in
2666 // UndefBits is set if the corresponding element of the vector is an
2667 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2668 // zero. Return true if this is not an array of constants, false if it is.
2670 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2671 uint64_t UndefBits[2]) {
2672 // Start with zero'd results.
2673 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2675 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2676 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2677 SDOperand OpVal = BV->getOperand(i);
2679 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2680 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2682 uint64_t EltBits = 0;
2683 if (OpVal.getOpcode() == ISD::UNDEF) {
2684 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2685 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2687 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2688 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2689 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2690 assert(CN->getValueType(0) == MVT::f32 &&
2691 "Only one legal FP vector type!");
2692 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2694 // Nonconstant element.
2698 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2701 //printf("%llx %llx %llx %llx\n",
2702 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2706 // If this is a splat (repetition) of a value across the whole vector, return
2707 // the smallest size that splats it. For example, "0x01010101010101..." is a
2708 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2709 // SplatSize = 1 byte.
2710 static bool isConstantSplat(const uint64_t Bits128[2],
2711 const uint64_t Undef128[2],
2712 unsigned &SplatBits, unsigned &SplatUndef,
2713 unsigned &SplatSize) {
2715 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2716 // the same as the lower 64-bits, ignoring undefs.
2717 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2718 return false; // Can't be a splat if two pieces don't match.
2720 uint64_t Bits64 = Bits128[0] | Bits128[1];
2721 uint64_t Undef64 = Undef128[0] & Undef128[1];
2723 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2725 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2726 return false; // Can't be a splat if two pieces don't match.
2728 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2729 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2731 // If the top 16-bits are different than the lower 16-bits, ignoring
2732 // undefs, we have an i32 splat.
2733 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2735 SplatUndef = Undef32;
2740 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2741 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2743 // If the top 8-bits are different than the lower 8-bits, ignoring
2744 // undefs, we have an i16 splat.
2745 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2747 SplatUndef = Undef16;
2752 // Otherwise, we have an 8-bit splat.
2753 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2754 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2759 /// BuildSplatI - Build a canonical splati of Val with an element size of
2760 /// SplatSize. Cast the result to VT.
2761 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2762 SelectionDAG &DAG) {
2763 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2765 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2766 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2769 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2771 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2775 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2777 // Build a canonical splat for this value.
2778 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2779 SmallVector<SDOperand, 8> Ops;
2780 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2781 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2782 &Ops[0], Ops.size());
2783 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2786 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2787 /// specified intrinsic ID.
2788 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2790 MVT::ValueType DestVT = MVT::Other) {
2791 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2792 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2793 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2796 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2797 /// specified intrinsic ID.
2798 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2799 SDOperand Op2, SelectionDAG &DAG,
2800 MVT::ValueType DestVT = MVT::Other) {
2801 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2802 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2803 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2807 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2808 /// amount. The result has the specified value type.
2809 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2810 MVT::ValueType VT, SelectionDAG &DAG) {
2811 // Force LHS/RHS to be the right type.
2812 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2813 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2816 for (unsigned i = 0; i != 16; ++i)
2817 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2818 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2819 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2820 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2823 // If this is a case we can't handle, return null and let the default
2824 // expansion code take care of it. If we CAN select this case, and if it
2825 // selects to a single instruction, return Op. Otherwise, if we can codegen
2826 // this case more efficiently than a constant pool load, lower it to the
2827 // sequence of ops that should be used.
2828 SDOperand PPCTargetLowering::LowerBUILD_VECTOR(SDOperand Op,
2829 SelectionDAG &DAG) {
2830 // If this is a vector of constants or undefs, get the bits. A bit in
2831 // UndefBits is set if the corresponding element of the vector is an
2832 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2834 uint64_t VectorBits[2];
2835 uint64_t UndefBits[2];
2836 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2837 return SDOperand(); // Not a constant vector.
2839 // If this is a splat (repetition) of a value across the whole vector, return
2840 // the smallest size that splats it. For example, "0x01010101010101..." is a
2841 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2842 // SplatSize = 1 byte.
2843 unsigned SplatBits, SplatUndef, SplatSize;
2844 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2845 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2847 // First, handle single instruction cases.
2850 if (SplatBits == 0) {
2851 // Canonicalize all zero vectors to be v4i32.
2852 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2853 SDOperand Z = DAG.getConstant(0, MVT::i32);
2854 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2855 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2860 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2861 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2862 if (SextVal >= -16 && SextVal <= 15)
2863 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2866 // Two instruction sequences.
2868 // If this value is in the range [-32,30] and is even, use:
2869 // tmp = VSPLTI[bhw], result = add tmp, tmp
2870 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2871 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2872 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2875 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2876 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2878 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2879 // Make -1 and vspltisw -1:
2880 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2882 // Make the VSLW intrinsic, computing 0x8000_0000.
2883 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2886 // xor by OnesV to invert it.
2887 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2888 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2891 // Check to see if this is a wide variety of vsplti*, binop self cases.
2892 unsigned SplatBitSize = SplatSize*8;
2893 static const signed char SplatCsts[] = {
2894 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2895 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2898 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2899 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2900 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2901 int i = SplatCsts[idx];
2903 // Figure out what shift amount will be used by altivec if shifted by i in
2905 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2907 // vsplti + shl self.
2908 if (SextVal == (i << (int)TypeShiftAmt)) {
2909 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2910 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2911 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2912 Intrinsic::ppc_altivec_vslw
2914 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2915 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2918 // vsplti + srl self.
2919 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2920 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2921 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2922 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2923 Intrinsic::ppc_altivec_vsrw
2925 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2926 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2929 // vsplti + sra self.
2930 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2931 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2932 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2933 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2934 Intrinsic::ppc_altivec_vsraw
2936 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2937 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2940 // vsplti + rol self.
2941 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2942 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2943 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2944 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2945 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2946 Intrinsic::ppc_altivec_vrlw
2948 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2949 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2952 // t = vsplti c, result = vsldoi t, t, 1
2953 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2954 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2955 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2957 // t = vsplti c, result = vsldoi t, t, 2
2958 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2959 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2960 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2962 // t = vsplti c, result = vsldoi t, t, 3
2963 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2964 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2965 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2969 // Three instruction sequences.
2971 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2972 if (SextVal >= 0 && SextVal <= 31) {
2973 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2974 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2975 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2976 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2978 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2979 if (SextVal >= -31 && SextVal <= 0) {
2980 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2981 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2982 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2983 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2990 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2991 /// the specified operations to build the shuffle.
2992 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2993 SDOperand RHS, SelectionDAG &DAG) {
2994 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2995 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2996 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2999 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3011 if (OpNum == OP_COPY) {
3012 if (LHSID == (1*9+2)*9+3) return LHS;
3013 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3017 SDOperand OpLHS, OpRHS;
3018 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
3019 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
3021 unsigned ShufIdxs[16];
3023 default: assert(0 && "Unknown i32 permute!");
3025 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3026 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3027 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3028 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3031 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3032 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3033 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3034 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3037 for (unsigned i = 0; i != 16; ++i)
3038 ShufIdxs[i] = (i&3)+0;
3041 for (unsigned i = 0; i != 16; ++i)
3042 ShufIdxs[i] = (i&3)+4;
3045 for (unsigned i = 0; i != 16; ++i)
3046 ShufIdxs[i] = (i&3)+8;
3049 for (unsigned i = 0; i != 16; ++i)
3050 ShufIdxs[i] = (i&3)+12;
3053 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
3055 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
3057 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
3060 for (unsigned i = 0; i != 16; ++i)
3061 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
3063 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
3064 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3067 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3068 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3069 /// return the code it can be lowered into. Worst case, it can always be
3070 /// lowered into a vperm.
3071 SDOperand PPCTargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op,
3072 SelectionDAG &DAG) {
3073 SDOperand V1 = Op.getOperand(0);
3074 SDOperand V2 = Op.getOperand(1);
3075 SDOperand PermMask = Op.getOperand(2);
3077 // Cases that are handled by instructions that take permute immediates
3078 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3079 // selected by the instruction selector.
3080 if (V2.getOpcode() == ISD::UNDEF) {
3081 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
3082 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
3083 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
3084 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
3085 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
3086 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
3087 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
3088 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
3089 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
3090 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
3091 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
3092 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
3097 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3098 // and produce a fixed permutation. If any of these match, do not lower to
3100 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
3101 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
3102 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
3103 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
3104 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
3105 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
3106 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
3107 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
3108 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
3111 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3112 // perfect shuffle table to emit an optimal matching sequence.
3113 unsigned PFIndexes[4];
3114 bool isFourElementShuffle = true;
3115 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3116 unsigned EltNo = 8; // Start out undef.
3117 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3118 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
3119 continue; // Undef, ignore it.
3121 unsigned ByteSource =
3122 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
3123 if ((ByteSource & 3) != j) {
3124 isFourElementShuffle = false;
3129 EltNo = ByteSource/4;
3130 } else if (EltNo != ByteSource/4) {
3131 isFourElementShuffle = false;
3135 PFIndexes[i] = EltNo;
3138 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3139 // perfect shuffle vector to determine if it is cost effective to do this as
3140 // discrete instructions, or whether we should use a vperm.
3141 if (isFourElementShuffle) {
3142 // Compute the index in the perfect shuffle table.
3143 unsigned PFTableIndex =
3144 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3146 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3147 unsigned Cost = (PFEntry >> 30);
3149 // Determining when to avoid vperm is tricky. Many things affect the cost
3150 // of vperm, particularly how many times the perm mask needs to be computed.
3151 // For example, if the perm mask can be hoisted out of a loop or is already
3152 // used (perhaps because there are multiple permutes with the same shuffle
3153 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3154 // the loop requires an extra register.
3156 // As a compromise, we only emit discrete instructions if the shuffle can be
3157 // generated in 3 or fewer operations. When we have loop information
3158 // available, if this block is within a loop, we should avoid using vperm
3159 // for 3-operation perms and use a constant pool load instead.
3161 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3164 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3165 // vector that will get spilled to the constant pool.
3166 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3168 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3169 // that it is in input element units, not in bytes. Convert now.
3170 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
3171 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
3173 SmallVector<SDOperand, 16> ResultMask;
3174 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3176 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3179 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
3181 for (unsigned j = 0; j != BytesPerElement; ++j)
3182 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3186 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3187 &ResultMask[0], ResultMask.size());
3188 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3191 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3192 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3193 /// information about the intrinsic.
3194 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
3196 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
3199 switch (IntrinsicID) {
3200 default: return false;
3201 // Comparison predicates.
3202 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3203 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3204 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3205 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3206 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3207 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3208 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3209 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3210 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3211 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3212 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3213 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3214 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3216 // Normal Comparisons.
3217 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3218 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3219 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3220 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3221 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3222 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3223 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3224 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3225 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3226 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3227 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3228 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3229 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3234 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3235 /// lower, do it, otherwise return null.
3236 SDOperand PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op,
3237 SelectionDAG &DAG) {
3238 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3239 // opcode number of the comparison.
3242 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3243 return SDOperand(); // Don't custom lower most intrinsics.
3245 // If this is a non-dot comparison, make the VCMP node and we are done.
3247 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3248 Op.getOperand(1), Op.getOperand(2),
3249 DAG.getConstant(CompareOpc, MVT::i32));
3250 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3253 // Create the PPCISD altivec 'dot' comparison node.
3255 Op.getOperand(2), // LHS
3256 Op.getOperand(3), // RHS
3257 DAG.getConstant(CompareOpc, MVT::i32)
3259 std::vector<MVT::ValueType> VTs;
3260 VTs.push_back(Op.getOperand(2).getValueType());
3261 VTs.push_back(MVT::Flag);
3262 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3264 // Now that we have the comparison, emit a copy from the CR to a GPR.
3265 // This is flagged to the above dot comparison.
3266 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3267 DAG.getRegister(PPC::CR6, MVT::i32),
3268 CompNode.getValue(1));
3270 // Unpack the result based on how the target uses it.
3271 unsigned BitNo; // Bit # of CR6.
3272 bool InvertBit; // Invert result?
3273 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
3274 default: // Can't happen, don't crash on invalid number though.
3275 case 0: // Return the value of the EQ bit of CR6.
3276 BitNo = 0; InvertBit = false;
3278 case 1: // Return the inverted value of the EQ bit of CR6.
3279 BitNo = 0; InvertBit = true;
3281 case 2: // Return the value of the LT bit of CR6.
3282 BitNo = 2; InvertBit = false;
3284 case 3: // Return the inverted value of the LT bit of CR6.
3285 BitNo = 2; InvertBit = true;
3289 // Shift the bit into the low position.
3290 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3291 DAG.getConstant(8-(3-BitNo), MVT::i32));
3293 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3294 DAG.getConstant(1, MVT::i32));
3296 // If we are supposed to, toggle the bit.
3298 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3299 DAG.getConstant(1, MVT::i32));
3303 SDOperand PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op,
3304 SelectionDAG &DAG) {
3305 // Create a stack slot that is 16-byte aligned.
3306 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3307 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3308 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3309 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3311 // Store the input value into Value#0 of the stack slot.
3312 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
3313 Op.getOperand(0), FIdx, NULL, 0);
3315 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3318 SDOperand PPCTargetLowering::LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3319 if (Op.getValueType() == MVT::v4i32) {
3320 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3322 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3323 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3325 SDOperand RHSSwap = // = vrlw RHS, 16
3326 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3328 // Shrinkify inputs to v8i16.
3329 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3330 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3331 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3333 // Low parts multiplied together, generating 32-bit results (we ignore the
3335 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3336 LHS, RHS, DAG, MVT::v4i32);
3338 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3339 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3340 // Shift the high parts up 16 bits.
3341 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3342 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3343 } else if (Op.getValueType() == MVT::v8i16) {
3344 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3346 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3348 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3349 LHS, RHS, Zero, DAG);
3350 } else if (Op.getValueType() == MVT::v16i8) {
3351 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3353 // Multiply the even 8-bit parts, producing 16-bit sums.
3354 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3355 LHS, RHS, DAG, MVT::v8i16);
3356 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3358 // Multiply the odd 8-bit parts, producing 16-bit sums.
3359 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3360 LHS, RHS, DAG, MVT::v8i16);
3361 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3363 // Merge the results together.
3365 for (unsigned i = 0; i != 8; ++i) {
3366 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3367 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3369 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3370 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3372 assert(0 && "Unknown mul to lower!");
3377 /// LowerOperation - Provide custom lowering hooks for some operations.
3379 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3380 switch (Op.getOpcode()) {
3381 default: assert(0 && "Wasn't expecting to be able to lower this!");
3382 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3383 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3384 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3385 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3386 case ISD::SETCC: return LowerSETCC(Op, DAG);
3388 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3389 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3392 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3393 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3395 case ISD::FORMAL_ARGUMENTS:
3396 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3397 VarArgsStackOffset, VarArgsNumGPR,
3398 VarArgsNumFPR, PPCSubTarget);
3400 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3401 getTargetMachine());
3402 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3403 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3404 case ISD::DYNAMIC_STACKALLOC:
3405 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3407 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3408 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3409 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3410 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3411 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3413 // Lower 64-bit shifts.
3414 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3415 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3416 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3418 // Vector-related lowering.
3419 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3420 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3421 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3422 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3423 case ISD::MUL: return LowerMUL(Op, DAG);
3425 // Frame & Return address.
3426 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3427 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3432 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3433 switch (N->getOpcode()) {
3434 default: assert(0 && "Wasn't expecting to be able to lower this!");
3435 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3440 //===----------------------------------------------------------------------===//
3441 // Other Lowering Code
3442 //===----------------------------------------------------------------------===//
3445 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3446 MachineBasicBlock *BB) {
3447 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3448 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3449 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3450 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3451 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3452 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3453 "Unexpected instr type to insert");
3455 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3456 // control-flow pattern. The incoming instruction knows the destination vreg
3457 // to set, the condition code register to branch on, the true/false values to
3458 // select between, and a branch opcode to use.
3459 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3460 ilist<MachineBasicBlock>::iterator It = BB;
3466 // cmpTY ccX, r1, r2
3468 // fallthrough --> copy0MBB
3469 MachineBasicBlock *thisMBB = BB;
3470 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3471 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3472 unsigned SelectPred = MI->getOperand(4).getImm();
3473 BuildMI(BB, TII->get(PPC::BCC))
3474 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3475 MachineFunction *F = BB->getParent();
3476 F->getBasicBlockList().insert(It, copy0MBB);
3477 F->getBasicBlockList().insert(It, sinkMBB);
3478 // Update machine-CFG edges by first adding all successors of the current
3479 // block to the new block which will contain the Phi node for the select.
3480 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3481 e = BB->succ_end(); i != e; ++i)
3482 sinkMBB->addSuccessor(*i);
3483 // Next, remove all successors of the current block, and add the true
3484 // and fallthrough blocks as its successors.
3485 while(!BB->succ_empty())
3486 BB->removeSuccessor(BB->succ_begin());
3487 BB->addSuccessor(copy0MBB);
3488 BB->addSuccessor(sinkMBB);
3491 // %FalseValue = ...
3492 // # fallthrough to sinkMBB
3495 // Update machine-CFG edges
3496 BB->addSuccessor(sinkMBB);
3499 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3502 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3503 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3504 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3506 delete MI; // The pseudo instruction is gone now.
3510 //===----------------------------------------------------------------------===//
3511 // Target Optimization Hooks
3512 //===----------------------------------------------------------------------===//
3514 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3515 DAGCombinerInfo &DCI) const {
3516 TargetMachine &TM = getTargetMachine();
3517 SelectionDAG &DAG = DCI.DAG;
3518 switch (N->getOpcode()) {
3521 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3522 if (C->getValue() == 0) // 0 << V -> 0.
3523 return N->getOperand(0);
3527 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3528 if (C->getValue() == 0) // 0 >>u V -> 0.
3529 return N->getOperand(0);
3533 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3534 if (C->getValue() == 0 || // 0 >>s V -> 0.
3535 C->isAllOnesValue()) // -1 >>s V -> -1.
3536 return N->getOperand(0);
3540 case ISD::SINT_TO_FP:
3541 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3542 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3543 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3544 // We allow the src/dst to be either f32/f64, but the intermediate
3545 // type must be i64.
3546 if (N->getOperand(0).getValueType() == MVT::i64 &&
3547 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3548 SDOperand Val = N->getOperand(0).getOperand(0);
3549 if (Val.getValueType() == MVT::f32) {
3550 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3551 DCI.AddToWorklist(Val.Val);
3554 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3555 DCI.AddToWorklist(Val.Val);
3556 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3557 DCI.AddToWorklist(Val.Val);
3558 if (N->getValueType(0) == MVT::f32) {
3559 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3560 DAG.getIntPtrConstant(0));
3561 DCI.AddToWorklist(Val.Val);
3564 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3565 // If the intermediate type is i32, we can avoid the load/store here
3572 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3573 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3574 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3575 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3576 N->getOperand(1).getValueType() == MVT::i32 &&
3577 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3578 SDOperand Val = N->getOperand(1).getOperand(0);
3579 if (Val.getValueType() == MVT::f32) {
3580 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3581 DCI.AddToWorklist(Val.Val);
3583 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3584 DCI.AddToWorklist(Val.Val);
3586 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3587 N->getOperand(2), N->getOperand(3));
3588 DCI.AddToWorklist(Val.Val);
3592 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3593 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3594 N->getOperand(1).Val->hasOneUse() &&
3595 (N->getOperand(1).getValueType() == MVT::i32 ||
3596 N->getOperand(1).getValueType() == MVT::i16)) {
3597 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3598 // Do an any-extend to 32-bits if this is a half-word input.
3599 if (BSwapOp.getValueType() == MVT::i16)
3600 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3602 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3603 N->getOperand(2), N->getOperand(3),
3604 DAG.getValueType(N->getOperand(1).getValueType()));
3608 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3609 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3610 N->getOperand(0).hasOneUse() &&
3611 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3612 SDOperand Load = N->getOperand(0);
3613 LoadSDNode *LD = cast<LoadSDNode>(Load);
3614 // Create the byte-swapping load.
3615 std::vector<MVT::ValueType> VTs;
3616 VTs.push_back(MVT::i32);
3617 VTs.push_back(MVT::Other);
3618 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3620 LD->getChain(), // Chain
3621 LD->getBasePtr(), // Ptr
3623 DAG.getValueType(N->getValueType(0)) // VT
3625 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3627 // If this is an i16 load, insert the truncate.
3628 SDOperand ResVal = BSLoad;
3629 if (N->getValueType(0) == MVT::i16)
3630 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3632 // First, combine the bswap away. This makes the value produced by the
3634 DCI.CombineTo(N, ResVal);
3636 // Next, combine the load away, we give it a bogus result value but a real
3637 // chain result. The result value is dead because the bswap is dead.
3638 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3640 // Return N so it doesn't get rechecked!
3641 return SDOperand(N, 0);
3645 case PPCISD::VCMP: {
3646 // If a VCMPo node already exists with exactly the same operands as this
3647 // node, use its result instead of this node (VCMPo computes both a CR6 and
3648 // a normal output).
3650 if (!N->getOperand(0).hasOneUse() &&
3651 !N->getOperand(1).hasOneUse() &&
3652 !N->getOperand(2).hasOneUse()) {
3654 // Scan all of the users of the LHS, looking for VCMPo's that match.
3655 SDNode *VCMPoNode = 0;
3657 SDNode *LHSN = N->getOperand(0).Val;
3658 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3660 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3661 (*UI)->getOperand(1) == N->getOperand(1) &&
3662 (*UI)->getOperand(2) == N->getOperand(2) &&
3663 (*UI)->getOperand(0) == N->getOperand(0)) {
3668 // If there is no VCMPo node, or if the flag value has a single use, don't
3670 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3673 // Look at the (necessarily single) use of the flag value. If it has a
3674 // chain, this transformation is more complex. Note that multiple things
3675 // could use the value result, which we should ignore.
3676 SDNode *FlagUser = 0;
3677 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3678 FlagUser == 0; ++UI) {
3679 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3681 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3682 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3689 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3690 // give up for right now.
3691 if (FlagUser->getOpcode() == PPCISD::MFCR)
3692 return SDOperand(VCMPoNode, 0);
3697 // If this is a branch on an altivec predicate comparison, lower this so
3698 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3699 // lowering is done pre-legalize, because the legalizer lowers the predicate
3700 // compare down to code that is difficult to reassemble.
3701 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3702 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3706 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3707 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3708 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3709 assert(isDot && "Can't compare against a vector result!");
3711 // If this is a comparison against something other than 0/1, then we know
3712 // that the condition is never/always true.
3713 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3714 if (Val != 0 && Val != 1) {
3715 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3716 return N->getOperand(0);
3717 // Always !=, turn it into an unconditional branch.
3718 return DAG.getNode(ISD::BR, MVT::Other,
3719 N->getOperand(0), N->getOperand(4));
3722 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3724 // Create the PPCISD altivec 'dot' comparison node.
3725 std::vector<MVT::ValueType> VTs;
3727 LHS.getOperand(2), // LHS of compare
3728 LHS.getOperand(3), // RHS of compare
3729 DAG.getConstant(CompareOpc, MVT::i32)
3731 VTs.push_back(LHS.getOperand(2).getValueType());
3732 VTs.push_back(MVT::Flag);
3733 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3735 // Unpack the result based on how the target uses it.
3736 PPC::Predicate CompOpc;
3737 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3738 default: // Can't happen, don't crash on invalid number though.
3739 case 0: // Branch on the value of the EQ bit of CR6.
3740 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3742 case 1: // Branch on the inverted value of the EQ bit of CR6.
3743 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3745 case 2: // Branch on the value of the LT bit of CR6.
3746 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3748 case 3: // Branch on the inverted value of the LT bit of CR6.
3749 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3753 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3754 DAG.getConstant(CompOpc, MVT::i32),
3755 DAG.getRegister(PPC::CR6, MVT::i32),
3756 N->getOperand(4), CompNode.getValue(1));
3765 //===----------------------------------------------------------------------===//
3766 // Inline Assembly Support
3767 //===----------------------------------------------------------------------===//
3769 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3773 const SelectionDAG &DAG,
3774 unsigned Depth) const {
3775 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3776 switch (Op.getOpcode()) {
3778 case PPCISD::LBRX: {
3779 // lhbrx is known to have the top bits cleared out.
3780 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3781 KnownZero = 0xFFFF0000;
3784 case ISD::INTRINSIC_WO_CHAIN: {
3785 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3787 case Intrinsic::ppc_altivec_vcmpbfp_p:
3788 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3789 case Intrinsic::ppc_altivec_vcmpequb_p:
3790 case Intrinsic::ppc_altivec_vcmpequh_p:
3791 case Intrinsic::ppc_altivec_vcmpequw_p:
3792 case Intrinsic::ppc_altivec_vcmpgefp_p:
3793 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3794 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3795 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3796 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3797 case Intrinsic::ppc_altivec_vcmpgtub_p:
3798 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3799 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3800 KnownZero = ~1U; // All bits but the low one are known to be zero.
3808 /// getConstraintType - Given a constraint, return the type of
3809 /// constraint it is for this target.
3810 PPCTargetLowering::ConstraintType
3811 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3812 if (Constraint.size() == 1) {
3813 switch (Constraint[0]) {
3820 return C_RegisterClass;
3823 return TargetLowering::getConstraintType(Constraint);
3826 std::pair<unsigned, const TargetRegisterClass*>
3827 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3828 MVT::ValueType VT) const {
3829 if (Constraint.size() == 1) {
3830 // GCC RS6000 Constraint Letters
3831 switch (Constraint[0]) {
3834 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3835 return std::make_pair(0U, PPC::G8RCRegisterClass);
3836 return std::make_pair(0U, PPC::GPRCRegisterClass);
3839 return std::make_pair(0U, PPC::F4RCRegisterClass);
3840 else if (VT == MVT::f64)
3841 return std::make_pair(0U, PPC::F8RCRegisterClass);
3844 return std::make_pair(0U, PPC::VRRCRegisterClass);
3846 return std::make_pair(0U, PPC::CRRCRegisterClass);
3850 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3854 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3855 /// vector. If it is invalid, don't add anything to Ops.
3856 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3857 std::vector<SDOperand>&Ops,
3858 SelectionDAG &DAG) {
3859 SDOperand Result(0,0);
3870 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3871 if (!CST) return; // Must be an immediate to match.
3872 unsigned Value = CST->getValue();
3874 default: assert(0 && "Unknown constraint letter!");
3875 case 'I': // "I" is a signed 16-bit constant.
3876 if ((short)Value == (int)Value)
3877 Result = DAG.getTargetConstant(Value, Op.getValueType());
3879 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3880 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3881 if ((short)Value == 0)
3882 Result = DAG.getTargetConstant(Value, Op.getValueType());
3884 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3885 if ((Value >> 16) == 0)
3886 Result = DAG.getTargetConstant(Value, Op.getValueType());
3888 case 'M': // "M" is a constant that is greater than 31.
3890 Result = DAG.getTargetConstant(Value, Op.getValueType());
3892 case 'N': // "N" is a positive constant that is an exact power of two.
3893 if ((int)Value > 0 && isPowerOf2_32(Value))
3894 Result = DAG.getTargetConstant(Value, Op.getValueType());
3896 case 'O': // "O" is the constant zero.
3898 Result = DAG.getTargetConstant(Value, Op.getValueType());
3900 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3901 if ((short)-Value == (int)-Value)
3902 Result = DAG.getTargetConstant(Value, Op.getValueType());
3910 Ops.push_back(Result);
3914 // Handle standard constraint letters.
3915 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3918 // isLegalAddressingMode - Return true if the addressing mode represented
3919 // by AM is legal for this target, for a load/store of the specified type.
3920 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3921 const Type *Ty) const {
3922 // FIXME: PPC does not allow r+i addressing modes for vectors!
3924 // PPC allows a sign-extended 16-bit immediate field.
3925 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3928 // No global is ever allowed as a base.
3932 // PPC only support r+r,
3934 case 0: // "r+i" or just "i", depending on HasBaseReg.
3937 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3939 // Otherwise we have r+r or r+i.
3942 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3944 // Allow 2*r as r+r.
3947 // No other scales are supported.
3954 /// isLegalAddressImmediate - Return true if the integer value can be used
3955 /// as the offset of the target addressing mode for load / store of the
3957 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3958 // PPC allows a sign-extended 16-bit immediate field.
3959 return (V > -(1 << 16) && V < (1 << 16)-1);
3962 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3966 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3967 // Depths > 0 not supported yet!
3968 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3971 MachineFunction &MF = DAG.getMachineFunction();
3972 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3973 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3975 bool isPPC64 = PPCSubTarget.isPPC64();
3977 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3979 // Set up a frame object for the return address.
3980 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3982 // Remember it for next time.
3983 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3985 // Make sure the function really does not optimize away the store of the RA
3987 FuncInfo->setLRStoreRequired();
3990 // Just load the return address off the stack.
3991 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3992 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3995 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3996 // Depths > 0 not supported yet!
3997 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4000 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4001 bool isPPC64 = PtrVT == MVT::i64;
4003 MachineFunction &MF = DAG.getMachineFunction();
4004 MachineFrameInfo *MFI = MF.getFrameInfo();
4005 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4006 && MFI->getStackSize();
4009 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
4012 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,