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(N.getOperand(0)
826 .getValueSizeInBits()),
827 LHSKnownZero, LHSKnownOne);
829 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
830 // If all of the bits are known zero on the LHS or RHS, the add won't
832 Base = N.getOperand(0);
833 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
837 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
838 // Loading from a constant address.
840 // If this address fits entirely in a 16-bit sext immediate field, codegen
843 if (isIntS16Immediate(CN, Imm)) {
844 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
845 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
849 // Handle 32-bit sext immediates with LIS + addr mode.
850 if (CN->getValueType(0) == MVT::i32 ||
851 (int64_t)CN->getValue() == (int)CN->getValue()) {
852 int Addr = (int)CN->getValue();
854 // Otherwise, break this down into an LIS + disp.
855 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
857 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
858 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
859 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
864 Disp = DAG.getTargetConstant(0, getPointerTy());
865 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
866 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
869 return true; // [r+0]
872 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
873 /// represented as an indexed [r+r] operation.
874 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
877 // Check to see if we can easily represent this as an [r+r] address. This
878 // will fail if it thinks that the address is more profitably represented as
879 // reg+imm, e.g. where imm = 0.
880 if (SelectAddressRegReg(N, Base, Index, DAG))
883 // If the operand is an addition, always emit this as [r+r], since this is
884 // better (for code size, and execution, as the memop does the add for free)
885 // than emitting an explicit add.
886 if (N.getOpcode() == ISD::ADD) {
887 Base = N.getOperand(0);
888 Index = N.getOperand(1);
892 // Otherwise, do it the hard way, using R0 as the base register.
893 Base = DAG.getRegister(PPC::R0, N.getValueType());
898 /// SelectAddressRegImmShift - Returns true if the address N can be
899 /// represented by a base register plus a signed 14-bit displacement
900 /// [r+imm*4]. Suitable for use by STD and friends.
901 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
904 // If this can be more profitably realized as r+r, fail.
905 if (SelectAddressRegReg(N, Disp, Base, DAG))
908 if (N.getOpcode() == ISD::ADD) {
910 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
911 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
912 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
913 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
915 Base = N.getOperand(0);
917 return true; // [r+i]
918 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
919 // Match LOAD (ADD (X, Lo(G))).
920 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
921 && "Cannot handle constant offsets yet!");
922 Disp = N.getOperand(1).getOperand(0); // The global address.
923 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
924 Disp.getOpcode() == ISD::TargetConstantPool ||
925 Disp.getOpcode() == ISD::TargetJumpTable);
926 Base = N.getOperand(0);
927 return true; // [&g+r]
929 } else if (N.getOpcode() == ISD::OR) {
931 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
932 // If this is an or of disjoint bitfields, we can codegen this as an add
933 // (for better address arithmetic) if the LHS and RHS of the OR are
934 // provably disjoint.
935 APInt LHSKnownZero, LHSKnownOne;
936 DAG.ComputeMaskedBits(N.getOperand(0),
937 APInt::getAllOnesValue(N.getOperand(0)
938 .getValueSizeInBits()),
939 LHSKnownZero, LHSKnownOne);
940 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
941 // If all of the bits are known zero on the LHS or RHS, the add won't
943 Base = N.getOperand(0);
944 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
948 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
949 // Loading from a constant address. Verify low two bits are clear.
950 if ((CN->getValue() & 3) == 0) {
951 // If this address fits entirely in a 14-bit sext immediate field, codegen
954 if (isIntS16Immediate(CN, Imm)) {
955 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
956 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
960 // Fold the low-part of 32-bit absolute addresses into addr mode.
961 if (CN->getValueType(0) == MVT::i32 ||
962 (int64_t)CN->getValue() == (int)CN->getValue()) {
963 int Addr = (int)CN->getValue();
965 // Otherwise, break this down into an LIS + disp.
966 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
968 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
969 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
970 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
976 Disp = DAG.getTargetConstant(0, getPointerTy());
977 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
978 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
981 return true; // [r+0]
985 /// getPreIndexedAddressParts - returns true by value, base pointer and
986 /// offset pointer and addressing mode by reference if the node's address
987 /// can be legally represented as pre-indexed load / store address.
988 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
990 ISD::MemIndexedMode &AM,
992 // Disabled by default for now.
993 if (!EnablePPCPreinc) return false;
997 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
998 Ptr = LD->getBasePtr();
999 VT = LD->getMemoryVT();
1001 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1003 Ptr = ST->getBasePtr();
1004 VT = ST->getMemoryVT();
1008 // PowerPC doesn't have preinc load/store instructions for vectors.
1009 if (MVT::isVector(VT))
1012 // TODO: Check reg+reg first.
1014 // LDU/STU use reg+imm*4, others use reg+imm.
1015 if (VT != MVT::i64) {
1017 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1021 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1025 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1026 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1027 // sext i32 to i64 when addr mode is r+i.
1028 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1029 LD->getExtensionType() == ISD::SEXTLOAD &&
1030 isa<ConstantSDNode>(Offset))
1038 //===----------------------------------------------------------------------===//
1039 // LowerOperation implementation
1040 //===----------------------------------------------------------------------===//
1042 SDOperand PPCTargetLowering::LowerConstantPool(SDOperand Op,
1043 SelectionDAG &DAG) {
1044 MVT::ValueType PtrVT = Op.getValueType();
1045 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1046 Constant *C = CP->getConstVal();
1047 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1048 SDOperand Zero = DAG.getConstant(0, PtrVT);
1050 const TargetMachine &TM = DAG.getTarget();
1052 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1053 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1055 // If this is a non-darwin platform, we don't support non-static relo models
1057 if (TM.getRelocationModel() == Reloc::Static ||
1058 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1059 // Generate non-pic code that has direct accesses to the constant pool.
1060 // The address of the global is just (hi(&g)+lo(&g)).
1061 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1064 if (TM.getRelocationModel() == Reloc::PIC_) {
1065 // With PIC, the first instruction is actually "GR+hi(&G)".
1066 Hi = DAG.getNode(ISD::ADD, PtrVT,
1067 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1070 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1074 SDOperand PPCTargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1075 MVT::ValueType PtrVT = Op.getValueType();
1076 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1077 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1078 SDOperand Zero = DAG.getConstant(0, PtrVT);
1080 const TargetMachine &TM = DAG.getTarget();
1082 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1083 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1085 // If this is a non-darwin platform, we don't support non-static relo models
1087 if (TM.getRelocationModel() == Reloc::Static ||
1088 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1089 // Generate non-pic code that has direct accesses to the constant pool.
1090 // The address of the global is just (hi(&g)+lo(&g)).
1091 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1094 if (TM.getRelocationModel() == Reloc::PIC_) {
1095 // With PIC, the first instruction is actually "GR+hi(&G)".
1096 Hi = DAG.getNode(ISD::ADD, PtrVT,
1097 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1100 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1104 SDOperand PPCTargetLowering::LowerGlobalTLSAddress(SDOperand Op,
1105 SelectionDAG &DAG) {
1106 assert(0 && "TLS not implemented for PPC.");
1109 SDOperand PPCTargetLowering::LowerGlobalAddress(SDOperand Op,
1110 SelectionDAG &DAG) {
1111 MVT::ValueType PtrVT = Op.getValueType();
1112 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1113 GlobalValue *GV = GSDN->getGlobal();
1114 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1115 // If it's a debug information descriptor, don't mess with it.
1116 if (DAG.isVerifiedDebugInfoDesc(Op))
1118 SDOperand Zero = DAG.getConstant(0, PtrVT);
1120 const TargetMachine &TM = DAG.getTarget();
1122 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1123 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1125 // If this is a non-darwin platform, we don't support non-static relo models
1127 if (TM.getRelocationModel() == Reloc::Static ||
1128 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1129 // Generate non-pic code that has direct accesses to globals.
1130 // The address of the global is just (hi(&g)+lo(&g)).
1131 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1134 if (TM.getRelocationModel() == Reloc::PIC_) {
1135 // With PIC, the first instruction is actually "GR+hi(&G)".
1136 Hi = DAG.getNode(ISD::ADD, PtrVT,
1137 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1140 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1142 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1145 // If the global is weak or external, we have to go through the lazy
1147 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1150 SDOperand PPCTargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1151 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1153 // If we're comparing for equality to zero, expose the fact that this is
1154 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1155 // fold the new nodes.
1156 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1157 if (C->isNullValue() && CC == ISD::SETEQ) {
1158 MVT::ValueType VT = Op.getOperand(0).getValueType();
1159 SDOperand Zext = Op.getOperand(0);
1160 if (VT < MVT::i32) {
1162 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1164 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1165 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1166 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1167 DAG.getConstant(Log2b, MVT::i32));
1168 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1170 // Leave comparisons against 0 and -1 alone for now, since they're usually
1171 // optimized. FIXME: revisit this when we can custom lower all setcc
1173 if (C->isAllOnesValue() || C->isNullValue())
1177 // If we have an integer seteq/setne, turn it into a compare against zero
1178 // by xor'ing the rhs with the lhs, which is faster than setting a
1179 // condition register, reading it back out, and masking the correct bit. The
1180 // normal approach here uses sub to do this instead of xor. Using xor exposes
1181 // the result to other bit-twiddling opportunities.
1182 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1183 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1184 MVT::ValueType VT = Op.getValueType();
1185 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1187 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1192 SDOperand PPCTargetLowering::LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1193 int VarArgsFrameIndex,
1194 int VarArgsStackOffset,
1195 unsigned VarArgsNumGPR,
1196 unsigned VarArgsNumFPR,
1197 const PPCSubtarget &Subtarget) {
1199 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1202 SDOperand PPCTargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1203 int VarArgsFrameIndex,
1204 int VarArgsStackOffset,
1205 unsigned VarArgsNumGPR,
1206 unsigned VarArgsNumFPR,
1207 const PPCSubtarget &Subtarget) {
1209 if (Subtarget.isMachoABI()) {
1210 // vastart just stores the address of the VarArgsFrameIndex slot into the
1211 // memory location argument.
1212 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1213 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1214 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1215 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1218 // For ELF 32 ABI we follow the layout of the va_list struct.
1219 // We suppose the given va_list is already allocated.
1222 // char gpr; /* index into the array of 8 GPRs
1223 // * stored in the register save area
1224 // * gpr=0 corresponds to r3,
1225 // * gpr=1 to r4, etc.
1227 // char fpr; /* index into the array of 8 FPRs
1228 // * stored in the register save area
1229 // * fpr=0 corresponds to f1,
1230 // * fpr=1 to f2, etc.
1232 // char *overflow_arg_area;
1233 // /* location on stack that holds
1234 // * the next overflow argument
1236 // char *reg_save_area;
1237 // /* where r3:r10 and f1:f8 (if saved)
1243 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1244 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1247 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1249 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1250 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1252 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1253 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1255 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1256 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1258 uint64_t FPROffset = 1;
1259 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1261 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1263 // Store first byte : number of int regs
1264 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1265 Op.getOperand(1), SV, 0);
1266 uint64_t nextOffset = FPROffset;
1267 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1270 // Store second byte : number of float regs
1271 SDOperand secondStore =
1272 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1273 nextOffset += StackOffset;
1274 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1276 // Store second word : arguments given on stack
1277 SDOperand thirdStore =
1278 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1279 nextOffset += FrameOffset;
1280 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1282 // Store third word : arguments given in registers
1283 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1287 #include "PPCGenCallingConv.inc"
1289 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1290 /// depending on which subtarget is selected.
1291 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1292 if (Subtarget.isMachoABI()) {
1293 static const unsigned FPR[] = {
1294 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1295 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1301 static const unsigned FPR[] = {
1302 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1309 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op,
1311 int &VarArgsFrameIndex,
1312 int &VarArgsStackOffset,
1313 unsigned &VarArgsNumGPR,
1314 unsigned &VarArgsNumFPR,
1315 const PPCSubtarget &Subtarget) {
1316 // TODO: add description of PPC stack frame format, or at least some docs.
1318 MachineFunction &MF = DAG.getMachineFunction();
1319 MachineFrameInfo *MFI = MF.getFrameInfo();
1320 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1321 SmallVector<SDOperand, 8> ArgValues;
1322 SDOperand Root = Op.getOperand(0);
1323 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1325 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1326 bool isPPC64 = PtrVT == MVT::i64;
1327 bool isMachoABI = Subtarget.isMachoABI();
1328 bool isELF32_ABI = Subtarget.isELF32_ABI();
1329 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1331 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1333 static const unsigned GPR_32[] = { // 32-bit registers.
1334 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1335 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1337 static const unsigned GPR_64[] = { // 64-bit registers.
1338 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1339 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1342 static const unsigned *FPR = GetFPR(Subtarget);
1344 static const unsigned VR[] = {
1345 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1346 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1349 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1350 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1351 const unsigned Num_VR_Regs = array_lengthof( VR);
1353 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1355 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1357 // In 32-bit non-varargs functions, the stack space for vectors is after the
1358 // stack space for non-vectors. We do not use this space unless we have
1359 // too many vectors to fit in registers, something that only occurs in
1360 // constructed examples:), but we have to walk the arglist to figure
1361 // that out...for the pathological case, compute VecArgOffset as the
1362 // start of the vector parameter area. Computing VecArgOffset is the
1363 // entire point of the following loop.
1364 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1365 // to handle Elf here.
1366 unsigned VecArgOffset = ArgOffset;
1367 if (!isVarArg && !isPPC64) {
1368 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e;
1370 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1371 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1372 ISD::ArgFlagsTy Flags =
1373 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1375 if (Flags.isByVal()) {
1376 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1377 ObjSize = Flags.getByValSize();
1379 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1380 VecArgOffset += ArgSize;
1385 default: assert(0 && "Unhandled argument type!");
1388 VecArgOffset += isPPC64 ? 8 : 4;
1390 case MVT::i64: // PPC64
1398 // Nothing to do, we're only looking at Nonvector args here.
1403 // We've found where the vector parameter area in memory is. Skip the
1404 // first 12 parameters; these don't use that memory.
1405 VecArgOffset = ((VecArgOffset+15)/16)*16;
1406 VecArgOffset += 12*16;
1408 // Add DAG nodes to load the arguments or copy them out of registers. On
1409 // entry to a function on PPC, the arguments start after the linkage area,
1410 // although the first ones are often in registers.
1412 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1413 // represented with two words (long long or double) must be copied to an
1414 // even GPR_idx value or to an even ArgOffset value. TODO: implement this.
1416 SmallVector<SDOperand, 8> MemOps;
1418 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1420 bool needsLoad = false;
1421 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1422 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1423 unsigned ArgSize = ObjSize;
1424 ISD::ArgFlagsTy Flags =
1425 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1426 // See if next argument requires stack alignment in ELF
1427 bool Expand = false; // TODO: implement this.
1429 unsigned CurArgOffset = ArgOffset;
1431 // FIXME alignment for ELF may not be right
1432 // FIXME the codegen can be much improved in some cases.
1433 // We do not have to keep everything in memory.
1434 if (Flags.isByVal()) {
1435 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1436 ObjSize = Flags.getByValSize();
1437 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1438 // Double word align in ELF
1439 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1440 // Objects of size 1 and 2 are right justified, everything else is
1441 // left justified. This means the memory address is adjusted forwards.
1442 if (ObjSize==1 || ObjSize==2) {
1443 CurArgOffset = CurArgOffset + (4 - ObjSize);
1445 // The value of the object is its address.
1446 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1447 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1448 ArgValues.push_back(FIN);
1449 if (ObjSize==1 || ObjSize==2) {
1450 if (GPR_idx != Num_GPR_Regs) {
1451 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1452 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1453 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1454 SDOperand Store = DAG.getTruncStore(Val.getValue(1), Val, FIN,
1455 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1456 MemOps.push_back(Store);
1458 if (isMachoABI) ArgOffset += PtrByteSize;
1460 ArgOffset += PtrByteSize;
1464 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1465 // Store whatever pieces of the object are in registers
1466 // to memory. ArgVal will be address of the beginning of
1468 if (GPR_idx != Num_GPR_Regs) {
1469 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1470 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1471 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1472 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1473 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1474 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1475 MemOps.push_back(Store);
1477 if (isMachoABI) ArgOffset += PtrByteSize;
1479 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1487 default: assert(0 && "Unhandled argument type!");
1490 // Double word align in ELF
1491 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1493 if (GPR_idx != Num_GPR_Regs) {
1494 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1495 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1496 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1500 ArgSize = PtrByteSize;
1502 // Stack align in ELF
1503 if (needsLoad && Expand && isELF32_ABI)
1504 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1505 // All int arguments reserve stack space in Macho ABI.
1506 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1510 case MVT::i64: // PPC64
1511 if (GPR_idx != Num_GPR_Regs) {
1512 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1513 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1514 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1516 if (ObjectVT == MVT::i32) {
1517 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1518 // value to MVT::i64 and then truncate to the correct register size.
1520 ArgVal = DAG.getNode(ISD::AssertSext, MVT::i64, ArgVal,
1521 DAG.getValueType(ObjectVT));
1522 else if (Flags.isZExt())
1523 ArgVal = DAG.getNode(ISD::AssertZext, MVT::i64, ArgVal,
1524 DAG.getValueType(ObjectVT));
1526 ArgVal = DAG.getNode(ISD::TRUNCATE, MVT::i32, ArgVal);
1533 // All int arguments reserve stack space in Macho ABI.
1534 if (isMachoABI || needsLoad) ArgOffset += 8;
1539 // Every 4 bytes of argument space consumes one of the GPRs available for
1540 // argument passing.
1541 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1543 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1546 if (FPR_idx != Num_FPR_Regs) {
1548 if (ObjectVT == MVT::f32)
1549 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1551 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1552 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1553 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1559 // Stack align in ELF
1560 if (needsLoad && Expand && isELF32_ABI)
1561 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1562 // All FP arguments reserve stack space in Macho ABI.
1563 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1569 // Note that vector arguments in registers don't reserve stack space,
1570 // except in varargs functions.
1571 if (VR_idx != Num_VR_Regs) {
1572 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1573 RegInfo.addLiveIn(VR[VR_idx], VReg);
1574 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1576 while ((ArgOffset % 16) != 0) {
1577 ArgOffset += PtrByteSize;
1578 if (GPR_idx != Num_GPR_Regs)
1582 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1586 if (!isVarArg && !isPPC64) {
1587 // Vectors go after all the nonvectors.
1588 CurArgOffset = VecArgOffset;
1591 // Vectors are aligned.
1592 ArgOffset = ((ArgOffset+15)/16)*16;
1593 CurArgOffset = ArgOffset;
1601 // We need to load the argument to a virtual register if we determined above
1602 // that we ran out of physical registers of the appropriate type.
1604 int FI = MFI->CreateFixedObject(ObjSize,
1605 CurArgOffset + (ArgSize - ObjSize));
1606 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1607 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1610 ArgValues.push_back(ArgVal);
1613 // If the function takes variable number of arguments, make a frame index for
1614 // the start of the first vararg value... for expansion of llvm.va_start.
1619 VarArgsNumGPR = GPR_idx;
1620 VarArgsNumFPR = FPR_idx;
1622 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1624 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1625 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1626 MVT::getSizeInBits(PtrVT)/8);
1628 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1635 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1637 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1639 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1640 // stored to the VarArgsFrameIndex on the stack.
1642 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1643 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1644 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1645 MemOps.push_back(Store);
1646 // Increment the address by four for the next argument to store
1647 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1648 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1652 // If this function is vararg, store any remaining integer argument regs
1653 // to their spots on the stack so that they may be loaded by deferencing the
1654 // result of va_next.
1655 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1658 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1660 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1662 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1663 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1664 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1665 MemOps.push_back(Store);
1666 // Increment the address by four for the next argument to store
1667 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1668 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1671 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1674 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1675 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1676 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1677 MemOps.push_back(Store);
1678 // Increment the address by eight for the next argument to store
1679 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1681 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1684 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1686 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1688 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1689 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1690 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1691 MemOps.push_back(Store);
1692 // Increment the address by eight for the next argument to store
1693 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1695 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1700 if (!MemOps.empty())
1701 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1703 ArgValues.push_back(Root);
1705 // Return the new list of results.
1706 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1707 Op.Val->value_end());
1708 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1711 /// isCallCompatibleAddress - Return the immediate to use if the specified
1712 /// 32-bit value is representable in the immediate field of a BxA instruction.
1713 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1714 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1717 int Addr = C->getValue();
1718 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1719 (Addr << 6 >> 6) != Addr)
1720 return 0; // Top 6 bits have to be sext of immediate.
1722 return DAG.getConstant((int)C->getValue() >> 2,
1723 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1726 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1727 /// by "Src" to address "Dst" of size "Size". Alignment information is
1728 /// specified by the specific parameter attribute. The copy will be passed as
1729 /// a byval function parameter.
1730 /// Sometimes what we are copying is the end of a larger object, the part that
1731 /// does not fit in registers.
1733 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1734 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1736 SDOperand AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);
1737 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1738 SDOperand AlwaysInline = DAG.getConstant(0, MVT::i32);
1739 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1742 SDOperand PPCTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG,
1743 const PPCSubtarget &Subtarget,
1744 TargetMachine &TM) {
1745 SDOperand Chain = Op.getOperand(0);
1746 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1747 SDOperand Callee = Op.getOperand(4);
1748 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1750 bool isMachoABI = Subtarget.isMachoABI();
1751 bool isELF32_ABI = Subtarget.isELF32_ABI();
1753 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1754 bool isPPC64 = PtrVT == MVT::i64;
1755 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1757 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1758 // SelectExpr to use to put the arguments in the appropriate registers.
1759 std::vector<SDOperand> args_to_use;
1761 // Count how many bytes are to be pushed on the stack, including the linkage
1762 // area, and parameter passing area. We start with 24/48 bytes, which is
1763 // prereserved space for [SP][CR][LR][3 x unused].
1764 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1766 // Add up all the space actually used.
1767 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1768 // they all go in registers, but we must reserve stack space for them for
1769 // possible use by the caller. In varargs or 64-bit calls, parameters are
1770 // assigned stack space in order, with padding so Altivec parameters are
1772 unsigned nAltivecParamsAtEnd = 0;
1773 for (unsigned i = 0; i != NumOps; ++i) {
1774 SDOperand Arg = Op.getOperand(5+2*i);
1775 MVT::ValueType ArgVT = Arg.getValueType();
1776 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1777 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1778 if (!isVarArg && !isPPC64) {
1779 // Non-varargs Altivec parameters go after all the non-Altivec parameters;
1780 // do those last so we know how much padding we need.
1781 nAltivecParamsAtEnd++;
1784 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1785 NumBytes = ((NumBytes+15)/16)*16;
1788 ISD::ArgFlagsTy Flags =
1789 cast<ARG_FLAGSSDNode>(Op.getOperand(5+2*i+1))->getArgFlags();
1790 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1791 if (Flags.isByVal())
1792 ArgSize = Flags.getByValSize();
1793 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1794 NumBytes += ArgSize;
1796 // Allow for Altivec parameters at the end, if needed.
1797 if (nAltivecParamsAtEnd) {
1798 NumBytes = ((NumBytes+15)/16)*16;
1799 NumBytes += 16*nAltivecParamsAtEnd;
1802 // The prolog code of the callee may store up to 8 GPR argument registers to
1803 // the stack, allowing va_start to index over them in memory if its varargs.
1804 // Because we cannot tell if this is needed on the caller side, we have to
1805 // conservatively assume that it is needed. As such, make sure we have at
1806 // least enough stack space for the caller to store the 8 GPRs.
1807 NumBytes = std::max(NumBytes,
1808 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1810 // Adjust the stack pointer for the new arguments...
1811 // These operations are automatically eliminated by the prolog/epilog pass
1812 Chain = DAG.getCALLSEQ_START(Chain,
1813 DAG.getConstant(NumBytes, PtrVT));
1814 SDOperand CallSeqStart = Chain;
1816 // Set up a copy of the stack pointer for use loading and storing any
1817 // arguments that may not fit in the registers available for argument
1821 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1823 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1825 // Figure out which arguments are going to go in registers, and which in
1826 // memory. Also, if this is a vararg function, floating point operations
1827 // must be stored to our stack, and loaded into integer regs as well, if
1828 // any integer regs are available for argument passing.
1829 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1830 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1832 static const unsigned GPR_32[] = { // 32-bit registers.
1833 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1834 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1836 static const unsigned GPR_64[] = { // 64-bit registers.
1837 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1838 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1840 static const unsigned *FPR = GetFPR(Subtarget);
1842 static const unsigned VR[] = {
1843 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1844 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1846 const unsigned NumGPRs = array_lengthof(GPR_32);
1847 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1848 const unsigned NumVRs = array_lengthof( VR);
1850 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1852 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1853 SmallVector<SDOperand, 8> MemOpChains;
1854 for (unsigned i = 0; i != NumOps; ++i) {
1856 SDOperand Arg = Op.getOperand(5+2*i);
1857 ISD::ArgFlagsTy Flags =
1858 cast<ARG_FLAGSSDNode>(Op.getOperand(5+2*i+1))->getArgFlags();
1859 // See if next argument requires stack alignment in ELF
1860 bool Expand = false; // TODO: implement this.
1862 // PtrOff will be used to store the current argument to the stack if a
1863 // register cannot be found for it.
1866 // Stack align in ELF 32
1867 if (isELF32_ABI && Expand)
1868 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1869 StackPtr.getValueType());
1871 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1873 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1875 // On PPC64, promote integers to 64-bit values.
1876 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1877 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
1878 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1879 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1882 // FIXME Elf untested, what are alignment rules?
1883 // FIXME memcpy is used way more than necessary. Correctness first.
1884 if (Flags.isByVal()) {
1885 unsigned Size = Flags.getByValSize();
1886 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1887 if (Size==1 || Size==2) {
1888 // Very small objects are passed right-justified.
1889 // Everything else is passed left-justified.
1890 MVT::ValueType VT = (Size==1) ? MVT::i8 : MVT::i16;
1891 if (GPR_idx != NumGPRs) {
1892 SDOperand Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
1894 MemOpChains.push_back(Load.getValue(1));
1895 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1897 ArgOffset += PtrByteSize;
1899 SDOperand Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
1900 SDOperand AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
1901 SDOperand MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
1902 CallSeqStart.Val->getOperand(0),
1904 // This must go outside the CALLSEQ_START..END.
1905 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1906 CallSeqStart.Val->getOperand(1));
1907 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1908 Chain = CallSeqStart = NewCallSeqStart;
1909 ArgOffset += PtrByteSize;
1913 // Copy entire object into memory. There are cases where gcc-generated
1914 // code assumes it is there, even if it could be put entirely into
1915 // registers. (This is not what the doc says.)
1916 SDOperand MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
1917 CallSeqStart.Val->getOperand(0),
1919 // This must go outside the CALLSEQ_START..END.
1920 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1921 CallSeqStart.Val->getOperand(1));
1922 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1923 Chain = CallSeqStart = NewCallSeqStart;
1924 // And copy the pieces of it that fit into registers.
1925 for (unsigned j=0; j<Size; j+=PtrByteSize) {
1926 SDOperand Const = DAG.getConstant(j, PtrOff.getValueType());
1927 SDOperand AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
1928 if (GPR_idx != NumGPRs) {
1929 SDOperand Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
1930 MemOpChains.push_back(Load.getValue(1));
1931 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1933 ArgOffset += PtrByteSize;
1935 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
1942 switch (Arg.getValueType()) {
1943 default: assert(0 && "Unexpected ValueType for argument!");
1946 // Double word align in ELF
1947 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1948 if (GPR_idx != NumGPRs) {
1949 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1951 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1954 if (inMem || isMachoABI) {
1955 // Stack align in ELF
1956 if (isELF32_ABI && Expand)
1957 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1959 ArgOffset += PtrByteSize;
1964 if (FPR_idx != NumFPRs) {
1965 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1968 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1969 MemOpChains.push_back(Store);
1971 // Float varargs are always shadowed in available integer registers
1972 if (GPR_idx != NumGPRs) {
1973 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1974 MemOpChains.push_back(Load.getValue(1));
1975 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1978 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1979 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1980 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1981 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1982 MemOpChains.push_back(Load.getValue(1));
1983 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1987 // If we have any FPRs remaining, we may also have GPRs remaining.
1988 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1991 if (GPR_idx != NumGPRs)
1993 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1994 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1999 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
2002 if (inMem || isMachoABI) {
2003 // Stack align in ELF
2004 if (isELF32_ABI && Expand)
2005 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2009 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2017 // These go aligned on the stack, or in the corresponding R registers
2018 // when within range. The Darwin PPC ABI doc claims they also go in
2019 // V registers; in fact gcc does this only for arguments that are
2020 // prototyped, not for those that match the ... We do it for all
2021 // arguments, seems to work.
2022 while (ArgOffset % 16 !=0) {
2023 ArgOffset += PtrByteSize;
2024 if (GPR_idx != NumGPRs)
2027 // We could elide this store in the case where the object fits
2028 // entirely in R registers. Maybe later.
2029 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2030 DAG.getConstant(ArgOffset, PtrVT));
2031 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2032 MemOpChains.push_back(Store);
2033 if (VR_idx != NumVRs) {
2034 SDOperand Load = DAG.getLoad(MVT::v4f32, Store, PtrOff, NULL, 0);
2035 MemOpChains.push_back(Load.getValue(1));
2036 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2039 for (unsigned i=0; i<16; i+=PtrByteSize) {
2040 if (GPR_idx == NumGPRs)
2042 SDOperand Ix = DAG.getNode(ISD::ADD, PtrVT, PtrOff,
2043 DAG.getConstant(i, PtrVT));
2044 SDOperand Load = DAG.getLoad(PtrVT, Store, Ix, NULL, 0);
2045 MemOpChains.push_back(Load.getValue(1));
2046 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2050 // Non-varargs Altivec params generally go in registers, but have
2051 // stack space allocated at the end.
2052 if (VR_idx != NumVRs) {
2053 // Doesn't have GPR space allocated.
2054 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2055 } else if (nAltivecParamsAtEnd==0) {
2056 // We are emitting Altivec params in order.
2057 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2058 DAG.getConstant(ArgOffset, PtrVT));
2059 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2060 MemOpChains.push_back(Store);
2066 // If all Altivec parameters fit in registers, as they usually do,
2067 // they get stack space following the non-Altivec parameters. We
2068 // don't track this here because nobody below needs it.
2069 // If there are more Altivec parameters than fit in registers emit
2071 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2073 // Offset is aligned; skip 1st 12 params which go in V registers.
2074 ArgOffset = ((ArgOffset+15)/16)*16;
2076 for (unsigned i = 0; i != NumOps; ++i) {
2077 SDOperand Arg = Op.getOperand(5+2*i);
2078 MVT::ValueType ArgType = Arg.getValueType();
2079 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2080 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2082 SDOperand PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2083 DAG.getConstant(ArgOffset, PtrVT));
2084 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2085 MemOpChains.push_back(Store);
2092 if (!MemOpChains.empty())
2093 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2094 &MemOpChains[0], MemOpChains.size());
2096 // Build a sequence of copy-to-reg nodes chained together with token chain
2097 // and flag operands which copy the outgoing args into the appropriate regs.
2099 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2100 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
2102 InFlag = Chain.getValue(1);
2105 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2106 if (isVarArg && isELF32_ABI) {
2107 SDOperand SetCR(DAG.getTargetNode(PPC::CRSET, MVT::i32), 0);
2108 Chain = DAG.getCopyToReg(Chain, PPC::CR1EQ, SetCR, InFlag);
2109 InFlag = Chain.getValue(1);
2112 std::vector<MVT::ValueType> NodeTys;
2113 NodeTys.push_back(MVT::Other); // Returns a chain
2114 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2116 SmallVector<SDOperand, 8> Ops;
2117 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2119 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2120 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2121 // node so that legalize doesn't hack it.
2122 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2123 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2124 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2125 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2126 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2127 // If this is an absolute destination address, use the munged value.
2128 Callee = SDOperand(Dest, 0);
2130 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2131 // to do the call, we can't use PPCISD::CALL.
2132 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
2133 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
2134 InFlag = Chain.getValue(1);
2136 // Copy the callee address into R12/X12 on darwin.
2138 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2139 Chain = DAG.getCopyToReg(Chain, Reg, Callee, InFlag);
2140 InFlag = Chain.getValue(1);
2144 NodeTys.push_back(MVT::Other);
2145 NodeTys.push_back(MVT::Flag);
2146 Ops.push_back(Chain);
2147 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2151 // If this is a direct call, pass the chain and the callee.
2153 Ops.push_back(Chain);
2154 Ops.push_back(Callee);
2157 // Add argument registers to the end of the list so that they are known live
2159 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2160 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2161 RegsToPass[i].second.getValueType()));
2164 Ops.push_back(InFlag);
2165 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
2166 InFlag = Chain.getValue(1);
2168 Chain = DAG.getCALLSEQ_END(Chain,
2169 DAG.getConstant(NumBytes, PtrVT),
2170 DAG.getConstant(0, PtrVT),
2172 if (Op.Val->getValueType(0) != MVT::Other)
2173 InFlag = Chain.getValue(1);
2175 SmallVector<SDOperand, 16> ResultVals;
2176 SmallVector<CCValAssign, 16> RVLocs;
2177 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2178 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2179 CCInfo.AnalyzeCallResult(Op.Val, RetCC_PPC);
2181 // Copy all of the result registers out of their specified physreg.
2182 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2183 CCValAssign &VA = RVLocs[i];
2184 MVT::ValueType VT = VA.getValVT();
2185 assert(VA.isRegLoc() && "Can only return in registers!");
2186 Chain = DAG.getCopyFromReg(Chain, VA.getLocReg(), VT, InFlag).getValue(1);
2187 ResultVals.push_back(Chain.getValue(0));
2188 InFlag = Chain.getValue(2);
2191 // If the function returns void, just return the chain.
2195 // Otherwise, merge everything together with a MERGE_VALUES node.
2196 ResultVals.push_back(Chain);
2197 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, Op.Val->getVTList(),
2198 &ResultVals[0], ResultVals.size());
2199 return Res.getValue(Op.ResNo);
2202 SDOperand PPCTargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG,
2203 TargetMachine &TM) {
2204 SmallVector<CCValAssign, 16> RVLocs;
2205 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2206 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2207 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2208 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
2210 // If this is the first return lowered for this function, add the regs to the
2211 // liveout set for the function.
2212 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2213 for (unsigned i = 0; i != RVLocs.size(); ++i)
2214 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2217 SDOperand Chain = Op.getOperand(0);
2220 // Copy the result values into the output registers.
2221 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2222 CCValAssign &VA = RVLocs[i];
2223 assert(VA.isRegLoc() && "Can only return in registers!");
2224 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2225 Flag = Chain.getValue(1);
2229 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2231 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2234 SDOperand PPCTargetLowering::LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
2235 const PPCSubtarget &Subtarget) {
2236 // When we pop the dynamic allocation we need to restore the SP link.
2238 // Get the corect type for pointers.
2239 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2241 // Construct the stack pointer operand.
2242 bool IsPPC64 = Subtarget.isPPC64();
2243 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2244 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
2246 // Get the operands for the STACKRESTORE.
2247 SDOperand Chain = Op.getOperand(0);
2248 SDOperand SaveSP = Op.getOperand(1);
2250 // Load the old link SP.
2251 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2253 // Restore the stack pointer.
2254 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2256 // Store the old link SP.
2257 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2260 SDOperand PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
2262 const PPCSubtarget &Subtarget) {
2263 MachineFunction &MF = DAG.getMachineFunction();
2264 bool IsPPC64 = Subtarget.isPPC64();
2265 bool isMachoABI = Subtarget.isMachoABI();
2267 // Get current frame pointer save index. The users of this index will be
2268 // primarily DYNALLOC instructions.
2269 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2270 int FPSI = FI->getFramePointerSaveIndex();
2272 // If the frame pointer save index hasn't been defined yet.
2274 // Find out what the fix offset of the frame pointer save area.
2275 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2277 // Allocate the frame index for frame pointer save area.
2278 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2280 FI->setFramePointerSaveIndex(FPSI);
2284 SDOperand Chain = Op.getOperand(0);
2285 SDOperand Size = Op.getOperand(1);
2287 // Get the corect type for pointers.
2288 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2290 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
2291 DAG.getConstant(0, PtrVT), Size);
2292 // Construct a node for the frame pointer save index.
2293 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
2294 // Build a DYNALLOC node.
2295 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
2296 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2297 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2301 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2303 SDOperand PPCTargetLowering::LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2304 // Not FP? Not a fsel.
2305 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2306 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2309 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2311 // Cannot handle SETEQ/SETNE.
2312 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2314 MVT::ValueType ResVT = Op.getValueType();
2315 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2316 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2317 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2319 // If the RHS of the comparison is a 0.0, we don't need to do the
2320 // subtraction at all.
2321 if (isFloatingPointZero(RHS))
2323 default: break; // SETUO etc aren't handled by fsel.
2327 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2331 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2332 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2333 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2337 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2341 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2342 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2343 return DAG.getNode(PPCISD::FSEL, ResVT,
2344 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2349 default: break; // SETUO etc aren't handled by fsel.
2353 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2354 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2355 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2356 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2360 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2361 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2362 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2363 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2367 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2368 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2369 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2370 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2374 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2375 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2376 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2377 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2382 // FIXME: Split this code up when LegalizeDAGTypes lands.
2383 SDOperand PPCTargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2384 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2385 SDOperand Src = Op.getOperand(0);
2386 if (Src.getValueType() == MVT::f32)
2387 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2390 switch (Op.getValueType()) {
2391 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2393 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2396 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2400 // Convert the FP value to an int value through memory.
2401 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2403 // Emit a store to the stack slot.
2404 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2406 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2408 if (Op.getValueType() == MVT::i32)
2409 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2410 DAG.getConstant(4, FIPtr.getValueType()));
2411 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2414 SDOperand PPCTargetLowering::LowerFP_ROUND_INREG(SDOperand Op,
2415 SelectionDAG &DAG) {
2416 assert(Op.getValueType() == MVT::ppcf128);
2417 SDNode *Node = Op.Val;
2418 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2419 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2420 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2421 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2423 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2424 // of the long double, and puts FPSCR back the way it was. We do not
2425 // actually model FPSCR.
2426 std::vector<MVT::ValueType> NodeTys;
2427 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2429 NodeTys.push_back(MVT::f64); // Return register
2430 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2431 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2432 MFFSreg = Result.getValue(0);
2433 InFlag = Result.getValue(1);
2436 NodeTys.push_back(MVT::Flag); // Returns a flag
2437 Ops[0] = DAG.getConstant(31, MVT::i32);
2439 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2440 InFlag = Result.getValue(0);
2443 NodeTys.push_back(MVT::Flag); // Returns a flag
2444 Ops[0] = DAG.getConstant(30, MVT::i32);
2446 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2447 InFlag = Result.getValue(0);
2450 NodeTys.push_back(MVT::f64); // result of add
2451 NodeTys.push_back(MVT::Flag); // Returns a flag
2455 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2456 FPreg = Result.getValue(0);
2457 InFlag = Result.getValue(1);
2460 NodeTys.push_back(MVT::f64);
2461 Ops[0] = DAG.getConstant(1, MVT::i32);
2465 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2466 FPreg = Result.getValue(0);
2468 // We know the low half is about to be thrown away, so just use something
2470 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2473 SDOperand PPCTargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2474 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2475 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2478 if (Op.getOperand(0).getValueType() == MVT::i64) {
2479 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2480 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2481 if (Op.getValueType() == MVT::f32)
2482 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2486 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2487 "Unhandled SINT_TO_FP type in custom expander!");
2488 // Since we only generate this in 64-bit mode, we can take advantage of
2489 // 64-bit registers. In particular, sign extend the input value into the
2490 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2491 // then lfd it and fcfid it.
2492 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2493 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2494 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2495 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2497 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2500 // STD the extended value into the stack slot.
2501 MemOperand MO(PseudoSourceValue::getFixedStack(),
2502 MemOperand::MOStore, FrameIdx, 8, 8);
2503 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2504 DAG.getEntryNode(), Ext64, FIdx,
2505 DAG.getMemOperand(MO));
2506 // Load the value as a double.
2507 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2509 // FCFID it and return it.
2510 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2511 if (Op.getValueType() == MVT::f32)
2512 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2516 SDOperand PPCTargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2518 The rounding mode is in bits 30:31 of FPSR, and has the following
2525 FLT_ROUNDS, on the other hand, expects the following:
2532 To perform the conversion, we do:
2533 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2536 MachineFunction &MF = DAG.getMachineFunction();
2537 MVT::ValueType VT = Op.getValueType();
2538 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2539 std::vector<MVT::ValueType> NodeTys;
2540 SDOperand MFFSreg, InFlag;
2542 // Save FP Control Word to register
2543 NodeTys.push_back(MVT::f64); // return register
2544 NodeTys.push_back(MVT::Flag); // unused in this context
2545 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2547 // Save FP register to stack slot
2548 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2549 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2550 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2551 StackSlot, NULL, 0);
2553 // Load FP Control Word from low 32 bits of stack slot.
2554 SDOperand Four = DAG.getConstant(4, PtrVT);
2555 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2556 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2558 // Transform as necessary
2560 DAG.getNode(ISD::AND, MVT::i32,
2561 CWD, DAG.getConstant(3, MVT::i32));
2563 DAG.getNode(ISD::SRL, MVT::i32,
2564 DAG.getNode(ISD::AND, MVT::i32,
2565 DAG.getNode(ISD::XOR, MVT::i32,
2566 CWD, DAG.getConstant(3, MVT::i32)),
2567 DAG.getConstant(3, MVT::i32)),
2568 DAG.getConstant(1, MVT::i8));
2571 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2573 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2574 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2577 SDOperand PPCTargetLowering::LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2578 MVT::ValueType VT = Op.getValueType();
2579 unsigned BitWidth = MVT::getSizeInBits(VT);
2580 assert(Op.getNumOperands() == 3 &&
2581 VT == Op.getOperand(1).getValueType() &&
2584 // Expand into a bunch of logical ops. Note that these ops
2585 // depend on the PPC behavior for oversized shift amounts.
2586 SDOperand Lo = Op.getOperand(0);
2587 SDOperand Hi = Op.getOperand(1);
2588 SDOperand Amt = Op.getOperand(2);
2589 MVT::ValueType AmtVT = Amt.getValueType();
2591 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2592 DAG.getConstant(BitWidth, AmtVT), Amt);
2593 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
2594 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
2595 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2596 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2597 DAG.getConstant(-BitWidth, AmtVT));
2598 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
2599 SDOperand OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2600 SDOperand OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
2601 SDOperand OutOps[] = { OutLo, OutHi };
2602 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2606 SDOperand PPCTargetLowering::LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2607 MVT::ValueType VT = Op.getValueType();
2608 unsigned BitWidth = MVT::getSizeInBits(VT);
2609 assert(Op.getNumOperands() == 3 &&
2610 VT == Op.getOperand(1).getValueType() &&
2613 // Expand into a bunch of logical ops. Note that these ops
2614 // depend on the PPC behavior for oversized shift amounts.
2615 SDOperand Lo = Op.getOperand(0);
2616 SDOperand Hi = Op.getOperand(1);
2617 SDOperand Amt = Op.getOperand(2);
2618 MVT::ValueType AmtVT = Amt.getValueType();
2620 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2621 DAG.getConstant(BitWidth, AmtVT), Amt);
2622 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2623 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2624 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2625 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2626 DAG.getConstant(-BitWidth, AmtVT));
2627 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
2628 SDOperand OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2629 SDOperand OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
2630 SDOperand OutOps[] = { OutLo, OutHi };
2631 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2635 SDOperand PPCTargetLowering::LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2636 MVT::ValueType VT = Op.getValueType();
2637 unsigned BitWidth = MVT::getSizeInBits(VT);
2638 assert(Op.getNumOperands() == 3 &&
2639 VT == Op.getOperand(1).getValueType() &&
2642 // Expand into a bunch of logical ops, followed by a select_cc.
2643 SDOperand Lo = Op.getOperand(0);
2644 SDOperand Hi = Op.getOperand(1);
2645 SDOperand Amt = Op.getOperand(2);
2646 MVT::ValueType AmtVT = Amt.getValueType();
2648 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2649 DAG.getConstant(BitWidth, AmtVT), Amt);
2650 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2651 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2652 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2653 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2654 DAG.getConstant(-BitWidth, AmtVT));
2655 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
2656 SDOperand OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
2657 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
2658 Tmp4, Tmp6, ISD::SETLE);
2659 SDOperand OutOps[] = { OutLo, OutHi };
2660 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2664 //===----------------------------------------------------------------------===//
2665 // Vector related lowering.
2668 // If this is a vector of constants or undefs, get the bits. A bit in
2669 // UndefBits is set if the corresponding element of the vector is an
2670 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2671 // zero. Return true if this is not an array of constants, false if it is.
2673 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2674 uint64_t UndefBits[2]) {
2675 // Start with zero'd results.
2676 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2678 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2679 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2680 SDOperand OpVal = BV->getOperand(i);
2682 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2683 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2685 uint64_t EltBits = 0;
2686 if (OpVal.getOpcode() == ISD::UNDEF) {
2687 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2688 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2690 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2691 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2692 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2693 assert(CN->getValueType(0) == MVT::f32 &&
2694 "Only one legal FP vector type!");
2695 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2697 // Nonconstant element.
2701 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2704 //printf("%llx %llx %llx %llx\n",
2705 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2709 // If this is a splat (repetition) of a value across the whole vector, return
2710 // the smallest size that splats it. For example, "0x01010101010101..." is a
2711 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2712 // SplatSize = 1 byte.
2713 static bool isConstantSplat(const uint64_t Bits128[2],
2714 const uint64_t Undef128[2],
2715 unsigned &SplatBits, unsigned &SplatUndef,
2716 unsigned &SplatSize) {
2718 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2719 // the same as the lower 64-bits, ignoring undefs.
2720 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2721 return false; // Can't be a splat if two pieces don't match.
2723 uint64_t Bits64 = Bits128[0] | Bits128[1];
2724 uint64_t Undef64 = Undef128[0] & Undef128[1];
2726 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2728 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2729 return false; // Can't be a splat if two pieces don't match.
2731 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2732 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2734 // If the top 16-bits are different than the lower 16-bits, ignoring
2735 // undefs, we have an i32 splat.
2736 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2738 SplatUndef = Undef32;
2743 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2744 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2746 // If the top 8-bits are different than the lower 8-bits, ignoring
2747 // undefs, we have an i16 splat.
2748 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2750 SplatUndef = Undef16;
2755 // Otherwise, we have an 8-bit splat.
2756 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2757 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2762 /// BuildSplatI - Build a canonical splati of Val with an element size of
2763 /// SplatSize. Cast the result to VT.
2764 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2765 SelectionDAG &DAG) {
2766 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2768 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2769 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2772 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2774 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2778 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2780 // Build a canonical splat for this value.
2781 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2782 SmallVector<SDOperand, 8> Ops;
2783 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2784 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2785 &Ops[0], Ops.size());
2786 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2789 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2790 /// specified intrinsic ID.
2791 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2793 MVT::ValueType DestVT = MVT::Other) {
2794 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2795 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2796 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2799 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2800 /// specified intrinsic ID.
2801 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2802 SDOperand Op2, SelectionDAG &DAG,
2803 MVT::ValueType DestVT = MVT::Other) {
2804 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2805 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2806 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2810 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2811 /// amount. The result has the specified value type.
2812 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2813 MVT::ValueType VT, SelectionDAG &DAG) {
2814 // Force LHS/RHS to be the right type.
2815 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2816 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2819 for (unsigned i = 0; i != 16; ++i)
2820 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2821 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2822 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2823 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2826 // If this is a case we can't handle, return null and let the default
2827 // expansion code take care of it. If we CAN select this case, and if it
2828 // selects to a single instruction, return Op. Otherwise, if we can codegen
2829 // this case more efficiently than a constant pool load, lower it to the
2830 // sequence of ops that should be used.
2831 SDOperand PPCTargetLowering::LowerBUILD_VECTOR(SDOperand Op,
2832 SelectionDAG &DAG) {
2833 // If this is a vector of constants or undefs, get the bits. A bit in
2834 // UndefBits is set if the corresponding element of the vector is an
2835 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2837 uint64_t VectorBits[2];
2838 uint64_t UndefBits[2];
2839 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2840 return SDOperand(); // Not a constant vector.
2842 // If this is a splat (repetition) of a value across the whole vector, return
2843 // the smallest size that splats it. For example, "0x01010101010101..." is a
2844 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2845 // SplatSize = 1 byte.
2846 unsigned SplatBits, SplatUndef, SplatSize;
2847 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2848 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2850 // First, handle single instruction cases.
2853 if (SplatBits == 0) {
2854 // Canonicalize all zero vectors to be v4i32.
2855 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2856 SDOperand Z = DAG.getConstant(0, MVT::i32);
2857 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2858 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2863 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2864 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2865 if (SextVal >= -16 && SextVal <= 15)
2866 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2869 // Two instruction sequences.
2871 // If this value is in the range [-32,30] and is even, use:
2872 // tmp = VSPLTI[bhw], result = add tmp, tmp
2873 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2874 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2875 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2878 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2879 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2881 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2882 // Make -1 and vspltisw -1:
2883 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2885 // Make the VSLW intrinsic, computing 0x8000_0000.
2886 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2889 // xor by OnesV to invert it.
2890 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2891 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2894 // Check to see if this is a wide variety of vsplti*, binop self cases.
2895 unsigned SplatBitSize = SplatSize*8;
2896 static const signed char SplatCsts[] = {
2897 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2898 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2901 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2902 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2903 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2904 int i = SplatCsts[idx];
2906 // Figure out what shift amount will be used by altivec if shifted by i in
2908 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2910 // vsplti + shl self.
2911 if (SextVal == (i << (int)TypeShiftAmt)) {
2912 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2913 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2914 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2915 Intrinsic::ppc_altivec_vslw
2917 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2918 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2921 // vsplti + srl self.
2922 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2923 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2924 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2925 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2926 Intrinsic::ppc_altivec_vsrw
2928 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2929 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2932 // vsplti + sra self.
2933 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2934 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2935 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2936 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2937 Intrinsic::ppc_altivec_vsraw
2939 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2940 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2943 // vsplti + rol self.
2944 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2945 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2946 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2947 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2948 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2949 Intrinsic::ppc_altivec_vrlw
2951 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2952 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2955 // t = vsplti c, result = vsldoi t, t, 1
2956 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2957 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2958 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2960 // t = vsplti c, result = vsldoi t, t, 2
2961 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2962 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2963 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2965 // t = vsplti c, result = vsldoi t, t, 3
2966 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2967 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2968 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2972 // Three instruction sequences.
2974 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2975 if (SextVal >= 0 && SextVal <= 31) {
2976 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2977 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2978 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2979 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2981 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2982 if (SextVal >= -31 && SextVal <= 0) {
2983 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2984 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2985 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2986 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2993 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2994 /// the specified operations to build the shuffle.
2995 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2996 SDOperand RHS, SelectionDAG &DAG) {
2997 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2998 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2999 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3002 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3014 if (OpNum == OP_COPY) {
3015 if (LHSID == (1*9+2)*9+3) return LHS;
3016 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3020 SDOperand OpLHS, OpRHS;
3021 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
3022 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
3024 unsigned ShufIdxs[16];
3026 default: assert(0 && "Unknown i32 permute!");
3028 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3029 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3030 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3031 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3034 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3035 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3036 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3037 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3040 for (unsigned i = 0; i != 16; ++i)
3041 ShufIdxs[i] = (i&3)+0;
3044 for (unsigned i = 0; i != 16; ++i)
3045 ShufIdxs[i] = (i&3)+4;
3048 for (unsigned i = 0; i != 16; ++i)
3049 ShufIdxs[i] = (i&3)+8;
3052 for (unsigned i = 0; i != 16; ++i)
3053 ShufIdxs[i] = (i&3)+12;
3056 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
3058 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
3060 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
3063 for (unsigned i = 0; i != 16; ++i)
3064 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
3066 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
3067 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3070 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3071 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3072 /// return the code it can be lowered into. Worst case, it can always be
3073 /// lowered into a vperm.
3074 SDOperand PPCTargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op,
3075 SelectionDAG &DAG) {
3076 SDOperand V1 = Op.getOperand(0);
3077 SDOperand V2 = Op.getOperand(1);
3078 SDOperand PermMask = Op.getOperand(2);
3080 // Cases that are handled by instructions that take permute immediates
3081 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3082 // selected by the instruction selector.
3083 if (V2.getOpcode() == ISD::UNDEF) {
3084 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
3085 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
3086 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
3087 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
3088 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
3089 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
3090 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
3091 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
3092 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
3093 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
3094 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
3095 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
3100 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3101 // and produce a fixed permutation. If any of these match, do not lower to
3103 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
3104 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
3105 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
3106 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
3107 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
3108 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
3109 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
3110 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
3111 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
3114 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3115 // perfect shuffle table to emit an optimal matching sequence.
3116 unsigned PFIndexes[4];
3117 bool isFourElementShuffle = true;
3118 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3119 unsigned EltNo = 8; // Start out undef.
3120 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3121 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
3122 continue; // Undef, ignore it.
3124 unsigned ByteSource =
3125 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
3126 if ((ByteSource & 3) != j) {
3127 isFourElementShuffle = false;
3132 EltNo = ByteSource/4;
3133 } else if (EltNo != ByteSource/4) {
3134 isFourElementShuffle = false;
3138 PFIndexes[i] = EltNo;
3141 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3142 // perfect shuffle vector to determine if it is cost effective to do this as
3143 // discrete instructions, or whether we should use a vperm.
3144 if (isFourElementShuffle) {
3145 // Compute the index in the perfect shuffle table.
3146 unsigned PFTableIndex =
3147 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3149 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3150 unsigned Cost = (PFEntry >> 30);
3152 // Determining when to avoid vperm is tricky. Many things affect the cost
3153 // of vperm, particularly how many times the perm mask needs to be computed.
3154 // For example, if the perm mask can be hoisted out of a loop or is already
3155 // used (perhaps because there are multiple permutes with the same shuffle
3156 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3157 // the loop requires an extra register.
3159 // As a compromise, we only emit discrete instructions if the shuffle can be
3160 // generated in 3 or fewer operations. When we have loop information
3161 // available, if this block is within a loop, we should avoid using vperm
3162 // for 3-operation perms and use a constant pool load instead.
3164 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3167 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3168 // vector that will get spilled to the constant pool.
3169 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3171 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3172 // that it is in input element units, not in bytes. Convert now.
3173 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
3174 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
3176 SmallVector<SDOperand, 16> ResultMask;
3177 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3179 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3182 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
3184 for (unsigned j = 0; j != BytesPerElement; ++j)
3185 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3189 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3190 &ResultMask[0], ResultMask.size());
3191 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3194 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3195 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3196 /// information about the intrinsic.
3197 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
3199 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
3202 switch (IntrinsicID) {
3203 default: return false;
3204 // Comparison predicates.
3205 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3206 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3207 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3208 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3209 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3210 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3211 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3212 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3213 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3214 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3215 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3216 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3217 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3219 // Normal Comparisons.
3220 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3221 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3222 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3223 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3224 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3225 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3226 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3227 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3228 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3229 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3230 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3231 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3232 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3237 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3238 /// lower, do it, otherwise return null.
3239 SDOperand PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op,
3240 SelectionDAG &DAG) {
3241 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3242 // opcode number of the comparison.
3245 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3246 return SDOperand(); // Don't custom lower most intrinsics.
3248 // If this is a non-dot comparison, make the VCMP node and we are done.
3250 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3251 Op.getOperand(1), Op.getOperand(2),
3252 DAG.getConstant(CompareOpc, MVT::i32));
3253 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3256 // Create the PPCISD altivec 'dot' comparison node.
3258 Op.getOperand(2), // LHS
3259 Op.getOperand(3), // RHS
3260 DAG.getConstant(CompareOpc, MVT::i32)
3262 std::vector<MVT::ValueType> VTs;
3263 VTs.push_back(Op.getOperand(2).getValueType());
3264 VTs.push_back(MVT::Flag);
3265 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3267 // Now that we have the comparison, emit a copy from the CR to a GPR.
3268 // This is flagged to the above dot comparison.
3269 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3270 DAG.getRegister(PPC::CR6, MVT::i32),
3271 CompNode.getValue(1));
3273 // Unpack the result based on how the target uses it.
3274 unsigned BitNo; // Bit # of CR6.
3275 bool InvertBit; // Invert result?
3276 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
3277 default: // Can't happen, don't crash on invalid number though.
3278 case 0: // Return the value of the EQ bit of CR6.
3279 BitNo = 0; InvertBit = false;
3281 case 1: // Return the inverted value of the EQ bit of CR6.
3282 BitNo = 0; InvertBit = true;
3284 case 2: // Return the value of the LT bit of CR6.
3285 BitNo = 2; InvertBit = false;
3287 case 3: // Return the inverted value of the LT bit of CR6.
3288 BitNo = 2; InvertBit = true;
3292 // Shift the bit into the low position.
3293 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3294 DAG.getConstant(8-(3-BitNo), MVT::i32));
3296 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3297 DAG.getConstant(1, MVT::i32));
3299 // If we are supposed to, toggle the bit.
3301 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3302 DAG.getConstant(1, MVT::i32));
3306 SDOperand PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op,
3307 SelectionDAG &DAG) {
3308 // Create a stack slot that is 16-byte aligned.
3309 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3310 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3311 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3312 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3314 // Store the input value into Value#0 of the stack slot.
3315 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
3316 Op.getOperand(0), FIdx, NULL, 0);
3318 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3321 SDOperand PPCTargetLowering::LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3322 if (Op.getValueType() == MVT::v4i32) {
3323 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3325 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3326 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3328 SDOperand RHSSwap = // = vrlw RHS, 16
3329 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3331 // Shrinkify inputs to v8i16.
3332 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3333 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3334 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3336 // Low parts multiplied together, generating 32-bit results (we ignore the
3338 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3339 LHS, RHS, DAG, MVT::v4i32);
3341 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3342 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3343 // Shift the high parts up 16 bits.
3344 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3345 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3346 } else if (Op.getValueType() == MVT::v8i16) {
3347 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3349 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3351 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3352 LHS, RHS, Zero, DAG);
3353 } else if (Op.getValueType() == MVT::v16i8) {
3354 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3356 // Multiply the even 8-bit parts, producing 16-bit sums.
3357 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3358 LHS, RHS, DAG, MVT::v8i16);
3359 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3361 // Multiply the odd 8-bit parts, producing 16-bit sums.
3362 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3363 LHS, RHS, DAG, MVT::v8i16);
3364 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3366 // Merge the results together.
3368 for (unsigned i = 0; i != 8; ++i) {
3369 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3370 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3372 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3373 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3375 assert(0 && "Unknown mul to lower!");
3380 /// LowerOperation - Provide custom lowering hooks for some operations.
3382 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3383 switch (Op.getOpcode()) {
3384 default: assert(0 && "Wasn't expecting to be able to lower this!");
3385 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3386 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3387 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3388 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3389 case ISD::SETCC: return LowerSETCC(Op, DAG);
3391 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3392 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3395 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3396 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3398 case ISD::FORMAL_ARGUMENTS:
3399 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3400 VarArgsStackOffset, VarArgsNumGPR,
3401 VarArgsNumFPR, PPCSubTarget);
3403 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3404 getTargetMachine());
3405 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3406 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3407 case ISD::DYNAMIC_STACKALLOC:
3408 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3410 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3411 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3412 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3413 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3414 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3416 // Lower 64-bit shifts.
3417 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3418 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3419 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3421 // Vector-related lowering.
3422 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3423 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3424 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3425 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3426 case ISD::MUL: return LowerMUL(Op, DAG);
3428 // Frame & Return address.
3429 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3430 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3435 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3436 switch (N->getOpcode()) {
3437 default: assert(0 && "Wasn't expecting to be able to lower this!");
3438 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3443 //===----------------------------------------------------------------------===//
3444 // Other Lowering Code
3445 //===----------------------------------------------------------------------===//
3448 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3449 MachineBasicBlock *BB) {
3450 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3451 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3452 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3453 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3454 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3455 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3456 "Unexpected instr type to insert");
3458 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3459 // control-flow pattern. The incoming instruction knows the destination vreg
3460 // to set, the condition code register to branch on, the true/false values to
3461 // select between, and a branch opcode to use.
3462 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3463 ilist<MachineBasicBlock>::iterator It = BB;
3469 // cmpTY ccX, r1, r2
3471 // fallthrough --> copy0MBB
3472 MachineBasicBlock *thisMBB = BB;
3473 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3474 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3475 unsigned SelectPred = MI->getOperand(4).getImm();
3476 BuildMI(BB, TII->get(PPC::BCC))
3477 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3478 MachineFunction *F = BB->getParent();
3479 F->getBasicBlockList().insert(It, copy0MBB);
3480 F->getBasicBlockList().insert(It, sinkMBB);
3481 // Update machine-CFG edges by first adding all successors of the current
3482 // block to the new block which will contain the Phi node for the select.
3483 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3484 e = BB->succ_end(); i != e; ++i)
3485 sinkMBB->addSuccessor(*i);
3486 // Next, remove all successors of the current block, and add the true
3487 // and fallthrough blocks as its successors.
3488 while(!BB->succ_empty())
3489 BB->removeSuccessor(BB->succ_begin());
3490 BB->addSuccessor(copy0MBB);
3491 BB->addSuccessor(sinkMBB);
3494 // %FalseValue = ...
3495 // # fallthrough to sinkMBB
3498 // Update machine-CFG edges
3499 BB->addSuccessor(sinkMBB);
3502 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3505 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3506 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3507 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3509 delete MI; // The pseudo instruction is gone now.
3513 //===----------------------------------------------------------------------===//
3514 // Target Optimization Hooks
3515 //===----------------------------------------------------------------------===//
3517 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3518 DAGCombinerInfo &DCI) const {
3519 TargetMachine &TM = getTargetMachine();
3520 SelectionDAG &DAG = DCI.DAG;
3521 switch (N->getOpcode()) {
3524 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3525 if (C->getValue() == 0) // 0 << V -> 0.
3526 return N->getOperand(0);
3530 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3531 if (C->getValue() == 0) // 0 >>u V -> 0.
3532 return N->getOperand(0);
3536 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3537 if (C->getValue() == 0 || // 0 >>s V -> 0.
3538 C->isAllOnesValue()) // -1 >>s V -> -1.
3539 return N->getOperand(0);
3543 case ISD::SINT_TO_FP:
3544 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3545 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3546 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3547 // We allow the src/dst to be either f32/f64, but the intermediate
3548 // type must be i64.
3549 if (N->getOperand(0).getValueType() == MVT::i64 &&
3550 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3551 SDOperand Val = N->getOperand(0).getOperand(0);
3552 if (Val.getValueType() == MVT::f32) {
3553 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3554 DCI.AddToWorklist(Val.Val);
3557 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3558 DCI.AddToWorklist(Val.Val);
3559 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3560 DCI.AddToWorklist(Val.Val);
3561 if (N->getValueType(0) == MVT::f32) {
3562 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3563 DAG.getIntPtrConstant(0));
3564 DCI.AddToWorklist(Val.Val);
3567 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3568 // If the intermediate type is i32, we can avoid the load/store here
3575 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3576 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3577 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3578 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3579 N->getOperand(1).getValueType() == MVT::i32 &&
3580 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3581 SDOperand Val = N->getOperand(1).getOperand(0);
3582 if (Val.getValueType() == MVT::f32) {
3583 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3584 DCI.AddToWorklist(Val.Val);
3586 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3587 DCI.AddToWorklist(Val.Val);
3589 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3590 N->getOperand(2), N->getOperand(3));
3591 DCI.AddToWorklist(Val.Val);
3595 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3596 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3597 N->getOperand(1).Val->hasOneUse() &&
3598 (N->getOperand(1).getValueType() == MVT::i32 ||
3599 N->getOperand(1).getValueType() == MVT::i16)) {
3600 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3601 // Do an any-extend to 32-bits if this is a half-word input.
3602 if (BSwapOp.getValueType() == MVT::i16)
3603 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3605 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3606 N->getOperand(2), N->getOperand(3),
3607 DAG.getValueType(N->getOperand(1).getValueType()));
3611 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3612 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3613 N->getOperand(0).hasOneUse() &&
3614 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3615 SDOperand Load = N->getOperand(0);
3616 LoadSDNode *LD = cast<LoadSDNode>(Load);
3617 // Create the byte-swapping load.
3618 std::vector<MVT::ValueType> VTs;
3619 VTs.push_back(MVT::i32);
3620 VTs.push_back(MVT::Other);
3621 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3623 LD->getChain(), // Chain
3624 LD->getBasePtr(), // Ptr
3626 DAG.getValueType(N->getValueType(0)) // VT
3628 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3630 // If this is an i16 load, insert the truncate.
3631 SDOperand ResVal = BSLoad;
3632 if (N->getValueType(0) == MVT::i16)
3633 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3635 // First, combine the bswap away. This makes the value produced by the
3637 DCI.CombineTo(N, ResVal);
3639 // Next, combine the load away, we give it a bogus result value but a real
3640 // chain result. The result value is dead because the bswap is dead.
3641 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3643 // Return N so it doesn't get rechecked!
3644 return SDOperand(N, 0);
3648 case PPCISD::VCMP: {
3649 // If a VCMPo node already exists with exactly the same operands as this
3650 // node, use its result instead of this node (VCMPo computes both a CR6 and
3651 // a normal output).
3653 if (!N->getOperand(0).hasOneUse() &&
3654 !N->getOperand(1).hasOneUse() &&
3655 !N->getOperand(2).hasOneUse()) {
3657 // Scan all of the users of the LHS, looking for VCMPo's that match.
3658 SDNode *VCMPoNode = 0;
3660 SDNode *LHSN = N->getOperand(0).Val;
3661 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3663 if ((*UI).getUser()->getOpcode() == PPCISD::VCMPo &&
3664 (*UI).getUser()->getOperand(1) == N->getOperand(1) &&
3665 (*UI).getUser()->getOperand(2) == N->getOperand(2) &&
3666 (*UI).getUser()->getOperand(0) == N->getOperand(0)) {
3667 VCMPoNode = UI->getUser();
3671 // If there is no VCMPo node, or if the flag value has a single use, don't
3673 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3676 // Look at the (necessarily single) use of the flag value. If it has a
3677 // chain, this transformation is more complex. Note that multiple things
3678 // could use the value result, which we should ignore.
3679 SDNode *FlagUser = 0;
3680 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3681 FlagUser == 0; ++UI) {
3682 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3683 SDNode *User = UI->getUser();
3684 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3685 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3692 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3693 // give up for right now.
3694 if (FlagUser->getOpcode() == PPCISD::MFCR)
3695 return SDOperand(VCMPoNode, 0);
3700 // If this is a branch on an altivec predicate comparison, lower this so
3701 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3702 // lowering is done pre-legalize, because the legalizer lowers the predicate
3703 // compare down to code that is difficult to reassemble.
3704 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3705 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3709 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3710 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3711 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3712 assert(isDot && "Can't compare against a vector result!");
3714 // If this is a comparison against something other than 0/1, then we know
3715 // that the condition is never/always true.
3716 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3717 if (Val != 0 && Val != 1) {
3718 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3719 return N->getOperand(0);
3720 // Always !=, turn it into an unconditional branch.
3721 return DAG.getNode(ISD::BR, MVT::Other,
3722 N->getOperand(0), N->getOperand(4));
3725 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3727 // Create the PPCISD altivec 'dot' comparison node.
3728 std::vector<MVT::ValueType> VTs;
3730 LHS.getOperand(2), // LHS of compare
3731 LHS.getOperand(3), // RHS of compare
3732 DAG.getConstant(CompareOpc, MVT::i32)
3734 VTs.push_back(LHS.getOperand(2).getValueType());
3735 VTs.push_back(MVT::Flag);
3736 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3738 // Unpack the result based on how the target uses it.
3739 PPC::Predicate CompOpc;
3740 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3741 default: // Can't happen, don't crash on invalid number though.
3742 case 0: // Branch on the value of the EQ bit of CR6.
3743 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3745 case 1: // Branch on the inverted value of the EQ bit of CR6.
3746 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3748 case 2: // Branch on the value of the LT bit of CR6.
3749 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3751 case 3: // Branch on the inverted value of the LT bit of CR6.
3752 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3756 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3757 DAG.getConstant(CompOpc, MVT::i32),
3758 DAG.getRegister(PPC::CR6, MVT::i32),
3759 N->getOperand(4), CompNode.getValue(1));
3768 //===----------------------------------------------------------------------===//
3769 // Inline Assembly Support
3770 //===----------------------------------------------------------------------===//
3772 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3776 const SelectionDAG &DAG,
3777 unsigned Depth) const {
3778 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3779 switch (Op.getOpcode()) {
3781 case PPCISD::LBRX: {
3782 // lhbrx is known to have the top bits cleared out.
3783 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3784 KnownZero = 0xFFFF0000;
3787 case ISD::INTRINSIC_WO_CHAIN: {
3788 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3790 case Intrinsic::ppc_altivec_vcmpbfp_p:
3791 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3792 case Intrinsic::ppc_altivec_vcmpequb_p:
3793 case Intrinsic::ppc_altivec_vcmpequh_p:
3794 case Intrinsic::ppc_altivec_vcmpequw_p:
3795 case Intrinsic::ppc_altivec_vcmpgefp_p:
3796 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3797 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3798 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3799 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3800 case Intrinsic::ppc_altivec_vcmpgtub_p:
3801 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3802 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3803 KnownZero = ~1U; // All bits but the low one are known to be zero.
3811 /// getConstraintType - Given a constraint, return the type of
3812 /// constraint it is for this target.
3813 PPCTargetLowering::ConstraintType
3814 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3815 if (Constraint.size() == 1) {
3816 switch (Constraint[0]) {
3823 return C_RegisterClass;
3826 return TargetLowering::getConstraintType(Constraint);
3829 std::pair<unsigned, const TargetRegisterClass*>
3830 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3831 MVT::ValueType VT) const {
3832 if (Constraint.size() == 1) {
3833 // GCC RS6000 Constraint Letters
3834 switch (Constraint[0]) {
3837 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3838 return std::make_pair(0U, PPC::G8RCRegisterClass);
3839 return std::make_pair(0U, PPC::GPRCRegisterClass);
3842 return std::make_pair(0U, PPC::F4RCRegisterClass);
3843 else if (VT == MVT::f64)
3844 return std::make_pair(0U, PPC::F8RCRegisterClass);
3847 return std::make_pair(0U, PPC::VRRCRegisterClass);
3849 return std::make_pair(0U, PPC::CRRCRegisterClass);
3853 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3857 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3858 /// vector. If it is invalid, don't add anything to Ops.
3859 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3860 std::vector<SDOperand>&Ops,
3861 SelectionDAG &DAG) {
3862 SDOperand Result(0,0);
3873 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3874 if (!CST) return; // Must be an immediate to match.
3875 unsigned Value = CST->getValue();
3877 default: assert(0 && "Unknown constraint letter!");
3878 case 'I': // "I" is a signed 16-bit constant.
3879 if ((short)Value == (int)Value)
3880 Result = DAG.getTargetConstant(Value, Op.getValueType());
3882 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3883 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3884 if ((short)Value == 0)
3885 Result = DAG.getTargetConstant(Value, Op.getValueType());
3887 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3888 if ((Value >> 16) == 0)
3889 Result = DAG.getTargetConstant(Value, Op.getValueType());
3891 case 'M': // "M" is a constant that is greater than 31.
3893 Result = DAG.getTargetConstant(Value, Op.getValueType());
3895 case 'N': // "N" is a positive constant that is an exact power of two.
3896 if ((int)Value > 0 && isPowerOf2_32(Value))
3897 Result = DAG.getTargetConstant(Value, Op.getValueType());
3899 case 'O': // "O" is the constant zero.
3901 Result = DAG.getTargetConstant(Value, Op.getValueType());
3903 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3904 if ((short)-Value == (int)-Value)
3905 Result = DAG.getTargetConstant(Value, Op.getValueType());
3913 Ops.push_back(Result);
3917 // Handle standard constraint letters.
3918 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3921 // isLegalAddressingMode - Return true if the addressing mode represented
3922 // by AM is legal for this target, for a load/store of the specified type.
3923 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3924 const Type *Ty) const {
3925 // FIXME: PPC does not allow r+i addressing modes for vectors!
3927 // PPC allows a sign-extended 16-bit immediate field.
3928 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3931 // No global is ever allowed as a base.
3935 // PPC only support r+r,
3937 case 0: // "r+i" or just "i", depending on HasBaseReg.
3940 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3942 // Otherwise we have r+r or r+i.
3945 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3947 // Allow 2*r as r+r.
3950 // No other scales are supported.
3957 /// isLegalAddressImmediate - Return true if the integer value can be used
3958 /// as the offset of the target addressing mode for load / store of the
3960 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3961 // PPC allows a sign-extended 16-bit immediate field.
3962 return (V > -(1 << 16) && V < (1 << 16)-1);
3965 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3969 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3970 // Depths > 0 not supported yet!
3971 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3974 MachineFunction &MF = DAG.getMachineFunction();
3975 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3976 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3978 bool isPPC64 = PPCSubTarget.isPPC64();
3980 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3982 // Set up a frame object for the return address.
3983 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3985 // Remember it for next time.
3986 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3988 // Make sure the function really does not optimize away the store of the RA
3990 FuncInfo->setLRStoreRequired();
3993 // Just load the return address off the stack.
3994 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3995 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3998 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3999 // Depths > 0 not supported yet!
4000 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
4003 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4004 bool isPPC64 = PtrVT == MVT::i64;
4006 MachineFunction &MF = DAG.getMachineFunction();
4007 MachineFrameInfo *MFI = MF.getFrameInfo();
4008 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4009 && MFI->getStackSize();
4012 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
4015 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,