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 setSetCCResultType(MVT::i32);
327 setShiftAmountType(MVT::i32);
328 setSetCCResultContents(ZeroOrOneSetCCResult);
330 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
331 setStackPointerRegisterToSaveRestore(PPC::X1);
332 setExceptionPointerRegister(PPC::X3);
333 setExceptionSelectorRegister(PPC::X4);
335 setStackPointerRegisterToSaveRestore(PPC::R1);
336 setExceptionPointerRegister(PPC::R3);
337 setExceptionSelectorRegister(PPC::R4);
340 // We have target-specific dag combine patterns for the following nodes:
341 setTargetDAGCombine(ISD::SINT_TO_FP);
342 setTargetDAGCombine(ISD::STORE);
343 setTargetDAGCombine(ISD::BR_CC);
344 setTargetDAGCombine(ISD::BSWAP);
346 // Darwin long double math library functions have $LDBL128 appended.
347 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
348 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
349 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
350 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
351 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
352 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
355 computeRegisterProperties();
358 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
359 /// function arguments in the caller parameter area.
360 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
361 TargetMachine &TM = getTargetMachine();
362 // Darwin passes everything on 4 byte boundary.
363 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
369 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
372 case PPCISD::FSEL: return "PPCISD::FSEL";
373 case PPCISD::FCFID: return "PPCISD::FCFID";
374 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
375 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
376 case PPCISD::STFIWX: return "PPCISD::STFIWX";
377 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
378 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
379 case PPCISD::VPERM: return "PPCISD::VPERM";
380 case PPCISD::Hi: return "PPCISD::Hi";
381 case PPCISD::Lo: return "PPCISD::Lo";
382 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
383 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
384 case PPCISD::SRL: return "PPCISD::SRL";
385 case PPCISD::SRA: return "PPCISD::SRA";
386 case PPCISD::SHL: return "PPCISD::SHL";
387 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
388 case PPCISD::STD_32: return "PPCISD::STD_32";
389 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
390 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
391 case PPCISD::MTCTR: return "PPCISD::MTCTR";
392 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
393 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
394 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
395 case PPCISD::MFCR: return "PPCISD::MFCR";
396 case PPCISD::VCMP: return "PPCISD::VCMP";
397 case PPCISD::VCMPo: return "PPCISD::VCMPo";
398 case PPCISD::LBRX: return "PPCISD::LBRX";
399 case PPCISD::STBRX: return "PPCISD::STBRX";
400 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
401 case PPCISD::MFFS: return "PPCISD::MFFS";
402 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
403 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
404 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
405 case PPCISD::MTFSF: return "PPCISD::MTFSF";
409 //===----------------------------------------------------------------------===//
410 // Node matching predicates, for use by the tblgen matching code.
411 //===----------------------------------------------------------------------===//
413 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
414 static bool isFloatingPointZero(SDOperand Op) {
415 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
416 return CFP->getValueAPF().isZero();
417 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
418 // Maybe this has already been legalized into the constant pool?
419 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
420 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
421 return CFP->getValueAPF().isZero();
426 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
427 /// true if Op is undef or if it matches the specified value.
428 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
429 return Op.getOpcode() == ISD::UNDEF ||
430 cast<ConstantSDNode>(Op)->getValue() == Val;
433 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
434 /// VPKUHUM instruction.
435 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
437 for (unsigned i = 0; i != 16; ++i)
438 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
441 for (unsigned i = 0; i != 8; ++i)
442 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
443 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
449 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
450 /// VPKUWUM instruction.
451 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
453 for (unsigned i = 0; i != 16; i += 2)
454 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
455 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
458 for (unsigned i = 0; i != 8; i += 2)
459 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
460 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
461 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
462 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
468 /// isVMerge - Common function, used to match vmrg* shuffles.
470 static bool isVMerge(SDNode *N, unsigned UnitSize,
471 unsigned LHSStart, unsigned RHSStart) {
472 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
473 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
474 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
475 "Unsupported merge size!");
477 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
478 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
479 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
480 LHSStart+j+i*UnitSize) ||
481 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
482 RHSStart+j+i*UnitSize))
488 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
489 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
490 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
492 return isVMerge(N, UnitSize, 8, 24);
493 return isVMerge(N, UnitSize, 8, 8);
496 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
497 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
498 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
500 return isVMerge(N, UnitSize, 0, 16);
501 return isVMerge(N, UnitSize, 0, 0);
505 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
506 /// amount, otherwise return -1.
507 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
508 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
509 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
510 // Find the first non-undef value in the shuffle mask.
512 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
515 if (i == 16) return -1; // all undef.
517 // Otherwise, check to see if the rest of the elements are consequtively
518 // numbered from this value.
519 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
520 if (ShiftAmt < i) return -1;
524 // Check the rest of the elements to see if they are consequtive.
525 for (++i; i != 16; ++i)
526 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
529 // Check the rest of the elements to see if they are consequtive.
530 for (++i; i != 16; ++i)
531 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
538 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
539 /// specifies a splat of a single element that is suitable for input to
540 /// VSPLTB/VSPLTH/VSPLTW.
541 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
542 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
543 N->getNumOperands() == 16 &&
544 (EltSize == 1 || EltSize == 2 || EltSize == 4));
546 // This is a splat operation if each element of the permute is the same, and
547 // if the value doesn't reference the second vector.
548 unsigned ElementBase = 0;
549 SDOperand Elt = N->getOperand(0);
550 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
551 ElementBase = EltV->getValue();
553 return false; // FIXME: Handle UNDEF elements too!
555 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
558 // Check that they are consequtive.
559 for (unsigned i = 1; i != EltSize; ++i) {
560 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
561 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
565 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
566 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
567 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
568 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
569 "Invalid VECTOR_SHUFFLE mask!");
570 for (unsigned j = 0; j != EltSize; ++j)
571 if (N->getOperand(i+j) != N->getOperand(j))
578 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
580 bool PPC::isAllNegativeZeroVector(SDNode *N) {
581 assert(N->getOpcode() == ISD::BUILD_VECTOR);
582 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
583 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
584 return CFP->getValueAPF().isNegZero();
588 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
589 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
590 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
591 assert(isSplatShuffleMask(N, EltSize));
592 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
595 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
596 /// by using a vspltis[bhw] instruction of the specified element size, return
597 /// the constant being splatted. The ByteSize field indicates the number of
598 /// bytes of each element [124] -> [bhw].
599 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
600 SDOperand OpVal(0, 0);
602 // If ByteSize of the splat is bigger than the element size of the
603 // build_vector, then we have a case where we are checking for a splat where
604 // multiple elements of the buildvector are folded together into a single
605 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
606 unsigned EltSize = 16/N->getNumOperands();
607 if (EltSize < ByteSize) {
608 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
609 SDOperand UniquedVals[4];
610 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
612 // See if all of the elements in the buildvector agree across.
613 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
614 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
615 // If the element isn't a constant, bail fully out.
616 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
619 if (UniquedVals[i&(Multiple-1)].Val == 0)
620 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
621 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
622 return SDOperand(); // no match.
625 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
626 // either constant or undef values that are identical for each chunk. See
627 // if these chunks can form into a larger vspltis*.
629 // Check to see if all of the leading entries are either 0 or -1. If
630 // neither, then this won't fit into the immediate field.
631 bool LeadingZero = true;
632 bool LeadingOnes = true;
633 for (unsigned i = 0; i != Multiple-1; ++i) {
634 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
636 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
637 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
639 // Finally, check the least significant entry.
641 if (UniquedVals[Multiple-1].Val == 0)
642 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
643 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
645 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
648 if (UniquedVals[Multiple-1].Val == 0)
649 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
650 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
651 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
652 return DAG.getTargetConstant(Val, MVT::i32);
658 // Check to see if this buildvec has a single non-undef value in its elements.
659 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
660 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
662 OpVal = N->getOperand(i);
663 else if (OpVal != N->getOperand(i))
667 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
669 unsigned ValSizeInBytes = 0;
671 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
672 Value = CN->getValue();
673 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
674 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
675 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
676 Value = FloatToBits(CN->getValueAPF().convertToFloat());
680 // If the splat value is larger than the element value, then we can never do
681 // this splat. The only case that we could fit the replicated bits into our
682 // immediate field for would be zero, and we prefer to use vxor for it.
683 if (ValSizeInBytes < ByteSize) return SDOperand();
685 // If the element value is larger than the splat value, cut it in half and
686 // check to see if the two halves are equal. Continue doing this until we
687 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
688 while (ValSizeInBytes > ByteSize) {
689 ValSizeInBytes >>= 1;
691 // If the top half equals the bottom half, we're still ok.
692 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
693 (Value & ((1 << (8*ValSizeInBytes))-1)))
697 // Properly sign extend the value.
698 int ShAmt = (4-ByteSize)*8;
699 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
701 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
702 if (MaskVal == 0) return SDOperand();
704 // Finally, if this value fits in a 5 bit sext field, return it
705 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
706 return DAG.getTargetConstant(MaskVal, MVT::i32);
710 //===----------------------------------------------------------------------===//
711 // Addressing Mode Selection
712 //===----------------------------------------------------------------------===//
714 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
715 /// or 64-bit immediate, and if the value can be accurately represented as a
716 /// sign extension from a 16-bit value. If so, this returns true and the
718 static bool isIntS16Immediate(SDNode *N, short &Imm) {
719 if (N->getOpcode() != ISD::Constant)
722 Imm = (short)cast<ConstantSDNode>(N)->getValue();
723 if (N->getValueType(0) == MVT::i32)
724 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
726 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
728 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
729 return isIntS16Immediate(Op.Val, Imm);
733 /// SelectAddressRegReg - Given the specified addressed, check to see if it
734 /// can be represented as an indexed [r+r] operation. Returns false if it
735 /// can be more efficiently represented with [r+imm].
736 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
740 if (N.getOpcode() == ISD::ADD) {
741 if (isIntS16Immediate(N.getOperand(1), imm))
743 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
746 Base = N.getOperand(0);
747 Index = N.getOperand(1);
749 } else if (N.getOpcode() == ISD::OR) {
750 if (isIntS16Immediate(N.getOperand(1), imm))
751 return false; // r+i can fold it if we can.
753 // If this is an or of disjoint bitfields, we can codegen this as an add
754 // (for better address arithmetic) if the LHS and RHS of the OR are provably
756 APInt LHSKnownZero, LHSKnownOne;
757 APInt RHSKnownZero, RHSKnownOne;
758 DAG.ComputeMaskedBits(N.getOperand(0),
759 APInt::getAllOnesValue(N.getOperand(0)
760 .getValueSizeInBits()),
761 LHSKnownZero, LHSKnownOne);
763 if (LHSKnownZero.getBoolValue()) {
764 DAG.ComputeMaskedBits(N.getOperand(1),
765 APInt::getAllOnesValue(N.getOperand(1)
766 .getValueSizeInBits()),
767 RHSKnownZero, RHSKnownOne);
768 // If all of the bits are known zero on the LHS or RHS, the add won't
770 if (~(LHSKnownZero | RHSKnownZero) == 0) {
771 Base = N.getOperand(0);
772 Index = N.getOperand(1);
781 /// Returns true if the address N can be represented by a base register plus
782 /// a signed 16-bit displacement [r+imm], and if it is not better
783 /// represented as reg+reg.
784 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
785 SDOperand &Base, SelectionDAG &DAG){
786 // If this can be more profitably realized as r+r, fail.
787 if (SelectAddressRegReg(N, Disp, Base, DAG))
790 if (N.getOpcode() == ISD::ADD) {
792 if (isIntS16Immediate(N.getOperand(1), imm)) {
793 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
794 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
795 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
797 Base = N.getOperand(0);
799 return true; // [r+i]
800 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
801 // Match LOAD (ADD (X, Lo(G))).
802 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
803 && "Cannot handle constant offsets yet!");
804 Disp = N.getOperand(1).getOperand(0); // The global address.
805 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
806 Disp.getOpcode() == ISD::TargetConstantPool ||
807 Disp.getOpcode() == ISD::TargetJumpTable);
808 Base = N.getOperand(0);
809 return true; // [&g+r]
811 } else if (N.getOpcode() == ISD::OR) {
813 if (isIntS16Immediate(N.getOperand(1), imm)) {
814 // If this is an or of disjoint bitfields, we can codegen this as an add
815 // (for better address arithmetic) if the LHS and RHS of the OR are
816 // provably disjoint.
817 APInt LHSKnownZero, LHSKnownOne;
818 DAG.ComputeMaskedBits(N.getOperand(0),
819 APInt::getAllOnesValue(32),
820 LHSKnownZero, LHSKnownOne);
821 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
822 // If all of the bits are known zero on the LHS or RHS, the add won't
824 Base = N.getOperand(0);
825 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
829 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
830 // Loading from a constant address.
832 // If this address fits entirely in a 16-bit sext immediate field, codegen
835 if (isIntS16Immediate(CN, Imm)) {
836 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
837 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
841 // Handle 32-bit sext immediates with LIS + addr mode.
842 if (CN->getValueType(0) == MVT::i32 ||
843 (int64_t)CN->getValue() == (int)CN->getValue()) {
844 int Addr = (int)CN->getValue();
846 // Otherwise, break this down into an LIS + disp.
847 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
849 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
850 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
851 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
856 Disp = DAG.getTargetConstant(0, getPointerTy());
857 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
858 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
861 return true; // [r+0]
864 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
865 /// represented as an indexed [r+r] operation.
866 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
869 // Check to see if we can easily represent this as an [r+r] address. This
870 // will fail if it thinks that the address is more profitably represented as
871 // reg+imm, e.g. where imm = 0.
872 if (SelectAddressRegReg(N, Base, Index, DAG))
875 // If the operand is an addition, always emit this as [r+r], since this is
876 // better (for code size, and execution, as the memop does the add for free)
877 // than emitting an explicit add.
878 if (N.getOpcode() == ISD::ADD) {
879 Base = N.getOperand(0);
880 Index = N.getOperand(1);
884 // Otherwise, do it the hard way, using R0 as the base register.
885 Base = DAG.getRegister(PPC::R0, N.getValueType());
890 /// SelectAddressRegImmShift - Returns true if the address N can be
891 /// represented by a base register plus a signed 14-bit displacement
892 /// [r+imm*4]. Suitable for use by STD and friends.
893 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
896 // If this can be more profitably realized as r+r, fail.
897 if (SelectAddressRegReg(N, Disp, Base, DAG))
900 if (N.getOpcode() == ISD::ADD) {
902 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
903 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
904 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
905 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
907 Base = N.getOperand(0);
909 return true; // [r+i]
910 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
911 // Match LOAD (ADD (X, Lo(G))).
912 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
913 && "Cannot handle constant offsets yet!");
914 Disp = N.getOperand(1).getOperand(0); // The global address.
915 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
916 Disp.getOpcode() == ISD::TargetConstantPool ||
917 Disp.getOpcode() == ISD::TargetJumpTable);
918 Base = N.getOperand(0);
919 return true; // [&g+r]
921 } else if (N.getOpcode() == ISD::OR) {
923 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
924 // If this is an or of disjoint bitfields, we can codegen this as an add
925 // (for better address arithmetic) if the LHS and RHS of the OR are
926 // provably disjoint.
927 APInt LHSKnownZero, LHSKnownOne;
928 DAG.ComputeMaskedBits(N.getOperand(0),
929 APInt::getAllOnesValue(32),
930 LHSKnownZero, LHSKnownOne);
931 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
932 // If all of the bits are known zero on the LHS or RHS, the add won't
934 Base = N.getOperand(0);
935 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
939 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
940 // Loading from a constant address. Verify low two bits are clear.
941 if ((CN->getValue() & 3) == 0) {
942 // If this address fits entirely in a 14-bit sext immediate field, codegen
945 if (isIntS16Immediate(CN, Imm)) {
946 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
947 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
951 // Fold the low-part of 32-bit absolute addresses into addr mode.
952 if (CN->getValueType(0) == MVT::i32 ||
953 (int64_t)CN->getValue() == (int)CN->getValue()) {
954 int Addr = (int)CN->getValue();
956 // Otherwise, break this down into an LIS + disp.
957 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
959 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
960 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
961 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
967 Disp = DAG.getTargetConstant(0, getPointerTy());
968 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
969 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
972 return true; // [r+0]
976 /// getPreIndexedAddressParts - returns true by value, base pointer and
977 /// offset pointer and addressing mode by reference if the node's address
978 /// can be legally represented as pre-indexed load / store address.
979 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
981 ISD::MemIndexedMode &AM,
983 // Disabled by default for now.
984 if (!EnablePPCPreinc) return false;
988 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
989 Ptr = LD->getBasePtr();
990 VT = LD->getMemoryVT();
992 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
994 Ptr = ST->getBasePtr();
995 VT = ST->getMemoryVT();
999 // PowerPC doesn't have preinc load/store instructions for vectors.
1000 if (MVT::isVector(VT))
1003 // TODO: Check reg+reg first.
1005 // LDU/STU use reg+imm*4, others use reg+imm.
1006 if (VT != MVT::i64) {
1008 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1012 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1016 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1017 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1018 // sext i32 to i64 when addr mode is r+i.
1019 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1020 LD->getExtensionType() == ISD::SEXTLOAD &&
1021 isa<ConstantSDNode>(Offset))
1029 //===----------------------------------------------------------------------===//
1030 // LowerOperation implementation
1031 //===----------------------------------------------------------------------===//
1033 SDOperand PPCTargetLowering::LowerConstantPool(SDOperand Op,
1034 SelectionDAG &DAG) {
1035 MVT::ValueType PtrVT = Op.getValueType();
1036 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1037 Constant *C = CP->getConstVal();
1038 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1039 SDOperand Zero = DAG.getConstant(0, PtrVT);
1041 const TargetMachine &TM = DAG.getTarget();
1043 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1044 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1046 // If this is a non-darwin platform, we don't support non-static relo models
1048 if (TM.getRelocationModel() == Reloc::Static ||
1049 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1050 // Generate non-pic code that has direct accesses to the constant pool.
1051 // The address of the global is just (hi(&g)+lo(&g)).
1052 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1055 if (TM.getRelocationModel() == Reloc::PIC_) {
1056 // With PIC, the first instruction is actually "GR+hi(&G)".
1057 Hi = DAG.getNode(ISD::ADD, PtrVT,
1058 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1061 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1065 SDOperand PPCTargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1066 MVT::ValueType PtrVT = Op.getValueType();
1067 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1068 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1069 SDOperand Zero = DAG.getConstant(0, PtrVT);
1071 const TargetMachine &TM = DAG.getTarget();
1073 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1074 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1076 // If this is a non-darwin platform, we don't support non-static relo models
1078 if (TM.getRelocationModel() == Reloc::Static ||
1079 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1080 // Generate non-pic code that has direct accesses to the constant pool.
1081 // The address of the global is just (hi(&g)+lo(&g)).
1082 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1085 if (TM.getRelocationModel() == Reloc::PIC_) {
1086 // With PIC, the first instruction is actually "GR+hi(&G)".
1087 Hi = DAG.getNode(ISD::ADD, PtrVT,
1088 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1091 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1095 SDOperand PPCTargetLowering::LowerGlobalTLSAddress(SDOperand Op,
1096 SelectionDAG &DAG) {
1097 assert(0 && "TLS not implemented for PPC.");
1100 SDOperand PPCTargetLowering::LowerGlobalAddress(SDOperand Op,
1101 SelectionDAG &DAG) {
1102 MVT::ValueType PtrVT = Op.getValueType();
1103 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1104 GlobalValue *GV = GSDN->getGlobal();
1105 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1106 // If it's a debug information descriptor, don't mess with it.
1107 if (DAG.isVerifiedDebugInfoDesc(Op))
1109 SDOperand Zero = DAG.getConstant(0, PtrVT);
1111 const TargetMachine &TM = DAG.getTarget();
1113 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1114 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1116 // If this is a non-darwin platform, we don't support non-static relo models
1118 if (TM.getRelocationModel() == Reloc::Static ||
1119 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1120 // Generate non-pic code that has direct accesses to globals.
1121 // The address of the global is just (hi(&g)+lo(&g)).
1122 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1125 if (TM.getRelocationModel() == Reloc::PIC_) {
1126 // With PIC, the first instruction is actually "GR+hi(&G)".
1127 Hi = DAG.getNode(ISD::ADD, PtrVT,
1128 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1131 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1133 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1136 // If the global is weak or external, we have to go through the lazy
1138 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1141 SDOperand PPCTargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1142 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1144 // If we're comparing for equality to zero, expose the fact that this is
1145 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1146 // fold the new nodes.
1147 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1148 if (C->isNullValue() && CC == ISD::SETEQ) {
1149 MVT::ValueType VT = Op.getOperand(0).getValueType();
1150 SDOperand Zext = Op.getOperand(0);
1151 if (VT < MVT::i32) {
1153 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1155 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1156 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1157 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1158 DAG.getConstant(Log2b, MVT::i32));
1159 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1161 // Leave comparisons against 0 and -1 alone for now, since they're usually
1162 // optimized. FIXME: revisit this when we can custom lower all setcc
1164 if (C->isAllOnesValue() || C->isNullValue())
1168 // If we have an integer seteq/setne, turn it into a compare against zero
1169 // by xor'ing the rhs with the lhs, which is faster than setting a
1170 // condition register, reading it back out, and masking the correct bit. The
1171 // normal approach here uses sub to do this instead of xor. Using xor exposes
1172 // the result to other bit-twiddling opportunities.
1173 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1174 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1175 MVT::ValueType VT = Op.getValueType();
1176 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1178 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1183 SDOperand PPCTargetLowering::LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1184 int VarArgsFrameIndex,
1185 int VarArgsStackOffset,
1186 unsigned VarArgsNumGPR,
1187 unsigned VarArgsNumFPR,
1188 const PPCSubtarget &Subtarget) {
1190 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1193 SDOperand PPCTargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1194 int VarArgsFrameIndex,
1195 int VarArgsStackOffset,
1196 unsigned VarArgsNumGPR,
1197 unsigned VarArgsNumFPR,
1198 const PPCSubtarget &Subtarget) {
1200 if (Subtarget.isMachoABI()) {
1201 // vastart just stores the address of the VarArgsFrameIndex slot into the
1202 // memory location argument.
1203 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1204 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1205 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1206 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1209 // For ELF 32 ABI we follow the layout of the va_list struct.
1210 // We suppose the given va_list is already allocated.
1213 // char gpr; /* index into the array of 8 GPRs
1214 // * stored in the register save area
1215 // * gpr=0 corresponds to r3,
1216 // * gpr=1 to r4, etc.
1218 // char fpr; /* index into the array of 8 FPRs
1219 // * stored in the register save area
1220 // * fpr=0 corresponds to f1,
1221 // * fpr=1 to f2, etc.
1223 // char *overflow_arg_area;
1224 // /* location on stack that holds
1225 // * the next overflow argument
1227 // char *reg_save_area;
1228 // /* where r3:r10 and f1:f8 (if saved)
1234 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1235 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1238 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1240 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1241 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1243 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1244 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1246 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1247 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1249 uint64_t FPROffset = 1;
1250 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1252 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1254 // Store first byte : number of int regs
1255 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1256 Op.getOperand(1), SV, 0);
1257 uint64_t nextOffset = FPROffset;
1258 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1261 // Store second byte : number of float regs
1262 SDOperand secondStore =
1263 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1264 nextOffset += StackOffset;
1265 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1267 // Store second word : arguments given on stack
1268 SDOperand thirdStore =
1269 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1270 nextOffset += FrameOffset;
1271 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1273 // Store third word : arguments given in registers
1274 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1278 #include "PPCGenCallingConv.inc"
1280 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1281 /// depending on which subtarget is selected.
1282 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1283 if (Subtarget.isMachoABI()) {
1284 static const unsigned FPR[] = {
1285 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1286 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1292 static const unsigned FPR[] = {
1293 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1299 SDOperand PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op,
1301 int &VarArgsFrameIndex,
1302 int &VarArgsStackOffset,
1303 unsigned &VarArgsNumGPR,
1304 unsigned &VarArgsNumFPR,
1305 const PPCSubtarget &Subtarget) {
1306 // TODO: add description of PPC stack frame format, or at least some docs.
1308 MachineFunction &MF = DAG.getMachineFunction();
1309 MachineFrameInfo *MFI = MF.getFrameInfo();
1310 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1311 SmallVector<SDOperand, 8> ArgValues;
1312 SDOperand Root = Op.getOperand(0);
1314 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1315 bool isPPC64 = PtrVT == MVT::i64;
1316 bool isMachoABI = Subtarget.isMachoABI();
1317 bool isELF32_ABI = Subtarget.isELF32_ABI();
1318 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1320 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1322 static const unsigned GPR_32[] = { // 32-bit registers.
1323 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1324 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1326 static const unsigned GPR_64[] = { // 64-bit registers.
1327 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1328 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1331 static const unsigned *FPR = GetFPR(Subtarget);
1333 static const unsigned VR[] = {
1334 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1335 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1338 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1339 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1340 const unsigned Num_VR_Regs = array_lengthof( VR);
1342 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1344 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1346 // Add DAG nodes to load the arguments or copy them out of registers. On
1347 // entry to a function on PPC, the arguments start after the linkage area,
1348 // although the first ones are often in registers.
1350 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1351 // represented with two words (long long or double) must be copied to an
1352 // even GPR_idx value or to an even ArgOffset value.
1354 SmallVector<SDOperand, 8> MemOps;
1356 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1358 bool needsLoad = false;
1359 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1360 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1361 unsigned ArgSize = ObjSize;
1362 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1363 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1364 unsigned isByVal = Flags & ISD::ParamFlags::ByVal;
1365 // See if next argument requires stack alignment in ELF
1366 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1367 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1368 (!(Flags & AlignFlag)));
1370 unsigned CurArgOffset = ArgOffset;
1372 // FIXME alignment for ELF may not be right
1373 // FIXME the codegen can be much improved in some cases.
1374 // We do not have to keep everything in memory.
1376 // Double word align in ELF
1377 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1378 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1379 ObjSize = (Flags & ISD::ParamFlags::ByValSize) >>
1380 ISD::ParamFlags::ByValSizeOffs;
1381 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1382 // The value of the object is its address.
1383 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1384 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1385 ArgValues.push_back(FIN);
1386 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1387 // Store whatever pieces of the object are in registers
1388 // to memory. ArgVal will be address of the beginning of
1390 if (GPR_idx != Num_GPR_Regs) {
1391 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1392 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1393 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1394 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1395 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1396 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1397 MemOps.push_back(Store);
1399 if (isMachoABI) ArgOffset += PtrByteSize;
1401 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1409 default: assert(0 && "Unhandled argument type!");
1411 // Double word align in ELF
1412 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1413 if (GPR_idx != Num_GPR_Regs) {
1414 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1415 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1416 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1420 ArgSize = PtrByteSize;
1422 // Stack align in ELF
1423 if (needsLoad && Expand && isELF32_ABI)
1424 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1425 // All int arguments reserve stack space in Macho ABI.
1426 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1429 case MVT::i64: // PPC64
1430 if (GPR_idx != Num_GPR_Regs) {
1431 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1432 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1433 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1438 // All int arguments reserve stack space in Macho ABI.
1439 if (isMachoABI || needsLoad) ArgOffset += 8;
1444 // Every 4 bytes of argument space consumes one of the GPRs available for
1445 // argument passing.
1446 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1448 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1451 if (FPR_idx != Num_FPR_Regs) {
1453 if (ObjectVT == MVT::f32)
1454 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1456 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1457 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1458 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1464 // Stack align in ELF
1465 if (needsLoad && Expand && isELF32_ABI)
1466 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1467 // All FP arguments reserve stack space in Macho ABI.
1468 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1474 // Note that vector arguments in registers don't reserve stack space.
1475 if (VR_idx != Num_VR_Regs) {
1476 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1477 RegInfo.addLiveIn(VR[VR_idx], VReg);
1478 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1481 // This should be simple, but requires getting 16-byte aligned stack
1483 assert(0 && "Loading VR argument not implemented yet!");
1489 // We need to load the argument to a virtual register if we determined above
1490 // that we ran out of physical registers of the appropriate type.
1492 int FI = MFI->CreateFixedObject(ObjSize,
1493 CurArgOffset + (ArgSize - ObjSize));
1494 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1495 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1498 ArgValues.push_back(ArgVal);
1501 // If the function takes variable number of arguments, make a frame index for
1502 // the start of the first vararg value... for expansion of llvm.va_start.
1503 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1508 VarArgsNumGPR = GPR_idx;
1509 VarArgsNumFPR = FPR_idx;
1511 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1513 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1514 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1515 MVT::getSizeInBits(PtrVT)/8);
1517 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1524 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1526 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1528 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1529 // stored to the VarArgsFrameIndex on the stack.
1531 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1532 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1533 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1534 MemOps.push_back(Store);
1535 // Increment the address by four for the next argument to store
1536 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1537 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1541 // If this function is vararg, store any remaining integer argument regs
1542 // to their spots on the stack so that they may be loaded by deferencing the
1543 // result of va_next.
1544 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1547 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1549 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1551 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1552 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1553 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1554 MemOps.push_back(Store);
1555 // Increment the address by four for the next argument to store
1556 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1557 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1560 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1563 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1564 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1565 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1566 MemOps.push_back(Store);
1567 // Increment the address by eight for the next argument to store
1568 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1570 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1573 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1575 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1577 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1578 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1579 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1580 MemOps.push_back(Store);
1581 // Increment the address by eight for the next argument to store
1582 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1584 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1589 if (!MemOps.empty())
1590 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1592 ArgValues.push_back(Root);
1594 // Return the new list of results.
1595 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1596 Op.Val->value_end());
1597 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1600 /// isCallCompatibleAddress - Return the immediate to use if the specified
1601 /// 32-bit value is representable in the immediate field of a BxA instruction.
1602 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1603 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1606 int Addr = C->getValue();
1607 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1608 (Addr << 6 >> 6) != Addr)
1609 return 0; // Top 6 bits have to be sext of immediate.
1611 return DAG.getConstant((int)C->getValue() >> 2,
1612 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1615 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1616 /// by "Src" to address "Dst" of size "Size". Alignment information is
1617 /// specified by the specific parameter attribute. The copy will be passed as
1618 /// a byval function parameter.
1619 /// Sometimes what we are copying is the end of a larger object, the part that
1620 /// does not fit in registers.
1622 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1623 unsigned Flags, SelectionDAG &DAG, unsigned Size) {
1624 unsigned Align = 1 <<
1625 ((Flags & ISD::ParamFlags::ByValAlign) >> ISD::ParamFlags::ByValAlignOffs);
1626 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1627 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1628 SDOperand AlwaysInline = DAG.getConstant(0, MVT::i32);
1629 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1632 SDOperand PPCTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG,
1633 const PPCSubtarget &Subtarget) {
1634 SDOperand Chain = Op.getOperand(0);
1635 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1636 SDOperand Callee = Op.getOperand(4);
1637 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1639 bool isMachoABI = Subtarget.isMachoABI();
1640 bool isELF32_ABI = Subtarget.isELF32_ABI();
1642 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1643 bool isPPC64 = PtrVT == MVT::i64;
1644 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1646 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1647 // SelectExpr to use to put the arguments in the appropriate registers.
1648 std::vector<SDOperand> args_to_use;
1650 // Count how many bytes are to be pushed on the stack, including the linkage
1651 // area, and parameter passing area. We start with 24/48 bytes, which is
1652 // prereserved space for [SP][CR][LR][3 x unused].
1653 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1655 // Add up all the space actually used.
1656 for (unsigned i = 0; i != NumOps; ++i) {
1657 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1658 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1659 if (Flags & ISD::ParamFlags::ByVal)
1660 ArgSize = (Flags & ISD::ParamFlags::ByValSize) >>
1661 ISD::ParamFlags::ByValSizeOffs;
1662 ArgSize = std::max(ArgSize, PtrByteSize);
1663 NumBytes += ArgSize;
1666 // The prolog code of the callee may store up to 8 GPR argument registers to
1667 // the stack, allowing va_start to index over them in memory if its varargs.
1668 // Because we cannot tell if this is needed on the caller side, we have to
1669 // conservatively assume that it is needed. As such, make sure we have at
1670 // least enough stack space for the caller to store the 8 GPRs.
1671 NumBytes = std::max(NumBytes,
1672 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1674 // Adjust the stack pointer for the new arguments...
1675 // These operations are automatically eliminated by the prolog/epilog pass
1676 Chain = DAG.getCALLSEQ_START(Chain,
1677 DAG.getConstant(NumBytes, PtrVT));
1678 SDOperand CallSeqStart = Chain;
1680 // Set up a copy of the stack pointer for use loading and storing any
1681 // arguments that may not fit in the registers available for argument
1685 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1687 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1689 // Figure out which arguments are going to go in registers, and which in
1690 // memory. Also, if this is a vararg function, floating point operations
1691 // must be stored to our stack, and loaded into integer regs as well, if
1692 // any integer regs are available for argument passing.
1693 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1694 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1696 static const unsigned GPR_32[] = { // 32-bit registers.
1697 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1698 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1700 static const unsigned GPR_64[] = { // 64-bit registers.
1701 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1702 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1704 static const unsigned *FPR = GetFPR(Subtarget);
1706 static const unsigned VR[] = {
1707 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1708 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1710 const unsigned NumGPRs = array_lengthof(GPR_32);
1711 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1712 const unsigned NumVRs = array_lengthof( VR);
1714 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1716 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1717 SmallVector<SDOperand, 8> MemOpChains;
1718 for (unsigned i = 0; i != NumOps; ++i) {
1720 SDOperand Arg = Op.getOperand(5+2*i);
1721 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1722 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1723 // See if next argument requires stack alignment in ELF
1724 unsigned next = 5+2*(i+1)+1;
1725 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1726 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1727 (!(Flags & AlignFlag)));
1729 // PtrOff will be used to store the current argument to the stack if a
1730 // register cannot be found for it.
1733 // Stack align in ELF 32
1734 if (isELF32_ABI && Expand)
1735 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1736 StackPtr.getValueType());
1738 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1740 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1742 // On PPC64, promote integers to 64-bit values.
1743 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1744 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1746 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1749 // FIXME Elf untested, what are alignment rules?
1750 // FIXME memcpy is used way more than necessary. Correctness first.
1751 if (Flags & ISD::ParamFlags::ByVal) {
1752 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1753 ISD::ParamFlags::ByValSizeOffs;
1754 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1755 if (Size==1 || Size==2) {
1756 // Very small objects are passed right-justified.
1757 // Everything else is passed left-justified.
1758 MVT::ValueType VT = (Size==1) ? MVT::i8 : MVT::i16;
1759 if (GPR_idx != NumGPRs) {
1760 SDOperand Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
1762 MemOpChains.push_back(Load.getValue(1));
1763 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1765 ArgOffset += PtrByteSize;
1767 SDOperand Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
1768 SDOperand AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
1769 SDOperand MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
1770 CallSeqStart.Val->getOperand(0),
1772 // This must go outside the CALLSEQ_START..END.
1773 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1774 CallSeqStart.Val->getOperand(1));
1775 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1776 Chain = CallSeqStart = NewCallSeqStart;
1777 ArgOffset += PtrByteSize;
1781 for (unsigned j=0; j<Size; j+=PtrByteSize) {
1782 SDOperand Const = DAG.getConstant(j, PtrOff.getValueType());
1783 SDOperand AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
1784 if (GPR_idx != NumGPRs) {
1785 SDOperand Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
1786 MemOpChains.push_back(Load.getValue(1));
1787 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1789 ArgOffset += PtrByteSize;
1791 SDOperand AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
1792 SDOperand MemcpyCall = CreateCopyOfByValArgument(AddArg, AddPtr,
1793 CallSeqStart.Val->getOperand(0),
1794 Flags, DAG, Size - j);
1795 // This must go outside the CALLSEQ_START..END.
1796 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1797 CallSeqStart.Val->getOperand(1));
1798 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1799 Chain = CallSeqStart = NewCallSeqStart;
1800 ArgOffset += ((Size - j + 3)/4)*4;
1807 switch (Arg.getValueType()) {
1808 default: assert(0 && "Unexpected ValueType for argument!");
1811 // Double word align in ELF
1812 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1813 if (GPR_idx != NumGPRs) {
1814 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1816 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1819 if (inMem || isMachoABI) {
1820 // Stack align in ELF
1821 if (isELF32_ABI && Expand)
1822 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1824 ArgOffset += PtrByteSize;
1830 // Float varargs need to be promoted to double.
1831 if (Arg.getValueType() == MVT::f32)
1832 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1835 if (FPR_idx != NumFPRs) {
1836 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1839 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1840 MemOpChains.push_back(Store);
1842 // Float varargs are always shadowed in available integer registers
1843 if (GPR_idx != NumGPRs) {
1844 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1845 MemOpChains.push_back(Load.getValue(1));
1846 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1849 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1850 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1851 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1852 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1853 MemOpChains.push_back(Load.getValue(1));
1854 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1858 // If we have any FPRs remaining, we may also have GPRs remaining.
1859 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1862 if (GPR_idx != NumGPRs)
1864 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1865 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1870 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1873 if (inMem || isMachoABI) {
1874 // Stack align in ELF
1875 if (isELF32_ABI && Expand)
1876 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1880 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1887 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1888 assert(VR_idx != NumVRs &&
1889 "Don't support passing more than 12 vector args yet!");
1890 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1894 if (!MemOpChains.empty())
1895 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1896 &MemOpChains[0], MemOpChains.size());
1898 // Build a sequence of copy-to-reg nodes chained together with token chain
1899 // and flag operands which copy the outgoing args into the appropriate regs.
1901 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1902 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1904 InFlag = Chain.getValue(1);
1907 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1908 if (isVarArg && isELF32_ABI) {
1909 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1910 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1911 InFlag = Chain.getValue(1);
1914 std::vector<MVT::ValueType> NodeTys;
1915 NodeTys.push_back(MVT::Other); // Returns a chain
1916 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1918 SmallVector<SDOperand, 8> Ops;
1919 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1921 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1922 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1923 // node so that legalize doesn't hack it.
1924 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1925 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1926 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1927 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1928 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1929 // If this is an absolute destination address, use the munged value.
1930 Callee = SDOperand(Dest, 0);
1932 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1933 // to do the call, we can't use PPCISD::CALL.
1934 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1935 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1936 InFlag = Chain.getValue(1);
1938 // Copy the callee address into R12 on darwin.
1940 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1941 InFlag = Chain.getValue(1);
1945 NodeTys.push_back(MVT::Other);
1946 NodeTys.push_back(MVT::Flag);
1947 Ops.push_back(Chain);
1948 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1952 // If this is a direct call, pass the chain and the callee.
1954 Ops.push_back(Chain);
1955 Ops.push_back(Callee);
1958 // Add argument registers to the end of the list so that they are known live
1960 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1961 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1962 RegsToPass[i].second.getValueType()));
1965 Ops.push_back(InFlag);
1966 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1967 InFlag = Chain.getValue(1);
1969 Chain = DAG.getCALLSEQ_END(Chain,
1970 DAG.getConstant(NumBytes, PtrVT),
1971 DAG.getConstant(0, PtrVT),
1973 if (Op.Val->getValueType(0) != MVT::Other)
1974 InFlag = Chain.getValue(1);
1976 SDOperand ResultVals[3];
1977 unsigned NumResults = 0;
1980 // If the call has results, copy the values out of the ret val registers.
1981 switch (Op.Val->getValueType(0)) {
1982 default: assert(0 && "Unexpected ret value!");
1983 case MVT::Other: break;
1985 if (Op.Val->getValueType(1) == MVT::i32) {
1986 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1987 ResultVals[0] = Chain.getValue(0);
1988 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
1989 Chain.getValue(2)).getValue(1);
1990 ResultVals[1] = Chain.getValue(0);
1992 NodeTys.push_back(MVT::i32);
1994 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1995 ResultVals[0] = Chain.getValue(0);
1998 NodeTys.push_back(MVT::i32);
2001 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
2002 ResultVals[0] = Chain.getValue(0);
2004 NodeTys.push_back(MVT::i64);
2007 if (Op.Val->getValueType(1) == MVT::f64) {
2008 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
2009 ResultVals[0] = Chain.getValue(0);
2010 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
2011 Chain.getValue(2)).getValue(1);
2012 ResultVals[1] = Chain.getValue(0);
2014 NodeTys.push_back(MVT::f64);
2015 NodeTys.push_back(MVT::f64);
2018 // else fall through
2020 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
2021 InFlag).getValue(1);
2022 ResultVals[0] = Chain.getValue(0);
2024 NodeTys.push_back(Op.Val->getValueType(0));
2030 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
2031 InFlag).getValue(1);
2032 ResultVals[0] = Chain.getValue(0);
2034 NodeTys.push_back(Op.Val->getValueType(0));
2038 NodeTys.push_back(MVT::Other);
2040 // If the function returns void, just return the chain.
2041 if (NumResults == 0)
2044 // Otherwise, merge everything together with a MERGE_VALUES node.
2045 ResultVals[NumResults++] = Chain;
2046 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
2047 ResultVals, NumResults);
2048 return Res.getValue(Op.ResNo);
2051 SDOperand PPCTargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG,
2052 TargetMachine &TM) {
2053 SmallVector<CCValAssign, 16> RVLocs;
2054 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2055 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2056 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2057 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
2059 // If this is the first return lowered for this function, add the regs to the
2060 // liveout set for the function.
2061 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2062 for (unsigned i = 0; i != RVLocs.size(); ++i)
2063 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2066 SDOperand Chain = Op.getOperand(0);
2069 // Copy the result values into the output registers.
2070 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2071 CCValAssign &VA = RVLocs[i];
2072 assert(VA.isRegLoc() && "Can only return in registers!");
2073 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2074 Flag = Chain.getValue(1);
2078 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2080 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2083 SDOperand PPCTargetLowering::LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
2084 const PPCSubtarget &Subtarget) {
2085 // When we pop the dynamic allocation we need to restore the SP link.
2087 // Get the corect type for pointers.
2088 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2090 // Construct the stack pointer operand.
2091 bool IsPPC64 = Subtarget.isPPC64();
2092 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2093 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
2095 // Get the operands for the STACKRESTORE.
2096 SDOperand Chain = Op.getOperand(0);
2097 SDOperand SaveSP = Op.getOperand(1);
2099 // Load the old link SP.
2100 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2102 // Restore the stack pointer.
2103 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2105 // Store the old link SP.
2106 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2109 SDOperand PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
2111 const PPCSubtarget &Subtarget) {
2112 MachineFunction &MF = DAG.getMachineFunction();
2113 bool IsPPC64 = Subtarget.isPPC64();
2114 bool isMachoABI = Subtarget.isMachoABI();
2116 // Get current frame pointer save index. The users of this index will be
2117 // primarily DYNALLOC instructions.
2118 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2119 int FPSI = FI->getFramePointerSaveIndex();
2121 // If the frame pointer save index hasn't been defined yet.
2123 // Find out what the fix offset of the frame pointer save area.
2124 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2126 // Allocate the frame index for frame pointer save area.
2127 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2129 FI->setFramePointerSaveIndex(FPSI);
2133 SDOperand Chain = Op.getOperand(0);
2134 SDOperand Size = Op.getOperand(1);
2136 // Get the corect type for pointers.
2137 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2139 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
2140 DAG.getConstant(0, PtrVT), Size);
2141 // Construct a node for the frame pointer save index.
2142 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
2143 // Build a DYNALLOC node.
2144 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
2145 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2146 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2150 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2152 SDOperand PPCTargetLowering::LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2153 // Not FP? Not a fsel.
2154 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2155 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2158 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2160 // Cannot handle SETEQ/SETNE.
2161 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2163 MVT::ValueType ResVT = Op.getValueType();
2164 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2165 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2166 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2168 // If the RHS of the comparison is a 0.0, we don't need to do the
2169 // subtraction at all.
2170 if (isFloatingPointZero(RHS))
2172 default: break; // SETUO etc aren't handled by fsel.
2176 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2180 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2181 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2182 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2186 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2190 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2191 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2192 return DAG.getNode(PPCISD::FSEL, ResVT,
2193 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2198 default: break; // SETUO etc aren't handled by fsel.
2202 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2203 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2204 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2205 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2209 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2210 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2211 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2212 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2216 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2217 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2218 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2219 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2223 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2224 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2225 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2226 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2231 // FIXME: Split this code up when LegalizeDAGTypes lands.
2232 SDOperand PPCTargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2233 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2234 SDOperand Src = Op.getOperand(0);
2235 if (Src.getValueType() == MVT::f32)
2236 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2239 switch (Op.getValueType()) {
2240 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2242 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2245 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2249 // Convert the FP value to an int value through memory.
2250 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2252 // Emit a store to the stack slot.
2253 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2255 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2257 if (Op.getValueType() == MVT::i32)
2258 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2259 DAG.getConstant(4, FIPtr.getValueType()));
2260 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2263 SDOperand PPCTargetLowering::LowerFP_ROUND_INREG(SDOperand Op,
2264 SelectionDAG &DAG) {
2265 assert(Op.getValueType() == MVT::ppcf128);
2266 SDNode *Node = Op.Val;
2267 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2268 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2269 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2270 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2272 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2273 // of the long double, and puts FPSCR back the way it was. We do not
2274 // actually model FPSCR.
2275 std::vector<MVT::ValueType> NodeTys;
2276 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2278 NodeTys.push_back(MVT::f64); // Return register
2279 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2280 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2281 MFFSreg = Result.getValue(0);
2282 InFlag = Result.getValue(1);
2285 NodeTys.push_back(MVT::Flag); // Returns a flag
2286 Ops[0] = DAG.getConstant(31, MVT::i32);
2288 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2289 InFlag = Result.getValue(0);
2292 NodeTys.push_back(MVT::Flag); // Returns a flag
2293 Ops[0] = DAG.getConstant(30, MVT::i32);
2295 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2296 InFlag = Result.getValue(0);
2299 NodeTys.push_back(MVT::f64); // result of add
2300 NodeTys.push_back(MVT::Flag); // Returns a flag
2304 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2305 FPreg = Result.getValue(0);
2306 InFlag = Result.getValue(1);
2309 NodeTys.push_back(MVT::f64);
2310 Ops[0] = DAG.getConstant(1, MVT::i32);
2314 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2315 FPreg = Result.getValue(0);
2317 // We know the low half is about to be thrown away, so just use something
2319 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2322 SDOperand PPCTargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2323 if (Op.getOperand(0).getValueType() == MVT::i64) {
2324 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2325 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2326 if (Op.getValueType() == MVT::f32)
2327 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2331 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2332 "Unhandled SINT_TO_FP type in custom expander!");
2333 // Since we only generate this in 64-bit mode, we can take advantage of
2334 // 64-bit registers. In particular, sign extend the input value into the
2335 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2336 // then lfd it and fcfid it.
2337 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2338 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2339 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2340 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2342 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2345 // STD the extended value into the stack slot.
2346 MemOperand MO(PseudoSourceValue::getFixedStack(),
2347 MemOperand::MOStore, FrameIdx, 8, 8);
2348 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2349 DAG.getEntryNode(), Ext64, FIdx,
2350 DAG.getMemOperand(MO));
2351 // Load the value as a double.
2352 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2354 // FCFID it and return it.
2355 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2356 if (Op.getValueType() == MVT::f32)
2357 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2361 SDOperand PPCTargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2363 The rounding mode is in bits 30:31 of FPSR, and has the following
2370 FLT_ROUNDS, on the other hand, expects the following:
2377 To perform the conversion, we do:
2378 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2381 MachineFunction &MF = DAG.getMachineFunction();
2382 MVT::ValueType VT = Op.getValueType();
2383 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2384 std::vector<MVT::ValueType> NodeTys;
2385 SDOperand MFFSreg, InFlag;
2387 // Save FP Control Word to register
2388 NodeTys.push_back(MVT::f64); // return register
2389 NodeTys.push_back(MVT::Flag); // unused in this context
2390 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2392 // Save FP register to stack slot
2393 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2394 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2395 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2396 StackSlot, NULL, 0);
2398 // Load FP Control Word from low 32 bits of stack slot.
2399 SDOperand Four = DAG.getConstant(4, PtrVT);
2400 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2401 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2403 // Transform as necessary
2405 DAG.getNode(ISD::AND, MVT::i32,
2406 CWD, DAG.getConstant(3, MVT::i32));
2408 DAG.getNode(ISD::SRL, MVT::i32,
2409 DAG.getNode(ISD::AND, MVT::i32,
2410 DAG.getNode(ISD::XOR, MVT::i32,
2411 CWD, DAG.getConstant(3, MVT::i32)),
2412 DAG.getConstant(3, MVT::i32)),
2413 DAG.getConstant(1, MVT::i8));
2416 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2418 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2419 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2422 SDOperand PPCTargetLowering::LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2423 MVT::ValueType VT = Op.getValueType();
2424 unsigned BitWidth = MVT::getSizeInBits(VT);
2425 assert(Op.getNumOperands() == 3 &&
2426 VT == Op.getOperand(1).getValueType() &&
2429 // Expand into a bunch of logical ops. Note that these ops
2430 // depend on the PPC behavior for oversized shift amounts.
2431 SDOperand Lo = Op.getOperand(0);
2432 SDOperand Hi = Op.getOperand(1);
2433 SDOperand Amt = Op.getOperand(2);
2434 MVT::ValueType AmtVT = Amt.getValueType();
2436 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2437 DAG.getConstant(BitWidth, AmtVT), Amt);
2438 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
2439 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
2440 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2441 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2442 DAG.getConstant(-BitWidth, AmtVT));
2443 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
2444 SDOperand OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2445 SDOperand OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
2446 SDOperand OutOps[] = { OutLo, OutHi };
2447 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2451 SDOperand PPCTargetLowering::LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2452 MVT::ValueType VT = Op.getValueType();
2453 unsigned BitWidth = MVT::getSizeInBits(VT);
2454 assert(Op.getNumOperands() == 3 &&
2455 VT == Op.getOperand(1).getValueType() &&
2458 // Expand into a bunch of logical ops. Note that these ops
2459 // depend on the PPC behavior for oversized shift amounts.
2460 SDOperand Lo = Op.getOperand(0);
2461 SDOperand Hi = Op.getOperand(1);
2462 SDOperand Amt = Op.getOperand(2);
2463 MVT::ValueType AmtVT = Amt.getValueType();
2465 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2466 DAG.getConstant(BitWidth, AmtVT), Amt);
2467 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2468 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2469 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2470 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2471 DAG.getConstant(-BitWidth, AmtVT));
2472 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
2473 SDOperand OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2474 SDOperand OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
2475 SDOperand OutOps[] = { OutLo, OutHi };
2476 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2480 SDOperand PPCTargetLowering::LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2481 MVT::ValueType VT = Op.getValueType();
2482 unsigned BitWidth = MVT::getSizeInBits(VT);
2483 assert(Op.getNumOperands() == 3 &&
2484 VT == Op.getOperand(1).getValueType() &&
2487 // Expand into a bunch of logical ops, followed by a select_cc.
2488 SDOperand Lo = Op.getOperand(0);
2489 SDOperand Hi = Op.getOperand(1);
2490 SDOperand Amt = Op.getOperand(2);
2491 MVT::ValueType AmtVT = Amt.getValueType();
2493 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2494 DAG.getConstant(BitWidth, AmtVT), Amt);
2495 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2496 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2497 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2498 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2499 DAG.getConstant(-BitWidth, AmtVT));
2500 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
2501 SDOperand OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
2502 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
2503 Tmp4, Tmp6, ISD::SETLE);
2504 SDOperand OutOps[] = { OutLo, OutHi };
2505 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2509 //===----------------------------------------------------------------------===//
2510 // Vector related lowering.
2513 // If this is a vector of constants or undefs, get the bits. A bit in
2514 // UndefBits is set if the corresponding element of the vector is an
2515 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2516 // zero. Return true if this is not an array of constants, false if it is.
2518 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2519 uint64_t UndefBits[2]) {
2520 // Start with zero'd results.
2521 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2523 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2524 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2525 SDOperand OpVal = BV->getOperand(i);
2527 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2528 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2530 uint64_t EltBits = 0;
2531 if (OpVal.getOpcode() == ISD::UNDEF) {
2532 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2533 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2535 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2536 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2537 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2538 assert(CN->getValueType(0) == MVT::f32 &&
2539 "Only one legal FP vector type!");
2540 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2542 // Nonconstant element.
2546 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2549 //printf("%llx %llx %llx %llx\n",
2550 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2554 // If this is a splat (repetition) of a value across the whole vector, return
2555 // the smallest size that splats it. For example, "0x01010101010101..." is a
2556 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2557 // SplatSize = 1 byte.
2558 static bool isConstantSplat(const uint64_t Bits128[2],
2559 const uint64_t Undef128[2],
2560 unsigned &SplatBits, unsigned &SplatUndef,
2561 unsigned &SplatSize) {
2563 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2564 // the same as the lower 64-bits, ignoring undefs.
2565 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2566 return false; // Can't be a splat if two pieces don't match.
2568 uint64_t Bits64 = Bits128[0] | Bits128[1];
2569 uint64_t Undef64 = Undef128[0] & Undef128[1];
2571 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2573 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2574 return false; // Can't be a splat if two pieces don't match.
2576 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2577 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2579 // If the top 16-bits are different than the lower 16-bits, ignoring
2580 // undefs, we have an i32 splat.
2581 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2583 SplatUndef = Undef32;
2588 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2589 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2591 // If the top 8-bits are different than the lower 8-bits, ignoring
2592 // undefs, we have an i16 splat.
2593 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2595 SplatUndef = Undef16;
2600 // Otherwise, we have an 8-bit splat.
2601 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2602 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2607 /// BuildSplatI - Build a canonical splati of Val with an element size of
2608 /// SplatSize. Cast the result to VT.
2609 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2610 SelectionDAG &DAG) {
2611 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2613 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2614 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2617 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2619 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2623 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2625 // Build a canonical splat for this value.
2626 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2627 SmallVector<SDOperand, 8> Ops;
2628 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2629 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2630 &Ops[0], Ops.size());
2631 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2634 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2635 /// specified intrinsic ID.
2636 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2638 MVT::ValueType DestVT = MVT::Other) {
2639 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2640 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2641 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2644 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2645 /// specified intrinsic ID.
2646 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2647 SDOperand Op2, SelectionDAG &DAG,
2648 MVT::ValueType DestVT = MVT::Other) {
2649 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2650 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2651 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2655 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2656 /// amount. The result has the specified value type.
2657 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2658 MVT::ValueType VT, SelectionDAG &DAG) {
2659 // Force LHS/RHS to be the right type.
2660 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2661 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2664 for (unsigned i = 0; i != 16; ++i)
2665 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2666 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2667 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2668 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2671 // If this is a case we can't handle, return null and let the default
2672 // expansion code take care of it. If we CAN select this case, and if it
2673 // selects to a single instruction, return Op. Otherwise, if we can codegen
2674 // this case more efficiently than a constant pool load, lower it to the
2675 // sequence of ops that should be used.
2676 SDOperand PPCTargetLowering::LowerBUILD_VECTOR(SDOperand Op,
2677 SelectionDAG &DAG) {
2678 // If this is a vector of constants or undefs, get the bits. A bit in
2679 // UndefBits is set if the corresponding element of the vector is an
2680 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2682 uint64_t VectorBits[2];
2683 uint64_t UndefBits[2];
2684 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2685 return SDOperand(); // Not a constant vector.
2687 // If this is a splat (repetition) of a value across the whole vector, return
2688 // the smallest size that splats it. For example, "0x01010101010101..." is a
2689 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2690 // SplatSize = 1 byte.
2691 unsigned SplatBits, SplatUndef, SplatSize;
2692 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2693 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2695 // First, handle single instruction cases.
2698 if (SplatBits == 0) {
2699 // Canonicalize all zero vectors to be v4i32.
2700 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2701 SDOperand Z = DAG.getConstant(0, MVT::i32);
2702 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2703 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2708 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2709 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2710 if (SextVal >= -16 && SextVal <= 15)
2711 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2714 // Two instruction sequences.
2716 // If this value is in the range [-32,30] and is even, use:
2717 // tmp = VSPLTI[bhw], result = add tmp, tmp
2718 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2719 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2720 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2723 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2724 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2726 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2727 // Make -1 and vspltisw -1:
2728 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2730 // Make the VSLW intrinsic, computing 0x8000_0000.
2731 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2734 // xor by OnesV to invert it.
2735 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2736 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2739 // Check to see if this is a wide variety of vsplti*, binop self cases.
2740 unsigned SplatBitSize = SplatSize*8;
2741 static const signed char SplatCsts[] = {
2742 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2743 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2746 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2747 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2748 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2749 int i = SplatCsts[idx];
2751 // Figure out what shift amount will be used by altivec if shifted by i in
2753 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2755 // vsplti + shl self.
2756 if (SextVal == (i << (int)TypeShiftAmt)) {
2757 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2758 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2759 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2760 Intrinsic::ppc_altivec_vslw
2762 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2763 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2766 // vsplti + srl self.
2767 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2768 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2769 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2770 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2771 Intrinsic::ppc_altivec_vsrw
2773 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2774 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2777 // vsplti + sra self.
2778 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2779 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2780 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2781 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2782 Intrinsic::ppc_altivec_vsraw
2784 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2785 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2788 // vsplti + rol self.
2789 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2790 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2791 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2792 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2793 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2794 Intrinsic::ppc_altivec_vrlw
2796 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2797 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2800 // t = vsplti c, result = vsldoi t, t, 1
2801 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2802 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2803 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2805 // t = vsplti c, result = vsldoi t, t, 2
2806 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2807 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2808 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2810 // t = vsplti c, result = vsldoi t, t, 3
2811 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2812 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2813 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2817 // Three instruction sequences.
2819 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2820 if (SextVal >= 0 && SextVal <= 31) {
2821 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2822 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2823 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2824 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2826 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2827 if (SextVal >= -31 && SextVal <= 0) {
2828 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2829 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2830 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2831 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2838 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2839 /// the specified operations to build the shuffle.
2840 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2841 SDOperand RHS, SelectionDAG &DAG) {
2842 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2843 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2844 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2847 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2859 if (OpNum == OP_COPY) {
2860 if (LHSID == (1*9+2)*9+3) return LHS;
2861 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2865 SDOperand OpLHS, OpRHS;
2866 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2867 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2869 unsigned ShufIdxs[16];
2871 default: assert(0 && "Unknown i32 permute!");
2873 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2874 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2875 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2876 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2879 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2880 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2881 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2882 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2885 for (unsigned i = 0; i != 16; ++i)
2886 ShufIdxs[i] = (i&3)+0;
2889 for (unsigned i = 0; i != 16; ++i)
2890 ShufIdxs[i] = (i&3)+4;
2893 for (unsigned i = 0; i != 16; ++i)
2894 ShufIdxs[i] = (i&3)+8;
2897 for (unsigned i = 0; i != 16; ++i)
2898 ShufIdxs[i] = (i&3)+12;
2901 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2903 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2905 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2908 for (unsigned i = 0; i != 16; ++i)
2909 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2911 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2912 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2915 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2916 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2917 /// return the code it can be lowered into. Worst case, it can always be
2918 /// lowered into a vperm.
2919 SDOperand PPCTargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op,
2920 SelectionDAG &DAG) {
2921 SDOperand V1 = Op.getOperand(0);
2922 SDOperand V2 = Op.getOperand(1);
2923 SDOperand PermMask = Op.getOperand(2);
2925 // Cases that are handled by instructions that take permute immediates
2926 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2927 // selected by the instruction selector.
2928 if (V2.getOpcode() == ISD::UNDEF) {
2929 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2930 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2931 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2932 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2933 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2934 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2935 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2936 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2937 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2938 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2939 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2940 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2945 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2946 // and produce a fixed permutation. If any of these match, do not lower to
2948 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2949 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2950 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2951 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2952 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2953 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2954 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2955 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2956 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2959 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2960 // perfect shuffle table to emit an optimal matching sequence.
2961 unsigned PFIndexes[4];
2962 bool isFourElementShuffle = true;
2963 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2964 unsigned EltNo = 8; // Start out undef.
2965 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2966 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2967 continue; // Undef, ignore it.
2969 unsigned ByteSource =
2970 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2971 if ((ByteSource & 3) != j) {
2972 isFourElementShuffle = false;
2977 EltNo = ByteSource/4;
2978 } else if (EltNo != ByteSource/4) {
2979 isFourElementShuffle = false;
2983 PFIndexes[i] = EltNo;
2986 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
2987 // perfect shuffle vector to determine if it is cost effective to do this as
2988 // discrete instructions, or whether we should use a vperm.
2989 if (isFourElementShuffle) {
2990 // Compute the index in the perfect shuffle table.
2991 unsigned PFTableIndex =
2992 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2994 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2995 unsigned Cost = (PFEntry >> 30);
2997 // Determining when to avoid vperm is tricky. Many things affect the cost
2998 // of vperm, particularly how many times the perm mask needs to be computed.
2999 // For example, if the perm mask can be hoisted out of a loop or is already
3000 // used (perhaps because there are multiple permutes with the same shuffle
3001 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3002 // the loop requires an extra register.
3004 // As a compromise, we only emit discrete instructions if the shuffle can be
3005 // generated in 3 or fewer operations. When we have loop information
3006 // available, if this block is within a loop, we should avoid using vperm
3007 // for 3-operation perms and use a constant pool load instead.
3009 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3012 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3013 // vector that will get spilled to the constant pool.
3014 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3016 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3017 // that it is in input element units, not in bytes. Convert now.
3018 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
3019 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
3021 SmallVector<SDOperand, 16> ResultMask;
3022 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3024 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3027 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
3029 for (unsigned j = 0; j != BytesPerElement; ++j)
3030 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3034 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3035 &ResultMask[0], ResultMask.size());
3036 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3039 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3040 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3041 /// information about the intrinsic.
3042 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
3044 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
3047 switch (IntrinsicID) {
3048 default: return false;
3049 // Comparison predicates.
3050 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3051 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3052 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3053 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3054 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3055 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3056 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3057 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3058 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3059 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3060 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3061 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3062 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3064 // Normal Comparisons.
3065 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3066 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3067 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3068 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3069 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3070 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3071 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3072 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3073 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3074 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3075 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3076 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3077 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3082 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3083 /// lower, do it, otherwise return null.
3084 SDOperand PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op,
3085 SelectionDAG &DAG) {
3086 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3087 // opcode number of the comparison.
3090 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3091 return SDOperand(); // Don't custom lower most intrinsics.
3093 // If this is a non-dot comparison, make the VCMP node and we are done.
3095 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3096 Op.getOperand(1), Op.getOperand(2),
3097 DAG.getConstant(CompareOpc, MVT::i32));
3098 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3101 // Create the PPCISD altivec 'dot' comparison node.
3103 Op.getOperand(2), // LHS
3104 Op.getOperand(3), // RHS
3105 DAG.getConstant(CompareOpc, MVT::i32)
3107 std::vector<MVT::ValueType> VTs;
3108 VTs.push_back(Op.getOperand(2).getValueType());
3109 VTs.push_back(MVT::Flag);
3110 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3112 // Now that we have the comparison, emit a copy from the CR to a GPR.
3113 // This is flagged to the above dot comparison.
3114 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3115 DAG.getRegister(PPC::CR6, MVT::i32),
3116 CompNode.getValue(1));
3118 // Unpack the result based on how the target uses it.
3119 unsigned BitNo; // Bit # of CR6.
3120 bool InvertBit; // Invert result?
3121 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
3122 default: // Can't happen, don't crash on invalid number though.
3123 case 0: // Return the value of the EQ bit of CR6.
3124 BitNo = 0; InvertBit = false;
3126 case 1: // Return the inverted value of the EQ bit of CR6.
3127 BitNo = 0; InvertBit = true;
3129 case 2: // Return the value of the LT bit of CR6.
3130 BitNo = 2; InvertBit = false;
3132 case 3: // Return the inverted value of the LT bit of CR6.
3133 BitNo = 2; InvertBit = true;
3137 // Shift the bit into the low position.
3138 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3139 DAG.getConstant(8-(3-BitNo), MVT::i32));
3141 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3142 DAG.getConstant(1, MVT::i32));
3144 // If we are supposed to, toggle the bit.
3146 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3147 DAG.getConstant(1, MVT::i32));
3151 SDOperand PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op,
3152 SelectionDAG &DAG) {
3153 // Create a stack slot that is 16-byte aligned.
3154 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3155 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3156 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3157 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3159 // Store the input value into Value#0 of the stack slot.
3160 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
3161 Op.getOperand(0), FIdx, NULL, 0);
3163 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3166 SDOperand PPCTargetLowering::LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3167 if (Op.getValueType() == MVT::v4i32) {
3168 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3170 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3171 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3173 SDOperand RHSSwap = // = vrlw RHS, 16
3174 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3176 // Shrinkify inputs to v8i16.
3177 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3178 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3179 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3181 // Low parts multiplied together, generating 32-bit results (we ignore the
3183 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3184 LHS, RHS, DAG, MVT::v4i32);
3186 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3187 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3188 // Shift the high parts up 16 bits.
3189 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3190 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3191 } else if (Op.getValueType() == MVT::v8i16) {
3192 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3194 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3196 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3197 LHS, RHS, Zero, DAG);
3198 } else if (Op.getValueType() == MVT::v16i8) {
3199 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3201 // Multiply the even 8-bit parts, producing 16-bit sums.
3202 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3203 LHS, RHS, DAG, MVT::v8i16);
3204 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3206 // Multiply the odd 8-bit parts, producing 16-bit sums.
3207 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3208 LHS, RHS, DAG, MVT::v8i16);
3209 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3211 // Merge the results together.
3213 for (unsigned i = 0; i != 8; ++i) {
3214 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3215 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3217 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3218 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3220 assert(0 && "Unknown mul to lower!");
3225 /// LowerOperation - Provide custom lowering hooks for some operations.
3227 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3228 switch (Op.getOpcode()) {
3229 default: assert(0 && "Wasn't expecting to be able to lower this!");
3230 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3231 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3232 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3233 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3234 case ISD::SETCC: return LowerSETCC(Op, DAG);
3236 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3237 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3240 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3241 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3243 case ISD::FORMAL_ARGUMENTS:
3244 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3245 VarArgsStackOffset, VarArgsNumGPR,
3246 VarArgsNumFPR, PPCSubTarget);
3248 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
3249 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3250 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3251 case ISD::DYNAMIC_STACKALLOC:
3252 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3254 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3255 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3256 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3257 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3258 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3260 // Lower 64-bit shifts.
3261 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3262 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3263 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3265 // Vector-related lowering.
3266 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3267 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3268 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3269 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3270 case ISD::MUL: return LowerMUL(Op, DAG);
3272 // Frame & Return address.
3273 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3274 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3279 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3280 switch (N->getOpcode()) {
3281 default: assert(0 && "Wasn't expecting to be able to lower this!");
3282 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3287 //===----------------------------------------------------------------------===//
3288 // Other Lowering Code
3289 //===----------------------------------------------------------------------===//
3292 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3293 MachineBasicBlock *BB) {
3294 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3295 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3296 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3297 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3298 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3299 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3300 "Unexpected instr type to insert");
3302 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3303 // control-flow pattern. The incoming instruction knows the destination vreg
3304 // to set, the condition code register to branch on, the true/false values to
3305 // select between, and a branch opcode to use.
3306 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3307 ilist<MachineBasicBlock>::iterator It = BB;
3313 // cmpTY ccX, r1, r2
3315 // fallthrough --> copy0MBB
3316 MachineBasicBlock *thisMBB = BB;
3317 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3318 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3319 unsigned SelectPred = MI->getOperand(4).getImm();
3320 BuildMI(BB, TII->get(PPC::BCC))
3321 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3322 MachineFunction *F = BB->getParent();
3323 F->getBasicBlockList().insert(It, copy0MBB);
3324 F->getBasicBlockList().insert(It, sinkMBB);
3325 // Update machine-CFG edges by first adding all successors of the current
3326 // block to the new block which will contain the Phi node for the select.
3327 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3328 e = BB->succ_end(); i != e; ++i)
3329 sinkMBB->addSuccessor(*i);
3330 // Next, remove all successors of the current block, and add the true
3331 // and fallthrough blocks as its successors.
3332 while(!BB->succ_empty())
3333 BB->removeSuccessor(BB->succ_begin());
3334 BB->addSuccessor(copy0MBB);
3335 BB->addSuccessor(sinkMBB);
3338 // %FalseValue = ...
3339 // # fallthrough to sinkMBB
3342 // Update machine-CFG edges
3343 BB->addSuccessor(sinkMBB);
3346 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3349 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3350 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3351 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3353 delete MI; // The pseudo instruction is gone now.
3357 //===----------------------------------------------------------------------===//
3358 // Target Optimization Hooks
3359 //===----------------------------------------------------------------------===//
3361 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3362 DAGCombinerInfo &DCI) const {
3363 TargetMachine &TM = getTargetMachine();
3364 SelectionDAG &DAG = DCI.DAG;
3365 switch (N->getOpcode()) {
3368 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3369 if (C->getValue() == 0) // 0 << V -> 0.
3370 return N->getOperand(0);
3374 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3375 if (C->getValue() == 0) // 0 >>u V -> 0.
3376 return N->getOperand(0);
3380 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3381 if (C->getValue() == 0 || // 0 >>s V -> 0.
3382 C->isAllOnesValue()) // -1 >>s V -> -1.
3383 return N->getOperand(0);
3387 case ISD::SINT_TO_FP:
3388 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3389 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3390 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3391 // We allow the src/dst to be either f32/f64, but the intermediate
3392 // type must be i64.
3393 if (N->getOperand(0).getValueType() == MVT::i64 &&
3394 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3395 SDOperand Val = N->getOperand(0).getOperand(0);
3396 if (Val.getValueType() == MVT::f32) {
3397 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3398 DCI.AddToWorklist(Val.Val);
3401 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3402 DCI.AddToWorklist(Val.Val);
3403 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3404 DCI.AddToWorklist(Val.Val);
3405 if (N->getValueType(0) == MVT::f32) {
3406 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3407 DAG.getIntPtrConstant(0));
3408 DCI.AddToWorklist(Val.Val);
3411 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3412 // If the intermediate type is i32, we can avoid the load/store here
3419 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3420 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3421 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3422 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3423 N->getOperand(1).getValueType() == MVT::i32 &&
3424 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3425 SDOperand Val = N->getOperand(1).getOperand(0);
3426 if (Val.getValueType() == MVT::f32) {
3427 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3428 DCI.AddToWorklist(Val.Val);
3430 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3431 DCI.AddToWorklist(Val.Val);
3433 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3434 N->getOperand(2), N->getOperand(3));
3435 DCI.AddToWorklist(Val.Val);
3439 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3440 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3441 N->getOperand(1).Val->hasOneUse() &&
3442 (N->getOperand(1).getValueType() == MVT::i32 ||
3443 N->getOperand(1).getValueType() == MVT::i16)) {
3444 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3445 // Do an any-extend to 32-bits if this is a half-word input.
3446 if (BSwapOp.getValueType() == MVT::i16)
3447 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3449 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3450 N->getOperand(2), N->getOperand(3),
3451 DAG.getValueType(N->getOperand(1).getValueType()));
3455 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3456 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3457 N->getOperand(0).hasOneUse() &&
3458 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3459 SDOperand Load = N->getOperand(0);
3460 LoadSDNode *LD = cast<LoadSDNode>(Load);
3461 // Create the byte-swapping load.
3462 std::vector<MVT::ValueType> VTs;
3463 VTs.push_back(MVT::i32);
3464 VTs.push_back(MVT::Other);
3465 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3467 LD->getChain(), // Chain
3468 LD->getBasePtr(), // Ptr
3470 DAG.getValueType(N->getValueType(0)) // VT
3472 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3474 // If this is an i16 load, insert the truncate.
3475 SDOperand ResVal = BSLoad;
3476 if (N->getValueType(0) == MVT::i16)
3477 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3479 // First, combine the bswap away. This makes the value produced by the
3481 DCI.CombineTo(N, ResVal);
3483 // Next, combine the load away, we give it a bogus result value but a real
3484 // chain result. The result value is dead because the bswap is dead.
3485 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3487 // Return N so it doesn't get rechecked!
3488 return SDOperand(N, 0);
3492 case PPCISD::VCMP: {
3493 // If a VCMPo node already exists with exactly the same operands as this
3494 // node, use its result instead of this node (VCMPo computes both a CR6 and
3495 // a normal output).
3497 if (!N->getOperand(0).hasOneUse() &&
3498 !N->getOperand(1).hasOneUse() &&
3499 !N->getOperand(2).hasOneUse()) {
3501 // Scan all of the users of the LHS, looking for VCMPo's that match.
3502 SDNode *VCMPoNode = 0;
3504 SDNode *LHSN = N->getOperand(0).Val;
3505 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3507 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3508 (*UI)->getOperand(1) == N->getOperand(1) &&
3509 (*UI)->getOperand(2) == N->getOperand(2) &&
3510 (*UI)->getOperand(0) == N->getOperand(0)) {
3515 // If there is no VCMPo node, or if the flag value has a single use, don't
3517 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3520 // Look at the (necessarily single) use of the flag value. If it has a
3521 // chain, this transformation is more complex. Note that multiple things
3522 // could use the value result, which we should ignore.
3523 SDNode *FlagUser = 0;
3524 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3525 FlagUser == 0; ++UI) {
3526 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3528 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3529 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3536 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3537 // give up for right now.
3538 if (FlagUser->getOpcode() == PPCISD::MFCR)
3539 return SDOperand(VCMPoNode, 0);
3544 // If this is a branch on an altivec predicate comparison, lower this so
3545 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3546 // lowering is done pre-legalize, because the legalizer lowers the predicate
3547 // compare down to code that is difficult to reassemble.
3548 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3549 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3553 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3554 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3555 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3556 assert(isDot && "Can't compare against a vector result!");
3558 // If this is a comparison against something other than 0/1, then we know
3559 // that the condition is never/always true.
3560 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3561 if (Val != 0 && Val != 1) {
3562 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3563 return N->getOperand(0);
3564 // Always !=, turn it into an unconditional branch.
3565 return DAG.getNode(ISD::BR, MVT::Other,
3566 N->getOperand(0), N->getOperand(4));
3569 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3571 // Create the PPCISD altivec 'dot' comparison node.
3572 std::vector<MVT::ValueType> VTs;
3574 LHS.getOperand(2), // LHS of compare
3575 LHS.getOperand(3), // RHS of compare
3576 DAG.getConstant(CompareOpc, MVT::i32)
3578 VTs.push_back(LHS.getOperand(2).getValueType());
3579 VTs.push_back(MVT::Flag);
3580 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3582 // Unpack the result based on how the target uses it.
3583 PPC::Predicate CompOpc;
3584 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3585 default: // Can't happen, don't crash on invalid number though.
3586 case 0: // Branch on the value of the EQ bit of CR6.
3587 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3589 case 1: // Branch on the inverted value of the EQ bit of CR6.
3590 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3592 case 2: // Branch on the value of the LT bit of CR6.
3593 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3595 case 3: // Branch on the inverted value of the LT bit of CR6.
3596 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3600 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3601 DAG.getConstant(CompOpc, MVT::i32),
3602 DAG.getRegister(PPC::CR6, MVT::i32),
3603 N->getOperand(4), CompNode.getValue(1));
3612 //===----------------------------------------------------------------------===//
3613 // Inline Assembly Support
3614 //===----------------------------------------------------------------------===//
3616 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3620 const SelectionDAG &DAG,
3621 unsigned Depth) const {
3622 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3623 switch (Op.getOpcode()) {
3625 case PPCISD::LBRX: {
3626 // lhbrx is known to have the top bits cleared out.
3627 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3628 KnownZero = 0xFFFF0000;
3631 case ISD::INTRINSIC_WO_CHAIN: {
3632 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3634 case Intrinsic::ppc_altivec_vcmpbfp_p:
3635 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3636 case Intrinsic::ppc_altivec_vcmpequb_p:
3637 case Intrinsic::ppc_altivec_vcmpequh_p:
3638 case Intrinsic::ppc_altivec_vcmpequw_p:
3639 case Intrinsic::ppc_altivec_vcmpgefp_p:
3640 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3641 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3642 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3643 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3644 case Intrinsic::ppc_altivec_vcmpgtub_p:
3645 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3646 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3647 KnownZero = ~1U; // All bits but the low one are known to be zero.
3655 /// getConstraintType - Given a constraint, return the type of
3656 /// constraint it is for this target.
3657 PPCTargetLowering::ConstraintType
3658 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3659 if (Constraint.size() == 1) {
3660 switch (Constraint[0]) {
3667 return C_RegisterClass;
3670 return TargetLowering::getConstraintType(Constraint);
3673 std::pair<unsigned, const TargetRegisterClass*>
3674 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3675 MVT::ValueType VT) const {
3676 if (Constraint.size() == 1) {
3677 // GCC RS6000 Constraint Letters
3678 switch (Constraint[0]) {
3681 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3682 return std::make_pair(0U, PPC::G8RCRegisterClass);
3683 return std::make_pair(0U, PPC::GPRCRegisterClass);
3686 return std::make_pair(0U, PPC::F4RCRegisterClass);
3687 else if (VT == MVT::f64)
3688 return std::make_pair(0U, PPC::F8RCRegisterClass);
3691 return std::make_pair(0U, PPC::VRRCRegisterClass);
3693 return std::make_pair(0U, PPC::CRRCRegisterClass);
3697 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3701 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3702 /// vector. If it is invalid, don't add anything to Ops.
3703 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3704 std::vector<SDOperand>&Ops,
3705 SelectionDAG &DAG) {
3706 SDOperand Result(0,0);
3717 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3718 if (!CST) return; // Must be an immediate to match.
3719 unsigned Value = CST->getValue();
3721 default: assert(0 && "Unknown constraint letter!");
3722 case 'I': // "I" is a signed 16-bit constant.
3723 if ((short)Value == (int)Value)
3724 Result = DAG.getTargetConstant(Value, Op.getValueType());
3726 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3727 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3728 if ((short)Value == 0)
3729 Result = DAG.getTargetConstant(Value, Op.getValueType());
3731 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3732 if ((Value >> 16) == 0)
3733 Result = DAG.getTargetConstant(Value, Op.getValueType());
3735 case 'M': // "M" is a constant that is greater than 31.
3737 Result = DAG.getTargetConstant(Value, Op.getValueType());
3739 case 'N': // "N" is a positive constant that is an exact power of two.
3740 if ((int)Value > 0 && isPowerOf2_32(Value))
3741 Result = DAG.getTargetConstant(Value, Op.getValueType());
3743 case 'O': // "O" is the constant zero.
3745 Result = DAG.getTargetConstant(Value, Op.getValueType());
3747 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3748 if ((short)-Value == (int)-Value)
3749 Result = DAG.getTargetConstant(Value, Op.getValueType());
3757 Ops.push_back(Result);
3761 // Handle standard constraint letters.
3762 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3765 // isLegalAddressingMode - Return true if the addressing mode represented
3766 // by AM is legal for this target, for a load/store of the specified type.
3767 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3768 const Type *Ty) const {
3769 // FIXME: PPC does not allow r+i addressing modes for vectors!
3771 // PPC allows a sign-extended 16-bit immediate field.
3772 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3775 // No global is ever allowed as a base.
3779 // PPC only support r+r,
3781 case 0: // "r+i" or just "i", depending on HasBaseReg.
3784 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3786 // Otherwise we have r+r or r+i.
3789 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3791 // Allow 2*r as r+r.
3794 // No other scales are supported.
3801 /// isLegalAddressImmediate - Return true if the integer value can be used
3802 /// as the offset of the target addressing mode for load / store of the
3804 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3805 // PPC allows a sign-extended 16-bit immediate field.
3806 return (V > -(1 << 16) && V < (1 << 16)-1);
3809 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3813 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3814 // Depths > 0 not supported yet!
3815 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3818 MachineFunction &MF = DAG.getMachineFunction();
3819 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3820 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3822 bool isPPC64 = PPCSubTarget.isPPC64();
3824 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3826 // Set up a frame object for the return address.
3827 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3829 // Remember it for next time.
3830 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3832 // Make sure the function really does not optimize away the store of the RA
3834 FuncInfo->setLRStoreRequired();
3837 // Just load the return address off the stack.
3838 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3839 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3842 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3843 // Depths > 0 not supported yet!
3844 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3847 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3848 bool isPPC64 = PtrVT == MVT::i64;
3850 MachineFunction &MF = DAG.getMachineFunction();
3851 MachineFrameInfo *MFI = MF.getFrameInfo();
3852 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3853 && MFI->getStackSize();
3856 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3859 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,