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);
237 // 32-bit PowerPC wants to expand i64 shifts itself.
238 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
239 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
240 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
243 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
244 // First set operation action for all vector types to expand. Then we
245 // will selectively turn on ones that can be effectively codegen'd.
246 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
247 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
248 // add/sub are legal for all supported vector VT's.
249 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
250 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
252 // We promote all shuffles to v16i8.
253 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
254 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
256 // We promote all non-typed operations to v4i32.
257 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
258 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
259 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
260 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
261 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
262 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
263 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
264 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
265 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
266 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
267 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
268 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
270 // No other operations are legal.
271 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
272 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
273 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
274 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
275 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
276 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
277 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
278 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
279 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
280 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
281 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
282 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
283 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
284 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
285 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
286 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
287 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
288 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
289 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
292 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
293 // with merges, splats, etc.
294 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
296 setOperationAction(ISD::AND , MVT::v4i32, Legal);
297 setOperationAction(ISD::OR , MVT::v4i32, Legal);
298 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
299 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
300 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
301 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
303 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
304 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
305 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
306 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
308 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
309 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
310 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
311 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
313 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
314 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
316 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
317 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
318 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
319 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
322 setSetCCResultType(MVT::i32);
323 setShiftAmountType(MVT::i32);
324 setSetCCResultContents(ZeroOrOneSetCCResult);
326 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
327 setStackPointerRegisterToSaveRestore(PPC::X1);
328 setExceptionPointerRegister(PPC::X3);
329 setExceptionSelectorRegister(PPC::X4);
331 setStackPointerRegisterToSaveRestore(PPC::R1);
332 setExceptionPointerRegister(PPC::R3);
333 setExceptionSelectorRegister(PPC::R4);
336 // We have target-specific dag combine patterns for the following nodes:
337 setTargetDAGCombine(ISD::SINT_TO_FP);
338 setTargetDAGCombine(ISD::STORE);
339 setTargetDAGCombine(ISD::BR_CC);
340 setTargetDAGCombine(ISD::BSWAP);
342 // Darwin long double math library functions have $LDBL128 appended.
343 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
344 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
345 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
346 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
347 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
348 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
351 computeRegisterProperties();
354 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
357 case PPCISD::FSEL: return "PPCISD::FSEL";
358 case PPCISD::FCFID: return "PPCISD::FCFID";
359 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
360 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
361 case PPCISD::STFIWX: return "PPCISD::STFIWX";
362 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
363 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
364 case PPCISD::VPERM: return "PPCISD::VPERM";
365 case PPCISD::Hi: return "PPCISD::Hi";
366 case PPCISD::Lo: return "PPCISD::Lo";
367 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
368 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
369 case PPCISD::SRL: return "PPCISD::SRL";
370 case PPCISD::SRA: return "PPCISD::SRA";
371 case PPCISD::SHL: return "PPCISD::SHL";
372 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
373 case PPCISD::STD_32: return "PPCISD::STD_32";
374 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
375 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
376 case PPCISD::MTCTR: return "PPCISD::MTCTR";
377 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
378 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
379 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
380 case PPCISD::MFCR: return "PPCISD::MFCR";
381 case PPCISD::VCMP: return "PPCISD::VCMP";
382 case PPCISD::VCMPo: return "PPCISD::VCMPo";
383 case PPCISD::LBRX: return "PPCISD::LBRX";
384 case PPCISD::STBRX: return "PPCISD::STBRX";
385 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
386 case PPCISD::MFFS: return "PPCISD::MFFS";
387 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
388 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
389 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
390 case PPCISD::MTFSF: return "PPCISD::MTFSF";
394 //===----------------------------------------------------------------------===//
395 // Node matching predicates, for use by the tblgen matching code.
396 //===----------------------------------------------------------------------===//
398 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
399 static bool isFloatingPointZero(SDOperand Op) {
400 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
401 return CFP->getValueAPF().isZero();
402 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
403 // Maybe this has already been legalized into the constant pool?
404 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
405 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
406 return CFP->getValueAPF().isZero();
411 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
412 /// true if Op is undef or if it matches the specified value.
413 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
414 return Op.getOpcode() == ISD::UNDEF ||
415 cast<ConstantSDNode>(Op)->getValue() == Val;
418 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
419 /// VPKUHUM instruction.
420 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
422 for (unsigned i = 0; i != 16; ++i)
423 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
426 for (unsigned i = 0; i != 8; ++i)
427 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
428 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
434 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
435 /// VPKUWUM instruction.
436 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
438 for (unsigned i = 0; i != 16; i += 2)
439 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
440 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
443 for (unsigned i = 0; i != 8; i += 2)
444 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
445 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
446 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
447 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
453 /// isVMerge - Common function, used to match vmrg* shuffles.
455 static bool isVMerge(SDNode *N, unsigned UnitSize,
456 unsigned LHSStart, unsigned RHSStart) {
457 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
458 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
459 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
460 "Unsupported merge size!");
462 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
463 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
464 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
465 LHSStart+j+i*UnitSize) ||
466 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
467 RHSStart+j+i*UnitSize))
473 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
474 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
475 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
477 return isVMerge(N, UnitSize, 8, 24);
478 return isVMerge(N, UnitSize, 8, 8);
481 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
482 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
483 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
485 return isVMerge(N, UnitSize, 0, 16);
486 return isVMerge(N, UnitSize, 0, 0);
490 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
491 /// amount, otherwise return -1.
492 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
493 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
494 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
495 // Find the first non-undef value in the shuffle mask.
497 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
500 if (i == 16) return -1; // all undef.
502 // Otherwise, check to see if the rest of the elements are consequtively
503 // numbered from this value.
504 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
505 if (ShiftAmt < i) return -1;
509 // Check the rest of the elements to see if they are consequtive.
510 for (++i; i != 16; ++i)
511 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
514 // Check the rest of the elements to see if they are consequtive.
515 for (++i; i != 16; ++i)
516 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
523 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
524 /// specifies a splat of a single element that is suitable for input to
525 /// VSPLTB/VSPLTH/VSPLTW.
526 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
527 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
528 N->getNumOperands() == 16 &&
529 (EltSize == 1 || EltSize == 2 || EltSize == 4));
531 // This is a splat operation if each element of the permute is the same, and
532 // if the value doesn't reference the second vector.
533 unsigned ElementBase = 0;
534 SDOperand Elt = N->getOperand(0);
535 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
536 ElementBase = EltV->getValue();
538 return false; // FIXME: Handle UNDEF elements too!
540 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
543 // Check that they are consequtive.
544 for (unsigned i = 1; i != EltSize; ++i) {
545 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
546 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
550 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
551 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
552 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
553 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
554 "Invalid VECTOR_SHUFFLE mask!");
555 for (unsigned j = 0; j != EltSize; ++j)
556 if (N->getOperand(i+j) != N->getOperand(j))
563 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
565 bool PPC::isAllNegativeZeroVector(SDNode *N) {
566 assert(N->getOpcode() == ISD::BUILD_VECTOR);
567 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
568 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
569 return CFP->getValueAPF().isNegZero();
573 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
574 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
575 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
576 assert(isSplatShuffleMask(N, EltSize));
577 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
580 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
581 /// by using a vspltis[bhw] instruction of the specified element size, return
582 /// the constant being splatted. The ByteSize field indicates the number of
583 /// bytes of each element [124] -> [bhw].
584 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
585 SDOperand OpVal(0, 0);
587 // If ByteSize of the splat is bigger than the element size of the
588 // build_vector, then we have a case where we are checking for a splat where
589 // multiple elements of the buildvector are folded together into a single
590 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
591 unsigned EltSize = 16/N->getNumOperands();
592 if (EltSize < ByteSize) {
593 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
594 SDOperand UniquedVals[4];
595 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
597 // See if all of the elements in the buildvector agree across.
598 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
599 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
600 // If the element isn't a constant, bail fully out.
601 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
604 if (UniquedVals[i&(Multiple-1)].Val == 0)
605 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
606 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
607 return SDOperand(); // no match.
610 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
611 // either constant or undef values that are identical for each chunk. See
612 // if these chunks can form into a larger vspltis*.
614 // Check to see if all of the leading entries are either 0 or -1. If
615 // neither, then this won't fit into the immediate field.
616 bool LeadingZero = true;
617 bool LeadingOnes = true;
618 for (unsigned i = 0; i != Multiple-1; ++i) {
619 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
621 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
622 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
624 // Finally, check the least significant entry.
626 if (UniquedVals[Multiple-1].Val == 0)
627 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
628 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
630 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
633 if (UniquedVals[Multiple-1].Val == 0)
634 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
635 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
636 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
637 return DAG.getTargetConstant(Val, MVT::i32);
643 // Check to see if this buildvec has a single non-undef value in its elements.
644 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
645 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
647 OpVal = N->getOperand(i);
648 else if (OpVal != N->getOperand(i))
652 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
654 unsigned ValSizeInBytes = 0;
656 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
657 Value = CN->getValue();
658 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
659 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
660 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
661 Value = FloatToBits(CN->getValueAPF().convertToFloat());
665 // If the splat value is larger than the element value, then we can never do
666 // this splat. The only case that we could fit the replicated bits into our
667 // immediate field for would be zero, and we prefer to use vxor for it.
668 if (ValSizeInBytes < ByteSize) return SDOperand();
670 // If the element value is larger than the splat value, cut it in half and
671 // check to see if the two halves are equal. Continue doing this until we
672 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
673 while (ValSizeInBytes > ByteSize) {
674 ValSizeInBytes >>= 1;
676 // If the top half equals the bottom half, we're still ok.
677 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
678 (Value & ((1 << (8*ValSizeInBytes))-1)))
682 // Properly sign extend the value.
683 int ShAmt = (4-ByteSize)*8;
684 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
686 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
687 if (MaskVal == 0) return SDOperand();
689 // Finally, if this value fits in a 5 bit sext field, return it
690 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
691 return DAG.getTargetConstant(MaskVal, MVT::i32);
695 //===----------------------------------------------------------------------===//
696 // Addressing Mode Selection
697 //===----------------------------------------------------------------------===//
699 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
700 /// or 64-bit immediate, and if the value can be accurately represented as a
701 /// sign extension from a 16-bit value. If so, this returns true and the
703 static bool isIntS16Immediate(SDNode *N, short &Imm) {
704 if (N->getOpcode() != ISD::Constant)
707 Imm = (short)cast<ConstantSDNode>(N)->getValue();
708 if (N->getValueType(0) == MVT::i32)
709 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
711 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
713 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
714 return isIntS16Immediate(Op.Val, Imm);
718 /// SelectAddressRegReg - Given the specified addressed, check to see if it
719 /// can be represented as an indexed [r+r] operation. Returns false if it
720 /// can be more efficiently represented with [r+imm].
721 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
725 if (N.getOpcode() == ISD::ADD) {
726 if (isIntS16Immediate(N.getOperand(1), imm))
728 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
731 Base = N.getOperand(0);
732 Index = N.getOperand(1);
734 } else if (N.getOpcode() == ISD::OR) {
735 if (isIntS16Immediate(N.getOperand(1), imm))
736 return false; // r+i can fold it if we can.
738 // If this is an or of disjoint bitfields, we can codegen this as an add
739 // (for better address arithmetic) if the LHS and RHS of the OR are provably
741 APInt LHSKnownZero, LHSKnownOne;
742 APInt RHSKnownZero, RHSKnownOne;
743 DAG.ComputeMaskedBits(N.getOperand(0),
744 APInt::getAllOnesValue(N.getOperand(0)
745 .getValueSizeInBits()),
746 LHSKnownZero, LHSKnownOne);
748 if (LHSKnownZero.getBoolValue()) {
749 DAG.ComputeMaskedBits(N.getOperand(1),
750 APInt::getAllOnesValue(N.getOperand(1)
751 .getValueSizeInBits()),
752 RHSKnownZero, RHSKnownOne);
753 // If all of the bits are known zero on the LHS or RHS, the add won't
755 if (~(LHSKnownZero | RHSKnownZero) == 0) {
756 Base = N.getOperand(0);
757 Index = N.getOperand(1);
766 /// Returns true if the address N can be represented by a base register plus
767 /// a signed 16-bit displacement [r+imm], and if it is not better
768 /// represented as reg+reg.
769 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
770 SDOperand &Base, SelectionDAG &DAG){
771 // If this can be more profitably realized as r+r, fail.
772 if (SelectAddressRegReg(N, Disp, Base, DAG))
775 if (N.getOpcode() == ISD::ADD) {
777 if (isIntS16Immediate(N.getOperand(1), imm)) {
778 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
779 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
780 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
782 Base = N.getOperand(0);
784 return true; // [r+i]
785 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
786 // Match LOAD (ADD (X, Lo(G))).
787 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
788 && "Cannot handle constant offsets yet!");
789 Disp = N.getOperand(1).getOperand(0); // The global address.
790 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
791 Disp.getOpcode() == ISD::TargetConstantPool ||
792 Disp.getOpcode() == ISD::TargetJumpTable);
793 Base = N.getOperand(0);
794 return true; // [&g+r]
796 } else if (N.getOpcode() == ISD::OR) {
798 if (isIntS16Immediate(N.getOperand(1), imm)) {
799 // If this is an or of disjoint bitfields, we can codegen this as an add
800 // (for better address arithmetic) if the LHS and RHS of the OR are
801 // provably disjoint.
802 APInt LHSKnownZero, LHSKnownOne;
803 DAG.ComputeMaskedBits(N.getOperand(0),
804 APInt::getAllOnesValue(32),
805 LHSKnownZero, LHSKnownOne);
806 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
807 // If all of the bits are known zero on the LHS or RHS, the add won't
809 Base = N.getOperand(0);
810 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
814 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
815 // Loading from a constant address.
817 // If this address fits entirely in a 16-bit sext immediate field, codegen
820 if (isIntS16Immediate(CN, Imm)) {
821 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
822 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
826 // Handle 32-bit sext immediates with LIS + addr mode.
827 if (CN->getValueType(0) == MVT::i32 ||
828 (int64_t)CN->getValue() == (int)CN->getValue()) {
829 int Addr = (int)CN->getValue();
831 // Otherwise, break this down into an LIS + disp.
832 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
834 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
835 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
836 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
841 Disp = DAG.getTargetConstant(0, getPointerTy());
842 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
843 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
846 return true; // [r+0]
849 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
850 /// represented as an indexed [r+r] operation.
851 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
854 // Check to see if we can easily represent this as an [r+r] address. This
855 // will fail if it thinks that the address is more profitably represented as
856 // reg+imm, e.g. where imm = 0.
857 if (SelectAddressRegReg(N, Base, Index, DAG))
860 // If the operand is an addition, always emit this as [r+r], since this is
861 // better (for code size, and execution, as the memop does the add for free)
862 // than emitting an explicit add.
863 if (N.getOpcode() == ISD::ADD) {
864 Base = N.getOperand(0);
865 Index = N.getOperand(1);
869 // Otherwise, do it the hard way, using R0 as the base register.
870 Base = DAG.getRegister(PPC::R0, N.getValueType());
875 /// SelectAddressRegImmShift - Returns true if the address N can be
876 /// represented by a base register plus a signed 14-bit displacement
877 /// [r+imm*4]. Suitable for use by STD and friends.
878 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
881 // If this can be more profitably realized as r+r, fail.
882 if (SelectAddressRegReg(N, Disp, Base, DAG))
885 if (N.getOpcode() == ISD::ADD) {
887 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
888 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
889 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
890 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
892 Base = N.getOperand(0);
894 return true; // [r+i]
895 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
896 // Match LOAD (ADD (X, Lo(G))).
897 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
898 && "Cannot handle constant offsets yet!");
899 Disp = N.getOperand(1).getOperand(0); // The global address.
900 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
901 Disp.getOpcode() == ISD::TargetConstantPool ||
902 Disp.getOpcode() == ISD::TargetJumpTable);
903 Base = N.getOperand(0);
904 return true; // [&g+r]
906 } else if (N.getOpcode() == ISD::OR) {
908 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
909 // If this is an or of disjoint bitfields, we can codegen this as an add
910 // (for better address arithmetic) if the LHS and RHS of the OR are
911 // provably disjoint.
912 APInt LHSKnownZero, LHSKnownOne;
913 DAG.ComputeMaskedBits(N.getOperand(0),
914 APInt::getAllOnesValue(32),
915 LHSKnownZero, LHSKnownOne);
916 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
917 // If all of the bits are known zero on the LHS or RHS, the add won't
919 Base = N.getOperand(0);
920 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
924 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
925 // Loading from a constant address. Verify low two bits are clear.
926 if ((CN->getValue() & 3) == 0) {
927 // If this address fits entirely in a 14-bit sext immediate field, codegen
930 if (isIntS16Immediate(CN, Imm)) {
931 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
932 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
936 // Fold the low-part of 32-bit absolute addresses into addr mode.
937 if (CN->getValueType(0) == MVT::i32 ||
938 (int64_t)CN->getValue() == (int)CN->getValue()) {
939 int Addr = (int)CN->getValue();
941 // Otherwise, break this down into an LIS + disp.
942 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
944 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
945 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
946 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
952 Disp = DAG.getTargetConstant(0, getPointerTy());
953 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
954 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
957 return true; // [r+0]
961 /// getPreIndexedAddressParts - returns true by value, base pointer and
962 /// offset pointer and addressing mode by reference if the node's address
963 /// can be legally represented as pre-indexed load / store address.
964 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
966 ISD::MemIndexedMode &AM,
968 // Disabled by default for now.
969 if (!EnablePPCPreinc) return false;
973 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
974 Ptr = LD->getBasePtr();
975 VT = LD->getMemoryVT();
977 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
979 Ptr = ST->getBasePtr();
980 VT = ST->getMemoryVT();
984 // PowerPC doesn't have preinc load/store instructions for vectors.
985 if (MVT::isVector(VT))
988 // TODO: Check reg+reg first.
990 // LDU/STU use reg+imm*4, others use reg+imm.
991 if (VT != MVT::i64) {
993 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
997 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1001 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1002 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1003 // sext i32 to i64 when addr mode is r+i.
1004 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1005 LD->getExtensionType() == ISD::SEXTLOAD &&
1006 isa<ConstantSDNode>(Offset))
1014 //===----------------------------------------------------------------------===//
1015 // LowerOperation implementation
1016 //===----------------------------------------------------------------------===//
1018 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
1019 MVT::ValueType PtrVT = Op.getValueType();
1020 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1021 Constant *C = CP->getConstVal();
1022 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1023 SDOperand Zero = DAG.getConstant(0, PtrVT);
1025 const TargetMachine &TM = DAG.getTarget();
1027 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1028 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1030 // If this is a non-darwin platform, we don't support non-static relo models
1032 if (TM.getRelocationModel() == Reloc::Static ||
1033 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1034 // Generate non-pic code that has direct accesses to the constant pool.
1035 // The address of the global is just (hi(&g)+lo(&g)).
1036 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1039 if (TM.getRelocationModel() == Reloc::PIC_) {
1040 // With PIC, the first instruction is actually "GR+hi(&G)".
1041 Hi = DAG.getNode(ISD::ADD, PtrVT,
1042 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1045 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1049 static SDOperand LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1050 MVT::ValueType PtrVT = Op.getValueType();
1051 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1052 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1053 SDOperand Zero = DAG.getConstant(0, PtrVT);
1055 const TargetMachine &TM = DAG.getTarget();
1057 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1058 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1060 // If this is a non-darwin platform, we don't support non-static relo models
1062 if (TM.getRelocationModel() == Reloc::Static ||
1063 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1064 // Generate non-pic code that has direct accesses to the constant pool.
1065 // The address of the global is just (hi(&g)+lo(&g)).
1066 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1069 if (TM.getRelocationModel() == Reloc::PIC_) {
1070 // With PIC, the first instruction is actually "GR+hi(&G)".
1071 Hi = DAG.getNode(ISD::ADD, PtrVT,
1072 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1075 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1079 static SDOperand LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
1080 assert(0 && "TLS not implemented for PPC.");
1083 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
1084 MVT::ValueType PtrVT = Op.getValueType();
1085 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1086 GlobalValue *GV = GSDN->getGlobal();
1087 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1088 // If it's a debug information descriptor, don't mess with it.
1089 if (DAG.isVerifiedDebugInfoDesc(Op))
1091 SDOperand Zero = DAG.getConstant(0, PtrVT);
1093 const TargetMachine &TM = DAG.getTarget();
1095 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1096 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1098 // If this is a non-darwin platform, we don't support non-static relo models
1100 if (TM.getRelocationModel() == Reloc::Static ||
1101 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1102 // Generate non-pic code that has direct accesses to globals.
1103 // The address of the global is just (hi(&g)+lo(&g)).
1104 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1107 if (TM.getRelocationModel() == Reloc::PIC_) {
1108 // With PIC, the first instruction is actually "GR+hi(&G)".
1109 Hi = DAG.getNode(ISD::ADD, PtrVT,
1110 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1113 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1115 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1118 // If the global is weak or external, we have to go through the lazy
1120 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1123 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1124 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1126 // If we're comparing for equality to zero, expose the fact that this is
1127 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1128 // fold the new nodes.
1129 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1130 if (C->isNullValue() && CC == ISD::SETEQ) {
1131 MVT::ValueType VT = Op.getOperand(0).getValueType();
1132 SDOperand Zext = Op.getOperand(0);
1133 if (VT < MVT::i32) {
1135 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1137 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1138 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1139 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1140 DAG.getConstant(Log2b, MVT::i32));
1141 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1143 // Leave comparisons against 0 and -1 alone for now, since they're usually
1144 // optimized. FIXME: revisit this when we can custom lower all setcc
1146 if (C->isAllOnesValue() || C->isNullValue())
1150 // If we have an integer seteq/setne, turn it into a compare against zero
1151 // by xor'ing the rhs with the lhs, which is faster than setting a
1152 // condition register, reading it back out, and masking the correct bit. The
1153 // normal approach here uses sub to do this instead of xor. Using xor exposes
1154 // the result to other bit-twiddling opportunities.
1155 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1156 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1157 MVT::ValueType VT = Op.getValueType();
1158 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1160 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1165 static SDOperand LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1166 int VarArgsFrameIndex,
1167 int VarArgsStackOffset,
1168 unsigned VarArgsNumGPR,
1169 unsigned VarArgsNumFPR,
1170 const PPCSubtarget &Subtarget) {
1172 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1175 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1176 int VarArgsFrameIndex,
1177 int VarArgsStackOffset,
1178 unsigned VarArgsNumGPR,
1179 unsigned VarArgsNumFPR,
1180 const PPCSubtarget &Subtarget) {
1182 if (Subtarget.isMachoABI()) {
1183 // vastart just stores the address of the VarArgsFrameIndex slot into the
1184 // memory location argument.
1185 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1186 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1187 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1188 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1191 // For ELF 32 ABI we follow the layout of the va_list struct.
1192 // We suppose the given va_list is already allocated.
1195 // char gpr; /* index into the array of 8 GPRs
1196 // * stored in the register save area
1197 // * gpr=0 corresponds to r3,
1198 // * gpr=1 to r4, etc.
1200 // char fpr; /* index into the array of 8 FPRs
1201 // * stored in the register save area
1202 // * fpr=0 corresponds to f1,
1203 // * fpr=1 to f2, etc.
1205 // char *overflow_arg_area;
1206 // /* location on stack that holds
1207 // * the next overflow argument
1209 // char *reg_save_area;
1210 // /* where r3:r10 and f1:f8 (if saved)
1216 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1217 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1220 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1222 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1223 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1225 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1226 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1228 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1229 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1231 uint64_t FPROffset = 1;
1232 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1234 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1236 // Store first byte : number of int regs
1237 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1238 Op.getOperand(1), SV, 0);
1239 uint64_t nextOffset = FPROffset;
1240 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1243 // Store second byte : number of float regs
1244 SDOperand secondStore =
1245 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1246 nextOffset += StackOffset;
1247 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1249 // Store second word : arguments given on stack
1250 SDOperand thirdStore =
1251 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1252 nextOffset += FrameOffset;
1253 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1255 // Store third word : arguments given in registers
1256 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1260 #include "PPCGenCallingConv.inc"
1262 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1263 /// depending on which subtarget is selected.
1264 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1265 if (Subtarget.isMachoABI()) {
1266 static const unsigned FPR[] = {
1267 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1268 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1274 static const unsigned FPR[] = {
1275 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1281 static SDOperand LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG,
1282 int &VarArgsFrameIndex,
1283 int &VarArgsStackOffset,
1284 unsigned &VarArgsNumGPR,
1285 unsigned &VarArgsNumFPR,
1286 const PPCSubtarget &Subtarget) {
1287 // TODO: add description of PPC stack frame format, or at least some docs.
1289 MachineFunction &MF = DAG.getMachineFunction();
1290 MachineFrameInfo *MFI = MF.getFrameInfo();
1291 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1292 SmallVector<SDOperand, 8> ArgValues;
1293 SDOperand Root = Op.getOperand(0);
1295 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1296 bool isPPC64 = PtrVT == MVT::i64;
1297 bool isMachoABI = Subtarget.isMachoABI();
1298 bool isELF32_ABI = Subtarget.isELF32_ABI();
1299 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1301 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1303 static const unsigned GPR_32[] = { // 32-bit registers.
1304 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1305 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1307 static const unsigned GPR_64[] = { // 64-bit registers.
1308 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1309 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1312 static const unsigned *FPR = GetFPR(Subtarget);
1314 static const unsigned VR[] = {
1315 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1316 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1319 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1320 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1321 const unsigned Num_VR_Regs = array_lengthof( VR);
1323 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1325 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1327 // Add DAG nodes to load the arguments or copy them out of registers. On
1328 // entry to a function on PPC, the arguments start after the linkage area,
1329 // although the first ones are often in registers.
1331 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1332 // represented with two words (long long or double) must be copied to an
1333 // even GPR_idx value or to an even ArgOffset value.
1335 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1337 bool needsLoad = false;
1338 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1339 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1340 unsigned ArgSize = ObjSize;
1341 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1342 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1343 // See if next argument requires stack alignment in ELF
1344 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1345 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1346 (!(Flags & AlignFlag)));
1348 unsigned CurArgOffset = ArgOffset;
1350 default: assert(0 && "Unhandled argument type!");
1352 // Double word align in ELF
1353 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1354 if (GPR_idx != Num_GPR_Regs) {
1355 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1356 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1357 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1361 ArgSize = PtrByteSize;
1363 // Stack align in ELF
1364 if (needsLoad && Expand && isELF32_ABI)
1365 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1366 // All int arguments reserve stack space in Macho ABI.
1367 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1370 case MVT::i64: // PPC64
1371 if (GPR_idx != Num_GPR_Regs) {
1372 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1373 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1374 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1379 // All int arguments reserve stack space in Macho ABI.
1380 if (isMachoABI || needsLoad) ArgOffset += 8;
1385 // Every 4 bytes of argument space consumes one of the GPRs available for
1386 // argument passing.
1387 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1389 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1392 if (FPR_idx != Num_FPR_Regs) {
1394 if (ObjectVT == MVT::f32)
1395 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1397 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1398 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1399 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1405 // Stack align in ELF
1406 if (needsLoad && Expand && isELF32_ABI)
1407 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1408 // All FP arguments reserve stack space in Macho ABI.
1409 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1415 // Note that vector arguments in registers don't reserve stack space.
1416 if (VR_idx != Num_VR_Regs) {
1417 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1418 RegInfo.addLiveIn(VR[VR_idx], VReg);
1419 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1422 // This should be simple, but requires getting 16-byte aligned stack
1424 assert(0 && "Loading VR argument not implemented yet!");
1430 // We need to load the argument to a virtual register if we determined above
1431 // that we ran out of physical registers of the appropriate type.
1433 int FI = MFI->CreateFixedObject(ObjSize,
1434 CurArgOffset + (ArgSize - ObjSize));
1435 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1436 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1439 ArgValues.push_back(ArgVal);
1442 // If the function takes variable number of arguments, make a frame index for
1443 // the start of the first vararg value... for expansion of llvm.va_start.
1444 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1449 VarArgsNumGPR = GPR_idx;
1450 VarArgsNumFPR = FPR_idx;
1452 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1454 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1455 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1456 MVT::getSizeInBits(PtrVT)/8);
1458 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1465 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1467 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1469 SmallVector<SDOperand, 8> MemOps;
1471 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1472 // stored to the VarArgsFrameIndex on the stack.
1474 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1475 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1476 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1477 MemOps.push_back(Store);
1478 // Increment the address by four for the next argument to store
1479 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1480 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1484 // If this function is vararg, store any remaining integer argument regs
1485 // to their spots on the stack so that they may be loaded by deferencing the
1486 // result of va_next.
1487 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1490 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1492 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1494 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1495 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1496 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1497 MemOps.push_back(Store);
1498 // Increment the address by four for the next argument to store
1499 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1500 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1503 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1506 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1507 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1508 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1509 MemOps.push_back(Store);
1510 // Increment the address by eight for the next argument to store
1511 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1513 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1516 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1518 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1520 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1521 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1522 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1523 MemOps.push_back(Store);
1524 // Increment the address by eight for the next argument to store
1525 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1527 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1531 if (!MemOps.empty())
1532 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1535 ArgValues.push_back(Root);
1537 // Return the new list of results.
1538 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1539 Op.Val->value_end());
1540 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1543 /// isCallCompatibleAddress - Return the immediate to use if the specified
1544 /// 32-bit value is representable in the immediate field of a BxA instruction.
1545 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1546 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1549 int Addr = C->getValue();
1550 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1551 (Addr << 6 >> 6) != Addr)
1552 return 0; // Top 6 bits have to be sext of immediate.
1554 return DAG.getConstant((int)C->getValue() >> 2,
1555 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1559 static SDOperand LowerCALL(SDOperand Op, SelectionDAG &DAG,
1560 const PPCSubtarget &Subtarget) {
1561 SDOperand Chain = Op.getOperand(0);
1562 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1563 SDOperand Callee = Op.getOperand(4);
1564 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1566 bool isMachoABI = Subtarget.isMachoABI();
1567 bool isELF32_ABI = Subtarget.isELF32_ABI();
1569 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1570 bool isPPC64 = PtrVT == MVT::i64;
1571 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1573 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1574 // SelectExpr to use to put the arguments in the appropriate registers.
1575 std::vector<SDOperand> args_to_use;
1577 // Count how many bytes are to be pushed on the stack, including the linkage
1578 // area, and parameter passing area. We start with 24/48 bytes, which is
1579 // prereserved space for [SP][CR][LR][3 x unused].
1580 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1582 // Add up all the space actually used.
1583 for (unsigned i = 0; i != NumOps; ++i) {
1584 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1585 ArgSize = std::max(ArgSize, PtrByteSize);
1586 NumBytes += ArgSize;
1589 // The prolog code of the callee may store up to 8 GPR argument registers to
1590 // the stack, allowing va_start to index over them in memory if its varargs.
1591 // Because we cannot tell if this is needed on the caller side, we have to
1592 // conservatively assume that it is needed. As such, make sure we have at
1593 // least enough stack space for the caller to store the 8 GPRs.
1594 NumBytes = std::max(NumBytes,
1595 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1597 // Adjust the stack pointer for the new arguments...
1598 // These operations are automatically eliminated by the prolog/epilog pass
1599 Chain = DAG.getCALLSEQ_START(Chain,
1600 DAG.getConstant(NumBytes, PtrVT));
1602 // Set up a copy of the stack pointer for use loading and storing any
1603 // arguments that may not fit in the registers available for argument
1607 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1609 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1611 // Figure out which arguments are going to go in registers, and which in
1612 // memory. Also, if this is a vararg function, floating point operations
1613 // must be stored to our stack, and loaded into integer regs as well, if
1614 // any integer regs are available for argument passing.
1615 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1616 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1618 static const unsigned GPR_32[] = { // 32-bit registers.
1619 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1620 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1622 static const unsigned GPR_64[] = { // 64-bit registers.
1623 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1624 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1626 static const unsigned *FPR = GetFPR(Subtarget);
1628 static const unsigned VR[] = {
1629 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1630 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1632 const unsigned NumGPRs = array_lengthof(GPR_32);
1633 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1634 const unsigned NumVRs = array_lengthof( VR);
1636 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1638 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1639 SmallVector<SDOperand, 8> MemOpChains;
1640 for (unsigned i = 0; i != NumOps; ++i) {
1642 SDOperand Arg = Op.getOperand(5+2*i);
1643 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1644 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1645 // See if next argument requires stack alignment in ELF
1646 unsigned next = 5+2*(i+1)+1;
1647 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1648 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1649 (!(Flags & AlignFlag)));
1651 // PtrOff will be used to store the current argument to the stack if a
1652 // register cannot be found for it.
1655 // Stack align in ELF 32
1656 if (isELF32_ABI && Expand)
1657 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1658 StackPtr.getValueType());
1660 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1662 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1664 // On PPC64, promote integers to 64-bit values.
1665 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1666 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1668 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1671 switch (Arg.getValueType()) {
1672 default: assert(0 && "Unexpected ValueType for argument!");
1675 // Double word align in ELF
1676 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1677 if (GPR_idx != NumGPRs) {
1678 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1680 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1683 if (inMem || isMachoABI) {
1684 // Stack align in ELF
1685 if (isELF32_ABI && Expand)
1686 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1688 ArgOffset += PtrByteSize;
1694 // Float varargs need to be promoted to double.
1695 if (Arg.getValueType() == MVT::f32)
1696 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1699 if (FPR_idx != NumFPRs) {
1700 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1703 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1704 MemOpChains.push_back(Store);
1706 // Float varargs are always shadowed in available integer registers
1707 if (GPR_idx != NumGPRs) {
1708 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1709 MemOpChains.push_back(Load.getValue(1));
1710 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1713 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1714 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1715 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1716 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1717 MemOpChains.push_back(Load.getValue(1));
1718 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1722 // If we have any FPRs remaining, we may also have GPRs remaining.
1723 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1726 if (GPR_idx != NumGPRs)
1728 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1729 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1734 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1737 if (inMem || isMachoABI) {
1738 // Stack align in ELF
1739 if (isELF32_ABI && Expand)
1740 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1744 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1751 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1752 assert(VR_idx != NumVRs &&
1753 "Don't support passing more than 12 vector args yet!");
1754 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1758 if (!MemOpChains.empty())
1759 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1760 &MemOpChains[0], MemOpChains.size());
1762 // Build a sequence of copy-to-reg nodes chained together with token chain
1763 // and flag operands which copy the outgoing args into the appropriate regs.
1765 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1766 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1768 InFlag = Chain.getValue(1);
1771 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1772 if (isVarArg && isELF32_ABI) {
1773 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1774 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1775 InFlag = Chain.getValue(1);
1778 std::vector<MVT::ValueType> NodeTys;
1779 NodeTys.push_back(MVT::Other); // Returns a chain
1780 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1782 SmallVector<SDOperand, 8> Ops;
1783 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1785 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1786 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1787 // node so that legalize doesn't hack it.
1788 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1789 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1790 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1791 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1792 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1793 // If this is an absolute destination address, use the munged value.
1794 Callee = SDOperand(Dest, 0);
1796 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1797 // to do the call, we can't use PPCISD::CALL.
1798 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1799 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1800 InFlag = Chain.getValue(1);
1802 // Copy the callee address into R12 on darwin.
1804 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1805 InFlag = Chain.getValue(1);
1809 NodeTys.push_back(MVT::Other);
1810 NodeTys.push_back(MVT::Flag);
1811 Ops.push_back(Chain);
1812 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1816 // If this is a direct call, pass the chain and the callee.
1818 Ops.push_back(Chain);
1819 Ops.push_back(Callee);
1822 // Add argument registers to the end of the list so that they are known live
1824 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1825 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1826 RegsToPass[i].second.getValueType()));
1829 Ops.push_back(InFlag);
1830 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1831 InFlag = Chain.getValue(1);
1833 Chain = DAG.getCALLSEQ_END(Chain,
1834 DAG.getConstant(NumBytes, PtrVT),
1835 DAG.getConstant(0, PtrVT),
1837 if (Op.Val->getValueType(0) != MVT::Other)
1838 InFlag = Chain.getValue(1);
1840 SDOperand ResultVals[3];
1841 unsigned NumResults = 0;
1844 // If the call has results, copy the values out of the ret val registers.
1845 switch (Op.Val->getValueType(0)) {
1846 default: assert(0 && "Unexpected ret value!");
1847 case MVT::Other: break;
1849 if (Op.Val->getValueType(1) == MVT::i32) {
1850 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1851 ResultVals[0] = Chain.getValue(0);
1852 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
1853 Chain.getValue(2)).getValue(1);
1854 ResultVals[1] = Chain.getValue(0);
1856 NodeTys.push_back(MVT::i32);
1858 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1859 ResultVals[0] = Chain.getValue(0);
1862 NodeTys.push_back(MVT::i32);
1865 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
1866 ResultVals[0] = Chain.getValue(0);
1868 NodeTys.push_back(MVT::i64);
1871 if (Op.Val->getValueType(1) == MVT::f64) {
1872 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
1873 ResultVals[0] = Chain.getValue(0);
1874 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
1875 Chain.getValue(2)).getValue(1);
1876 ResultVals[1] = Chain.getValue(0);
1878 NodeTys.push_back(MVT::f64);
1879 NodeTys.push_back(MVT::f64);
1882 // else fall through
1884 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
1885 InFlag).getValue(1);
1886 ResultVals[0] = Chain.getValue(0);
1888 NodeTys.push_back(Op.Val->getValueType(0));
1894 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
1895 InFlag).getValue(1);
1896 ResultVals[0] = Chain.getValue(0);
1898 NodeTys.push_back(Op.Val->getValueType(0));
1902 NodeTys.push_back(MVT::Other);
1904 // If the function returns void, just return the chain.
1905 if (NumResults == 0)
1908 // Otherwise, merge everything together with a MERGE_VALUES node.
1909 ResultVals[NumResults++] = Chain;
1910 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1911 ResultVals, NumResults);
1912 return Res.getValue(Op.ResNo);
1915 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1916 SmallVector<CCValAssign, 16> RVLocs;
1917 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1918 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1919 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1920 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
1922 // If this is the first return lowered for this function, add the regs to the
1923 // liveout set for the function.
1924 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1925 for (unsigned i = 0; i != RVLocs.size(); ++i)
1926 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1929 SDOperand Chain = Op.getOperand(0);
1932 // Copy the result values into the output registers.
1933 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1934 CCValAssign &VA = RVLocs[i];
1935 assert(VA.isRegLoc() && "Can only return in registers!");
1936 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1937 Flag = Chain.getValue(1);
1941 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
1943 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
1946 static SDOperand LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
1947 const PPCSubtarget &Subtarget) {
1948 // When we pop the dynamic allocation we need to restore the SP link.
1950 // Get the corect type for pointers.
1951 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1953 // Construct the stack pointer operand.
1954 bool IsPPC64 = Subtarget.isPPC64();
1955 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
1956 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
1958 // Get the operands for the STACKRESTORE.
1959 SDOperand Chain = Op.getOperand(0);
1960 SDOperand SaveSP = Op.getOperand(1);
1962 // Load the old link SP.
1963 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
1965 // Restore the stack pointer.
1966 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
1968 // Store the old link SP.
1969 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
1972 static SDOperand LowerDYNAMIC_STACKALLOC(SDOperand Op, SelectionDAG &DAG,
1973 const PPCSubtarget &Subtarget) {
1974 MachineFunction &MF = DAG.getMachineFunction();
1975 bool IsPPC64 = Subtarget.isPPC64();
1976 bool isMachoABI = Subtarget.isMachoABI();
1978 // Get current frame pointer save index. The users of this index will be
1979 // primarily DYNALLOC instructions.
1980 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1981 int FPSI = FI->getFramePointerSaveIndex();
1983 // If the frame pointer save index hasn't been defined yet.
1985 // Find out what the fix offset of the frame pointer save area.
1986 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
1988 // Allocate the frame index for frame pointer save area.
1989 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
1991 FI->setFramePointerSaveIndex(FPSI);
1995 SDOperand Chain = Op.getOperand(0);
1996 SDOperand Size = Op.getOperand(1);
1998 // Get the corect type for pointers.
1999 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2001 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
2002 DAG.getConstant(0, PtrVT), Size);
2003 // Construct a node for the frame pointer save index.
2004 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
2005 // Build a DYNALLOC node.
2006 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
2007 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2008 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2012 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2014 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2015 // Not FP? Not a fsel.
2016 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2017 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2020 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2022 // Cannot handle SETEQ/SETNE.
2023 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2025 MVT::ValueType ResVT = Op.getValueType();
2026 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2027 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2028 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2030 // If the RHS of the comparison is a 0.0, we don't need to do the
2031 // subtraction at all.
2032 if (isFloatingPointZero(RHS))
2034 default: break; // SETUO etc aren't handled by fsel.
2038 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2042 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2043 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2044 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2048 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2052 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2053 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2054 return DAG.getNode(PPCISD::FSEL, ResVT,
2055 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2060 default: break; // SETUO etc aren't handled by fsel.
2064 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2065 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2066 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2067 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2071 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2072 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2073 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2074 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2078 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2079 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2080 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2081 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2085 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2086 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2087 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2088 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2093 // FIXME: Split this code up when LegalizeDAGTypes lands.
2094 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2095 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2096 SDOperand Src = Op.getOperand(0);
2097 if (Src.getValueType() == MVT::f32)
2098 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2101 switch (Op.getValueType()) {
2102 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2104 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2107 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2111 // Convert the FP value to an int value through memory.
2112 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2114 // Emit a store to the stack slot.
2115 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2117 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2119 if (Op.getValueType() == MVT::i32)
2120 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2121 DAG.getConstant(4, FIPtr.getValueType()));
2122 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2125 static SDOperand LowerFP_ROUND_INREG(SDOperand Op, SelectionDAG &DAG) {
2126 assert(Op.getValueType() == MVT::ppcf128);
2127 SDNode *Node = Op.Val;
2128 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2129 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2130 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2131 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2133 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2134 // of the long double, and puts FPSCR back the way it was. We do not
2135 // actually model FPSCR.
2136 std::vector<MVT::ValueType> NodeTys;
2137 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2139 NodeTys.push_back(MVT::f64); // Return register
2140 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2141 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2142 MFFSreg = Result.getValue(0);
2143 InFlag = Result.getValue(1);
2146 NodeTys.push_back(MVT::Flag); // Returns a flag
2147 Ops[0] = DAG.getConstant(31, MVT::i32);
2149 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2150 InFlag = Result.getValue(0);
2153 NodeTys.push_back(MVT::Flag); // Returns a flag
2154 Ops[0] = DAG.getConstant(30, MVT::i32);
2156 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2157 InFlag = Result.getValue(0);
2160 NodeTys.push_back(MVT::f64); // result of add
2161 NodeTys.push_back(MVT::Flag); // Returns a flag
2165 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2166 FPreg = Result.getValue(0);
2167 InFlag = Result.getValue(1);
2170 NodeTys.push_back(MVT::f64);
2171 Ops[0] = DAG.getConstant(1, MVT::i32);
2175 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2176 FPreg = Result.getValue(0);
2178 // We know the low half is about to be thrown away, so just use something
2180 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2183 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2184 if (Op.getOperand(0).getValueType() == MVT::i64) {
2185 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2186 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2187 if (Op.getValueType() == MVT::f32)
2188 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2192 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2193 "Unhandled SINT_TO_FP type in custom expander!");
2194 // Since we only generate this in 64-bit mode, we can take advantage of
2195 // 64-bit registers. In particular, sign extend the input value into the
2196 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2197 // then lfd it and fcfid it.
2198 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2199 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2200 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2201 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2203 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2206 // STD the extended value into the stack slot.
2207 MemOperand MO(PseudoSourceValue::getFixedStack(),
2208 MemOperand::MOStore, FrameIdx, 8, 8);
2209 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2210 DAG.getEntryNode(), Ext64, FIdx,
2211 DAG.getMemOperand(MO));
2212 // Load the value as a double.
2213 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2215 // FCFID it and return it.
2216 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2217 if (Op.getValueType() == MVT::f32)
2218 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2222 static SDOperand LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2224 The rounding mode is in bits 30:31 of FPSR, and has the following
2231 FLT_ROUNDS, on the other hand, expects the following:
2238 To perform the conversion, we do:
2239 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2242 MachineFunction &MF = DAG.getMachineFunction();
2243 MVT::ValueType VT = Op.getValueType();
2244 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2245 std::vector<MVT::ValueType> NodeTys;
2246 SDOperand MFFSreg, InFlag;
2248 // Save FP Control Word to register
2249 NodeTys.push_back(MVT::f64); // return register
2250 NodeTys.push_back(MVT::Flag); // unused in this context
2251 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2253 // Save FP register to stack slot
2254 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2255 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2256 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2257 StackSlot, NULL, 0);
2259 // Load FP Control Word from low 32 bits of stack slot.
2260 SDOperand Four = DAG.getConstant(4, PtrVT);
2261 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2262 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2264 // Transform as necessary
2266 DAG.getNode(ISD::AND, MVT::i32,
2267 CWD, DAG.getConstant(3, MVT::i32));
2269 DAG.getNode(ISD::SRL, MVT::i32,
2270 DAG.getNode(ISD::AND, MVT::i32,
2271 DAG.getNode(ISD::XOR, MVT::i32,
2272 CWD, DAG.getConstant(3, MVT::i32)),
2273 DAG.getConstant(3, MVT::i32)),
2274 DAG.getConstant(1, MVT::i8));
2277 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2279 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2280 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2283 static SDOperand LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2284 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2285 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
2287 // Expand into a bunch of logical ops. Note that these ops
2288 // depend on the PPC behavior for oversized shift amounts.
2289 SDOperand Lo = Op.getOperand(0);
2290 SDOperand Hi = Op.getOperand(1);
2291 SDOperand Amt = Op.getOperand(2);
2293 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2294 DAG.getConstant(32, MVT::i32), Amt);
2295 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
2296 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
2297 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2298 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2299 DAG.getConstant(-32U, MVT::i32));
2300 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
2301 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2302 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
2303 SDOperand OutOps[] = { OutLo, OutHi };
2304 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2308 static SDOperand LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2309 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2310 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRL!");
2312 // Otherwise, expand into a bunch of logical ops. Note that these ops
2313 // depend on the PPC behavior for oversized shift amounts.
2314 SDOperand Lo = Op.getOperand(0);
2315 SDOperand Hi = Op.getOperand(1);
2316 SDOperand Amt = Op.getOperand(2);
2318 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2319 DAG.getConstant(32, MVT::i32), Amt);
2320 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2321 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2322 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2323 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2324 DAG.getConstant(-32U, MVT::i32));
2325 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
2326 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2327 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
2328 SDOperand OutOps[] = { OutLo, OutHi };
2329 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2333 static SDOperand LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2334 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2335 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
2337 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
2338 SDOperand Lo = Op.getOperand(0);
2339 SDOperand Hi = Op.getOperand(1);
2340 SDOperand Amt = Op.getOperand(2);
2342 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2343 DAG.getConstant(32, MVT::i32), Amt);
2344 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2345 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2346 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2347 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2348 DAG.getConstant(-32U, MVT::i32));
2349 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
2350 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
2351 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
2352 Tmp4, Tmp6, ISD::SETLE);
2353 SDOperand OutOps[] = { OutLo, OutHi };
2354 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2358 //===----------------------------------------------------------------------===//
2359 // Vector related lowering.
2362 // If this is a vector of constants or undefs, get the bits. A bit in
2363 // UndefBits is set if the corresponding element of the vector is an
2364 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2365 // zero. Return true if this is not an array of constants, false if it is.
2367 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2368 uint64_t UndefBits[2]) {
2369 // Start with zero'd results.
2370 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2372 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2373 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2374 SDOperand OpVal = BV->getOperand(i);
2376 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2377 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2379 uint64_t EltBits = 0;
2380 if (OpVal.getOpcode() == ISD::UNDEF) {
2381 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2382 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2384 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2385 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2386 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2387 assert(CN->getValueType(0) == MVT::f32 &&
2388 "Only one legal FP vector type!");
2389 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2391 // Nonconstant element.
2395 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2398 //printf("%llx %llx %llx %llx\n",
2399 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2403 // If this is a splat (repetition) of a value across the whole vector, return
2404 // the smallest size that splats it. For example, "0x01010101010101..." is a
2405 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2406 // SplatSize = 1 byte.
2407 static bool isConstantSplat(const uint64_t Bits128[2],
2408 const uint64_t Undef128[2],
2409 unsigned &SplatBits, unsigned &SplatUndef,
2410 unsigned &SplatSize) {
2412 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2413 // the same as the lower 64-bits, ignoring undefs.
2414 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2415 return false; // Can't be a splat if two pieces don't match.
2417 uint64_t Bits64 = Bits128[0] | Bits128[1];
2418 uint64_t Undef64 = Undef128[0] & Undef128[1];
2420 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2422 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2423 return false; // Can't be a splat if two pieces don't match.
2425 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2426 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2428 // If the top 16-bits are different than the lower 16-bits, ignoring
2429 // undefs, we have an i32 splat.
2430 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2432 SplatUndef = Undef32;
2437 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2438 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2440 // If the top 8-bits are different than the lower 8-bits, ignoring
2441 // undefs, we have an i16 splat.
2442 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2444 SplatUndef = Undef16;
2449 // Otherwise, we have an 8-bit splat.
2450 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2451 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2456 /// BuildSplatI - Build a canonical splati of Val with an element size of
2457 /// SplatSize. Cast the result to VT.
2458 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2459 SelectionDAG &DAG) {
2460 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2462 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2463 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2466 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2468 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2472 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2474 // Build a canonical splat for this value.
2475 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2476 SmallVector<SDOperand, 8> Ops;
2477 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2478 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2479 &Ops[0], Ops.size());
2480 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2483 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2484 /// specified intrinsic ID.
2485 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2487 MVT::ValueType DestVT = MVT::Other) {
2488 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2489 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2490 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2493 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2494 /// specified intrinsic ID.
2495 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2496 SDOperand Op2, SelectionDAG &DAG,
2497 MVT::ValueType DestVT = MVT::Other) {
2498 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2499 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2500 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2504 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2505 /// amount. The result has the specified value type.
2506 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2507 MVT::ValueType VT, SelectionDAG &DAG) {
2508 // Force LHS/RHS to be the right type.
2509 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2510 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2513 for (unsigned i = 0; i != 16; ++i)
2514 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2515 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2516 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2517 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2520 // If this is a case we can't handle, return null and let the default
2521 // expansion code take care of it. If we CAN select this case, and if it
2522 // selects to a single instruction, return Op. Otherwise, if we can codegen
2523 // this case more efficiently than a constant pool load, lower it to the
2524 // sequence of ops that should be used.
2525 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2526 // If this is a vector of constants or undefs, get the bits. A bit in
2527 // UndefBits is set if the corresponding element of the vector is an
2528 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2530 uint64_t VectorBits[2];
2531 uint64_t UndefBits[2];
2532 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2533 return SDOperand(); // Not a constant vector.
2535 // If this is a splat (repetition) of a value across the whole vector, return
2536 // the smallest size that splats it. For example, "0x01010101010101..." is a
2537 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2538 // SplatSize = 1 byte.
2539 unsigned SplatBits, SplatUndef, SplatSize;
2540 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2541 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2543 // First, handle single instruction cases.
2546 if (SplatBits == 0) {
2547 // Canonicalize all zero vectors to be v4i32.
2548 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2549 SDOperand Z = DAG.getConstant(0, MVT::i32);
2550 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2551 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2556 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2557 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2558 if (SextVal >= -16 && SextVal <= 15)
2559 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2562 // Two instruction sequences.
2564 // If this value is in the range [-32,30] and is even, use:
2565 // tmp = VSPLTI[bhw], result = add tmp, tmp
2566 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2567 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2568 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2571 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2572 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2574 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2575 // Make -1 and vspltisw -1:
2576 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2578 // Make the VSLW intrinsic, computing 0x8000_0000.
2579 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2582 // xor by OnesV to invert it.
2583 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2584 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2587 // Check to see if this is a wide variety of vsplti*, binop self cases.
2588 unsigned SplatBitSize = SplatSize*8;
2589 static const signed char SplatCsts[] = {
2590 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2591 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2594 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2595 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2596 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2597 int i = SplatCsts[idx];
2599 // Figure out what shift amount will be used by altivec if shifted by i in
2601 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2603 // vsplti + shl self.
2604 if (SextVal == (i << (int)TypeShiftAmt)) {
2605 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2606 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2607 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2608 Intrinsic::ppc_altivec_vslw
2610 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2611 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2614 // vsplti + srl self.
2615 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2616 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2617 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2618 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2619 Intrinsic::ppc_altivec_vsrw
2621 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2622 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2625 // vsplti + sra self.
2626 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2627 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2628 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2629 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2630 Intrinsic::ppc_altivec_vsraw
2632 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2633 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2636 // vsplti + rol self.
2637 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2638 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2639 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2640 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2641 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2642 Intrinsic::ppc_altivec_vrlw
2644 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2645 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2648 // t = vsplti c, result = vsldoi t, t, 1
2649 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2650 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2651 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2653 // t = vsplti c, result = vsldoi t, t, 2
2654 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2655 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2656 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2658 // t = vsplti c, result = vsldoi t, t, 3
2659 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2660 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2661 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2665 // Three instruction sequences.
2667 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2668 if (SextVal >= 0 && SextVal <= 31) {
2669 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2670 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2671 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2672 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2674 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2675 if (SextVal >= -31 && SextVal <= 0) {
2676 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2677 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2678 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2679 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2686 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2687 /// the specified operations to build the shuffle.
2688 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2689 SDOperand RHS, SelectionDAG &DAG) {
2690 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2691 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2692 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2695 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2707 if (OpNum == OP_COPY) {
2708 if (LHSID == (1*9+2)*9+3) return LHS;
2709 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2713 SDOperand OpLHS, OpRHS;
2714 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2715 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2717 unsigned ShufIdxs[16];
2719 default: assert(0 && "Unknown i32 permute!");
2721 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2722 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2723 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2724 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2727 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2728 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2729 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2730 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2733 for (unsigned i = 0; i != 16; ++i)
2734 ShufIdxs[i] = (i&3)+0;
2737 for (unsigned i = 0; i != 16; ++i)
2738 ShufIdxs[i] = (i&3)+4;
2741 for (unsigned i = 0; i != 16; ++i)
2742 ShufIdxs[i] = (i&3)+8;
2745 for (unsigned i = 0; i != 16; ++i)
2746 ShufIdxs[i] = (i&3)+12;
2749 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2751 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2753 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2756 for (unsigned i = 0; i != 16; ++i)
2757 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2759 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2760 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2763 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2764 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2765 /// return the code it can be lowered into. Worst case, it can always be
2766 /// lowered into a vperm.
2767 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2768 SDOperand V1 = Op.getOperand(0);
2769 SDOperand V2 = Op.getOperand(1);
2770 SDOperand PermMask = Op.getOperand(2);
2772 // Cases that are handled by instructions that take permute immediates
2773 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2774 // selected by the instruction selector.
2775 if (V2.getOpcode() == ISD::UNDEF) {
2776 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2777 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2778 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2779 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2780 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2781 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2782 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2783 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2784 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2785 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2786 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2787 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2792 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2793 // and produce a fixed permutation. If any of these match, do not lower to
2795 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2796 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2797 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2798 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2799 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2800 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2801 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2802 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2803 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2806 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2807 // perfect shuffle table to emit an optimal matching sequence.
2808 unsigned PFIndexes[4];
2809 bool isFourElementShuffle = true;
2810 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2811 unsigned EltNo = 8; // Start out undef.
2812 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2813 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2814 continue; // Undef, ignore it.
2816 unsigned ByteSource =
2817 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2818 if ((ByteSource & 3) != j) {
2819 isFourElementShuffle = false;
2824 EltNo = ByteSource/4;
2825 } else if (EltNo != ByteSource/4) {
2826 isFourElementShuffle = false;
2830 PFIndexes[i] = EltNo;
2833 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
2834 // perfect shuffle vector to determine if it is cost effective to do this as
2835 // discrete instructions, or whether we should use a vperm.
2836 if (isFourElementShuffle) {
2837 // Compute the index in the perfect shuffle table.
2838 unsigned PFTableIndex =
2839 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2841 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2842 unsigned Cost = (PFEntry >> 30);
2844 // Determining when to avoid vperm is tricky. Many things affect the cost
2845 // of vperm, particularly how many times the perm mask needs to be computed.
2846 // For example, if the perm mask can be hoisted out of a loop or is already
2847 // used (perhaps because there are multiple permutes with the same shuffle
2848 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
2849 // the loop requires an extra register.
2851 // As a compromise, we only emit discrete instructions if the shuffle can be
2852 // generated in 3 or fewer operations. When we have loop information
2853 // available, if this block is within a loop, we should avoid using vperm
2854 // for 3-operation perms and use a constant pool load instead.
2856 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
2859 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
2860 // vector that will get spilled to the constant pool.
2861 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
2863 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
2864 // that it is in input element units, not in bytes. Convert now.
2865 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
2866 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
2868 SmallVector<SDOperand, 16> ResultMask;
2869 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
2871 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
2874 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
2876 for (unsigned j = 0; j != BytesPerElement; ++j)
2877 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
2881 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2882 &ResultMask[0], ResultMask.size());
2883 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
2886 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
2887 /// altivec comparison. If it is, return true and fill in Opc/isDot with
2888 /// information about the intrinsic.
2889 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
2891 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
2894 switch (IntrinsicID) {
2895 default: return false;
2896 // Comparison predicates.
2897 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
2898 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
2899 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
2900 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
2901 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
2902 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
2903 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
2904 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
2905 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
2906 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
2907 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
2908 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
2909 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
2911 // Normal Comparisons.
2912 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
2913 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
2914 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
2915 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
2916 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
2917 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
2918 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
2919 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
2920 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
2921 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
2922 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
2923 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
2924 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
2929 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
2930 /// lower, do it, otherwise return null.
2931 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
2932 // If this is a lowered altivec predicate compare, CompareOpc is set to the
2933 // opcode number of the comparison.
2936 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
2937 return SDOperand(); // Don't custom lower most intrinsics.
2939 // If this is a non-dot comparison, make the VCMP node and we are done.
2941 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
2942 Op.getOperand(1), Op.getOperand(2),
2943 DAG.getConstant(CompareOpc, MVT::i32));
2944 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
2947 // Create the PPCISD altivec 'dot' comparison node.
2949 Op.getOperand(2), // LHS
2950 Op.getOperand(3), // RHS
2951 DAG.getConstant(CompareOpc, MVT::i32)
2953 std::vector<MVT::ValueType> VTs;
2954 VTs.push_back(Op.getOperand(2).getValueType());
2955 VTs.push_back(MVT::Flag);
2956 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
2958 // Now that we have the comparison, emit a copy from the CR to a GPR.
2959 // This is flagged to the above dot comparison.
2960 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
2961 DAG.getRegister(PPC::CR6, MVT::i32),
2962 CompNode.getValue(1));
2964 // Unpack the result based on how the target uses it.
2965 unsigned BitNo; // Bit # of CR6.
2966 bool InvertBit; // Invert result?
2967 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
2968 default: // Can't happen, don't crash on invalid number though.
2969 case 0: // Return the value of the EQ bit of CR6.
2970 BitNo = 0; InvertBit = false;
2972 case 1: // Return the inverted value of the EQ bit of CR6.
2973 BitNo = 0; InvertBit = true;
2975 case 2: // Return the value of the LT bit of CR6.
2976 BitNo = 2; InvertBit = false;
2978 case 3: // Return the inverted value of the LT bit of CR6.
2979 BitNo = 2; InvertBit = true;
2983 // Shift the bit into the low position.
2984 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
2985 DAG.getConstant(8-(3-BitNo), MVT::i32));
2987 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
2988 DAG.getConstant(1, MVT::i32));
2990 // If we are supposed to, toggle the bit.
2992 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
2993 DAG.getConstant(1, MVT::i32));
2997 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2998 // Create a stack slot that is 16-byte aligned.
2999 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3000 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3001 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3002 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3004 // Store the input value into Value#0 of the stack slot.
3005 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
3006 Op.getOperand(0), FIdx, NULL, 0);
3008 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3011 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3012 if (Op.getValueType() == MVT::v4i32) {
3013 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3015 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3016 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3018 SDOperand RHSSwap = // = vrlw RHS, 16
3019 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3021 // Shrinkify inputs to v8i16.
3022 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3023 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3024 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3026 // Low parts multiplied together, generating 32-bit results (we ignore the
3028 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3029 LHS, RHS, DAG, MVT::v4i32);
3031 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3032 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3033 // Shift the high parts up 16 bits.
3034 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3035 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3036 } else if (Op.getValueType() == MVT::v8i16) {
3037 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3039 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3041 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3042 LHS, RHS, Zero, DAG);
3043 } else if (Op.getValueType() == MVT::v16i8) {
3044 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3046 // Multiply the even 8-bit parts, producing 16-bit sums.
3047 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3048 LHS, RHS, DAG, MVT::v8i16);
3049 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3051 // Multiply the odd 8-bit parts, producing 16-bit sums.
3052 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3053 LHS, RHS, DAG, MVT::v8i16);
3054 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3056 // Merge the results together.
3058 for (unsigned i = 0; i != 8; ++i) {
3059 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3060 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3062 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3063 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3065 assert(0 && "Unknown mul to lower!");
3070 /// LowerOperation - Provide custom lowering hooks for some operations.
3072 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3073 switch (Op.getOpcode()) {
3074 default: assert(0 && "Wasn't expecting to be able to lower this!");
3075 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3076 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3077 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3078 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3079 case ISD::SETCC: return LowerSETCC(Op, DAG);
3081 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3082 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3085 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3086 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3088 case ISD::FORMAL_ARGUMENTS:
3089 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3090 VarArgsStackOffset, VarArgsNumGPR,
3091 VarArgsNumFPR, PPCSubTarget);
3093 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
3094 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3095 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3096 case ISD::DYNAMIC_STACKALLOC:
3097 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3099 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3100 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3101 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3102 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3103 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3105 // Lower 64-bit shifts.
3106 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3107 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3108 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3110 // Vector-related lowering.
3111 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3112 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3113 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3114 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3115 case ISD::MUL: return LowerMUL(Op, DAG);
3117 // Frame & Return address.
3118 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3119 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3124 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3125 switch (N->getOpcode()) {
3126 default: assert(0 && "Wasn't expecting to be able to lower this!");
3127 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3132 //===----------------------------------------------------------------------===//
3133 // Other Lowering Code
3134 //===----------------------------------------------------------------------===//
3137 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3138 MachineBasicBlock *BB) {
3139 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3140 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3141 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3142 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3143 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3144 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3145 "Unexpected instr type to insert");
3147 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3148 // control-flow pattern. The incoming instruction knows the destination vreg
3149 // to set, the condition code register to branch on, the true/false values to
3150 // select between, and a branch opcode to use.
3151 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3152 ilist<MachineBasicBlock>::iterator It = BB;
3158 // cmpTY ccX, r1, r2
3160 // fallthrough --> copy0MBB
3161 MachineBasicBlock *thisMBB = BB;
3162 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3163 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3164 unsigned SelectPred = MI->getOperand(4).getImm();
3165 BuildMI(BB, TII->get(PPC::BCC))
3166 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3167 MachineFunction *F = BB->getParent();
3168 F->getBasicBlockList().insert(It, copy0MBB);
3169 F->getBasicBlockList().insert(It, sinkMBB);
3170 // Update machine-CFG edges by first adding all successors of the current
3171 // block to the new block which will contain the Phi node for the select.
3172 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3173 e = BB->succ_end(); i != e; ++i)
3174 sinkMBB->addSuccessor(*i);
3175 // Next, remove all successors of the current block, and add the true
3176 // and fallthrough blocks as its successors.
3177 while(!BB->succ_empty())
3178 BB->removeSuccessor(BB->succ_begin());
3179 BB->addSuccessor(copy0MBB);
3180 BB->addSuccessor(sinkMBB);
3183 // %FalseValue = ...
3184 // # fallthrough to sinkMBB
3187 // Update machine-CFG edges
3188 BB->addSuccessor(sinkMBB);
3191 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3194 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3195 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3196 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3198 delete MI; // The pseudo instruction is gone now.
3202 //===----------------------------------------------------------------------===//
3203 // Target Optimization Hooks
3204 //===----------------------------------------------------------------------===//
3206 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3207 DAGCombinerInfo &DCI) const {
3208 TargetMachine &TM = getTargetMachine();
3209 SelectionDAG &DAG = DCI.DAG;
3210 switch (N->getOpcode()) {
3213 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3214 if (C->getValue() == 0) // 0 << V -> 0.
3215 return N->getOperand(0);
3219 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3220 if (C->getValue() == 0) // 0 >>u V -> 0.
3221 return N->getOperand(0);
3225 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3226 if (C->getValue() == 0 || // 0 >>s V -> 0.
3227 C->isAllOnesValue()) // -1 >>s V -> -1.
3228 return N->getOperand(0);
3232 case ISD::SINT_TO_FP:
3233 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3234 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3235 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3236 // We allow the src/dst to be either f32/f64, but the intermediate
3237 // type must be i64.
3238 if (N->getOperand(0).getValueType() == MVT::i64 &&
3239 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3240 SDOperand Val = N->getOperand(0).getOperand(0);
3241 if (Val.getValueType() == MVT::f32) {
3242 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3243 DCI.AddToWorklist(Val.Val);
3246 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3247 DCI.AddToWorklist(Val.Val);
3248 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3249 DCI.AddToWorklist(Val.Val);
3250 if (N->getValueType(0) == MVT::f32) {
3251 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3252 DAG.getIntPtrConstant(0));
3253 DCI.AddToWorklist(Val.Val);
3256 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3257 // If the intermediate type is i32, we can avoid the load/store here
3264 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3265 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3266 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3267 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3268 N->getOperand(1).getValueType() == MVT::i32 &&
3269 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3270 SDOperand Val = N->getOperand(1).getOperand(0);
3271 if (Val.getValueType() == MVT::f32) {
3272 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3273 DCI.AddToWorklist(Val.Val);
3275 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3276 DCI.AddToWorklist(Val.Val);
3278 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3279 N->getOperand(2), N->getOperand(3));
3280 DCI.AddToWorklist(Val.Val);
3284 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3285 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3286 N->getOperand(1).Val->hasOneUse() &&
3287 (N->getOperand(1).getValueType() == MVT::i32 ||
3288 N->getOperand(1).getValueType() == MVT::i16)) {
3289 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3290 // Do an any-extend to 32-bits if this is a half-word input.
3291 if (BSwapOp.getValueType() == MVT::i16)
3292 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3294 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3295 N->getOperand(2), N->getOperand(3),
3296 DAG.getValueType(N->getOperand(1).getValueType()));
3300 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3301 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3302 N->getOperand(0).hasOneUse() &&
3303 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3304 SDOperand Load = N->getOperand(0);
3305 LoadSDNode *LD = cast<LoadSDNode>(Load);
3306 // Create the byte-swapping load.
3307 std::vector<MVT::ValueType> VTs;
3308 VTs.push_back(MVT::i32);
3309 VTs.push_back(MVT::Other);
3310 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3312 LD->getChain(), // Chain
3313 LD->getBasePtr(), // Ptr
3315 DAG.getValueType(N->getValueType(0)) // VT
3317 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3319 // If this is an i16 load, insert the truncate.
3320 SDOperand ResVal = BSLoad;
3321 if (N->getValueType(0) == MVT::i16)
3322 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3324 // First, combine the bswap away. This makes the value produced by the
3326 DCI.CombineTo(N, ResVal);
3328 // Next, combine the load away, we give it a bogus result value but a real
3329 // chain result. The result value is dead because the bswap is dead.
3330 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3332 // Return N so it doesn't get rechecked!
3333 return SDOperand(N, 0);
3337 case PPCISD::VCMP: {
3338 // If a VCMPo node already exists with exactly the same operands as this
3339 // node, use its result instead of this node (VCMPo computes both a CR6 and
3340 // a normal output).
3342 if (!N->getOperand(0).hasOneUse() &&
3343 !N->getOperand(1).hasOneUse() &&
3344 !N->getOperand(2).hasOneUse()) {
3346 // Scan all of the users of the LHS, looking for VCMPo's that match.
3347 SDNode *VCMPoNode = 0;
3349 SDNode *LHSN = N->getOperand(0).Val;
3350 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3352 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3353 (*UI)->getOperand(1) == N->getOperand(1) &&
3354 (*UI)->getOperand(2) == N->getOperand(2) &&
3355 (*UI)->getOperand(0) == N->getOperand(0)) {
3360 // If there is no VCMPo node, or if the flag value has a single use, don't
3362 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3365 // Look at the (necessarily single) use of the flag value. If it has a
3366 // chain, this transformation is more complex. Note that multiple things
3367 // could use the value result, which we should ignore.
3368 SDNode *FlagUser = 0;
3369 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3370 FlagUser == 0; ++UI) {
3371 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3373 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3374 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3381 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3382 // give up for right now.
3383 if (FlagUser->getOpcode() == PPCISD::MFCR)
3384 return SDOperand(VCMPoNode, 0);
3389 // If this is a branch on an altivec predicate comparison, lower this so
3390 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3391 // lowering is done pre-legalize, because the legalizer lowers the predicate
3392 // compare down to code that is difficult to reassemble.
3393 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3394 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3398 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3399 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3400 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3401 assert(isDot && "Can't compare against a vector result!");
3403 // If this is a comparison against something other than 0/1, then we know
3404 // that the condition is never/always true.
3405 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3406 if (Val != 0 && Val != 1) {
3407 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3408 return N->getOperand(0);
3409 // Always !=, turn it into an unconditional branch.
3410 return DAG.getNode(ISD::BR, MVT::Other,
3411 N->getOperand(0), N->getOperand(4));
3414 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3416 // Create the PPCISD altivec 'dot' comparison node.
3417 std::vector<MVT::ValueType> VTs;
3419 LHS.getOperand(2), // LHS of compare
3420 LHS.getOperand(3), // RHS of compare
3421 DAG.getConstant(CompareOpc, MVT::i32)
3423 VTs.push_back(LHS.getOperand(2).getValueType());
3424 VTs.push_back(MVT::Flag);
3425 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3427 // Unpack the result based on how the target uses it.
3428 PPC::Predicate CompOpc;
3429 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3430 default: // Can't happen, don't crash on invalid number though.
3431 case 0: // Branch on the value of the EQ bit of CR6.
3432 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3434 case 1: // Branch on the inverted value of the EQ bit of CR6.
3435 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3437 case 2: // Branch on the value of the LT bit of CR6.
3438 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3440 case 3: // Branch on the inverted value of the LT bit of CR6.
3441 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3445 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3446 DAG.getConstant(CompOpc, MVT::i32),
3447 DAG.getRegister(PPC::CR6, MVT::i32),
3448 N->getOperand(4), CompNode.getValue(1));
3457 //===----------------------------------------------------------------------===//
3458 // Inline Assembly Support
3459 //===----------------------------------------------------------------------===//
3461 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3465 const SelectionDAG &DAG,
3466 unsigned Depth) const {
3467 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3468 switch (Op.getOpcode()) {
3470 case PPCISD::LBRX: {
3471 // lhbrx is known to have the top bits cleared out.
3472 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3473 KnownZero = 0xFFFF0000;
3476 case ISD::INTRINSIC_WO_CHAIN: {
3477 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3479 case Intrinsic::ppc_altivec_vcmpbfp_p:
3480 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3481 case Intrinsic::ppc_altivec_vcmpequb_p:
3482 case Intrinsic::ppc_altivec_vcmpequh_p:
3483 case Intrinsic::ppc_altivec_vcmpequw_p:
3484 case Intrinsic::ppc_altivec_vcmpgefp_p:
3485 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3486 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3487 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3488 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3489 case Intrinsic::ppc_altivec_vcmpgtub_p:
3490 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3491 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3492 KnownZero = ~1U; // All bits but the low one are known to be zero.
3500 /// getConstraintType - Given a constraint, return the type of
3501 /// constraint it is for this target.
3502 PPCTargetLowering::ConstraintType
3503 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3504 if (Constraint.size() == 1) {
3505 switch (Constraint[0]) {
3512 return C_RegisterClass;
3515 return TargetLowering::getConstraintType(Constraint);
3518 std::pair<unsigned, const TargetRegisterClass*>
3519 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3520 MVT::ValueType VT) const {
3521 if (Constraint.size() == 1) {
3522 // GCC RS6000 Constraint Letters
3523 switch (Constraint[0]) {
3526 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3527 return std::make_pair(0U, PPC::G8RCRegisterClass);
3528 return std::make_pair(0U, PPC::GPRCRegisterClass);
3531 return std::make_pair(0U, PPC::F4RCRegisterClass);
3532 else if (VT == MVT::f64)
3533 return std::make_pair(0U, PPC::F8RCRegisterClass);
3536 return std::make_pair(0U, PPC::VRRCRegisterClass);
3538 return std::make_pair(0U, PPC::CRRCRegisterClass);
3542 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3546 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3547 /// vector. If it is invalid, don't add anything to Ops.
3548 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3549 std::vector<SDOperand>&Ops,
3550 SelectionDAG &DAG) {
3551 SDOperand Result(0,0);
3562 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3563 if (!CST) return; // Must be an immediate to match.
3564 unsigned Value = CST->getValue();
3566 default: assert(0 && "Unknown constraint letter!");
3567 case 'I': // "I" is a signed 16-bit constant.
3568 if ((short)Value == (int)Value)
3569 Result = DAG.getTargetConstant(Value, Op.getValueType());
3571 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3572 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3573 if ((short)Value == 0)
3574 Result = DAG.getTargetConstant(Value, Op.getValueType());
3576 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3577 if ((Value >> 16) == 0)
3578 Result = DAG.getTargetConstant(Value, Op.getValueType());
3580 case 'M': // "M" is a constant that is greater than 31.
3582 Result = DAG.getTargetConstant(Value, Op.getValueType());
3584 case 'N': // "N" is a positive constant that is an exact power of two.
3585 if ((int)Value > 0 && isPowerOf2_32(Value))
3586 Result = DAG.getTargetConstant(Value, Op.getValueType());
3588 case 'O': // "O" is the constant zero.
3590 Result = DAG.getTargetConstant(Value, Op.getValueType());
3592 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3593 if ((short)-Value == (int)-Value)
3594 Result = DAG.getTargetConstant(Value, Op.getValueType());
3602 Ops.push_back(Result);
3606 // Handle standard constraint letters.
3607 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3610 // isLegalAddressingMode - Return true if the addressing mode represented
3611 // by AM is legal for this target, for a load/store of the specified type.
3612 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3613 const Type *Ty) const {
3614 // FIXME: PPC does not allow r+i addressing modes for vectors!
3616 // PPC allows a sign-extended 16-bit immediate field.
3617 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3620 // No global is ever allowed as a base.
3624 // PPC only support r+r,
3626 case 0: // "r+i" or just "i", depending on HasBaseReg.
3629 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3631 // Otherwise we have r+r or r+i.
3634 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3636 // Allow 2*r as r+r.
3639 // No other scales are supported.
3646 /// isLegalAddressImmediate - Return true if the integer value can be used
3647 /// as the offset of the target addressing mode for load / store of the
3649 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3650 // PPC allows a sign-extended 16-bit immediate field.
3651 return (V > -(1 << 16) && V < (1 << 16)-1);
3654 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3658 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3659 // Depths > 0 not supported yet!
3660 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3663 MachineFunction &MF = DAG.getMachineFunction();
3664 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3665 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3667 bool isPPC64 = PPCSubTarget.isPPC64();
3669 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3671 // Set up a frame object for the return address.
3672 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3674 // Remember it for next time.
3675 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3677 // Make sure the function really does not optimize away the store of the RA
3679 FuncInfo->setLRStoreRequired();
3682 // Just load the return address off the stack.
3683 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3684 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3687 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3688 // Depths > 0 not supported yet!
3689 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3692 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3693 bool isPPC64 = PtrVT == MVT::i64;
3695 MachineFunction &MF = DAG.getMachineFunction();
3696 MachineFrameInfo *MFI = MF.getFrameInfo();
3697 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3698 && MFI->getStackSize();
3701 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3704 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,