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 uint64_t LHSKnownZero, LHSKnownOne;
742 uint64_t RHSKnownZero, RHSKnownOne;
743 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
746 DAG.ComputeMaskedBits(N.getOperand(1), ~0U, RHSKnownZero, RHSKnownOne);
747 // If all of the bits are known zero on the LHS or RHS, the add won't
749 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
750 Base = N.getOperand(0);
751 Index = N.getOperand(1);
760 /// Returns true if the address N can be represented by a base register plus
761 /// a signed 16-bit displacement [r+imm], and if it is not better
762 /// represented as reg+reg.
763 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
764 SDOperand &Base, SelectionDAG &DAG){
765 // If this can be more profitably realized as r+r, fail.
766 if (SelectAddressRegReg(N, Disp, Base, DAG))
769 if (N.getOpcode() == ISD::ADD) {
771 if (isIntS16Immediate(N.getOperand(1), imm)) {
772 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
773 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
774 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
776 Base = N.getOperand(0);
778 return true; // [r+i]
779 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
780 // Match LOAD (ADD (X, Lo(G))).
781 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
782 && "Cannot handle constant offsets yet!");
783 Disp = N.getOperand(1).getOperand(0); // The global address.
784 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
785 Disp.getOpcode() == ISD::TargetConstantPool ||
786 Disp.getOpcode() == ISD::TargetJumpTable);
787 Base = N.getOperand(0);
788 return true; // [&g+r]
790 } else if (N.getOpcode() == ISD::OR) {
792 if (isIntS16Immediate(N.getOperand(1), imm)) {
793 // If this is an or of disjoint bitfields, we can codegen this as an add
794 // (for better address arithmetic) if the LHS and RHS of the OR are
795 // provably disjoint.
796 uint64_t LHSKnownZero, LHSKnownOne;
797 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
798 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
799 // If all of the bits are known zero on the LHS or RHS, the add won't
801 Base = N.getOperand(0);
802 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
806 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
807 // Loading from a constant address.
809 // If this address fits entirely in a 16-bit sext immediate field, codegen
812 if (isIntS16Immediate(CN, Imm)) {
813 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
814 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
818 // Handle 32-bit sext immediates with LIS + addr mode.
819 if (CN->getValueType(0) == MVT::i32 ||
820 (int64_t)CN->getValue() == (int)CN->getValue()) {
821 int Addr = (int)CN->getValue();
823 // Otherwise, break this down into an LIS + disp.
824 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
826 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
827 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
828 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
833 Disp = DAG.getTargetConstant(0, getPointerTy());
834 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
835 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
838 return true; // [r+0]
841 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
842 /// represented as an indexed [r+r] operation.
843 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
846 // Check to see if we can easily represent this as an [r+r] address. This
847 // will fail if it thinks that the address is more profitably represented as
848 // reg+imm, e.g. where imm = 0.
849 if (SelectAddressRegReg(N, Base, Index, DAG))
852 // If the operand is an addition, always emit this as [r+r], since this is
853 // better (for code size, and execution, as the memop does the add for free)
854 // than emitting an explicit add.
855 if (N.getOpcode() == ISD::ADD) {
856 Base = N.getOperand(0);
857 Index = N.getOperand(1);
861 // Otherwise, do it the hard way, using R0 as the base register.
862 Base = DAG.getRegister(PPC::R0, N.getValueType());
867 /// SelectAddressRegImmShift - Returns true if the address N can be
868 /// represented by a base register plus a signed 14-bit displacement
869 /// [r+imm*4]. Suitable for use by STD and friends.
870 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
873 // If this can be more profitably realized as r+r, fail.
874 if (SelectAddressRegReg(N, Disp, Base, DAG))
877 if (N.getOpcode() == ISD::ADD) {
879 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
880 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
881 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
882 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
884 Base = N.getOperand(0);
886 return true; // [r+i]
887 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
888 // Match LOAD (ADD (X, Lo(G))).
889 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
890 && "Cannot handle constant offsets yet!");
891 Disp = N.getOperand(1).getOperand(0); // The global address.
892 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
893 Disp.getOpcode() == ISD::TargetConstantPool ||
894 Disp.getOpcode() == ISD::TargetJumpTable);
895 Base = N.getOperand(0);
896 return true; // [&g+r]
898 } else if (N.getOpcode() == ISD::OR) {
900 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
901 // If this is an or of disjoint bitfields, we can codegen this as an add
902 // (for better address arithmetic) if the LHS and RHS of the OR are
903 // provably disjoint.
904 uint64_t LHSKnownZero, LHSKnownOne;
905 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
906 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
907 // If all of the bits are known zero on the LHS or RHS, the add won't
909 Base = N.getOperand(0);
910 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
914 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
915 // Loading from a constant address. Verify low two bits are clear.
916 if ((CN->getValue() & 3) == 0) {
917 // If this address fits entirely in a 14-bit sext immediate field, codegen
920 if (isIntS16Immediate(CN, Imm)) {
921 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
922 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
926 // Fold the low-part of 32-bit absolute addresses into addr mode.
927 if (CN->getValueType(0) == MVT::i32 ||
928 (int64_t)CN->getValue() == (int)CN->getValue()) {
929 int Addr = (int)CN->getValue();
931 // Otherwise, break this down into an LIS + disp.
932 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
934 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
935 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
936 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
942 Disp = DAG.getTargetConstant(0, getPointerTy());
943 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
944 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
947 return true; // [r+0]
951 /// getPreIndexedAddressParts - returns true by value, base pointer and
952 /// offset pointer and addressing mode by reference if the node's address
953 /// can be legally represented as pre-indexed load / store address.
954 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
956 ISD::MemIndexedMode &AM,
958 // Disabled by default for now.
959 if (!EnablePPCPreinc) return false;
963 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
964 Ptr = LD->getBasePtr();
965 VT = LD->getMemoryVT();
967 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
969 Ptr = ST->getBasePtr();
970 VT = ST->getMemoryVT();
974 // PowerPC doesn't have preinc load/store instructions for vectors.
975 if (MVT::isVector(VT))
978 // TODO: Check reg+reg first.
980 // LDU/STU use reg+imm*4, others use reg+imm.
981 if (VT != MVT::i64) {
983 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
987 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
991 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
992 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
993 // sext i32 to i64 when addr mode is r+i.
994 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
995 LD->getExtensionType() == ISD::SEXTLOAD &&
996 isa<ConstantSDNode>(Offset))
1004 //===----------------------------------------------------------------------===//
1005 // LowerOperation implementation
1006 //===----------------------------------------------------------------------===//
1008 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
1009 MVT::ValueType PtrVT = Op.getValueType();
1010 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1011 Constant *C = CP->getConstVal();
1012 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1013 SDOperand Zero = DAG.getConstant(0, PtrVT);
1015 const TargetMachine &TM = DAG.getTarget();
1017 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1018 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1020 // If this is a non-darwin platform, we don't support non-static relo models
1022 if (TM.getRelocationModel() == Reloc::Static ||
1023 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1024 // Generate non-pic code that has direct accesses to the constant pool.
1025 // The address of the global is just (hi(&g)+lo(&g)).
1026 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1029 if (TM.getRelocationModel() == Reloc::PIC_) {
1030 // With PIC, the first instruction is actually "GR+hi(&G)".
1031 Hi = DAG.getNode(ISD::ADD, PtrVT,
1032 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1035 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1039 static SDOperand LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1040 MVT::ValueType PtrVT = Op.getValueType();
1041 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1042 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1043 SDOperand Zero = DAG.getConstant(0, PtrVT);
1045 const TargetMachine &TM = DAG.getTarget();
1047 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1048 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1050 // If this is a non-darwin platform, we don't support non-static relo models
1052 if (TM.getRelocationModel() == Reloc::Static ||
1053 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1054 // Generate non-pic code that has direct accesses to the constant pool.
1055 // The address of the global is just (hi(&g)+lo(&g)).
1056 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1059 if (TM.getRelocationModel() == Reloc::PIC_) {
1060 // With PIC, the first instruction is actually "GR+hi(&G)".
1061 Hi = DAG.getNode(ISD::ADD, PtrVT,
1062 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1065 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1069 static SDOperand LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
1070 assert(0 && "TLS not implemented for PPC.");
1073 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
1074 MVT::ValueType PtrVT = Op.getValueType();
1075 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1076 GlobalValue *GV = GSDN->getGlobal();
1077 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1078 // If it's a debug information descriptor, don't mess with it.
1079 if (DAG.isVerifiedDebugInfoDesc(Op))
1081 SDOperand Zero = DAG.getConstant(0, PtrVT);
1083 const TargetMachine &TM = DAG.getTarget();
1085 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1086 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1088 // If this is a non-darwin platform, we don't support non-static relo models
1090 if (TM.getRelocationModel() == Reloc::Static ||
1091 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1092 // Generate non-pic code that has direct accesses to globals.
1093 // The address of the global is just (hi(&g)+lo(&g)).
1094 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1097 if (TM.getRelocationModel() == Reloc::PIC_) {
1098 // With PIC, the first instruction is actually "GR+hi(&G)".
1099 Hi = DAG.getNode(ISD::ADD, PtrVT,
1100 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1103 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1105 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1108 // If the global is weak or external, we have to go through the lazy
1110 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1113 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1114 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1116 // If we're comparing for equality to zero, expose the fact that this is
1117 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1118 // fold the new nodes.
1119 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1120 if (C->isNullValue() && CC == ISD::SETEQ) {
1121 MVT::ValueType VT = Op.getOperand(0).getValueType();
1122 SDOperand Zext = Op.getOperand(0);
1123 if (VT < MVT::i32) {
1125 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1127 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1128 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1129 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1130 DAG.getConstant(Log2b, MVT::i32));
1131 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1133 // Leave comparisons against 0 and -1 alone for now, since they're usually
1134 // optimized. FIXME: revisit this when we can custom lower all setcc
1136 if (C->isAllOnesValue() || C->isNullValue())
1140 // If we have an integer seteq/setne, turn it into a compare against zero
1141 // by xor'ing the rhs with the lhs, which is faster than setting a
1142 // condition register, reading it back out, and masking the correct bit. The
1143 // normal approach here uses sub to do this instead of xor. Using xor exposes
1144 // the result to other bit-twiddling opportunities.
1145 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1146 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1147 MVT::ValueType VT = Op.getValueType();
1148 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1150 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1155 static SDOperand LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1156 int VarArgsFrameIndex,
1157 int VarArgsStackOffset,
1158 unsigned VarArgsNumGPR,
1159 unsigned VarArgsNumFPR,
1160 const PPCSubtarget &Subtarget) {
1162 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1165 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1166 int VarArgsFrameIndex,
1167 int VarArgsStackOffset,
1168 unsigned VarArgsNumGPR,
1169 unsigned VarArgsNumFPR,
1170 const PPCSubtarget &Subtarget) {
1172 if (Subtarget.isMachoABI()) {
1173 // vastart just stores the address of the VarArgsFrameIndex slot into the
1174 // memory location argument.
1175 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1176 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1177 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1178 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1181 // For ELF 32 ABI we follow the layout of the va_list struct.
1182 // We suppose the given va_list is already allocated.
1185 // char gpr; /* index into the array of 8 GPRs
1186 // * stored in the register save area
1187 // * gpr=0 corresponds to r3,
1188 // * gpr=1 to r4, etc.
1190 // char fpr; /* index into the array of 8 FPRs
1191 // * stored in the register save area
1192 // * fpr=0 corresponds to f1,
1193 // * fpr=1 to f2, etc.
1195 // char *overflow_arg_area;
1196 // /* location on stack that holds
1197 // * the next overflow argument
1199 // char *reg_save_area;
1200 // /* where r3:r10 and f1:f8 (if saved)
1206 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1207 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1210 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1212 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1213 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1215 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1216 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1218 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1219 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1221 uint64_t FPROffset = 1;
1222 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1224 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1226 // Store first byte : number of int regs
1227 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1228 Op.getOperand(1), SV, 0);
1229 uint64_t nextOffset = FPROffset;
1230 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1233 // Store second byte : number of float regs
1234 SDOperand secondStore =
1235 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1236 nextOffset += StackOffset;
1237 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1239 // Store second word : arguments given on stack
1240 SDOperand thirdStore =
1241 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1242 nextOffset += FrameOffset;
1243 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1245 // Store third word : arguments given in registers
1246 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1250 #include "PPCGenCallingConv.inc"
1252 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1253 /// depending on which subtarget is selected.
1254 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1255 if (Subtarget.isMachoABI()) {
1256 static const unsigned FPR[] = {
1257 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1258 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1264 static const unsigned FPR[] = {
1265 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1271 static SDOperand LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG,
1272 int &VarArgsFrameIndex,
1273 int &VarArgsStackOffset,
1274 unsigned &VarArgsNumGPR,
1275 unsigned &VarArgsNumFPR,
1276 const PPCSubtarget &Subtarget) {
1277 // TODO: add description of PPC stack frame format, or at least some docs.
1279 MachineFunction &MF = DAG.getMachineFunction();
1280 MachineFrameInfo *MFI = MF.getFrameInfo();
1281 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1282 SmallVector<SDOperand, 8> ArgValues;
1283 SDOperand Root = Op.getOperand(0);
1285 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1286 bool isPPC64 = PtrVT == MVT::i64;
1287 bool isMachoABI = Subtarget.isMachoABI();
1288 bool isELF32_ABI = Subtarget.isELF32_ABI();
1289 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1291 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1293 static const unsigned GPR_32[] = { // 32-bit registers.
1294 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1295 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1297 static const unsigned GPR_64[] = { // 64-bit registers.
1298 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1299 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1302 static const unsigned *FPR = GetFPR(Subtarget);
1304 static const unsigned VR[] = {
1305 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1306 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1309 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1310 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1311 const unsigned Num_VR_Regs = array_lengthof( VR);
1313 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1315 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1317 // Add DAG nodes to load the arguments or copy them out of registers. On
1318 // entry to a function on PPC, the arguments start after the linkage area,
1319 // although the first ones are often in registers.
1321 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1322 // represented with two words (long long or double) must be copied to an
1323 // even GPR_idx value or to an even ArgOffset value.
1325 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1327 bool needsLoad = false;
1328 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1329 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1330 unsigned ArgSize = ObjSize;
1331 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1332 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1333 // See if next argument requires stack alignment in ELF
1334 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1335 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1336 (!(Flags & AlignFlag)));
1338 unsigned CurArgOffset = ArgOffset;
1340 default: assert(0 && "Unhandled argument type!");
1342 // Double word align in ELF
1343 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1344 if (GPR_idx != Num_GPR_Regs) {
1345 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1346 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1347 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1351 ArgSize = PtrByteSize;
1353 // Stack align in ELF
1354 if (needsLoad && Expand && isELF32_ABI)
1355 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1356 // All int arguments reserve stack space in Macho ABI.
1357 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1360 case MVT::i64: // PPC64
1361 if (GPR_idx != Num_GPR_Regs) {
1362 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1363 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1364 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1369 // All int arguments reserve stack space in Macho ABI.
1370 if (isMachoABI || needsLoad) ArgOffset += 8;
1375 // Every 4 bytes of argument space consumes one of the GPRs available for
1376 // argument passing.
1377 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1379 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1382 if (FPR_idx != Num_FPR_Regs) {
1384 if (ObjectVT == MVT::f32)
1385 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1387 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1388 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1389 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1395 // Stack align in ELF
1396 if (needsLoad && Expand && isELF32_ABI)
1397 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1398 // All FP arguments reserve stack space in Macho ABI.
1399 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1405 // Note that vector arguments in registers don't reserve stack space.
1406 if (VR_idx != Num_VR_Regs) {
1407 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1408 RegInfo.addLiveIn(VR[VR_idx], VReg);
1409 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1412 // This should be simple, but requires getting 16-byte aligned stack
1414 assert(0 && "Loading VR argument not implemented yet!");
1420 // We need to load the argument to a virtual register if we determined above
1421 // that we ran out of physical registers of the appropriate type.
1423 int FI = MFI->CreateFixedObject(ObjSize,
1424 CurArgOffset + (ArgSize - ObjSize));
1425 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1426 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1429 ArgValues.push_back(ArgVal);
1432 // If the function takes variable number of arguments, make a frame index for
1433 // the start of the first vararg value... for expansion of llvm.va_start.
1434 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1439 VarArgsNumGPR = GPR_idx;
1440 VarArgsNumFPR = FPR_idx;
1442 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1444 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1445 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1446 MVT::getSizeInBits(PtrVT)/8);
1448 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1455 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1457 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1459 SmallVector<SDOperand, 8> MemOps;
1461 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1462 // stored to the VarArgsFrameIndex on the stack.
1464 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1465 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1466 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1467 MemOps.push_back(Store);
1468 // Increment the address by four for the next argument to store
1469 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1470 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1474 // If this function is vararg, store any remaining integer argument regs
1475 // to their spots on the stack so that they may be loaded by deferencing the
1476 // result of va_next.
1477 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1480 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1482 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1484 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1485 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1486 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1487 MemOps.push_back(Store);
1488 // Increment the address by four for the next argument to store
1489 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1490 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1493 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1496 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1497 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1498 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1499 MemOps.push_back(Store);
1500 // Increment the address by eight for the next argument to store
1501 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1503 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1506 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1508 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1510 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1511 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1512 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1513 MemOps.push_back(Store);
1514 // Increment the address by eight for the next argument to store
1515 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1517 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1521 if (!MemOps.empty())
1522 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1525 ArgValues.push_back(Root);
1527 // Return the new list of results.
1528 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1529 Op.Val->value_end());
1530 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1533 /// isCallCompatibleAddress - Return the immediate to use if the specified
1534 /// 32-bit value is representable in the immediate field of a BxA instruction.
1535 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1536 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1539 int Addr = C->getValue();
1540 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1541 (Addr << 6 >> 6) != Addr)
1542 return 0; // Top 6 bits have to be sext of immediate.
1544 return DAG.getConstant((int)C->getValue() >> 2,
1545 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1549 static SDOperand LowerCALL(SDOperand Op, SelectionDAG &DAG,
1550 const PPCSubtarget &Subtarget) {
1551 SDOperand Chain = Op.getOperand(0);
1552 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1553 SDOperand Callee = Op.getOperand(4);
1554 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1556 bool isMachoABI = Subtarget.isMachoABI();
1557 bool isELF32_ABI = Subtarget.isELF32_ABI();
1559 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1560 bool isPPC64 = PtrVT == MVT::i64;
1561 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1563 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1564 // SelectExpr to use to put the arguments in the appropriate registers.
1565 std::vector<SDOperand> args_to_use;
1567 // Count how many bytes are to be pushed on the stack, including the linkage
1568 // area, and parameter passing area. We start with 24/48 bytes, which is
1569 // prereserved space for [SP][CR][LR][3 x unused].
1570 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1572 // Add up all the space actually used.
1573 for (unsigned i = 0; i != NumOps; ++i) {
1574 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1575 ArgSize = std::max(ArgSize, PtrByteSize);
1576 NumBytes += ArgSize;
1579 // The prolog code of the callee may store up to 8 GPR argument registers to
1580 // the stack, allowing va_start to index over them in memory if its varargs.
1581 // Because we cannot tell if this is needed on the caller side, we have to
1582 // conservatively assume that it is needed. As such, make sure we have at
1583 // least enough stack space for the caller to store the 8 GPRs.
1584 NumBytes = std::max(NumBytes,
1585 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1587 // Adjust the stack pointer for the new arguments...
1588 // These operations are automatically eliminated by the prolog/epilog pass
1589 Chain = DAG.getCALLSEQ_START(Chain,
1590 DAG.getConstant(NumBytes, PtrVT));
1592 // Set up a copy of the stack pointer for use loading and storing any
1593 // arguments that may not fit in the registers available for argument
1597 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1599 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1601 // Figure out which arguments are going to go in registers, and which in
1602 // memory. Also, if this is a vararg function, floating point operations
1603 // must be stored to our stack, and loaded into integer regs as well, if
1604 // any integer regs are available for argument passing.
1605 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1606 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1608 static const unsigned GPR_32[] = { // 32-bit registers.
1609 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1610 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1612 static const unsigned GPR_64[] = { // 64-bit registers.
1613 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1614 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1616 static const unsigned *FPR = GetFPR(Subtarget);
1618 static const unsigned VR[] = {
1619 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1620 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1622 const unsigned NumGPRs = array_lengthof(GPR_32);
1623 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1624 const unsigned NumVRs = array_lengthof( VR);
1626 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1628 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1629 SmallVector<SDOperand, 8> MemOpChains;
1630 for (unsigned i = 0; i != NumOps; ++i) {
1632 SDOperand Arg = Op.getOperand(5+2*i);
1633 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1634 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1635 // See if next argument requires stack alignment in ELF
1636 unsigned next = 5+2*(i+1)+1;
1637 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1638 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1639 (!(Flags & AlignFlag)));
1641 // PtrOff will be used to store the current argument to the stack if a
1642 // register cannot be found for it.
1645 // Stack align in ELF 32
1646 if (isELF32_ABI && Expand)
1647 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1648 StackPtr.getValueType());
1650 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1652 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1654 // On PPC64, promote integers to 64-bit values.
1655 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1656 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1658 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1661 switch (Arg.getValueType()) {
1662 default: assert(0 && "Unexpected ValueType for argument!");
1665 // Double word align in ELF
1666 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1667 if (GPR_idx != NumGPRs) {
1668 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1670 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1673 if (inMem || isMachoABI) {
1674 // Stack align in ELF
1675 if (isELF32_ABI && Expand)
1676 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1678 ArgOffset += PtrByteSize;
1684 // Float varargs need to be promoted to double.
1685 if (Arg.getValueType() == MVT::f32)
1686 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1689 if (FPR_idx != NumFPRs) {
1690 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1693 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1694 MemOpChains.push_back(Store);
1696 // Float varargs are always shadowed in available integer registers
1697 if (GPR_idx != NumGPRs) {
1698 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1699 MemOpChains.push_back(Load.getValue(1));
1700 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1703 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1704 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1705 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1706 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1707 MemOpChains.push_back(Load.getValue(1));
1708 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1712 // If we have any FPRs remaining, we may also have GPRs remaining.
1713 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1716 if (GPR_idx != NumGPRs)
1718 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1719 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1724 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1727 if (inMem || isMachoABI) {
1728 // Stack align in ELF
1729 if (isELF32_ABI && Expand)
1730 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1734 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1741 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1742 assert(VR_idx != NumVRs &&
1743 "Don't support passing more than 12 vector args yet!");
1744 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1748 if (!MemOpChains.empty())
1749 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1750 &MemOpChains[0], MemOpChains.size());
1752 // Build a sequence of copy-to-reg nodes chained together with token chain
1753 // and flag operands which copy the outgoing args into the appropriate regs.
1755 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1756 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1758 InFlag = Chain.getValue(1);
1761 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1762 if (isVarArg && isELF32_ABI) {
1763 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1764 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1765 InFlag = Chain.getValue(1);
1768 std::vector<MVT::ValueType> NodeTys;
1769 NodeTys.push_back(MVT::Other); // Returns a chain
1770 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1772 SmallVector<SDOperand, 8> Ops;
1773 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1775 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1776 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1777 // node so that legalize doesn't hack it.
1778 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1779 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1780 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1781 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1782 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1783 // If this is an absolute destination address, use the munged value.
1784 Callee = SDOperand(Dest, 0);
1786 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1787 // to do the call, we can't use PPCISD::CALL.
1788 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1789 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1790 InFlag = Chain.getValue(1);
1792 // Copy the callee address into R12 on darwin.
1794 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1795 InFlag = Chain.getValue(1);
1799 NodeTys.push_back(MVT::Other);
1800 NodeTys.push_back(MVT::Flag);
1801 Ops.push_back(Chain);
1802 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1806 // If this is a direct call, pass the chain and the callee.
1808 Ops.push_back(Chain);
1809 Ops.push_back(Callee);
1812 // Add argument registers to the end of the list so that they are known live
1814 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1815 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1816 RegsToPass[i].second.getValueType()));
1819 Ops.push_back(InFlag);
1820 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1821 InFlag = Chain.getValue(1);
1823 Chain = DAG.getCALLSEQ_END(Chain,
1824 DAG.getConstant(NumBytes, PtrVT),
1825 DAG.getConstant(0, PtrVT),
1827 if (Op.Val->getValueType(0) != MVT::Other)
1828 InFlag = Chain.getValue(1);
1830 SDOperand ResultVals[3];
1831 unsigned NumResults = 0;
1834 // If the call has results, copy the values out of the ret val registers.
1835 switch (Op.Val->getValueType(0)) {
1836 default: assert(0 && "Unexpected ret value!");
1837 case MVT::Other: break;
1839 if (Op.Val->getValueType(1) == MVT::i32) {
1840 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1841 ResultVals[0] = Chain.getValue(0);
1842 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
1843 Chain.getValue(2)).getValue(1);
1844 ResultVals[1] = Chain.getValue(0);
1846 NodeTys.push_back(MVT::i32);
1848 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1849 ResultVals[0] = Chain.getValue(0);
1852 NodeTys.push_back(MVT::i32);
1855 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
1856 ResultVals[0] = Chain.getValue(0);
1858 NodeTys.push_back(MVT::i64);
1861 if (Op.Val->getValueType(1) == MVT::f64) {
1862 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
1863 ResultVals[0] = Chain.getValue(0);
1864 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
1865 Chain.getValue(2)).getValue(1);
1866 ResultVals[1] = Chain.getValue(0);
1868 NodeTys.push_back(MVT::f64);
1869 NodeTys.push_back(MVT::f64);
1872 // else fall through
1874 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
1875 InFlag).getValue(1);
1876 ResultVals[0] = Chain.getValue(0);
1878 NodeTys.push_back(Op.Val->getValueType(0));
1884 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
1885 InFlag).getValue(1);
1886 ResultVals[0] = Chain.getValue(0);
1888 NodeTys.push_back(Op.Val->getValueType(0));
1892 NodeTys.push_back(MVT::Other);
1894 // If the function returns void, just return the chain.
1895 if (NumResults == 0)
1898 // Otherwise, merge everything together with a MERGE_VALUES node.
1899 ResultVals[NumResults++] = Chain;
1900 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1901 ResultVals, NumResults);
1902 return Res.getValue(Op.ResNo);
1905 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1906 SmallVector<CCValAssign, 16> RVLocs;
1907 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1908 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1909 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1910 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
1912 // If this is the first return lowered for this function, add the regs to the
1913 // liveout set for the function.
1914 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1915 for (unsigned i = 0; i != RVLocs.size(); ++i)
1916 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1919 SDOperand Chain = Op.getOperand(0);
1922 // Copy the result values into the output registers.
1923 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1924 CCValAssign &VA = RVLocs[i];
1925 assert(VA.isRegLoc() && "Can only return in registers!");
1926 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1927 Flag = Chain.getValue(1);
1931 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
1933 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
1936 static SDOperand LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
1937 const PPCSubtarget &Subtarget) {
1938 // When we pop the dynamic allocation we need to restore the SP link.
1940 // Get the corect type for pointers.
1941 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1943 // Construct the stack pointer operand.
1944 bool IsPPC64 = Subtarget.isPPC64();
1945 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
1946 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
1948 // Get the operands for the STACKRESTORE.
1949 SDOperand Chain = Op.getOperand(0);
1950 SDOperand SaveSP = Op.getOperand(1);
1952 // Load the old link SP.
1953 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
1955 // Restore the stack pointer.
1956 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
1958 // Store the old link SP.
1959 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
1962 static SDOperand LowerDYNAMIC_STACKALLOC(SDOperand Op, SelectionDAG &DAG,
1963 const PPCSubtarget &Subtarget) {
1964 MachineFunction &MF = DAG.getMachineFunction();
1965 bool IsPPC64 = Subtarget.isPPC64();
1966 bool isMachoABI = Subtarget.isMachoABI();
1968 // Get current frame pointer save index. The users of this index will be
1969 // primarily DYNALLOC instructions.
1970 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1971 int FPSI = FI->getFramePointerSaveIndex();
1973 // If the frame pointer save index hasn't been defined yet.
1975 // Find out what the fix offset of the frame pointer save area.
1976 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
1978 // Allocate the frame index for frame pointer save area.
1979 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
1981 FI->setFramePointerSaveIndex(FPSI);
1985 SDOperand Chain = Op.getOperand(0);
1986 SDOperand Size = Op.getOperand(1);
1988 // Get the corect type for pointers.
1989 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1991 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
1992 DAG.getConstant(0, PtrVT), Size);
1993 // Construct a node for the frame pointer save index.
1994 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
1995 // Build a DYNALLOC node.
1996 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
1997 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
1998 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2002 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2004 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2005 // Not FP? Not a fsel.
2006 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2007 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2010 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2012 // Cannot handle SETEQ/SETNE.
2013 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2015 MVT::ValueType ResVT = Op.getValueType();
2016 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2017 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2018 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2020 // If the RHS of the comparison is a 0.0, we don't need to do the
2021 // subtraction at all.
2022 if (isFloatingPointZero(RHS))
2024 default: break; // SETUO etc aren't handled by fsel.
2028 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2032 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2033 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2034 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
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,
2045 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2050 default: break; // SETUO etc aren't handled by fsel.
2054 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2055 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2056 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2057 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2061 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2062 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2063 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2064 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2068 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2069 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2070 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2071 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2075 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2076 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2077 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2078 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2083 // FIXME: Split this code up when LegalizeDAGTypes lands.
2084 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2085 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2086 SDOperand Src = Op.getOperand(0);
2087 if (Src.getValueType() == MVT::f32)
2088 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2091 switch (Op.getValueType()) {
2092 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2094 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2097 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2101 // Convert the FP value to an int value through memory.
2102 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2104 // Emit a store to the stack slot.
2105 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2107 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2109 if (Op.getValueType() == MVT::i32)
2110 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2111 DAG.getConstant(4, FIPtr.getValueType()));
2112 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2115 static SDOperand LowerFP_ROUND_INREG(SDOperand Op, SelectionDAG &DAG) {
2116 assert(Op.getValueType() == MVT::ppcf128);
2117 SDNode *Node = Op.Val;
2118 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2119 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2120 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2121 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2123 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2124 // of the long double, and puts FPSCR back the way it was. We do not
2125 // actually model FPSCR.
2126 std::vector<MVT::ValueType> NodeTys;
2127 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2129 NodeTys.push_back(MVT::f64); // Return register
2130 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2131 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2132 MFFSreg = Result.getValue(0);
2133 InFlag = Result.getValue(1);
2136 NodeTys.push_back(MVT::Flag); // Returns a flag
2137 Ops[0] = DAG.getConstant(31, MVT::i32);
2139 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2140 InFlag = Result.getValue(0);
2143 NodeTys.push_back(MVT::Flag); // Returns a flag
2144 Ops[0] = DAG.getConstant(30, MVT::i32);
2146 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2147 InFlag = Result.getValue(0);
2150 NodeTys.push_back(MVT::f64); // result of add
2151 NodeTys.push_back(MVT::Flag); // Returns a flag
2155 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2156 FPreg = Result.getValue(0);
2157 InFlag = Result.getValue(1);
2160 NodeTys.push_back(MVT::f64);
2161 Ops[0] = DAG.getConstant(1, MVT::i32);
2165 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2166 FPreg = Result.getValue(0);
2168 // We know the low half is about to be thrown away, so just use something
2170 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2173 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2174 if (Op.getOperand(0).getValueType() == MVT::i64) {
2175 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2176 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2177 if (Op.getValueType() == MVT::f32)
2178 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2182 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2183 "Unhandled SINT_TO_FP type in custom expander!");
2184 // Since we only generate this in 64-bit mode, we can take advantage of
2185 // 64-bit registers. In particular, sign extend the input value into the
2186 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2187 // then lfd it and fcfid it.
2188 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2189 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2190 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2191 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2193 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2196 // STD the extended value into the stack slot.
2197 MemOperand MO(PseudoSourceValue::getFixedStack(),
2198 MemOperand::MOStore, FrameIdx, 8, 8);
2199 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2200 DAG.getEntryNode(), Ext64, FIdx,
2201 DAG.getMemOperand(MO));
2202 // Load the value as a double.
2203 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2205 // FCFID it and return it.
2206 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2207 if (Op.getValueType() == MVT::f32)
2208 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2212 static SDOperand LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2214 The rounding mode is in bits 30:31 of FPSR, and has the following
2221 FLT_ROUNDS, on the other hand, expects the following:
2228 To perform the conversion, we do:
2229 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2232 MachineFunction &MF = DAG.getMachineFunction();
2233 MVT::ValueType VT = Op.getValueType();
2234 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2235 std::vector<MVT::ValueType> NodeTys;
2236 SDOperand MFFSreg, InFlag;
2238 // Save FP Control Word to register
2239 NodeTys.push_back(MVT::f64); // return register
2240 NodeTys.push_back(MVT::Flag); // unused in this context
2241 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2243 // Save FP register to stack slot
2244 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2245 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2246 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2247 StackSlot, NULL, 0);
2249 // Load FP Control Word from low 32 bits of stack slot.
2250 SDOperand Four = DAG.getConstant(4, PtrVT);
2251 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2252 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2254 // Transform as necessary
2256 DAG.getNode(ISD::AND, MVT::i32,
2257 CWD, DAG.getConstant(3, MVT::i32));
2259 DAG.getNode(ISD::SRL, MVT::i32,
2260 DAG.getNode(ISD::AND, MVT::i32,
2261 DAG.getNode(ISD::XOR, MVT::i32,
2262 CWD, DAG.getConstant(3, MVT::i32)),
2263 DAG.getConstant(3, MVT::i32)),
2264 DAG.getConstant(1, MVT::i8));
2267 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2269 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2270 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2273 static SDOperand LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2274 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2275 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
2277 // Expand into a bunch of logical ops. Note that these ops
2278 // depend on the PPC behavior for oversized shift amounts.
2279 SDOperand Lo = Op.getOperand(0);
2280 SDOperand Hi = Op.getOperand(1);
2281 SDOperand Amt = Op.getOperand(2);
2283 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2284 DAG.getConstant(32, MVT::i32), Amt);
2285 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
2286 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
2287 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2288 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2289 DAG.getConstant(-32U, MVT::i32));
2290 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
2291 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2292 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
2293 SDOperand OutOps[] = { OutLo, OutHi };
2294 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2298 static SDOperand LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2299 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2300 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRL!");
2302 // Otherwise, expand into a bunch of logical ops. Note that these ops
2303 // depend on the PPC behavior for oversized shift amounts.
2304 SDOperand Lo = Op.getOperand(0);
2305 SDOperand Hi = Op.getOperand(1);
2306 SDOperand Amt = Op.getOperand(2);
2308 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2309 DAG.getConstant(32, MVT::i32), Amt);
2310 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2311 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2312 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2313 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2314 DAG.getConstant(-32U, MVT::i32));
2315 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
2316 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2317 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
2318 SDOperand OutOps[] = { OutLo, OutHi };
2319 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2323 static SDOperand LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2324 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2325 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
2327 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
2328 SDOperand Lo = Op.getOperand(0);
2329 SDOperand Hi = Op.getOperand(1);
2330 SDOperand Amt = Op.getOperand(2);
2332 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2333 DAG.getConstant(32, MVT::i32), Amt);
2334 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2335 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2336 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2337 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2338 DAG.getConstant(-32U, MVT::i32));
2339 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
2340 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
2341 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
2342 Tmp4, Tmp6, ISD::SETLE);
2343 SDOperand OutOps[] = { OutLo, OutHi };
2344 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2348 //===----------------------------------------------------------------------===//
2349 // Vector related lowering.
2352 // If this is a vector of constants or undefs, get the bits. A bit in
2353 // UndefBits is set if the corresponding element of the vector is an
2354 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2355 // zero. Return true if this is not an array of constants, false if it is.
2357 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2358 uint64_t UndefBits[2]) {
2359 // Start with zero'd results.
2360 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2362 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2363 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2364 SDOperand OpVal = BV->getOperand(i);
2366 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2367 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2369 uint64_t EltBits = 0;
2370 if (OpVal.getOpcode() == ISD::UNDEF) {
2371 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2372 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2374 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2375 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2376 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2377 assert(CN->getValueType(0) == MVT::f32 &&
2378 "Only one legal FP vector type!");
2379 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2381 // Nonconstant element.
2385 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2388 //printf("%llx %llx %llx %llx\n",
2389 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2393 // If this is a splat (repetition) of a value across the whole vector, return
2394 // the smallest size that splats it. For example, "0x01010101010101..." is a
2395 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2396 // SplatSize = 1 byte.
2397 static bool isConstantSplat(const uint64_t Bits128[2],
2398 const uint64_t Undef128[2],
2399 unsigned &SplatBits, unsigned &SplatUndef,
2400 unsigned &SplatSize) {
2402 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2403 // the same as the lower 64-bits, ignoring undefs.
2404 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2405 return false; // Can't be a splat if two pieces don't match.
2407 uint64_t Bits64 = Bits128[0] | Bits128[1];
2408 uint64_t Undef64 = Undef128[0] & Undef128[1];
2410 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2412 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2413 return false; // Can't be a splat if two pieces don't match.
2415 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2416 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2418 // If the top 16-bits are different than the lower 16-bits, ignoring
2419 // undefs, we have an i32 splat.
2420 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2422 SplatUndef = Undef32;
2427 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2428 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2430 // If the top 8-bits are different than the lower 8-bits, ignoring
2431 // undefs, we have an i16 splat.
2432 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2434 SplatUndef = Undef16;
2439 // Otherwise, we have an 8-bit splat.
2440 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2441 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2446 /// BuildSplatI - Build a canonical splati of Val with an element size of
2447 /// SplatSize. Cast the result to VT.
2448 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2449 SelectionDAG &DAG) {
2450 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2452 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2453 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2456 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2458 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2462 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2464 // Build a canonical splat for this value.
2465 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2466 SmallVector<SDOperand, 8> Ops;
2467 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2468 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2469 &Ops[0], Ops.size());
2470 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2473 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2474 /// specified intrinsic ID.
2475 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2477 MVT::ValueType DestVT = MVT::Other) {
2478 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2479 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2480 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2483 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2484 /// specified intrinsic ID.
2485 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2486 SDOperand Op2, SelectionDAG &DAG,
2487 MVT::ValueType DestVT = MVT::Other) {
2488 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2489 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2490 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2494 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2495 /// amount. The result has the specified value type.
2496 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2497 MVT::ValueType VT, SelectionDAG &DAG) {
2498 // Force LHS/RHS to be the right type.
2499 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2500 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2503 for (unsigned i = 0; i != 16; ++i)
2504 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2505 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2506 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2507 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2510 // If this is a case we can't handle, return null and let the default
2511 // expansion code take care of it. If we CAN select this case, and if it
2512 // selects to a single instruction, return Op. Otherwise, if we can codegen
2513 // this case more efficiently than a constant pool load, lower it to the
2514 // sequence of ops that should be used.
2515 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2516 // If this is a vector of constants or undefs, get the bits. A bit in
2517 // UndefBits is set if the corresponding element of the vector is an
2518 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2520 uint64_t VectorBits[2];
2521 uint64_t UndefBits[2];
2522 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2523 return SDOperand(); // Not a constant vector.
2525 // If this is a splat (repetition) of a value across the whole vector, return
2526 // the smallest size that splats it. For example, "0x01010101010101..." is a
2527 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2528 // SplatSize = 1 byte.
2529 unsigned SplatBits, SplatUndef, SplatSize;
2530 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2531 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2533 // First, handle single instruction cases.
2536 if (SplatBits == 0) {
2537 // Canonicalize all zero vectors to be v4i32.
2538 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2539 SDOperand Z = DAG.getConstant(0, MVT::i32);
2540 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2541 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2546 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2547 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2548 if (SextVal >= -16 && SextVal <= 15)
2549 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2552 // Two instruction sequences.
2554 // If this value is in the range [-32,30] and is even, use:
2555 // tmp = VSPLTI[bhw], result = add tmp, tmp
2556 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2557 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2558 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2561 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2562 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2564 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2565 // Make -1 and vspltisw -1:
2566 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2568 // Make the VSLW intrinsic, computing 0x8000_0000.
2569 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2572 // xor by OnesV to invert it.
2573 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2574 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2577 // Check to see if this is a wide variety of vsplti*, binop self cases.
2578 unsigned SplatBitSize = SplatSize*8;
2579 static const signed char SplatCsts[] = {
2580 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2581 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2584 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2585 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2586 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2587 int i = SplatCsts[idx];
2589 // Figure out what shift amount will be used by altivec if shifted by i in
2591 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2593 // vsplti + shl self.
2594 if (SextVal == (i << (int)TypeShiftAmt)) {
2595 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2596 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2597 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2598 Intrinsic::ppc_altivec_vslw
2600 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2601 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2604 // vsplti + srl self.
2605 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2606 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2607 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2608 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2609 Intrinsic::ppc_altivec_vsrw
2611 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2612 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2615 // vsplti + sra self.
2616 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2617 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2618 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2619 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2620 Intrinsic::ppc_altivec_vsraw
2622 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2623 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2626 // vsplti + rol self.
2627 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2628 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2629 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2630 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2631 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2632 Intrinsic::ppc_altivec_vrlw
2634 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2635 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2638 // t = vsplti c, result = vsldoi t, t, 1
2639 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2640 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2641 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2643 // t = vsplti c, result = vsldoi t, t, 2
2644 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2645 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2646 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2648 // t = vsplti c, result = vsldoi t, t, 3
2649 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2650 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2651 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2655 // Three instruction sequences.
2657 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2658 if (SextVal >= 0 && SextVal <= 31) {
2659 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2660 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2661 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2662 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2664 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2665 if (SextVal >= -31 && SextVal <= 0) {
2666 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2667 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2668 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2669 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2676 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2677 /// the specified operations to build the shuffle.
2678 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2679 SDOperand RHS, SelectionDAG &DAG) {
2680 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2681 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2682 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2685 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2697 if (OpNum == OP_COPY) {
2698 if (LHSID == (1*9+2)*9+3) return LHS;
2699 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2703 SDOperand OpLHS, OpRHS;
2704 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2705 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2707 unsigned ShufIdxs[16];
2709 default: assert(0 && "Unknown i32 permute!");
2711 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2712 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2713 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2714 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2717 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2718 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2719 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2720 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2723 for (unsigned i = 0; i != 16; ++i)
2724 ShufIdxs[i] = (i&3)+0;
2727 for (unsigned i = 0; i != 16; ++i)
2728 ShufIdxs[i] = (i&3)+4;
2731 for (unsigned i = 0; i != 16; ++i)
2732 ShufIdxs[i] = (i&3)+8;
2735 for (unsigned i = 0; i != 16; ++i)
2736 ShufIdxs[i] = (i&3)+12;
2739 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2741 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2743 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2746 for (unsigned i = 0; i != 16; ++i)
2747 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2749 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2750 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2753 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2754 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2755 /// return the code it can be lowered into. Worst case, it can always be
2756 /// lowered into a vperm.
2757 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2758 SDOperand V1 = Op.getOperand(0);
2759 SDOperand V2 = Op.getOperand(1);
2760 SDOperand PermMask = Op.getOperand(2);
2762 // Cases that are handled by instructions that take permute immediates
2763 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2764 // selected by the instruction selector.
2765 if (V2.getOpcode() == ISD::UNDEF) {
2766 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2767 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2768 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2769 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2770 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2771 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2772 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2773 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2774 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2775 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2776 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2777 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2782 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2783 // and produce a fixed permutation. If any of these match, do not lower to
2785 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2786 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2787 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2788 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2789 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2790 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2791 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2792 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2793 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2796 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2797 // perfect shuffle table to emit an optimal matching sequence.
2798 unsigned PFIndexes[4];
2799 bool isFourElementShuffle = true;
2800 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2801 unsigned EltNo = 8; // Start out undef.
2802 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2803 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2804 continue; // Undef, ignore it.
2806 unsigned ByteSource =
2807 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2808 if ((ByteSource & 3) != j) {
2809 isFourElementShuffle = false;
2814 EltNo = ByteSource/4;
2815 } else if (EltNo != ByteSource/4) {
2816 isFourElementShuffle = false;
2820 PFIndexes[i] = EltNo;
2823 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
2824 // perfect shuffle vector to determine if it is cost effective to do this as
2825 // discrete instructions, or whether we should use a vperm.
2826 if (isFourElementShuffle) {
2827 // Compute the index in the perfect shuffle table.
2828 unsigned PFTableIndex =
2829 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2831 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2832 unsigned Cost = (PFEntry >> 30);
2834 // Determining when to avoid vperm is tricky. Many things affect the cost
2835 // of vperm, particularly how many times the perm mask needs to be computed.
2836 // For example, if the perm mask can be hoisted out of a loop or is already
2837 // used (perhaps because there are multiple permutes with the same shuffle
2838 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
2839 // the loop requires an extra register.
2841 // As a compromise, we only emit discrete instructions if the shuffle can be
2842 // generated in 3 or fewer operations. When we have loop information
2843 // available, if this block is within a loop, we should avoid using vperm
2844 // for 3-operation perms and use a constant pool load instead.
2846 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
2849 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
2850 // vector that will get spilled to the constant pool.
2851 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
2853 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
2854 // that it is in input element units, not in bytes. Convert now.
2855 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
2856 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
2858 SmallVector<SDOperand, 16> ResultMask;
2859 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
2861 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
2864 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
2866 for (unsigned j = 0; j != BytesPerElement; ++j)
2867 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
2871 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2872 &ResultMask[0], ResultMask.size());
2873 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
2876 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
2877 /// altivec comparison. If it is, return true and fill in Opc/isDot with
2878 /// information about the intrinsic.
2879 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
2881 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
2884 switch (IntrinsicID) {
2885 default: return false;
2886 // Comparison predicates.
2887 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
2888 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
2889 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
2890 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
2891 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
2892 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
2893 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
2894 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
2895 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
2896 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
2897 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
2898 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
2899 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
2901 // Normal Comparisons.
2902 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
2903 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
2904 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
2905 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
2906 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
2907 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
2908 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
2909 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
2910 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
2911 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
2912 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
2913 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
2914 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
2919 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
2920 /// lower, do it, otherwise return null.
2921 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
2922 // If this is a lowered altivec predicate compare, CompareOpc is set to the
2923 // opcode number of the comparison.
2926 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
2927 return SDOperand(); // Don't custom lower most intrinsics.
2929 // If this is a non-dot comparison, make the VCMP node and we are done.
2931 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
2932 Op.getOperand(1), Op.getOperand(2),
2933 DAG.getConstant(CompareOpc, MVT::i32));
2934 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
2937 // Create the PPCISD altivec 'dot' comparison node.
2939 Op.getOperand(2), // LHS
2940 Op.getOperand(3), // RHS
2941 DAG.getConstant(CompareOpc, MVT::i32)
2943 std::vector<MVT::ValueType> VTs;
2944 VTs.push_back(Op.getOperand(2).getValueType());
2945 VTs.push_back(MVT::Flag);
2946 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
2948 // Now that we have the comparison, emit a copy from the CR to a GPR.
2949 // This is flagged to the above dot comparison.
2950 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
2951 DAG.getRegister(PPC::CR6, MVT::i32),
2952 CompNode.getValue(1));
2954 // Unpack the result based on how the target uses it.
2955 unsigned BitNo; // Bit # of CR6.
2956 bool InvertBit; // Invert result?
2957 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
2958 default: // Can't happen, don't crash on invalid number though.
2959 case 0: // Return the value of the EQ bit of CR6.
2960 BitNo = 0; InvertBit = false;
2962 case 1: // Return the inverted value of the EQ bit of CR6.
2963 BitNo = 0; InvertBit = true;
2965 case 2: // Return the value of the LT bit of CR6.
2966 BitNo = 2; InvertBit = false;
2968 case 3: // Return the inverted value of the LT bit of CR6.
2969 BitNo = 2; InvertBit = true;
2973 // Shift the bit into the low position.
2974 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
2975 DAG.getConstant(8-(3-BitNo), MVT::i32));
2977 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
2978 DAG.getConstant(1, MVT::i32));
2980 // If we are supposed to, toggle the bit.
2982 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
2983 DAG.getConstant(1, MVT::i32));
2987 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2988 // Create a stack slot that is 16-byte aligned.
2989 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2990 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
2991 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2992 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2994 // Store the input value into Value#0 of the stack slot.
2995 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
2996 Op.getOperand(0), FIdx, NULL, 0);
2998 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3001 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3002 if (Op.getValueType() == MVT::v4i32) {
3003 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3005 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3006 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3008 SDOperand RHSSwap = // = vrlw RHS, 16
3009 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3011 // Shrinkify inputs to v8i16.
3012 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3013 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3014 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3016 // Low parts multiplied together, generating 32-bit results (we ignore the
3018 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3019 LHS, RHS, DAG, MVT::v4i32);
3021 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3022 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3023 // Shift the high parts up 16 bits.
3024 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3025 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3026 } else if (Op.getValueType() == MVT::v8i16) {
3027 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3029 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3031 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3032 LHS, RHS, Zero, DAG);
3033 } else if (Op.getValueType() == MVT::v16i8) {
3034 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3036 // Multiply the even 8-bit parts, producing 16-bit sums.
3037 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3038 LHS, RHS, DAG, MVT::v8i16);
3039 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3041 // Multiply the odd 8-bit parts, producing 16-bit sums.
3042 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3043 LHS, RHS, DAG, MVT::v8i16);
3044 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3046 // Merge the results together.
3048 for (unsigned i = 0; i != 8; ++i) {
3049 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3050 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3052 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3053 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3055 assert(0 && "Unknown mul to lower!");
3060 /// LowerOperation - Provide custom lowering hooks for some operations.
3062 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3063 switch (Op.getOpcode()) {
3064 default: assert(0 && "Wasn't expecting to be able to lower this!");
3065 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3066 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3067 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3068 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3069 case ISD::SETCC: return LowerSETCC(Op, DAG);
3071 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3072 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3075 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3076 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3078 case ISD::FORMAL_ARGUMENTS:
3079 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3080 VarArgsStackOffset, VarArgsNumGPR,
3081 VarArgsNumFPR, PPCSubTarget);
3083 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
3084 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3085 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3086 case ISD::DYNAMIC_STACKALLOC:
3087 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3089 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3090 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3091 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3092 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3093 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3095 // Lower 64-bit shifts.
3096 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3097 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3098 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3100 // Vector-related lowering.
3101 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3102 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3103 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3104 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3105 case ISD::MUL: return LowerMUL(Op, DAG);
3107 // Frame & Return address.
3108 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3109 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3114 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3115 switch (N->getOpcode()) {
3116 default: assert(0 && "Wasn't expecting to be able to lower this!");
3117 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3122 //===----------------------------------------------------------------------===//
3123 // Other Lowering Code
3124 //===----------------------------------------------------------------------===//
3127 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3128 MachineBasicBlock *BB) {
3129 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3130 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3131 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3132 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3133 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3134 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3135 "Unexpected instr type to insert");
3137 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3138 // control-flow pattern. The incoming instruction knows the destination vreg
3139 // to set, the condition code register to branch on, the true/false values to
3140 // select between, and a branch opcode to use.
3141 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3142 ilist<MachineBasicBlock>::iterator It = BB;
3148 // cmpTY ccX, r1, r2
3150 // fallthrough --> copy0MBB
3151 MachineBasicBlock *thisMBB = BB;
3152 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3153 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3154 unsigned SelectPred = MI->getOperand(4).getImm();
3155 BuildMI(BB, TII->get(PPC::BCC))
3156 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3157 MachineFunction *F = BB->getParent();
3158 F->getBasicBlockList().insert(It, copy0MBB);
3159 F->getBasicBlockList().insert(It, sinkMBB);
3160 // Update machine-CFG edges by first adding all successors of the current
3161 // block to the new block which will contain the Phi node for the select.
3162 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3163 e = BB->succ_end(); i != e; ++i)
3164 sinkMBB->addSuccessor(*i);
3165 // Next, remove all successors of the current block, and add the true
3166 // and fallthrough blocks as its successors.
3167 while(!BB->succ_empty())
3168 BB->removeSuccessor(BB->succ_begin());
3169 BB->addSuccessor(copy0MBB);
3170 BB->addSuccessor(sinkMBB);
3173 // %FalseValue = ...
3174 // # fallthrough to sinkMBB
3177 // Update machine-CFG edges
3178 BB->addSuccessor(sinkMBB);
3181 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3184 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3185 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3186 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3188 delete MI; // The pseudo instruction is gone now.
3192 //===----------------------------------------------------------------------===//
3193 // Target Optimization Hooks
3194 //===----------------------------------------------------------------------===//
3196 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3197 DAGCombinerInfo &DCI) const {
3198 TargetMachine &TM = getTargetMachine();
3199 SelectionDAG &DAG = DCI.DAG;
3200 switch (N->getOpcode()) {
3203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3204 if (C->getValue() == 0) // 0 << V -> 0.
3205 return N->getOperand(0);
3209 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3210 if (C->getValue() == 0) // 0 >>u V -> 0.
3211 return N->getOperand(0);
3215 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3216 if (C->getValue() == 0 || // 0 >>s V -> 0.
3217 C->isAllOnesValue()) // -1 >>s V -> -1.
3218 return N->getOperand(0);
3222 case ISD::SINT_TO_FP:
3223 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3224 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3225 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3226 // We allow the src/dst to be either f32/f64, but the intermediate
3227 // type must be i64.
3228 if (N->getOperand(0).getValueType() == MVT::i64 &&
3229 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3230 SDOperand Val = N->getOperand(0).getOperand(0);
3231 if (Val.getValueType() == MVT::f32) {
3232 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3233 DCI.AddToWorklist(Val.Val);
3236 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3237 DCI.AddToWorklist(Val.Val);
3238 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3239 DCI.AddToWorklist(Val.Val);
3240 if (N->getValueType(0) == MVT::f32) {
3241 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3242 DAG.getIntPtrConstant(0));
3243 DCI.AddToWorklist(Val.Val);
3246 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3247 // If the intermediate type is i32, we can avoid the load/store here
3254 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3255 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3256 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3257 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3258 N->getOperand(1).getValueType() == MVT::i32 &&
3259 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3260 SDOperand Val = N->getOperand(1).getOperand(0);
3261 if (Val.getValueType() == MVT::f32) {
3262 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3263 DCI.AddToWorklist(Val.Val);
3265 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3266 DCI.AddToWorklist(Val.Val);
3268 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3269 N->getOperand(2), N->getOperand(3));
3270 DCI.AddToWorklist(Val.Val);
3274 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3275 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3276 N->getOperand(1).Val->hasOneUse() &&
3277 (N->getOperand(1).getValueType() == MVT::i32 ||
3278 N->getOperand(1).getValueType() == MVT::i16)) {
3279 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3280 // Do an any-extend to 32-bits if this is a half-word input.
3281 if (BSwapOp.getValueType() == MVT::i16)
3282 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3284 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3285 N->getOperand(2), N->getOperand(3),
3286 DAG.getValueType(N->getOperand(1).getValueType()));
3290 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3291 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3292 N->getOperand(0).hasOneUse() &&
3293 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3294 SDOperand Load = N->getOperand(0);
3295 LoadSDNode *LD = cast<LoadSDNode>(Load);
3296 // Create the byte-swapping load.
3297 std::vector<MVT::ValueType> VTs;
3298 VTs.push_back(MVT::i32);
3299 VTs.push_back(MVT::Other);
3300 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3302 LD->getChain(), // Chain
3303 LD->getBasePtr(), // Ptr
3305 DAG.getValueType(N->getValueType(0)) // VT
3307 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3309 // If this is an i16 load, insert the truncate.
3310 SDOperand ResVal = BSLoad;
3311 if (N->getValueType(0) == MVT::i16)
3312 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3314 // First, combine the bswap away. This makes the value produced by the
3316 DCI.CombineTo(N, ResVal);
3318 // Next, combine the load away, we give it a bogus result value but a real
3319 // chain result. The result value is dead because the bswap is dead.
3320 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3322 // Return N so it doesn't get rechecked!
3323 return SDOperand(N, 0);
3327 case PPCISD::VCMP: {
3328 // If a VCMPo node already exists with exactly the same operands as this
3329 // node, use its result instead of this node (VCMPo computes both a CR6 and
3330 // a normal output).
3332 if (!N->getOperand(0).hasOneUse() &&
3333 !N->getOperand(1).hasOneUse() &&
3334 !N->getOperand(2).hasOneUse()) {
3336 // Scan all of the users of the LHS, looking for VCMPo's that match.
3337 SDNode *VCMPoNode = 0;
3339 SDNode *LHSN = N->getOperand(0).Val;
3340 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3342 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3343 (*UI)->getOperand(1) == N->getOperand(1) &&
3344 (*UI)->getOperand(2) == N->getOperand(2) &&
3345 (*UI)->getOperand(0) == N->getOperand(0)) {
3350 // If there is no VCMPo node, or if the flag value has a single use, don't
3352 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3355 // Look at the (necessarily single) use of the flag value. If it has a
3356 // chain, this transformation is more complex. Note that multiple things
3357 // could use the value result, which we should ignore.
3358 SDNode *FlagUser = 0;
3359 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3360 FlagUser == 0; ++UI) {
3361 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3363 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3364 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3371 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3372 // give up for right now.
3373 if (FlagUser->getOpcode() == PPCISD::MFCR)
3374 return SDOperand(VCMPoNode, 0);
3379 // If this is a branch on an altivec predicate comparison, lower this so
3380 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3381 // lowering is done pre-legalize, because the legalizer lowers the predicate
3382 // compare down to code that is difficult to reassemble.
3383 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3384 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3388 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3389 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3390 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3391 assert(isDot && "Can't compare against a vector result!");
3393 // If this is a comparison against something other than 0/1, then we know
3394 // that the condition is never/always true.
3395 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3396 if (Val != 0 && Val != 1) {
3397 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3398 return N->getOperand(0);
3399 // Always !=, turn it into an unconditional branch.
3400 return DAG.getNode(ISD::BR, MVT::Other,
3401 N->getOperand(0), N->getOperand(4));
3404 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3406 // Create the PPCISD altivec 'dot' comparison node.
3407 std::vector<MVT::ValueType> VTs;
3409 LHS.getOperand(2), // LHS of compare
3410 LHS.getOperand(3), // RHS of compare
3411 DAG.getConstant(CompareOpc, MVT::i32)
3413 VTs.push_back(LHS.getOperand(2).getValueType());
3414 VTs.push_back(MVT::Flag);
3415 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3417 // Unpack the result based on how the target uses it.
3418 PPC::Predicate CompOpc;
3419 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3420 default: // Can't happen, don't crash on invalid number though.
3421 case 0: // Branch on the value of the EQ bit of CR6.
3422 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3424 case 1: // Branch on the inverted value of the EQ bit of CR6.
3425 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3427 case 2: // Branch on the value of the LT bit of CR6.
3428 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3430 case 3: // Branch on the inverted value of the LT bit of CR6.
3431 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3435 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3436 DAG.getConstant(CompOpc, MVT::i32),
3437 DAG.getRegister(PPC::CR6, MVT::i32),
3438 N->getOperand(4), CompNode.getValue(1));
3447 //===----------------------------------------------------------------------===//
3448 // Inline Assembly Support
3449 //===----------------------------------------------------------------------===//
3451 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3455 const SelectionDAG &DAG,
3456 unsigned Depth) const {
3457 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3458 switch (Op.getOpcode()) {
3460 case PPCISD::LBRX: {
3461 // lhbrx is known to have the top bits cleared out.
3462 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3463 KnownZero = 0xFFFF0000;
3466 case ISD::INTRINSIC_WO_CHAIN: {
3467 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3469 case Intrinsic::ppc_altivec_vcmpbfp_p:
3470 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3471 case Intrinsic::ppc_altivec_vcmpequb_p:
3472 case Intrinsic::ppc_altivec_vcmpequh_p:
3473 case Intrinsic::ppc_altivec_vcmpequw_p:
3474 case Intrinsic::ppc_altivec_vcmpgefp_p:
3475 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3476 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3477 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3478 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3479 case Intrinsic::ppc_altivec_vcmpgtub_p:
3480 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3481 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3482 KnownZero = ~1U; // All bits but the low one are known to be zero.
3490 /// getConstraintType - Given a constraint, return the type of
3491 /// constraint it is for this target.
3492 PPCTargetLowering::ConstraintType
3493 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3494 if (Constraint.size() == 1) {
3495 switch (Constraint[0]) {
3502 return C_RegisterClass;
3505 return TargetLowering::getConstraintType(Constraint);
3508 std::pair<unsigned, const TargetRegisterClass*>
3509 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3510 MVT::ValueType VT) const {
3511 if (Constraint.size() == 1) {
3512 // GCC RS6000 Constraint Letters
3513 switch (Constraint[0]) {
3516 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3517 return std::make_pair(0U, PPC::G8RCRegisterClass);
3518 return std::make_pair(0U, PPC::GPRCRegisterClass);
3521 return std::make_pair(0U, PPC::F4RCRegisterClass);
3522 else if (VT == MVT::f64)
3523 return std::make_pair(0U, PPC::F8RCRegisterClass);
3526 return std::make_pair(0U, PPC::VRRCRegisterClass);
3528 return std::make_pair(0U, PPC::CRRCRegisterClass);
3532 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3536 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3537 /// vector. If it is invalid, don't add anything to Ops.
3538 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3539 std::vector<SDOperand>&Ops,
3540 SelectionDAG &DAG) {
3541 SDOperand Result(0,0);
3552 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3553 if (!CST) return; // Must be an immediate to match.
3554 unsigned Value = CST->getValue();
3556 default: assert(0 && "Unknown constraint letter!");
3557 case 'I': // "I" is a signed 16-bit constant.
3558 if ((short)Value == (int)Value)
3559 Result = DAG.getTargetConstant(Value, Op.getValueType());
3561 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3562 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3563 if ((short)Value == 0)
3564 Result = DAG.getTargetConstant(Value, Op.getValueType());
3566 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3567 if ((Value >> 16) == 0)
3568 Result = DAG.getTargetConstant(Value, Op.getValueType());
3570 case 'M': // "M" is a constant that is greater than 31.
3572 Result = DAG.getTargetConstant(Value, Op.getValueType());
3574 case 'N': // "N" is a positive constant that is an exact power of two.
3575 if ((int)Value > 0 && isPowerOf2_32(Value))
3576 Result = DAG.getTargetConstant(Value, Op.getValueType());
3578 case 'O': // "O" is the constant zero.
3580 Result = DAG.getTargetConstant(Value, Op.getValueType());
3582 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3583 if ((short)-Value == (int)-Value)
3584 Result = DAG.getTargetConstant(Value, Op.getValueType());
3592 Ops.push_back(Result);
3596 // Handle standard constraint letters.
3597 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3600 // isLegalAddressingMode - Return true if the addressing mode represented
3601 // by AM is legal for this target, for a load/store of the specified type.
3602 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3603 const Type *Ty) const {
3604 // FIXME: PPC does not allow r+i addressing modes for vectors!
3606 // PPC allows a sign-extended 16-bit immediate field.
3607 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3610 // No global is ever allowed as a base.
3614 // PPC only support r+r,
3616 case 0: // "r+i" or just "i", depending on HasBaseReg.
3619 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3621 // Otherwise we have r+r or r+i.
3624 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3626 // Allow 2*r as r+r.
3629 // No other scales are supported.
3636 /// isLegalAddressImmediate - Return true if the integer value can be used
3637 /// as the offset of the target addressing mode for load / store of the
3639 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3640 // PPC allows a sign-extended 16-bit immediate field.
3641 return (V > -(1 << 16) && V < (1 << 16)-1);
3644 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3648 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3649 // Depths > 0 not supported yet!
3650 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3653 MachineFunction &MF = DAG.getMachineFunction();
3654 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3655 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3657 bool isPPC64 = PPCSubTarget.isPPC64();
3659 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3661 // Set up a frame object for the return address.
3662 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3664 // Remember it for next time.
3665 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3667 // Make sure the function really does not optimize away the store of the RA
3669 FuncInfo->setLRStoreRequired();
3672 // Just load the return address off the stack.
3673 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3674 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3677 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3678 // Depths > 0 not supported yet!
3679 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3682 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3683 bool isPPC64 = PtrVT == MVT::i64;
3685 MachineFunction &MF = DAG.getMachineFunction();
3686 MachineFrameInfo *MFI = MF.getFrameInfo();
3687 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3688 && MFI->getStackSize();
3691 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3694 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,