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);
85 setOperationAction(ISD::PREFETCH, MVT::Other, Expand);
87 // PowerPC has no SREM/UREM instructions
88 setOperationAction(ISD::SREM, MVT::i32, Expand);
89 setOperationAction(ISD::UREM, MVT::i32, Expand);
90 setOperationAction(ISD::SREM, MVT::i64, Expand);
91 setOperationAction(ISD::UREM, MVT::i64, Expand);
93 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
94 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
95 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
96 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
97 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
98 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
99 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
100 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
101 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
103 // We don't support sin/cos/sqrt/fmod/pow
104 setOperationAction(ISD::FSIN , MVT::f64, Expand);
105 setOperationAction(ISD::FCOS , MVT::f64, Expand);
106 setOperationAction(ISD::FREM , MVT::f64, Expand);
107 setOperationAction(ISD::FPOW , MVT::f64, Expand);
108 setOperationAction(ISD::FSIN , MVT::f32, Expand);
109 setOperationAction(ISD::FCOS , MVT::f32, Expand);
110 setOperationAction(ISD::FREM , MVT::f32, Expand);
111 setOperationAction(ISD::FPOW , MVT::f32, Expand);
113 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
115 // If we're enabling GP optimizations, use hardware square root
116 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
117 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
118 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
121 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
122 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
124 // PowerPC does not have BSWAP, CTPOP or CTTZ
125 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
126 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
127 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
128 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
129 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
130 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
132 // PowerPC does not have ROTR
133 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
135 // PowerPC does not have Select
136 setOperationAction(ISD::SELECT, MVT::i32, Expand);
137 setOperationAction(ISD::SELECT, MVT::i64, Expand);
138 setOperationAction(ISD::SELECT, MVT::f32, Expand);
139 setOperationAction(ISD::SELECT, MVT::f64, Expand);
141 // PowerPC wants to turn select_cc of FP into fsel when possible.
142 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
143 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
145 // PowerPC wants to optimize integer setcc a bit
146 setOperationAction(ISD::SETCC, MVT::i32, Custom);
148 // PowerPC does not have BRCOND which requires SetCC
149 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
151 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
153 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
154 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
156 // PowerPC does not have [U|S]INT_TO_FP
157 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
158 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
160 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
161 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
162 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
163 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
165 // We cannot sextinreg(i1). Expand to shifts.
166 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
168 // Support label based line numbers.
169 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
170 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
172 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
173 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
174 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
175 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
178 // We want to legalize GlobalAddress and ConstantPool nodes into the
179 // appropriate instructions to materialize the address.
180 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
181 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
182 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
183 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
184 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
185 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
186 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
187 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
189 // RET must be custom lowered, to meet ABI requirements
190 setOperationAction(ISD::RET , MVT::Other, Custom);
192 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
193 setOperationAction(ISD::VASTART , MVT::Other, Custom);
195 // VAARG is custom lowered with ELF 32 ABI
196 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
197 setOperationAction(ISD::VAARG, MVT::Other, Custom);
199 setOperationAction(ISD::VAARG, MVT::Other, Expand);
201 // Use the default implementation.
202 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
203 setOperationAction(ISD::VAEND , MVT::Other, Expand);
204 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
205 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
206 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
207 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
209 // We want to custom lower some of our intrinsics.
210 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
212 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
213 // They also have instructions for converting between i64 and fp.
214 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
215 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
216 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
217 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
218 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
220 // FIXME: disable this lowered code. This generates 64-bit register values,
221 // and we don't model the fact that the top part is clobbered by calls. We
222 // need to flag these together so that the value isn't live across a call.
223 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
225 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
226 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
228 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
229 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
232 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
233 // 64-bit PowerPC implementations can support i64 types directly
234 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
235 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
236 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
237 // 64-bit PowerPC wants to expand i128 shifts itself.
238 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
239 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
240 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
242 // 32-bit PowerPC wants to expand i64 shifts itself.
243 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
244 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
245 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
248 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
249 // First set operation action for all vector types to expand. Then we
250 // will selectively turn on ones that can be effectively codegen'd.
251 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
252 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
253 // add/sub are legal for all supported vector VT's.
254 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
255 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
257 // We promote all shuffles to v16i8.
258 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
259 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
261 // We promote all non-typed operations to v4i32.
262 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
263 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
264 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
265 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
266 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
267 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
268 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
269 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
270 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
271 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
272 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
273 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
275 // No other operations are legal.
276 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
277 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
278 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
279 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
280 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
281 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
282 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
283 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
284 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
285 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
286 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
287 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
288 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
289 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
290 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
291 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
292 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
293 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
294 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
297 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
298 // with merges, splats, etc.
299 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
301 setOperationAction(ISD::AND , MVT::v4i32, Legal);
302 setOperationAction(ISD::OR , MVT::v4i32, Legal);
303 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
304 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
305 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
306 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
308 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
309 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
310 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
311 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
313 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
314 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
315 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
316 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
318 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
319 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
321 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
322 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
323 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
324 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
327 setSetCCResultType(MVT::i32);
328 setShiftAmountType(MVT::i32);
329 setSetCCResultContents(ZeroOrOneSetCCResult);
331 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
332 setStackPointerRegisterToSaveRestore(PPC::X1);
333 setExceptionPointerRegister(PPC::X3);
334 setExceptionSelectorRegister(PPC::X4);
336 setStackPointerRegisterToSaveRestore(PPC::R1);
337 setExceptionPointerRegister(PPC::R3);
338 setExceptionSelectorRegister(PPC::R4);
341 // We have target-specific dag combine patterns for the following nodes:
342 setTargetDAGCombine(ISD::SINT_TO_FP);
343 setTargetDAGCombine(ISD::STORE);
344 setTargetDAGCombine(ISD::BR_CC);
345 setTargetDAGCombine(ISD::BSWAP);
347 // Darwin long double math library functions have $LDBL128 appended.
348 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
349 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
350 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
351 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
352 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
353 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
356 computeRegisterProperties();
359 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
360 /// function arguments in the caller parameter area.
361 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
362 TargetMachine &TM = getTargetMachine();
363 // Darwin passes everything on 4 byte boundary.
364 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
370 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
373 case PPCISD::FSEL: return "PPCISD::FSEL";
374 case PPCISD::FCFID: return "PPCISD::FCFID";
375 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
376 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
377 case PPCISD::STFIWX: return "PPCISD::STFIWX";
378 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
379 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
380 case PPCISD::VPERM: return "PPCISD::VPERM";
381 case PPCISD::Hi: return "PPCISD::Hi";
382 case PPCISD::Lo: return "PPCISD::Lo";
383 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
384 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
385 case PPCISD::SRL: return "PPCISD::SRL";
386 case PPCISD::SRA: return "PPCISD::SRA";
387 case PPCISD::SHL: return "PPCISD::SHL";
388 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
389 case PPCISD::STD_32: return "PPCISD::STD_32";
390 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
391 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
392 case PPCISD::MTCTR: return "PPCISD::MTCTR";
393 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
394 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
395 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
396 case PPCISD::MFCR: return "PPCISD::MFCR";
397 case PPCISD::VCMP: return "PPCISD::VCMP";
398 case PPCISD::VCMPo: return "PPCISD::VCMPo";
399 case PPCISD::LBRX: return "PPCISD::LBRX";
400 case PPCISD::STBRX: return "PPCISD::STBRX";
401 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
402 case PPCISD::MFFS: return "PPCISD::MFFS";
403 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
404 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
405 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
406 case PPCISD::MTFSF: return "PPCISD::MTFSF";
410 //===----------------------------------------------------------------------===//
411 // Node matching predicates, for use by the tblgen matching code.
412 //===----------------------------------------------------------------------===//
414 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
415 static bool isFloatingPointZero(SDOperand Op) {
416 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
417 return CFP->getValueAPF().isZero();
418 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
419 // Maybe this has already been legalized into the constant pool?
420 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
421 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
422 return CFP->getValueAPF().isZero();
427 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
428 /// true if Op is undef or if it matches the specified value.
429 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
430 return Op.getOpcode() == ISD::UNDEF ||
431 cast<ConstantSDNode>(Op)->getValue() == Val;
434 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
435 /// VPKUHUM instruction.
436 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
438 for (unsigned i = 0; i != 16; ++i)
439 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
442 for (unsigned i = 0; i != 8; ++i)
443 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
444 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
450 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
451 /// VPKUWUM instruction.
452 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
454 for (unsigned i = 0; i != 16; i += 2)
455 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
456 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
459 for (unsigned i = 0; i != 8; i += 2)
460 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
461 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
462 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
463 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
469 /// isVMerge - Common function, used to match vmrg* shuffles.
471 static bool isVMerge(SDNode *N, unsigned UnitSize,
472 unsigned LHSStart, unsigned RHSStart) {
473 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
474 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
475 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
476 "Unsupported merge size!");
478 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
479 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
480 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
481 LHSStart+j+i*UnitSize) ||
482 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
483 RHSStart+j+i*UnitSize))
489 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
490 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
491 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
493 return isVMerge(N, UnitSize, 8, 24);
494 return isVMerge(N, UnitSize, 8, 8);
497 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
498 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
499 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
501 return isVMerge(N, UnitSize, 0, 16);
502 return isVMerge(N, UnitSize, 0, 0);
506 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
507 /// amount, otherwise return -1.
508 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
509 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
510 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
511 // Find the first non-undef value in the shuffle mask.
513 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
516 if (i == 16) return -1; // all undef.
518 // Otherwise, check to see if the rest of the elements are consequtively
519 // numbered from this value.
520 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
521 if (ShiftAmt < i) return -1;
525 // Check the rest of the elements to see if they are consequtive.
526 for (++i; i != 16; ++i)
527 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
530 // Check the rest of the elements to see if they are consequtive.
531 for (++i; i != 16; ++i)
532 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
539 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
540 /// specifies a splat of a single element that is suitable for input to
541 /// VSPLTB/VSPLTH/VSPLTW.
542 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
543 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
544 N->getNumOperands() == 16 &&
545 (EltSize == 1 || EltSize == 2 || EltSize == 4));
547 // This is a splat operation if each element of the permute is the same, and
548 // if the value doesn't reference the second vector.
549 unsigned ElementBase = 0;
550 SDOperand Elt = N->getOperand(0);
551 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
552 ElementBase = EltV->getValue();
554 return false; // FIXME: Handle UNDEF elements too!
556 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
559 // Check that they are consequtive.
560 for (unsigned i = 1; i != EltSize; ++i) {
561 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
562 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
566 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
567 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
568 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
569 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
570 "Invalid VECTOR_SHUFFLE mask!");
571 for (unsigned j = 0; j != EltSize; ++j)
572 if (N->getOperand(i+j) != N->getOperand(j))
579 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
581 bool PPC::isAllNegativeZeroVector(SDNode *N) {
582 assert(N->getOpcode() == ISD::BUILD_VECTOR);
583 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
584 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
585 return CFP->getValueAPF().isNegZero();
589 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
590 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
591 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
592 assert(isSplatShuffleMask(N, EltSize));
593 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
596 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
597 /// by using a vspltis[bhw] instruction of the specified element size, return
598 /// the constant being splatted. The ByteSize field indicates the number of
599 /// bytes of each element [124] -> [bhw].
600 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
601 SDOperand OpVal(0, 0);
603 // If ByteSize of the splat is bigger than the element size of the
604 // build_vector, then we have a case where we are checking for a splat where
605 // multiple elements of the buildvector are folded together into a single
606 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
607 unsigned EltSize = 16/N->getNumOperands();
608 if (EltSize < ByteSize) {
609 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
610 SDOperand UniquedVals[4];
611 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
613 // See if all of the elements in the buildvector agree across.
614 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
615 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
616 // If the element isn't a constant, bail fully out.
617 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
620 if (UniquedVals[i&(Multiple-1)].Val == 0)
621 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
622 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
623 return SDOperand(); // no match.
626 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
627 // either constant or undef values that are identical for each chunk. See
628 // if these chunks can form into a larger vspltis*.
630 // Check to see if all of the leading entries are either 0 or -1. If
631 // neither, then this won't fit into the immediate field.
632 bool LeadingZero = true;
633 bool LeadingOnes = true;
634 for (unsigned i = 0; i != Multiple-1; ++i) {
635 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
637 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
638 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
640 // Finally, check the least significant entry.
642 if (UniquedVals[Multiple-1].Val == 0)
643 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
644 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
646 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
649 if (UniquedVals[Multiple-1].Val == 0)
650 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
651 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
652 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
653 return DAG.getTargetConstant(Val, MVT::i32);
659 // Check to see if this buildvec has a single non-undef value in its elements.
660 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
661 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
663 OpVal = N->getOperand(i);
664 else if (OpVal != N->getOperand(i))
668 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
670 unsigned ValSizeInBytes = 0;
672 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
673 Value = CN->getValue();
674 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
675 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
676 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
677 Value = FloatToBits(CN->getValueAPF().convertToFloat());
681 // If the splat value is larger than the element value, then we can never do
682 // this splat. The only case that we could fit the replicated bits into our
683 // immediate field for would be zero, and we prefer to use vxor for it.
684 if (ValSizeInBytes < ByteSize) return SDOperand();
686 // If the element value is larger than the splat value, cut it in half and
687 // check to see if the two halves are equal. Continue doing this until we
688 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
689 while (ValSizeInBytes > ByteSize) {
690 ValSizeInBytes >>= 1;
692 // If the top half equals the bottom half, we're still ok.
693 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
694 (Value & ((1 << (8*ValSizeInBytes))-1)))
698 // Properly sign extend the value.
699 int ShAmt = (4-ByteSize)*8;
700 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
702 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
703 if (MaskVal == 0) return SDOperand();
705 // Finally, if this value fits in a 5 bit sext field, return it
706 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
707 return DAG.getTargetConstant(MaskVal, MVT::i32);
711 //===----------------------------------------------------------------------===//
712 // Addressing Mode Selection
713 //===----------------------------------------------------------------------===//
715 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
716 /// or 64-bit immediate, and if the value can be accurately represented as a
717 /// sign extension from a 16-bit value. If so, this returns true and the
719 static bool isIntS16Immediate(SDNode *N, short &Imm) {
720 if (N->getOpcode() != ISD::Constant)
723 Imm = (short)cast<ConstantSDNode>(N)->getValue();
724 if (N->getValueType(0) == MVT::i32)
725 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
727 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
729 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
730 return isIntS16Immediate(Op.Val, Imm);
734 /// SelectAddressRegReg - Given the specified addressed, check to see if it
735 /// can be represented as an indexed [r+r] operation. Returns false if it
736 /// can be more efficiently represented with [r+imm].
737 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
741 if (N.getOpcode() == ISD::ADD) {
742 if (isIntS16Immediate(N.getOperand(1), imm))
744 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
747 Base = N.getOperand(0);
748 Index = N.getOperand(1);
750 } else if (N.getOpcode() == ISD::OR) {
751 if (isIntS16Immediate(N.getOperand(1), imm))
752 return false; // r+i can fold it if we can.
754 // If this is an or of disjoint bitfields, we can codegen this as an add
755 // (for better address arithmetic) if the LHS and RHS of the OR are provably
757 APInt LHSKnownZero, LHSKnownOne;
758 APInt RHSKnownZero, RHSKnownOne;
759 DAG.ComputeMaskedBits(N.getOperand(0),
760 APInt::getAllOnesValue(N.getOperand(0)
761 .getValueSizeInBits()),
762 LHSKnownZero, LHSKnownOne);
764 if (LHSKnownZero.getBoolValue()) {
765 DAG.ComputeMaskedBits(N.getOperand(1),
766 APInt::getAllOnesValue(N.getOperand(1)
767 .getValueSizeInBits()),
768 RHSKnownZero, RHSKnownOne);
769 // If all of the bits are known zero on the LHS or RHS, the add won't
771 if (~(LHSKnownZero | RHSKnownZero) == 0) {
772 Base = N.getOperand(0);
773 Index = N.getOperand(1);
782 /// Returns true if the address N can be represented by a base register plus
783 /// a signed 16-bit displacement [r+imm], and if it is not better
784 /// represented as reg+reg.
785 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
786 SDOperand &Base, SelectionDAG &DAG){
787 // If this can be more profitably realized as r+r, fail.
788 if (SelectAddressRegReg(N, Disp, Base, DAG))
791 if (N.getOpcode() == ISD::ADD) {
793 if (isIntS16Immediate(N.getOperand(1), imm)) {
794 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
795 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
796 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
798 Base = N.getOperand(0);
800 return true; // [r+i]
801 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
802 // Match LOAD (ADD (X, Lo(G))).
803 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
804 && "Cannot handle constant offsets yet!");
805 Disp = N.getOperand(1).getOperand(0); // The global address.
806 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
807 Disp.getOpcode() == ISD::TargetConstantPool ||
808 Disp.getOpcode() == ISD::TargetJumpTable);
809 Base = N.getOperand(0);
810 return true; // [&g+r]
812 } else if (N.getOpcode() == ISD::OR) {
814 if (isIntS16Immediate(N.getOperand(1), imm)) {
815 // If this is an or of disjoint bitfields, we can codegen this as an add
816 // (for better address arithmetic) if the LHS and RHS of the OR are
817 // provably disjoint.
818 APInt LHSKnownZero, LHSKnownOne;
819 DAG.ComputeMaskedBits(N.getOperand(0),
820 APInt::getAllOnesValue(32),
821 LHSKnownZero, LHSKnownOne);
822 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
823 // If all of the bits are known zero on the LHS or RHS, the add won't
825 Base = N.getOperand(0);
826 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
830 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
831 // Loading from a constant address.
833 // If this address fits entirely in a 16-bit sext immediate field, codegen
836 if (isIntS16Immediate(CN, Imm)) {
837 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
838 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
842 // Handle 32-bit sext immediates with LIS + addr mode.
843 if (CN->getValueType(0) == MVT::i32 ||
844 (int64_t)CN->getValue() == (int)CN->getValue()) {
845 int Addr = (int)CN->getValue();
847 // Otherwise, break this down into an LIS + disp.
848 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
850 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
851 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
852 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
857 Disp = DAG.getTargetConstant(0, getPointerTy());
858 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
859 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
862 return true; // [r+0]
865 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
866 /// represented as an indexed [r+r] operation.
867 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
870 // Check to see if we can easily represent this as an [r+r] address. This
871 // will fail if it thinks that the address is more profitably represented as
872 // reg+imm, e.g. where imm = 0.
873 if (SelectAddressRegReg(N, Base, Index, DAG))
876 // If the operand is an addition, always emit this as [r+r], since this is
877 // better (for code size, and execution, as the memop does the add for free)
878 // than emitting an explicit add.
879 if (N.getOpcode() == ISD::ADD) {
880 Base = N.getOperand(0);
881 Index = N.getOperand(1);
885 // Otherwise, do it the hard way, using R0 as the base register.
886 Base = DAG.getRegister(PPC::R0, N.getValueType());
891 /// SelectAddressRegImmShift - Returns true if the address N can be
892 /// represented by a base register plus a signed 14-bit displacement
893 /// [r+imm*4]. Suitable for use by STD and friends.
894 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
897 // If this can be more profitably realized as r+r, fail.
898 if (SelectAddressRegReg(N, Disp, Base, DAG))
901 if (N.getOpcode() == ISD::ADD) {
903 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
904 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
905 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
906 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
908 Base = N.getOperand(0);
910 return true; // [r+i]
911 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
912 // Match LOAD (ADD (X, Lo(G))).
913 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
914 && "Cannot handle constant offsets yet!");
915 Disp = N.getOperand(1).getOperand(0); // The global address.
916 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
917 Disp.getOpcode() == ISD::TargetConstantPool ||
918 Disp.getOpcode() == ISD::TargetJumpTable);
919 Base = N.getOperand(0);
920 return true; // [&g+r]
922 } else if (N.getOpcode() == ISD::OR) {
924 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
925 // If this is an or of disjoint bitfields, we can codegen this as an add
926 // (for better address arithmetic) if the LHS and RHS of the OR are
927 // provably disjoint.
928 APInt LHSKnownZero, LHSKnownOne;
929 DAG.ComputeMaskedBits(N.getOperand(0),
930 APInt::getAllOnesValue(32),
931 LHSKnownZero, LHSKnownOne);
932 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
933 // If all of the bits are known zero on the LHS or RHS, the add won't
935 Base = N.getOperand(0);
936 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
940 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
941 // Loading from a constant address. Verify low two bits are clear.
942 if ((CN->getValue() & 3) == 0) {
943 // If this address fits entirely in a 14-bit sext immediate field, codegen
946 if (isIntS16Immediate(CN, Imm)) {
947 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
948 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
952 // Fold the low-part of 32-bit absolute addresses into addr mode.
953 if (CN->getValueType(0) == MVT::i32 ||
954 (int64_t)CN->getValue() == (int)CN->getValue()) {
955 int Addr = (int)CN->getValue();
957 // Otherwise, break this down into an LIS + disp.
958 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
960 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
961 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
962 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
968 Disp = DAG.getTargetConstant(0, getPointerTy());
969 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
970 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
973 return true; // [r+0]
977 /// getPreIndexedAddressParts - returns true by value, base pointer and
978 /// offset pointer and addressing mode by reference if the node's address
979 /// can be legally represented as pre-indexed load / store address.
980 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
982 ISD::MemIndexedMode &AM,
984 // Disabled by default for now.
985 if (!EnablePPCPreinc) return false;
989 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
990 Ptr = LD->getBasePtr();
991 VT = LD->getMemoryVT();
993 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
995 Ptr = ST->getBasePtr();
996 VT = ST->getMemoryVT();
1000 // PowerPC doesn't have preinc load/store instructions for vectors.
1001 if (MVT::isVector(VT))
1004 // TODO: Check reg+reg first.
1006 // LDU/STU use reg+imm*4, others use reg+imm.
1007 if (VT != MVT::i64) {
1009 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1013 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1017 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1018 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1019 // sext i32 to i64 when addr mode is r+i.
1020 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1021 LD->getExtensionType() == ISD::SEXTLOAD &&
1022 isa<ConstantSDNode>(Offset))
1030 //===----------------------------------------------------------------------===//
1031 // LowerOperation implementation
1032 //===----------------------------------------------------------------------===//
1034 SDOperand PPCTargetLowering::LowerConstantPool(SDOperand Op,
1035 SelectionDAG &DAG) {
1036 MVT::ValueType PtrVT = Op.getValueType();
1037 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1038 Constant *C = CP->getConstVal();
1039 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1040 SDOperand Zero = DAG.getConstant(0, PtrVT);
1042 const TargetMachine &TM = DAG.getTarget();
1044 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1045 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1047 // If this is a non-darwin platform, we don't support non-static relo models
1049 if (TM.getRelocationModel() == Reloc::Static ||
1050 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1051 // Generate non-pic code that has direct accesses to the constant pool.
1052 // The address of the global is just (hi(&g)+lo(&g)).
1053 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1056 if (TM.getRelocationModel() == Reloc::PIC_) {
1057 // With PIC, the first instruction is actually "GR+hi(&G)".
1058 Hi = DAG.getNode(ISD::ADD, PtrVT,
1059 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1062 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1066 SDOperand PPCTargetLowering::LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1067 MVT::ValueType PtrVT = Op.getValueType();
1068 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1069 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1070 SDOperand Zero = DAG.getConstant(0, PtrVT);
1072 const TargetMachine &TM = DAG.getTarget();
1074 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1075 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1077 // If this is a non-darwin platform, we don't support non-static relo models
1079 if (TM.getRelocationModel() == Reloc::Static ||
1080 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1081 // Generate non-pic code that has direct accesses to the constant pool.
1082 // The address of the global is just (hi(&g)+lo(&g)).
1083 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1086 if (TM.getRelocationModel() == Reloc::PIC_) {
1087 // With PIC, the first instruction is actually "GR+hi(&G)".
1088 Hi = DAG.getNode(ISD::ADD, PtrVT,
1089 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1092 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1096 SDOperand PPCTargetLowering::LowerGlobalTLSAddress(SDOperand Op,
1097 SelectionDAG &DAG) {
1098 assert(0 && "TLS not implemented for PPC.");
1101 SDOperand PPCTargetLowering::LowerGlobalAddress(SDOperand Op,
1102 SelectionDAG &DAG) {
1103 MVT::ValueType PtrVT = Op.getValueType();
1104 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1105 GlobalValue *GV = GSDN->getGlobal();
1106 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1107 // If it's a debug information descriptor, don't mess with it.
1108 if (DAG.isVerifiedDebugInfoDesc(Op))
1110 SDOperand Zero = DAG.getConstant(0, PtrVT);
1112 const TargetMachine &TM = DAG.getTarget();
1114 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1115 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1117 // If this is a non-darwin platform, we don't support non-static relo models
1119 if (TM.getRelocationModel() == Reloc::Static ||
1120 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1121 // Generate non-pic code that has direct accesses to globals.
1122 // The address of the global is just (hi(&g)+lo(&g)).
1123 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1126 if (TM.getRelocationModel() == Reloc::PIC_) {
1127 // With PIC, the first instruction is actually "GR+hi(&G)".
1128 Hi = DAG.getNode(ISD::ADD, PtrVT,
1129 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1132 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1134 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1137 // If the global is weak or external, we have to go through the lazy
1139 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1142 SDOperand PPCTargetLowering::LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1143 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1145 // If we're comparing for equality to zero, expose the fact that this is
1146 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1147 // fold the new nodes.
1148 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1149 if (C->isNullValue() && CC == ISD::SETEQ) {
1150 MVT::ValueType VT = Op.getOperand(0).getValueType();
1151 SDOperand Zext = Op.getOperand(0);
1152 if (VT < MVT::i32) {
1154 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1156 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1157 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1158 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1159 DAG.getConstant(Log2b, MVT::i32));
1160 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1162 // Leave comparisons against 0 and -1 alone for now, since they're usually
1163 // optimized. FIXME: revisit this when we can custom lower all setcc
1165 if (C->isAllOnesValue() || C->isNullValue())
1169 // If we have an integer seteq/setne, turn it into a compare against zero
1170 // by xor'ing the rhs with the lhs, which is faster than setting a
1171 // condition register, reading it back out, and masking the correct bit. The
1172 // normal approach here uses sub to do this instead of xor. Using xor exposes
1173 // the result to other bit-twiddling opportunities.
1174 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1175 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1176 MVT::ValueType VT = Op.getValueType();
1177 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1179 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1184 SDOperand PPCTargetLowering::LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1185 int VarArgsFrameIndex,
1186 int VarArgsStackOffset,
1187 unsigned VarArgsNumGPR,
1188 unsigned VarArgsNumFPR,
1189 const PPCSubtarget &Subtarget) {
1191 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1194 SDOperand PPCTargetLowering::LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1195 int VarArgsFrameIndex,
1196 int VarArgsStackOffset,
1197 unsigned VarArgsNumGPR,
1198 unsigned VarArgsNumFPR,
1199 const PPCSubtarget &Subtarget) {
1201 if (Subtarget.isMachoABI()) {
1202 // vastart just stores the address of the VarArgsFrameIndex slot into the
1203 // memory location argument.
1204 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1205 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1206 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1207 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1210 // For ELF 32 ABI we follow the layout of the va_list struct.
1211 // We suppose the given va_list is already allocated.
1214 // char gpr; /* index into the array of 8 GPRs
1215 // * stored in the register save area
1216 // * gpr=0 corresponds to r3,
1217 // * gpr=1 to r4, etc.
1219 // char fpr; /* index into the array of 8 FPRs
1220 // * stored in the register save area
1221 // * fpr=0 corresponds to f1,
1222 // * fpr=1 to f2, etc.
1224 // char *overflow_arg_area;
1225 // /* location on stack that holds
1226 // * the next overflow argument
1228 // char *reg_save_area;
1229 // /* where r3:r10 and f1:f8 (if saved)
1235 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1236 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1239 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1241 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1242 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1244 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1245 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1247 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1248 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1250 uint64_t FPROffset = 1;
1251 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1253 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1255 // Store first byte : number of int regs
1256 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1257 Op.getOperand(1), SV, 0);
1258 uint64_t nextOffset = FPROffset;
1259 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1262 // Store second byte : number of float regs
1263 SDOperand secondStore =
1264 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1265 nextOffset += StackOffset;
1266 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1268 // Store second word : arguments given on stack
1269 SDOperand thirdStore =
1270 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1271 nextOffset += FrameOffset;
1272 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1274 // Store third word : arguments given in registers
1275 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1279 #include "PPCGenCallingConv.inc"
1281 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1282 /// depending on which subtarget is selected.
1283 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1284 if (Subtarget.isMachoABI()) {
1285 static const unsigned FPR[] = {
1286 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1287 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1293 static const unsigned FPR[] = {
1294 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1301 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDOperand Op,
1303 int &VarArgsFrameIndex,
1304 int &VarArgsStackOffset,
1305 unsigned &VarArgsNumGPR,
1306 unsigned &VarArgsNumFPR,
1307 const PPCSubtarget &Subtarget) {
1308 // TODO: add description of PPC stack frame format, or at least some docs.
1310 MachineFunction &MF = DAG.getMachineFunction();
1311 MachineFrameInfo *MFI = MF.getFrameInfo();
1312 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1313 SmallVector<SDOperand, 8> ArgValues;
1314 SDOperand Root = Op.getOperand(0);
1316 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1317 bool isPPC64 = PtrVT == MVT::i64;
1318 bool isMachoABI = Subtarget.isMachoABI();
1319 bool isELF32_ABI = Subtarget.isELF32_ABI();
1320 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1322 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1324 static const unsigned GPR_32[] = { // 32-bit registers.
1325 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1326 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1328 static const unsigned GPR_64[] = { // 64-bit registers.
1329 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1330 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1333 static const unsigned *FPR = GetFPR(Subtarget);
1335 static const unsigned VR[] = {
1336 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1337 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1340 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1341 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1342 const unsigned Num_VR_Regs = array_lengthof( VR);
1344 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1346 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1348 // Add DAG nodes to load the arguments or copy them out of registers. On
1349 // entry to a function on PPC, the arguments start after the linkage area,
1350 // although the first ones are often in registers.
1352 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1353 // represented with two words (long long or double) must be copied to an
1354 // even GPR_idx value or to an even ArgOffset value.
1356 SmallVector<SDOperand, 8> MemOps;
1358 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1360 bool needsLoad = false;
1361 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1362 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1363 unsigned ArgSize = ObjSize;
1364 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1365 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1366 unsigned isByVal = Flags & ISD::ParamFlags::ByVal;
1367 // See if next argument requires stack alignment in ELF
1368 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1369 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1370 (!(Flags & AlignFlag)));
1372 unsigned CurArgOffset = ArgOffset;
1374 // FIXME alignment for ELF may not be right
1375 // FIXME the codegen can be much improved in some cases.
1376 // We do not have to keep everything in memory.
1378 // Double word align in ELF
1379 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1380 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1381 ObjSize = (Flags & ISD::ParamFlags::ByValSize) >>
1382 ISD::ParamFlags::ByValSizeOffs;
1383 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1384 // The value of the object is its address.
1385 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1386 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1387 ArgValues.push_back(FIN);
1388 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1389 // Store whatever pieces of the object are in registers
1390 // to memory. ArgVal will be address of the beginning of
1392 if (GPR_idx != Num_GPR_Regs) {
1393 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1394 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1395 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1396 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1397 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1398 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1399 MemOps.push_back(Store);
1401 if (isMachoABI) ArgOffset += PtrByteSize;
1403 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1411 default: assert(0 && "Unhandled argument type!");
1414 // Double word align in ELF
1415 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1417 if (GPR_idx != Num_GPR_Regs) {
1418 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1419 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1420 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1424 ArgSize = PtrByteSize;
1426 // Stack align in ELF
1427 if (needsLoad && Expand && isELF32_ABI)
1428 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1429 // All int arguments reserve stack space in Macho ABI.
1430 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1434 case MVT::i64: // PPC64
1435 if (GPR_idx != Num_GPR_Regs) {
1436 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1437 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1438 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1440 if (ObjectVT == MVT::i32) {
1441 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1442 // value to MVT::i64 and then truncate to the correct register size.
1443 if (Flags & ISD::ParamFlags::SExt)
1444 ArgVal = DAG.getNode(ISD::AssertSext, MVT::i64, ArgVal,
1445 DAG.getValueType(ObjectVT));
1446 else if (Flags & ISD::ParamFlags::ZExt)
1447 ArgVal = DAG.getNode(ISD::AssertZext, MVT::i64, ArgVal,
1448 DAG.getValueType(ObjectVT));
1450 ArgVal = DAG.getNode(ISD::TRUNCATE, MVT::i32, ArgVal);
1457 // All int arguments reserve stack space in Macho ABI.
1458 if (isMachoABI || needsLoad) ArgOffset += 8;
1463 // Every 4 bytes of argument space consumes one of the GPRs available for
1464 // argument passing.
1465 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1467 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1470 if (FPR_idx != Num_FPR_Regs) {
1472 if (ObjectVT == MVT::f32)
1473 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1475 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1476 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1477 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1483 // Stack align in ELF
1484 if (needsLoad && Expand && isELF32_ABI)
1485 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1486 // All FP arguments reserve stack space in Macho ABI.
1487 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1493 // Note that vector arguments in registers don't reserve stack space.
1494 if (VR_idx != Num_VR_Regs) {
1495 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1496 RegInfo.addLiveIn(VR[VR_idx], VReg);
1497 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1500 // This should be simple, but requires getting 16-byte aligned stack
1502 assert(0 && "Loading VR argument not implemented yet!");
1508 // We need to load the argument to a virtual register if we determined above
1509 // that we ran out of physical registers of the appropriate type.
1511 int FI = MFI->CreateFixedObject(ObjSize,
1512 CurArgOffset + (ArgSize - ObjSize));
1513 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1514 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1517 ArgValues.push_back(ArgVal);
1520 // If the function takes variable number of arguments, make a frame index for
1521 // the start of the first vararg value... for expansion of llvm.va_start.
1522 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1527 VarArgsNumGPR = GPR_idx;
1528 VarArgsNumFPR = FPR_idx;
1530 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1532 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1533 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1534 MVT::getSizeInBits(PtrVT)/8);
1536 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1543 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1545 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1547 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1548 // stored to the VarArgsFrameIndex on the stack.
1550 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1551 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1552 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1553 MemOps.push_back(Store);
1554 // Increment the address by four for the next argument to store
1555 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1556 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1560 // If this function is vararg, store any remaining integer argument regs
1561 // to their spots on the stack so that they may be loaded by deferencing the
1562 // result of va_next.
1563 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1566 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1568 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1570 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1571 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1572 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1573 MemOps.push_back(Store);
1574 // Increment the address by four for the next argument to store
1575 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1576 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1579 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1582 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1583 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1584 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1585 MemOps.push_back(Store);
1586 // Increment the address by eight for the next argument to store
1587 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1589 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1592 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1594 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1596 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1597 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1598 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1599 MemOps.push_back(Store);
1600 // Increment the address by eight for the next argument to store
1601 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1603 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1608 if (!MemOps.empty())
1609 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1611 ArgValues.push_back(Root);
1613 // Return the new list of results.
1614 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1615 Op.Val->value_end());
1616 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1619 /// isCallCompatibleAddress - Return the immediate to use if the specified
1620 /// 32-bit value is representable in the immediate field of a BxA instruction.
1621 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1622 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1625 int Addr = C->getValue();
1626 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1627 (Addr << 6 >> 6) != Addr)
1628 return 0; // Top 6 bits have to be sext of immediate.
1630 return DAG.getConstant((int)C->getValue() >> 2,
1631 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1634 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1635 /// by "Src" to address "Dst" of size "Size". Alignment information is
1636 /// specified by the specific parameter attribute. The copy will be passed as
1637 /// a byval function parameter.
1638 /// Sometimes what we are copying is the end of a larger object, the part that
1639 /// does not fit in registers.
1641 CreateCopyOfByValArgument(SDOperand Src, SDOperand Dst, SDOperand Chain,
1642 unsigned Flags, SelectionDAG &DAG, unsigned Size) {
1643 unsigned Align = 1 <<
1644 ((Flags & ISD::ParamFlags::ByValAlign) >> ISD::ParamFlags::ByValAlignOffs);
1645 SDOperand AlignNode = DAG.getConstant(Align, MVT::i32);
1646 SDOperand SizeNode = DAG.getConstant(Size, MVT::i32);
1647 SDOperand AlwaysInline = DAG.getConstant(0, MVT::i32);
1648 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, AlignNode, AlwaysInline);
1651 SDOperand PPCTargetLowering::LowerCALL(SDOperand Op, SelectionDAG &DAG,
1652 const PPCSubtarget &Subtarget) {
1653 SDOperand Chain = Op.getOperand(0);
1654 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1655 SDOperand Callee = Op.getOperand(4);
1656 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1658 bool isMachoABI = Subtarget.isMachoABI();
1659 bool isELF32_ABI = Subtarget.isELF32_ABI();
1661 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1662 bool isPPC64 = PtrVT == MVT::i64;
1663 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1665 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1666 // SelectExpr to use to put the arguments in the appropriate registers.
1667 std::vector<SDOperand> args_to_use;
1669 // Count how many bytes are to be pushed on the stack, including the linkage
1670 // area, and parameter passing area. We start with 24/48 bytes, which is
1671 // prereserved space for [SP][CR][LR][3 x unused].
1672 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1674 // Add up all the space actually used.
1675 for (unsigned i = 0; i != NumOps; ++i) {
1676 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1677 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1678 if (Flags & ISD::ParamFlags::ByVal)
1679 ArgSize = (Flags & ISD::ParamFlags::ByValSize) >>
1680 ISD::ParamFlags::ByValSizeOffs;
1681 ArgSize = std::max(ArgSize, PtrByteSize);
1682 NumBytes += ArgSize;
1685 // The prolog code of the callee may store up to 8 GPR argument registers to
1686 // the stack, allowing va_start to index over them in memory if its varargs.
1687 // Because we cannot tell if this is needed on the caller side, we have to
1688 // conservatively assume that it is needed. As such, make sure we have at
1689 // least enough stack space for the caller to store the 8 GPRs.
1690 NumBytes = std::max(NumBytes,
1691 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1693 // Adjust the stack pointer for the new arguments...
1694 // These operations are automatically eliminated by the prolog/epilog pass
1695 Chain = DAG.getCALLSEQ_START(Chain,
1696 DAG.getConstant(NumBytes, PtrVT));
1697 SDOperand CallSeqStart = Chain;
1699 // Set up a copy of the stack pointer for use loading and storing any
1700 // arguments that may not fit in the registers available for argument
1704 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1706 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1708 // Figure out which arguments are going to go in registers, and which in
1709 // memory. Also, if this is a vararg function, floating point operations
1710 // must be stored to our stack, and loaded into integer regs as well, if
1711 // any integer regs are available for argument passing.
1712 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1713 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1715 static const unsigned GPR_32[] = { // 32-bit registers.
1716 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1717 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1719 static const unsigned GPR_64[] = { // 64-bit registers.
1720 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1721 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1723 static const unsigned *FPR = GetFPR(Subtarget);
1725 static const unsigned VR[] = {
1726 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1727 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1729 const unsigned NumGPRs = array_lengthof(GPR_32);
1730 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1731 const unsigned NumVRs = array_lengthof( VR);
1733 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1735 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1736 SmallVector<SDOperand, 8> MemOpChains;
1737 for (unsigned i = 0; i != NumOps; ++i) {
1739 SDOperand Arg = Op.getOperand(5+2*i);
1740 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1741 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1742 // See if next argument requires stack alignment in ELF
1743 unsigned next = 5+2*(i+1)+1;
1744 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1745 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1746 (!(Flags & AlignFlag)));
1748 // PtrOff will be used to store the current argument to the stack if a
1749 // register cannot be found for it.
1752 // Stack align in ELF 32
1753 if (isELF32_ABI && Expand)
1754 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1755 StackPtr.getValueType());
1757 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1759 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1761 // On PPC64, promote integers to 64-bit values.
1762 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1763 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1764 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1767 // FIXME Elf untested, what are alignment rules?
1768 // FIXME memcpy is used way more than necessary. Correctness first.
1769 if (Flags & ISD::ParamFlags::ByVal) {
1770 unsigned Size = (Flags & ISD::ParamFlags::ByValSize) >>
1771 ISD::ParamFlags::ByValSizeOffs;
1772 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1773 if (Size==1 || Size==2) {
1774 // Very small objects are passed right-justified.
1775 // Everything else is passed left-justified.
1776 MVT::ValueType VT = (Size==1) ? MVT::i8 : MVT::i16;
1777 if (GPR_idx != NumGPRs) {
1778 SDOperand Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
1780 MemOpChains.push_back(Load.getValue(1));
1781 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1783 ArgOffset += PtrByteSize;
1785 SDOperand Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
1786 SDOperand AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
1787 SDOperand MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
1788 CallSeqStart.Val->getOperand(0),
1790 // This must go outside the CALLSEQ_START..END.
1791 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1792 CallSeqStart.Val->getOperand(1));
1793 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1794 Chain = CallSeqStart = NewCallSeqStart;
1795 ArgOffset += PtrByteSize;
1799 for (unsigned j=0; j<Size; j+=PtrByteSize) {
1800 SDOperand Const = DAG.getConstant(j, PtrOff.getValueType());
1801 SDOperand AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
1802 if (GPR_idx != NumGPRs) {
1803 SDOperand Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
1804 MemOpChains.push_back(Load.getValue(1));
1805 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
1807 ArgOffset += PtrByteSize;
1809 SDOperand AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
1810 SDOperand MemcpyCall = CreateCopyOfByValArgument(AddArg, AddPtr,
1811 CallSeqStart.Val->getOperand(0),
1812 Flags, DAG, Size - j);
1813 // This must go outside the CALLSEQ_START..END.
1814 SDOperand NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
1815 CallSeqStart.Val->getOperand(1));
1816 DAG.ReplaceAllUsesWith(CallSeqStart.Val, NewCallSeqStart.Val);
1817 Chain = CallSeqStart = NewCallSeqStart;
1818 ArgOffset += ((Size - j + 3)/4)*4;
1825 switch (Arg.getValueType()) {
1826 default: assert(0 && "Unexpected ValueType for argument!");
1829 // Double word align in ELF
1830 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1831 if (GPR_idx != NumGPRs) {
1832 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1834 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1837 if (inMem || isMachoABI) {
1838 // Stack align in ELF
1839 if (isELF32_ABI && Expand)
1840 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1842 ArgOffset += PtrByteSize;
1848 // Float varargs need to be promoted to double.
1849 if (Arg.getValueType() == MVT::f32)
1850 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1853 if (FPR_idx != NumFPRs) {
1854 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1857 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1858 MemOpChains.push_back(Store);
1860 // Float varargs are always shadowed in available integer registers
1861 if (GPR_idx != NumGPRs) {
1862 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1863 MemOpChains.push_back(Load.getValue(1));
1864 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1867 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1868 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1869 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1870 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1871 MemOpChains.push_back(Load.getValue(1));
1872 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1876 // If we have any FPRs remaining, we may also have GPRs remaining.
1877 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1880 if (GPR_idx != NumGPRs)
1882 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1883 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1888 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1891 if (inMem || isMachoABI) {
1892 // Stack align in ELF
1893 if (isELF32_ABI && Expand)
1894 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1898 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1905 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1906 assert(VR_idx != NumVRs &&
1907 "Don't support passing more than 12 vector args yet!");
1908 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1912 if (!MemOpChains.empty())
1913 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1914 &MemOpChains[0], MemOpChains.size());
1916 // Build a sequence of copy-to-reg nodes chained together with token chain
1917 // and flag operands which copy the outgoing args into the appropriate regs.
1919 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1920 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1922 InFlag = Chain.getValue(1);
1925 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1926 if (isVarArg && isELF32_ABI) {
1927 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1928 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1929 InFlag = Chain.getValue(1);
1932 std::vector<MVT::ValueType> NodeTys;
1933 NodeTys.push_back(MVT::Other); // Returns a chain
1934 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1936 SmallVector<SDOperand, 8> Ops;
1937 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1939 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1940 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1941 // node so that legalize doesn't hack it.
1942 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1943 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1944 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1945 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1946 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1947 // If this is an absolute destination address, use the munged value.
1948 Callee = SDOperand(Dest, 0);
1950 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1951 // to do the call, we can't use PPCISD::CALL.
1952 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1953 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1954 InFlag = Chain.getValue(1);
1956 // Copy the callee address into R12 on darwin.
1958 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1959 InFlag = Chain.getValue(1);
1963 NodeTys.push_back(MVT::Other);
1964 NodeTys.push_back(MVT::Flag);
1965 Ops.push_back(Chain);
1966 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1970 // If this is a direct call, pass the chain and the callee.
1972 Ops.push_back(Chain);
1973 Ops.push_back(Callee);
1976 // Add argument registers to the end of the list so that they are known live
1978 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1979 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1980 RegsToPass[i].second.getValueType()));
1983 Ops.push_back(InFlag);
1984 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1985 InFlag = Chain.getValue(1);
1987 Chain = DAG.getCALLSEQ_END(Chain,
1988 DAG.getConstant(NumBytes, PtrVT),
1989 DAG.getConstant(0, PtrVT),
1991 if (Op.Val->getValueType(0) != MVT::Other)
1992 InFlag = Chain.getValue(1);
1994 SDOperand ResultVals[3];
1995 unsigned NumResults = 0;
1998 // If the call has results, copy the values out of the ret val registers.
1999 switch (Op.Val->getValueType(0)) {
2000 default: assert(0 && "Unexpected ret value!");
2001 case MVT::Other: break;
2003 if (Op.Val->getValueType(1) == MVT::i32) {
2004 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
2005 ResultVals[0] = Chain.getValue(0);
2006 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
2007 Chain.getValue(2)).getValue(1);
2008 ResultVals[1] = Chain.getValue(0);
2010 NodeTys.push_back(MVT::i32);
2012 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
2013 ResultVals[0] = Chain.getValue(0);
2016 NodeTys.push_back(MVT::i32);
2019 if (Op.Val->getValueType(1) == MVT::i64) {
2020 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
2021 ResultVals[0] = Chain.getValue(0);
2022 Chain = DAG.getCopyFromReg(Chain, PPC::X4, MVT::i64,
2023 Chain.getValue(2)).getValue(1);
2024 ResultVals[1] = Chain.getValue(0);
2026 NodeTys.push_back(MVT::i64);
2028 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
2029 ResultVals[0] = Chain.getValue(0);
2032 NodeTys.push_back(MVT::i64);
2035 if (Op.Val->getValueType(1) == MVT::f64) {
2036 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
2037 ResultVals[0] = Chain.getValue(0);
2038 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
2039 Chain.getValue(2)).getValue(1);
2040 ResultVals[1] = Chain.getValue(0);
2042 NodeTys.push_back(MVT::f64);
2043 NodeTys.push_back(MVT::f64);
2046 // else fall through
2048 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
2049 InFlag).getValue(1);
2050 ResultVals[0] = Chain.getValue(0);
2052 NodeTys.push_back(Op.Val->getValueType(0));
2058 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
2059 InFlag).getValue(1);
2060 ResultVals[0] = Chain.getValue(0);
2062 NodeTys.push_back(Op.Val->getValueType(0));
2066 NodeTys.push_back(MVT::Other);
2068 // If the function returns void, just return the chain.
2069 if (NumResults == 0)
2072 // Otherwise, merge everything together with a MERGE_VALUES node.
2073 ResultVals[NumResults++] = Chain;
2074 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
2075 ResultVals, NumResults);
2076 return Res.getValue(Op.ResNo);
2079 SDOperand PPCTargetLowering::LowerRET(SDOperand Op, SelectionDAG &DAG,
2080 TargetMachine &TM) {
2081 SmallVector<CCValAssign, 16> RVLocs;
2082 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2083 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2084 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2085 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
2087 // If this is the first return lowered for this function, add the regs to the
2088 // liveout set for the function.
2089 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2090 for (unsigned i = 0; i != RVLocs.size(); ++i)
2091 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2094 SDOperand Chain = Op.getOperand(0);
2097 // Copy the result values into the output registers.
2098 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2099 CCValAssign &VA = RVLocs[i];
2100 assert(VA.isRegLoc() && "Can only return in registers!");
2101 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2102 Flag = Chain.getValue(1);
2106 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2108 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2111 SDOperand PPCTargetLowering::LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
2112 const PPCSubtarget &Subtarget) {
2113 // When we pop the dynamic allocation we need to restore the SP link.
2115 // Get the corect type for pointers.
2116 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2118 // Construct the stack pointer operand.
2119 bool IsPPC64 = Subtarget.isPPC64();
2120 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2121 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
2123 // Get the operands for the STACKRESTORE.
2124 SDOperand Chain = Op.getOperand(0);
2125 SDOperand SaveSP = Op.getOperand(1);
2127 // Load the old link SP.
2128 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2130 // Restore the stack pointer.
2131 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2133 // Store the old link SP.
2134 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2137 SDOperand PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDOperand Op,
2139 const PPCSubtarget &Subtarget) {
2140 MachineFunction &MF = DAG.getMachineFunction();
2141 bool IsPPC64 = Subtarget.isPPC64();
2142 bool isMachoABI = Subtarget.isMachoABI();
2144 // Get current frame pointer save index. The users of this index will be
2145 // primarily DYNALLOC instructions.
2146 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2147 int FPSI = FI->getFramePointerSaveIndex();
2149 // If the frame pointer save index hasn't been defined yet.
2151 // Find out what the fix offset of the frame pointer save area.
2152 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2154 // Allocate the frame index for frame pointer save area.
2155 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2157 FI->setFramePointerSaveIndex(FPSI);
2161 SDOperand Chain = Op.getOperand(0);
2162 SDOperand Size = Op.getOperand(1);
2164 // Get the corect type for pointers.
2165 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2167 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
2168 DAG.getConstant(0, PtrVT), Size);
2169 // Construct a node for the frame pointer save index.
2170 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
2171 // Build a DYNALLOC node.
2172 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
2173 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2174 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2178 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2180 SDOperand PPCTargetLowering::LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2181 // Not FP? Not a fsel.
2182 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2183 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2186 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2188 // Cannot handle SETEQ/SETNE.
2189 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2191 MVT::ValueType ResVT = Op.getValueType();
2192 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2193 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2194 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2196 // If the RHS of the comparison is a 0.0, we don't need to do the
2197 // subtraction at all.
2198 if (isFloatingPointZero(RHS))
2200 default: break; // SETUO etc aren't handled by fsel.
2204 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2208 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2209 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2210 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2214 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2218 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2219 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2220 return DAG.getNode(PPCISD::FSEL, ResVT,
2221 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2226 default: break; // SETUO etc aren't handled by fsel.
2230 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2231 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2232 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2233 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2237 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2238 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2239 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2240 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2244 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2245 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2246 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2247 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2251 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2252 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2253 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2254 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2259 // FIXME: Split this code up when LegalizeDAGTypes lands.
2260 SDOperand PPCTargetLowering::LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2261 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2262 SDOperand Src = Op.getOperand(0);
2263 if (Src.getValueType() == MVT::f32)
2264 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2267 switch (Op.getValueType()) {
2268 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2270 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2273 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2277 // Convert the FP value to an int value through memory.
2278 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2280 // Emit a store to the stack slot.
2281 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2283 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2285 if (Op.getValueType() == MVT::i32)
2286 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2287 DAG.getConstant(4, FIPtr.getValueType()));
2288 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2291 SDOperand PPCTargetLowering::LowerFP_ROUND_INREG(SDOperand Op,
2292 SelectionDAG &DAG) {
2293 assert(Op.getValueType() == MVT::ppcf128);
2294 SDNode *Node = Op.Val;
2295 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2296 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2297 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2298 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2300 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2301 // of the long double, and puts FPSCR back the way it was. We do not
2302 // actually model FPSCR.
2303 std::vector<MVT::ValueType> NodeTys;
2304 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2306 NodeTys.push_back(MVT::f64); // Return register
2307 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2308 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2309 MFFSreg = Result.getValue(0);
2310 InFlag = Result.getValue(1);
2313 NodeTys.push_back(MVT::Flag); // Returns a flag
2314 Ops[0] = DAG.getConstant(31, MVT::i32);
2316 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2317 InFlag = Result.getValue(0);
2320 NodeTys.push_back(MVT::Flag); // Returns a flag
2321 Ops[0] = DAG.getConstant(30, MVT::i32);
2323 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2324 InFlag = Result.getValue(0);
2327 NodeTys.push_back(MVT::f64); // result of add
2328 NodeTys.push_back(MVT::Flag); // Returns a flag
2332 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2333 FPreg = Result.getValue(0);
2334 InFlag = Result.getValue(1);
2337 NodeTys.push_back(MVT::f64);
2338 Ops[0] = DAG.getConstant(1, MVT::i32);
2342 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2343 FPreg = Result.getValue(0);
2345 // We know the low half is about to be thrown away, so just use something
2347 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2350 SDOperand PPCTargetLowering::LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2351 if (Op.getOperand(0).getValueType() == MVT::i64) {
2352 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2353 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2354 if (Op.getValueType() == MVT::f32)
2355 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2359 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2360 "Unhandled SINT_TO_FP type in custom expander!");
2361 // Since we only generate this in 64-bit mode, we can take advantage of
2362 // 64-bit registers. In particular, sign extend the input value into the
2363 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2364 // then lfd it and fcfid it.
2365 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2366 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2367 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2368 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2370 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2373 // STD the extended value into the stack slot.
2374 MemOperand MO(PseudoSourceValue::getFixedStack(),
2375 MemOperand::MOStore, FrameIdx, 8, 8);
2376 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2377 DAG.getEntryNode(), Ext64, FIdx,
2378 DAG.getMemOperand(MO));
2379 // Load the value as a double.
2380 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2382 // FCFID it and return it.
2383 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2384 if (Op.getValueType() == MVT::f32)
2385 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2389 SDOperand PPCTargetLowering::LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2391 The rounding mode is in bits 30:31 of FPSR, and has the following
2398 FLT_ROUNDS, on the other hand, expects the following:
2405 To perform the conversion, we do:
2406 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2409 MachineFunction &MF = DAG.getMachineFunction();
2410 MVT::ValueType VT = Op.getValueType();
2411 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2412 std::vector<MVT::ValueType> NodeTys;
2413 SDOperand MFFSreg, InFlag;
2415 // Save FP Control Word to register
2416 NodeTys.push_back(MVT::f64); // return register
2417 NodeTys.push_back(MVT::Flag); // unused in this context
2418 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2420 // Save FP register to stack slot
2421 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2422 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2423 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2424 StackSlot, NULL, 0);
2426 // Load FP Control Word from low 32 bits of stack slot.
2427 SDOperand Four = DAG.getConstant(4, PtrVT);
2428 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2429 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2431 // Transform as necessary
2433 DAG.getNode(ISD::AND, MVT::i32,
2434 CWD, DAG.getConstant(3, MVT::i32));
2436 DAG.getNode(ISD::SRL, MVT::i32,
2437 DAG.getNode(ISD::AND, MVT::i32,
2438 DAG.getNode(ISD::XOR, MVT::i32,
2439 CWD, DAG.getConstant(3, MVT::i32)),
2440 DAG.getConstant(3, MVT::i32)),
2441 DAG.getConstant(1, MVT::i8));
2444 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2446 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2447 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2450 SDOperand PPCTargetLowering::LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2451 MVT::ValueType VT = Op.getValueType();
2452 unsigned BitWidth = MVT::getSizeInBits(VT);
2453 assert(Op.getNumOperands() == 3 &&
2454 VT == Op.getOperand(1).getValueType() &&
2457 // Expand into a bunch of logical ops. Note that these ops
2458 // depend on the PPC behavior for oversized shift amounts.
2459 SDOperand Lo = Op.getOperand(0);
2460 SDOperand Hi = Op.getOperand(1);
2461 SDOperand Amt = Op.getOperand(2);
2462 MVT::ValueType AmtVT = Amt.getValueType();
2464 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2465 DAG.getConstant(BitWidth, AmtVT), Amt);
2466 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
2467 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
2468 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2469 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2470 DAG.getConstant(-BitWidth, AmtVT));
2471 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
2472 SDOperand OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2473 SDOperand OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
2474 SDOperand OutOps[] = { OutLo, OutHi };
2475 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2479 SDOperand PPCTargetLowering::LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2480 MVT::ValueType VT = Op.getValueType();
2481 unsigned BitWidth = MVT::getSizeInBits(VT);
2482 assert(Op.getNumOperands() == 3 &&
2483 VT == Op.getOperand(1).getValueType() &&
2486 // Expand into a bunch of logical ops. Note that these ops
2487 // depend on the PPC behavior for oversized shift amounts.
2488 SDOperand Lo = Op.getOperand(0);
2489 SDOperand Hi = Op.getOperand(1);
2490 SDOperand Amt = Op.getOperand(2);
2491 MVT::ValueType AmtVT = Amt.getValueType();
2493 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2494 DAG.getConstant(BitWidth, AmtVT), Amt);
2495 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2496 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2497 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2498 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2499 DAG.getConstant(-BitWidth, AmtVT));
2500 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
2501 SDOperand OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
2502 SDOperand OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
2503 SDOperand OutOps[] = { OutLo, OutHi };
2504 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2508 SDOperand PPCTargetLowering::LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2509 MVT::ValueType VT = Op.getValueType();
2510 unsigned BitWidth = MVT::getSizeInBits(VT);
2511 assert(Op.getNumOperands() == 3 &&
2512 VT == Op.getOperand(1).getValueType() &&
2515 // Expand into a bunch of logical ops, followed by a select_cc.
2516 SDOperand Lo = Op.getOperand(0);
2517 SDOperand Hi = Op.getOperand(1);
2518 SDOperand Amt = Op.getOperand(2);
2519 MVT::ValueType AmtVT = Amt.getValueType();
2521 SDOperand Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
2522 DAG.getConstant(BitWidth, AmtVT), Amt);
2523 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
2524 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
2525 SDOperand Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
2526 SDOperand Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
2527 DAG.getConstant(-BitWidth, AmtVT));
2528 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
2529 SDOperand OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
2530 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
2531 Tmp4, Tmp6, ISD::SETLE);
2532 SDOperand OutOps[] = { OutLo, OutHi };
2533 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(VT, VT),
2537 //===----------------------------------------------------------------------===//
2538 // Vector related lowering.
2541 // If this is a vector of constants or undefs, get the bits. A bit in
2542 // UndefBits is set if the corresponding element of the vector is an
2543 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2544 // zero. Return true if this is not an array of constants, false if it is.
2546 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2547 uint64_t UndefBits[2]) {
2548 // Start with zero'd results.
2549 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2551 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2552 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2553 SDOperand OpVal = BV->getOperand(i);
2555 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2556 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2558 uint64_t EltBits = 0;
2559 if (OpVal.getOpcode() == ISD::UNDEF) {
2560 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2561 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2563 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2564 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2565 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2566 assert(CN->getValueType(0) == MVT::f32 &&
2567 "Only one legal FP vector type!");
2568 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2570 // Nonconstant element.
2574 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2577 //printf("%llx %llx %llx %llx\n",
2578 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2582 // If this is a splat (repetition) of a value across the whole vector, return
2583 // the smallest size that splats it. For example, "0x01010101010101..." is a
2584 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2585 // SplatSize = 1 byte.
2586 static bool isConstantSplat(const uint64_t Bits128[2],
2587 const uint64_t Undef128[2],
2588 unsigned &SplatBits, unsigned &SplatUndef,
2589 unsigned &SplatSize) {
2591 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2592 // the same as the lower 64-bits, ignoring undefs.
2593 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2594 return false; // Can't be a splat if two pieces don't match.
2596 uint64_t Bits64 = Bits128[0] | Bits128[1];
2597 uint64_t Undef64 = Undef128[0] & Undef128[1];
2599 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2601 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2602 return false; // Can't be a splat if two pieces don't match.
2604 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2605 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2607 // If the top 16-bits are different than the lower 16-bits, ignoring
2608 // undefs, we have an i32 splat.
2609 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2611 SplatUndef = Undef32;
2616 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2617 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2619 // If the top 8-bits are different than the lower 8-bits, ignoring
2620 // undefs, we have an i16 splat.
2621 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2623 SplatUndef = Undef16;
2628 // Otherwise, we have an 8-bit splat.
2629 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2630 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2635 /// BuildSplatI - Build a canonical splati of Val with an element size of
2636 /// SplatSize. Cast the result to VT.
2637 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2638 SelectionDAG &DAG) {
2639 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2641 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2642 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2645 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2647 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2651 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2653 // Build a canonical splat for this value.
2654 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2655 SmallVector<SDOperand, 8> Ops;
2656 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2657 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2658 &Ops[0], Ops.size());
2659 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2662 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2663 /// specified intrinsic ID.
2664 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2666 MVT::ValueType DestVT = MVT::Other) {
2667 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2668 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2669 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2672 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2673 /// specified intrinsic ID.
2674 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2675 SDOperand Op2, SelectionDAG &DAG,
2676 MVT::ValueType DestVT = MVT::Other) {
2677 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2678 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2679 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2683 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2684 /// amount. The result has the specified value type.
2685 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2686 MVT::ValueType VT, SelectionDAG &DAG) {
2687 // Force LHS/RHS to be the right type.
2688 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2689 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2692 for (unsigned i = 0; i != 16; ++i)
2693 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2694 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2695 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2696 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2699 // If this is a case we can't handle, return null and let the default
2700 // expansion code take care of it. If we CAN select this case, and if it
2701 // selects to a single instruction, return Op. Otherwise, if we can codegen
2702 // this case more efficiently than a constant pool load, lower it to the
2703 // sequence of ops that should be used.
2704 SDOperand PPCTargetLowering::LowerBUILD_VECTOR(SDOperand Op,
2705 SelectionDAG &DAG) {
2706 // If this is a vector of constants or undefs, get the bits. A bit in
2707 // UndefBits is set if the corresponding element of the vector is an
2708 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2710 uint64_t VectorBits[2];
2711 uint64_t UndefBits[2];
2712 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2713 return SDOperand(); // Not a constant vector.
2715 // If this is a splat (repetition) of a value across the whole vector, return
2716 // the smallest size that splats it. For example, "0x01010101010101..." is a
2717 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2718 // SplatSize = 1 byte.
2719 unsigned SplatBits, SplatUndef, SplatSize;
2720 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2721 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2723 // First, handle single instruction cases.
2726 if (SplatBits == 0) {
2727 // Canonicalize all zero vectors to be v4i32.
2728 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2729 SDOperand Z = DAG.getConstant(0, MVT::i32);
2730 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2731 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2736 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2737 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2738 if (SextVal >= -16 && SextVal <= 15)
2739 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2742 // Two instruction sequences.
2744 // If this value is in the range [-32,30] and is even, use:
2745 // tmp = VSPLTI[bhw], result = add tmp, tmp
2746 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2747 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2748 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2751 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2752 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2754 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2755 // Make -1 and vspltisw -1:
2756 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2758 // Make the VSLW intrinsic, computing 0x8000_0000.
2759 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2762 // xor by OnesV to invert it.
2763 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2764 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2767 // Check to see if this is a wide variety of vsplti*, binop self cases.
2768 unsigned SplatBitSize = SplatSize*8;
2769 static const signed char SplatCsts[] = {
2770 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2771 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2774 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2775 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2776 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2777 int i = SplatCsts[idx];
2779 // Figure out what shift amount will be used by altivec if shifted by i in
2781 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2783 // vsplti + shl self.
2784 if (SextVal == (i << (int)TypeShiftAmt)) {
2785 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2786 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2787 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2788 Intrinsic::ppc_altivec_vslw
2790 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2791 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2794 // vsplti + srl self.
2795 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2796 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2797 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2798 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2799 Intrinsic::ppc_altivec_vsrw
2801 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2802 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2805 // vsplti + sra self.
2806 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2807 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2808 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2809 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2810 Intrinsic::ppc_altivec_vsraw
2812 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2813 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2816 // vsplti + rol self.
2817 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2818 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2819 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2820 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2821 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2822 Intrinsic::ppc_altivec_vrlw
2824 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2825 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2828 // t = vsplti c, result = vsldoi t, t, 1
2829 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2830 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2831 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2833 // t = vsplti c, result = vsldoi t, t, 2
2834 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2835 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2836 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2838 // t = vsplti c, result = vsldoi t, t, 3
2839 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2840 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2841 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2845 // Three instruction sequences.
2847 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2848 if (SextVal >= 0 && SextVal <= 31) {
2849 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2850 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2851 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2852 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2854 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2855 if (SextVal >= -31 && SextVal <= 0) {
2856 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2857 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2858 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2859 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2866 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2867 /// the specified operations to build the shuffle.
2868 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2869 SDOperand RHS, SelectionDAG &DAG) {
2870 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2871 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2872 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2875 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2887 if (OpNum == OP_COPY) {
2888 if (LHSID == (1*9+2)*9+3) return LHS;
2889 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2893 SDOperand OpLHS, OpRHS;
2894 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2895 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2897 unsigned ShufIdxs[16];
2899 default: assert(0 && "Unknown i32 permute!");
2901 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2902 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2903 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2904 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2907 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2908 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2909 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2910 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2913 for (unsigned i = 0; i != 16; ++i)
2914 ShufIdxs[i] = (i&3)+0;
2917 for (unsigned i = 0; i != 16; ++i)
2918 ShufIdxs[i] = (i&3)+4;
2921 for (unsigned i = 0; i != 16; ++i)
2922 ShufIdxs[i] = (i&3)+8;
2925 for (unsigned i = 0; i != 16; ++i)
2926 ShufIdxs[i] = (i&3)+12;
2929 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2931 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2933 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2936 for (unsigned i = 0; i != 16; ++i)
2937 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2939 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2940 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2943 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2944 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2945 /// return the code it can be lowered into. Worst case, it can always be
2946 /// lowered into a vperm.
2947 SDOperand PPCTargetLowering::LowerVECTOR_SHUFFLE(SDOperand Op,
2948 SelectionDAG &DAG) {
2949 SDOperand V1 = Op.getOperand(0);
2950 SDOperand V2 = Op.getOperand(1);
2951 SDOperand PermMask = Op.getOperand(2);
2953 // Cases that are handled by instructions that take permute immediates
2954 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2955 // selected by the instruction selector.
2956 if (V2.getOpcode() == ISD::UNDEF) {
2957 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2958 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2959 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2960 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2961 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2962 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2963 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2964 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2965 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2966 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2967 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2968 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2973 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2974 // and produce a fixed permutation. If any of these match, do not lower to
2976 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2977 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2978 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2979 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2980 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2981 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2982 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2983 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2984 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2987 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2988 // perfect shuffle table to emit an optimal matching sequence.
2989 unsigned PFIndexes[4];
2990 bool isFourElementShuffle = true;
2991 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2992 unsigned EltNo = 8; // Start out undef.
2993 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2994 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2995 continue; // Undef, ignore it.
2997 unsigned ByteSource =
2998 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2999 if ((ByteSource & 3) != j) {
3000 isFourElementShuffle = false;
3005 EltNo = ByteSource/4;
3006 } else if (EltNo != ByteSource/4) {
3007 isFourElementShuffle = false;
3011 PFIndexes[i] = EltNo;
3014 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3015 // perfect shuffle vector to determine if it is cost effective to do this as
3016 // discrete instructions, or whether we should use a vperm.
3017 if (isFourElementShuffle) {
3018 // Compute the index in the perfect shuffle table.
3019 unsigned PFTableIndex =
3020 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3022 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3023 unsigned Cost = (PFEntry >> 30);
3025 // Determining when to avoid vperm is tricky. Many things affect the cost
3026 // of vperm, particularly how many times the perm mask needs to be computed.
3027 // For example, if the perm mask can be hoisted out of a loop or is already
3028 // used (perhaps because there are multiple permutes with the same shuffle
3029 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3030 // the loop requires an extra register.
3032 // As a compromise, we only emit discrete instructions if the shuffle can be
3033 // generated in 3 or fewer operations. When we have loop information
3034 // available, if this block is within a loop, we should avoid using vperm
3035 // for 3-operation perms and use a constant pool load instead.
3037 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3040 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3041 // vector that will get spilled to the constant pool.
3042 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3044 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3045 // that it is in input element units, not in bytes. Convert now.
3046 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
3047 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
3049 SmallVector<SDOperand, 16> ResultMask;
3050 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3052 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3055 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
3057 for (unsigned j = 0; j != BytesPerElement; ++j)
3058 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3062 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3063 &ResultMask[0], ResultMask.size());
3064 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3067 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3068 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3069 /// information about the intrinsic.
3070 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
3072 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
3075 switch (IntrinsicID) {
3076 default: return false;
3077 // Comparison predicates.
3078 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3079 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3080 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3081 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3082 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3083 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3084 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3085 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3086 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3087 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3088 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3089 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3090 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3092 // Normal Comparisons.
3093 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3094 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3095 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3096 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3097 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3098 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3099 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3100 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3101 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3102 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3103 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3104 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3105 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3110 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3111 /// lower, do it, otherwise return null.
3112 SDOperand PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDOperand Op,
3113 SelectionDAG &DAG) {
3114 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3115 // opcode number of the comparison.
3118 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3119 return SDOperand(); // Don't custom lower most intrinsics.
3121 // If this is a non-dot comparison, make the VCMP node and we are done.
3123 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3124 Op.getOperand(1), Op.getOperand(2),
3125 DAG.getConstant(CompareOpc, MVT::i32));
3126 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3129 // Create the PPCISD altivec 'dot' comparison node.
3131 Op.getOperand(2), // LHS
3132 Op.getOperand(3), // RHS
3133 DAG.getConstant(CompareOpc, MVT::i32)
3135 std::vector<MVT::ValueType> VTs;
3136 VTs.push_back(Op.getOperand(2).getValueType());
3137 VTs.push_back(MVT::Flag);
3138 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3140 // Now that we have the comparison, emit a copy from the CR to a GPR.
3141 // This is flagged to the above dot comparison.
3142 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3143 DAG.getRegister(PPC::CR6, MVT::i32),
3144 CompNode.getValue(1));
3146 // Unpack the result based on how the target uses it.
3147 unsigned BitNo; // Bit # of CR6.
3148 bool InvertBit; // Invert result?
3149 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
3150 default: // Can't happen, don't crash on invalid number though.
3151 case 0: // Return the value of the EQ bit of CR6.
3152 BitNo = 0; InvertBit = false;
3154 case 1: // Return the inverted value of the EQ bit of CR6.
3155 BitNo = 0; InvertBit = true;
3157 case 2: // Return the value of the LT bit of CR6.
3158 BitNo = 2; InvertBit = false;
3160 case 3: // Return the inverted value of the LT bit of CR6.
3161 BitNo = 2; InvertBit = true;
3165 // Shift the bit into the low position.
3166 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3167 DAG.getConstant(8-(3-BitNo), MVT::i32));
3169 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3170 DAG.getConstant(1, MVT::i32));
3172 // If we are supposed to, toggle the bit.
3174 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3175 DAG.getConstant(1, MVT::i32));
3179 SDOperand PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDOperand Op,
3180 SelectionDAG &DAG) {
3181 // Create a stack slot that is 16-byte aligned.
3182 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3183 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3184 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3185 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3187 // Store the input value into Value#0 of the stack slot.
3188 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
3189 Op.getOperand(0), FIdx, NULL, 0);
3191 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3194 SDOperand PPCTargetLowering::LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3195 if (Op.getValueType() == MVT::v4i32) {
3196 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3198 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3199 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3201 SDOperand RHSSwap = // = vrlw RHS, 16
3202 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3204 // Shrinkify inputs to v8i16.
3205 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3206 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3207 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3209 // Low parts multiplied together, generating 32-bit results (we ignore the
3211 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3212 LHS, RHS, DAG, MVT::v4i32);
3214 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3215 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3216 // Shift the high parts up 16 bits.
3217 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3218 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3219 } else if (Op.getValueType() == MVT::v8i16) {
3220 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3222 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3224 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3225 LHS, RHS, Zero, DAG);
3226 } else if (Op.getValueType() == MVT::v16i8) {
3227 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3229 // Multiply the even 8-bit parts, producing 16-bit sums.
3230 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3231 LHS, RHS, DAG, MVT::v8i16);
3232 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3234 // Multiply the odd 8-bit parts, producing 16-bit sums.
3235 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3236 LHS, RHS, DAG, MVT::v8i16);
3237 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3239 // Merge the results together.
3241 for (unsigned i = 0; i != 8; ++i) {
3242 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3243 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3245 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3246 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3248 assert(0 && "Unknown mul to lower!");
3253 /// LowerOperation - Provide custom lowering hooks for some operations.
3255 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3256 switch (Op.getOpcode()) {
3257 default: assert(0 && "Wasn't expecting to be able to lower this!");
3258 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3259 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3260 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3261 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3262 case ISD::SETCC: return LowerSETCC(Op, DAG);
3264 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3265 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3268 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3269 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3271 case ISD::FORMAL_ARGUMENTS:
3272 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3273 VarArgsStackOffset, VarArgsNumGPR,
3274 VarArgsNumFPR, PPCSubTarget);
3276 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
3277 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3278 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3279 case ISD::DYNAMIC_STACKALLOC:
3280 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3282 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3283 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3284 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3285 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3286 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3288 // Lower 64-bit shifts.
3289 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3290 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3291 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3293 // Vector-related lowering.
3294 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3295 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3296 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3297 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3298 case ISD::MUL: return LowerMUL(Op, DAG);
3300 // Frame & Return address.
3301 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3302 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3307 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3308 switch (N->getOpcode()) {
3309 default: assert(0 && "Wasn't expecting to be able to lower this!");
3310 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3315 //===----------------------------------------------------------------------===//
3316 // Other Lowering Code
3317 //===----------------------------------------------------------------------===//
3320 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3321 MachineBasicBlock *BB) {
3322 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3323 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3324 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3325 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3326 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3327 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3328 "Unexpected instr type to insert");
3330 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3331 // control-flow pattern. The incoming instruction knows the destination vreg
3332 // to set, the condition code register to branch on, the true/false values to
3333 // select between, and a branch opcode to use.
3334 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3335 ilist<MachineBasicBlock>::iterator It = BB;
3341 // cmpTY ccX, r1, r2
3343 // fallthrough --> copy0MBB
3344 MachineBasicBlock *thisMBB = BB;
3345 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3346 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3347 unsigned SelectPred = MI->getOperand(4).getImm();
3348 BuildMI(BB, TII->get(PPC::BCC))
3349 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3350 MachineFunction *F = BB->getParent();
3351 F->getBasicBlockList().insert(It, copy0MBB);
3352 F->getBasicBlockList().insert(It, sinkMBB);
3353 // Update machine-CFG edges by first adding all successors of the current
3354 // block to the new block which will contain the Phi node for the select.
3355 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3356 e = BB->succ_end(); i != e; ++i)
3357 sinkMBB->addSuccessor(*i);
3358 // Next, remove all successors of the current block, and add the true
3359 // and fallthrough blocks as its successors.
3360 while(!BB->succ_empty())
3361 BB->removeSuccessor(BB->succ_begin());
3362 BB->addSuccessor(copy0MBB);
3363 BB->addSuccessor(sinkMBB);
3366 // %FalseValue = ...
3367 // # fallthrough to sinkMBB
3370 // Update machine-CFG edges
3371 BB->addSuccessor(sinkMBB);
3374 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3377 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3378 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3379 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3381 delete MI; // The pseudo instruction is gone now.
3385 //===----------------------------------------------------------------------===//
3386 // Target Optimization Hooks
3387 //===----------------------------------------------------------------------===//
3389 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3390 DAGCombinerInfo &DCI) const {
3391 TargetMachine &TM = getTargetMachine();
3392 SelectionDAG &DAG = DCI.DAG;
3393 switch (N->getOpcode()) {
3396 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3397 if (C->getValue() == 0) // 0 << V -> 0.
3398 return N->getOperand(0);
3402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3403 if (C->getValue() == 0) // 0 >>u V -> 0.
3404 return N->getOperand(0);
3408 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3409 if (C->getValue() == 0 || // 0 >>s V -> 0.
3410 C->isAllOnesValue()) // -1 >>s V -> -1.
3411 return N->getOperand(0);
3415 case ISD::SINT_TO_FP:
3416 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3417 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3418 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3419 // We allow the src/dst to be either f32/f64, but the intermediate
3420 // type must be i64.
3421 if (N->getOperand(0).getValueType() == MVT::i64 &&
3422 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3423 SDOperand Val = N->getOperand(0).getOperand(0);
3424 if (Val.getValueType() == MVT::f32) {
3425 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3426 DCI.AddToWorklist(Val.Val);
3429 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3430 DCI.AddToWorklist(Val.Val);
3431 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3432 DCI.AddToWorklist(Val.Val);
3433 if (N->getValueType(0) == MVT::f32) {
3434 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3435 DAG.getIntPtrConstant(0));
3436 DCI.AddToWorklist(Val.Val);
3439 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3440 // If the intermediate type is i32, we can avoid the load/store here
3447 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3448 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3449 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3450 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3451 N->getOperand(1).getValueType() == MVT::i32 &&
3452 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3453 SDOperand Val = N->getOperand(1).getOperand(0);
3454 if (Val.getValueType() == MVT::f32) {
3455 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3456 DCI.AddToWorklist(Val.Val);
3458 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3459 DCI.AddToWorklist(Val.Val);
3461 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3462 N->getOperand(2), N->getOperand(3));
3463 DCI.AddToWorklist(Val.Val);
3467 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3468 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3469 N->getOperand(1).Val->hasOneUse() &&
3470 (N->getOperand(1).getValueType() == MVT::i32 ||
3471 N->getOperand(1).getValueType() == MVT::i16)) {
3472 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3473 // Do an any-extend to 32-bits if this is a half-word input.
3474 if (BSwapOp.getValueType() == MVT::i16)
3475 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3477 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3478 N->getOperand(2), N->getOperand(3),
3479 DAG.getValueType(N->getOperand(1).getValueType()));
3483 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3484 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3485 N->getOperand(0).hasOneUse() &&
3486 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3487 SDOperand Load = N->getOperand(0);
3488 LoadSDNode *LD = cast<LoadSDNode>(Load);
3489 // Create the byte-swapping load.
3490 std::vector<MVT::ValueType> VTs;
3491 VTs.push_back(MVT::i32);
3492 VTs.push_back(MVT::Other);
3493 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3495 LD->getChain(), // Chain
3496 LD->getBasePtr(), // Ptr
3498 DAG.getValueType(N->getValueType(0)) // VT
3500 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3502 // If this is an i16 load, insert the truncate.
3503 SDOperand ResVal = BSLoad;
3504 if (N->getValueType(0) == MVT::i16)
3505 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3507 // First, combine the bswap away. This makes the value produced by the
3509 DCI.CombineTo(N, ResVal);
3511 // Next, combine the load away, we give it a bogus result value but a real
3512 // chain result. The result value is dead because the bswap is dead.
3513 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3515 // Return N so it doesn't get rechecked!
3516 return SDOperand(N, 0);
3520 case PPCISD::VCMP: {
3521 // If a VCMPo node already exists with exactly the same operands as this
3522 // node, use its result instead of this node (VCMPo computes both a CR6 and
3523 // a normal output).
3525 if (!N->getOperand(0).hasOneUse() &&
3526 !N->getOperand(1).hasOneUse() &&
3527 !N->getOperand(2).hasOneUse()) {
3529 // Scan all of the users of the LHS, looking for VCMPo's that match.
3530 SDNode *VCMPoNode = 0;
3532 SDNode *LHSN = N->getOperand(0).Val;
3533 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3535 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3536 (*UI)->getOperand(1) == N->getOperand(1) &&
3537 (*UI)->getOperand(2) == N->getOperand(2) &&
3538 (*UI)->getOperand(0) == N->getOperand(0)) {
3543 // If there is no VCMPo node, or if the flag value has a single use, don't
3545 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3548 // Look at the (necessarily single) use of the flag value. If it has a
3549 // chain, this transformation is more complex. Note that multiple things
3550 // could use the value result, which we should ignore.
3551 SDNode *FlagUser = 0;
3552 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3553 FlagUser == 0; ++UI) {
3554 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3556 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3557 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3564 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3565 // give up for right now.
3566 if (FlagUser->getOpcode() == PPCISD::MFCR)
3567 return SDOperand(VCMPoNode, 0);
3572 // If this is a branch on an altivec predicate comparison, lower this so
3573 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3574 // lowering is done pre-legalize, because the legalizer lowers the predicate
3575 // compare down to code that is difficult to reassemble.
3576 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3577 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3581 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3582 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3583 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3584 assert(isDot && "Can't compare against a vector result!");
3586 // If this is a comparison against something other than 0/1, then we know
3587 // that the condition is never/always true.
3588 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3589 if (Val != 0 && Val != 1) {
3590 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3591 return N->getOperand(0);
3592 // Always !=, turn it into an unconditional branch.
3593 return DAG.getNode(ISD::BR, MVT::Other,
3594 N->getOperand(0), N->getOperand(4));
3597 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3599 // Create the PPCISD altivec 'dot' comparison node.
3600 std::vector<MVT::ValueType> VTs;
3602 LHS.getOperand(2), // LHS of compare
3603 LHS.getOperand(3), // RHS of compare
3604 DAG.getConstant(CompareOpc, MVT::i32)
3606 VTs.push_back(LHS.getOperand(2).getValueType());
3607 VTs.push_back(MVT::Flag);
3608 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3610 // Unpack the result based on how the target uses it.
3611 PPC::Predicate CompOpc;
3612 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3613 default: // Can't happen, don't crash on invalid number though.
3614 case 0: // Branch on the value of the EQ bit of CR6.
3615 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3617 case 1: // Branch on the inverted value of the EQ bit of CR6.
3618 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3620 case 2: // Branch on the value of the LT bit of CR6.
3621 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3623 case 3: // Branch on the inverted value of the LT bit of CR6.
3624 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3628 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3629 DAG.getConstant(CompOpc, MVT::i32),
3630 DAG.getRegister(PPC::CR6, MVT::i32),
3631 N->getOperand(4), CompNode.getValue(1));
3640 //===----------------------------------------------------------------------===//
3641 // Inline Assembly Support
3642 //===----------------------------------------------------------------------===//
3644 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3648 const SelectionDAG &DAG,
3649 unsigned Depth) const {
3650 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3651 switch (Op.getOpcode()) {
3653 case PPCISD::LBRX: {
3654 // lhbrx is known to have the top bits cleared out.
3655 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3656 KnownZero = 0xFFFF0000;
3659 case ISD::INTRINSIC_WO_CHAIN: {
3660 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3662 case Intrinsic::ppc_altivec_vcmpbfp_p:
3663 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3664 case Intrinsic::ppc_altivec_vcmpequb_p:
3665 case Intrinsic::ppc_altivec_vcmpequh_p:
3666 case Intrinsic::ppc_altivec_vcmpequw_p:
3667 case Intrinsic::ppc_altivec_vcmpgefp_p:
3668 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3669 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3670 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3671 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3672 case Intrinsic::ppc_altivec_vcmpgtub_p:
3673 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3674 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3675 KnownZero = ~1U; // All bits but the low one are known to be zero.
3683 /// getConstraintType - Given a constraint, return the type of
3684 /// constraint it is for this target.
3685 PPCTargetLowering::ConstraintType
3686 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3687 if (Constraint.size() == 1) {
3688 switch (Constraint[0]) {
3695 return C_RegisterClass;
3698 return TargetLowering::getConstraintType(Constraint);
3701 std::pair<unsigned, const TargetRegisterClass*>
3702 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3703 MVT::ValueType VT) const {
3704 if (Constraint.size() == 1) {
3705 // GCC RS6000 Constraint Letters
3706 switch (Constraint[0]) {
3709 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3710 return std::make_pair(0U, PPC::G8RCRegisterClass);
3711 return std::make_pair(0U, PPC::GPRCRegisterClass);
3714 return std::make_pair(0U, PPC::F4RCRegisterClass);
3715 else if (VT == MVT::f64)
3716 return std::make_pair(0U, PPC::F8RCRegisterClass);
3719 return std::make_pair(0U, PPC::VRRCRegisterClass);
3721 return std::make_pair(0U, PPC::CRRCRegisterClass);
3725 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3729 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3730 /// vector. If it is invalid, don't add anything to Ops.
3731 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3732 std::vector<SDOperand>&Ops,
3733 SelectionDAG &DAG) {
3734 SDOperand Result(0,0);
3745 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3746 if (!CST) return; // Must be an immediate to match.
3747 unsigned Value = CST->getValue();
3749 default: assert(0 && "Unknown constraint letter!");
3750 case 'I': // "I" is a signed 16-bit constant.
3751 if ((short)Value == (int)Value)
3752 Result = DAG.getTargetConstant(Value, Op.getValueType());
3754 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3755 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3756 if ((short)Value == 0)
3757 Result = DAG.getTargetConstant(Value, Op.getValueType());
3759 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3760 if ((Value >> 16) == 0)
3761 Result = DAG.getTargetConstant(Value, Op.getValueType());
3763 case 'M': // "M" is a constant that is greater than 31.
3765 Result = DAG.getTargetConstant(Value, Op.getValueType());
3767 case 'N': // "N" is a positive constant that is an exact power of two.
3768 if ((int)Value > 0 && isPowerOf2_32(Value))
3769 Result = DAG.getTargetConstant(Value, Op.getValueType());
3771 case 'O': // "O" is the constant zero.
3773 Result = DAG.getTargetConstant(Value, Op.getValueType());
3775 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3776 if ((short)-Value == (int)-Value)
3777 Result = DAG.getTargetConstant(Value, Op.getValueType());
3785 Ops.push_back(Result);
3789 // Handle standard constraint letters.
3790 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3793 // isLegalAddressingMode - Return true if the addressing mode represented
3794 // by AM is legal for this target, for a load/store of the specified type.
3795 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3796 const Type *Ty) const {
3797 // FIXME: PPC does not allow r+i addressing modes for vectors!
3799 // PPC allows a sign-extended 16-bit immediate field.
3800 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3803 // No global is ever allowed as a base.
3807 // PPC only support r+r,
3809 case 0: // "r+i" or just "i", depending on HasBaseReg.
3812 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3814 // Otherwise we have r+r or r+i.
3817 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3819 // Allow 2*r as r+r.
3822 // No other scales are supported.
3829 /// isLegalAddressImmediate - Return true if the integer value can be used
3830 /// as the offset of the target addressing mode for load / store of the
3832 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3833 // PPC allows a sign-extended 16-bit immediate field.
3834 return (V > -(1 << 16) && V < (1 << 16)-1);
3837 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3841 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3842 // Depths > 0 not supported yet!
3843 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3846 MachineFunction &MF = DAG.getMachineFunction();
3847 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3848 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3850 bool isPPC64 = PPCSubTarget.isPPC64();
3852 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3854 // Set up a frame object for the return address.
3855 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3857 // Remember it for next time.
3858 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3860 // Make sure the function really does not optimize away the store of the RA
3862 FuncInfo->setLRStoreRequired();
3865 // Just load the return address off the stack.
3866 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3867 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3870 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3871 // Depths > 0 not supported yet!
3872 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3875 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3876 bool isPPC64 = PtrVT == MVT::i64;
3878 MachineFunction &MF = DAG.getMachineFunction();
3879 MachineFrameInfo *MFI = MF.getFrameInfo();
3880 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3881 && MFI->getStackSize();
3884 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3887 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,