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);
85 // PowerPC has no SREM/UREM instructions
86 setOperationAction(ISD::SREM, MVT::i32, Expand);
87 setOperationAction(ISD::UREM, MVT::i32, Expand);
88 setOperationAction(ISD::SREM, MVT::i64, Expand);
89 setOperationAction(ISD::UREM, MVT::i64, Expand);
91 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
92 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
93 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
94 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
95 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
96 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
97 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
98 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
99 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
101 // We don't support sin/cos/sqrt/fmod/pow
102 setOperationAction(ISD::FSIN , MVT::f64, Expand);
103 setOperationAction(ISD::FCOS , MVT::f64, Expand);
104 setOperationAction(ISD::FREM , MVT::f64, Expand);
105 setOperationAction(ISD::FPOW , MVT::f64, Expand);
106 setOperationAction(ISD::FSIN , MVT::f32, Expand);
107 setOperationAction(ISD::FCOS , MVT::f32, Expand);
108 setOperationAction(ISD::FREM , MVT::f32, Expand);
109 setOperationAction(ISD::FPOW , MVT::f32, Expand);
111 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
113 // If we're enabling GP optimizations, use hardware square root
114 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
115 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
116 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
119 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
120 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
122 // PowerPC does not have BSWAP, CTPOP or CTTZ
123 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
124 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
125 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
126 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
127 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
128 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
130 // PowerPC does not have ROTR
131 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
133 // PowerPC does not have Select
134 setOperationAction(ISD::SELECT, MVT::i32, Expand);
135 setOperationAction(ISD::SELECT, MVT::i64, Expand);
136 setOperationAction(ISD::SELECT, MVT::f32, Expand);
137 setOperationAction(ISD::SELECT, MVT::f64, Expand);
139 // PowerPC wants to turn select_cc of FP into fsel when possible.
140 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
141 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
143 // PowerPC wants to optimize integer setcc a bit
144 setOperationAction(ISD::SETCC, MVT::i32, Custom);
146 // PowerPC does not have BRCOND which requires SetCC
147 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
149 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
151 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
152 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
154 // PowerPC does not have [U|S]INT_TO_FP
155 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
156 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
158 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
159 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
160 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
161 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
163 // We cannot sextinreg(i1). Expand to shifts.
164 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
166 // Support label based line numbers.
167 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
168 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
170 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
171 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
172 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
173 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
176 // We want to legalize GlobalAddress and ConstantPool nodes into the
177 // appropriate instructions to materialize the address.
178 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
179 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
180 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
181 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
182 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
183 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
184 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
185 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
187 // RET must be custom lowered, to meet ABI requirements
188 setOperationAction(ISD::RET , MVT::Other, Custom);
190 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
191 setOperationAction(ISD::VASTART , MVT::Other, Custom);
193 // VAARG is custom lowered with ELF 32 ABI
194 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
195 setOperationAction(ISD::VAARG, MVT::Other, Custom);
197 setOperationAction(ISD::VAARG, MVT::Other, Expand);
199 // Use the default implementation.
200 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
201 setOperationAction(ISD::VAEND , MVT::Other, Expand);
202 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
203 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
204 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
205 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
207 // We want to custom lower some of our intrinsics.
208 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
210 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
211 // They also have instructions for converting between i64 and fp.
212 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
213 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
214 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
215 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
216 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
218 // FIXME: disable this lowered code. This generates 64-bit register values,
219 // and we don't model the fact that the top part is clobbered by calls. We
220 // need to flag these together so that the value isn't live across a call.
221 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
223 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
224 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
226 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
227 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
230 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
231 // 64-bit PowerPC implementations can support i64 types directly
232 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
233 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
234 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
236 // 32-bit PowerPC wants to expand i64 shifts itself.
237 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
238 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
239 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
242 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
243 // First set operation action for all vector types to expand. Then we
244 // will selectively turn on ones that can be effectively codegen'd.
245 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
246 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
247 // add/sub are legal for all supported vector VT's.
248 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
249 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
251 // We promote all shuffles to v16i8.
252 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
253 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
255 // We promote all non-typed operations to v4i32.
256 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
257 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
258 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
259 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
260 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
261 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
262 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
263 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
264 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
265 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
266 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
267 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
269 // No other operations are legal.
270 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
271 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
272 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
273 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
274 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
275 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
276 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
277 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
278 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
279 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
280 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
281 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
282 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
283 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
284 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
285 setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
286 setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
287 setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
288 setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
291 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
292 // with merges, splats, etc.
293 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
295 setOperationAction(ISD::AND , MVT::v4i32, Legal);
296 setOperationAction(ISD::OR , MVT::v4i32, Legal);
297 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
298 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
299 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
300 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
302 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
303 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
304 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
305 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
307 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
308 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
309 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
310 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
312 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
313 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
315 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
316 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
317 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
318 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
321 setSetCCResultType(MVT::i32);
322 setShiftAmountType(MVT::i32);
323 setSetCCResultContents(ZeroOrOneSetCCResult);
325 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
326 setStackPointerRegisterToSaveRestore(PPC::X1);
327 setExceptionPointerRegister(PPC::X3);
328 setExceptionSelectorRegister(PPC::X4);
330 setStackPointerRegisterToSaveRestore(PPC::R1);
331 setExceptionPointerRegister(PPC::R3);
332 setExceptionSelectorRegister(PPC::R4);
335 // We have target-specific dag combine patterns for the following nodes:
336 setTargetDAGCombine(ISD::SINT_TO_FP);
337 setTargetDAGCombine(ISD::STORE);
338 setTargetDAGCombine(ISD::BR_CC);
339 setTargetDAGCombine(ISD::BSWAP);
341 // Darwin long double math library functions have $LDBL128 appended.
342 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
343 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
344 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
345 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
346 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
347 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
350 computeRegisterProperties();
353 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
356 case PPCISD::FSEL: return "PPCISD::FSEL";
357 case PPCISD::FCFID: return "PPCISD::FCFID";
358 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
359 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
360 case PPCISD::STFIWX: return "PPCISD::STFIWX";
361 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
362 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
363 case PPCISD::VPERM: return "PPCISD::VPERM";
364 case PPCISD::Hi: return "PPCISD::Hi";
365 case PPCISD::Lo: return "PPCISD::Lo";
366 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
367 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
368 case PPCISD::SRL: return "PPCISD::SRL";
369 case PPCISD::SRA: return "PPCISD::SRA";
370 case PPCISD::SHL: return "PPCISD::SHL";
371 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
372 case PPCISD::STD_32: return "PPCISD::STD_32";
373 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
374 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
375 case PPCISD::MTCTR: return "PPCISD::MTCTR";
376 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
377 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
378 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
379 case PPCISD::MFCR: return "PPCISD::MFCR";
380 case PPCISD::VCMP: return "PPCISD::VCMP";
381 case PPCISD::VCMPo: return "PPCISD::VCMPo";
382 case PPCISD::LBRX: return "PPCISD::LBRX";
383 case PPCISD::STBRX: return "PPCISD::STBRX";
384 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
385 case PPCISD::MFFS: return "PPCISD::MFFS";
386 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
387 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
388 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
389 case PPCISD::MTFSF: return "PPCISD::MTFSF";
393 //===----------------------------------------------------------------------===//
394 // Node matching predicates, for use by the tblgen matching code.
395 //===----------------------------------------------------------------------===//
397 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
398 static bool isFloatingPointZero(SDOperand Op) {
399 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
400 return CFP->getValueAPF().isZero();
401 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
402 // Maybe this has already been legalized into the constant pool?
403 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
404 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
405 return CFP->getValueAPF().isZero();
410 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
411 /// true if Op is undef or if it matches the specified value.
412 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
413 return Op.getOpcode() == ISD::UNDEF ||
414 cast<ConstantSDNode>(Op)->getValue() == Val;
417 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
418 /// VPKUHUM instruction.
419 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
421 for (unsigned i = 0; i != 16; ++i)
422 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
425 for (unsigned i = 0; i != 8; ++i)
426 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
427 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
433 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
434 /// VPKUWUM instruction.
435 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
437 for (unsigned i = 0; i != 16; i += 2)
438 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
439 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
442 for (unsigned i = 0; i != 8; i += 2)
443 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
444 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
445 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
446 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
452 /// isVMerge - Common function, used to match vmrg* shuffles.
454 static bool isVMerge(SDNode *N, unsigned UnitSize,
455 unsigned LHSStart, unsigned RHSStart) {
456 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
457 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
458 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
459 "Unsupported merge size!");
461 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
462 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
463 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
464 LHSStart+j+i*UnitSize) ||
465 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
466 RHSStart+j+i*UnitSize))
472 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
473 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
474 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
476 return isVMerge(N, UnitSize, 8, 24);
477 return isVMerge(N, UnitSize, 8, 8);
480 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
481 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
482 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
484 return isVMerge(N, UnitSize, 0, 16);
485 return isVMerge(N, UnitSize, 0, 0);
489 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
490 /// amount, otherwise return -1.
491 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
492 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
493 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
494 // Find the first non-undef value in the shuffle mask.
496 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
499 if (i == 16) return -1; // all undef.
501 // Otherwise, check to see if the rest of the elements are consequtively
502 // numbered from this value.
503 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
504 if (ShiftAmt < i) return -1;
508 // Check the rest of the elements to see if they are consequtive.
509 for (++i; i != 16; ++i)
510 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
513 // Check the rest of the elements to see if they are consequtive.
514 for (++i; i != 16; ++i)
515 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
522 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
523 /// specifies a splat of a single element that is suitable for input to
524 /// VSPLTB/VSPLTH/VSPLTW.
525 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
526 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
527 N->getNumOperands() == 16 &&
528 (EltSize == 1 || EltSize == 2 || EltSize == 4));
530 // This is a splat operation if each element of the permute is the same, and
531 // if the value doesn't reference the second vector.
532 unsigned ElementBase = 0;
533 SDOperand Elt = N->getOperand(0);
534 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
535 ElementBase = EltV->getValue();
537 return false; // FIXME: Handle UNDEF elements too!
539 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
542 // Check that they are consequtive.
543 for (unsigned i = 1; i != EltSize; ++i) {
544 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
545 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
549 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
550 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
551 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
552 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
553 "Invalid VECTOR_SHUFFLE mask!");
554 for (unsigned j = 0; j != EltSize; ++j)
555 if (N->getOperand(i+j) != N->getOperand(j))
562 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
564 bool PPC::isAllNegativeZeroVector(SDNode *N) {
565 assert(N->getOpcode() == ISD::BUILD_VECTOR);
566 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
567 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
568 return CFP->getValueAPF().isNegZero();
572 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
573 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
574 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
575 assert(isSplatShuffleMask(N, EltSize));
576 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
579 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
580 /// by using a vspltis[bhw] instruction of the specified element size, return
581 /// the constant being splatted. The ByteSize field indicates the number of
582 /// bytes of each element [124] -> [bhw].
583 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
584 SDOperand OpVal(0, 0);
586 // If ByteSize of the splat is bigger than the element size of the
587 // build_vector, then we have a case where we are checking for a splat where
588 // multiple elements of the buildvector are folded together into a single
589 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
590 unsigned EltSize = 16/N->getNumOperands();
591 if (EltSize < ByteSize) {
592 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
593 SDOperand UniquedVals[4];
594 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
596 // See if all of the elements in the buildvector agree across.
597 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
598 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
599 // If the element isn't a constant, bail fully out.
600 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
603 if (UniquedVals[i&(Multiple-1)].Val == 0)
604 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
605 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
606 return SDOperand(); // no match.
609 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
610 // either constant or undef values that are identical for each chunk. See
611 // if these chunks can form into a larger vspltis*.
613 // Check to see if all of the leading entries are either 0 or -1. If
614 // neither, then this won't fit into the immediate field.
615 bool LeadingZero = true;
616 bool LeadingOnes = true;
617 for (unsigned i = 0; i != Multiple-1; ++i) {
618 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
620 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
621 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
623 // Finally, check the least significant entry.
625 if (UniquedVals[Multiple-1].Val == 0)
626 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
627 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
629 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
632 if (UniquedVals[Multiple-1].Val == 0)
633 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
634 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
635 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
636 return DAG.getTargetConstant(Val, MVT::i32);
642 // Check to see if this buildvec has a single non-undef value in its elements.
643 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
644 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
646 OpVal = N->getOperand(i);
647 else if (OpVal != N->getOperand(i))
651 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
653 unsigned ValSizeInBytes = 0;
655 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
656 Value = CN->getValue();
657 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
658 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
659 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
660 Value = FloatToBits(CN->getValueAPF().convertToFloat());
664 // If the splat value is larger than the element value, then we can never do
665 // this splat. The only case that we could fit the replicated bits into our
666 // immediate field for would be zero, and we prefer to use vxor for it.
667 if (ValSizeInBytes < ByteSize) return SDOperand();
669 // If the element value is larger than the splat value, cut it in half and
670 // check to see if the two halves are equal. Continue doing this until we
671 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
672 while (ValSizeInBytes > ByteSize) {
673 ValSizeInBytes >>= 1;
675 // If the top half equals the bottom half, we're still ok.
676 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
677 (Value & ((1 << (8*ValSizeInBytes))-1)))
681 // Properly sign extend the value.
682 int ShAmt = (4-ByteSize)*8;
683 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
685 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
686 if (MaskVal == 0) return SDOperand();
688 // Finally, if this value fits in a 5 bit sext field, return it
689 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
690 return DAG.getTargetConstant(MaskVal, MVT::i32);
694 //===----------------------------------------------------------------------===//
695 // Addressing Mode Selection
696 //===----------------------------------------------------------------------===//
698 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
699 /// or 64-bit immediate, and if the value can be accurately represented as a
700 /// sign extension from a 16-bit value. If so, this returns true and the
702 static bool isIntS16Immediate(SDNode *N, short &Imm) {
703 if (N->getOpcode() != ISD::Constant)
706 Imm = (short)cast<ConstantSDNode>(N)->getValue();
707 if (N->getValueType(0) == MVT::i32)
708 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
710 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
712 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
713 return isIntS16Immediate(Op.Val, Imm);
717 /// SelectAddressRegReg - Given the specified addressed, check to see if it
718 /// can be represented as an indexed [r+r] operation. Returns false if it
719 /// can be more efficiently represented with [r+imm].
720 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
724 if (N.getOpcode() == ISD::ADD) {
725 if (isIntS16Immediate(N.getOperand(1), imm))
727 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
730 Base = N.getOperand(0);
731 Index = N.getOperand(1);
733 } else if (N.getOpcode() == ISD::OR) {
734 if (isIntS16Immediate(N.getOperand(1), imm))
735 return false; // r+i can fold it if we can.
737 // If this is an or of disjoint bitfields, we can codegen this as an add
738 // (for better address arithmetic) if the LHS and RHS of the OR are provably
740 uint64_t LHSKnownZero, LHSKnownOne;
741 uint64_t RHSKnownZero, RHSKnownOne;
742 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
745 DAG.ComputeMaskedBits(N.getOperand(1), ~0U, RHSKnownZero, RHSKnownOne);
746 // If all of the bits are known zero on the LHS or RHS, the add won't
748 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
749 Base = N.getOperand(0);
750 Index = N.getOperand(1);
759 /// Returns true if the address N can be represented by a base register plus
760 /// a signed 16-bit displacement [r+imm], and if it is not better
761 /// represented as reg+reg.
762 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
763 SDOperand &Base, SelectionDAG &DAG){
764 // If this can be more profitably realized as r+r, fail.
765 if (SelectAddressRegReg(N, Disp, Base, DAG))
768 if (N.getOpcode() == ISD::ADD) {
770 if (isIntS16Immediate(N.getOperand(1), imm)) {
771 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
772 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
773 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
775 Base = N.getOperand(0);
777 return true; // [r+i]
778 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
779 // Match LOAD (ADD (X, Lo(G))).
780 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
781 && "Cannot handle constant offsets yet!");
782 Disp = N.getOperand(1).getOperand(0); // The global address.
783 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
784 Disp.getOpcode() == ISD::TargetConstantPool ||
785 Disp.getOpcode() == ISD::TargetJumpTable);
786 Base = N.getOperand(0);
787 return true; // [&g+r]
789 } else if (N.getOpcode() == ISD::OR) {
791 if (isIntS16Immediate(N.getOperand(1), imm)) {
792 // If this is an or of disjoint bitfields, we can codegen this as an add
793 // (for better address arithmetic) if the LHS and RHS of the OR are
794 // provably disjoint.
795 uint64_t LHSKnownZero, LHSKnownOne;
796 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
797 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
798 // If all of the bits are known zero on the LHS or RHS, the add won't
800 Base = N.getOperand(0);
801 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
805 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
806 // Loading from a constant address.
808 // If this address fits entirely in a 16-bit sext immediate field, codegen
811 if (isIntS16Immediate(CN, Imm)) {
812 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
813 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
817 // Handle 32-bit sext immediates with LIS + addr mode.
818 if (CN->getValueType(0) == MVT::i32 ||
819 (int64_t)CN->getValue() == (int)CN->getValue()) {
820 int Addr = (int)CN->getValue();
822 // Otherwise, break this down into an LIS + disp.
823 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
825 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
826 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
827 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
832 Disp = DAG.getTargetConstant(0, getPointerTy());
833 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
834 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
837 return true; // [r+0]
840 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
841 /// represented as an indexed [r+r] operation.
842 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
845 // Check to see if we can easily represent this as an [r+r] address. This
846 // will fail if it thinks that the address is more profitably represented as
847 // reg+imm, e.g. where imm = 0.
848 if (SelectAddressRegReg(N, Base, Index, DAG))
851 // If the operand is an addition, always emit this as [r+r], since this is
852 // better (for code size, and execution, as the memop does the add for free)
853 // than emitting an explicit add.
854 if (N.getOpcode() == ISD::ADD) {
855 Base = N.getOperand(0);
856 Index = N.getOperand(1);
860 // Otherwise, do it the hard way, using R0 as the base register.
861 Base = DAG.getRegister(PPC::R0, N.getValueType());
866 /// SelectAddressRegImmShift - Returns true if the address N can be
867 /// represented by a base register plus a signed 14-bit displacement
868 /// [r+imm*4]. Suitable for use by STD and friends.
869 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
872 // If this can be more profitably realized as r+r, fail.
873 if (SelectAddressRegReg(N, Disp, Base, DAG))
876 if (N.getOpcode() == ISD::ADD) {
878 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
879 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
880 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
881 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
883 Base = N.getOperand(0);
885 return true; // [r+i]
886 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
887 // Match LOAD (ADD (X, Lo(G))).
888 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
889 && "Cannot handle constant offsets yet!");
890 Disp = N.getOperand(1).getOperand(0); // The global address.
891 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
892 Disp.getOpcode() == ISD::TargetConstantPool ||
893 Disp.getOpcode() == ISD::TargetJumpTable);
894 Base = N.getOperand(0);
895 return true; // [&g+r]
897 } else if (N.getOpcode() == ISD::OR) {
899 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
900 // If this is an or of disjoint bitfields, we can codegen this as an add
901 // (for better address arithmetic) if the LHS and RHS of the OR are
902 // provably disjoint.
903 uint64_t LHSKnownZero, LHSKnownOne;
904 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
905 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
906 // If all of the bits are known zero on the LHS or RHS, the add won't
908 Base = N.getOperand(0);
909 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
913 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
914 // Loading from a constant address. Verify low two bits are clear.
915 if ((CN->getValue() & 3) == 0) {
916 // If this address fits entirely in a 14-bit sext immediate field, codegen
919 if (isIntS16Immediate(CN, Imm)) {
920 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
921 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
925 // Fold the low-part of 32-bit absolute addresses into addr mode.
926 if (CN->getValueType(0) == MVT::i32 ||
927 (int64_t)CN->getValue() == (int)CN->getValue()) {
928 int Addr = (int)CN->getValue();
930 // Otherwise, break this down into an LIS + disp.
931 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
933 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
934 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
935 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
941 Disp = DAG.getTargetConstant(0, getPointerTy());
942 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
943 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
946 return true; // [r+0]
950 /// getPreIndexedAddressParts - returns true by value, base pointer and
951 /// offset pointer and addressing mode by reference if the node's address
952 /// can be legally represented as pre-indexed load / store address.
953 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
955 ISD::MemIndexedMode &AM,
957 // Disabled by default for now.
958 if (!EnablePPCPreinc) return false;
962 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
963 Ptr = LD->getBasePtr();
964 VT = LD->getMemoryVT();
966 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
968 Ptr = ST->getBasePtr();
969 VT = ST->getMemoryVT();
973 // PowerPC doesn't have preinc load/store instructions for vectors.
974 if (MVT::isVector(VT))
977 // TODO: Check reg+reg first.
979 // LDU/STU use reg+imm*4, others use reg+imm.
980 if (VT != MVT::i64) {
982 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
986 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
990 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
991 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
992 // sext i32 to i64 when addr mode is r+i.
993 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
994 LD->getExtensionType() == ISD::SEXTLOAD &&
995 isa<ConstantSDNode>(Offset))
1003 //===----------------------------------------------------------------------===//
1004 // LowerOperation implementation
1005 //===----------------------------------------------------------------------===//
1007 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
1008 MVT::ValueType PtrVT = Op.getValueType();
1009 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1010 Constant *C = CP->getConstVal();
1011 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1012 SDOperand Zero = DAG.getConstant(0, PtrVT);
1014 const TargetMachine &TM = DAG.getTarget();
1016 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1017 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1019 // If this is a non-darwin platform, we don't support non-static relo models
1021 if (TM.getRelocationModel() == Reloc::Static ||
1022 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1023 // Generate non-pic code that has direct accesses to the constant pool.
1024 // The address of the global is just (hi(&g)+lo(&g)).
1025 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1028 if (TM.getRelocationModel() == Reloc::PIC_) {
1029 // With PIC, the first instruction is actually "GR+hi(&G)".
1030 Hi = DAG.getNode(ISD::ADD, PtrVT,
1031 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1034 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1038 static SDOperand LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1039 MVT::ValueType PtrVT = Op.getValueType();
1040 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1041 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1042 SDOperand Zero = DAG.getConstant(0, PtrVT);
1044 const TargetMachine &TM = DAG.getTarget();
1046 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1047 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1049 // If this is a non-darwin platform, we don't support non-static relo models
1051 if (TM.getRelocationModel() == Reloc::Static ||
1052 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1053 // Generate non-pic code that has direct accesses to the constant pool.
1054 // The address of the global is just (hi(&g)+lo(&g)).
1055 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1058 if (TM.getRelocationModel() == Reloc::PIC_) {
1059 // With PIC, the first instruction is actually "GR+hi(&G)".
1060 Hi = DAG.getNode(ISD::ADD, PtrVT,
1061 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1064 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1068 static SDOperand LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
1069 assert(0 && "TLS not implemented for PPC.");
1072 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
1073 MVT::ValueType PtrVT = Op.getValueType();
1074 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1075 GlobalValue *GV = GSDN->getGlobal();
1076 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1077 // If it's a debug information descriptor, don't mess with it.
1078 if (DAG.isVerifiedDebugInfoDesc(Op))
1080 SDOperand Zero = DAG.getConstant(0, PtrVT);
1082 const TargetMachine &TM = DAG.getTarget();
1084 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1085 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1087 // If this is a non-darwin platform, we don't support non-static relo models
1089 if (TM.getRelocationModel() == Reloc::Static ||
1090 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1091 // Generate non-pic code that has direct accesses to globals.
1092 // The address of the global is just (hi(&g)+lo(&g)).
1093 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1096 if (TM.getRelocationModel() == Reloc::PIC_) {
1097 // With PIC, the first instruction is actually "GR+hi(&G)".
1098 Hi = DAG.getNode(ISD::ADD, PtrVT,
1099 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1102 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1104 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1107 // If the global is weak or external, we have to go through the lazy
1109 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1112 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1113 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1115 // If we're comparing for equality to zero, expose the fact that this is
1116 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1117 // fold the new nodes.
1118 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1119 if (C->isNullValue() && CC == ISD::SETEQ) {
1120 MVT::ValueType VT = Op.getOperand(0).getValueType();
1121 SDOperand Zext = Op.getOperand(0);
1122 if (VT < MVT::i32) {
1124 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1126 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1127 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1128 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1129 DAG.getConstant(Log2b, MVT::i32));
1130 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1132 // Leave comparisons against 0 and -1 alone for now, since they're usually
1133 // optimized. FIXME: revisit this when we can custom lower all setcc
1135 if (C->isAllOnesValue() || C->isNullValue())
1139 // If we have an integer seteq/setne, turn it into a compare against zero
1140 // by xor'ing the rhs with the lhs, which is faster than setting a
1141 // condition register, reading it back out, and masking the correct bit. The
1142 // normal approach here uses sub to do this instead of xor. Using xor exposes
1143 // the result to other bit-twiddling opportunities.
1144 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1145 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1146 MVT::ValueType VT = Op.getValueType();
1147 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1149 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1154 static SDOperand LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1155 int VarArgsFrameIndex,
1156 int VarArgsStackOffset,
1157 unsigned VarArgsNumGPR,
1158 unsigned VarArgsNumFPR,
1159 const PPCSubtarget &Subtarget) {
1161 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1164 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1165 int VarArgsFrameIndex,
1166 int VarArgsStackOffset,
1167 unsigned VarArgsNumGPR,
1168 unsigned VarArgsNumFPR,
1169 const PPCSubtarget &Subtarget) {
1171 if (Subtarget.isMachoABI()) {
1172 // vastart just stores the address of the VarArgsFrameIndex slot into the
1173 // memory location argument.
1174 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1175 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1176 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1177 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1180 // For ELF 32 ABI we follow the layout of the va_list struct.
1181 // We suppose the given va_list is already allocated.
1184 // char gpr; /* index into the array of 8 GPRs
1185 // * stored in the register save area
1186 // * gpr=0 corresponds to r3,
1187 // * gpr=1 to r4, etc.
1189 // char fpr; /* index into the array of 8 FPRs
1190 // * stored in the register save area
1191 // * fpr=0 corresponds to f1,
1192 // * fpr=1 to f2, etc.
1194 // char *overflow_arg_area;
1195 // /* location on stack that holds
1196 // * the next overflow argument
1198 // char *reg_save_area;
1199 // /* where r3:r10 and f1:f8 (if saved)
1205 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1206 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1209 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1211 SDOperand StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1212 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1214 uint64_t FrameOffset = MVT::getSizeInBits(PtrVT)/8;
1215 SDOperand ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1217 uint64_t StackOffset = MVT::getSizeInBits(PtrVT)/8 - 1;
1218 SDOperand ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1220 uint64_t FPROffset = 1;
1221 SDOperand ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1223 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1225 // Store first byte : number of int regs
1226 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1227 Op.getOperand(1), SV, 0);
1228 uint64_t nextOffset = FPROffset;
1229 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1232 // Store second byte : number of float regs
1233 SDOperand secondStore =
1234 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1235 nextOffset += StackOffset;
1236 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1238 // Store second word : arguments given on stack
1239 SDOperand thirdStore =
1240 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1241 nextOffset += FrameOffset;
1242 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1244 // Store third word : arguments given in registers
1245 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1249 #include "PPCGenCallingConv.inc"
1251 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1252 /// depending on which subtarget is selected.
1253 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1254 if (Subtarget.isMachoABI()) {
1255 static const unsigned FPR[] = {
1256 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1257 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1263 static const unsigned FPR[] = {
1264 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1270 static SDOperand LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG,
1271 int &VarArgsFrameIndex,
1272 int &VarArgsStackOffset,
1273 unsigned &VarArgsNumGPR,
1274 unsigned &VarArgsNumFPR,
1275 const PPCSubtarget &Subtarget) {
1276 // TODO: add description of PPC stack frame format, or at least some docs.
1278 MachineFunction &MF = DAG.getMachineFunction();
1279 MachineFrameInfo *MFI = MF.getFrameInfo();
1280 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1281 SmallVector<SDOperand, 8> ArgValues;
1282 SDOperand Root = Op.getOperand(0);
1284 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1285 bool isPPC64 = PtrVT == MVT::i64;
1286 bool isMachoABI = Subtarget.isMachoABI();
1287 bool isELF32_ABI = Subtarget.isELF32_ABI();
1288 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1290 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1292 static const unsigned GPR_32[] = { // 32-bit registers.
1293 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1294 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1296 static const unsigned GPR_64[] = { // 64-bit registers.
1297 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1298 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1301 static const unsigned *FPR = GetFPR(Subtarget);
1303 static const unsigned VR[] = {
1304 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1305 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1308 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1309 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1310 const unsigned Num_VR_Regs = array_lengthof( VR);
1312 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1314 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1316 // Add DAG nodes to load the arguments or copy them out of registers. On
1317 // entry to a function on PPC, the arguments start after the linkage area,
1318 // although the first ones are often in registers.
1320 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1321 // represented with two words (long long or double) must be copied to an
1322 // even GPR_idx value or to an even ArgOffset value.
1324 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1326 bool needsLoad = false;
1327 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1328 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1329 unsigned ArgSize = ObjSize;
1330 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1331 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1332 // See if next argument requires stack alignment in ELF
1333 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1334 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1335 (!(Flags & AlignFlag)));
1337 unsigned CurArgOffset = ArgOffset;
1339 default: assert(0 && "Unhandled argument type!");
1341 // Double word align in ELF
1342 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1343 if (GPR_idx != Num_GPR_Regs) {
1344 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1345 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1346 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1350 ArgSize = PtrByteSize;
1352 // Stack align in ELF
1353 if (needsLoad && Expand && isELF32_ABI)
1354 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1355 // All int arguments reserve stack space in Macho ABI.
1356 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1359 case MVT::i64: // PPC64
1360 if (GPR_idx != Num_GPR_Regs) {
1361 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1362 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1363 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1368 // All int arguments reserve stack space in Macho ABI.
1369 if (isMachoABI || needsLoad) ArgOffset += 8;
1374 // Every 4 bytes of argument space consumes one of the GPRs available for
1375 // argument passing.
1376 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1378 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1381 if (FPR_idx != Num_FPR_Regs) {
1383 if (ObjectVT == MVT::f32)
1384 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1386 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1387 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1388 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1394 // Stack align in ELF
1395 if (needsLoad && Expand && isELF32_ABI)
1396 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1397 // All FP arguments reserve stack space in Macho ABI.
1398 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1404 // Note that vector arguments in registers don't reserve stack space.
1405 if (VR_idx != Num_VR_Regs) {
1406 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1407 RegInfo.addLiveIn(VR[VR_idx], VReg);
1408 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1411 // This should be simple, but requires getting 16-byte aligned stack
1413 assert(0 && "Loading VR argument not implemented yet!");
1419 // We need to load the argument to a virtual register if we determined above
1420 // that we ran out of physical registers of the appropriate type.
1422 int FI = MFI->CreateFixedObject(ObjSize,
1423 CurArgOffset + (ArgSize - ObjSize));
1424 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1425 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1428 ArgValues.push_back(ArgVal);
1431 // If the function takes variable number of arguments, make a frame index for
1432 // the start of the first vararg value... for expansion of llvm.va_start.
1433 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1438 VarArgsNumGPR = GPR_idx;
1439 VarArgsNumFPR = FPR_idx;
1441 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1443 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1444 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1445 MVT::getSizeInBits(PtrVT)/8);
1447 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1454 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1456 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1458 SmallVector<SDOperand, 8> MemOps;
1460 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1461 // stored to the VarArgsFrameIndex on the stack.
1463 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1464 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1465 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1466 MemOps.push_back(Store);
1467 // Increment the address by four for the next argument to store
1468 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1469 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1473 // If this function is vararg, store any remaining integer argument regs
1474 // to their spots on the stack so that they may be loaded by deferencing the
1475 // result of va_next.
1476 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1479 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1481 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1483 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1484 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1485 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1486 MemOps.push_back(Store);
1487 // Increment the address by four for the next argument to store
1488 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1489 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1492 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1495 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1496 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1497 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1498 MemOps.push_back(Store);
1499 // Increment the address by eight for the next argument to store
1500 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1502 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1505 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1507 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1509 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1510 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1511 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1512 MemOps.push_back(Store);
1513 // Increment the address by eight for the next argument to store
1514 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1516 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1520 if (!MemOps.empty())
1521 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1524 ArgValues.push_back(Root);
1526 // Return the new list of results.
1527 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1528 Op.Val->value_end());
1529 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1532 /// isCallCompatibleAddress - Return the immediate to use if the specified
1533 /// 32-bit value is representable in the immediate field of a BxA instruction.
1534 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1535 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1538 int Addr = C->getValue();
1539 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1540 (Addr << 6 >> 6) != Addr)
1541 return 0; // Top 6 bits have to be sext of immediate.
1543 return DAG.getConstant((int)C->getValue() >> 2,
1544 DAG.getTargetLoweringInfo().getPointerTy()).Val;
1548 static SDOperand LowerCALL(SDOperand Op, SelectionDAG &DAG,
1549 const PPCSubtarget &Subtarget) {
1550 SDOperand Chain = Op.getOperand(0);
1551 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1552 SDOperand Callee = Op.getOperand(4);
1553 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1555 bool isMachoABI = Subtarget.isMachoABI();
1556 bool isELF32_ABI = Subtarget.isELF32_ABI();
1558 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1559 bool isPPC64 = PtrVT == MVT::i64;
1560 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1562 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1563 // SelectExpr to use to put the arguments in the appropriate registers.
1564 std::vector<SDOperand> args_to_use;
1566 // Count how many bytes are to be pushed on the stack, including the linkage
1567 // area, and parameter passing area. We start with 24/48 bytes, which is
1568 // prereserved space for [SP][CR][LR][3 x unused].
1569 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1571 // Add up all the space actually used.
1572 for (unsigned i = 0; i != NumOps; ++i) {
1573 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1574 ArgSize = std::max(ArgSize, PtrByteSize);
1575 NumBytes += ArgSize;
1578 // The prolog code of the callee may store up to 8 GPR argument registers to
1579 // the stack, allowing va_start to index over them in memory if its varargs.
1580 // Because we cannot tell if this is needed on the caller side, we have to
1581 // conservatively assume that it is needed. As such, make sure we have at
1582 // least enough stack space for the caller to store the 8 GPRs.
1583 NumBytes = std::max(NumBytes,
1584 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1586 // Adjust the stack pointer for the new arguments...
1587 // These operations are automatically eliminated by the prolog/epilog pass
1588 Chain = DAG.getCALLSEQ_START(Chain,
1589 DAG.getConstant(NumBytes, PtrVT));
1591 // Set up a copy of the stack pointer for use loading and storing any
1592 // arguments that may not fit in the registers available for argument
1596 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1598 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1600 // Figure out which arguments are going to go in registers, and which in
1601 // memory. Also, if this is a vararg function, floating point operations
1602 // must be stored to our stack, and loaded into integer regs as well, if
1603 // any integer regs are available for argument passing.
1604 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1605 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1607 static const unsigned GPR_32[] = { // 32-bit registers.
1608 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1609 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1611 static const unsigned GPR_64[] = { // 64-bit registers.
1612 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1613 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1615 static const unsigned *FPR = GetFPR(Subtarget);
1617 static const unsigned VR[] = {
1618 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1619 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1621 const unsigned NumGPRs = array_lengthof(GPR_32);
1622 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1623 const unsigned NumVRs = array_lengthof( VR);
1625 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1627 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1628 SmallVector<SDOperand, 8> MemOpChains;
1629 for (unsigned i = 0; i != NumOps; ++i) {
1631 SDOperand Arg = Op.getOperand(5+2*i);
1632 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1633 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1634 // See if next argument requires stack alignment in ELF
1635 unsigned next = 5+2*(i+1)+1;
1636 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1637 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1638 (!(Flags & AlignFlag)));
1640 // PtrOff will be used to store the current argument to the stack if a
1641 // register cannot be found for it.
1644 // Stack align in ELF 32
1645 if (isELF32_ABI && Expand)
1646 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1647 StackPtr.getValueType());
1649 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1651 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1653 // On PPC64, promote integers to 64-bit values.
1654 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1655 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1657 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1660 switch (Arg.getValueType()) {
1661 default: assert(0 && "Unexpected ValueType for argument!");
1664 // Double word align in ELF
1665 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1666 if (GPR_idx != NumGPRs) {
1667 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1669 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1672 if (inMem || isMachoABI) {
1673 // Stack align in ELF
1674 if (isELF32_ABI && Expand)
1675 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1677 ArgOffset += PtrByteSize;
1683 // Float varargs need to be promoted to double.
1684 if (Arg.getValueType() == MVT::f32)
1685 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1688 if (FPR_idx != NumFPRs) {
1689 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1692 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1693 MemOpChains.push_back(Store);
1695 // Float varargs are always shadowed in available integer registers
1696 if (GPR_idx != NumGPRs) {
1697 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1698 MemOpChains.push_back(Load.getValue(1));
1699 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1702 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1703 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1704 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1705 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1706 MemOpChains.push_back(Load.getValue(1));
1707 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1711 // If we have any FPRs remaining, we may also have GPRs remaining.
1712 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1715 if (GPR_idx != NumGPRs)
1717 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1718 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1723 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1726 if (inMem || isMachoABI) {
1727 // Stack align in ELF
1728 if (isELF32_ABI && Expand)
1729 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1733 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1740 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1741 assert(VR_idx != NumVRs &&
1742 "Don't support passing more than 12 vector args yet!");
1743 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1747 if (!MemOpChains.empty())
1748 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1749 &MemOpChains[0], MemOpChains.size());
1751 // Build a sequence of copy-to-reg nodes chained together with token chain
1752 // and flag operands which copy the outgoing args into the appropriate regs.
1754 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1755 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1757 InFlag = Chain.getValue(1);
1760 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1761 if (isVarArg && isELF32_ABI) {
1762 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1763 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1764 InFlag = Chain.getValue(1);
1767 std::vector<MVT::ValueType> NodeTys;
1768 NodeTys.push_back(MVT::Other); // Returns a chain
1769 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1771 SmallVector<SDOperand, 8> Ops;
1772 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1774 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1775 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1776 // node so that legalize doesn't hack it.
1777 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1778 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1779 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1780 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1781 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1782 // If this is an absolute destination address, use the munged value.
1783 Callee = SDOperand(Dest, 0);
1785 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1786 // to do the call, we can't use PPCISD::CALL.
1787 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1788 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1789 InFlag = Chain.getValue(1);
1791 // Copy the callee address into R12 on darwin.
1793 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1794 InFlag = Chain.getValue(1);
1798 NodeTys.push_back(MVT::Other);
1799 NodeTys.push_back(MVT::Flag);
1800 Ops.push_back(Chain);
1801 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1805 // If this is a direct call, pass the chain and the callee.
1807 Ops.push_back(Chain);
1808 Ops.push_back(Callee);
1811 // Add argument registers to the end of the list so that they are known live
1813 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1814 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1815 RegsToPass[i].second.getValueType()));
1818 Ops.push_back(InFlag);
1819 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1820 InFlag = Chain.getValue(1);
1822 Chain = DAG.getCALLSEQ_END(Chain,
1823 DAG.getConstant(NumBytes, PtrVT),
1824 DAG.getConstant(0, PtrVT),
1826 if (Op.Val->getValueType(0) != MVT::Other)
1827 InFlag = Chain.getValue(1);
1829 SDOperand ResultVals[3];
1830 unsigned NumResults = 0;
1833 // If the call has results, copy the values out of the ret val registers.
1834 switch (Op.Val->getValueType(0)) {
1835 default: assert(0 && "Unexpected ret value!");
1836 case MVT::Other: break;
1838 if (Op.Val->getValueType(1) == MVT::i32) {
1839 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1840 ResultVals[0] = Chain.getValue(0);
1841 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
1842 Chain.getValue(2)).getValue(1);
1843 ResultVals[1] = Chain.getValue(0);
1845 NodeTys.push_back(MVT::i32);
1847 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1848 ResultVals[0] = Chain.getValue(0);
1851 NodeTys.push_back(MVT::i32);
1854 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
1855 ResultVals[0] = Chain.getValue(0);
1857 NodeTys.push_back(MVT::i64);
1860 if (Op.Val->getValueType(1) == MVT::f64) {
1861 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
1862 ResultVals[0] = Chain.getValue(0);
1863 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
1864 Chain.getValue(2)).getValue(1);
1865 ResultVals[1] = Chain.getValue(0);
1867 NodeTys.push_back(MVT::f64);
1868 NodeTys.push_back(MVT::f64);
1871 // else fall through
1873 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
1874 InFlag).getValue(1);
1875 ResultVals[0] = Chain.getValue(0);
1877 NodeTys.push_back(Op.Val->getValueType(0));
1883 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
1884 InFlag).getValue(1);
1885 ResultVals[0] = Chain.getValue(0);
1887 NodeTys.push_back(Op.Val->getValueType(0));
1891 NodeTys.push_back(MVT::Other);
1893 // If the function returns void, just return the chain.
1894 if (NumResults == 0)
1897 // Otherwise, merge everything together with a MERGE_VALUES node.
1898 ResultVals[NumResults++] = Chain;
1899 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1900 ResultVals, NumResults);
1901 return Res.getValue(Op.ResNo);
1904 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1905 SmallVector<CCValAssign, 16> RVLocs;
1906 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1907 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1908 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1909 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
1911 // If this is the first return lowered for this function, add the regs to the
1912 // liveout set for the function.
1913 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1914 for (unsigned i = 0; i != RVLocs.size(); ++i)
1915 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1918 SDOperand Chain = Op.getOperand(0);
1921 // Copy the result values into the output registers.
1922 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1923 CCValAssign &VA = RVLocs[i];
1924 assert(VA.isRegLoc() && "Can only return in registers!");
1925 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1926 Flag = Chain.getValue(1);
1930 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
1932 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
1935 static SDOperand LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
1936 const PPCSubtarget &Subtarget) {
1937 // When we pop the dynamic allocation we need to restore the SP link.
1939 // Get the corect type for pointers.
1940 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1942 // Construct the stack pointer operand.
1943 bool IsPPC64 = Subtarget.isPPC64();
1944 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
1945 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
1947 // Get the operands for the STACKRESTORE.
1948 SDOperand Chain = Op.getOperand(0);
1949 SDOperand SaveSP = Op.getOperand(1);
1951 // Load the old link SP.
1952 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
1954 // Restore the stack pointer.
1955 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
1957 // Store the old link SP.
1958 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
1961 static SDOperand LowerDYNAMIC_STACKALLOC(SDOperand Op, SelectionDAG &DAG,
1962 const PPCSubtarget &Subtarget) {
1963 MachineFunction &MF = DAG.getMachineFunction();
1964 bool IsPPC64 = Subtarget.isPPC64();
1965 bool isMachoABI = Subtarget.isMachoABI();
1967 // Get current frame pointer save index. The users of this index will be
1968 // primarily DYNALLOC instructions.
1969 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1970 int FPSI = FI->getFramePointerSaveIndex();
1972 // If the frame pointer save index hasn't been defined yet.
1974 // Find out what the fix offset of the frame pointer save area.
1975 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
1977 // Allocate the frame index for frame pointer save area.
1978 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
1980 FI->setFramePointerSaveIndex(FPSI);
1984 SDOperand Chain = Op.getOperand(0);
1985 SDOperand Size = Op.getOperand(1);
1987 // Get the corect type for pointers.
1988 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1990 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
1991 DAG.getConstant(0, PtrVT), Size);
1992 // Construct a node for the frame pointer save index.
1993 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
1994 // Build a DYNALLOC node.
1995 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
1996 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
1997 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2001 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2003 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
2004 // Not FP? Not a fsel.
2005 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
2006 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
2009 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2011 // Cannot handle SETEQ/SETNE.
2012 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
2014 MVT::ValueType ResVT = Op.getValueType();
2015 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
2016 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2017 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2019 // If the RHS of the comparison is a 0.0, we don't need to do the
2020 // subtraction at all.
2021 if (isFloatingPointZero(RHS))
2023 default: break; // SETUO etc aren't handled by fsel.
2027 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2031 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2032 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2033 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2037 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2041 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2042 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2043 return DAG.getNode(PPCISD::FSEL, ResVT,
2044 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2049 default: break; // SETUO etc aren't handled by fsel.
2053 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2054 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2055 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2056 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2060 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2061 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2062 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2063 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2067 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2068 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2069 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2070 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2074 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2075 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2076 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2077 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2082 // FIXME: Split this code up when LegalizeDAGTypes lands.
2083 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2084 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2085 SDOperand Src = Op.getOperand(0);
2086 if (Src.getValueType() == MVT::f32)
2087 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2090 switch (Op.getValueType()) {
2091 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2093 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2096 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2100 // Convert the FP value to an int value through memory.
2101 SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
2103 // Emit a store to the stack slot.
2104 SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2106 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2108 if (Op.getValueType() == MVT::i32)
2109 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2110 DAG.getConstant(4, FIPtr.getValueType()));
2111 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2114 static SDOperand LowerFP_ROUND_INREG(SDOperand Op, SelectionDAG &DAG) {
2115 assert(Op.getValueType() == MVT::ppcf128);
2116 SDNode *Node = Op.Val;
2117 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2118 assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
2119 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2120 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2122 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2123 // of the long double, and puts FPSCR back the way it was. We do not
2124 // actually model FPSCR.
2125 std::vector<MVT::ValueType> NodeTys;
2126 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2128 NodeTys.push_back(MVT::f64); // Return register
2129 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2130 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2131 MFFSreg = Result.getValue(0);
2132 InFlag = Result.getValue(1);
2135 NodeTys.push_back(MVT::Flag); // Returns a flag
2136 Ops[0] = DAG.getConstant(31, MVT::i32);
2138 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2139 InFlag = Result.getValue(0);
2142 NodeTys.push_back(MVT::Flag); // Returns a flag
2143 Ops[0] = DAG.getConstant(30, MVT::i32);
2145 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2146 InFlag = Result.getValue(0);
2149 NodeTys.push_back(MVT::f64); // result of add
2150 NodeTys.push_back(MVT::Flag); // Returns a flag
2154 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2155 FPreg = Result.getValue(0);
2156 InFlag = Result.getValue(1);
2159 NodeTys.push_back(MVT::f64);
2160 Ops[0] = DAG.getConstant(1, MVT::i32);
2164 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2165 FPreg = Result.getValue(0);
2167 // We know the low half is about to be thrown away, so just use something
2169 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2172 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2173 if (Op.getOperand(0).getValueType() == MVT::i64) {
2174 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2175 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2176 if (Op.getValueType() == MVT::f32)
2177 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2181 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2182 "Unhandled SINT_TO_FP type in custom expander!");
2183 // Since we only generate this in 64-bit mode, we can take advantage of
2184 // 64-bit registers. In particular, sign extend the input value into the
2185 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2186 // then lfd it and fcfid it.
2187 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2188 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2189 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2190 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2192 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2195 // STD the extended value into the stack slot.
2196 MemOperand MO(PseudoSourceValue::getFixedStack(),
2197 MemOperand::MOStore, FrameIdx, 8, 8);
2198 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2199 DAG.getEntryNode(), Ext64, FIdx,
2200 DAG.getMemOperand(MO));
2201 // Load the value as a double.
2202 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2204 // FCFID it and return it.
2205 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2206 if (Op.getValueType() == MVT::f32)
2207 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2211 static SDOperand LowerFLT_ROUNDS_(SDOperand Op, SelectionDAG &DAG) {
2213 The rounding mode is in bits 30:31 of FPSR, and has the following
2220 FLT_ROUNDS, on the other hand, expects the following:
2227 To perform the conversion, we do:
2228 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2231 MachineFunction &MF = DAG.getMachineFunction();
2232 MVT::ValueType VT = Op.getValueType();
2233 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2234 std::vector<MVT::ValueType> NodeTys;
2235 SDOperand MFFSreg, InFlag;
2237 // Save FP Control Word to register
2238 NodeTys.push_back(MVT::f64); // return register
2239 NodeTys.push_back(MVT::Flag); // unused in this context
2240 SDOperand Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2242 // Save FP register to stack slot
2243 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
2244 SDOperand StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
2245 SDOperand Store = DAG.getStore(DAG.getEntryNode(), Chain,
2246 StackSlot, NULL, 0);
2248 // Load FP Control Word from low 32 bits of stack slot.
2249 SDOperand Four = DAG.getConstant(4, PtrVT);
2250 SDOperand Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
2251 SDOperand CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
2253 // Transform as necessary
2255 DAG.getNode(ISD::AND, MVT::i32,
2256 CWD, DAG.getConstant(3, MVT::i32));
2258 DAG.getNode(ISD::SRL, MVT::i32,
2259 DAG.getNode(ISD::AND, MVT::i32,
2260 DAG.getNode(ISD::XOR, MVT::i32,
2261 CWD, DAG.getConstant(3, MVT::i32)),
2262 DAG.getConstant(3, MVT::i32)),
2263 DAG.getConstant(1, MVT::i8));
2266 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
2268 return DAG.getNode((MVT::getSizeInBits(VT) < 16 ?
2269 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
2272 static SDOperand LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2273 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2274 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
2276 // Expand into a bunch of logical ops. Note that these ops
2277 // depend on the PPC behavior for oversized shift amounts.
2278 SDOperand Lo = Op.getOperand(0);
2279 SDOperand Hi = Op.getOperand(1);
2280 SDOperand Amt = Op.getOperand(2);
2282 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2283 DAG.getConstant(32, MVT::i32), Amt);
2284 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
2285 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
2286 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2287 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2288 DAG.getConstant(-32U, MVT::i32));
2289 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
2290 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2291 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
2292 SDOperand OutOps[] = { OutLo, OutHi };
2293 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2297 static SDOperand LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2298 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2299 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRL!");
2301 // Otherwise, expand into a bunch of logical ops. Note that these ops
2302 // depend on the PPC behavior for oversized shift amounts.
2303 SDOperand Lo = Op.getOperand(0);
2304 SDOperand Hi = Op.getOperand(1);
2305 SDOperand Amt = Op.getOperand(2);
2307 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2308 DAG.getConstant(32, MVT::i32), Amt);
2309 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2310 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2311 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2312 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2313 DAG.getConstant(-32U, MVT::i32));
2314 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
2315 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2316 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
2317 SDOperand OutOps[] = { OutLo, OutHi };
2318 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2322 static SDOperand LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2323 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2324 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
2326 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
2327 SDOperand Lo = Op.getOperand(0);
2328 SDOperand Hi = Op.getOperand(1);
2329 SDOperand Amt = Op.getOperand(2);
2331 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2332 DAG.getConstant(32, MVT::i32), Amt);
2333 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2334 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2335 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2336 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2337 DAG.getConstant(-32U, MVT::i32));
2338 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
2339 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
2340 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
2341 Tmp4, Tmp6, ISD::SETLE);
2342 SDOperand OutOps[] = { OutLo, OutHi };
2343 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2347 //===----------------------------------------------------------------------===//
2348 // Vector related lowering.
2351 // If this is a vector of constants or undefs, get the bits. A bit in
2352 // UndefBits is set if the corresponding element of the vector is an
2353 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2354 // zero. Return true if this is not an array of constants, false if it is.
2356 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2357 uint64_t UndefBits[2]) {
2358 // Start with zero'd results.
2359 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2361 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2362 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2363 SDOperand OpVal = BV->getOperand(i);
2365 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2366 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2368 uint64_t EltBits = 0;
2369 if (OpVal.getOpcode() == ISD::UNDEF) {
2370 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2371 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2373 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2374 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2375 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2376 assert(CN->getValueType(0) == MVT::f32 &&
2377 "Only one legal FP vector type!");
2378 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2380 // Nonconstant element.
2384 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2387 //printf("%llx %llx %llx %llx\n",
2388 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2392 // If this is a splat (repetition) of a value across the whole vector, return
2393 // the smallest size that splats it. For example, "0x01010101010101..." is a
2394 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2395 // SplatSize = 1 byte.
2396 static bool isConstantSplat(const uint64_t Bits128[2],
2397 const uint64_t Undef128[2],
2398 unsigned &SplatBits, unsigned &SplatUndef,
2399 unsigned &SplatSize) {
2401 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2402 // the same as the lower 64-bits, ignoring undefs.
2403 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2404 return false; // Can't be a splat if two pieces don't match.
2406 uint64_t Bits64 = Bits128[0] | Bits128[1];
2407 uint64_t Undef64 = Undef128[0] & Undef128[1];
2409 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2411 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2412 return false; // Can't be a splat if two pieces don't match.
2414 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2415 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2417 // If the top 16-bits are different than the lower 16-bits, ignoring
2418 // undefs, we have an i32 splat.
2419 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2421 SplatUndef = Undef32;
2426 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2427 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2429 // If the top 8-bits are different than the lower 8-bits, ignoring
2430 // undefs, we have an i16 splat.
2431 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2433 SplatUndef = Undef16;
2438 // Otherwise, we have an 8-bit splat.
2439 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2440 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2445 /// BuildSplatI - Build a canonical splati of Val with an element size of
2446 /// SplatSize. Cast the result to VT.
2447 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2448 SelectionDAG &DAG) {
2449 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2451 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2452 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2455 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2457 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2461 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2463 // Build a canonical splat for this value.
2464 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2465 SmallVector<SDOperand, 8> Ops;
2466 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2467 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2468 &Ops[0], Ops.size());
2469 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2472 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2473 /// specified intrinsic ID.
2474 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2476 MVT::ValueType DestVT = MVT::Other) {
2477 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2478 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2479 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2482 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2483 /// specified intrinsic ID.
2484 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2485 SDOperand Op2, SelectionDAG &DAG,
2486 MVT::ValueType DestVT = MVT::Other) {
2487 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2488 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2489 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2493 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2494 /// amount. The result has the specified value type.
2495 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2496 MVT::ValueType VT, SelectionDAG &DAG) {
2497 // Force LHS/RHS to be the right type.
2498 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2499 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2502 for (unsigned i = 0; i != 16; ++i)
2503 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2504 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2505 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2506 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2509 // If this is a case we can't handle, return null and let the default
2510 // expansion code take care of it. If we CAN select this case, and if it
2511 // selects to a single instruction, return Op. Otherwise, if we can codegen
2512 // this case more efficiently than a constant pool load, lower it to the
2513 // sequence of ops that should be used.
2514 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2515 // If this is a vector of constants or undefs, get the bits. A bit in
2516 // UndefBits is set if the corresponding element of the vector is an
2517 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2519 uint64_t VectorBits[2];
2520 uint64_t UndefBits[2];
2521 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2522 return SDOperand(); // Not a constant vector.
2524 // If this is a splat (repetition) of a value across the whole vector, return
2525 // the smallest size that splats it. For example, "0x01010101010101..." is a
2526 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2527 // SplatSize = 1 byte.
2528 unsigned SplatBits, SplatUndef, SplatSize;
2529 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2530 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2532 // First, handle single instruction cases.
2535 if (SplatBits == 0) {
2536 // Canonicalize all zero vectors to be v4i32.
2537 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2538 SDOperand Z = DAG.getConstant(0, MVT::i32);
2539 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2540 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2545 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2546 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2547 if (SextVal >= -16 && SextVal <= 15)
2548 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2551 // Two instruction sequences.
2553 // If this value is in the range [-32,30] and is even, use:
2554 // tmp = VSPLTI[bhw], result = add tmp, tmp
2555 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2556 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2557 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2560 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2561 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2563 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2564 // Make -1 and vspltisw -1:
2565 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2567 // Make the VSLW intrinsic, computing 0x8000_0000.
2568 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2571 // xor by OnesV to invert it.
2572 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2573 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2576 // Check to see if this is a wide variety of vsplti*, binop self cases.
2577 unsigned SplatBitSize = SplatSize*8;
2578 static const signed char SplatCsts[] = {
2579 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2580 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2583 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2584 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2585 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2586 int i = SplatCsts[idx];
2588 // Figure out what shift amount will be used by altivec if shifted by i in
2590 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2592 // vsplti + shl self.
2593 if (SextVal == (i << (int)TypeShiftAmt)) {
2594 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2595 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2596 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2597 Intrinsic::ppc_altivec_vslw
2599 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2600 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2603 // vsplti + srl self.
2604 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2605 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2606 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2607 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2608 Intrinsic::ppc_altivec_vsrw
2610 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2611 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2614 // vsplti + sra self.
2615 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2616 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2617 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2618 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2619 Intrinsic::ppc_altivec_vsraw
2621 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2622 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2625 // vsplti + rol self.
2626 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2627 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2628 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2629 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2630 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2631 Intrinsic::ppc_altivec_vrlw
2633 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2634 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2637 // t = vsplti c, result = vsldoi t, t, 1
2638 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2639 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2640 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2642 // t = vsplti c, result = vsldoi t, t, 2
2643 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2644 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2645 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2647 // t = vsplti c, result = vsldoi t, t, 3
2648 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2649 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2650 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2654 // Three instruction sequences.
2656 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2657 if (SextVal >= 0 && SextVal <= 31) {
2658 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2659 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2660 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
2661 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2663 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2664 if (SextVal >= -31 && SextVal <= 0) {
2665 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2666 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2667 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
2668 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2675 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2676 /// the specified operations to build the shuffle.
2677 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2678 SDOperand RHS, SelectionDAG &DAG) {
2679 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2680 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2681 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2684 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2696 if (OpNum == OP_COPY) {
2697 if (LHSID == (1*9+2)*9+3) return LHS;
2698 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2702 SDOperand OpLHS, OpRHS;
2703 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2704 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2706 unsigned ShufIdxs[16];
2708 default: assert(0 && "Unknown i32 permute!");
2710 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2711 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2712 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2713 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2716 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2717 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2718 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2719 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2722 for (unsigned i = 0; i != 16; ++i)
2723 ShufIdxs[i] = (i&3)+0;
2726 for (unsigned i = 0; i != 16; ++i)
2727 ShufIdxs[i] = (i&3)+4;
2730 for (unsigned i = 0; i != 16; ++i)
2731 ShufIdxs[i] = (i&3)+8;
2734 for (unsigned i = 0; i != 16; ++i)
2735 ShufIdxs[i] = (i&3)+12;
2738 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2740 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2742 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2745 for (unsigned i = 0; i != 16; ++i)
2746 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2748 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2749 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2752 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2753 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2754 /// return the code it can be lowered into. Worst case, it can always be
2755 /// lowered into a vperm.
2756 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2757 SDOperand V1 = Op.getOperand(0);
2758 SDOperand V2 = Op.getOperand(1);
2759 SDOperand PermMask = Op.getOperand(2);
2761 // Cases that are handled by instructions that take permute immediates
2762 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2763 // selected by the instruction selector.
2764 if (V2.getOpcode() == ISD::UNDEF) {
2765 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2766 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2767 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2768 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2769 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2770 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2771 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2772 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2773 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2774 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2775 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2776 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2781 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2782 // and produce a fixed permutation. If any of these match, do not lower to
2784 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2785 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2786 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2787 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2788 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2789 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2790 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2791 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2792 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2795 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2796 // perfect shuffle table to emit an optimal matching sequence.
2797 unsigned PFIndexes[4];
2798 bool isFourElementShuffle = true;
2799 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2800 unsigned EltNo = 8; // Start out undef.
2801 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2802 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2803 continue; // Undef, ignore it.
2805 unsigned ByteSource =
2806 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2807 if ((ByteSource & 3) != j) {
2808 isFourElementShuffle = false;
2813 EltNo = ByteSource/4;
2814 } else if (EltNo != ByteSource/4) {
2815 isFourElementShuffle = false;
2819 PFIndexes[i] = EltNo;
2822 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
2823 // perfect shuffle vector to determine if it is cost effective to do this as
2824 // discrete instructions, or whether we should use a vperm.
2825 if (isFourElementShuffle) {
2826 // Compute the index in the perfect shuffle table.
2827 unsigned PFTableIndex =
2828 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2830 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2831 unsigned Cost = (PFEntry >> 30);
2833 // Determining when to avoid vperm is tricky. Many things affect the cost
2834 // of vperm, particularly how many times the perm mask needs to be computed.
2835 // For example, if the perm mask can be hoisted out of a loop or is already
2836 // used (perhaps because there are multiple permutes with the same shuffle
2837 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
2838 // the loop requires an extra register.
2840 // As a compromise, we only emit discrete instructions if the shuffle can be
2841 // generated in 3 or fewer operations. When we have loop information
2842 // available, if this block is within a loop, we should avoid using vperm
2843 // for 3-operation perms and use a constant pool load instead.
2845 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
2848 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
2849 // vector that will get spilled to the constant pool.
2850 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
2852 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
2853 // that it is in input element units, not in bytes. Convert now.
2854 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
2855 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
2857 SmallVector<SDOperand, 16> ResultMask;
2858 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
2860 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
2863 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
2865 for (unsigned j = 0; j != BytesPerElement; ++j)
2866 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
2870 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2871 &ResultMask[0], ResultMask.size());
2872 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
2875 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
2876 /// altivec comparison. If it is, return true and fill in Opc/isDot with
2877 /// information about the intrinsic.
2878 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
2880 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
2883 switch (IntrinsicID) {
2884 default: return false;
2885 // Comparison predicates.
2886 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
2887 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
2888 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
2889 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
2890 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
2891 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
2892 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
2893 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
2894 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
2895 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
2896 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
2897 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
2898 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
2900 // Normal Comparisons.
2901 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
2902 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
2903 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
2904 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
2905 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
2906 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
2907 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
2908 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
2909 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
2910 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
2911 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
2912 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
2913 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
2918 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
2919 /// lower, do it, otherwise return null.
2920 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
2921 // If this is a lowered altivec predicate compare, CompareOpc is set to the
2922 // opcode number of the comparison.
2925 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
2926 return SDOperand(); // Don't custom lower most intrinsics.
2928 // If this is a non-dot comparison, make the VCMP node and we are done.
2930 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
2931 Op.getOperand(1), Op.getOperand(2),
2932 DAG.getConstant(CompareOpc, MVT::i32));
2933 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
2936 // Create the PPCISD altivec 'dot' comparison node.
2938 Op.getOperand(2), // LHS
2939 Op.getOperand(3), // RHS
2940 DAG.getConstant(CompareOpc, MVT::i32)
2942 std::vector<MVT::ValueType> VTs;
2943 VTs.push_back(Op.getOperand(2).getValueType());
2944 VTs.push_back(MVT::Flag);
2945 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
2947 // Now that we have the comparison, emit a copy from the CR to a GPR.
2948 // This is flagged to the above dot comparison.
2949 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
2950 DAG.getRegister(PPC::CR6, MVT::i32),
2951 CompNode.getValue(1));
2953 // Unpack the result based on how the target uses it.
2954 unsigned BitNo; // Bit # of CR6.
2955 bool InvertBit; // Invert result?
2956 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
2957 default: // Can't happen, don't crash on invalid number though.
2958 case 0: // Return the value of the EQ bit of CR6.
2959 BitNo = 0; InvertBit = false;
2961 case 1: // Return the inverted value of the EQ bit of CR6.
2962 BitNo = 0; InvertBit = true;
2964 case 2: // Return the value of the LT bit of CR6.
2965 BitNo = 2; InvertBit = false;
2967 case 3: // Return the inverted value of the LT bit of CR6.
2968 BitNo = 2; InvertBit = true;
2972 // Shift the bit into the low position.
2973 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
2974 DAG.getConstant(8-(3-BitNo), MVT::i32));
2976 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
2977 DAG.getConstant(1, MVT::i32));
2979 // If we are supposed to, toggle the bit.
2981 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
2982 DAG.getConstant(1, MVT::i32));
2986 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2987 // Create a stack slot that is 16-byte aligned.
2988 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2989 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
2990 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2991 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2993 // Store the input value into Value#0 of the stack slot.
2994 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
2995 Op.getOperand(0), FIdx, NULL, 0);
2997 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3000 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
3001 if (Op.getValueType() == MVT::v4i32) {
3002 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3004 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3005 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3007 SDOperand RHSSwap = // = vrlw RHS, 16
3008 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3010 // Shrinkify inputs to v8i16.
3011 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3012 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3013 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3015 // Low parts multiplied together, generating 32-bit results (we ignore the
3017 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3018 LHS, RHS, DAG, MVT::v4i32);
3020 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3021 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3022 // Shift the high parts up 16 bits.
3023 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3024 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3025 } else if (Op.getValueType() == MVT::v8i16) {
3026 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3028 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3030 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3031 LHS, RHS, Zero, DAG);
3032 } else if (Op.getValueType() == MVT::v16i8) {
3033 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3035 // Multiply the even 8-bit parts, producing 16-bit sums.
3036 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3037 LHS, RHS, DAG, MVT::v8i16);
3038 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3040 // Multiply the odd 8-bit parts, producing 16-bit sums.
3041 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3042 LHS, RHS, DAG, MVT::v8i16);
3043 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3045 // Merge the results together.
3047 for (unsigned i = 0; i != 8; ++i) {
3048 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3049 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3051 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3052 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3054 assert(0 && "Unknown mul to lower!");
3059 /// LowerOperation - Provide custom lowering hooks for some operations.
3061 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3062 switch (Op.getOpcode()) {
3063 default: assert(0 && "Wasn't expecting to be able to lower this!");
3064 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3065 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3066 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3067 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3068 case ISD::SETCC: return LowerSETCC(Op, DAG);
3070 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3071 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3074 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3075 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3077 case ISD::FORMAL_ARGUMENTS:
3078 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3079 VarArgsStackOffset, VarArgsNumGPR,
3080 VarArgsNumFPR, PPCSubTarget);
3082 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
3083 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3084 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3085 case ISD::DYNAMIC_STACKALLOC:
3086 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3088 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3089 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3090 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3091 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3092 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3094 // Lower 64-bit shifts.
3095 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3096 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3097 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3099 // Vector-related lowering.
3100 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3101 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3102 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3103 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3104 case ISD::MUL: return LowerMUL(Op, DAG);
3106 // Frame & Return address.
3107 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3108 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3113 SDNode *PPCTargetLowering::ExpandOperationResult(SDNode *N, SelectionDAG &DAG) {
3114 switch (N->getOpcode()) {
3115 default: assert(0 && "Wasn't expecting to be able to lower this!");
3116 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(SDOperand(N, 0), DAG).Val;
3121 //===----------------------------------------------------------------------===//
3122 // Other Lowering Code
3123 //===----------------------------------------------------------------------===//
3126 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
3127 MachineBasicBlock *BB) {
3128 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3129 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3130 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3131 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3132 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3133 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3134 "Unexpected instr type to insert");
3136 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3137 // control-flow pattern. The incoming instruction knows the destination vreg
3138 // to set, the condition code register to branch on, the true/false values to
3139 // select between, and a branch opcode to use.
3140 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3141 ilist<MachineBasicBlock>::iterator It = BB;
3147 // cmpTY ccX, r1, r2
3149 // fallthrough --> copy0MBB
3150 MachineBasicBlock *thisMBB = BB;
3151 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3152 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3153 unsigned SelectPred = MI->getOperand(4).getImm();
3154 BuildMI(BB, TII->get(PPC::BCC))
3155 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3156 MachineFunction *F = BB->getParent();
3157 F->getBasicBlockList().insert(It, copy0MBB);
3158 F->getBasicBlockList().insert(It, sinkMBB);
3159 // Update machine-CFG edges by first adding all successors of the current
3160 // block to the new block which will contain the Phi node for the select.
3161 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3162 e = BB->succ_end(); i != e; ++i)
3163 sinkMBB->addSuccessor(*i);
3164 // Next, remove all successors of the current block, and add the true
3165 // and fallthrough blocks as its successors.
3166 while(!BB->succ_empty())
3167 BB->removeSuccessor(BB->succ_begin());
3168 BB->addSuccessor(copy0MBB);
3169 BB->addSuccessor(sinkMBB);
3172 // %FalseValue = ...
3173 // # fallthrough to sinkMBB
3176 // Update machine-CFG edges
3177 BB->addSuccessor(sinkMBB);
3180 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3183 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3184 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3185 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3187 delete MI; // The pseudo instruction is gone now.
3191 //===----------------------------------------------------------------------===//
3192 // Target Optimization Hooks
3193 //===----------------------------------------------------------------------===//
3195 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3196 DAGCombinerInfo &DCI) const {
3197 TargetMachine &TM = getTargetMachine();
3198 SelectionDAG &DAG = DCI.DAG;
3199 switch (N->getOpcode()) {
3202 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3203 if (C->getValue() == 0) // 0 << V -> 0.
3204 return N->getOperand(0);
3208 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3209 if (C->getValue() == 0) // 0 >>u V -> 0.
3210 return N->getOperand(0);
3214 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3215 if (C->getValue() == 0 || // 0 >>s V -> 0.
3216 C->isAllOnesValue()) // -1 >>s V -> -1.
3217 return N->getOperand(0);
3221 case ISD::SINT_TO_FP:
3222 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3223 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3224 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3225 // We allow the src/dst to be either f32/f64, but the intermediate
3226 // type must be i64.
3227 if (N->getOperand(0).getValueType() == MVT::i64 &&
3228 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
3229 SDOperand Val = N->getOperand(0).getOperand(0);
3230 if (Val.getValueType() == MVT::f32) {
3231 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3232 DCI.AddToWorklist(Val.Val);
3235 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3236 DCI.AddToWorklist(Val.Val);
3237 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3238 DCI.AddToWorklist(Val.Val);
3239 if (N->getValueType(0) == MVT::f32) {
3240 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
3241 DAG.getIntPtrConstant(0));
3242 DCI.AddToWorklist(Val.Val);
3245 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3246 // If the intermediate type is i32, we can avoid the load/store here
3253 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3254 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3255 !cast<StoreSDNode>(N)->isTruncatingStore() &&
3256 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3257 N->getOperand(1).getValueType() == MVT::i32 &&
3258 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
3259 SDOperand Val = N->getOperand(1).getOperand(0);
3260 if (Val.getValueType() == MVT::f32) {
3261 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3262 DCI.AddToWorklist(Val.Val);
3264 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3265 DCI.AddToWorklist(Val.Val);
3267 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3268 N->getOperand(2), N->getOperand(3));
3269 DCI.AddToWorklist(Val.Val);
3273 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3274 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3275 N->getOperand(1).Val->hasOneUse() &&
3276 (N->getOperand(1).getValueType() == MVT::i32 ||
3277 N->getOperand(1).getValueType() == MVT::i16)) {
3278 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3279 // Do an any-extend to 32-bits if this is a half-word input.
3280 if (BSwapOp.getValueType() == MVT::i16)
3281 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3283 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3284 N->getOperand(2), N->getOperand(3),
3285 DAG.getValueType(N->getOperand(1).getValueType()));
3289 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3290 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3291 N->getOperand(0).hasOneUse() &&
3292 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3293 SDOperand Load = N->getOperand(0);
3294 LoadSDNode *LD = cast<LoadSDNode>(Load);
3295 // Create the byte-swapping load.
3296 std::vector<MVT::ValueType> VTs;
3297 VTs.push_back(MVT::i32);
3298 VTs.push_back(MVT::Other);
3299 SDOperand MO = DAG.getMemOperand(LD->getMemOperand());
3301 LD->getChain(), // Chain
3302 LD->getBasePtr(), // Ptr
3304 DAG.getValueType(N->getValueType(0)) // VT
3306 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3308 // If this is an i16 load, insert the truncate.
3309 SDOperand ResVal = BSLoad;
3310 if (N->getValueType(0) == MVT::i16)
3311 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3313 // First, combine the bswap away. This makes the value produced by the
3315 DCI.CombineTo(N, ResVal);
3317 // Next, combine the load away, we give it a bogus result value but a real
3318 // chain result. The result value is dead because the bswap is dead.
3319 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3321 // Return N so it doesn't get rechecked!
3322 return SDOperand(N, 0);
3326 case PPCISD::VCMP: {
3327 // If a VCMPo node already exists with exactly the same operands as this
3328 // node, use its result instead of this node (VCMPo computes both a CR6 and
3329 // a normal output).
3331 if (!N->getOperand(0).hasOneUse() &&
3332 !N->getOperand(1).hasOneUse() &&
3333 !N->getOperand(2).hasOneUse()) {
3335 // Scan all of the users of the LHS, looking for VCMPo's that match.
3336 SDNode *VCMPoNode = 0;
3338 SDNode *LHSN = N->getOperand(0).Val;
3339 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3341 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3342 (*UI)->getOperand(1) == N->getOperand(1) &&
3343 (*UI)->getOperand(2) == N->getOperand(2) &&
3344 (*UI)->getOperand(0) == N->getOperand(0)) {
3349 // If there is no VCMPo node, or if the flag value has a single use, don't
3351 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3354 // Look at the (necessarily single) use of the flag value. If it has a
3355 // chain, this transformation is more complex. Note that multiple things
3356 // could use the value result, which we should ignore.
3357 SDNode *FlagUser = 0;
3358 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3359 FlagUser == 0; ++UI) {
3360 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3362 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3363 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3370 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3371 // give up for right now.
3372 if (FlagUser->getOpcode() == PPCISD::MFCR)
3373 return SDOperand(VCMPoNode, 0);
3378 // If this is a branch on an altivec predicate comparison, lower this so
3379 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3380 // lowering is done pre-legalize, because the legalizer lowers the predicate
3381 // compare down to code that is difficult to reassemble.
3382 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3383 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3387 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3388 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3389 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3390 assert(isDot && "Can't compare against a vector result!");
3392 // If this is a comparison against something other than 0/1, then we know
3393 // that the condition is never/always true.
3394 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3395 if (Val != 0 && Val != 1) {
3396 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3397 return N->getOperand(0);
3398 // Always !=, turn it into an unconditional branch.
3399 return DAG.getNode(ISD::BR, MVT::Other,
3400 N->getOperand(0), N->getOperand(4));
3403 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3405 // Create the PPCISD altivec 'dot' comparison node.
3406 std::vector<MVT::ValueType> VTs;
3408 LHS.getOperand(2), // LHS of compare
3409 LHS.getOperand(3), // RHS of compare
3410 DAG.getConstant(CompareOpc, MVT::i32)
3412 VTs.push_back(LHS.getOperand(2).getValueType());
3413 VTs.push_back(MVT::Flag);
3414 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3416 // Unpack the result based on how the target uses it.
3417 PPC::Predicate CompOpc;
3418 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3419 default: // Can't happen, don't crash on invalid number though.
3420 case 0: // Branch on the value of the EQ bit of CR6.
3421 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3423 case 1: // Branch on the inverted value of the EQ bit of CR6.
3424 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3426 case 2: // Branch on the value of the LT bit of CR6.
3427 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3429 case 3: // Branch on the inverted value of the LT bit of CR6.
3430 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3434 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3435 DAG.getConstant(CompOpc, MVT::i32),
3436 DAG.getRegister(PPC::CR6, MVT::i32),
3437 N->getOperand(4), CompNode.getValue(1));
3446 //===----------------------------------------------------------------------===//
3447 // Inline Assembly Support
3448 //===----------------------------------------------------------------------===//
3450 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3454 const SelectionDAG &DAG,
3455 unsigned Depth) const {
3456 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3457 switch (Op.getOpcode()) {
3459 case PPCISD::LBRX: {
3460 // lhbrx is known to have the top bits cleared out.
3461 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3462 KnownZero = 0xFFFF0000;
3465 case ISD::INTRINSIC_WO_CHAIN: {
3466 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3468 case Intrinsic::ppc_altivec_vcmpbfp_p:
3469 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3470 case Intrinsic::ppc_altivec_vcmpequb_p:
3471 case Intrinsic::ppc_altivec_vcmpequh_p:
3472 case Intrinsic::ppc_altivec_vcmpequw_p:
3473 case Intrinsic::ppc_altivec_vcmpgefp_p:
3474 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3475 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3476 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3477 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3478 case Intrinsic::ppc_altivec_vcmpgtub_p:
3479 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3480 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3481 KnownZero = ~1U; // All bits but the low one are known to be zero.
3489 /// getConstraintType - Given a constraint, return the type of
3490 /// constraint it is for this target.
3491 PPCTargetLowering::ConstraintType
3492 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3493 if (Constraint.size() == 1) {
3494 switch (Constraint[0]) {
3501 return C_RegisterClass;
3504 return TargetLowering::getConstraintType(Constraint);
3507 std::pair<unsigned, const TargetRegisterClass*>
3508 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3509 MVT::ValueType VT) const {
3510 if (Constraint.size() == 1) {
3511 // GCC RS6000 Constraint Letters
3512 switch (Constraint[0]) {
3515 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3516 return std::make_pair(0U, PPC::G8RCRegisterClass);
3517 return std::make_pair(0U, PPC::GPRCRegisterClass);
3520 return std::make_pair(0U, PPC::F4RCRegisterClass);
3521 else if (VT == MVT::f64)
3522 return std::make_pair(0U, PPC::F8RCRegisterClass);
3525 return std::make_pair(0U, PPC::VRRCRegisterClass);
3527 return std::make_pair(0U, PPC::CRRCRegisterClass);
3531 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3535 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3536 /// vector. If it is invalid, don't add anything to Ops.
3537 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3538 std::vector<SDOperand>&Ops,
3539 SelectionDAG &DAG) {
3540 SDOperand Result(0,0);
3551 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3552 if (!CST) return; // Must be an immediate to match.
3553 unsigned Value = CST->getValue();
3555 default: assert(0 && "Unknown constraint letter!");
3556 case 'I': // "I" is a signed 16-bit constant.
3557 if ((short)Value == (int)Value)
3558 Result = DAG.getTargetConstant(Value, Op.getValueType());
3560 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3561 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3562 if ((short)Value == 0)
3563 Result = DAG.getTargetConstant(Value, Op.getValueType());
3565 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3566 if ((Value >> 16) == 0)
3567 Result = DAG.getTargetConstant(Value, Op.getValueType());
3569 case 'M': // "M" is a constant that is greater than 31.
3571 Result = DAG.getTargetConstant(Value, Op.getValueType());
3573 case 'N': // "N" is a positive constant that is an exact power of two.
3574 if ((int)Value > 0 && isPowerOf2_32(Value))
3575 Result = DAG.getTargetConstant(Value, Op.getValueType());
3577 case 'O': // "O" is the constant zero.
3579 Result = DAG.getTargetConstant(Value, Op.getValueType());
3581 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3582 if ((short)-Value == (int)-Value)
3583 Result = DAG.getTargetConstant(Value, Op.getValueType());
3591 Ops.push_back(Result);
3595 // Handle standard constraint letters.
3596 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3599 // isLegalAddressingMode - Return true if the addressing mode represented
3600 // by AM is legal for this target, for a load/store of the specified type.
3601 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3602 const Type *Ty) const {
3603 // FIXME: PPC does not allow r+i addressing modes for vectors!
3605 // PPC allows a sign-extended 16-bit immediate field.
3606 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3609 // No global is ever allowed as a base.
3613 // PPC only support r+r,
3615 case 0: // "r+i" or just "i", depending on HasBaseReg.
3618 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3620 // Otherwise we have r+r or r+i.
3623 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3625 // Allow 2*r as r+r.
3628 // No other scales are supported.
3635 /// isLegalAddressImmediate - Return true if the integer value can be used
3636 /// as the offset of the target addressing mode for load / store of the
3638 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3639 // PPC allows a sign-extended 16-bit immediate field.
3640 return (V > -(1 << 16) && V < (1 << 16)-1);
3643 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3647 SDOperand PPCTargetLowering::LowerRETURNADDR(SDOperand Op, SelectionDAG &DAG) {
3648 // Depths > 0 not supported yet!
3649 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3652 MachineFunction &MF = DAG.getMachineFunction();
3653 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
3654 int RAIdx = FuncInfo->getReturnAddrSaveIndex();
3656 bool isPPC64 = PPCSubTarget.isPPC64();
3658 PPCFrameInfo::getReturnSaveOffset(isPPC64, PPCSubTarget.isMachoABI());
3660 // Set up a frame object for the return address.
3661 RAIdx = MF.getFrameInfo()->CreateFixedObject(isPPC64 ? 8 : 4, Offset);
3663 // Remember it for next time.
3664 FuncInfo->setReturnAddrSaveIndex(RAIdx);
3666 // Make sure the function really does not optimize away the store of the RA
3668 FuncInfo->setLRStoreRequired();
3671 // Just load the return address off the stack.
3672 SDOperand RetAddrFI = DAG.getFrameIndex(RAIdx, getPointerTy());
3673 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
3676 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG) {
3677 // Depths > 0 not supported yet!
3678 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3681 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3682 bool isPPC64 = PtrVT == MVT::i64;
3684 MachineFunction &MF = DAG.getMachineFunction();
3685 MachineFrameInfo *MFI = MF.getFrameInfo();
3686 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3687 && MFI->getStackSize();
3690 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3693 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,