1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source 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/SelectionDAG.h"
27 #include "llvm/CodeGen/SSARegMap.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include "llvm/Support/CommandLine.h"
36 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
37 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
40 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
41 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
45 // Use _setjmp/_longjmp instead of setjmp/longjmp.
46 setUseUnderscoreSetJmp(true);
47 setUseUnderscoreLongJmp(true);
49 // Set up the register classes.
50 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
51 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
52 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
54 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
55 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
56 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Expand);
58 // PowerPC does not have truncstore for i1.
59 setStoreXAction(MVT::i1, Promote);
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 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
74 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
76 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
77 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
78 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
79 // This is used in the ppcf128->int sequence. Note it has different semantics
80 // from FP_ROUND: that rounds to nearest, this rounds to zero.
81 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
83 // PowerPC has no intrinsics for these particular operations
84 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
85 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
86 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
88 // PowerPC has no SREM/UREM instructions
89 setOperationAction(ISD::SREM, MVT::i32, Expand);
90 setOperationAction(ISD::UREM, MVT::i32, Expand);
91 setOperationAction(ISD::SREM, MVT::i64, Expand);
92 setOperationAction(ISD::UREM, MVT::i64, Expand);
94 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
95 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
96 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
97 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
98 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
99 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
100 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
101 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
102 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
104 // We don't support sin/cos/sqrt/fmod/pow
105 setOperationAction(ISD::FSIN , MVT::f64, Expand);
106 setOperationAction(ISD::FCOS , MVT::f64, Expand);
107 setOperationAction(ISD::FREM , MVT::f64, Expand);
108 setOperationAction(ISD::FPOW , MVT::f64, Expand);
109 setOperationAction(ISD::FSIN , MVT::f32, Expand);
110 setOperationAction(ISD::FCOS , MVT::f32, Expand);
111 setOperationAction(ISD::FREM , MVT::f32, Expand);
112 setOperationAction(ISD::FPOW , MVT::f32, Expand);
114 // If we're enabling GP optimizations, use hardware square root
115 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
116 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
117 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
120 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
121 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
123 // PowerPC does not have BSWAP, CTPOP or CTTZ
124 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
125 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
126 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
127 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
128 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
129 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
131 // PowerPC does not have ROTR
132 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
134 // PowerPC does not have Select
135 setOperationAction(ISD::SELECT, MVT::i32, Expand);
136 setOperationAction(ISD::SELECT, MVT::i64, Expand);
137 setOperationAction(ISD::SELECT, MVT::f32, Expand);
138 setOperationAction(ISD::SELECT, MVT::f64, Expand);
140 // PowerPC wants to turn select_cc of FP into fsel when possible.
141 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
142 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
144 // PowerPC wants to optimize integer setcc a bit
145 setOperationAction(ISD::SETCC, MVT::i32, Custom);
147 // PowerPC does not have BRCOND which requires SetCC
148 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
150 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
152 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
153 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
155 // PowerPC does not have [U|S]INT_TO_FP
156 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
157 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
159 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
160 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
161 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
162 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
164 // We cannot sextinreg(i1). Expand to shifts.
165 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
167 // Support label based line numbers.
168 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
169 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
170 if (!TM.getSubtarget<PPCSubtarget>().isDarwin()) {
171 setOperationAction(ISD::LABEL, MVT::Other, Expand);
173 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
174 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
175 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
176 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
179 // We want to legalize GlobalAddress and ConstantPool nodes into the
180 // appropriate instructions to materialize the address.
181 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
182 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
183 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
184 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
185 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
186 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
187 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
188 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
190 // RET must be custom lowered, to meet ABI requirements
191 setOperationAction(ISD::RET , MVT::Other, Custom);
193 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
194 setOperationAction(ISD::VASTART , MVT::Other, Custom);
196 // VAARG is custom lowered with ELF 32 ABI
197 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
198 setOperationAction(ISD::VAARG, MVT::Other, Custom);
200 setOperationAction(ISD::VAARG, MVT::Other, Expand);
202 // Use the default implementation.
203 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
204 setOperationAction(ISD::VAEND , MVT::Other, Expand);
205 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
206 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
207 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
208 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
210 // We want to custom lower some of our intrinsics.
211 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
213 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
214 // They also have instructions for converting between i64 and fp.
215 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
216 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
217 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
218 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
219 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
221 // FIXME: disable this lowered code. This generates 64-bit register values,
222 // and we don't model the fact that the top part is clobbered by calls. We
223 // need to flag these together so that the value isn't live across a call.
224 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
226 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
227 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
229 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
230 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
233 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
234 // 64 bit PowerPC implementations can support i64 types directly
235 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
236 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
237 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
239 // 32 bit PowerPC wants to expand i64 shifts itself.
240 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
241 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
242 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
245 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
246 // First set operation action for all vector types to expand. Then we
247 // will selectively turn on ones that can be effectively codegen'd.
248 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
249 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
250 // add/sub are legal for all supported vector VT's.
251 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
252 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
254 // We promote all shuffles to v16i8.
255 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
256 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
258 // We promote all non-typed operations to v4i32.
259 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
260 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
261 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
262 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
263 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
264 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
265 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
266 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
267 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
268 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
269 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
270 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
272 // No other operations are legal.
273 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
274 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
275 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
276 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
277 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
278 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
279 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
280 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
281 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
282 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
283 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
284 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
285 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
286 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
288 setOperationAction(ISD::SCALAR_TO_VECTOR, (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 computeRegisterProperties();
344 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
347 case PPCISD::FSEL: return "PPCISD::FSEL";
348 case PPCISD::FCFID: return "PPCISD::FCFID";
349 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
350 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
351 case PPCISD::STFIWX: return "PPCISD::STFIWX";
352 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
353 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
354 case PPCISD::VPERM: return "PPCISD::VPERM";
355 case PPCISD::Hi: return "PPCISD::Hi";
356 case PPCISD::Lo: return "PPCISD::Lo";
357 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
358 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
359 case PPCISD::SRL: return "PPCISD::SRL";
360 case PPCISD::SRA: return "PPCISD::SRA";
361 case PPCISD::SHL: return "PPCISD::SHL";
362 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
363 case PPCISD::STD_32: return "PPCISD::STD_32";
364 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
365 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
366 case PPCISD::MTCTR: return "PPCISD::MTCTR";
367 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
368 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
369 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
370 case PPCISD::MFCR: return "PPCISD::MFCR";
371 case PPCISD::VCMP: return "PPCISD::VCMP";
372 case PPCISD::VCMPo: return "PPCISD::VCMPo";
373 case PPCISD::LBRX: return "PPCISD::LBRX";
374 case PPCISD::STBRX: return "PPCISD::STBRX";
375 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
379 //===----------------------------------------------------------------------===//
380 // Node matching predicates, for use by the tblgen matching code.
381 //===----------------------------------------------------------------------===//
383 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
384 static bool isFloatingPointZero(SDOperand Op) {
385 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
386 return CFP->getValueAPF().isZero();
387 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
388 // Maybe this has already been legalized into the constant pool?
389 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
390 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
391 return CFP->getValueAPF().isZero();
396 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
397 /// true if Op is undef or if it matches the specified value.
398 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
399 return Op.getOpcode() == ISD::UNDEF ||
400 cast<ConstantSDNode>(Op)->getValue() == Val;
403 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
404 /// VPKUHUM instruction.
405 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
407 for (unsigned i = 0; i != 16; ++i)
408 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
411 for (unsigned i = 0; i != 8; ++i)
412 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
413 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
419 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
420 /// VPKUWUM instruction.
421 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
423 for (unsigned i = 0; i != 16; i += 2)
424 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
425 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
428 for (unsigned i = 0; i != 8; i += 2)
429 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
430 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
431 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
432 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
438 /// isVMerge - Common function, used to match vmrg* shuffles.
440 static bool isVMerge(SDNode *N, unsigned UnitSize,
441 unsigned LHSStart, unsigned RHSStart) {
442 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
443 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
444 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
445 "Unsupported merge size!");
447 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
448 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
449 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
450 LHSStart+j+i*UnitSize) ||
451 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
452 RHSStart+j+i*UnitSize))
458 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
459 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
460 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
462 return isVMerge(N, UnitSize, 8, 24);
463 return isVMerge(N, UnitSize, 8, 8);
466 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
467 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
468 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
470 return isVMerge(N, UnitSize, 0, 16);
471 return isVMerge(N, UnitSize, 0, 0);
475 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
476 /// amount, otherwise return -1.
477 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
478 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
479 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
480 // Find the first non-undef value in the shuffle mask.
482 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
485 if (i == 16) return -1; // all undef.
487 // Otherwise, check to see if the rest of the elements are consequtively
488 // numbered from this value.
489 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
490 if (ShiftAmt < i) return -1;
494 // Check the rest of the elements to see if they are consequtive.
495 for (++i; i != 16; ++i)
496 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
499 // Check the rest of the elements to see if they are consequtive.
500 for (++i; i != 16; ++i)
501 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
508 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
509 /// specifies a splat of a single element that is suitable for input to
510 /// VSPLTB/VSPLTH/VSPLTW.
511 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
512 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
513 N->getNumOperands() == 16 &&
514 (EltSize == 1 || EltSize == 2 || EltSize == 4));
516 // This is a splat operation if each element of the permute is the same, and
517 // if the value doesn't reference the second vector.
518 unsigned ElementBase = 0;
519 SDOperand Elt = N->getOperand(0);
520 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
521 ElementBase = EltV->getValue();
523 return false; // FIXME: Handle UNDEF elements too!
525 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
528 // Check that they are consequtive.
529 for (unsigned i = 1; i != EltSize; ++i) {
530 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
531 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
535 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
536 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
537 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
538 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
539 "Invalid VECTOR_SHUFFLE mask!");
540 for (unsigned j = 0; j != EltSize; ++j)
541 if (N->getOperand(i+j) != N->getOperand(j))
548 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
550 bool PPC::isAllNegativeZeroVector(SDNode *N) {
551 assert(N->getOpcode() == ISD::BUILD_VECTOR);
552 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
553 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
554 return CFP->getValueAPF().isNegZero();
558 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
559 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
560 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
561 assert(isSplatShuffleMask(N, EltSize));
562 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
565 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
566 /// by using a vspltis[bhw] instruction of the specified element size, return
567 /// the constant being splatted. The ByteSize field indicates the number of
568 /// bytes of each element [124] -> [bhw].
569 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
570 SDOperand OpVal(0, 0);
572 // If ByteSize of the splat is bigger than the element size of the
573 // build_vector, then we have a case where we are checking for a splat where
574 // multiple elements of the buildvector are folded together into a single
575 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
576 unsigned EltSize = 16/N->getNumOperands();
577 if (EltSize < ByteSize) {
578 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
579 SDOperand UniquedVals[4];
580 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
582 // See if all of the elements in the buildvector agree across.
583 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
584 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
585 // If the element isn't a constant, bail fully out.
586 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
589 if (UniquedVals[i&(Multiple-1)].Val == 0)
590 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
591 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
592 return SDOperand(); // no match.
595 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
596 // either constant or undef values that are identical for each chunk. See
597 // if these chunks can form into a larger vspltis*.
599 // Check to see if all of the leading entries are either 0 or -1. If
600 // neither, then this won't fit into the immediate field.
601 bool LeadingZero = true;
602 bool LeadingOnes = true;
603 for (unsigned i = 0; i != Multiple-1; ++i) {
604 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
606 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
607 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
609 // Finally, check the least significant entry.
611 if (UniquedVals[Multiple-1].Val == 0)
612 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
613 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
615 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
618 if (UniquedVals[Multiple-1].Val == 0)
619 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
620 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
621 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
622 return DAG.getTargetConstant(Val, MVT::i32);
628 // Check to see if this buildvec has a single non-undef value in its elements.
629 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
630 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
632 OpVal = N->getOperand(i);
633 else if (OpVal != N->getOperand(i))
637 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
639 unsigned ValSizeInBytes = 0;
641 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
642 Value = CN->getValue();
643 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
644 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
645 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
646 Value = FloatToBits(CN->getValueAPF().convertToFloat());
650 // If the splat value is larger than the element value, then we can never do
651 // this splat. The only case that we could fit the replicated bits into our
652 // immediate field for would be zero, and we prefer to use vxor for it.
653 if (ValSizeInBytes < ByteSize) return SDOperand();
655 // If the element value is larger than the splat value, cut it in half and
656 // check to see if the two halves are equal. Continue doing this until we
657 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
658 while (ValSizeInBytes > ByteSize) {
659 ValSizeInBytes >>= 1;
661 // If the top half equals the bottom half, we're still ok.
662 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
663 (Value & ((1 << (8*ValSizeInBytes))-1)))
667 // Properly sign extend the value.
668 int ShAmt = (4-ByteSize)*8;
669 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
671 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
672 if (MaskVal == 0) return SDOperand();
674 // Finally, if this value fits in a 5 bit sext field, return it
675 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
676 return DAG.getTargetConstant(MaskVal, MVT::i32);
680 //===----------------------------------------------------------------------===//
681 // Addressing Mode Selection
682 //===----------------------------------------------------------------------===//
684 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
685 /// or 64-bit immediate, and if the value can be accurately represented as a
686 /// sign extension from a 16-bit value. If so, this returns true and the
688 static bool isIntS16Immediate(SDNode *N, short &Imm) {
689 if (N->getOpcode() != ISD::Constant)
692 Imm = (short)cast<ConstantSDNode>(N)->getValue();
693 if (N->getValueType(0) == MVT::i32)
694 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
696 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
698 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
699 return isIntS16Immediate(Op.Val, Imm);
703 /// SelectAddressRegReg - Given the specified addressed, check to see if it
704 /// can be represented as an indexed [r+r] operation. Returns false if it
705 /// can be more efficiently represented with [r+imm].
706 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
710 if (N.getOpcode() == ISD::ADD) {
711 if (isIntS16Immediate(N.getOperand(1), imm))
713 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
716 Base = N.getOperand(0);
717 Index = N.getOperand(1);
719 } else if (N.getOpcode() == ISD::OR) {
720 if (isIntS16Immediate(N.getOperand(1), imm))
721 return false; // r+i can fold it if we can.
723 // If this is an or of disjoint bitfields, we can codegen this as an add
724 // (for better address arithmetic) if the LHS and RHS of the OR are provably
726 uint64_t LHSKnownZero, LHSKnownOne;
727 uint64_t RHSKnownZero, RHSKnownOne;
728 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
731 DAG.ComputeMaskedBits(N.getOperand(1), ~0U, RHSKnownZero, RHSKnownOne);
732 // If all of the bits are known zero on the LHS or RHS, the add won't
734 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
735 Base = N.getOperand(0);
736 Index = N.getOperand(1);
745 /// Returns true if the address N can be represented by a base register plus
746 /// a signed 16-bit displacement [r+imm], and if it is not better
747 /// represented as reg+reg.
748 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
749 SDOperand &Base, SelectionDAG &DAG){
750 // If this can be more profitably realized as r+r, fail.
751 if (SelectAddressRegReg(N, Disp, Base, DAG))
754 if (N.getOpcode() == ISD::ADD) {
756 if (isIntS16Immediate(N.getOperand(1), imm)) {
757 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
758 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
759 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
761 Base = N.getOperand(0);
763 return true; // [r+i]
764 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
765 // Match LOAD (ADD (X, Lo(G))).
766 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
767 && "Cannot handle constant offsets yet!");
768 Disp = N.getOperand(1).getOperand(0); // The global address.
769 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
770 Disp.getOpcode() == ISD::TargetConstantPool ||
771 Disp.getOpcode() == ISD::TargetJumpTable);
772 Base = N.getOperand(0);
773 return true; // [&g+r]
775 } else if (N.getOpcode() == ISD::OR) {
777 if (isIntS16Immediate(N.getOperand(1), imm)) {
778 // If this is an or of disjoint bitfields, we can codegen this as an add
779 // (for better address arithmetic) if the LHS and RHS of the OR are
780 // provably disjoint.
781 uint64_t LHSKnownZero, LHSKnownOne;
782 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
783 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
784 // If all of the bits are known zero on the LHS or RHS, the add won't
786 Base = N.getOperand(0);
787 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
791 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
792 // Loading from a constant address.
794 // If this address fits entirely in a 16-bit sext immediate field, codegen
797 if (isIntS16Immediate(CN, Imm)) {
798 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
799 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
803 // Handle 32-bit sext immediates with LIS + addr mode.
804 if (CN->getValueType(0) == MVT::i32 ||
805 (int64_t)CN->getValue() == (int)CN->getValue()) {
806 int Addr = (int)CN->getValue();
808 // Otherwise, break this down into an LIS + disp.
809 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
811 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
812 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
813 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
818 Disp = DAG.getTargetConstant(0, getPointerTy());
819 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
820 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
823 return true; // [r+0]
826 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
827 /// represented as an indexed [r+r] operation.
828 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
831 // Check to see if we can easily represent this as an [r+r] address. This
832 // will fail if it thinks that the address is more profitably represented as
833 // reg+imm, e.g. where imm = 0.
834 if (SelectAddressRegReg(N, Base, Index, DAG))
837 // If the operand is an addition, always emit this as [r+r], since this is
838 // better (for code size, and execution, as the memop does the add for free)
839 // than emitting an explicit add.
840 if (N.getOpcode() == ISD::ADD) {
841 Base = N.getOperand(0);
842 Index = N.getOperand(1);
846 // Otherwise, do it the hard way, using R0 as the base register.
847 Base = DAG.getRegister(PPC::R0, N.getValueType());
852 /// SelectAddressRegImmShift - Returns true if the address N can be
853 /// represented by a base register plus a signed 14-bit displacement
854 /// [r+imm*4]. Suitable for use by STD and friends.
855 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
858 // If this can be more profitably realized as r+r, fail.
859 if (SelectAddressRegReg(N, Disp, Base, DAG))
862 if (N.getOpcode() == ISD::ADD) {
864 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
865 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
866 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
867 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
869 Base = N.getOperand(0);
871 return true; // [r+i]
872 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
873 // Match LOAD (ADD (X, Lo(G))).
874 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
875 && "Cannot handle constant offsets yet!");
876 Disp = N.getOperand(1).getOperand(0); // The global address.
877 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
878 Disp.getOpcode() == ISD::TargetConstantPool ||
879 Disp.getOpcode() == ISD::TargetJumpTable);
880 Base = N.getOperand(0);
881 return true; // [&g+r]
883 } else if (N.getOpcode() == ISD::OR) {
885 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
886 // If this is an or of disjoint bitfields, we can codegen this as an add
887 // (for better address arithmetic) if the LHS and RHS of the OR are
888 // provably disjoint.
889 uint64_t LHSKnownZero, LHSKnownOne;
890 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
891 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
892 // If all of the bits are known zero on the LHS or RHS, the add won't
894 Base = N.getOperand(0);
895 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
899 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
900 // Loading from a constant address. Verify low two bits are clear.
901 if ((CN->getValue() & 3) == 0) {
902 // If this address fits entirely in a 14-bit sext immediate field, codegen
905 if (isIntS16Immediate(CN, Imm)) {
906 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
907 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
911 // Fold the low-part of 32-bit absolute addresses into addr mode.
912 if (CN->getValueType(0) == MVT::i32 ||
913 (int64_t)CN->getValue() == (int)CN->getValue()) {
914 int Addr = (int)CN->getValue();
916 // Otherwise, break this down into an LIS + disp.
917 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
919 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
920 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
921 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
927 Disp = DAG.getTargetConstant(0, getPointerTy());
928 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
929 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
932 return true; // [r+0]
936 /// getPreIndexedAddressParts - returns true by value, base pointer and
937 /// offset pointer and addressing mode by reference if the node's address
938 /// can be legally represented as pre-indexed load / store address.
939 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
941 ISD::MemIndexedMode &AM,
943 // Disabled by default for now.
944 if (!EnablePPCPreinc) return false;
948 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
949 Ptr = LD->getBasePtr();
950 VT = LD->getLoadedVT();
952 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
954 Ptr = ST->getBasePtr();
955 VT = ST->getStoredVT();
959 // PowerPC doesn't have preinc load/store instructions for vectors.
960 if (MVT::isVector(VT))
963 // TODO: Check reg+reg first.
965 // LDU/STU use reg+imm*4, others use reg+imm.
966 if (VT != MVT::i64) {
968 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
972 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
976 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
977 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
978 // sext i32 to i64 when addr mode is r+i.
979 if (LD->getValueType(0) == MVT::i64 && LD->getLoadedVT() == MVT::i32 &&
980 LD->getExtensionType() == ISD::SEXTLOAD &&
981 isa<ConstantSDNode>(Offset))
989 //===----------------------------------------------------------------------===//
990 // LowerOperation implementation
991 //===----------------------------------------------------------------------===//
993 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
994 MVT::ValueType PtrVT = Op.getValueType();
995 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
996 Constant *C = CP->getConstVal();
997 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
998 SDOperand Zero = DAG.getConstant(0, PtrVT);
1000 const TargetMachine &TM = DAG.getTarget();
1002 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1003 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1005 // If this is a non-darwin platform, we don't support non-static relo models
1007 if (TM.getRelocationModel() == Reloc::Static ||
1008 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1009 // Generate non-pic code that has direct accesses to the constant pool.
1010 // The address of the global is just (hi(&g)+lo(&g)).
1011 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1014 if (TM.getRelocationModel() == Reloc::PIC_) {
1015 // With PIC, the first instruction is actually "GR+hi(&G)".
1016 Hi = DAG.getNode(ISD::ADD, PtrVT,
1017 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1020 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1024 static SDOperand LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1025 MVT::ValueType PtrVT = Op.getValueType();
1026 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1027 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1028 SDOperand Zero = DAG.getConstant(0, PtrVT);
1030 const TargetMachine &TM = DAG.getTarget();
1032 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1033 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1035 // If this is a non-darwin platform, we don't support non-static relo models
1037 if (TM.getRelocationModel() == Reloc::Static ||
1038 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1039 // Generate non-pic code that has direct accesses to the constant pool.
1040 // The address of the global is just (hi(&g)+lo(&g)).
1041 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1044 if (TM.getRelocationModel() == Reloc::PIC_) {
1045 // With PIC, the first instruction is actually "GR+hi(&G)".
1046 Hi = DAG.getNode(ISD::ADD, PtrVT,
1047 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1050 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1054 static SDOperand LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
1055 assert(0 && "TLS not implemented for PPC.");
1058 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
1059 MVT::ValueType PtrVT = Op.getValueType();
1060 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1061 GlobalValue *GV = GSDN->getGlobal();
1062 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1063 SDOperand Zero = DAG.getConstant(0, PtrVT);
1065 const TargetMachine &TM = DAG.getTarget();
1067 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1068 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1070 // If this is a non-darwin platform, we don't support non-static relo models
1072 if (TM.getRelocationModel() == Reloc::Static ||
1073 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1074 // Generate non-pic code that has direct accesses to globals.
1075 // The address of the global is just (hi(&g)+lo(&g)).
1076 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1079 if (TM.getRelocationModel() == Reloc::PIC_) {
1080 // With PIC, the first instruction is actually "GR+hi(&G)".
1081 Hi = DAG.getNode(ISD::ADD, PtrVT,
1082 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1085 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1087 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1090 // If the global is weak or external, we have to go through the lazy
1092 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1095 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1096 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1098 // If we're comparing for equality to zero, expose the fact that this is
1099 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1100 // fold the new nodes.
1101 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1102 if (C->isNullValue() && CC == ISD::SETEQ) {
1103 MVT::ValueType VT = Op.getOperand(0).getValueType();
1104 SDOperand Zext = Op.getOperand(0);
1105 if (VT < MVT::i32) {
1107 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1109 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1110 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1111 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1112 DAG.getConstant(Log2b, MVT::i32));
1113 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1115 // Leave comparisons against 0 and -1 alone for now, since they're usually
1116 // optimized. FIXME: revisit this when we can custom lower all setcc
1118 if (C->isAllOnesValue() || C->isNullValue())
1122 // If we have an integer seteq/setne, turn it into a compare against zero
1123 // by xor'ing the rhs with the lhs, which is faster than setting a
1124 // condition register, reading it back out, and masking the correct bit. The
1125 // normal approach here uses sub to do this instead of xor. Using xor exposes
1126 // the result to other bit-twiddling opportunities.
1127 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1128 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1129 MVT::ValueType VT = Op.getValueType();
1130 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1132 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1137 static SDOperand LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1138 int VarArgsFrameIndex,
1139 int VarArgsStackOffset,
1140 unsigned VarArgsNumGPR,
1141 unsigned VarArgsNumFPR,
1142 const PPCSubtarget &Subtarget) {
1144 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1147 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1148 int VarArgsFrameIndex,
1149 int VarArgsStackOffset,
1150 unsigned VarArgsNumGPR,
1151 unsigned VarArgsNumFPR,
1152 const PPCSubtarget &Subtarget) {
1154 if (Subtarget.isMachoABI()) {
1155 // vastart just stores the address of the VarArgsFrameIndex slot into the
1156 // memory location argument.
1157 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1158 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1159 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
1160 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV->getValue(),
1164 // For ELF 32 ABI we follow the layout of the va_list struct.
1165 // We suppose the given va_list is already allocated.
1168 // char gpr; /* index into the array of 8 GPRs
1169 // * stored in the register save area
1170 // * gpr=0 corresponds to r3,
1171 // * gpr=1 to r4, etc.
1173 // char fpr; /* index into the array of 8 FPRs
1174 // * stored in the register save area
1175 // * fpr=0 corresponds to f1,
1176 // * fpr=1 to f2, etc.
1178 // char *overflow_arg_area;
1179 // /* location on stack that holds
1180 // * the next overflow argument
1182 // char *reg_save_area;
1183 // /* where r3:r10 and f1:f8 (if saved)
1189 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1190 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1193 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1195 SDOperand StackOffset = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1196 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1198 SDOperand ConstFrameOffset = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8,
1200 SDOperand ConstStackOffset = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8 - 1,
1202 SDOperand ConstFPROffset = DAG.getConstant(1, PtrVT);
1204 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
1206 // Store first byte : number of int regs
1207 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1208 Op.getOperand(1), SV->getValue(),
1210 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1213 // Store second byte : number of float regs
1214 SDOperand secondStore = DAG.getStore(firstStore, ArgFPR, nextPtr,
1215 SV->getValue(), SV->getOffset());
1216 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1218 // Store second word : arguments given on stack
1219 SDOperand thirdStore = DAG.getStore(secondStore, StackOffset, nextPtr,
1220 SV->getValue(), SV->getOffset());
1221 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1223 // Store third word : arguments given in registers
1224 return DAG.getStore(thirdStore, FR, nextPtr, SV->getValue(),
1229 #include "PPCGenCallingConv.inc"
1231 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1232 /// depending on which subtarget is selected.
1233 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1234 if (Subtarget.isMachoABI()) {
1235 static const unsigned FPR[] = {
1236 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1237 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1243 static const unsigned FPR[] = {
1244 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1250 static SDOperand LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG,
1251 int &VarArgsFrameIndex,
1252 int &VarArgsStackOffset,
1253 unsigned &VarArgsNumGPR,
1254 unsigned &VarArgsNumFPR,
1255 const PPCSubtarget &Subtarget) {
1256 // TODO: add description of PPC stack frame format, or at least some docs.
1258 MachineFunction &MF = DAG.getMachineFunction();
1259 MachineFrameInfo *MFI = MF.getFrameInfo();
1260 SSARegMap *RegMap = MF.getSSARegMap();
1261 SmallVector<SDOperand, 8> ArgValues;
1262 SDOperand Root = Op.getOperand(0);
1264 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1265 bool isPPC64 = PtrVT == MVT::i64;
1266 bool isMachoABI = Subtarget.isMachoABI();
1267 bool isELF32_ABI = Subtarget.isELF32_ABI();
1268 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1270 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1272 static const unsigned GPR_32[] = { // 32-bit registers.
1273 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1274 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1276 static const unsigned GPR_64[] = { // 64-bit registers.
1277 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1278 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1281 static const unsigned *FPR = GetFPR(Subtarget);
1283 static const unsigned VR[] = {
1284 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1285 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1288 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1289 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1290 const unsigned Num_VR_Regs = array_lengthof( VR);
1292 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1294 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1296 // Add DAG nodes to load the arguments or copy them out of registers. On
1297 // entry to a function on PPC, the arguments start after the linkage area,
1298 // although the first ones are often in registers.
1300 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1301 // represented with two words (long long or double) must be copied to an
1302 // even GPR_idx value or to an even ArgOffset value.
1304 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1306 bool needsLoad = false;
1307 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1308 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1309 unsigned ArgSize = ObjSize;
1310 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1311 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1312 // See if next argument requires stack alignment in ELF
1313 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1314 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1315 (!(Flags & AlignFlag)));
1317 unsigned CurArgOffset = ArgOffset;
1319 default: assert(0 && "Unhandled argument type!");
1321 // Double word align in ELF
1322 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1323 if (GPR_idx != Num_GPR_Regs) {
1324 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1325 MF.addLiveIn(GPR[GPR_idx], VReg);
1326 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1330 ArgSize = PtrByteSize;
1332 // Stack align in ELF
1333 if (needsLoad && Expand && isELF32_ABI)
1334 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1335 // All int arguments reserve stack space in Macho ABI.
1336 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1339 case MVT::i64: // PPC64
1340 if (GPR_idx != Num_GPR_Regs) {
1341 unsigned VReg = RegMap->createVirtualRegister(&PPC::G8RCRegClass);
1342 MF.addLiveIn(GPR[GPR_idx], VReg);
1343 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1348 // All int arguments reserve stack space in Macho ABI.
1349 if (isMachoABI || needsLoad) ArgOffset += 8;
1354 // Every 4 bytes of argument space consumes one of the GPRs available for
1355 // argument passing.
1356 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1358 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1361 if (FPR_idx != Num_FPR_Regs) {
1363 if (ObjectVT == MVT::f32)
1364 VReg = RegMap->createVirtualRegister(&PPC::F4RCRegClass);
1366 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1367 MF.addLiveIn(FPR[FPR_idx], VReg);
1368 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1374 // Stack align in ELF
1375 if (needsLoad && Expand && isELF32_ABI)
1376 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1377 // All FP arguments reserve stack space in Macho ABI.
1378 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1384 // Note that vector arguments in registers don't reserve stack space.
1385 if (VR_idx != Num_VR_Regs) {
1386 unsigned VReg = RegMap->createVirtualRegister(&PPC::VRRCRegClass);
1387 MF.addLiveIn(VR[VR_idx], VReg);
1388 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1391 // This should be simple, but requires getting 16-byte aligned stack
1393 assert(0 && "Loading VR argument not implemented yet!");
1399 // We need to load the argument to a virtual register if we determined above
1400 // that we ran out of physical registers of the appropriate type
1402 // If the argument is actually used, emit a load from the right stack
1404 if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
1405 int FI = MFI->CreateFixedObject(ObjSize,
1406 CurArgOffset + (ArgSize - ObjSize));
1407 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1408 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1410 // Don't emit a dead load.
1411 ArgVal = DAG.getNode(ISD::UNDEF, ObjectVT);
1415 ArgValues.push_back(ArgVal);
1418 // If the function takes variable number of arguments, make a frame index for
1419 // the start of the first vararg value... for expansion of llvm.va_start.
1420 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1425 VarArgsNumGPR = GPR_idx;
1426 VarArgsNumFPR = FPR_idx;
1428 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1430 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1431 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1432 MVT::getSizeInBits(PtrVT)/8);
1434 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1441 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1443 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1445 SmallVector<SDOperand, 8> MemOps;
1447 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1448 // stored to the VarArgsFrameIndex on the stack.
1450 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1451 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1452 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1453 MemOps.push_back(Store);
1454 // Increment the address by four for the next argument to store
1455 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1456 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1460 // If this function is vararg, store any remaining integer argument regs
1461 // to their spots on the stack so that they may be loaded by deferencing the
1462 // result of va_next.
1463 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1466 VReg = RegMap->createVirtualRegister(&PPC::G8RCRegClass);
1468 VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1470 MF.addLiveIn(GPR[GPR_idx], VReg);
1471 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1472 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1473 MemOps.push_back(Store);
1474 // Increment the address by four for the next argument to store
1475 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1476 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1479 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1482 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1483 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1484 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1485 MemOps.push_back(Store);
1486 // Increment the address by eight for the next argument to store
1487 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1489 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1492 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1494 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1496 MF.addLiveIn(FPR[FPR_idx], VReg);
1497 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1498 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1499 MemOps.push_back(Store);
1500 // Increment the address by eight for the next argument to store
1501 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1503 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1507 if (!MemOps.empty())
1508 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1511 ArgValues.push_back(Root);
1513 // Return the new list of results.
1514 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1515 Op.Val->value_end());
1516 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1519 /// isCallCompatibleAddress - Return the immediate to use if the specified
1520 /// 32-bit value is representable in the immediate field of a BxA instruction.
1521 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1522 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1525 int Addr = C->getValue();
1526 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1527 (Addr << 6 >> 6) != Addr)
1528 return 0; // Top 6 bits have to be sext of immediate.
1530 return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).Val;
1534 static SDOperand LowerCALL(SDOperand Op, SelectionDAG &DAG,
1535 const PPCSubtarget &Subtarget) {
1536 SDOperand Chain = Op.getOperand(0);
1537 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1538 SDOperand Callee = Op.getOperand(4);
1539 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1541 bool isMachoABI = Subtarget.isMachoABI();
1542 bool isELF32_ABI = Subtarget.isELF32_ABI();
1544 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1545 bool isPPC64 = PtrVT == MVT::i64;
1546 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1548 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1549 // SelectExpr to use to put the arguments in the appropriate registers.
1550 std::vector<SDOperand> args_to_use;
1552 // Count how many bytes are to be pushed on the stack, including the linkage
1553 // area, and parameter passing area. We start with 24/48 bytes, which is
1554 // prereserved space for [SP][CR][LR][3 x unused].
1555 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1557 // Add up all the space actually used.
1558 for (unsigned i = 0; i != NumOps; ++i) {
1559 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1560 ArgSize = std::max(ArgSize, PtrByteSize);
1561 NumBytes += ArgSize;
1564 // The prolog code of the callee may store up to 8 GPR argument registers to
1565 // the stack, allowing va_start to index over them in memory if its varargs.
1566 // Because we cannot tell if this is needed on the caller side, we have to
1567 // conservatively assume that it is needed. As such, make sure we have at
1568 // least enough stack space for the caller to store the 8 GPRs.
1569 NumBytes = std::max(NumBytes,
1570 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1572 // Adjust the stack pointer for the new arguments...
1573 // These operations are automatically eliminated by the prolog/epilog pass
1574 Chain = DAG.getCALLSEQ_START(Chain,
1575 DAG.getConstant(NumBytes, PtrVT));
1577 // Set up a copy of the stack pointer for use loading and storing any
1578 // arguments that may not fit in the registers available for argument
1582 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1584 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1586 // Figure out which arguments are going to go in registers, and which in
1587 // memory. Also, if this is a vararg function, floating point operations
1588 // must be stored to our stack, and loaded into integer regs as well, if
1589 // any integer regs are available for argument passing.
1590 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1591 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1593 static const unsigned GPR_32[] = { // 32-bit registers.
1594 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1595 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1597 static const unsigned GPR_64[] = { // 64-bit registers.
1598 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1599 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1601 static const unsigned *FPR = GetFPR(Subtarget);
1603 static const unsigned VR[] = {
1604 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1605 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1607 const unsigned NumGPRs = array_lengthof(GPR_32);
1608 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1609 const unsigned NumVRs = array_lengthof( VR);
1611 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1613 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1614 SmallVector<SDOperand, 8> MemOpChains;
1615 for (unsigned i = 0; i != NumOps; ++i) {
1617 SDOperand Arg = Op.getOperand(5+2*i);
1618 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1619 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1620 // See if next argument requires stack alignment in ELF
1621 unsigned next = 5+2*(i+1)+1;
1622 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1623 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1624 (!(Flags & AlignFlag)));
1626 // PtrOff will be used to store the current argument to the stack if a
1627 // register cannot be found for it.
1630 // Stack align in ELF 32
1631 if (isELF32_ABI && Expand)
1632 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1633 StackPtr.getValueType());
1635 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1637 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1639 // On PPC64, promote integers to 64-bit values.
1640 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1641 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1643 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1646 switch (Arg.getValueType()) {
1647 default: assert(0 && "Unexpected ValueType for argument!");
1650 // Double word align in ELF
1651 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1652 if (GPR_idx != NumGPRs) {
1653 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1655 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1658 if (inMem || isMachoABI) {
1659 // Stack align in ELF
1660 if (isELF32_ABI && Expand)
1661 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1663 ArgOffset += PtrByteSize;
1669 // Float varargs need to be promoted to double.
1670 if (Arg.getValueType() == MVT::f32)
1671 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1674 if (FPR_idx != NumFPRs) {
1675 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1678 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1679 MemOpChains.push_back(Store);
1681 // Float varargs are always shadowed in available integer registers
1682 if (GPR_idx != NumGPRs) {
1683 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1684 MemOpChains.push_back(Load.getValue(1));
1685 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1688 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1689 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1690 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1691 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1692 MemOpChains.push_back(Load.getValue(1));
1693 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1697 // If we have any FPRs remaining, we may also have GPRs remaining.
1698 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1701 if (GPR_idx != NumGPRs)
1703 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1704 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1709 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1712 if (inMem || isMachoABI) {
1713 // Stack align in ELF
1714 if (isELF32_ABI && Expand)
1715 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1719 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1726 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1727 assert(VR_idx != NumVRs &&
1728 "Don't support passing more than 12 vector args yet!");
1729 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1733 if (!MemOpChains.empty())
1734 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1735 &MemOpChains[0], MemOpChains.size());
1737 // Build a sequence of copy-to-reg nodes chained together with token chain
1738 // and flag operands which copy the outgoing args into the appropriate regs.
1740 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1741 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1743 InFlag = Chain.getValue(1);
1746 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1747 if (isVarArg && isELF32_ABI) {
1748 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1749 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1750 InFlag = Chain.getValue(1);
1753 std::vector<MVT::ValueType> NodeTys;
1754 NodeTys.push_back(MVT::Other); // Returns a chain
1755 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1757 SmallVector<SDOperand, 8> Ops;
1758 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1760 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1761 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1762 // node so that legalize doesn't hack it.
1763 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1764 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1765 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1766 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1767 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1768 // If this is an absolute destination address, use the munged value.
1769 Callee = SDOperand(Dest, 0);
1771 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1772 // to do the call, we can't use PPCISD::CALL.
1773 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1774 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1775 InFlag = Chain.getValue(1);
1777 // Copy the callee address into R12 on darwin.
1779 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1780 InFlag = Chain.getValue(1);
1784 NodeTys.push_back(MVT::Other);
1785 NodeTys.push_back(MVT::Flag);
1786 Ops.push_back(Chain);
1787 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1791 // If this is a direct call, pass the chain and the callee.
1793 Ops.push_back(Chain);
1794 Ops.push_back(Callee);
1797 // Add argument registers to the end of the list so that they are known live
1799 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1800 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1801 RegsToPass[i].second.getValueType()));
1804 Ops.push_back(InFlag);
1805 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1806 InFlag = Chain.getValue(1);
1808 SDOperand ResultVals[3];
1809 unsigned NumResults = 0;
1812 // If the call has results, copy the values out of the ret val registers.
1813 switch (Op.Val->getValueType(0)) {
1814 default: assert(0 && "Unexpected ret value!");
1815 case MVT::Other: break;
1817 if (Op.Val->getValueType(1) == MVT::i32) {
1818 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1819 ResultVals[0] = Chain.getValue(0);
1820 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
1821 Chain.getValue(2)).getValue(1);
1822 ResultVals[1] = Chain.getValue(0);
1824 NodeTys.push_back(MVT::i32);
1826 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1827 ResultVals[0] = Chain.getValue(0);
1830 NodeTys.push_back(MVT::i32);
1833 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
1834 ResultVals[0] = Chain.getValue(0);
1836 NodeTys.push_back(MVT::i64);
1839 if (Op.Val->getValueType(1) == MVT::f64) {
1840 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
1841 ResultVals[0] = Chain.getValue(0);
1842 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
1843 Chain.getValue(2)).getValue(1);
1844 ResultVals[1] = Chain.getValue(0);
1846 NodeTys.push_back(MVT::f64);
1847 NodeTys.push_back(MVT::f64);
1850 // else fall through
1852 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
1853 InFlag).getValue(1);
1854 ResultVals[0] = Chain.getValue(0);
1856 NodeTys.push_back(Op.Val->getValueType(0));
1862 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
1863 InFlag).getValue(1);
1864 ResultVals[0] = Chain.getValue(0);
1866 NodeTys.push_back(Op.Val->getValueType(0));
1870 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
1871 DAG.getConstant(NumBytes, PtrVT));
1872 NodeTys.push_back(MVT::Other);
1874 // If the function returns void, just return the chain.
1875 if (NumResults == 0)
1878 // Otherwise, merge everything together with a MERGE_VALUES node.
1879 ResultVals[NumResults++] = Chain;
1880 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1881 ResultVals, NumResults);
1882 return Res.getValue(Op.ResNo);
1885 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1886 SmallVector<CCValAssign, 16> RVLocs;
1887 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1888 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1889 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1890 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
1892 // If this is the first return lowered for this function, add the regs to the
1893 // liveout set for the function.
1894 if (DAG.getMachineFunction().liveout_empty()) {
1895 for (unsigned i = 0; i != RVLocs.size(); ++i)
1896 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
1899 SDOperand Chain = Op.getOperand(0);
1902 // Copy the result values into the output registers.
1903 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1904 CCValAssign &VA = RVLocs[i];
1905 assert(VA.isRegLoc() && "Can only return in registers!");
1906 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1907 Flag = Chain.getValue(1);
1911 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
1913 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
1916 static SDOperand LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
1917 const PPCSubtarget &Subtarget) {
1918 // When we pop the dynamic allocation we need to restore the SP link.
1920 // Get the corect type for pointers.
1921 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1923 // Construct the stack pointer operand.
1924 bool IsPPC64 = Subtarget.isPPC64();
1925 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
1926 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
1928 // Get the operands for the STACKRESTORE.
1929 SDOperand Chain = Op.getOperand(0);
1930 SDOperand SaveSP = Op.getOperand(1);
1932 // Load the old link SP.
1933 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
1935 // Restore the stack pointer.
1936 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
1938 // Store the old link SP.
1939 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
1942 static SDOperand LowerDYNAMIC_STACKALLOC(SDOperand Op, SelectionDAG &DAG,
1943 const PPCSubtarget &Subtarget) {
1944 MachineFunction &MF = DAG.getMachineFunction();
1945 bool IsPPC64 = Subtarget.isPPC64();
1946 bool isMachoABI = Subtarget.isMachoABI();
1948 // Get current frame pointer save index. The users of this index will be
1949 // primarily DYNALLOC instructions.
1950 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1951 int FPSI = FI->getFramePointerSaveIndex();
1953 // If the frame pointer save index hasn't been defined yet.
1955 // Find out what the fix offset of the frame pointer save area.
1956 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
1958 // Allocate the frame index for frame pointer save area.
1959 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
1961 FI->setFramePointerSaveIndex(FPSI);
1965 SDOperand Chain = Op.getOperand(0);
1966 SDOperand Size = Op.getOperand(1);
1968 // Get the corect type for pointers.
1969 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1971 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
1972 DAG.getConstant(0, PtrVT), Size);
1973 // Construct a node for the frame pointer save index.
1974 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
1975 // Build a DYNALLOC node.
1976 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
1977 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
1978 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
1982 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
1984 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
1985 // Not FP? Not a fsel.
1986 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
1987 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
1990 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1992 // Cannot handle SETEQ/SETNE.
1993 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
1995 MVT::ValueType ResVT = Op.getValueType();
1996 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
1997 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
1998 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
2000 // If the RHS of the comparison is a 0.0, we don't need to do the
2001 // subtraction at all.
2002 if (isFloatingPointZero(RHS))
2004 default: break; // SETUO etc aren't handled by fsel.
2008 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2012 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2013 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2014 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2018 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2022 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2023 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2024 return DAG.getNode(PPCISD::FSEL, ResVT,
2025 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2030 default: break; // SETUO etc aren't handled by fsel.
2034 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2035 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2036 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2037 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2041 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2042 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2043 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2044 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2048 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2049 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2050 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2051 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2055 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2056 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2057 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2058 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2063 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2064 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2065 SDOperand Src = Op.getOperand(0);
2066 if (Src.getValueType() == MVT::f32)
2067 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2070 switch (Op.getValueType()) {
2071 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2073 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2076 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2080 // Convert the FP value to an int value through memory.
2081 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
2082 if (Op.getValueType() == MVT::i32)
2083 Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
2087 static SDOperand LowerFP_ROUND_INREG(SDOperand Op, SelectionDAG &DAG) {
2088 assert(Op.getValueType() == MVT::ppcf128);
2089 SDNode *Node = Op.Val;
2090 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2091 assert(Node->getOperand(0).Val->getOpcode()==ISD::BUILD_PAIR);
2092 SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
2093 SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
2095 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2096 // of the long double, and puts FPSCR back the way it was. We do not
2097 // actually model FPSCR.
2098 std::vector<MVT::ValueType> NodeTys;
2099 SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
2101 NodeTys.push_back(MVT::f64); // Return register
2102 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2103 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2104 MFFSreg = Result.getValue(0);
2105 InFlag = Result.getValue(1);
2108 NodeTys.push_back(MVT::Flag); // Returns a flag
2109 Ops[0] = DAG.getConstant(31, MVT::i32);
2111 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2112 InFlag = Result.getValue(0);
2115 NodeTys.push_back(MVT::Flag); // Returns a flag
2116 Ops[0] = DAG.getConstant(30, MVT::i32);
2118 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2119 InFlag = Result.getValue(0);
2122 NodeTys.push_back(MVT::f64); // result of add
2123 NodeTys.push_back(MVT::Flag); // Returns a flag
2127 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2128 FPreg = Result.getValue(0);
2129 InFlag = Result.getValue(1);
2132 NodeTys.push_back(MVT::f64);
2133 Ops[0] = DAG.getConstant(1, MVT::i32);
2137 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2138 FPreg = Result.getValue(0);
2140 // We know the low half is about to be thrown away, so just use something
2142 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2145 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2146 if (Op.getOperand(0).getValueType() == MVT::i64) {
2147 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2148 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2149 if (Op.getValueType() == MVT::f32)
2150 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
2154 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2155 "Unhandled SINT_TO_FP type in custom expander!");
2156 // Since we only generate this in 64-bit mode, we can take advantage of
2157 // 64-bit registers. In particular, sign extend the input value into the
2158 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2159 // then lfd it and fcfid it.
2160 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2161 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2162 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2163 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2165 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2168 // STD the extended value into the stack slot.
2169 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2170 DAG.getEntryNode(), Ext64, FIdx,
2171 DAG.getSrcValue(NULL));
2172 // Load the value as a double.
2173 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2175 // FCFID it and return it.
2176 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2177 if (Op.getValueType() == MVT::f32)
2178 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
2182 static SDOperand LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2183 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2184 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
2186 // Expand into a bunch of logical ops. Note that these ops
2187 // depend on the PPC behavior for oversized shift amounts.
2188 SDOperand Lo = Op.getOperand(0);
2189 SDOperand Hi = Op.getOperand(1);
2190 SDOperand Amt = Op.getOperand(2);
2192 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2193 DAG.getConstant(32, MVT::i32), Amt);
2194 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
2195 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
2196 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2197 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2198 DAG.getConstant(-32U, MVT::i32));
2199 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
2200 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2201 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
2202 SDOperand OutOps[] = { OutLo, OutHi };
2203 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2207 static SDOperand LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2208 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2209 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRL!");
2211 // Otherwise, expand into a bunch of logical ops. Note that these ops
2212 // depend on the PPC behavior for oversized shift amounts.
2213 SDOperand Lo = Op.getOperand(0);
2214 SDOperand Hi = Op.getOperand(1);
2215 SDOperand Amt = Op.getOperand(2);
2217 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2218 DAG.getConstant(32, MVT::i32), Amt);
2219 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2220 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2221 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2222 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2223 DAG.getConstant(-32U, MVT::i32));
2224 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
2225 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2226 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
2227 SDOperand OutOps[] = { OutLo, OutHi };
2228 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2232 static SDOperand LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2233 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2234 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
2236 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
2237 SDOperand Lo = Op.getOperand(0);
2238 SDOperand Hi = Op.getOperand(1);
2239 SDOperand Amt = Op.getOperand(2);
2241 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2242 DAG.getConstant(32, MVT::i32), Amt);
2243 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2244 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2245 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2246 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2247 DAG.getConstant(-32U, MVT::i32));
2248 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
2249 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
2250 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
2251 Tmp4, Tmp6, ISD::SETLE);
2252 SDOperand OutOps[] = { OutLo, OutHi };
2253 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2257 //===----------------------------------------------------------------------===//
2258 // Vector related lowering.
2261 // If this is a vector of constants or undefs, get the bits. A bit in
2262 // UndefBits is set if the corresponding element of the vector is an
2263 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2264 // zero. Return true if this is not an array of constants, false if it is.
2266 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2267 uint64_t UndefBits[2]) {
2268 // Start with zero'd results.
2269 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2271 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2272 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2273 SDOperand OpVal = BV->getOperand(i);
2275 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2276 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2278 uint64_t EltBits = 0;
2279 if (OpVal.getOpcode() == ISD::UNDEF) {
2280 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2281 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2283 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2284 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2285 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2286 assert(CN->getValueType(0) == MVT::f32 &&
2287 "Only one legal FP vector type!");
2288 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2290 // Nonconstant element.
2294 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2297 //printf("%llx %llx %llx %llx\n",
2298 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2302 // If this is a splat (repetition) of a value across the whole vector, return
2303 // the smallest size that splats it. For example, "0x01010101010101..." is a
2304 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2305 // SplatSize = 1 byte.
2306 static bool isConstantSplat(const uint64_t Bits128[2],
2307 const uint64_t Undef128[2],
2308 unsigned &SplatBits, unsigned &SplatUndef,
2309 unsigned &SplatSize) {
2311 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2312 // the same as the lower 64-bits, ignoring undefs.
2313 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2314 return false; // Can't be a splat if two pieces don't match.
2316 uint64_t Bits64 = Bits128[0] | Bits128[1];
2317 uint64_t Undef64 = Undef128[0] & Undef128[1];
2319 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2321 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2322 return false; // Can't be a splat if two pieces don't match.
2324 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2325 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2327 // If the top 16-bits are different than the lower 16-bits, ignoring
2328 // undefs, we have an i32 splat.
2329 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2331 SplatUndef = Undef32;
2336 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2337 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2339 // If the top 8-bits are different than the lower 8-bits, ignoring
2340 // undefs, we have an i16 splat.
2341 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2343 SplatUndef = Undef16;
2348 // Otherwise, we have an 8-bit splat.
2349 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2350 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2355 /// BuildSplatI - Build a canonical splati of Val with an element size of
2356 /// SplatSize. Cast the result to VT.
2357 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2358 SelectionDAG &DAG) {
2359 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2361 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2362 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2365 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2367 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2371 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2373 // Build a canonical splat for this value.
2374 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2375 SmallVector<SDOperand, 8> Ops;
2376 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2377 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2378 &Ops[0], Ops.size());
2379 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2382 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2383 /// specified intrinsic ID.
2384 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2386 MVT::ValueType DestVT = MVT::Other) {
2387 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2388 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2389 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2392 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2393 /// specified intrinsic ID.
2394 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2395 SDOperand Op2, SelectionDAG &DAG,
2396 MVT::ValueType DestVT = MVT::Other) {
2397 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2398 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2399 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2403 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2404 /// amount. The result has the specified value type.
2405 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2406 MVT::ValueType VT, SelectionDAG &DAG) {
2407 // Force LHS/RHS to be the right type.
2408 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2409 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2412 for (unsigned i = 0; i != 16; ++i)
2413 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2414 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2415 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2416 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2419 // If this is a case we can't handle, return null and let the default
2420 // expansion code take care of it. If we CAN select this case, and if it
2421 // selects to a single instruction, return Op. Otherwise, if we can codegen
2422 // this case more efficiently than a constant pool load, lower it to the
2423 // sequence of ops that should be used.
2424 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2425 // If this is a vector of constants or undefs, get the bits. A bit in
2426 // UndefBits is set if the corresponding element of the vector is an
2427 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2429 uint64_t VectorBits[2];
2430 uint64_t UndefBits[2];
2431 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2432 return SDOperand(); // Not a constant vector.
2434 // If this is a splat (repetition) of a value across the whole vector, return
2435 // the smallest size that splats it. For example, "0x01010101010101..." is a
2436 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2437 // SplatSize = 1 byte.
2438 unsigned SplatBits, SplatUndef, SplatSize;
2439 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2440 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2442 // First, handle single instruction cases.
2445 if (SplatBits == 0) {
2446 // Canonicalize all zero vectors to be v4i32.
2447 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2448 SDOperand Z = DAG.getConstant(0, MVT::i32);
2449 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2450 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2455 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2456 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2457 if (SextVal >= -16 && SextVal <= 15)
2458 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2461 // Two instruction sequences.
2463 // If this value is in the range [-32,30] and is even, use:
2464 // tmp = VSPLTI[bhw], result = add tmp, tmp
2465 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2466 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2467 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2470 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2471 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2473 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2474 // Make -1 and vspltisw -1:
2475 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2477 // Make the VSLW intrinsic, computing 0x8000_0000.
2478 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2481 // xor by OnesV to invert it.
2482 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2483 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2486 // Check to see if this is a wide variety of vsplti*, binop self cases.
2487 unsigned SplatBitSize = SplatSize*8;
2488 static const signed char SplatCsts[] = {
2489 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2490 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2493 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2494 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2495 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2496 int i = SplatCsts[idx];
2498 // Figure out what shift amount will be used by altivec if shifted by i in
2500 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2502 // vsplti + shl self.
2503 if (SextVal == (i << (int)TypeShiftAmt)) {
2504 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2505 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2506 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2507 Intrinsic::ppc_altivec_vslw
2509 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2510 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2513 // vsplti + srl self.
2514 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2515 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2516 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2517 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2518 Intrinsic::ppc_altivec_vsrw
2520 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2521 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2524 // vsplti + sra self.
2525 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2526 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2527 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2528 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2529 Intrinsic::ppc_altivec_vsraw
2531 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2532 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2535 // vsplti + rol self.
2536 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2537 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2538 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2539 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2540 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2541 Intrinsic::ppc_altivec_vrlw
2543 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2544 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2547 // t = vsplti c, result = vsldoi t, t, 1
2548 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2549 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2550 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2552 // t = vsplti c, result = vsldoi t, t, 2
2553 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2554 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2555 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2557 // t = vsplti c, result = vsldoi t, t, 3
2558 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2559 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2560 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2564 // Three instruction sequences.
2566 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2567 if (SextVal >= 0 && SextVal <= 31) {
2568 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2569 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2570 LHS = DAG.getNode(ISD::SUB, Op.getValueType(), LHS, RHS);
2571 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2573 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2574 if (SextVal >= -31 && SextVal <= 0) {
2575 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2576 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2577 LHS = DAG.getNode(ISD::ADD, Op.getValueType(), LHS, RHS);
2578 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2585 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2586 /// the specified operations to build the shuffle.
2587 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2588 SDOperand RHS, SelectionDAG &DAG) {
2589 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2590 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2591 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2594 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2606 if (OpNum == OP_COPY) {
2607 if (LHSID == (1*9+2)*9+3) return LHS;
2608 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2612 SDOperand OpLHS, OpRHS;
2613 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2614 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2616 unsigned ShufIdxs[16];
2618 default: assert(0 && "Unknown i32 permute!");
2620 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2621 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2622 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2623 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2626 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2627 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2628 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2629 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2632 for (unsigned i = 0; i != 16; ++i)
2633 ShufIdxs[i] = (i&3)+0;
2636 for (unsigned i = 0; i != 16; ++i)
2637 ShufIdxs[i] = (i&3)+4;
2640 for (unsigned i = 0; i != 16; ++i)
2641 ShufIdxs[i] = (i&3)+8;
2644 for (unsigned i = 0; i != 16; ++i)
2645 ShufIdxs[i] = (i&3)+12;
2648 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2650 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2652 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2655 for (unsigned i = 0; i != 16; ++i)
2656 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2658 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2659 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2662 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2663 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2664 /// return the code it can be lowered into. Worst case, it can always be
2665 /// lowered into a vperm.
2666 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2667 SDOperand V1 = Op.getOperand(0);
2668 SDOperand V2 = Op.getOperand(1);
2669 SDOperand PermMask = Op.getOperand(2);
2671 // Cases that are handled by instructions that take permute immediates
2672 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2673 // selected by the instruction selector.
2674 if (V2.getOpcode() == ISD::UNDEF) {
2675 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2676 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2677 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2678 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2679 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2680 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2681 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2682 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2683 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2684 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2685 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2686 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2691 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2692 // and produce a fixed permutation. If any of these match, do not lower to
2694 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2695 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2696 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2697 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2698 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2699 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2700 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2701 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2702 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2705 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2706 // perfect shuffle table to emit an optimal matching sequence.
2707 unsigned PFIndexes[4];
2708 bool isFourElementShuffle = true;
2709 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2710 unsigned EltNo = 8; // Start out undef.
2711 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2712 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2713 continue; // Undef, ignore it.
2715 unsigned ByteSource =
2716 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2717 if ((ByteSource & 3) != j) {
2718 isFourElementShuffle = false;
2723 EltNo = ByteSource/4;
2724 } else if (EltNo != ByteSource/4) {
2725 isFourElementShuffle = false;
2729 PFIndexes[i] = EltNo;
2732 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
2733 // perfect shuffle vector to determine if it is cost effective to do this as
2734 // discrete instructions, or whether we should use a vperm.
2735 if (isFourElementShuffle) {
2736 // Compute the index in the perfect shuffle table.
2737 unsigned PFTableIndex =
2738 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2740 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2741 unsigned Cost = (PFEntry >> 30);
2743 // Determining when to avoid vperm is tricky. Many things affect the cost
2744 // of vperm, particularly how many times the perm mask needs to be computed.
2745 // For example, if the perm mask can be hoisted out of a loop or is already
2746 // used (perhaps because there are multiple permutes with the same shuffle
2747 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
2748 // the loop requires an extra register.
2750 // As a compromise, we only emit discrete instructions if the shuffle can be
2751 // generated in 3 or fewer operations. When we have loop information
2752 // available, if this block is within a loop, we should avoid using vperm
2753 // for 3-operation perms and use a constant pool load instead.
2755 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
2758 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
2759 // vector that will get spilled to the constant pool.
2760 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
2762 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
2763 // that it is in input element units, not in bytes. Convert now.
2764 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
2765 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
2767 SmallVector<SDOperand, 16> ResultMask;
2768 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
2770 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
2773 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
2775 for (unsigned j = 0; j != BytesPerElement; ++j)
2776 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
2780 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2781 &ResultMask[0], ResultMask.size());
2782 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
2785 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
2786 /// altivec comparison. If it is, return true and fill in Opc/isDot with
2787 /// information about the intrinsic.
2788 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
2790 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
2793 switch (IntrinsicID) {
2794 default: return false;
2795 // Comparison predicates.
2796 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
2797 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
2798 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
2799 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
2800 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
2801 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
2802 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
2803 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
2804 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
2805 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
2806 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
2807 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
2808 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
2810 // Normal Comparisons.
2811 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
2812 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
2813 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
2814 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
2815 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
2816 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
2817 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
2818 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
2819 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
2820 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
2821 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
2822 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
2823 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
2828 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
2829 /// lower, do it, otherwise return null.
2830 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
2831 // If this is a lowered altivec predicate compare, CompareOpc is set to the
2832 // opcode number of the comparison.
2835 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
2836 return SDOperand(); // Don't custom lower most intrinsics.
2838 // If this is a non-dot comparison, make the VCMP node and we are done.
2840 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
2841 Op.getOperand(1), Op.getOperand(2),
2842 DAG.getConstant(CompareOpc, MVT::i32));
2843 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
2846 // Create the PPCISD altivec 'dot' comparison node.
2848 Op.getOperand(2), // LHS
2849 Op.getOperand(3), // RHS
2850 DAG.getConstant(CompareOpc, MVT::i32)
2852 std::vector<MVT::ValueType> VTs;
2853 VTs.push_back(Op.getOperand(2).getValueType());
2854 VTs.push_back(MVT::Flag);
2855 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
2857 // Now that we have the comparison, emit a copy from the CR to a GPR.
2858 // This is flagged to the above dot comparison.
2859 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
2860 DAG.getRegister(PPC::CR6, MVT::i32),
2861 CompNode.getValue(1));
2863 // Unpack the result based on how the target uses it.
2864 unsigned BitNo; // Bit # of CR6.
2865 bool InvertBit; // Invert result?
2866 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
2867 default: // Can't happen, don't crash on invalid number though.
2868 case 0: // Return the value of the EQ bit of CR6.
2869 BitNo = 0; InvertBit = false;
2871 case 1: // Return the inverted value of the EQ bit of CR6.
2872 BitNo = 0; InvertBit = true;
2874 case 2: // Return the value of the LT bit of CR6.
2875 BitNo = 2; InvertBit = false;
2877 case 3: // Return the inverted value of the LT bit of CR6.
2878 BitNo = 2; InvertBit = true;
2882 // Shift the bit into the low position.
2883 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
2884 DAG.getConstant(8-(3-BitNo), MVT::i32));
2886 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
2887 DAG.getConstant(1, MVT::i32));
2889 // If we are supposed to, toggle the bit.
2891 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
2892 DAG.getConstant(1, MVT::i32));
2896 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2897 // Create a stack slot that is 16-byte aligned.
2898 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2899 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
2900 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2901 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2903 // Store the input value into Value#0 of the stack slot.
2904 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
2905 Op.getOperand(0), FIdx, NULL, 0);
2907 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
2910 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
2911 if (Op.getValueType() == MVT::v4i32) {
2912 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2914 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
2915 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
2917 SDOperand RHSSwap = // = vrlw RHS, 16
2918 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
2920 // Shrinkify inputs to v8i16.
2921 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
2922 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
2923 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
2925 // Low parts multiplied together, generating 32-bit results (we ignore the
2927 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
2928 LHS, RHS, DAG, MVT::v4i32);
2930 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
2931 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
2932 // Shift the high parts up 16 bits.
2933 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
2934 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
2935 } else if (Op.getValueType() == MVT::v8i16) {
2936 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2938 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
2940 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
2941 LHS, RHS, Zero, DAG);
2942 } else if (Op.getValueType() == MVT::v16i8) {
2943 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2945 // Multiply the even 8-bit parts, producing 16-bit sums.
2946 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
2947 LHS, RHS, DAG, MVT::v8i16);
2948 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
2950 // Multiply the odd 8-bit parts, producing 16-bit sums.
2951 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
2952 LHS, RHS, DAG, MVT::v8i16);
2953 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
2955 // Merge the results together.
2957 for (unsigned i = 0; i != 8; ++i) {
2958 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
2959 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
2961 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
2962 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2964 assert(0 && "Unknown mul to lower!");
2969 /// LowerOperation - Provide custom lowering hooks for some operations.
2971 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
2972 switch (Op.getOpcode()) {
2973 default: assert(0 && "Wasn't expecting to be able to lower this!");
2974 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2975 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
2976 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2977 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
2978 case ISD::SETCC: return LowerSETCC(Op, DAG);
2980 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
2981 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
2984 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
2985 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
2987 case ISD::FORMAL_ARGUMENTS:
2988 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
2989 VarArgsStackOffset, VarArgsNumGPR,
2990 VarArgsNumFPR, PPCSubTarget);
2992 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
2993 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
2994 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
2995 case ISD::DYNAMIC_STACKALLOC:
2996 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
2998 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
2999 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3000 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3001 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3003 // Lower 64-bit shifts.
3004 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3005 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3006 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3008 // Vector-related lowering.
3009 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3010 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3011 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3012 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3013 case ISD::MUL: return LowerMUL(Op, DAG);
3015 // Frame & Return address. Currently unimplemented
3016 case ISD::RETURNADDR: break;
3017 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3022 //===----------------------------------------------------------------------===//
3023 // Other Lowering Code
3024 //===----------------------------------------------------------------------===//
3027 PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
3028 MachineBasicBlock *BB) {
3029 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3030 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
3031 MI->getOpcode() == PPC::SELECT_CC_I8 ||
3032 MI->getOpcode() == PPC::SELECT_CC_F4 ||
3033 MI->getOpcode() == PPC::SELECT_CC_F8 ||
3034 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
3035 "Unexpected instr type to insert");
3037 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
3038 // control-flow pattern. The incoming instruction knows the destination vreg
3039 // to set, the condition code register to branch on, the true/false values to
3040 // select between, and a branch opcode to use.
3041 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3042 ilist<MachineBasicBlock>::iterator It = BB;
3048 // cmpTY ccX, r1, r2
3050 // fallthrough --> copy0MBB
3051 MachineBasicBlock *thisMBB = BB;
3052 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
3053 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
3054 unsigned SelectPred = MI->getOperand(4).getImm();
3055 BuildMI(BB, TII->get(PPC::BCC))
3056 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
3057 MachineFunction *F = BB->getParent();
3058 F->getBasicBlockList().insert(It, copy0MBB);
3059 F->getBasicBlockList().insert(It, sinkMBB);
3060 // Update machine-CFG edges by first adding all successors of the current
3061 // block to the new block which will contain the Phi node for the select.
3062 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
3063 e = BB->succ_end(); i != e; ++i)
3064 sinkMBB->addSuccessor(*i);
3065 // Next, remove all successors of the current block, and add the true
3066 // and fallthrough blocks as its successors.
3067 while(!BB->succ_empty())
3068 BB->removeSuccessor(BB->succ_begin());
3069 BB->addSuccessor(copy0MBB);
3070 BB->addSuccessor(sinkMBB);
3073 // %FalseValue = ...
3074 // # fallthrough to sinkMBB
3077 // Update machine-CFG edges
3078 BB->addSuccessor(sinkMBB);
3081 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3084 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3085 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3086 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3088 delete MI; // The pseudo instruction is gone now.
3092 //===----------------------------------------------------------------------===//
3093 // Target Optimization Hooks
3094 //===----------------------------------------------------------------------===//
3096 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3097 DAGCombinerInfo &DCI) const {
3098 TargetMachine &TM = getTargetMachine();
3099 SelectionDAG &DAG = DCI.DAG;
3100 switch (N->getOpcode()) {
3103 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3104 if (C->getValue() == 0) // 0 << V -> 0.
3105 return N->getOperand(0);
3109 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3110 if (C->getValue() == 0) // 0 >>u V -> 0.
3111 return N->getOperand(0);
3115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3116 if (C->getValue() == 0 || // 0 >>s V -> 0.
3117 C->isAllOnesValue()) // -1 >>s V -> -1.
3118 return N->getOperand(0);
3122 case ISD::SINT_TO_FP:
3123 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3124 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3125 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3126 // We allow the src/dst to be either f32/f64, but the intermediate
3127 // type must be i64.
3128 if (N->getOperand(0).getValueType() == MVT::i64) {
3129 SDOperand Val = N->getOperand(0).getOperand(0);
3130 if (Val.getValueType() == MVT::f32) {
3131 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3132 DCI.AddToWorklist(Val.Val);
3135 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3136 DCI.AddToWorklist(Val.Val);
3137 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3138 DCI.AddToWorklist(Val.Val);
3139 if (N->getValueType(0) == MVT::f32) {
3140 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
3141 DCI.AddToWorklist(Val.Val);
3144 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3145 // If the intermediate type is i32, we can avoid the load/store here
3152 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3153 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3154 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3155 N->getOperand(1).getValueType() == MVT::i32) {
3156 SDOperand Val = N->getOperand(1).getOperand(0);
3157 if (Val.getValueType() == MVT::f32) {
3158 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3159 DCI.AddToWorklist(Val.Val);
3161 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3162 DCI.AddToWorklist(Val.Val);
3164 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3165 N->getOperand(2), N->getOperand(3));
3166 DCI.AddToWorklist(Val.Val);
3170 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3171 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3172 N->getOperand(1).Val->hasOneUse() &&
3173 (N->getOperand(1).getValueType() == MVT::i32 ||
3174 N->getOperand(1).getValueType() == MVT::i16)) {
3175 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3176 // Do an any-extend to 32-bits if this is a half-word input.
3177 if (BSwapOp.getValueType() == MVT::i16)
3178 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3180 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3181 N->getOperand(2), N->getOperand(3),
3182 DAG.getValueType(N->getOperand(1).getValueType()));
3186 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3187 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3188 N->getOperand(0).hasOneUse() &&
3189 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3190 SDOperand Load = N->getOperand(0);
3191 LoadSDNode *LD = cast<LoadSDNode>(Load);
3192 // Create the byte-swapping load.
3193 std::vector<MVT::ValueType> VTs;
3194 VTs.push_back(MVT::i32);
3195 VTs.push_back(MVT::Other);
3196 SDOperand SV = DAG.getSrcValue(LD->getSrcValue(), LD->getSrcValueOffset());
3198 LD->getChain(), // Chain
3199 LD->getBasePtr(), // Ptr
3201 DAG.getValueType(N->getValueType(0)) // VT
3203 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3205 // If this is an i16 load, insert the truncate.
3206 SDOperand ResVal = BSLoad;
3207 if (N->getValueType(0) == MVT::i16)
3208 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3210 // First, combine the bswap away. This makes the value produced by the
3212 DCI.CombineTo(N, ResVal);
3214 // Next, combine the load away, we give it a bogus result value but a real
3215 // chain result. The result value is dead because the bswap is dead.
3216 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3218 // Return N so it doesn't get rechecked!
3219 return SDOperand(N, 0);
3223 case PPCISD::VCMP: {
3224 // If a VCMPo node already exists with exactly the same operands as this
3225 // node, use its result instead of this node (VCMPo computes both a CR6 and
3226 // a normal output).
3228 if (!N->getOperand(0).hasOneUse() &&
3229 !N->getOperand(1).hasOneUse() &&
3230 !N->getOperand(2).hasOneUse()) {
3232 // Scan all of the users of the LHS, looking for VCMPo's that match.
3233 SDNode *VCMPoNode = 0;
3235 SDNode *LHSN = N->getOperand(0).Val;
3236 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3238 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3239 (*UI)->getOperand(1) == N->getOperand(1) &&
3240 (*UI)->getOperand(2) == N->getOperand(2) &&
3241 (*UI)->getOperand(0) == N->getOperand(0)) {
3246 // If there is no VCMPo node, or if the flag value has a single use, don't
3248 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3251 // Look at the (necessarily single) use of the flag value. If it has a
3252 // chain, this transformation is more complex. Note that multiple things
3253 // could use the value result, which we should ignore.
3254 SDNode *FlagUser = 0;
3255 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3256 FlagUser == 0; ++UI) {
3257 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3259 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3260 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3267 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3268 // give up for right now.
3269 if (FlagUser->getOpcode() == PPCISD::MFCR)
3270 return SDOperand(VCMPoNode, 0);
3275 // If this is a branch on an altivec predicate comparison, lower this so
3276 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3277 // lowering is done pre-legalize, because the legalizer lowers the predicate
3278 // compare down to code that is difficult to reassemble.
3279 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3280 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3284 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3285 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3286 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3287 assert(isDot && "Can't compare against a vector result!");
3289 // If this is a comparison against something other than 0/1, then we know
3290 // that the condition is never/always true.
3291 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3292 if (Val != 0 && Val != 1) {
3293 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3294 return N->getOperand(0);
3295 // Always !=, turn it into an unconditional branch.
3296 return DAG.getNode(ISD::BR, MVT::Other,
3297 N->getOperand(0), N->getOperand(4));
3300 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3302 // Create the PPCISD altivec 'dot' comparison node.
3303 std::vector<MVT::ValueType> VTs;
3305 LHS.getOperand(2), // LHS of compare
3306 LHS.getOperand(3), // RHS of compare
3307 DAG.getConstant(CompareOpc, MVT::i32)
3309 VTs.push_back(LHS.getOperand(2).getValueType());
3310 VTs.push_back(MVT::Flag);
3311 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3313 // Unpack the result based on how the target uses it.
3314 PPC::Predicate CompOpc;
3315 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3316 default: // Can't happen, don't crash on invalid number though.
3317 case 0: // Branch on the value of the EQ bit of CR6.
3318 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3320 case 1: // Branch on the inverted value of the EQ bit of CR6.
3321 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3323 case 2: // Branch on the value of the LT bit of CR6.
3324 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3326 case 3: // Branch on the inverted value of the LT bit of CR6.
3327 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3331 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3332 DAG.getConstant(CompOpc, MVT::i32),
3333 DAG.getRegister(PPC::CR6, MVT::i32),
3334 N->getOperand(4), CompNode.getValue(1));
3343 //===----------------------------------------------------------------------===//
3344 // Inline Assembly Support
3345 //===----------------------------------------------------------------------===//
3347 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3349 uint64_t &KnownZero,
3351 const SelectionDAG &DAG,
3352 unsigned Depth) const {
3355 switch (Op.getOpcode()) {
3357 case PPCISD::LBRX: {
3358 // lhbrx is known to have the top bits cleared out.
3359 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3360 KnownZero = 0xFFFF0000;
3363 case ISD::INTRINSIC_WO_CHAIN: {
3364 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3366 case Intrinsic::ppc_altivec_vcmpbfp_p:
3367 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3368 case Intrinsic::ppc_altivec_vcmpequb_p:
3369 case Intrinsic::ppc_altivec_vcmpequh_p:
3370 case Intrinsic::ppc_altivec_vcmpequw_p:
3371 case Intrinsic::ppc_altivec_vcmpgefp_p:
3372 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3373 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3374 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3375 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3376 case Intrinsic::ppc_altivec_vcmpgtub_p:
3377 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3378 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3379 KnownZero = ~1U; // All bits but the low one are known to be zero.
3387 /// getConstraintType - Given a constraint, return the type of
3388 /// constraint it is for this target.
3389 PPCTargetLowering::ConstraintType
3390 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3391 if (Constraint.size() == 1) {
3392 switch (Constraint[0]) {
3399 return C_RegisterClass;
3402 return TargetLowering::getConstraintType(Constraint);
3405 std::pair<unsigned, const TargetRegisterClass*>
3406 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3407 MVT::ValueType VT) const {
3408 if (Constraint.size() == 1) {
3409 // GCC RS6000 Constraint Letters
3410 switch (Constraint[0]) {
3413 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3414 return std::make_pair(0U, PPC::G8RCRegisterClass);
3415 return std::make_pair(0U, PPC::GPRCRegisterClass);
3418 return std::make_pair(0U, PPC::F4RCRegisterClass);
3419 else if (VT == MVT::f64)
3420 return std::make_pair(0U, PPC::F8RCRegisterClass);
3423 return std::make_pair(0U, PPC::VRRCRegisterClass);
3425 return std::make_pair(0U, PPC::CRRCRegisterClass);
3429 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3433 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3434 /// vector. If it is invalid, don't add anything to Ops.
3435 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3436 std::vector<SDOperand>&Ops,
3437 SelectionDAG &DAG) {
3438 SDOperand Result(0,0);
3449 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3450 if (!CST) return; // Must be an immediate to match.
3451 unsigned Value = CST->getValue();
3453 default: assert(0 && "Unknown constraint letter!");
3454 case 'I': // "I" is a signed 16-bit constant.
3455 if ((short)Value == (int)Value)
3456 Result = DAG.getTargetConstant(Value, Op.getValueType());
3458 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3459 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3460 if ((short)Value == 0)
3461 Result = DAG.getTargetConstant(Value, Op.getValueType());
3463 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3464 if ((Value >> 16) == 0)
3465 Result = DAG.getTargetConstant(Value, Op.getValueType());
3467 case 'M': // "M" is a constant that is greater than 31.
3469 Result = DAG.getTargetConstant(Value, Op.getValueType());
3471 case 'N': // "N" is a positive constant that is an exact power of two.
3472 if ((int)Value > 0 && isPowerOf2_32(Value))
3473 Result = DAG.getTargetConstant(Value, Op.getValueType());
3475 case 'O': // "O" is the constant zero.
3477 Result = DAG.getTargetConstant(Value, Op.getValueType());
3479 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3480 if ((short)-Value == (int)-Value)
3481 Result = DAG.getTargetConstant(Value, Op.getValueType());
3489 Ops.push_back(Result);
3493 // Handle standard constraint letters.
3494 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3497 // isLegalAddressingMode - Return true if the addressing mode represented
3498 // by AM is legal for this target, for a load/store of the specified type.
3499 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3500 const Type *Ty) const {
3501 // FIXME: PPC does not allow r+i addressing modes for vectors!
3503 // PPC allows a sign-extended 16-bit immediate field.
3504 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3507 // No global is ever allowed as a base.
3511 // PPC only support r+r,
3513 case 0: // "r+i" or just "i", depending on HasBaseReg.
3516 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3518 // Otherwise we have r+r or r+i.
3521 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3523 // Allow 2*r as r+r.
3526 // No other scales are supported.
3533 /// isLegalAddressImmediate - Return true if the integer value can be used
3534 /// as the offset of the target addressing mode for load / store of the
3536 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3537 // PPC allows a sign-extended 16-bit immediate field.
3538 return (V > -(1 << 16) && V < (1 << 16)-1);
3541 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3545 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG)
3547 // Depths > 0 not supported yet!
3548 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3551 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3552 bool isPPC64 = PtrVT == MVT::i64;
3554 MachineFunction &MF = DAG.getMachineFunction();
3555 MachineFrameInfo *MFI = MF.getFrameInfo();
3556 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3557 && MFI->getStackSize();
3560 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3563 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,