1 //===-- SPUISelLowering.cpp - Cell SPU DAG Lowering Implementation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SPUTargetLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "SPURegisterNames.h"
15 #include "SPUISelLowering.h"
16 #include "SPUTargetMachine.h"
17 #include "llvm/ADT/VectorExtras.h"
18 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/Constants.h"
26 #include "llvm/Function.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Support/MathExtras.h"
30 #include "llvm/Target/TargetOptions.h"
36 // Used in getTargetNodeName() below
38 std::map<unsigned, const char *> node_names;
40 //! MVT::ValueType mapping to useful data for Cell SPU
41 struct valtype_map_s {
42 const MVT::ValueType valtype;
43 const int prefslot_byte;
46 const valtype_map_s valtype_map[] = {
57 const size_t n_valtype_map = sizeof(valtype_map) / sizeof(valtype_map[0]);
59 const valtype_map_s *getValueTypeMapEntry(MVT::ValueType VT) {
60 const valtype_map_s *retval = 0;
62 for (size_t i = 0; i < n_valtype_map; ++i) {
63 if (valtype_map[i].valtype == VT) {
64 retval = valtype_map + i;
71 cerr << "getValueTypeMapEntry returns NULL for "
72 << MVT::getValueTypeString(VT)
81 //! Predicate that returns true if operand is a memory target
83 \arg Op Operand to test
84 \return true if the operand is a memory target (i.e., global
85 address, external symbol, constant pool) or an A-form
88 bool isMemoryOperand(const SDOperand &Op)
90 const unsigned Opc = Op.getOpcode();
91 return (Opc == ISD::GlobalAddress
92 || Opc == ISD::GlobalTLSAddress
93 || Opc == ISD::JumpTable
94 || Opc == ISD::ConstantPool
95 || Opc == ISD::ExternalSymbol
96 || Opc == ISD::TargetGlobalAddress
97 || Opc == ISD::TargetGlobalTLSAddress
98 || Opc == ISD::TargetJumpTable
99 || Opc == ISD::TargetConstantPool
100 || Opc == ISD::TargetExternalSymbol
101 || Opc == SPUISD::AFormAddr);
104 //! Predicate that returns true if the operand is an indirect target
105 bool isIndirectOperand(const SDOperand &Op)
107 const unsigned Opc = Op.getOpcode();
108 return (Opc == ISD::Register
109 || Opc == SPUISD::LDRESULT);
113 SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
114 : TargetLowering(TM),
117 // Fold away setcc operations if possible.
120 // Use _setjmp/_longjmp instead of setjmp/longjmp.
121 setUseUnderscoreSetJmp(true);
122 setUseUnderscoreLongJmp(true);
124 // Set up the SPU's register classes:
125 // NOTE: i8 register class is not registered because we cannot determine when
126 // we need to zero or sign extend for custom-lowered loads and stores.
127 // NOTE: Ignore the previous note. For now. :-)
128 addRegisterClass(MVT::i8, SPU::R8CRegisterClass);
129 addRegisterClass(MVT::i16, SPU::R16CRegisterClass);
130 addRegisterClass(MVT::i32, SPU::R32CRegisterClass);
131 addRegisterClass(MVT::i64, SPU::R64CRegisterClass);
132 addRegisterClass(MVT::f32, SPU::R32FPRegisterClass);
133 addRegisterClass(MVT::f64, SPU::R64FPRegisterClass);
134 addRegisterClass(MVT::i128, SPU::GPRCRegisterClass);
136 // SPU has no sign or zero extended loads for i1, i8, i16:
137 setLoadXAction(ISD::EXTLOAD, MVT::i1, Promote);
138 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
139 setLoadXAction(ISD::ZEXTLOAD, MVT::i1, Promote);
140 setTruncStoreAction(MVT::i8, MVT::i1, Custom);
141 setTruncStoreAction(MVT::i16, MVT::i1, Custom);
142 setTruncStoreAction(MVT::i32, MVT::i1, Custom);
143 setTruncStoreAction(MVT::i64, MVT::i1, Custom);
144 setTruncStoreAction(MVT::i128, MVT::i1, Custom);
146 setLoadXAction(ISD::EXTLOAD, MVT::i8, Custom);
147 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Custom);
148 setLoadXAction(ISD::ZEXTLOAD, MVT::i8, Custom);
149 setTruncStoreAction(MVT::i8 , MVT::i8, Custom);
150 setTruncStoreAction(MVT::i16 , MVT::i8, Custom);
151 setTruncStoreAction(MVT::i32 , MVT::i8, Custom);
152 setTruncStoreAction(MVT::i64 , MVT::i8, Custom);
153 setTruncStoreAction(MVT::i128, MVT::i8, Custom);
155 setLoadXAction(ISD::EXTLOAD, MVT::i16, Custom);
156 setLoadXAction(ISD::SEXTLOAD, MVT::i16, Custom);
157 setLoadXAction(ISD::ZEXTLOAD, MVT::i16, Custom);
159 // SPU constant load actions are custom lowered:
160 setOperationAction(ISD::Constant, MVT::i64, Custom);
161 setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
162 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
164 // SPU's loads and stores have to be custom lowered:
165 for (unsigned sctype = (unsigned) MVT::i1; sctype < (unsigned) MVT::f128;
167 setOperationAction(ISD::LOAD, sctype, Custom);
168 setOperationAction(ISD::STORE, sctype, Custom);
171 // Custom lower BRCOND for i1, i8 to "promote" the result to
172 // i32 and i16, respectively.
173 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
175 // Expand the jumptable branches
176 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
177 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
178 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
180 // SPU has no intrinsics for these particular operations:
181 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
182 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
183 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
185 // PowerPC has no SREM/UREM instructions
186 setOperationAction(ISD::SREM, MVT::i32, Expand);
187 setOperationAction(ISD::UREM, MVT::i32, Expand);
188 setOperationAction(ISD::SREM, MVT::i64, Expand);
189 setOperationAction(ISD::UREM, MVT::i64, Expand);
191 // We don't support sin/cos/sqrt/fmod
192 setOperationAction(ISD::FSIN , MVT::f64, Expand);
193 setOperationAction(ISD::FCOS , MVT::f64, Expand);
194 setOperationAction(ISD::FREM , MVT::f64, Expand);
195 setOperationAction(ISD::FSIN , MVT::f32, Expand);
196 setOperationAction(ISD::FCOS , MVT::f32, Expand);
197 setOperationAction(ISD::FREM , MVT::f32, Expand);
199 // If we're enabling GP optimizations, use hardware square root
200 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
201 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
203 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
204 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
206 // SPU can do rotate right and left, so legalize it... but customize for i8
207 // because instructions don't exist.
208 setOperationAction(ISD::ROTR, MVT::i32, Legal);
209 setOperationAction(ISD::ROTR, MVT::i16, Legal);
210 setOperationAction(ISD::ROTR, MVT::i8, Custom);
211 setOperationAction(ISD::ROTL, MVT::i32, Legal);
212 setOperationAction(ISD::ROTL, MVT::i16, Legal);
213 setOperationAction(ISD::ROTL, MVT::i8, Custom);
214 // SPU has no native version of shift left/right for i8
215 setOperationAction(ISD::SHL, MVT::i8, Custom);
216 setOperationAction(ISD::SRL, MVT::i8, Custom);
217 setOperationAction(ISD::SRA, MVT::i8, Custom);
219 // Custom lower i32 multiplications
220 setOperationAction(ISD::MUL, MVT::i32, Custom);
222 // Need to custom handle (some) common i8 math ops
223 setOperationAction(ISD::SUB, MVT::i8, Custom);
224 setOperationAction(ISD::MUL, MVT::i8, Custom);
226 // SPU does not have BSWAP. It does have i32 support CTLZ.
227 // CTPOP has to be custom lowered.
228 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
229 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
231 setOperationAction(ISD::CTPOP, MVT::i8, Custom);
232 setOperationAction(ISD::CTPOP, MVT::i16, Custom);
233 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
234 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
236 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
237 setOperationAction(ISD::CTTZ , MVT::i64, Expand);
239 setOperationAction(ISD::CTLZ , MVT::i32, Legal);
241 // SPU does not have select or setcc
242 setOperationAction(ISD::SELECT, MVT::i1, Expand);
243 setOperationAction(ISD::SELECT, MVT::i8, Expand);
244 setOperationAction(ISD::SELECT, MVT::i16, Expand);
245 setOperationAction(ISD::SELECT, MVT::i32, Expand);
246 setOperationAction(ISD::SELECT, MVT::i64, Expand);
247 setOperationAction(ISD::SELECT, MVT::f32, Expand);
248 setOperationAction(ISD::SELECT, MVT::f64, Expand);
250 setOperationAction(ISD::SETCC, MVT::i1, Expand);
251 setOperationAction(ISD::SETCC, MVT::i8, Expand);
252 setOperationAction(ISD::SETCC, MVT::i16, Expand);
253 setOperationAction(ISD::SETCC, MVT::i32, Expand);
254 setOperationAction(ISD::SETCC, MVT::i64, Expand);
255 setOperationAction(ISD::SETCC, MVT::f32, Expand);
256 setOperationAction(ISD::SETCC, MVT::f64, Expand);
258 // SPU has a legal FP -> signed INT instruction
259 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
260 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
261 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
262 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
264 // FDIV on SPU requires custom lowering
265 setOperationAction(ISD::FDIV, MVT::f32, Custom);
266 //setOperationAction(ISD::FDIV, MVT::f64, Custom);
268 // SPU has [U|S]INT_TO_FP
269 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
270 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
271 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
272 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
273 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
274 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
275 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
276 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
278 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Legal);
279 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Legal);
280 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Legal);
281 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Legal);
283 // We cannot sextinreg(i1). Expand to shifts.
284 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
286 // Support label based line numbers.
287 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
288 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
290 // We want to legalize GlobalAddress and ConstantPool nodes into the
291 // appropriate instructions to materialize the address.
292 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
293 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
294 setOperationAction(ISD::ConstantPool, MVT::f32, Custom);
295 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
296 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
297 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
298 setOperationAction(ISD::ConstantPool, MVT::f64, Custom);
299 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
301 // RET must be custom lowered, to meet ABI requirements
302 setOperationAction(ISD::RET, MVT::Other, Custom);
304 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
305 setOperationAction(ISD::VASTART , MVT::Other, Custom);
307 // Use the default implementation.
308 setOperationAction(ISD::VAARG , MVT::Other, Expand);
309 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
310 setOperationAction(ISD::VAEND , MVT::Other, Expand);
311 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
312 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
313 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
314 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Expand);
316 // Cell SPU has instructions for converting between i64 and fp.
317 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
318 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
320 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
321 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
323 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
324 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
326 // First set operation action for all vector types to expand. Then we
327 // will selectively turn on ones that can be effectively codegen'd.
328 addRegisterClass(MVT::v16i8, SPU::VECREGRegisterClass);
329 addRegisterClass(MVT::v8i16, SPU::VECREGRegisterClass);
330 addRegisterClass(MVT::v4i32, SPU::VECREGRegisterClass);
331 addRegisterClass(MVT::v2i64, SPU::VECREGRegisterClass);
332 addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
333 addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
335 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
336 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
337 // add/sub are legal for all supported vector VT's.
338 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
339 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
340 // mul has to be custom lowered.
341 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Custom);
343 setOperationAction(ISD::AND , (MVT::ValueType)VT, Legal);
344 setOperationAction(ISD::OR , (MVT::ValueType)VT, Legal);
345 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Legal);
346 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Legal);
347 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Legal);
348 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Legal);
350 // These operations need to be expanded:
351 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
352 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
353 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
354 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
355 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Custom);
357 // Custom lower build_vector, constant pool spills, insert and
358 // extract vector elements:
359 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
360 setOperationAction(ISD::ConstantPool, (MVT::ValueType)VT, Custom);
361 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Custom);
362 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
363 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
364 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
367 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
368 setOperationAction(ISD::AND, MVT::v16i8, Custom);
369 setOperationAction(ISD::OR, MVT::v16i8, Custom);
370 setOperationAction(ISD::XOR, MVT::v16i8, Custom);
371 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
373 setSetCCResultType(MVT::i32);
374 setShiftAmountType(MVT::i32);
375 setSetCCResultContents(ZeroOrOneSetCCResult);
377 setStackPointerRegisterToSaveRestore(SPU::R1);
379 // We have target-specific dag combine patterns for the following nodes:
380 // e.g., setTargetDAGCombine(ISD::SUB);
382 computeRegisterProperties();
386 SPUTargetLowering::getTargetNodeName(unsigned Opcode) const
388 if (node_names.empty()) {
389 node_names[(unsigned) SPUISD::RET_FLAG] = "SPUISD::RET_FLAG";
390 node_names[(unsigned) SPUISD::Hi] = "SPUISD::Hi";
391 node_names[(unsigned) SPUISD::Lo] = "SPUISD::Lo";
392 node_names[(unsigned) SPUISD::PCRelAddr] = "SPUISD::PCRelAddr";
393 node_names[(unsigned) SPUISD::AFormAddr] = "SPUISD::AFormAddr";
394 node_names[(unsigned) SPUISD::DFormAddr] = "SPUISD::DFormAddr";
395 node_names[(unsigned) SPUISD::XFormAddr] = "SPUISD::XFormAddr";
396 node_names[(unsigned) SPUISD::LDRESULT] = "SPUISD::LDRESULT";
397 node_names[(unsigned) SPUISD::CALL] = "SPUISD::CALL";
398 node_names[(unsigned) SPUISD::SHUFB] = "SPUISD::SHUFB";
399 node_names[(unsigned) SPUISD::INSERT_MASK] = "SPUISD::INSERT_MASK";
400 node_names[(unsigned) SPUISD::CNTB] = "SPUISD::CNTB";
401 node_names[(unsigned) SPUISD::PROMOTE_SCALAR] = "SPUISD::PROMOTE_SCALAR";
402 node_names[(unsigned) SPUISD::EXTRACT_ELT0] = "SPUISD::EXTRACT_ELT0";
403 node_names[(unsigned) SPUISD::EXTRACT_ELT0_CHAINED] = "SPUISD::EXTRACT_ELT0_CHAINED";
404 node_names[(unsigned) SPUISD::EXTRACT_I1_ZEXT] = "SPUISD::EXTRACT_I1_ZEXT";
405 node_names[(unsigned) SPUISD::EXTRACT_I1_SEXT] = "SPUISD::EXTRACT_I1_SEXT";
406 node_names[(unsigned) SPUISD::EXTRACT_I8_ZEXT] = "SPUISD::EXTRACT_I8_ZEXT";
407 node_names[(unsigned) SPUISD::EXTRACT_I8_SEXT] = "SPUISD::EXTRACT_I8_SEXT";
408 node_names[(unsigned) SPUISD::MPY] = "SPUISD::MPY";
409 node_names[(unsigned) SPUISD::MPYU] = "SPUISD::MPYU";
410 node_names[(unsigned) SPUISD::MPYH] = "SPUISD::MPYH";
411 node_names[(unsigned) SPUISD::MPYHH] = "SPUISD::MPYHH";
412 node_names[(unsigned) SPUISD::VEC_SHL] = "SPUISD::VEC_SHL";
413 node_names[(unsigned) SPUISD::VEC_SRL] = "SPUISD::VEC_SRL";
414 node_names[(unsigned) SPUISD::VEC_SRA] = "SPUISD::VEC_SRA";
415 node_names[(unsigned) SPUISD::VEC_ROTL] = "SPUISD::VEC_ROTL";
416 node_names[(unsigned) SPUISD::VEC_ROTR] = "SPUISD::VEC_ROTR";
417 node_names[(unsigned) SPUISD::ROTBYTES_RIGHT_Z] =
418 "SPUISD::ROTBYTES_RIGHT_Z";
419 node_names[(unsigned) SPUISD::ROTBYTES_RIGHT_S] =
420 "SPUISD::ROTBYTES_RIGHT_S";
421 node_names[(unsigned) SPUISD::ROTBYTES_LEFT] = "SPUISD::ROTBYTES_LEFT";
422 node_names[(unsigned) SPUISD::ROTBYTES_LEFT_CHAINED] =
423 "SPUISD::ROTBYTES_LEFT_CHAINED";
424 node_names[(unsigned) SPUISD::FSMBI] = "SPUISD::FSMBI";
425 node_names[(unsigned) SPUISD::SELB] = "SPUISD::SELB";
426 node_names[(unsigned) SPUISD::SFPConstant] = "SPUISD::SFPConstant";
427 node_names[(unsigned) SPUISD::FPInterp] = "SPUISD::FPInterp";
428 node_names[(unsigned) SPUISD::FPRecipEst] = "SPUISD::FPRecipEst";
429 node_names[(unsigned) SPUISD::SEXT32TO64] = "SPUISD::SEXT32TO64";
432 std::map<unsigned, const char *>::iterator i = node_names.find(Opcode);
434 return ((i != node_names.end()) ? i->second : 0);
437 //===----------------------------------------------------------------------===//
438 // Calling convention code:
439 //===----------------------------------------------------------------------===//
441 #include "SPUGenCallingConv.inc"
443 //===----------------------------------------------------------------------===//
444 // LowerOperation implementation
445 //===----------------------------------------------------------------------===//
447 /// Aligned load common code for CellSPU
449 \param[in] Op The SelectionDAG load or store operand
450 \param[in] DAG The selection DAG
451 \param[in] ST CellSPU subtarget information structure
452 \param[in,out] alignment Caller initializes this to the load or store node's
453 value from getAlignment(), may be updated while generating the aligned load
454 \param[in,out] alignOffs Aligned offset; set by AlignedLoad to the aligned
455 offset (divisible by 16, modulo 16 == 0)
456 \param[in,out] prefSlotOffs Preferred slot offset; set by AlignedLoad to the
457 offset of the preferred slot (modulo 16 != 0)
458 \param[in,out] VT Caller initializes this value type to the the load or store
459 node's loaded or stored value type; may be updated if an i1-extended load or
461 \param[out] was16aligned true if the base pointer had 16-byte alignment,
462 otherwise false. Can help to determine if the chunk needs to be rotated.
464 Both load and store lowering load a block of data aligned on a 16-byte
465 boundary. This is the common aligned load code shared between both.
468 AlignedLoad(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST,
470 unsigned &alignment, int &alignOffs, int &prefSlotOffs,
471 MVT::ValueType &VT, bool &was16aligned)
473 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
474 const valtype_map_s *vtm = getValueTypeMapEntry(VT);
475 SDOperand basePtr = LSN->getBasePtr();
476 SDOperand chain = LSN->getChain();
478 if (basePtr.getOpcode() == ISD::ADD) {
479 SDOperand Op1 = basePtr.Val->getOperand(1);
481 if (Op1.getOpcode() == ISD::Constant || Op1.getOpcode() == ISD::TargetConstant) {
482 const ConstantSDNode *CN = cast<ConstantSDNode>(basePtr.getOperand(1));
484 alignOffs = (int) CN->getValue();
485 prefSlotOffs = (int) (alignOffs & 0xf);
487 // Adjust the rotation amount to ensure that the final result ends up in
488 // the preferred slot:
489 prefSlotOffs -= vtm->prefslot_byte;
490 basePtr = basePtr.getOperand(0);
492 // Loading from memory, can we adjust alignment?
493 if (basePtr.getOpcode() == SPUISD::AFormAddr) {
494 SDOperand APtr = basePtr.getOperand(0);
495 if (APtr.getOpcode() == ISD::TargetGlobalAddress) {
496 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(APtr);
497 alignment = GSDN->getGlobal()->getAlignment();
502 prefSlotOffs = -vtm->prefslot_byte;
506 prefSlotOffs = -vtm->prefslot_byte;
509 if (alignment == 16) {
510 // Realign the base pointer as a D-Form address:
511 if (!isMemoryOperand(basePtr) || (alignOffs & ~0xf) != 0) {
512 basePtr = DAG.getNode(ISD::ADD, PtrVT,
514 DAG.getConstant((alignOffs & ~0xf), PtrVT));
517 // Emit the vector load:
519 return DAG.getLoad(MVT::v16i8, chain, basePtr,
520 LSN->getSrcValue(), LSN->getSrcValueOffset(),
521 LSN->isVolatile(), 16);
524 // Unaligned load or we're using the "large memory" model, which means that
525 // we have to be very pessimistic:
526 if (isMemoryOperand(basePtr) || isIndirectOperand(basePtr)) {
527 basePtr = DAG.getNode(SPUISD::XFormAddr, PtrVT, basePtr, DAG.getConstant(0, PtrVT));
531 basePtr = DAG.getNode(ISD::ADD, PtrVT, basePtr, DAG.getConstant(alignOffs, PtrVT));
532 was16aligned = false;
533 return DAG.getLoad(MVT::v16i8, chain, basePtr,
534 LSN->getSrcValue(), LSN->getSrcValueOffset(),
535 LSN->isVolatile(), 16);
538 /// Custom lower loads for CellSPU
540 All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
541 within a 16-byte block, we have to rotate to extract the requested element.
544 LowerLOAD(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
545 LoadSDNode *LN = cast<LoadSDNode>(Op);
546 SDOperand the_chain = LN->getChain();
547 MVT::ValueType VT = LN->getLoadedVT();
548 MVT::ValueType OpVT = Op.Val->getValueType(0);
549 ISD::LoadExtType ExtType = LN->getExtensionType();
550 unsigned alignment = LN->getAlignment();
553 switch (LN->getAddressingMode()) {
554 case ISD::UNINDEXED: {
558 AlignedLoad(Op, DAG, ST, LN,alignment, offset, rotamt, VT, was16aligned);
563 the_chain = result.getValue(1);
564 // Rotate the chunk if necessary
567 if (rotamt != 0 || !was16aligned) {
568 SDVTList vecvts = DAG.getVTList(MVT::v16i8, MVT::Other);
573 Ops[2] = DAG.getConstant(rotamt, MVT::i16);
575 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
576 LoadSDNode *LN1 = cast<LoadSDNode>(result);
577 Ops[2] = DAG.getNode(ISD::ADD, PtrVT, LN1->getBasePtr(),
578 DAG.getConstant(rotamt, PtrVT));
581 result = DAG.getNode(SPUISD::ROTBYTES_LEFT_CHAINED, vecvts, Ops, 3);
582 the_chain = result.getValue(1);
585 if (VT == OpVT || ExtType == ISD::EXTLOAD) {
587 MVT::ValueType vecVT = MVT::v16i8;
589 // Convert the loaded v16i8 vector to the appropriate vector type
590 // specified by the operand:
593 vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
595 vecVT = MVT::getVectorType(OpVT, (128 / MVT::getSizeInBits(OpVT)));
598 Ops[1] = DAG.getNode(ISD::BIT_CONVERT, vecVT, result);
599 scalarvts = DAG.getVTList((OpVT == VT ? VT : OpVT), MVT::Other);
600 result = DAG.getNode(SPUISD::EXTRACT_ELT0_CHAINED, scalarvts, Ops, 2);
601 the_chain = result.getValue(1);
603 // Handle the sign and zero-extending loads for i1 and i8:
606 if (ExtType == ISD::SEXTLOAD) {
607 NewOpC = (OpVT == MVT::i1
608 ? SPUISD::EXTRACT_I1_SEXT
609 : SPUISD::EXTRACT_I8_SEXT);
611 assert(ExtType == ISD::ZEXTLOAD);
612 NewOpC = (OpVT == MVT::i1
613 ? SPUISD::EXTRACT_I1_ZEXT
614 : SPUISD::EXTRACT_I8_ZEXT);
617 result = DAG.getNode(NewOpC, OpVT, result);
620 SDVTList retvts = DAG.getVTList(OpVT, MVT::Other);
621 SDOperand retops[3] = {
624 DAG.getConstant(alignment, MVT::i32)
627 result = DAG.getNode(SPUISD::LDRESULT, retvts,
628 retops, sizeof(retops) / sizeof(retops[0]));
635 case ISD::LAST_INDEXED_MODE:
636 cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
638 cerr << (unsigned) LN->getAddressingMode() << "\n";
646 /// Custom lower stores for CellSPU
648 All CellSPU stores are aligned to 16-byte boundaries, so for elements
649 within a 16-byte block, we have to generate a shuffle to insert the
650 requested element into its place, then store the resulting block.
653 LowerSTORE(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
654 StoreSDNode *SN = cast<StoreSDNode>(Op);
655 SDOperand Value = SN->getValue();
656 MVT::ValueType VT = Value.getValueType();
657 MVT::ValueType StVT = (!SN->isTruncatingStore() ? VT : SN->getStoredVT());
658 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
659 unsigned alignment = SN->getAlignment();
661 switch (SN->getAddressingMode()) {
662 case ISD::UNINDEXED: {
663 int chunk_offset, slot_offset;
666 // The vector type we really want to load from the 16-byte chunk, except
667 // in the case of MVT::i1, which has to be v16i8.
668 unsigned vecVT, stVecVT = MVT::v16i8;
671 stVecVT = MVT::getVectorType(StVT, (128 / MVT::getSizeInBits(StVT)));
672 vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
674 SDOperand alignLoadVec =
675 AlignedLoad(Op, DAG, ST, SN, alignment,
676 chunk_offset, slot_offset, VT, was16aligned);
678 if (alignLoadVec.Val == 0)
681 LoadSDNode *LN = cast<LoadSDNode>(alignLoadVec);
682 SDOperand basePtr = LN->getBasePtr();
683 SDOperand the_chain = alignLoadVec.getValue(1);
684 SDOperand theValue = SN->getValue();
688 && (theValue.getOpcode() == ISD::AssertZext
689 || theValue.getOpcode() == ISD::AssertSext)) {
690 // Drill down and get the value for zero- and sign-extended
692 theValue = theValue.getOperand(0);
697 SDOperand insertEltOffs = DAG.getConstant(chunk_offset, PtrVT);
698 SDOperand insertEltPtr;
699 SDOperand insertEltOp;
701 // If the base pointer is already a D-form address, then just create
702 // a new D-form address with a slot offset and the orignal base pointer.
703 // Otherwise generate a D-form address with the slot offset relative
704 // to the stack pointer, which is always aligned.
705 DEBUG(cerr << "CellSPU LowerSTORE: basePtr = ");
706 DEBUG(basePtr.Val->dump(&DAG));
709 if (basePtr.getOpcode() == SPUISD::DFormAddr) {
710 // Hmmmm... do we ever actually hit this code?
711 insertEltPtr = DAG.getNode(SPUISD::DFormAddr, PtrVT,
712 basePtr.getOperand(0),
714 } else if (basePtr.getOpcode() == SPUISD::XFormAddr ||
715 (basePtr.getOpcode() == ISD::ADD
716 && basePtr.getOperand(0).getOpcode() == SPUISD::XFormAddr)) {
717 insertEltPtr = basePtr;
719 // $sp is always aligned, so use it instead of potentially loading an
720 // address into a new register:
721 insertEltPtr = DAG.getNode(SPUISD::DFormAddr, PtrVT,
722 DAG.getRegister(SPU::R1, PtrVT),
726 insertEltOp = DAG.getNode(SPUISD::INSERT_MASK, stVecVT, insertEltPtr);
727 result = DAG.getNode(SPUISD::SHUFB, vecVT,
728 DAG.getNode(ISD::SCALAR_TO_VECTOR, vecVT, theValue),
730 DAG.getNode(ISD::BIT_CONVERT, vecVT, insertEltOp));
732 result = DAG.getStore(the_chain, result, basePtr,
733 LN->getSrcValue(), LN->getSrcValueOffset(),
734 LN->isVolatile(), LN->getAlignment());
743 case ISD::LAST_INDEXED_MODE:
744 cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
746 cerr << (unsigned) SN->getAddressingMode() << "\n";
754 /// Generate the address of a constant pool entry.
756 LowerConstantPool(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
757 MVT::ValueType PtrVT = Op.getValueType();
758 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
759 Constant *C = CP->getConstVal();
760 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
761 SDOperand Zero = DAG.getConstant(0, PtrVT);
762 const TargetMachine &TM = DAG.getTarget();
764 if (TM.getRelocationModel() == Reloc::Static) {
765 if (!ST->usingLargeMem()) {
766 // Just return the SDOperand with the constant pool address in it.
767 return DAG.getNode(SPUISD::AFormAddr, PtrVT, CPI, Zero);
770 SDOperand Hi = DAG.getNode(SPUISD::Hi, PtrVT, CPI, Zero);
771 SDOperand Lo = DAG.getNode(SPUISD::Lo, PtrVT, CPI, Zero);
773 return DAG.getNode(ISD::ADD, PtrVT, Lo, Hi);
775 return DAG.getNode(SPUISD::XFormAddr, PtrVT, CPI, Zero);
781 "LowerConstantPool: Relocation model other than static not supported.");
786 LowerJumpTable(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
787 MVT::ValueType PtrVT = Op.getValueType();
788 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
789 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
790 SDOperand Zero = DAG.getConstant(0, PtrVT);
791 const TargetMachine &TM = DAG.getTarget();
793 if (TM.getRelocationModel() == Reloc::Static) {
794 return (!ST->usingLargeMem()
795 ? DAG.getNode(SPUISD::AFormAddr, PtrVT, JTI, Zero)
796 : DAG.getNode(SPUISD::XFormAddr, PtrVT, JTI, Zero));
800 "LowerJumpTable: Relocation model other than static not supported.");
805 LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
806 MVT::ValueType PtrVT = Op.getValueType();
807 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
808 GlobalValue *GV = GSDN->getGlobal();
809 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
810 const TargetMachine &TM = DAG.getTarget();
811 SDOperand Zero = DAG.getConstant(0, PtrVT);
813 if (TM.getRelocationModel() == Reloc::Static) {
814 return (!ST->usingLargeMem()
815 ? DAG.getNode(SPUISD::AFormAddr, PtrVT, GA, Zero)
816 : DAG.getNode(SPUISD::XFormAddr, PtrVT, GA, Zero));
818 cerr << "LowerGlobalAddress: Relocation model other than static not "
827 //! Custom lower i64 integer constants
829 This code inserts all of the necessary juggling that needs to occur to load
830 a 64-bit constant into a register.
833 LowerConstant(SDOperand Op, SelectionDAG &DAG) {
834 unsigned VT = Op.getValueType();
835 ConstantSDNode *CN = cast<ConstantSDNode>(Op.Val);
837 if (VT == MVT::i64) {
838 SDOperand T = DAG.getConstant(CN->getValue(), MVT::i64);
839 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
840 DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
843 cerr << "LowerConstant: unhandled constant type "
844 << MVT::getValueTypeString(VT)
853 //! Custom lower single precision floating point constants
855 "float" immediates can be lowered as if they were unsigned 32-bit integers.
856 The SPUISD::SFPConstant pseudo-instruction handles this in the instruction
860 LowerConstantFP(SDOperand Op, SelectionDAG &DAG) {
861 unsigned VT = Op.getValueType();
862 ConstantFPSDNode *FP = cast<ConstantFPSDNode>(Op.Val);
865 "LowerConstantFP: Node is not ConstantFPSDNode");
867 if (VT == MVT::f32) {
868 float targetConst = FP->getValueAPF().convertToFloat();
869 return DAG.getNode(SPUISD::SFPConstant, VT,
870 DAG.getTargetConstantFP(targetConst, VT));
871 } else if (VT == MVT::f64) {
872 uint64_t dbits = DoubleToBits(FP->getValueAPF().convertToDouble());
873 return DAG.getNode(ISD::BIT_CONVERT, VT,
874 LowerConstant(DAG.getConstant(dbits, MVT::i64), DAG));
880 //! Lower MVT::i1, MVT::i8 brcond to a promoted type (MVT::i32, MVT::i16)
882 LowerBRCOND(SDOperand Op, SelectionDAG &DAG)
884 SDOperand Cond = Op.getOperand(1);
885 MVT::ValueType CondVT = Cond.getValueType();
886 MVT::ValueType CondNVT;
888 if (CondVT == MVT::i1 || CondVT == MVT::i8) {
889 CondNVT = (CondVT == MVT::i1 ? MVT::i32 : MVT::i16);
890 return DAG.getNode(ISD::BRCOND, Op.getValueType(),
892 DAG.getNode(ISD::ZERO_EXTEND, CondNVT, Op.getOperand(1)),
895 return SDOperand(); // Unchanged
899 LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG, int &VarArgsFrameIndex)
901 MachineFunction &MF = DAG.getMachineFunction();
902 MachineFrameInfo *MFI = MF.getFrameInfo();
903 MachineRegisterInfo &RegInfo = MF.getRegInfo();
904 SmallVector<SDOperand, 8> ArgValues;
905 SDOperand Root = Op.getOperand(0);
906 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
908 const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
909 const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
911 unsigned ArgOffset = SPUFrameInfo::minStackSize();
912 unsigned ArgRegIdx = 0;
913 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
915 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
917 // Add DAG nodes to load the arguments or copy them out of registers.
918 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
920 bool needsLoad = false;
921 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
922 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
926 cerr << "LowerFORMAL_ARGUMENTS Unhandled argument type: "
927 << MVT::getValueTypeString(ObjectVT)
932 if (!isVarArg && ArgRegIdx < NumArgRegs) {
933 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R8CRegClass);
934 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
935 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i8);
942 if (!isVarArg && ArgRegIdx < NumArgRegs) {
943 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
944 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
945 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i16);
952 if (!isVarArg && ArgRegIdx < NumArgRegs) {
953 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
954 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
955 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
962 if (!isVarArg && ArgRegIdx < NumArgRegs) {
963 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64CRegClass);
964 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
965 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
972 if (!isVarArg && ArgRegIdx < NumArgRegs) {
973 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
974 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
975 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f32);
982 if (!isVarArg && ArgRegIdx < NumArgRegs) {
983 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64FPRegClass);
984 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
985 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f64);
996 if (!isVarArg && ArgRegIdx < NumArgRegs) {
997 unsigned VReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
998 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
999 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1007 // We need to load the argument to a virtual register if we determined above
1008 // that we ran out of physical registers of the appropriate type
1010 // If the argument is actually used, emit a load from the right stack
1012 if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
1013 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1014 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1015 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1017 // Don't emit a dead load.
1018 ArgVal = DAG.getNode(ISD::UNDEF, ObjectVT);
1021 ArgOffset += StackSlotSize;
1024 ArgValues.push_back(ArgVal);
1027 // If the function takes variable number of arguments, make a frame index for
1028 // the start of the first vararg value... for expansion of llvm.va_start.
1030 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1032 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1033 // If this function is vararg, store any remaining integer argument regs to
1034 // their spots on the stack so that they may be loaded by deferencing the
1035 // result of va_next.
1036 SmallVector<SDOperand, 8> MemOps;
1037 for (; ArgRegIdx != NumArgRegs; ++ArgRegIdx) {
1038 unsigned VReg = RegInfo.createVirtualRegister(&SPU::GPRCRegClass);
1039 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
1040 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1041 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1042 MemOps.push_back(Store);
1043 // Increment the address by four for the next argument to store
1044 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1045 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1047 if (!MemOps.empty())
1048 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1051 ArgValues.push_back(Root);
1053 // Return the new list of results.
1054 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1055 Op.Val->value_end());
1056 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1059 /// isLSAAddress - Return the immediate to use if the specified
1060 /// value is representable as a LSA address.
1061 static SDNode *isLSAAddress(SDOperand Op, SelectionDAG &DAG) {
1062 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1065 int Addr = C->getValue();
1066 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1067 (Addr << 14 >> 14) != Addr)
1068 return 0; // Top 14 bits have to be sext of immediate.
1070 return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).Val;
1075 LowerCALL(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1076 SDOperand Chain = Op.getOperand(0);
1078 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1079 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1081 SDOperand Callee = Op.getOperand(4);
1082 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1083 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
1084 const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
1085 const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
1087 // Handy pointer type
1088 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1090 // Accumulate how many bytes are to be pushed on the stack, including the
1091 // linkage area, and parameter passing area. According to the SPU ABI,
1092 // we minimally need space for [LR] and [SP]
1093 unsigned NumStackBytes = SPUFrameInfo::minStackSize();
1095 // Set up a copy of the stack pointer for use loading and storing any
1096 // arguments that may not fit in the registers available for argument
1098 SDOperand StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
1100 // Figure out which arguments are going to go in registers, and which in
1102 unsigned ArgOffset = SPUFrameInfo::minStackSize(); // Just below [LR]
1103 unsigned ArgRegIdx = 0;
1105 // Keep track of registers passing arguments
1106 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1107 // And the arguments passed on the stack
1108 SmallVector<SDOperand, 8> MemOpChains;
1110 for (unsigned i = 0; i != NumOps; ++i) {
1111 SDOperand Arg = Op.getOperand(5+2*i);
1113 // PtrOff will be used to store the current argument to the stack if a
1114 // register cannot be found for it.
1115 SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1116 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1118 switch (Arg.getValueType()) {
1119 default: assert(0 && "Unexpected ValueType for argument!");
1123 if (ArgRegIdx != NumArgRegs) {
1124 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1126 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1127 ArgOffset += StackSlotSize;
1132 if (ArgRegIdx != NumArgRegs) {
1133 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1135 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1136 ArgOffset += StackSlotSize;
1143 if (ArgRegIdx != NumArgRegs) {
1144 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1146 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1147 ArgOffset += StackSlotSize;
1153 // Update number of stack bytes actually used, insert a call sequence start
1154 NumStackBytes = (ArgOffset - SPUFrameInfo::minStackSize());
1155 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumStackBytes, PtrVT));
1157 if (!MemOpChains.empty()) {
1158 // Adjust the stack pointer for the stack arguments.
1159 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1160 &MemOpChains[0], MemOpChains.size());
1163 // Build a sequence of copy-to-reg nodes chained together with token chain
1164 // and flag operands which copy the outgoing args into the appropriate regs.
1166 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1167 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1169 InFlag = Chain.getValue(1);
1172 std::vector<MVT::ValueType> NodeTys;
1173 NodeTys.push_back(MVT::Other); // Returns a chain
1174 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1176 SmallVector<SDOperand, 8> Ops;
1177 unsigned CallOpc = SPUISD::CALL;
1179 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1180 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1181 // node so that legalize doesn't hack it.
1182 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1183 GlobalValue *GV = G->getGlobal();
1184 unsigned CalleeVT = Callee.getValueType();
1185 SDOperand Zero = DAG.getConstant(0, PtrVT);
1186 SDOperand GA = DAG.getTargetGlobalAddress(GV, CalleeVT);
1188 if (!ST->usingLargeMem()) {
1189 // Turn calls to targets that are defined (i.e., have bodies) into BRSL
1190 // style calls, otherwise, external symbols are BRASL calls. This assumes
1191 // that declared/defined symbols are in the same compilation unit and can
1192 // be reached through PC-relative jumps.
1195 // This may be an unsafe assumption for JIT and really large compilation
1197 if (GV->isDeclaration()) {
1198 Callee = DAG.getNode(SPUISD::AFormAddr, CalleeVT, GA, Zero);
1200 Callee = DAG.getNode(SPUISD::PCRelAddr, CalleeVT, GA, Zero);
1203 // "Large memory" mode: Turn all calls into indirect calls with a X-form
1205 Callee = DAG.getNode(SPUISD::XFormAddr, PtrVT, GA, Zero);
1207 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1208 Callee = DAG.getExternalSymbol(S->getSymbol(), Callee.getValueType());
1209 else if (SDNode *Dest = isLSAAddress(Callee, DAG)) {
1210 // If this is an absolute destination address that appears to be a legal
1211 // local store address, use the munged value.
1212 Callee = SDOperand(Dest, 0);
1215 Ops.push_back(Chain);
1216 Ops.push_back(Callee);
1218 // Add argument registers to the end of the list so that they are known live
1220 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1221 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1222 RegsToPass[i].second.getValueType()));
1225 Ops.push_back(InFlag);
1226 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1227 InFlag = Chain.getValue(1);
1229 SDOperand ResultVals[3];
1230 unsigned NumResults = 0;
1233 // If the call has results, copy the values out of the ret val registers.
1234 switch (Op.Val->getValueType(0)) {
1235 default: assert(0 && "Unexpected ret value!");
1236 case MVT::Other: break;
1238 if (Op.Val->getValueType(1) == MVT::i32) {
1239 Chain = DAG.getCopyFromReg(Chain, SPU::R4, MVT::i32, InFlag).getValue(1);
1240 ResultVals[0] = Chain.getValue(0);
1241 Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i32,
1242 Chain.getValue(2)).getValue(1);
1243 ResultVals[1] = Chain.getValue(0);
1245 NodeTys.push_back(MVT::i32);
1247 Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i32, InFlag).getValue(1);
1248 ResultVals[0] = Chain.getValue(0);
1251 NodeTys.push_back(MVT::i32);
1254 Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i64, InFlag).getValue(1);
1255 ResultVals[0] = Chain.getValue(0);
1257 NodeTys.push_back(MVT::i64);
1261 Chain = DAG.getCopyFromReg(Chain, SPU::R3, Op.Val->getValueType(0),
1262 InFlag).getValue(1);
1263 ResultVals[0] = Chain.getValue(0);
1265 NodeTys.push_back(Op.Val->getValueType(0));
1272 Chain = DAG.getCopyFromReg(Chain, SPU::R3, Op.Val->getValueType(0),
1273 InFlag).getValue(1);
1274 ResultVals[0] = Chain.getValue(0);
1276 NodeTys.push_back(Op.Val->getValueType(0));
1280 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
1281 DAG.getConstant(NumStackBytes, PtrVT));
1282 NodeTys.push_back(MVT::Other);
1284 // If the function returns void, just return the chain.
1285 if (NumResults == 0)
1288 // Otherwise, merge everything together with a MERGE_VALUES node.
1289 ResultVals[NumResults++] = Chain;
1290 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1291 ResultVals, NumResults);
1292 return Res.getValue(Op.ResNo);
1296 LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1297 SmallVector<CCValAssign, 16> RVLocs;
1298 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1299 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1300 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1301 CCInfo.AnalyzeReturn(Op.Val, RetCC_SPU);
1303 // If this is the first return lowered for this function, add the regs to the
1304 // liveout set for the function.
1305 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1306 for (unsigned i = 0; i != RVLocs.size(); ++i)
1307 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1310 SDOperand Chain = Op.getOperand(0);
1313 // Copy the result values into the output registers.
1314 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1315 CCValAssign &VA = RVLocs[i];
1316 assert(VA.isRegLoc() && "Can only return in registers!");
1317 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1318 Flag = Chain.getValue(1);
1322 return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain, Flag);
1324 return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain);
1328 //===----------------------------------------------------------------------===//
1329 // Vector related lowering:
1330 //===----------------------------------------------------------------------===//
1332 static ConstantSDNode *
1333 getVecImm(SDNode *N) {
1334 SDOperand OpVal(0, 0);
1336 // Check to see if this buildvec has a single non-undef value in its elements.
1337 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1338 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1340 OpVal = N->getOperand(i);
1341 else if (OpVal != N->getOperand(i))
1345 if (OpVal.Val != 0) {
1346 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1351 return 0; // All UNDEF: use implicit def.; not Constant node
1354 /// get_vec_i18imm - Test if this vector is a vector filled with the same value
1355 /// and the value fits into an unsigned 18-bit constant, and if so, return the
1357 SDOperand SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
1358 MVT::ValueType ValueType) {
1359 if (ConstantSDNode *CN = getVecImm(N)) {
1360 uint64_t Value = CN->getValue();
1361 if (Value <= 0x3ffff)
1362 return DAG.getConstant(Value, ValueType);
1368 /// get_vec_i16imm - Test if this vector is a vector filled with the same value
1369 /// and the value fits into a signed 16-bit constant, and if so, return the
1371 SDOperand SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
1372 MVT::ValueType ValueType) {
1373 if (ConstantSDNode *CN = getVecImm(N)) {
1374 if (ValueType == MVT::i32) {
1375 int Value = (int) CN->getValue();
1376 int SExtValue = ((Value & 0xffff) << 16) >> 16;
1378 if (Value == SExtValue)
1379 return DAG.getConstant(Value, ValueType);
1380 } else if (ValueType == MVT::i16) {
1381 short Value = (short) CN->getValue();
1382 int SExtValue = ((int) Value << 16) >> 16;
1384 if (Value == (short) SExtValue)
1385 return DAG.getConstant(Value, ValueType);
1386 } else if (ValueType == MVT::i64) {
1387 int64_t Value = CN->getValue();
1388 int64_t SExtValue = ((Value & 0xffff) << (64 - 16)) >> (64 - 16);
1390 if (Value == SExtValue)
1391 return DAG.getConstant(Value, ValueType);
1398 /// get_vec_i10imm - Test if this vector is a vector filled with the same value
1399 /// and the value fits into a signed 10-bit constant, and if so, return the
1401 SDOperand SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
1402 MVT::ValueType ValueType) {
1403 if (ConstantSDNode *CN = getVecImm(N)) {
1404 int Value = (int) CN->getValue();
1405 if ((ValueType == MVT::i32 && isS10Constant(Value))
1406 || (ValueType == MVT::i16 && isS10Constant((short) Value)))
1407 return DAG.getConstant(Value, ValueType);
1413 /// get_vec_i8imm - Test if this vector is a vector filled with the same value
1414 /// and the value fits into a signed 8-bit constant, and if so, return the
1417 /// @note: The incoming vector is v16i8 because that's the only way we can load
1418 /// constant vectors. Thus, we test to see if the upper and lower bytes are the
1420 SDOperand SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
1421 MVT::ValueType ValueType) {
1422 if (ConstantSDNode *CN = getVecImm(N)) {
1423 int Value = (int) CN->getValue();
1424 if (ValueType == MVT::i16
1425 && Value <= 0xffff /* truncated from uint64_t */
1426 && ((short) Value >> 8) == ((short) Value & 0xff))
1427 return DAG.getConstant(Value & 0xff, ValueType);
1428 else if (ValueType == MVT::i8
1429 && (Value & 0xff) == Value)
1430 return DAG.getConstant(Value, ValueType);
1436 /// get_ILHUvec_imm - Test if this vector is a vector filled with the same value
1437 /// and the value fits into a signed 16-bit constant, and if so, return the
1439 SDOperand SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
1440 MVT::ValueType ValueType) {
1441 if (ConstantSDNode *CN = getVecImm(N)) {
1442 uint64_t Value = CN->getValue();
1443 if ((ValueType == MVT::i32
1444 && ((unsigned) Value & 0xffff0000) == (unsigned) Value)
1445 || (ValueType == MVT::i64 && (Value & 0xffff0000) == Value))
1446 return DAG.getConstant(Value >> 16, ValueType);
1452 /// get_v4i32_imm - Catch-all for general 32-bit constant vectors
1453 SDOperand SPU::get_v4i32_imm(SDNode *N, SelectionDAG &DAG) {
1454 if (ConstantSDNode *CN = getVecImm(N)) {
1455 return DAG.getConstant((unsigned) CN->getValue(), MVT::i32);
1461 /// get_v4i32_imm - Catch-all for general 64-bit constant vectors
1462 SDOperand SPU::get_v2i64_imm(SDNode *N, SelectionDAG &DAG) {
1463 if (ConstantSDNode *CN = getVecImm(N)) {
1464 return DAG.getConstant((unsigned) CN->getValue(), MVT::i64);
1470 // If this is a vector of constants or undefs, get the bits. A bit in
1471 // UndefBits is set if the corresponding element of the vector is an
1472 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1473 // zero. Return true if this is not an array of constants, false if it is.
1475 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
1476 uint64_t UndefBits[2]) {
1477 // Start with zero'd results.
1478 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
1480 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
1481 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
1482 SDOperand OpVal = BV->getOperand(i);
1484 unsigned PartNo = i >= e/2; // In the upper 128 bits?
1485 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
1487 uint64_t EltBits = 0;
1488 if (OpVal.getOpcode() == ISD::UNDEF) {
1489 uint64_t EltUndefBits = ~0ULL >> (64-EltBitSize);
1490 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
1492 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1493 EltBits = CN->getValue() & (~0ULL >> (64-EltBitSize));
1494 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
1495 const APFloat &apf = CN->getValueAPF();
1496 EltBits = (CN->getValueType(0) == MVT::f32
1497 ? FloatToBits(apf.convertToFloat())
1498 : DoubleToBits(apf.convertToDouble()));
1500 // Nonconstant element.
1504 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
1507 //printf("%llx %llx %llx %llx\n",
1508 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
1512 /// If this is a splat (repetition) of a value across the whole vector, return
1513 /// the smallest size that splats it. For example, "0x01010101010101..." is a
1514 /// splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
1515 /// SplatSize = 1 byte.
1516 static bool isConstantSplat(const uint64_t Bits128[2],
1517 const uint64_t Undef128[2],
1519 uint64_t &SplatBits, uint64_t &SplatUndef,
1521 // Don't let undefs prevent splats from matching. See if the top 64-bits are
1522 // the same as the lower 64-bits, ignoring undefs.
1523 uint64_t Bits64 = Bits128[0] | Bits128[1];
1524 uint64_t Undef64 = Undef128[0] & Undef128[1];
1525 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
1526 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
1527 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
1528 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
1530 if ((Bits128[0] & ~Undef128[1]) == (Bits128[1] & ~Undef128[0])) {
1531 if (MinSplatBits < 64) {
1533 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
1535 if ((Bits64 & (~Undef64 >> 32)) == ((Bits64 >> 32) & ~Undef64)) {
1536 if (MinSplatBits < 32) {
1538 // If the top 16-bits are different than the lower 16-bits, ignoring
1539 // undefs, we have an i32 splat.
1540 if ((Bits32 & (~Undef32 >> 16)) == ((Bits32 >> 16) & ~Undef32)) {
1541 if (MinSplatBits < 16) {
1542 // If the top 8-bits are different than the lower 8-bits, ignoring
1543 // undefs, we have an i16 splat.
1544 if ((Bits16 & (uint16_t(~Undef16) >> 8)) == ((Bits16 >> 8) & ~Undef16)) {
1545 // Otherwise, we have an 8-bit splat.
1546 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
1547 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
1553 SplatUndef = Undef16;
1560 SplatUndef = Undef32;
1566 SplatBits = Bits128[0];
1567 SplatUndef = Undef128[0];
1573 return false; // Can't be a splat if two pieces don't match.
1576 // If this is a case we can't handle, return null and let the default
1577 // expansion code take care of it. If we CAN select this case, and if it
1578 // selects to a single instruction, return Op. Otherwise, if we can codegen
1579 // this case more efficiently than a constant pool load, lower it to the
1580 // sequence of ops that should be used.
1581 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1582 MVT::ValueType VT = Op.getValueType();
1583 // If this is a vector of constants or undefs, get the bits. A bit in
1584 // UndefBits is set if the corresponding element of the vector is an
1585 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1587 uint64_t VectorBits[2];
1588 uint64_t UndefBits[2];
1589 uint64_t SplatBits, SplatUndef;
1591 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits)
1592 || !isConstantSplat(VectorBits, UndefBits,
1593 MVT::getSizeInBits(MVT::getVectorElementType(VT)),
1594 SplatBits, SplatUndef, SplatSize))
1595 return SDOperand(); // Not a constant vector, not a splat.
1600 uint32_t Value32 = SplatBits;
1601 assert(SplatSize == 4
1602 && "LowerBUILD_VECTOR: Unexpected floating point vector element.");
1603 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1604 SDOperand T = DAG.getConstant(Value32, MVT::i32);
1605 return DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32,
1606 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, T, T, T, T));
1610 uint64_t f64val = SplatBits;
1611 assert(SplatSize == 8
1612 && "LowerBUILD_VECTOR: 64-bit float vector element: unexpected size.");
1613 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1614 SDOperand T = DAG.getConstant(f64val, MVT::i64);
1615 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64,
1616 DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
1620 // 8-bit constants have to be expanded to 16-bits
1621 unsigned short Value16 = SplatBits | (SplatBits << 8);
1623 for (int i = 0; i < 8; ++i)
1624 Ops[i] = DAG.getConstant(Value16, MVT::i16);
1625 return DAG.getNode(ISD::BIT_CONVERT, VT,
1626 DAG.getNode(ISD::BUILD_VECTOR, MVT::v8i16, Ops, 8));
1629 unsigned short Value16;
1631 Value16 = (unsigned short) (SplatBits & 0xffff);
1633 Value16 = (unsigned short) (SplatBits | (SplatBits << 8));
1634 SDOperand T = DAG.getConstant(Value16, MVT::getVectorElementType(VT));
1636 for (int i = 0; i < 8; ++i) Ops[i] = T;
1637 return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops, 8);
1640 unsigned int Value = SplatBits;
1641 SDOperand T = DAG.getConstant(Value, MVT::getVectorElementType(VT));
1642 return DAG.getNode(ISD::BUILD_VECTOR, VT, T, T, T, T);
1645 uint64_t val = SplatBits;
1646 uint32_t upper = uint32_t(val >> 32);
1647 uint32_t lower = uint32_t(val);
1652 SmallVector<SDOperand, 16> ShufBytes;
1654 bool upper_special, lower_special;
1656 // NOTE: This code creates common-case shuffle masks that can be easily
1657 // detected as common expressions. It is not attempting to create highly
1658 // specialized masks to replace any and all 0's, 0xff's and 0x80's.
1660 // Detect if the upper or lower half is a special shuffle mask pattern:
1661 upper_special = (upper == 0 || upper == 0xffffffff || upper == 0x80000000);
1662 lower_special = (lower == 0 || lower == 0xffffffff || lower == 0x80000000);
1664 // Create lower vector if not a special pattern
1665 if (!lower_special) {
1666 SDOperand LO32C = DAG.getConstant(lower, MVT::i32);
1667 LO32 = DAG.getNode(ISD::BIT_CONVERT, VT,
1668 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1669 LO32C, LO32C, LO32C, LO32C));
1672 // Create upper vector if not a special pattern
1673 if (!upper_special) {
1674 SDOperand HI32C = DAG.getConstant(upper, MVT::i32);
1675 HI32 = DAG.getNode(ISD::BIT_CONVERT, VT,
1676 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1677 HI32C, HI32C, HI32C, HI32C));
1680 // If either upper or lower are special, then the two input operands are
1681 // the same (basically, one of them is a "don't care")
1686 if (lower_special && upper_special) {
1687 // Unhappy situation... both upper and lower are special, so punt with
1688 // a target constant:
1689 SDOperand Zero = DAG.getConstant(0, MVT::i32);
1690 HI32 = LO32 = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Zero, Zero,
1694 for (int i = 0; i < 4; ++i) {
1695 for (int j = 0; j < 4; ++j) {
1697 bool process_upper, process_lower;
1700 process_upper = (upper_special && (i & 1) == 0);
1701 process_lower = (lower_special && (i & 1) == 1);
1703 if (process_upper || process_lower) {
1704 if ((process_upper && upper == 0)
1705 || (process_lower && lower == 0))
1707 else if ((process_upper && upper == 0xffffffff)
1708 || (process_lower && lower == 0xffffffff))
1710 else if ((process_upper && upper == 0x80000000)
1711 || (process_lower && lower == 0x80000000))
1712 val = (j == 0 ? 0xe0 : 0x80);
1714 val = i * 4 + j + ((i & 1) * 16);
1716 ShufBytes.push_back(DAG.getConstant(val, MVT::i8));
1720 return DAG.getNode(SPUISD::SHUFB, VT, HI32, LO32,
1721 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
1722 &ShufBytes[0], ShufBytes.size()));
1724 // For zero, this can be lowered efficiently via v4i32 BUILD_VECTOR
1725 SDOperand Zero = DAG.getConstant(0, MVT::i32);
1726 return DAG.getNode(ISD::BIT_CONVERT, VT,
1727 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1728 Zero, Zero, Zero, Zero));
1736 /// LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3) to something on
1737 /// which the Cell can operate. The code inspects V3 to ascertain whether the
1738 /// permutation vector, V3, is monotonically increasing with one "exception"
1739 /// element, e.g., (0, 1, _, 3). If this is the case, then generate a
1740 /// INSERT_MASK synthetic instruction. Otherwise, spill V3 to the constant pool.
1741 /// In either case, the net result is going to eventually invoke SHUFB to
1742 /// permute/shuffle the bytes from V1 and V2.
1744 /// INSERT_MASK is eventually selected as one of the C*D instructions, generate
1745 /// control word for byte/halfword/word insertion. This takes care of a single
1746 /// element move from V2 into V1.
1748 /// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
1749 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
1750 SDOperand V1 = Op.getOperand(0);
1751 SDOperand V2 = Op.getOperand(1);
1752 SDOperand PermMask = Op.getOperand(2);
1754 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1756 // If we have a single element being moved from V1 to V2, this can be handled
1757 // using the C*[DX] compute mask instructions, but the vector elements have
1758 // to be monotonically increasing with one exception element.
1759 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
1760 unsigned EltsFromV2 = 0;
1762 unsigned V2EltIdx0 = 0;
1763 unsigned CurrElt = 0;
1764 bool monotonic = true;
1765 if (EltVT == MVT::i8)
1767 else if (EltVT == MVT::i16)
1769 else if (EltVT == MVT::i32)
1772 assert(0 && "Unhandled vector type in LowerVECTOR_SHUFFLE");
1774 for (unsigned i = 0, e = PermMask.getNumOperands();
1775 EltsFromV2 <= 1 && monotonic && i != e;
1778 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1781 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1783 if (SrcElt >= V2EltIdx0) {
1785 V2Elt = (V2EltIdx0 - SrcElt) << 2;
1786 } else if (CurrElt != SrcElt) {
1793 if (EltsFromV2 == 1 && monotonic) {
1794 // Compute mask and shuffle
1795 MachineFunction &MF = DAG.getMachineFunction();
1796 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1797 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
1798 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1799 // Initialize temporary register to 0
1800 SDOperand InitTempReg =
1801 DAG.getCopyToReg(DAG.getEntryNode(), VReg, DAG.getConstant(0, PtrVT));
1802 // Copy register's contents as index in INSERT_MASK:
1803 SDOperand ShufMaskOp =
1804 DAG.getNode(SPUISD::INSERT_MASK, V1.getValueType(),
1805 DAG.getTargetConstant(V2Elt, MVT::i32),
1806 DAG.getCopyFromReg(InitTempReg, VReg, PtrVT));
1807 // Use shuffle mask in SHUFB synthetic instruction:
1808 return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V2, V1, ShufMaskOp);
1810 // Convert the SHUFFLE_VECTOR mask's input element units to the actual bytes.
1811 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
1813 SmallVector<SDOperand, 16> ResultMask;
1814 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
1816 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1819 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1821 for (unsigned j = 0; j != BytesPerElement; ++j) {
1822 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
1827 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
1828 &ResultMask[0], ResultMask.size());
1829 return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V1, V2, VPermMask);
1833 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1834 SDOperand Op0 = Op.getOperand(0); // Op0 = the scalar
1836 if (Op0.Val->getOpcode() == ISD::Constant) {
1837 // For a constant, build the appropriate constant vector, which will
1838 // eventually simplify to a vector register load.
1840 ConstantSDNode *CN = cast<ConstantSDNode>(Op0.Val);
1841 SmallVector<SDOperand, 16> ConstVecValues;
1845 // Create a constant vector:
1846 switch (Op.getValueType()) {
1847 default: assert(0 && "Unexpected constant value type in "
1848 "LowerSCALAR_TO_VECTOR");
1849 case MVT::v16i8: n_copies = 16; VT = MVT::i8; break;
1850 case MVT::v8i16: n_copies = 8; VT = MVT::i16; break;
1851 case MVT::v4i32: n_copies = 4; VT = MVT::i32; break;
1852 case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
1853 case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
1854 case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
1857 SDOperand CValue = DAG.getConstant(CN->getValue(), VT);
1858 for (size_t j = 0; j < n_copies; ++j)
1859 ConstVecValues.push_back(CValue);
1861 return DAG.getNode(ISD::BUILD_VECTOR, Op.getValueType(),
1862 &ConstVecValues[0], ConstVecValues.size());
1864 // Otherwise, copy the value from one register to another:
1865 switch (Op0.getValueType()) {
1866 default: assert(0 && "Unexpected value type in LowerSCALAR_TO_VECTOR");
1873 return DAG.getNode(SPUISD::PROMOTE_SCALAR, Op.getValueType(), Op0, Op0);
1880 static SDOperand LowerVectorMUL(SDOperand Op, SelectionDAG &DAG) {
1881 switch (Op.getValueType()) {
1883 SDOperand rA = Op.getOperand(0);
1884 SDOperand rB = Op.getOperand(1);
1885 SDOperand HiProd1 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rA, rB);
1886 SDOperand HiProd2 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rB, rA);
1887 SDOperand LoProd = DAG.getNode(SPUISD::MPYU, MVT::v4i32, rA, rB);
1888 SDOperand Residual1 = DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd1);
1890 return DAG.getNode(ISD::ADD, MVT::v4i32, Residual1, HiProd2);
1894 // Multiply two v8i16 vectors (pipeline friendly version):
1895 // a) multiply lower halves, mask off upper 16-bit of 32-bit product
1896 // b) multiply upper halves, rotate left by 16 bits (inserts 16 lower zeroes)
1897 // c) Use SELB to select upper and lower halves from the intermediate results
1899 // NOTE: We really want to move the FSMBI to earlier to actually get the
1900 // dual-issue. This code does manage to do this, even if it's a little on
1903 MachineFunction &MF = DAG.getMachineFunction();
1904 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1905 SDOperand Chain = Op.getOperand(0);
1906 SDOperand rA = Op.getOperand(0);
1907 SDOperand rB = Op.getOperand(1);
1908 unsigned FSMBIreg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1909 unsigned HiProdReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1912 DAG.getCopyToReg(Chain, FSMBIreg,
1913 DAG.getNode(SPUISD::FSMBI, MVT::v8i16,
1914 DAG.getConstant(0xcccc, MVT::i32)));
1917 DAG.getCopyToReg(FSMBOp, HiProdReg,
1918 DAG.getNode(SPUISD::MPYHH, MVT::v8i16, rA, rB));
1920 SDOperand HHProd_v4i32 =
1921 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
1922 DAG.getCopyFromReg(HHProd, HiProdReg, MVT::v4i32));
1924 return DAG.getNode(SPUISD::SELB, MVT::v8i16,
1925 DAG.getNode(SPUISD::MPY, MVT::v8i16, rA, rB),
1926 DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
1927 DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32,
1929 DAG.getConstant(16, MVT::i16))),
1930 DAG.getCopyFromReg(FSMBOp, FSMBIreg, MVT::v4i32));
1933 // This M00sE is N@stI! (apologies to Monty Python)
1935 // SPU doesn't know how to do any 8-bit multiplication, so the solution
1936 // is to break it all apart, sign extend, and reassemble the various
1937 // intermediate products.
1939 MachineFunction &MF = DAG.getMachineFunction();
1940 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1941 SDOperand Chain = Op.getOperand(0);
1942 SDOperand rA = Op.getOperand(0);
1943 SDOperand rB = Op.getOperand(1);
1944 SDOperand c8 = DAG.getConstant(8, MVT::i8);
1945 SDOperand c16 = DAG.getConstant(16, MVT::i8);
1947 unsigned FSMBreg_2222 = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1948 unsigned LoProd_reg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1949 unsigned HiProd_reg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1952 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1953 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rA),
1954 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rB));
1956 SDOperand rALH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rA, c8);
1958 SDOperand rBLH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rB, c8);
1961 DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16,
1962 DAG.getNode(SPUISD::MPY, MVT::v8i16, rALH, rBLH), c8);
1964 SDOperand FSMBdef_2222 =
1965 DAG.getCopyToReg(Chain, FSMBreg_2222,
1966 DAG.getNode(SPUISD::FSMBI, MVT::v8i16,
1967 DAG.getConstant(0x2222, MVT::i32)));
1969 SDOperand FSMBuse_2222 =
1970 DAG.getCopyFromReg(FSMBdef_2222, FSMBreg_2222, MVT::v4i32);
1972 SDOperand LoProd_1 =
1973 DAG.getCopyToReg(Chain, LoProd_reg,
1974 DAG.getNode(SPUISD::SELB, MVT::v8i16, LLProd, LHProd,
1977 SDOperand LoProdMask = DAG.getConstant(0xffff, MVT::i32);
1980 DAG.getNode(ISD::AND, MVT::v4i32,
1981 DAG.getCopyFromReg(LoProd_1, LoProd_reg, MVT::v4i32),
1982 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1983 LoProdMask, LoProdMask,
1984 LoProdMask, LoProdMask));
1987 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
1988 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rA), c16);
1991 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
1992 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rB), c16);
1995 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1996 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rAH),
1997 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rBH));
1999 SDOperand HHProd_1 =
2000 DAG.getNode(SPUISD::MPY, MVT::v8i16,
2001 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
2002 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rAH, c8)),
2003 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
2004 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rBH, c8)));
2007 DAG.getCopyToReg(Chain, HiProd_reg,
2008 DAG.getNode(SPUISD::SELB, MVT::v8i16,
2010 DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16, HHProd_1, c8),
2014 DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32,
2015 DAG.getCopyFromReg(HHProd, HiProd_reg, MVT::v4i32), c16);
2017 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8,
2018 DAG.getNode(ISD::OR, MVT::v4i32,
2023 cerr << "CellSPU: Unknown vector multiplication, got "
2024 << MVT::getValueTypeString(Op.getValueType())
2033 static SDOperand LowerFDIVf32(SDOperand Op, SelectionDAG &DAG) {
2034 MachineFunction &MF = DAG.getMachineFunction();
2035 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2037 SDOperand A = Op.getOperand(0);
2038 SDOperand B = Op.getOperand(1);
2039 unsigned VT = Op.getValueType();
2041 unsigned VRegBR, VRegC;
2043 if (VT == MVT::f32) {
2044 VRegBR = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
2045 VRegC = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
2047 VRegBR = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
2048 VRegC = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
2050 // TODO: make sure we're feeding FPInterp the right arguments
2051 // Right now: fi B, frest(B)
2054 // (Floating Interpolate (FP Reciprocal Estimate B))
2056 DAG.getCopyToReg(DAG.getEntryNode(), VRegBR,
2057 DAG.getNode(SPUISD::FPInterp, VT, B,
2058 DAG.getNode(SPUISD::FPRecipEst, VT, B)));
2060 // Computes A * BRcpl and stores in a temporary register
2062 DAG.getCopyToReg(BRcpl, VRegC,
2063 DAG.getNode(ISD::FMUL, VT, A,
2064 DAG.getCopyFromReg(BRcpl, VRegBR, VT)));
2065 // What's the Chain variable do? It's magic!
2066 // TODO: set Chain = Op(0).getEntryNode()
2068 return DAG.getNode(ISD::FADD, VT,
2069 DAG.getCopyFromReg(AxBRcpl, VRegC, VT),
2070 DAG.getNode(ISD::FMUL, VT,
2071 DAG.getCopyFromReg(AxBRcpl, VRegBR, VT),
2072 DAG.getNode(ISD::FSUB, VT, A,
2073 DAG.getNode(ISD::FMUL, VT, B,
2074 DAG.getCopyFromReg(AxBRcpl, VRegC, VT)))));
2077 static SDOperand LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2078 unsigned VT = Op.getValueType();
2079 SDOperand N = Op.getOperand(0);
2080 SDOperand Elt = Op.getOperand(1);
2081 SDOperand ShufMask[16];
2082 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt);
2084 assert(C != 0 && "LowerEXTRACT_VECTOR_ELT expecting constant SDNode");
2086 int EltNo = (int) C->getValue();
2089 if (VT == MVT::i8 && EltNo >= 16)
2090 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i8 extraction slot > 15");
2091 else if (VT == MVT::i16 && EltNo >= 8)
2092 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i16 extraction slot > 7");
2093 else if (VT == MVT::i32 && EltNo >= 4)
2094 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i32 extraction slot > 4");
2095 else if (VT == MVT::i64 && EltNo >= 2)
2096 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i64 extraction slot > 2");
2098 if (EltNo == 0 && (VT == MVT::i32 || VT == MVT::i64)) {
2099 // i32 and i64: Element 0 is the preferred slot
2100 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT, N);
2103 // Need to generate shuffle mask and extract:
2104 int prefslot_begin = -1, prefslot_end = -1;
2105 int elt_byte = EltNo * MVT::getSizeInBits(VT) / 8;
2109 prefslot_begin = prefslot_end = 3;
2113 prefslot_begin = 2; prefslot_end = 3;
2117 prefslot_begin = 0; prefslot_end = 3;
2121 prefslot_begin = 0; prefslot_end = 7;
2126 assert(prefslot_begin != -1 && prefslot_end != -1 &&
2127 "LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
2129 for (int i = 0; i < 16; ++i) {
2130 // zero fill uppper part of preferred slot, don't care about the
2132 unsigned int mask_val;
2134 if (i <= prefslot_end) {
2136 ((i < prefslot_begin)
2138 : elt_byte + (i - prefslot_begin));
2140 ShufMask[i] = DAG.getConstant(mask_val, MVT::i8);
2142 ShufMask[i] = ShufMask[i % (prefslot_end + 1)];
2145 SDOperand ShufMaskVec =
2146 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2148 sizeof(ShufMask) / sizeof(ShufMask[0]));
2150 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
2151 DAG.getNode(SPUISD::SHUFB, N.getValueType(),
2152 N, N, ShufMaskVec));
2156 static SDOperand LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2157 SDOperand VecOp = Op.getOperand(0);
2158 SDOperand ValOp = Op.getOperand(1);
2159 SDOperand IdxOp = Op.getOperand(2);
2160 MVT::ValueType VT = Op.getValueType();
2162 ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
2163 assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
2165 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2166 // Use $2 because it's always 16-byte aligned and it's available:
2167 SDOperand PtrBase = DAG.getRegister(SPU::R2, PtrVT);
2170 DAG.getNode(SPUISD::SHUFB, VT,
2171 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, ValOp),
2173 DAG.getNode(SPUISD::INSERT_MASK, VT,
2174 DAG.getNode(ISD::ADD, PtrVT,
2176 DAG.getConstant(CN->getValue(),
2182 static SDOperand LowerI8Math(SDOperand Op, SelectionDAG &DAG, unsigned Opc) {
2183 SDOperand N0 = Op.getOperand(0); // Everything has at least one operand
2185 assert(Op.getValueType() == MVT::i8);
2188 assert(0 && "Unhandled i8 math operator");
2192 // 8-bit subtraction: Promote the arguments up to 16-bits and truncate
2194 SDOperand N1 = Op.getOperand(1);
2195 N0 = (N0.getOpcode() != ISD::Constant
2196 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2197 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2198 N1 = (N1.getOpcode() != ISD::Constant
2199 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N1)
2200 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2201 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2202 DAG.getNode(Opc, MVT::i16, N0, N1));
2206 SDOperand N1 = Op.getOperand(1);
2208 N0 = (N0.getOpcode() != ISD::Constant
2209 ? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
2210 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2211 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::ZERO_EXTEND : ISD::TRUNCATE);
2212 N1 = (N1.getOpcode() != ISD::Constant
2213 ? DAG.getNode(N1Opc, MVT::i16, N1)
2214 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2215 SDOperand ExpandArg =
2216 DAG.getNode(ISD::OR, MVT::i16, N0,
2217 DAG.getNode(ISD::SHL, MVT::i16,
2218 N0, DAG.getConstant(8, MVT::i16)));
2219 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2220 DAG.getNode(Opc, MVT::i16, ExpandArg, N1));
2224 SDOperand N1 = Op.getOperand(1);
2226 N0 = (N0.getOpcode() != ISD::Constant
2227 ? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
2228 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2229 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::ZERO_EXTEND : ISD::TRUNCATE);
2230 N1 = (N1.getOpcode() != ISD::Constant
2231 ? DAG.getNode(N1Opc, MVT::i16, N1)
2232 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2233 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2234 DAG.getNode(Opc, MVT::i16, N0, N1));
2237 SDOperand N1 = Op.getOperand(1);
2239 N0 = (N0.getOpcode() != ISD::Constant
2240 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2241 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2242 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::SIGN_EXTEND : ISD::TRUNCATE);
2243 N1 = (N1.getOpcode() != ISD::Constant
2244 ? DAG.getNode(N1Opc, MVT::i16, N1)
2245 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2246 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2247 DAG.getNode(Opc, MVT::i16, N0, N1));
2250 SDOperand N1 = Op.getOperand(1);
2252 N0 = (N0.getOpcode() != ISD::Constant
2253 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2254 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2255 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::SIGN_EXTEND : ISD::TRUNCATE);
2256 N1 = (N1.getOpcode() != ISD::Constant
2257 ? DAG.getNode(N1Opc, MVT::i16, N1)
2258 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2259 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2260 DAG.getNode(Opc, MVT::i16, N0, N1));
2268 //! Lower byte immediate operations for v16i8 vectors:
2270 LowerByteImmed(SDOperand Op, SelectionDAG &DAG) {
2273 MVT::ValueType VT = Op.getValueType();
2275 ConstVec = Op.getOperand(0);
2276 Arg = Op.getOperand(1);
2277 if (ConstVec.Val->getOpcode() != ISD::BUILD_VECTOR) {
2278 if (ConstVec.Val->getOpcode() == ISD::BIT_CONVERT) {
2279 ConstVec = ConstVec.getOperand(0);
2281 ConstVec = Op.getOperand(1);
2282 Arg = Op.getOperand(0);
2283 if (ConstVec.Val->getOpcode() == ISD::BIT_CONVERT) {
2284 ConstVec = ConstVec.getOperand(0);
2289 if (ConstVec.Val->getOpcode() == ISD::BUILD_VECTOR) {
2290 uint64_t VectorBits[2];
2291 uint64_t UndefBits[2];
2292 uint64_t SplatBits, SplatUndef;
2295 if (!GetConstantBuildVectorBits(ConstVec.Val, VectorBits, UndefBits)
2296 && isConstantSplat(VectorBits, UndefBits,
2297 MVT::getSizeInBits(MVT::getVectorElementType(VT)),
2298 SplatBits, SplatUndef, SplatSize)) {
2299 SDOperand tcVec[16];
2300 SDOperand tc = DAG.getTargetConstant(SplatBits & 0xff, MVT::i8);
2301 const size_t tcVecSize = sizeof(tcVec) / sizeof(tcVec[0]);
2303 // Turn the BUILD_VECTOR into a set of target constants:
2304 for (size_t i = 0; i < tcVecSize; ++i)
2307 return DAG.getNode(Op.Val->getOpcode(), VT, Arg,
2308 DAG.getNode(ISD::BUILD_VECTOR, VT, tcVec, tcVecSize));
2315 //! Lower i32 multiplication
2316 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG, unsigned VT,
2320 cerr << "CellSPU: Unknown LowerMUL value type, got "
2321 << MVT::getValueTypeString(Op.getValueType())
2327 SDOperand rA = Op.getOperand(0);
2328 SDOperand rB = Op.getOperand(1);
2330 return DAG.getNode(ISD::ADD, MVT::i32,
2331 DAG.getNode(ISD::ADD, MVT::i32,
2332 DAG.getNode(SPUISD::MPYH, MVT::i32, rA, rB),
2333 DAG.getNode(SPUISD::MPYH, MVT::i32, rB, rA)),
2334 DAG.getNode(SPUISD::MPYU, MVT::i32, rA, rB));
2341 //! Custom lowering for CTPOP (count population)
2343 Custom lowering code that counts the number ones in the input
2344 operand. SPU has such an instruction, but it counts the number of
2345 ones per byte, which then have to be accumulated.
2347 static SDOperand LowerCTPOP(SDOperand Op, SelectionDAG &DAG) {
2348 unsigned VT = Op.getValueType();
2349 unsigned vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
2353 SDOperand N = Op.getOperand(0);
2354 SDOperand Elt0 = DAG.getConstant(0, MVT::i32);
2356 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2357 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2359 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i8, CNTB, Elt0);
2363 MachineFunction &MF = DAG.getMachineFunction();
2364 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2366 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
2368 SDOperand N = Op.getOperand(0);
2369 SDOperand Elt0 = DAG.getConstant(0, MVT::i16);
2370 SDOperand Mask0 = DAG.getConstant(0x0f, MVT::i16);
2371 SDOperand Shift1 = DAG.getConstant(8, MVT::i16);
2373 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2374 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2376 // CNTB_result becomes the chain to which all of the virtual registers
2377 // CNTB_reg, SUM1_reg become associated:
2378 SDOperand CNTB_result =
2379 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, CNTB, Elt0);
2381 SDOperand CNTB_rescopy =
2382 DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
2384 SDOperand Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i16);
2386 return DAG.getNode(ISD::AND, MVT::i16,
2387 DAG.getNode(ISD::ADD, MVT::i16,
2388 DAG.getNode(ISD::SRL, MVT::i16,
2395 MachineFunction &MF = DAG.getMachineFunction();
2396 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2398 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2399 unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2401 SDOperand N = Op.getOperand(0);
2402 SDOperand Elt0 = DAG.getConstant(0, MVT::i32);
2403 SDOperand Mask0 = DAG.getConstant(0xff, MVT::i32);
2404 SDOperand Shift1 = DAG.getConstant(16, MVT::i32);
2405 SDOperand Shift2 = DAG.getConstant(8, MVT::i32);
2407 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2408 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2410 // CNTB_result becomes the chain to which all of the virtual registers
2411 // CNTB_reg, SUM1_reg become associated:
2412 SDOperand CNTB_result =
2413 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, CNTB, Elt0);
2415 SDOperand CNTB_rescopy =
2416 DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
2419 DAG.getNode(ISD::SRL, MVT::i32,
2420 DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32), Shift1);
2423 DAG.getNode(ISD::ADD, MVT::i32,
2424 Comp1, DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32));
2426 SDOperand Sum1_rescopy =
2427 DAG.getCopyToReg(CNTB_result, SUM1_reg, Sum1);
2430 DAG.getNode(ISD::SRL, MVT::i32,
2431 DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32),
2434 DAG.getNode(ISD::ADD, MVT::i32, Comp2,
2435 DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32));
2437 return DAG.getNode(ISD::AND, MVT::i32, Sum2, Mask0);
2447 /// LowerOperation - Provide custom lowering hooks for some operations.
2450 SPUTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG)
2452 switch (Op.getOpcode()) {
2454 cerr << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
2455 cerr << "Op.getOpcode() = " << Op.getOpcode() << "\n";
2456 cerr << "*Op.Val:\n";
2463 return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
2465 return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
2466 case ISD::ConstantPool:
2467 return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
2468 case ISD::GlobalAddress:
2469 return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
2470 case ISD::JumpTable:
2471 return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
2473 return LowerConstant(Op, DAG);
2474 case ISD::ConstantFP:
2475 return LowerConstantFP(Op, DAG);
2477 return LowerBRCOND(Op, DAG);
2478 case ISD::FORMAL_ARGUMENTS:
2479 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex);
2481 return LowerCALL(Op, DAG, SPUTM.getSubtargetImpl());
2483 return LowerRET(Op, DAG, getTargetMachine());
2492 return LowerI8Math(Op, DAG, Op.getOpcode());
2494 // Vector-related lowering.
2495 case ISD::BUILD_VECTOR:
2496 return LowerBUILD_VECTOR(Op, DAG);
2497 case ISD::SCALAR_TO_VECTOR:
2498 return LowerSCALAR_TO_VECTOR(Op, DAG);
2499 case ISD::VECTOR_SHUFFLE:
2500 return LowerVECTOR_SHUFFLE(Op, DAG);
2501 case ISD::EXTRACT_VECTOR_ELT:
2502 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2503 case ISD::INSERT_VECTOR_ELT:
2504 return LowerINSERT_VECTOR_ELT(Op, DAG);
2506 // Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
2510 return LowerByteImmed(Op, DAG);
2512 // Vector and i8 multiply:
2514 if (MVT::isVector(Op.getValueType()))
2515 return LowerVectorMUL(Op, DAG);
2516 else if (Op.getValueType() == MVT::i8)
2517 return LowerI8Math(Op, DAG, Op.getOpcode());
2519 return LowerMUL(Op, DAG, Op.getValueType(), Op.getOpcode());
2522 if (Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::v4f32)
2523 return LowerFDIVf32(Op, DAG);
2524 // else if (Op.getValueType() == MVT::f64)
2525 // return LowerFDIVf64(Op, DAG);
2527 assert(0 && "Calling FDIV on unsupported MVT");
2530 return LowerCTPOP(Op, DAG);
2536 //===----------------------------------------------------------------------===//
2537 // Other Lowering Code
2538 //===----------------------------------------------------------------------===//
2541 SPUTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
2542 MachineBasicBlock *BB)
2547 //===----------------------------------------------------------------------===//
2548 // Target Optimization Hooks
2549 //===----------------------------------------------------------------------===//
2552 SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
2555 TargetMachine &TM = getTargetMachine();
2556 SelectionDAG &DAG = DCI.DAG;
2557 SDOperand N0 = N->getOperand(0); // everything has at least one operand
2559 switch (N->getOpcode()) {
2561 // Do something creative here for ISD nodes that can be coalesced in unique
2566 // Otherwise, return unchanged.
2570 //===----------------------------------------------------------------------===//
2571 // Inline Assembly Support
2572 //===----------------------------------------------------------------------===//
2574 /// getConstraintType - Given a constraint letter, return the type of
2575 /// constraint it is for this target.
2576 SPUTargetLowering::ConstraintType
2577 SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
2578 if (ConstraintLetter.size() == 1) {
2579 switch (ConstraintLetter[0]) {
2586 return C_RegisterClass;
2589 return TargetLowering::getConstraintType(ConstraintLetter);
2592 std::pair<unsigned, const TargetRegisterClass*>
2593 SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
2594 MVT::ValueType VT) const
2596 if (Constraint.size() == 1) {
2597 // GCC RS6000 Constraint Letters
2598 switch (Constraint[0]) {
2602 return std::make_pair(0U, SPU::R64CRegisterClass);
2603 return std::make_pair(0U, SPU::R32CRegisterClass);
2606 return std::make_pair(0U, SPU::R32FPRegisterClass);
2607 else if (VT == MVT::f64)
2608 return std::make_pair(0U, SPU::R64FPRegisterClass);
2611 return std::make_pair(0U, SPU::GPRCRegisterClass);
2615 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2619 SPUTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
2621 uint64_t &KnownZero,
2623 const SelectionDAG &DAG,
2624 unsigned Depth ) const {
2629 // LowerAsmOperandForConstraint
2631 SPUTargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
2632 char ConstraintLetter,
2633 std::vector<SDOperand> &Ops,
2634 SelectionDAG &DAG) {
2635 // Default, for the time being, to the base class handler
2636 TargetLowering::LowerAsmOperandForConstraint(Op, ConstraintLetter, Ops, DAG);
2639 /// isLegalAddressImmediate - Return true if the integer value can be used
2640 /// as the offset of the target addressing mode.
2641 bool SPUTargetLowering::isLegalAddressImmediate(int64_t V, const Type *Ty) const {
2642 // SPU's addresses are 256K:
2643 return (V > -(1 << 18) && V < (1 << 18) - 1);
2646 bool SPUTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {