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 for (unsigned sctype = (unsigned) MVT::i1; sctype < (unsigned) MVT::f128;
294 setOperationAction(ISD::GlobalAddress, sctype, Custom);
295 setOperationAction(ISD::ConstantPool, sctype, Custom);
296 setOperationAction(ISD::JumpTable, sctype, Custom);
299 // RET must be custom lowered, to meet ABI requirements
300 setOperationAction(ISD::RET, MVT::Other, Custom);
302 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
303 setOperationAction(ISD::VASTART , MVT::Other, Custom);
305 // Use the default implementation.
306 setOperationAction(ISD::VAARG , MVT::Other, Expand);
307 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
308 setOperationAction(ISD::VAEND , MVT::Other, Expand);
309 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
310 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
311 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
312 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Expand);
314 // Cell SPU has instructions for converting between i64 and fp.
315 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
316 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
318 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
319 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
321 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
322 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
324 // First set operation action for all vector types to expand. Then we
325 // will selectively turn on ones that can be effectively codegen'd.
326 addRegisterClass(MVT::v16i8, SPU::VECREGRegisterClass);
327 addRegisterClass(MVT::v8i16, SPU::VECREGRegisterClass);
328 addRegisterClass(MVT::v4i32, SPU::VECREGRegisterClass);
329 addRegisterClass(MVT::v2i64, SPU::VECREGRegisterClass);
330 addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
331 addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
333 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
334 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
335 // add/sub are legal for all supported vector VT's.
336 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
337 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
338 // mul has to be custom lowered.
339 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Custom);
341 setOperationAction(ISD::AND , (MVT::ValueType)VT, Legal);
342 setOperationAction(ISD::OR , (MVT::ValueType)VT, Legal);
343 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Legal);
344 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Legal);
345 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Legal);
346 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Legal);
348 // These operations need to be expanded:
349 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
350 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
351 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
352 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
353 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Custom);
355 // Custom lower build_vector, constant pool spills, insert and
356 // extract vector elements:
357 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Custom);
358 setOperationAction(ISD::ConstantPool, (MVT::ValueType)VT, Custom);
359 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Custom);
360 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
361 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Custom);
362 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
365 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
366 setOperationAction(ISD::AND, MVT::v16i8, Custom);
367 setOperationAction(ISD::OR, MVT::v16i8, Custom);
368 setOperationAction(ISD::XOR, MVT::v16i8, Custom);
369 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
371 setSetCCResultType(MVT::i32);
372 setShiftAmountType(MVT::i32);
373 setSetCCResultContents(ZeroOrOneSetCCResult);
375 setStackPointerRegisterToSaveRestore(SPU::R1);
377 // We have target-specific dag combine patterns for the following nodes:
378 setTargetDAGCombine(ISD::ADD);
380 computeRegisterProperties();
384 SPUTargetLowering::getTargetNodeName(unsigned Opcode) const
386 if (node_names.empty()) {
387 node_names[(unsigned) SPUISD::RET_FLAG] = "SPUISD::RET_FLAG";
388 node_names[(unsigned) SPUISD::Hi] = "SPUISD::Hi";
389 node_names[(unsigned) SPUISD::Lo] = "SPUISD::Lo";
390 node_names[(unsigned) SPUISD::PCRelAddr] = "SPUISD::PCRelAddr";
391 node_names[(unsigned) SPUISD::AFormAddr] = "SPUISD::AFormAddr";
392 node_names[(unsigned) SPUISD::IndirectAddr] = "SPUISD::IndirectAddr";
393 node_names[(unsigned) SPUISD::LDRESULT] = "SPUISD::LDRESULT";
394 node_names[(unsigned) SPUISD::CALL] = "SPUISD::CALL";
395 node_names[(unsigned) SPUISD::SHUFB] = "SPUISD::SHUFB";
396 node_names[(unsigned) SPUISD::INSERT_MASK] = "SPUISD::INSERT_MASK";
397 node_names[(unsigned) SPUISD::CNTB] = "SPUISD::CNTB";
398 node_names[(unsigned) SPUISD::PROMOTE_SCALAR] = "SPUISD::PROMOTE_SCALAR";
399 node_names[(unsigned) SPUISD::EXTRACT_ELT0] = "SPUISD::EXTRACT_ELT0";
400 node_names[(unsigned) SPUISD::EXTRACT_ELT0_CHAINED] = "SPUISD::EXTRACT_ELT0_CHAINED";
401 node_names[(unsigned) SPUISD::EXTRACT_I1_ZEXT] = "SPUISD::EXTRACT_I1_ZEXT";
402 node_names[(unsigned) SPUISD::EXTRACT_I1_SEXT] = "SPUISD::EXTRACT_I1_SEXT";
403 node_names[(unsigned) SPUISD::EXTRACT_I8_ZEXT] = "SPUISD::EXTRACT_I8_ZEXT";
404 node_names[(unsigned) SPUISD::EXTRACT_I8_SEXT] = "SPUISD::EXTRACT_I8_SEXT";
405 node_names[(unsigned) SPUISD::MPY] = "SPUISD::MPY";
406 node_names[(unsigned) SPUISD::MPYU] = "SPUISD::MPYU";
407 node_names[(unsigned) SPUISD::MPYH] = "SPUISD::MPYH";
408 node_names[(unsigned) SPUISD::MPYHH] = "SPUISD::MPYHH";
409 node_names[(unsigned) SPUISD::VEC_SHL] = "SPUISD::VEC_SHL";
410 node_names[(unsigned) SPUISD::VEC_SRL] = "SPUISD::VEC_SRL";
411 node_names[(unsigned) SPUISD::VEC_SRA] = "SPUISD::VEC_SRA";
412 node_names[(unsigned) SPUISD::VEC_ROTL] = "SPUISD::VEC_ROTL";
413 node_names[(unsigned) SPUISD::VEC_ROTR] = "SPUISD::VEC_ROTR";
414 node_names[(unsigned) SPUISD::ROTBYTES_RIGHT_Z] =
415 "SPUISD::ROTBYTES_RIGHT_Z";
416 node_names[(unsigned) SPUISD::ROTBYTES_RIGHT_S] =
417 "SPUISD::ROTBYTES_RIGHT_S";
418 node_names[(unsigned) SPUISD::ROTBYTES_LEFT] = "SPUISD::ROTBYTES_LEFT";
419 node_names[(unsigned) SPUISD::ROTBYTES_LEFT_CHAINED] =
420 "SPUISD::ROTBYTES_LEFT_CHAINED";
421 node_names[(unsigned) SPUISD::FSMBI] = "SPUISD::FSMBI";
422 node_names[(unsigned) SPUISD::SELB] = "SPUISD::SELB";
423 node_names[(unsigned) SPUISD::SFPConstant] = "SPUISD::SFPConstant";
424 node_names[(unsigned) SPUISD::FPInterp] = "SPUISD::FPInterp";
425 node_names[(unsigned) SPUISD::FPRecipEst] = "SPUISD::FPRecipEst";
426 node_names[(unsigned) SPUISD::SEXT32TO64] = "SPUISD::SEXT32TO64";
429 std::map<unsigned, const char *>::iterator i = node_names.find(Opcode);
431 return ((i != node_names.end()) ? i->second : 0);
434 //===----------------------------------------------------------------------===//
435 // Calling convention code:
436 //===----------------------------------------------------------------------===//
438 #include "SPUGenCallingConv.inc"
440 //===----------------------------------------------------------------------===//
441 // LowerOperation implementation
442 //===----------------------------------------------------------------------===//
444 /// Aligned load common code for CellSPU
446 \param[in] Op The SelectionDAG load or store operand
447 \param[in] DAG The selection DAG
448 \param[in] ST CellSPU subtarget information structure
449 \param[in,out] alignment Caller initializes this to the load or store node's
450 value from getAlignment(), may be updated while generating the aligned load
451 \param[in,out] alignOffs Aligned offset; set by AlignedLoad to the aligned
452 offset (divisible by 16, modulo 16 == 0)
453 \param[in,out] prefSlotOffs Preferred slot offset; set by AlignedLoad to the
454 offset of the preferred slot (modulo 16 != 0)
455 \param[in,out] VT Caller initializes this value type to the the load or store
456 node's loaded or stored value type; may be updated if an i1-extended load or
458 \param[out] was16aligned true if the base pointer had 16-byte alignment,
459 otherwise false. Can help to determine if the chunk needs to be rotated.
461 Both load and store lowering load a block of data aligned on a 16-byte
462 boundary. This is the common aligned load code shared between both.
465 AlignedLoad(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST,
467 unsigned &alignment, int &alignOffs, int &prefSlotOffs,
468 MVT::ValueType &VT, bool &was16aligned)
470 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
471 const valtype_map_s *vtm = getValueTypeMapEntry(VT);
472 SDOperand basePtr = LSN->getBasePtr();
473 SDOperand chain = LSN->getChain();
475 if (basePtr.getOpcode() == ISD::ADD) {
476 SDOperand Op1 = basePtr.Val->getOperand(1);
478 if (Op1.getOpcode() == ISD::Constant || Op1.getOpcode() == ISD::TargetConstant) {
479 const ConstantSDNode *CN = cast<ConstantSDNode>(basePtr.getOperand(1));
481 alignOffs = (int) CN->getValue();
482 prefSlotOffs = (int) (alignOffs & 0xf);
484 // Adjust the rotation amount to ensure that the final result ends up in
485 // the preferred slot:
486 prefSlotOffs -= vtm->prefslot_byte;
487 basePtr = basePtr.getOperand(0);
489 // Loading from memory, can we adjust alignment?
490 if (basePtr.getOpcode() == SPUISD::AFormAddr) {
491 SDOperand APtr = basePtr.getOperand(0);
492 if (APtr.getOpcode() == ISD::TargetGlobalAddress) {
493 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(APtr);
494 alignment = GSDN->getGlobal()->getAlignment();
499 prefSlotOffs = -vtm->prefslot_byte;
503 prefSlotOffs = -vtm->prefslot_byte;
506 if (alignment == 16) {
507 // Realign the base pointer as a D-Form address:
508 if (!isMemoryOperand(basePtr) || (alignOffs & ~0xf) != 0) {
509 basePtr = DAG.getNode(ISD::ADD, PtrVT,
511 DAG.getConstant((alignOffs & ~0xf), PtrVT));
514 // Emit the vector load:
516 return DAG.getLoad(MVT::v16i8, chain, basePtr,
517 LSN->getSrcValue(), LSN->getSrcValueOffset(),
518 LSN->isVolatile(), 16);
521 // Unaligned load or we're using the "large memory" model, which means that
522 // we have to be very pessimistic:
523 if (isMemoryOperand(basePtr) || isIndirectOperand(basePtr)) {
524 basePtr = DAG.getNode(SPUISD::IndirectAddr, PtrVT, basePtr, DAG.getConstant(0, PtrVT));
528 basePtr = DAG.getNode(ISD::ADD, PtrVT, basePtr,
529 DAG.getConstant((alignOffs & ~0xf), PtrVT));
530 was16aligned = false;
531 return DAG.getLoad(MVT::v16i8, chain, basePtr,
532 LSN->getSrcValue(), LSN->getSrcValueOffset(),
533 LSN->isVolatile(), 16);
536 /// Custom lower loads for CellSPU
538 All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
539 within a 16-byte block, we have to rotate to extract the requested element.
542 LowerLOAD(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
543 LoadSDNode *LN = cast<LoadSDNode>(Op);
544 SDOperand the_chain = LN->getChain();
545 MVT::ValueType VT = LN->getMemoryVT();
546 MVT::ValueType OpVT = Op.Val->getValueType(0);
547 ISD::LoadExtType ExtType = LN->getExtensionType();
548 unsigned alignment = LN->getAlignment();
551 switch (LN->getAddressingMode()) {
552 case ISD::UNINDEXED: {
556 AlignedLoad(Op, DAG, ST, LN,alignment, offset, rotamt, VT, was16aligned);
561 the_chain = result.getValue(1);
562 // Rotate the chunk if necessary
565 if (rotamt != 0 || !was16aligned) {
566 SDVTList vecvts = DAG.getVTList(MVT::v16i8, MVT::Other);
571 Ops[2] = DAG.getConstant(rotamt, MVT::i16);
573 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
574 LoadSDNode *LN1 = cast<LoadSDNode>(result);
575 Ops[2] = DAG.getNode(ISD::ADD, PtrVT, LN1->getBasePtr(),
576 DAG.getConstant(rotamt, PtrVT));
579 result = DAG.getNode(SPUISD::ROTBYTES_LEFT_CHAINED, vecvts, Ops, 3);
580 the_chain = result.getValue(1);
583 if (VT == OpVT || ExtType == ISD::EXTLOAD) {
585 MVT::ValueType vecVT = MVT::v16i8;
587 // Convert the loaded v16i8 vector to the appropriate vector type
588 // specified by the operand:
591 vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
593 vecVT = MVT::getVectorType(OpVT, (128 / MVT::getSizeInBits(OpVT)));
596 Ops[1] = DAG.getNode(ISD::BIT_CONVERT, vecVT, result);
597 scalarvts = DAG.getVTList((OpVT == VT ? VT : OpVT), MVT::Other);
598 result = DAG.getNode(SPUISD::EXTRACT_ELT0_CHAINED, scalarvts, Ops, 2);
599 the_chain = result.getValue(1);
601 // Handle the sign and zero-extending loads for i1 and i8:
604 if (ExtType == ISD::SEXTLOAD) {
605 NewOpC = (OpVT == MVT::i1
606 ? SPUISD::EXTRACT_I1_SEXT
607 : SPUISD::EXTRACT_I8_SEXT);
609 assert(ExtType == ISD::ZEXTLOAD);
610 NewOpC = (OpVT == MVT::i1
611 ? SPUISD::EXTRACT_I1_ZEXT
612 : SPUISD::EXTRACT_I8_ZEXT);
615 result = DAG.getNode(NewOpC, OpVT, result);
618 SDVTList retvts = DAG.getVTList(OpVT, MVT::Other);
619 SDOperand retops[2] = {
624 result = DAG.getNode(SPUISD::LDRESULT, retvts,
625 retops, sizeof(retops) / sizeof(retops[0]));
632 case ISD::LAST_INDEXED_MODE:
633 cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
635 cerr << (unsigned) LN->getAddressingMode() << "\n";
643 /// Custom lower stores for CellSPU
645 All CellSPU stores are aligned to 16-byte boundaries, so for elements
646 within a 16-byte block, we have to generate a shuffle to insert the
647 requested element into its place, then store the resulting block.
650 LowerSTORE(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
651 StoreSDNode *SN = cast<StoreSDNode>(Op);
652 SDOperand Value = SN->getValue();
653 MVT::ValueType VT = Value.getValueType();
654 MVT::ValueType StVT = (!SN->isTruncatingStore() ? VT : SN->getMemoryVT());
655 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
656 unsigned alignment = SN->getAlignment();
658 switch (SN->getAddressingMode()) {
659 case ISD::UNINDEXED: {
660 int chunk_offset, slot_offset;
663 // The vector type we really want to load from the 16-byte chunk, except
664 // in the case of MVT::i1, which has to be v16i8.
665 unsigned vecVT, stVecVT = MVT::v16i8;
668 stVecVT = MVT::getVectorType(StVT, (128 / MVT::getSizeInBits(StVT)));
669 vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
671 SDOperand alignLoadVec =
672 AlignedLoad(Op, DAG, ST, SN, alignment,
673 chunk_offset, slot_offset, VT, was16aligned);
675 if (alignLoadVec.Val == 0)
678 LoadSDNode *LN = cast<LoadSDNode>(alignLoadVec);
679 SDOperand basePtr = LN->getBasePtr();
680 SDOperand the_chain = alignLoadVec.getValue(1);
681 SDOperand theValue = SN->getValue();
685 && (theValue.getOpcode() == ISD::AssertZext
686 || theValue.getOpcode() == ISD::AssertSext)) {
687 // Drill down and get the value for zero- and sign-extended
689 theValue = theValue.getOperand(0);
694 SDOperand insertEltOffs = DAG.getConstant(chunk_offset, PtrVT);
695 SDOperand insertEltPtr;
696 SDOperand insertEltOp;
698 // If the base pointer is already a D-form address, then just create
699 // a new D-form address with a slot offset and the orignal base pointer.
700 // Otherwise generate a D-form address with the slot offset relative
701 // to the stack pointer, which is always aligned.
702 DEBUG(cerr << "CellSPU LowerSTORE: basePtr = ");
703 DEBUG(basePtr.Val->dump(&DAG));
706 if (basePtr.getOpcode() == SPUISD::IndirectAddr ||
707 (basePtr.getOpcode() == ISD::ADD
708 && basePtr.getOperand(0).getOpcode() == SPUISD::IndirectAddr)) {
709 insertEltPtr = basePtr;
712 // $sp is always aligned, so use it when necessary to avoid loading
715 basePtr.Val->hasOneUse() ? DAG.getRegister(SPU::R1, PtrVT) : basePtr;
716 insertEltPtr = DAG.getNode(ISD::ADD, PtrVT, ptrP, insertEltOffs);
718 insertEltPtr = DAG.getNode(ISD::ADD, PtrVT, basePtr, insertEltOffs);
722 insertEltOp = DAG.getNode(SPUISD::INSERT_MASK, stVecVT, insertEltPtr);
723 result = DAG.getNode(SPUISD::SHUFB, vecVT,
724 DAG.getNode(ISD::SCALAR_TO_VECTOR, vecVT, theValue),
726 DAG.getNode(ISD::BIT_CONVERT, vecVT, insertEltOp));
728 result = DAG.getStore(the_chain, result, basePtr,
729 LN->getSrcValue(), LN->getSrcValueOffset(),
730 LN->isVolatile(), LN->getAlignment());
739 case ISD::LAST_INDEXED_MODE:
740 cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
742 cerr << (unsigned) SN->getAddressingMode() << "\n";
750 /// Generate the address of a constant pool entry.
752 LowerConstantPool(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
753 MVT::ValueType PtrVT = Op.getValueType();
754 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
755 Constant *C = CP->getConstVal();
756 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
757 SDOperand Zero = DAG.getConstant(0, PtrVT);
758 const TargetMachine &TM = DAG.getTarget();
760 if (TM.getRelocationModel() == Reloc::Static) {
761 if (!ST->usingLargeMem()) {
762 // Just return the SDOperand with the constant pool address in it.
763 return DAG.getNode(SPUISD::AFormAddr, PtrVT, CPI, Zero);
766 SDOperand Hi = DAG.getNode(SPUISD::Hi, PtrVT, CPI, Zero);
767 SDOperand Lo = DAG.getNode(SPUISD::Lo, PtrVT, CPI, Zero);
769 return DAG.getNode(ISD::ADD, PtrVT, Lo, Hi);
771 return DAG.getNode(SPUISD::IndirectAddr, PtrVT, CPI, Zero);
777 "LowerConstantPool: Relocation model other than static not supported.");
782 LowerJumpTable(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
783 MVT::ValueType PtrVT = Op.getValueType();
784 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
785 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
786 SDOperand Zero = DAG.getConstant(0, PtrVT);
787 const TargetMachine &TM = DAG.getTarget();
789 if (TM.getRelocationModel() == Reloc::Static) {
790 SDOperand JmpAForm = DAG.getNode(SPUISD::AFormAddr, PtrVT, JTI, Zero);
791 return (!ST->usingLargeMem()
793 : DAG.getNode(SPUISD::IndirectAddr, PtrVT, JmpAForm, Zero));
797 "LowerJumpTable: Relocation model other than static not supported.");
802 LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
803 MVT::ValueType PtrVT = Op.getValueType();
804 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
805 GlobalValue *GV = GSDN->getGlobal();
806 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
807 const TargetMachine &TM = DAG.getTarget();
808 SDOperand Zero = DAG.getConstant(0, PtrVT);
810 if (TM.getRelocationModel() == Reloc::Static) {
811 if (!ST->usingLargeMem()) {
812 return DAG.getNode(SPUISD::AFormAddr, PtrVT, GA, Zero);
814 SDOperand Hi = DAG.getNode(SPUISD::Hi, PtrVT, GA, Zero);
815 SDOperand Lo = DAG.getNode(SPUISD::Lo, PtrVT, GA, Zero);
816 return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
819 cerr << "LowerGlobalAddress: Relocation model other than static not "
828 //! Custom lower i64 integer constants
830 This code inserts all of the necessary juggling that needs to occur to load
831 a 64-bit constant into a register.
834 LowerConstant(SDOperand Op, SelectionDAG &DAG) {
835 unsigned VT = Op.getValueType();
836 ConstantSDNode *CN = cast<ConstantSDNode>(Op.Val);
838 if (VT == MVT::i64) {
839 SDOperand T = DAG.getConstant(CN->getValue(), MVT::i64);
840 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
841 DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
844 cerr << "LowerConstant: unhandled constant type "
845 << MVT::getValueTypeString(VT)
854 //! Custom lower single precision floating point constants
856 "float" immediates can be lowered as if they were unsigned 32-bit integers.
857 The SPUISD::SFPConstant pseudo-instruction handles this in the instruction
861 LowerConstantFP(SDOperand Op, SelectionDAG &DAG) {
862 unsigned VT = Op.getValueType();
863 ConstantFPSDNode *FP = cast<ConstantFPSDNode>(Op.Val);
866 "LowerConstantFP: Node is not ConstantFPSDNode");
868 if (VT == MVT::f32) {
869 float targetConst = FP->getValueAPF().convertToFloat();
870 return DAG.getNode(SPUISD::SFPConstant, VT,
871 DAG.getTargetConstantFP(targetConst, VT));
872 } else if (VT == MVT::f64) {
873 uint64_t dbits = DoubleToBits(FP->getValueAPF().convertToDouble());
874 return DAG.getNode(ISD::BIT_CONVERT, VT,
875 LowerConstant(DAG.getConstant(dbits, MVT::i64), DAG));
881 //! Lower MVT::i1, MVT::i8 brcond to a promoted type (MVT::i32, MVT::i16)
883 LowerBRCOND(SDOperand Op, SelectionDAG &DAG)
885 SDOperand Cond = Op.getOperand(1);
886 MVT::ValueType CondVT = Cond.getValueType();
887 MVT::ValueType CondNVT;
889 if (CondVT == MVT::i1 || CondVT == MVT::i8) {
890 CondNVT = (CondVT == MVT::i1 ? MVT::i32 : MVT::i16);
891 return DAG.getNode(ISD::BRCOND, Op.getValueType(),
893 DAG.getNode(ISD::ZERO_EXTEND, CondNVT, Op.getOperand(1)),
896 return SDOperand(); // Unchanged
900 LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG, int &VarArgsFrameIndex)
902 MachineFunction &MF = DAG.getMachineFunction();
903 MachineFrameInfo *MFI = MF.getFrameInfo();
904 MachineRegisterInfo &RegInfo = MF.getRegInfo();
905 SmallVector<SDOperand, 8> ArgValues;
906 SDOperand Root = Op.getOperand(0);
907 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
909 const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
910 const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
912 unsigned ArgOffset = SPUFrameInfo::minStackSize();
913 unsigned ArgRegIdx = 0;
914 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
916 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
918 // Add DAG nodes to load the arguments or copy them out of registers.
919 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
921 bool needsLoad = false;
922 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
923 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
927 cerr << "LowerFORMAL_ARGUMENTS Unhandled argument type: "
928 << MVT::getValueTypeString(ObjectVT)
933 if (!isVarArg && ArgRegIdx < NumArgRegs) {
934 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R8CRegClass);
935 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
936 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i8);
943 if (!isVarArg && ArgRegIdx < NumArgRegs) {
944 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
945 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
946 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i16);
953 if (!isVarArg && ArgRegIdx < NumArgRegs) {
954 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
955 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
956 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
963 if (!isVarArg && ArgRegIdx < NumArgRegs) {
964 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64CRegClass);
965 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
966 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
973 if (!isVarArg && ArgRegIdx < NumArgRegs) {
974 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
975 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
976 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f32);
983 if (!isVarArg && ArgRegIdx < NumArgRegs) {
984 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64FPRegClass);
985 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
986 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f64);
997 if (!isVarArg && ArgRegIdx < NumArgRegs) {
998 unsigned VReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
999 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
1000 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1008 // We need to load the argument to a virtual register if we determined above
1009 // that we ran out of physical registers of the appropriate type
1011 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1012 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1013 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1014 ArgOffset += StackSlotSize;
1017 ArgValues.push_back(ArgVal);
1020 // If the function takes variable number of arguments, make a frame index for
1021 // the start of the first vararg value... for expansion of llvm.va_start.
1023 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1025 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1026 // If this function is vararg, store any remaining integer argument regs to
1027 // their spots on the stack so that they may be loaded by deferencing the
1028 // result of va_next.
1029 SmallVector<SDOperand, 8> MemOps;
1030 for (; ArgRegIdx != NumArgRegs; ++ArgRegIdx) {
1031 unsigned VReg = RegInfo.createVirtualRegister(&SPU::GPRCRegClass);
1032 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
1033 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1034 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1035 MemOps.push_back(Store);
1036 // Increment the address by four for the next argument to store
1037 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1038 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1040 if (!MemOps.empty())
1041 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1044 ArgValues.push_back(Root);
1046 // Return the new list of results.
1047 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1048 Op.Val->value_end());
1049 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1052 /// isLSAAddress - Return the immediate to use if the specified
1053 /// value is representable as a LSA address.
1054 static SDNode *isLSAAddress(SDOperand Op, SelectionDAG &DAG) {
1055 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1058 int Addr = C->getValue();
1059 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1060 (Addr << 14 >> 14) != Addr)
1061 return 0; // Top 14 bits have to be sext of immediate.
1063 return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).Val;
1068 LowerCALL(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1069 SDOperand Chain = Op.getOperand(0);
1071 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1072 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1074 SDOperand Callee = Op.getOperand(4);
1075 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1076 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
1077 const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
1078 const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
1080 // Handy pointer type
1081 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1083 // Accumulate how many bytes are to be pushed on the stack, including the
1084 // linkage area, and parameter passing area. According to the SPU ABI,
1085 // we minimally need space for [LR] and [SP]
1086 unsigned NumStackBytes = SPUFrameInfo::minStackSize();
1088 // Set up a copy of the stack pointer for use loading and storing any
1089 // arguments that may not fit in the registers available for argument
1091 SDOperand StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
1093 // Figure out which arguments are going to go in registers, and which in
1095 unsigned ArgOffset = SPUFrameInfo::minStackSize(); // Just below [LR]
1096 unsigned ArgRegIdx = 0;
1098 // Keep track of registers passing arguments
1099 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1100 // And the arguments passed on the stack
1101 SmallVector<SDOperand, 8> MemOpChains;
1103 for (unsigned i = 0; i != NumOps; ++i) {
1104 SDOperand Arg = Op.getOperand(5+2*i);
1106 // PtrOff will be used to store the current argument to the stack if a
1107 // register cannot be found for it.
1108 SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1109 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1111 switch (Arg.getValueType()) {
1112 default: assert(0 && "Unexpected ValueType for argument!");
1116 if (ArgRegIdx != NumArgRegs) {
1117 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1119 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1120 ArgOffset += StackSlotSize;
1125 if (ArgRegIdx != NumArgRegs) {
1126 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1128 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1129 ArgOffset += StackSlotSize;
1136 if (ArgRegIdx != NumArgRegs) {
1137 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1139 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1140 ArgOffset += StackSlotSize;
1146 // Update number of stack bytes actually used, insert a call sequence start
1147 NumStackBytes = (ArgOffset - SPUFrameInfo::minStackSize());
1148 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumStackBytes, PtrVT));
1150 if (!MemOpChains.empty()) {
1151 // Adjust the stack pointer for the stack arguments.
1152 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1153 &MemOpChains[0], MemOpChains.size());
1156 // Build a sequence of copy-to-reg nodes chained together with token chain
1157 // and flag operands which copy the outgoing args into the appropriate regs.
1159 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1160 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1162 InFlag = Chain.getValue(1);
1165 std::vector<MVT::ValueType> NodeTys;
1166 NodeTys.push_back(MVT::Other); // Returns a chain
1167 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1169 SmallVector<SDOperand, 8> Ops;
1170 unsigned CallOpc = SPUISD::CALL;
1172 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1173 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1174 // node so that legalize doesn't hack it.
1175 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1176 GlobalValue *GV = G->getGlobal();
1177 unsigned CalleeVT = Callee.getValueType();
1178 SDOperand Zero = DAG.getConstant(0, PtrVT);
1179 SDOperand GA = DAG.getTargetGlobalAddress(GV, CalleeVT);
1181 if (!ST->usingLargeMem()) {
1182 // Turn calls to targets that are defined (i.e., have bodies) into BRSL
1183 // style calls, otherwise, external symbols are BRASL calls. This assumes
1184 // that declared/defined symbols are in the same compilation unit and can
1185 // be reached through PC-relative jumps.
1188 // This may be an unsafe assumption for JIT and really large compilation
1190 if (GV->isDeclaration()) {
1191 Callee = DAG.getNode(SPUISD::AFormAddr, CalleeVT, GA, Zero);
1193 Callee = DAG.getNode(SPUISD::PCRelAddr, CalleeVT, GA, Zero);
1196 // "Large memory" mode: Turn all calls into indirect calls with a X-form
1198 Callee = DAG.getNode(SPUISD::IndirectAddr, PtrVT, GA, Zero);
1200 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1201 Callee = DAG.getExternalSymbol(S->getSymbol(), Callee.getValueType());
1202 else if (SDNode *Dest = isLSAAddress(Callee, DAG)) {
1203 // If this is an absolute destination address that appears to be a legal
1204 // local store address, use the munged value.
1205 Callee = SDOperand(Dest, 0);
1208 Ops.push_back(Chain);
1209 Ops.push_back(Callee);
1211 // Add argument registers to the end of the list so that they are known live
1213 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1214 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1215 RegsToPass[i].second.getValueType()));
1218 Ops.push_back(InFlag);
1219 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1220 InFlag = Chain.getValue(1);
1222 Chain = DAG.getCALLSEQ_END(Chain,
1223 DAG.getConstant(NumStackBytes, PtrVT),
1224 DAG.getConstant(0, PtrVT),
1226 if (Op.Val->getValueType(0) != MVT::Other)
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 NodeTys.push_back(MVT::Other);
1282 // If the function returns void, just return the chain.
1283 if (NumResults == 0)
1286 // Otherwise, merge everything together with a MERGE_VALUES node.
1287 ResultVals[NumResults++] = Chain;
1288 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1289 ResultVals, NumResults);
1290 return Res.getValue(Op.ResNo);
1294 LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1295 SmallVector<CCValAssign, 16> RVLocs;
1296 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1297 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1298 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1299 CCInfo.AnalyzeReturn(Op.Val, RetCC_SPU);
1301 // If this is the first return lowered for this function, add the regs to the
1302 // liveout set for the function.
1303 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1304 for (unsigned i = 0; i != RVLocs.size(); ++i)
1305 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1308 SDOperand Chain = Op.getOperand(0);
1311 // Copy the result values into the output registers.
1312 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1313 CCValAssign &VA = RVLocs[i];
1314 assert(VA.isRegLoc() && "Can only return in registers!");
1315 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1316 Flag = Chain.getValue(1);
1320 return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain, Flag);
1322 return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain);
1326 //===----------------------------------------------------------------------===//
1327 // Vector related lowering:
1328 //===----------------------------------------------------------------------===//
1330 static ConstantSDNode *
1331 getVecImm(SDNode *N) {
1332 SDOperand OpVal(0, 0);
1334 // Check to see if this buildvec has a single non-undef value in its elements.
1335 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1336 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1338 OpVal = N->getOperand(i);
1339 else if (OpVal != N->getOperand(i))
1343 if (OpVal.Val != 0) {
1344 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1349 return 0; // All UNDEF: use implicit def.; not Constant node
1352 /// get_vec_i18imm - Test if this vector is a vector filled with the same value
1353 /// and the value fits into an unsigned 18-bit constant, and if so, return the
1355 SDOperand SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
1356 MVT::ValueType ValueType) {
1357 if (ConstantSDNode *CN = getVecImm(N)) {
1358 uint64_t Value = CN->getValue();
1359 if (Value <= 0x3ffff)
1360 return DAG.getConstant(Value, ValueType);
1366 /// get_vec_i16imm - Test if this vector is a vector filled with the same value
1367 /// and the value fits into a signed 16-bit constant, and if so, return the
1369 SDOperand SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
1370 MVT::ValueType ValueType) {
1371 if (ConstantSDNode *CN = getVecImm(N)) {
1372 if (ValueType == MVT::i32) {
1373 int Value = (int) CN->getValue();
1374 int SExtValue = ((Value & 0xffff) << 16) >> 16;
1376 if (Value == SExtValue)
1377 return DAG.getConstant(Value, ValueType);
1378 } else if (ValueType == MVT::i16) {
1379 short Value = (short) CN->getValue();
1380 int SExtValue = ((int) Value << 16) >> 16;
1382 if (Value == (short) SExtValue)
1383 return DAG.getConstant(Value, ValueType);
1384 } else if (ValueType == MVT::i64) {
1385 int64_t Value = CN->getValue();
1386 int64_t SExtValue = ((Value & 0xffff) << (64 - 16)) >> (64 - 16);
1388 if (Value == SExtValue)
1389 return DAG.getConstant(Value, ValueType);
1396 /// get_vec_i10imm - Test if this vector is a vector filled with the same value
1397 /// and the value fits into a signed 10-bit constant, and if so, return the
1399 SDOperand SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
1400 MVT::ValueType ValueType) {
1401 if (ConstantSDNode *CN = getVecImm(N)) {
1402 int Value = (int) CN->getValue();
1403 if ((ValueType == MVT::i32 && isS10Constant(Value))
1404 || (ValueType == MVT::i16 && isS10Constant((short) Value)))
1405 return DAG.getConstant(Value, ValueType);
1411 /// get_vec_i8imm - Test if this vector is a vector filled with the same value
1412 /// and the value fits into a signed 8-bit constant, and if so, return the
1415 /// @note: The incoming vector is v16i8 because that's the only way we can load
1416 /// constant vectors. Thus, we test to see if the upper and lower bytes are the
1418 SDOperand SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
1419 MVT::ValueType ValueType) {
1420 if (ConstantSDNode *CN = getVecImm(N)) {
1421 int Value = (int) CN->getValue();
1422 if (ValueType == MVT::i16
1423 && Value <= 0xffff /* truncated from uint64_t */
1424 && ((short) Value >> 8) == ((short) Value & 0xff))
1425 return DAG.getConstant(Value & 0xff, ValueType);
1426 else if (ValueType == MVT::i8
1427 && (Value & 0xff) == Value)
1428 return DAG.getConstant(Value, ValueType);
1434 /// get_ILHUvec_imm - Test if this vector is a vector filled with the same value
1435 /// and the value fits into a signed 16-bit constant, and if so, return the
1437 SDOperand SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
1438 MVT::ValueType ValueType) {
1439 if (ConstantSDNode *CN = getVecImm(N)) {
1440 uint64_t Value = CN->getValue();
1441 if ((ValueType == MVT::i32
1442 && ((unsigned) Value & 0xffff0000) == (unsigned) Value)
1443 || (ValueType == MVT::i64 && (Value & 0xffff0000) == Value))
1444 return DAG.getConstant(Value >> 16, ValueType);
1450 /// get_v4i32_imm - Catch-all for general 32-bit constant vectors
1451 SDOperand SPU::get_v4i32_imm(SDNode *N, SelectionDAG &DAG) {
1452 if (ConstantSDNode *CN = getVecImm(N)) {
1453 return DAG.getConstant((unsigned) CN->getValue(), MVT::i32);
1459 /// get_v4i32_imm - Catch-all for general 64-bit constant vectors
1460 SDOperand SPU::get_v2i64_imm(SDNode *N, SelectionDAG &DAG) {
1461 if (ConstantSDNode *CN = getVecImm(N)) {
1462 return DAG.getConstant((unsigned) CN->getValue(), MVT::i64);
1468 // If this is a vector of constants or undefs, get the bits. A bit in
1469 // UndefBits is set if the corresponding element of the vector is an
1470 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1471 // zero. Return true if this is not an array of constants, false if it is.
1473 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
1474 uint64_t UndefBits[2]) {
1475 // Start with zero'd results.
1476 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
1478 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
1479 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
1480 SDOperand OpVal = BV->getOperand(i);
1482 unsigned PartNo = i >= e/2; // In the upper 128 bits?
1483 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
1485 uint64_t EltBits = 0;
1486 if (OpVal.getOpcode() == ISD::UNDEF) {
1487 uint64_t EltUndefBits = ~0ULL >> (64-EltBitSize);
1488 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
1490 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1491 EltBits = CN->getValue() & (~0ULL >> (64-EltBitSize));
1492 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
1493 const APFloat &apf = CN->getValueAPF();
1494 EltBits = (CN->getValueType(0) == MVT::f32
1495 ? FloatToBits(apf.convertToFloat())
1496 : DoubleToBits(apf.convertToDouble()));
1498 // Nonconstant element.
1502 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
1505 //printf("%llx %llx %llx %llx\n",
1506 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
1510 /// If this is a splat (repetition) of a value across the whole vector, return
1511 /// the smallest size that splats it. For example, "0x01010101010101..." is a
1512 /// splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
1513 /// SplatSize = 1 byte.
1514 static bool isConstantSplat(const uint64_t Bits128[2],
1515 const uint64_t Undef128[2],
1517 uint64_t &SplatBits, uint64_t &SplatUndef,
1519 // Don't let undefs prevent splats from matching. See if the top 64-bits are
1520 // the same as the lower 64-bits, ignoring undefs.
1521 uint64_t Bits64 = Bits128[0] | Bits128[1];
1522 uint64_t Undef64 = Undef128[0] & Undef128[1];
1523 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
1524 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
1525 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
1526 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
1528 if ((Bits128[0] & ~Undef128[1]) == (Bits128[1] & ~Undef128[0])) {
1529 if (MinSplatBits < 64) {
1531 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
1533 if ((Bits64 & (~Undef64 >> 32)) == ((Bits64 >> 32) & ~Undef64)) {
1534 if (MinSplatBits < 32) {
1536 // If the top 16-bits are different than the lower 16-bits, ignoring
1537 // undefs, we have an i32 splat.
1538 if ((Bits32 & (~Undef32 >> 16)) == ((Bits32 >> 16) & ~Undef32)) {
1539 if (MinSplatBits < 16) {
1540 // If the top 8-bits are different than the lower 8-bits, ignoring
1541 // undefs, we have an i16 splat.
1542 if ((Bits16 & (uint16_t(~Undef16) >> 8)) == ((Bits16 >> 8) & ~Undef16)) {
1543 // Otherwise, we have an 8-bit splat.
1544 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
1545 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
1551 SplatUndef = Undef16;
1558 SplatUndef = Undef32;
1564 SplatBits = Bits128[0];
1565 SplatUndef = Undef128[0];
1571 return false; // Can't be a splat if two pieces don't match.
1574 // If this is a case we can't handle, return null and let the default
1575 // expansion code take care of it. If we CAN select this case, and if it
1576 // selects to a single instruction, return Op. Otherwise, if we can codegen
1577 // this case more efficiently than a constant pool load, lower it to the
1578 // sequence of ops that should be used.
1579 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1580 MVT::ValueType VT = Op.getValueType();
1581 // If this is a vector of constants or undefs, get the bits. A bit in
1582 // UndefBits is set if the corresponding element of the vector is an
1583 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1585 uint64_t VectorBits[2];
1586 uint64_t UndefBits[2];
1587 uint64_t SplatBits, SplatUndef;
1589 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits)
1590 || !isConstantSplat(VectorBits, UndefBits,
1591 MVT::getSizeInBits(MVT::getVectorElementType(VT)),
1592 SplatBits, SplatUndef, SplatSize))
1593 return SDOperand(); // Not a constant vector, not a splat.
1598 uint32_t Value32 = SplatBits;
1599 assert(SplatSize == 4
1600 && "LowerBUILD_VECTOR: Unexpected floating point vector element.");
1601 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1602 SDOperand T = DAG.getConstant(Value32, MVT::i32);
1603 return DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32,
1604 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, T, T, T, T));
1608 uint64_t f64val = SplatBits;
1609 assert(SplatSize == 8
1610 && "LowerBUILD_VECTOR: 64-bit float vector element: unexpected size.");
1611 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1612 SDOperand T = DAG.getConstant(f64val, MVT::i64);
1613 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64,
1614 DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
1618 // 8-bit constants have to be expanded to 16-bits
1619 unsigned short Value16 = SplatBits | (SplatBits << 8);
1621 for (int i = 0; i < 8; ++i)
1622 Ops[i] = DAG.getConstant(Value16, MVT::i16);
1623 return DAG.getNode(ISD::BIT_CONVERT, VT,
1624 DAG.getNode(ISD::BUILD_VECTOR, MVT::v8i16, Ops, 8));
1627 unsigned short Value16;
1629 Value16 = (unsigned short) (SplatBits & 0xffff);
1631 Value16 = (unsigned short) (SplatBits | (SplatBits << 8));
1632 SDOperand T = DAG.getConstant(Value16, MVT::getVectorElementType(VT));
1634 for (int i = 0; i < 8; ++i) Ops[i] = T;
1635 return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops, 8);
1638 unsigned int Value = SplatBits;
1639 SDOperand T = DAG.getConstant(Value, MVT::getVectorElementType(VT));
1640 return DAG.getNode(ISD::BUILD_VECTOR, VT, T, T, T, T);
1643 uint64_t val = SplatBits;
1644 uint32_t upper = uint32_t(val >> 32);
1645 uint32_t lower = uint32_t(val);
1650 SmallVector<SDOperand, 16> ShufBytes;
1652 bool upper_special, lower_special;
1654 // NOTE: This code creates common-case shuffle masks that can be easily
1655 // detected as common expressions. It is not attempting to create highly
1656 // specialized masks to replace any and all 0's, 0xff's and 0x80's.
1658 // Detect if the upper or lower half is a special shuffle mask pattern:
1659 upper_special = (upper == 0 || upper == 0xffffffff || upper == 0x80000000);
1660 lower_special = (lower == 0 || lower == 0xffffffff || lower == 0x80000000);
1662 // Create lower vector if not a special pattern
1663 if (!lower_special) {
1664 SDOperand LO32C = DAG.getConstant(lower, MVT::i32);
1665 LO32 = DAG.getNode(ISD::BIT_CONVERT, VT,
1666 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1667 LO32C, LO32C, LO32C, LO32C));
1670 // Create upper vector if not a special pattern
1671 if (!upper_special) {
1672 SDOperand HI32C = DAG.getConstant(upper, MVT::i32);
1673 HI32 = DAG.getNode(ISD::BIT_CONVERT, VT,
1674 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1675 HI32C, HI32C, HI32C, HI32C));
1678 // If either upper or lower are special, then the two input operands are
1679 // the same (basically, one of them is a "don't care")
1684 if (lower_special && upper_special) {
1685 // Unhappy situation... both upper and lower are special, so punt with
1686 // a target constant:
1687 SDOperand Zero = DAG.getConstant(0, MVT::i32);
1688 HI32 = LO32 = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Zero, Zero,
1692 for (int i = 0; i < 4; ++i) {
1693 for (int j = 0; j < 4; ++j) {
1695 bool process_upper, process_lower;
1698 process_upper = (upper_special && (i & 1) == 0);
1699 process_lower = (lower_special && (i & 1) == 1);
1701 if (process_upper || process_lower) {
1702 if ((process_upper && upper == 0)
1703 || (process_lower && lower == 0))
1705 else if ((process_upper && upper == 0xffffffff)
1706 || (process_lower && lower == 0xffffffff))
1708 else if ((process_upper && upper == 0x80000000)
1709 || (process_lower && lower == 0x80000000))
1710 val = (j == 0 ? 0xe0 : 0x80);
1712 val = i * 4 + j + ((i & 1) * 16);
1714 ShufBytes.push_back(DAG.getConstant(val, MVT::i8));
1718 return DAG.getNode(SPUISD::SHUFB, VT, HI32, LO32,
1719 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
1720 &ShufBytes[0], ShufBytes.size()));
1722 // For zero, this can be lowered efficiently via v4i32 BUILD_VECTOR
1723 SDOperand Zero = DAG.getConstant(0, MVT::i32);
1724 return DAG.getNode(ISD::BIT_CONVERT, VT,
1725 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1726 Zero, Zero, Zero, Zero));
1734 /// LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3) to something on
1735 /// which the Cell can operate. The code inspects V3 to ascertain whether the
1736 /// permutation vector, V3, is monotonically increasing with one "exception"
1737 /// element, e.g., (0, 1, _, 3). If this is the case, then generate a
1738 /// INSERT_MASK synthetic instruction. Otherwise, spill V3 to the constant pool.
1739 /// In either case, the net result is going to eventually invoke SHUFB to
1740 /// permute/shuffle the bytes from V1 and V2.
1742 /// INSERT_MASK is eventually selected as one of the C*D instructions, generate
1743 /// control word for byte/halfword/word insertion. This takes care of a single
1744 /// element move from V2 into V1.
1746 /// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
1747 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
1748 SDOperand V1 = Op.getOperand(0);
1749 SDOperand V2 = Op.getOperand(1);
1750 SDOperand PermMask = Op.getOperand(2);
1752 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1754 // If we have a single element being moved from V1 to V2, this can be handled
1755 // using the C*[DX] compute mask instructions, but the vector elements have
1756 // to be monotonically increasing with one exception element.
1757 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
1758 unsigned EltsFromV2 = 0;
1760 unsigned V2EltIdx0 = 0;
1761 unsigned CurrElt = 0;
1762 bool monotonic = true;
1763 if (EltVT == MVT::i8)
1765 else if (EltVT == MVT::i16)
1767 else if (EltVT == MVT::i32)
1770 assert(0 && "Unhandled vector type in LowerVECTOR_SHUFFLE");
1772 for (unsigned i = 0, e = PermMask.getNumOperands();
1773 EltsFromV2 <= 1 && monotonic && i != e;
1776 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1779 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1781 if (SrcElt >= V2EltIdx0) {
1783 V2Elt = (V2EltIdx0 - SrcElt) << 2;
1784 } else if (CurrElt != SrcElt) {
1791 if (EltsFromV2 == 1 && monotonic) {
1792 // Compute mask and shuffle
1793 MachineFunction &MF = DAG.getMachineFunction();
1794 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1795 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
1796 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1797 // Initialize temporary register to 0
1798 SDOperand InitTempReg =
1799 DAG.getCopyToReg(DAG.getEntryNode(), VReg, DAG.getConstant(0, PtrVT));
1800 // Copy register's contents as index in INSERT_MASK:
1801 SDOperand ShufMaskOp =
1802 DAG.getNode(SPUISD::INSERT_MASK, V1.getValueType(),
1803 DAG.getTargetConstant(V2Elt, MVT::i32),
1804 DAG.getCopyFromReg(InitTempReg, VReg, PtrVT));
1805 // Use shuffle mask in SHUFB synthetic instruction:
1806 return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V2, V1, ShufMaskOp);
1808 // Convert the SHUFFLE_VECTOR mask's input element units to the actual bytes.
1809 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
1811 SmallVector<SDOperand, 16> ResultMask;
1812 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
1814 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1817 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1819 for (unsigned j = 0; j != BytesPerElement; ++j) {
1820 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
1825 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
1826 &ResultMask[0], ResultMask.size());
1827 return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V1, V2, VPermMask);
1831 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1832 SDOperand Op0 = Op.getOperand(0); // Op0 = the scalar
1834 if (Op0.Val->getOpcode() == ISD::Constant) {
1835 // For a constant, build the appropriate constant vector, which will
1836 // eventually simplify to a vector register load.
1838 ConstantSDNode *CN = cast<ConstantSDNode>(Op0.Val);
1839 SmallVector<SDOperand, 16> ConstVecValues;
1843 // Create a constant vector:
1844 switch (Op.getValueType()) {
1845 default: assert(0 && "Unexpected constant value type in "
1846 "LowerSCALAR_TO_VECTOR");
1847 case MVT::v16i8: n_copies = 16; VT = MVT::i8; break;
1848 case MVT::v8i16: n_copies = 8; VT = MVT::i16; break;
1849 case MVT::v4i32: n_copies = 4; VT = MVT::i32; break;
1850 case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
1851 case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
1852 case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
1855 SDOperand CValue = DAG.getConstant(CN->getValue(), VT);
1856 for (size_t j = 0; j < n_copies; ++j)
1857 ConstVecValues.push_back(CValue);
1859 return DAG.getNode(ISD::BUILD_VECTOR, Op.getValueType(),
1860 &ConstVecValues[0], ConstVecValues.size());
1862 // Otherwise, copy the value from one register to another:
1863 switch (Op0.getValueType()) {
1864 default: assert(0 && "Unexpected value type in LowerSCALAR_TO_VECTOR");
1871 return DAG.getNode(SPUISD::PROMOTE_SCALAR, Op.getValueType(), Op0, Op0);
1878 static SDOperand LowerVectorMUL(SDOperand Op, SelectionDAG &DAG) {
1879 switch (Op.getValueType()) {
1881 SDOperand rA = Op.getOperand(0);
1882 SDOperand rB = Op.getOperand(1);
1883 SDOperand HiProd1 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rA, rB);
1884 SDOperand HiProd2 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rB, rA);
1885 SDOperand LoProd = DAG.getNode(SPUISD::MPYU, MVT::v4i32, rA, rB);
1886 SDOperand Residual1 = DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd1);
1888 return DAG.getNode(ISD::ADD, MVT::v4i32, Residual1, HiProd2);
1892 // Multiply two v8i16 vectors (pipeline friendly version):
1893 // a) multiply lower halves, mask off upper 16-bit of 32-bit product
1894 // b) multiply upper halves, rotate left by 16 bits (inserts 16 lower zeroes)
1895 // c) Use SELB to select upper and lower halves from the intermediate results
1897 // NOTE: We really want to move the FSMBI to earlier to actually get the
1898 // dual-issue. This code does manage to do this, even if it's a little on
1901 MachineFunction &MF = DAG.getMachineFunction();
1902 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1903 SDOperand Chain = Op.getOperand(0);
1904 SDOperand rA = Op.getOperand(0);
1905 SDOperand rB = Op.getOperand(1);
1906 unsigned FSMBIreg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1907 unsigned HiProdReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1910 DAG.getCopyToReg(Chain, FSMBIreg,
1911 DAG.getNode(SPUISD::FSMBI, MVT::v8i16,
1912 DAG.getConstant(0xcccc, MVT::i32)));
1915 DAG.getCopyToReg(FSMBOp, HiProdReg,
1916 DAG.getNode(SPUISD::MPYHH, MVT::v8i16, rA, rB));
1918 SDOperand HHProd_v4i32 =
1919 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
1920 DAG.getCopyFromReg(HHProd, HiProdReg, MVT::v4i32));
1922 return DAG.getNode(SPUISD::SELB, MVT::v8i16,
1923 DAG.getNode(SPUISD::MPY, MVT::v8i16, rA, rB),
1924 DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
1925 DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32,
1927 DAG.getConstant(16, MVT::i16))),
1928 DAG.getCopyFromReg(FSMBOp, FSMBIreg, MVT::v4i32));
1931 // This M00sE is N@stI! (apologies to Monty Python)
1933 // SPU doesn't know how to do any 8-bit multiplication, so the solution
1934 // is to break it all apart, sign extend, and reassemble the various
1935 // intermediate products.
1937 MachineFunction &MF = DAG.getMachineFunction();
1938 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1939 SDOperand Chain = Op.getOperand(0);
1940 SDOperand rA = Op.getOperand(0);
1941 SDOperand rB = Op.getOperand(1);
1942 SDOperand c8 = DAG.getConstant(8, MVT::i8);
1943 SDOperand c16 = DAG.getConstant(16, MVT::i8);
1945 unsigned FSMBreg_2222 = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1946 unsigned LoProd_reg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1947 unsigned HiProd_reg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1950 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1951 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rA),
1952 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rB));
1954 SDOperand rALH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rA, c8);
1956 SDOperand rBLH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rB, c8);
1959 DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16,
1960 DAG.getNode(SPUISD::MPY, MVT::v8i16, rALH, rBLH), c8);
1962 SDOperand FSMBdef_2222 =
1963 DAG.getCopyToReg(Chain, FSMBreg_2222,
1964 DAG.getNode(SPUISD::FSMBI, MVT::v8i16,
1965 DAG.getConstant(0x2222, MVT::i32)));
1967 SDOperand FSMBuse_2222 =
1968 DAG.getCopyFromReg(FSMBdef_2222, FSMBreg_2222, MVT::v4i32);
1970 SDOperand LoProd_1 =
1971 DAG.getCopyToReg(Chain, LoProd_reg,
1972 DAG.getNode(SPUISD::SELB, MVT::v8i16, LLProd, LHProd,
1975 SDOperand LoProdMask = DAG.getConstant(0xffff, MVT::i32);
1978 DAG.getNode(ISD::AND, MVT::v4i32,
1979 DAG.getCopyFromReg(LoProd_1, LoProd_reg, MVT::v4i32),
1980 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1981 LoProdMask, LoProdMask,
1982 LoProdMask, LoProdMask));
1985 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
1986 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rA), c16);
1989 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
1990 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rB), c16);
1993 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1994 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rAH),
1995 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rBH));
1997 SDOperand HHProd_1 =
1998 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1999 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
2000 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rAH, c8)),
2001 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
2002 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rBH, c8)));
2005 DAG.getCopyToReg(Chain, HiProd_reg,
2006 DAG.getNode(SPUISD::SELB, MVT::v8i16,
2008 DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16, HHProd_1, c8),
2012 DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32,
2013 DAG.getCopyFromReg(HHProd, HiProd_reg, MVT::v4i32), c16);
2015 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8,
2016 DAG.getNode(ISD::OR, MVT::v4i32,
2021 cerr << "CellSPU: Unknown vector multiplication, got "
2022 << MVT::getValueTypeString(Op.getValueType())
2031 static SDOperand LowerFDIVf32(SDOperand Op, SelectionDAG &DAG) {
2032 MachineFunction &MF = DAG.getMachineFunction();
2033 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2035 SDOperand A = Op.getOperand(0);
2036 SDOperand B = Op.getOperand(1);
2037 unsigned VT = Op.getValueType();
2039 unsigned VRegBR, VRegC;
2041 if (VT == MVT::f32) {
2042 VRegBR = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
2043 VRegC = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
2045 VRegBR = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
2046 VRegC = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
2048 // TODO: make sure we're feeding FPInterp the right arguments
2049 // Right now: fi B, frest(B)
2052 // (Floating Interpolate (FP Reciprocal Estimate B))
2054 DAG.getCopyToReg(DAG.getEntryNode(), VRegBR,
2055 DAG.getNode(SPUISD::FPInterp, VT, B,
2056 DAG.getNode(SPUISD::FPRecipEst, VT, B)));
2058 // Computes A * BRcpl and stores in a temporary register
2060 DAG.getCopyToReg(BRcpl, VRegC,
2061 DAG.getNode(ISD::FMUL, VT, A,
2062 DAG.getCopyFromReg(BRcpl, VRegBR, VT)));
2063 // What's the Chain variable do? It's magic!
2064 // TODO: set Chain = Op(0).getEntryNode()
2066 return DAG.getNode(ISD::FADD, VT,
2067 DAG.getCopyFromReg(AxBRcpl, VRegC, VT),
2068 DAG.getNode(ISD::FMUL, VT,
2069 DAG.getCopyFromReg(AxBRcpl, VRegBR, VT),
2070 DAG.getNode(ISD::FSUB, VT, A,
2071 DAG.getNode(ISD::FMUL, VT, B,
2072 DAG.getCopyFromReg(AxBRcpl, VRegC, VT)))));
2075 static SDOperand LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2076 unsigned VT = Op.getValueType();
2077 SDOperand N = Op.getOperand(0);
2078 SDOperand Elt = Op.getOperand(1);
2079 SDOperand ShufMask[16];
2080 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt);
2082 assert(C != 0 && "LowerEXTRACT_VECTOR_ELT expecting constant SDNode");
2084 int EltNo = (int) C->getValue();
2087 if (VT == MVT::i8 && EltNo >= 16)
2088 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i8 extraction slot > 15");
2089 else if (VT == MVT::i16 && EltNo >= 8)
2090 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i16 extraction slot > 7");
2091 else if (VT == MVT::i32 && EltNo >= 4)
2092 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i32 extraction slot > 4");
2093 else if (VT == MVT::i64 && EltNo >= 2)
2094 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i64 extraction slot > 2");
2096 if (EltNo == 0 && (VT == MVT::i32 || VT == MVT::i64)) {
2097 // i32 and i64: Element 0 is the preferred slot
2098 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT, N);
2101 // Need to generate shuffle mask and extract:
2102 int prefslot_begin = -1, prefslot_end = -1;
2103 int elt_byte = EltNo * MVT::getSizeInBits(VT) / 8;
2107 prefslot_begin = prefslot_end = 3;
2111 prefslot_begin = 2; prefslot_end = 3;
2115 prefslot_begin = 0; prefslot_end = 3;
2119 prefslot_begin = 0; prefslot_end = 7;
2124 assert(prefslot_begin != -1 && prefslot_end != -1 &&
2125 "LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
2127 for (int i = 0; i < 16; ++i) {
2128 // zero fill uppper part of preferred slot, don't care about the
2130 unsigned int mask_val;
2132 if (i <= prefslot_end) {
2134 ((i < prefslot_begin)
2136 : elt_byte + (i - prefslot_begin));
2138 ShufMask[i] = DAG.getConstant(mask_val, MVT::i8);
2140 ShufMask[i] = ShufMask[i % (prefslot_end + 1)];
2143 SDOperand ShufMaskVec =
2144 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2146 sizeof(ShufMask) / sizeof(ShufMask[0]));
2148 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
2149 DAG.getNode(SPUISD::SHUFB, N.getValueType(),
2150 N, N, ShufMaskVec));
2154 static SDOperand LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2155 SDOperand VecOp = Op.getOperand(0);
2156 SDOperand ValOp = Op.getOperand(1);
2157 SDOperand IdxOp = Op.getOperand(2);
2158 MVT::ValueType VT = Op.getValueType();
2160 ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
2161 assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
2163 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2164 // Use $2 because it's always 16-byte aligned and it's available:
2165 SDOperand PtrBase = DAG.getRegister(SPU::R2, PtrVT);
2168 DAG.getNode(SPUISD::SHUFB, VT,
2169 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, ValOp),
2171 DAG.getNode(SPUISD::INSERT_MASK, VT,
2172 DAG.getNode(ISD::ADD, PtrVT,
2174 DAG.getConstant(CN->getValue(),
2180 static SDOperand LowerI8Math(SDOperand Op, SelectionDAG &DAG, unsigned Opc) {
2181 SDOperand N0 = Op.getOperand(0); // Everything has at least one operand
2183 assert(Op.getValueType() == MVT::i8);
2186 assert(0 && "Unhandled i8 math operator");
2190 // 8-bit subtraction: Promote the arguments up to 16-bits and truncate
2192 SDOperand N1 = Op.getOperand(1);
2193 N0 = (N0.getOpcode() != ISD::Constant
2194 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2195 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2196 N1 = (N1.getOpcode() != ISD::Constant
2197 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N1)
2198 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2199 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2200 DAG.getNode(Opc, MVT::i16, N0, N1));
2204 SDOperand N1 = Op.getOperand(1);
2206 N0 = (N0.getOpcode() != ISD::Constant
2207 ? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
2208 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2209 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::ZERO_EXTEND : ISD::TRUNCATE);
2210 N1 = (N1.getOpcode() != ISD::Constant
2211 ? DAG.getNode(N1Opc, MVT::i16, N1)
2212 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2213 SDOperand ExpandArg =
2214 DAG.getNode(ISD::OR, MVT::i16, N0,
2215 DAG.getNode(ISD::SHL, MVT::i16,
2216 N0, DAG.getConstant(8, MVT::i16)));
2217 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2218 DAG.getNode(Opc, MVT::i16, ExpandArg, N1));
2222 SDOperand N1 = Op.getOperand(1);
2224 N0 = (N0.getOpcode() != ISD::Constant
2225 ? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
2226 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2227 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::ZERO_EXTEND : ISD::TRUNCATE);
2228 N1 = (N1.getOpcode() != ISD::Constant
2229 ? DAG.getNode(N1Opc, MVT::i16, N1)
2230 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2231 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2232 DAG.getNode(Opc, MVT::i16, N0, N1));
2235 SDOperand N1 = Op.getOperand(1);
2237 N0 = (N0.getOpcode() != ISD::Constant
2238 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2239 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2240 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::SIGN_EXTEND : ISD::TRUNCATE);
2241 N1 = (N1.getOpcode() != ISD::Constant
2242 ? DAG.getNode(N1Opc, MVT::i16, N1)
2243 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2244 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2245 DAG.getNode(Opc, MVT::i16, N0, N1));
2248 SDOperand N1 = Op.getOperand(1);
2250 N0 = (N0.getOpcode() != ISD::Constant
2251 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2252 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2253 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::SIGN_EXTEND : ISD::TRUNCATE);
2254 N1 = (N1.getOpcode() != ISD::Constant
2255 ? DAG.getNode(N1Opc, MVT::i16, N1)
2256 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2257 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2258 DAG.getNode(Opc, MVT::i16, N0, N1));
2266 //! Lower byte immediate operations for v16i8 vectors:
2268 LowerByteImmed(SDOperand Op, SelectionDAG &DAG) {
2271 MVT::ValueType VT = Op.getValueType();
2273 ConstVec = Op.getOperand(0);
2274 Arg = Op.getOperand(1);
2275 if (ConstVec.Val->getOpcode() != ISD::BUILD_VECTOR) {
2276 if (ConstVec.Val->getOpcode() == ISD::BIT_CONVERT) {
2277 ConstVec = ConstVec.getOperand(0);
2279 ConstVec = Op.getOperand(1);
2280 Arg = Op.getOperand(0);
2281 if (ConstVec.Val->getOpcode() == ISD::BIT_CONVERT) {
2282 ConstVec = ConstVec.getOperand(0);
2287 if (ConstVec.Val->getOpcode() == ISD::BUILD_VECTOR) {
2288 uint64_t VectorBits[2];
2289 uint64_t UndefBits[2];
2290 uint64_t SplatBits, SplatUndef;
2293 if (!GetConstantBuildVectorBits(ConstVec.Val, VectorBits, UndefBits)
2294 && isConstantSplat(VectorBits, UndefBits,
2295 MVT::getSizeInBits(MVT::getVectorElementType(VT)),
2296 SplatBits, SplatUndef, SplatSize)) {
2297 SDOperand tcVec[16];
2298 SDOperand tc = DAG.getTargetConstant(SplatBits & 0xff, MVT::i8);
2299 const size_t tcVecSize = sizeof(tcVec) / sizeof(tcVec[0]);
2301 // Turn the BUILD_VECTOR into a set of target constants:
2302 for (size_t i = 0; i < tcVecSize; ++i)
2305 return DAG.getNode(Op.Val->getOpcode(), VT, Arg,
2306 DAG.getNode(ISD::BUILD_VECTOR, VT, tcVec, tcVecSize));
2313 //! Lower i32 multiplication
2314 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG, unsigned VT,
2318 cerr << "CellSPU: Unknown LowerMUL value type, got "
2319 << MVT::getValueTypeString(Op.getValueType())
2325 SDOperand rA = Op.getOperand(0);
2326 SDOperand rB = Op.getOperand(1);
2328 return DAG.getNode(ISD::ADD, MVT::i32,
2329 DAG.getNode(ISD::ADD, MVT::i32,
2330 DAG.getNode(SPUISD::MPYH, MVT::i32, rA, rB),
2331 DAG.getNode(SPUISD::MPYH, MVT::i32, rB, rA)),
2332 DAG.getNode(SPUISD::MPYU, MVT::i32, rA, rB));
2339 //! Custom lowering for CTPOP (count population)
2341 Custom lowering code that counts the number ones in the input
2342 operand. SPU has such an instruction, but it counts the number of
2343 ones per byte, which then have to be accumulated.
2345 static SDOperand LowerCTPOP(SDOperand Op, SelectionDAG &DAG) {
2346 unsigned VT = Op.getValueType();
2347 unsigned vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
2351 SDOperand N = Op.getOperand(0);
2352 SDOperand Elt0 = DAG.getConstant(0, MVT::i32);
2354 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2355 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2357 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i8, CNTB, Elt0);
2361 MachineFunction &MF = DAG.getMachineFunction();
2362 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2364 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
2366 SDOperand N = Op.getOperand(0);
2367 SDOperand Elt0 = DAG.getConstant(0, MVT::i16);
2368 SDOperand Mask0 = DAG.getConstant(0x0f, MVT::i16);
2369 SDOperand Shift1 = DAG.getConstant(8, MVT::i16);
2371 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2372 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2374 // CNTB_result becomes the chain to which all of the virtual registers
2375 // CNTB_reg, SUM1_reg become associated:
2376 SDOperand CNTB_result =
2377 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, CNTB, Elt0);
2379 SDOperand CNTB_rescopy =
2380 DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
2382 SDOperand Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i16);
2384 return DAG.getNode(ISD::AND, MVT::i16,
2385 DAG.getNode(ISD::ADD, MVT::i16,
2386 DAG.getNode(ISD::SRL, MVT::i16,
2393 MachineFunction &MF = DAG.getMachineFunction();
2394 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2396 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2397 unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2399 SDOperand N = Op.getOperand(0);
2400 SDOperand Elt0 = DAG.getConstant(0, MVT::i32);
2401 SDOperand Mask0 = DAG.getConstant(0xff, MVT::i32);
2402 SDOperand Shift1 = DAG.getConstant(16, MVT::i32);
2403 SDOperand Shift2 = DAG.getConstant(8, MVT::i32);
2405 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2406 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2408 // CNTB_result becomes the chain to which all of the virtual registers
2409 // CNTB_reg, SUM1_reg become associated:
2410 SDOperand CNTB_result =
2411 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, CNTB, Elt0);
2413 SDOperand CNTB_rescopy =
2414 DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
2417 DAG.getNode(ISD::SRL, MVT::i32,
2418 DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32), Shift1);
2421 DAG.getNode(ISD::ADD, MVT::i32,
2422 Comp1, DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32));
2424 SDOperand Sum1_rescopy =
2425 DAG.getCopyToReg(CNTB_result, SUM1_reg, Sum1);
2428 DAG.getNode(ISD::SRL, MVT::i32,
2429 DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32),
2432 DAG.getNode(ISD::ADD, MVT::i32, Comp2,
2433 DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32));
2435 return DAG.getNode(ISD::AND, MVT::i32, Sum2, Mask0);
2445 /// LowerOperation - Provide custom lowering hooks for some operations.
2448 SPUTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG)
2450 switch (Op.getOpcode()) {
2452 cerr << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
2453 cerr << "Op.getOpcode() = " << Op.getOpcode() << "\n";
2454 cerr << "*Op.Val:\n";
2461 return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
2463 return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
2464 case ISD::ConstantPool:
2465 return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
2466 case ISD::GlobalAddress:
2467 return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
2468 case ISD::JumpTable:
2469 return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
2471 return LowerConstant(Op, DAG);
2472 case ISD::ConstantFP:
2473 return LowerConstantFP(Op, DAG);
2475 return LowerBRCOND(Op, DAG);
2476 case ISD::FORMAL_ARGUMENTS:
2477 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex);
2479 return LowerCALL(Op, DAG, SPUTM.getSubtargetImpl());
2481 return LowerRET(Op, DAG, getTargetMachine());
2490 return LowerI8Math(Op, DAG, Op.getOpcode());
2492 // Vector-related lowering.
2493 case ISD::BUILD_VECTOR:
2494 return LowerBUILD_VECTOR(Op, DAG);
2495 case ISD::SCALAR_TO_VECTOR:
2496 return LowerSCALAR_TO_VECTOR(Op, DAG);
2497 case ISD::VECTOR_SHUFFLE:
2498 return LowerVECTOR_SHUFFLE(Op, DAG);
2499 case ISD::EXTRACT_VECTOR_ELT:
2500 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2501 case ISD::INSERT_VECTOR_ELT:
2502 return LowerINSERT_VECTOR_ELT(Op, DAG);
2504 // Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
2508 return LowerByteImmed(Op, DAG);
2510 // Vector and i8 multiply:
2512 if (MVT::isVector(Op.getValueType()))
2513 return LowerVectorMUL(Op, DAG);
2514 else if (Op.getValueType() == MVT::i8)
2515 return LowerI8Math(Op, DAG, Op.getOpcode());
2517 return LowerMUL(Op, DAG, Op.getValueType(), Op.getOpcode());
2520 if (Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::v4f32)
2521 return LowerFDIVf32(Op, DAG);
2522 // else if (Op.getValueType() == MVT::f64)
2523 // return LowerFDIVf64(Op, DAG);
2525 assert(0 && "Calling FDIV on unsupported MVT");
2528 return LowerCTPOP(Op, DAG);
2534 //===----------------------------------------------------------------------===//
2535 // Target Optimization Hooks
2536 //===----------------------------------------------------------------------===//
2539 SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
2542 TargetMachine &TM = getTargetMachine();
2544 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
2545 SelectionDAG &DAG = DCI.DAG;
2546 SDOperand N0 = N->getOperand(0); // everything has at least one operand
2548 switch (N->getOpcode()) {
2550 case SPUISD::IndirectAddr: {
2551 if (!ST->usingLargeMem() && N0.getOpcode() == SPUISD::AFormAddr) {
2552 ConstantSDNode *CN = cast<ConstantSDNode>(N->getOperand(1));
2553 if (CN->getValue() == 0) {
2554 // (SPUindirect (SPUaform <addr>, 0), 0) ->
2555 // (SPUaform <addr>, 0)
2557 DEBUG(cerr << "Replace: ");
2558 DEBUG(N->dump(&DAG));
2559 DEBUG(cerr << "\nWith: ");
2560 DEBUG(N0.Val->dump(&DAG));
2561 DEBUG(cerr << "\n");
2568 SDOperand Op0 = N->getOperand(0);
2569 SDOperand Op1 = N->getOperand(1);
2571 if ((Op1.getOpcode() == ISD::Constant
2572 || Op1.getOpcode() == ISD::TargetConstant)
2573 && Op0.getOpcode() == SPUISD::IndirectAddr) {
2574 SDOperand Op01 = Op0.getOperand(1);
2575 if (Op01.getOpcode() == ISD::Constant
2576 || Op01.getOpcode() == ISD::TargetConstant) {
2577 // (add <const>, (SPUindirect <arg>, <const>)) ->
2578 // (SPUindirect <arg>, <const + const>)
2579 ConstantSDNode *CN0 = cast<ConstantSDNode>(Op1);
2580 ConstantSDNode *CN1 = cast<ConstantSDNode>(Op01);
2581 SDOperand combinedConst =
2582 DAG.getConstant(CN0->getValue() + CN1->getValue(),
2583 Op0.getValueType());
2585 DEBUG(cerr << "Replace: (add " << CN0->getValue() << ", "
2586 << "(SPUindirect <arg>, " << CN1->getValue() << "))\n");
2587 DEBUG(cerr << "With: (SPUindirect <arg>, "
2588 << CN0->getValue() + CN1->getValue() << ")\n");
2589 return DAG.getNode(SPUISD::IndirectAddr, Op0.getValueType(),
2590 Op0.getOperand(0), combinedConst);
2592 } else if ((Op0.getOpcode() == ISD::Constant
2593 || Op0.getOpcode() == ISD::TargetConstant)
2594 && Op1.getOpcode() == SPUISD::IndirectAddr) {
2595 SDOperand Op11 = Op1.getOperand(1);
2596 if (Op11.getOpcode() == ISD::Constant
2597 || Op11.getOpcode() == ISD::TargetConstant) {
2598 // (add (SPUindirect <arg>, <const>), <const>) ->
2599 // (SPUindirect <arg>, <const + const>)
2600 ConstantSDNode *CN0 = cast<ConstantSDNode>(Op0);
2601 ConstantSDNode *CN1 = cast<ConstantSDNode>(Op11);
2602 SDOperand combinedConst =
2603 DAG.getConstant(CN0->getValue() + CN1->getValue(),
2604 Op0.getValueType());
2606 DEBUG(cerr << "Replace: (add " << CN0->getValue() << ", "
2607 << "(SPUindirect <arg>, " << CN1->getValue() << "))\n");
2608 DEBUG(cerr << "With: (SPUindirect <arg>, "
2609 << CN0->getValue() + CN1->getValue() << ")\n");
2611 return DAG.getNode(SPUISD::IndirectAddr, Op1.getValueType(),
2612 Op1.getOperand(0), combinedConst);
2617 // Otherwise, return unchanged.
2621 //===----------------------------------------------------------------------===//
2622 // Inline Assembly Support
2623 //===----------------------------------------------------------------------===//
2625 /// getConstraintType - Given a constraint letter, return the type of
2626 /// constraint it is for this target.
2627 SPUTargetLowering::ConstraintType
2628 SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
2629 if (ConstraintLetter.size() == 1) {
2630 switch (ConstraintLetter[0]) {
2637 return C_RegisterClass;
2640 return TargetLowering::getConstraintType(ConstraintLetter);
2643 std::pair<unsigned, const TargetRegisterClass*>
2644 SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
2645 MVT::ValueType VT) const
2647 if (Constraint.size() == 1) {
2648 // GCC RS6000 Constraint Letters
2649 switch (Constraint[0]) {
2653 return std::make_pair(0U, SPU::R64CRegisterClass);
2654 return std::make_pair(0U, SPU::R32CRegisterClass);
2657 return std::make_pair(0U, SPU::R32FPRegisterClass);
2658 else if (VT == MVT::f64)
2659 return std::make_pair(0U, SPU::R64FPRegisterClass);
2662 return std::make_pair(0U, SPU::GPRCRegisterClass);
2666 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2670 SPUTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
2674 const SelectionDAG &DAG,
2675 unsigned Depth ) const {
2676 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
2679 // LowerAsmOperandForConstraint
2681 SPUTargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
2682 char ConstraintLetter,
2683 std::vector<SDOperand> &Ops,
2684 SelectionDAG &DAG) {
2685 // Default, for the time being, to the base class handler
2686 TargetLowering::LowerAsmOperandForConstraint(Op, ConstraintLetter, Ops, DAG);
2689 /// isLegalAddressImmediate - Return true if the integer value can be used
2690 /// as the offset of the target addressing mode.
2691 bool SPUTargetLowering::isLegalAddressImmediate(int64_t V, const Type *Ty) const {
2692 // SPU's addresses are 256K:
2693 return (V > -(1 << 18) && V < (1 << 18) - 1);
2696 bool SPUTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {