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 addRegisterClass(MVT::i8, SPU::R8CRegisterClass);
126 addRegisterClass(MVT::i16, SPU::R16CRegisterClass);
127 addRegisterClass(MVT::i32, SPU::R32CRegisterClass);
128 addRegisterClass(MVT::i64, SPU::R64CRegisterClass);
129 addRegisterClass(MVT::f32, SPU::R32FPRegisterClass);
130 addRegisterClass(MVT::f64, SPU::R64FPRegisterClass);
131 addRegisterClass(MVT::i128, SPU::GPRCRegisterClass);
133 // SPU has no sign or zero extended loads for i1, i8, i16:
134 setLoadXAction(ISD::EXTLOAD, MVT::i1, Promote);
135 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
136 setLoadXAction(ISD::ZEXTLOAD, MVT::i1, Promote);
137 setTruncStoreAction(MVT::i8, MVT::i1, Custom);
138 setTruncStoreAction(MVT::i16, MVT::i1, Custom);
139 setTruncStoreAction(MVT::i32, MVT::i1, Custom);
140 setTruncStoreAction(MVT::i64, MVT::i1, Custom);
141 setTruncStoreAction(MVT::i128, MVT::i1, Custom);
143 setLoadXAction(ISD::EXTLOAD, MVT::i8, Custom);
144 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Custom);
145 setLoadXAction(ISD::ZEXTLOAD, MVT::i8, Custom);
146 setTruncStoreAction(MVT::i8 , MVT::i8, Custom);
147 setTruncStoreAction(MVT::i16 , MVT::i8, Custom);
148 setTruncStoreAction(MVT::i32 , MVT::i8, Custom);
149 setTruncStoreAction(MVT::i64 , MVT::i8, Custom);
150 setTruncStoreAction(MVT::i128, MVT::i8, Custom);
152 setLoadXAction(ISD::EXTLOAD, MVT::i16, Custom);
153 setLoadXAction(ISD::SEXTLOAD, MVT::i16, Custom);
154 setLoadXAction(ISD::ZEXTLOAD, MVT::i16, Custom);
156 // SPU constant load actions are custom lowered:
157 setOperationAction(ISD::Constant, MVT::i64, Custom);
158 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
159 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
161 // SPU's loads and stores have to be custom lowered:
162 for (unsigned sctype = (unsigned) MVT::i1; sctype < (unsigned) MVT::f128;
164 setOperationAction(ISD::LOAD, sctype, Custom);
165 setOperationAction(ISD::STORE, sctype, Custom);
168 // Custom lower BRCOND for i1, i8 to "promote" the result to
169 // i32 and i16, respectively.
170 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
172 // Expand the jumptable branches
173 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
174 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
175 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
177 // SPU has no intrinsics for these particular operations:
178 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
180 // PowerPC has no SREM/UREM instructions
181 setOperationAction(ISD::SREM, MVT::i32, Expand);
182 setOperationAction(ISD::UREM, MVT::i32, Expand);
183 setOperationAction(ISD::SREM, MVT::i64, Expand);
184 setOperationAction(ISD::UREM, MVT::i64, Expand);
186 // We don't support sin/cos/sqrt/fmod
187 setOperationAction(ISD::FSIN , MVT::f64, Expand);
188 setOperationAction(ISD::FCOS , MVT::f64, Expand);
189 setOperationAction(ISD::FREM , MVT::f64, Expand);
190 setOperationAction(ISD::FSIN , MVT::f32, Expand);
191 setOperationAction(ISD::FCOS , MVT::f32, Expand);
192 setOperationAction(ISD::FREM , MVT::f32, Expand);
194 // If we're enabling GP optimizations, use hardware square root
195 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
196 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
198 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
199 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
201 // SPU can do rotate right and left, so legalize it... but customize for i8
202 // because instructions don't exist.
203 setOperationAction(ISD::ROTR, MVT::i32, Legal);
204 setOperationAction(ISD::ROTR, MVT::i16, Legal);
205 setOperationAction(ISD::ROTR, MVT::i8, Custom);
206 setOperationAction(ISD::ROTL, MVT::i32, Legal);
207 setOperationAction(ISD::ROTL, MVT::i16, Legal);
208 setOperationAction(ISD::ROTL, MVT::i8, Custom);
209 // SPU has no native version of shift left/right for i8
210 setOperationAction(ISD::SHL, MVT::i8, Custom);
211 setOperationAction(ISD::SRL, MVT::i8, Custom);
212 setOperationAction(ISD::SRA, MVT::i8, Custom);
213 // And SPU needs custom lowering for shift left/right for i64
214 setOperationAction(ISD::SHL, MVT::i64, Custom);
215 setOperationAction(ISD::SRL, MVT::i64, Custom);
216 setOperationAction(ISD::SRA, MVT::i64, Custom);
218 // Custom lower i32 multiplications
219 setOperationAction(ISD::MUL, MVT::i32, Custom);
221 // Need to custom handle (some) common i8 math ops
222 setOperationAction(ISD::SUB, MVT::i8, Custom);
223 setOperationAction(ISD::MUL, MVT::i8, Custom);
225 // SPU does not have BSWAP. It does have i32 support CTLZ.
226 // CTPOP has to be custom lowered.
227 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
228 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
230 setOperationAction(ISD::CTPOP, MVT::i8, Custom);
231 setOperationAction(ISD::CTPOP, MVT::i16, Custom);
232 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
233 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
235 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
236 setOperationAction(ISD::CTTZ , MVT::i64, Expand);
238 setOperationAction(ISD::CTLZ , MVT::i32, Legal);
240 // SPU has a version of select that implements (a&~c)|(b|c), just like
241 // select ought to work:
242 setOperationAction(ISD::SELECT, MVT::i1, Promote);
243 setOperationAction(ISD::SELECT, MVT::i8, Legal);
244 setOperationAction(ISD::SELECT, MVT::i16, Legal);
245 setOperationAction(ISD::SELECT, MVT::i32, Legal);
246 setOperationAction(ISD::SELECT, MVT::i64, Expand);
248 setOperationAction(ISD::SETCC, MVT::i1, Promote);
249 setOperationAction(ISD::SETCC, MVT::i8, Legal);
250 setOperationAction(ISD::SETCC, MVT::i16, Legal);
251 setOperationAction(ISD::SETCC, MVT::i32, Legal);
252 setOperationAction(ISD::SETCC, MVT::i64, Expand);
254 // Zero extension and sign extension for i64 have to be
256 setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
257 setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
258 setOperationAction(ISD::ANY_EXTEND, MVT::i64, Custom);
260 // SPU has a legal FP -> signed INT instruction
261 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
262 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
263 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
264 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
266 // FDIV on SPU requires custom lowering
267 setOperationAction(ISD::FDIV, MVT::f32, Custom);
268 //setOperationAction(ISD::FDIV, MVT::f64, Custom);
270 // SPU has [U|S]INT_TO_FP
271 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
272 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
273 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
274 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
275 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
276 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
277 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
278 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
280 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Legal);
281 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Legal);
282 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Legal);
283 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Legal);
285 // We cannot sextinreg(i1). Expand to shifts.
286 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
288 // Support label based line numbers.
289 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
290 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
292 // We want to legalize GlobalAddress and ConstantPool nodes into the
293 // appropriate instructions to materialize the address.
294 for (unsigned sctype = (unsigned) MVT::i1; sctype < (unsigned) MVT::f128;
296 setOperationAction(ISD::GlobalAddress, sctype, Custom);
297 setOperationAction(ISD::ConstantPool, sctype, Custom);
298 setOperationAction(ISD::JumpTable, sctype, 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 setShiftAmountType(MVT::i32);
374 setSetCCResultContents(ZeroOrOneSetCCResult);
376 setStackPointerRegisterToSaveRestore(SPU::R1);
378 // We have target-specific dag combine patterns for the following nodes:
379 setTargetDAGCombine(ISD::ADD);
380 setTargetDAGCombine(ISD::ZERO_EXTEND);
381 setTargetDAGCombine(ISD::SIGN_EXTEND);
382 setTargetDAGCombine(ISD::ANY_EXTEND);
384 computeRegisterProperties();
388 SPUTargetLowering::getTargetNodeName(unsigned Opcode) const
390 if (node_names.empty()) {
391 node_names[(unsigned) SPUISD::RET_FLAG] = "SPUISD::RET_FLAG";
392 node_names[(unsigned) SPUISD::Hi] = "SPUISD::Hi";
393 node_names[(unsigned) SPUISD::Lo] = "SPUISD::Lo";
394 node_names[(unsigned) SPUISD::PCRelAddr] = "SPUISD::PCRelAddr";
395 node_names[(unsigned) SPUISD::AFormAddr] = "SPUISD::AFormAddr";
396 node_names[(unsigned) SPUISD::IndirectAddr] = "SPUISD::IndirectAddr";
397 node_names[(unsigned) SPUISD::LDRESULT] = "SPUISD::LDRESULT";
398 node_names[(unsigned) SPUISD::CALL] = "SPUISD::CALL";
399 node_names[(unsigned) SPUISD::SHUFB] = "SPUISD::SHUFB";
400 node_names[(unsigned) SPUISD::INSERT_MASK] = "SPUISD::INSERT_MASK";
401 node_names[(unsigned) SPUISD::CNTB] = "SPUISD::CNTB";
402 node_names[(unsigned) SPUISD::PROMOTE_SCALAR] = "SPUISD::PROMOTE_SCALAR";
403 node_names[(unsigned) SPUISD::EXTRACT_ELT0] = "SPUISD::EXTRACT_ELT0";
404 node_names[(unsigned) SPUISD::EXTRACT_ELT0_CHAINED] = "SPUISD::EXTRACT_ELT0_CHAINED";
405 node_names[(unsigned) SPUISD::EXTRACT_I1_ZEXT] = "SPUISD::EXTRACT_I1_ZEXT";
406 node_names[(unsigned) SPUISD::EXTRACT_I1_SEXT] = "SPUISD::EXTRACT_I1_SEXT";
407 node_names[(unsigned) SPUISD::EXTRACT_I8_ZEXT] = "SPUISD::EXTRACT_I8_ZEXT";
408 node_names[(unsigned) SPUISD::EXTRACT_I8_SEXT] = "SPUISD::EXTRACT_I8_SEXT";
409 node_names[(unsigned) SPUISD::MPY] = "SPUISD::MPY";
410 node_names[(unsigned) SPUISD::MPYU] = "SPUISD::MPYU";
411 node_names[(unsigned) SPUISD::MPYH] = "SPUISD::MPYH";
412 node_names[(unsigned) SPUISD::MPYHH] = "SPUISD::MPYHH";
413 node_names[(unsigned) SPUISD::SHLQUAD_L_BITS] = "SPUISD::SHLQUAD_L_BITS";
414 node_names[(unsigned) SPUISD::SHLQUAD_L_BYTES] = "SPUISD::SHLQUAD_L_BYTES";
415 node_names[(unsigned) SPUISD::VEC_SHL] = "SPUISD::VEC_SHL";
416 node_names[(unsigned) SPUISD::VEC_SRL] = "SPUISD::VEC_SRL";
417 node_names[(unsigned) SPUISD::VEC_SRA] = "SPUISD::VEC_SRA";
418 node_names[(unsigned) SPUISD::VEC_ROTL] = "SPUISD::VEC_ROTL";
419 node_names[(unsigned) SPUISD::VEC_ROTR] = "SPUISD::VEC_ROTR";
420 node_names[(unsigned) SPUISD::ROTQUAD_RZ_BYTES] =
421 "SPUISD::ROTQUAD_RZ_BYTES";
422 node_names[(unsigned) SPUISD::ROTQUAD_RZ_BITS] =
423 "SPUISD::ROTQUAD_RZ_BITS";
424 node_names[(unsigned) SPUISD::ROTBYTES_RIGHT_S] =
425 "SPUISD::ROTBYTES_RIGHT_S";
426 node_names[(unsigned) SPUISD::ROTBYTES_LEFT] = "SPUISD::ROTBYTES_LEFT";
427 node_names[(unsigned) SPUISD::ROTBYTES_LEFT_CHAINED] =
428 "SPUISD::ROTBYTES_LEFT_CHAINED";
429 node_names[(unsigned) SPUISD::FSMBI] = "SPUISD::FSMBI";
430 node_names[(unsigned) SPUISD::SELB] = "SPUISD::SELB";
431 node_names[(unsigned) SPUISD::FPInterp] = "SPUISD::FPInterp";
432 node_names[(unsigned) SPUISD::FPRecipEst] = "SPUISD::FPRecipEst";
433 node_names[(unsigned) SPUISD::SEXT32TO64] = "SPUISD::SEXT32TO64";
436 std::map<unsigned, const char *>::iterator i = node_names.find(Opcode);
438 return ((i != node_names.end()) ? i->second : 0);
442 SPUTargetLowering::getSetCCResultType(const SDOperand &Op) const {
443 MVT::ValueType VT = Op.getValueType();
444 if (MVT::isInteger(VT))
450 //===----------------------------------------------------------------------===//
451 // Calling convention code:
452 //===----------------------------------------------------------------------===//
454 #include "SPUGenCallingConv.inc"
456 //===----------------------------------------------------------------------===//
457 // LowerOperation implementation
458 //===----------------------------------------------------------------------===//
460 /// Aligned load common code for CellSPU
462 \param[in] Op The SelectionDAG load or store operand
463 \param[in] DAG The selection DAG
464 \param[in] ST CellSPU subtarget information structure
465 \param[in,out] alignment Caller initializes this to the load or store node's
466 value from getAlignment(), may be updated while generating the aligned load
467 \param[in,out] alignOffs Aligned offset; set by AlignedLoad to the aligned
468 offset (divisible by 16, modulo 16 == 0)
469 \param[in,out] prefSlotOffs Preferred slot offset; set by AlignedLoad to the
470 offset of the preferred slot (modulo 16 != 0)
471 \param[in,out] VT Caller initializes this value type to the the load or store
472 node's loaded or stored value type; may be updated if an i1-extended load or
474 \param[out] was16aligned true if the base pointer had 16-byte alignment,
475 otherwise false. Can help to determine if the chunk needs to be rotated.
477 Both load and store lowering load a block of data aligned on a 16-byte
478 boundary. This is the common aligned load code shared between both.
481 AlignedLoad(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST,
483 unsigned &alignment, int &alignOffs, int &prefSlotOffs,
484 MVT::ValueType &VT, bool &was16aligned)
486 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
487 const valtype_map_s *vtm = getValueTypeMapEntry(VT);
488 SDOperand basePtr = LSN->getBasePtr();
489 SDOperand chain = LSN->getChain();
491 if (basePtr.getOpcode() == ISD::ADD) {
492 SDOperand Op1 = basePtr.Val->getOperand(1);
494 if (Op1.getOpcode() == ISD::Constant || Op1.getOpcode() == ISD::TargetConstant) {
495 const ConstantSDNode *CN = cast<ConstantSDNode>(basePtr.getOperand(1));
497 alignOffs = (int) CN->getValue();
498 prefSlotOffs = (int) (alignOffs & 0xf);
500 // Adjust the rotation amount to ensure that the final result ends up in
501 // the preferred slot:
502 prefSlotOffs -= vtm->prefslot_byte;
503 basePtr = basePtr.getOperand(0);
505 // Loading from memory, can we adjust alignment?
506 if (basePtr.getOpcode() == SPUISD::AFormAddr) {
507 SDOperand APtr = basePtr.getOperand(0);
508 if (APtr.getOpcode() == ISD::TargetGlobalAddress) {
509 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(APtr);
510 alignment = GSDN->getGlobal()->getAlignment();
515 prefSlotOffs = -vtm->prefslot_byte;
519 prefSlotOffs = -vtm->prefslot_byte;
522 if (alignment == 16) {
523 // Realign the base pointer as a D-Form address:
524 if (!isMemoryOperand(basePtr) || (alignOffs & ~0xf) != 0) {
525 basePtr = DAG.getNode(ISD::ADD, PtrVT,
527 DAG.getConstant((alignOffs & ~0xf), PtrVT));
530 // Emit the vector load:
532 return DAG.getLoad(MVT::v16i8, chain, basePtr,
533 LSN->getSrcValue(), LSN->getSrcValueOffset(),
534 LSN->isVolatile(), 16);
537 // Unaligned load or we're using the "large memory" model, which means that
538 // we have to be very pessimistic:
539 if (isMemoryOperand(basePtr) || isIndirectOperand(basePtr)) {
540 basePtr = DAG.getNode(SPUISD::IndirectAddr, PtrVT, basePtr, DAG.getConstant(0, PtrVT));
544 basePtr = DAG.getNode(ISD::ADD, PtrVT, basePtr,
545 DAG.getConstant((alignOffs & ~0xf), PtrVT));
546 was16aligned = false;
547 return DAG.getLoad(MVT::v16i8, chain, basePtr,
548 LSN->getSrcValue(), LSN->getSrcValueOffset(),
549 LSN->isVolatile(), 16);
552 /// Custom lower loads for CellSPU
554 All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
555 within a 16-byte block, we have to rotate to extract the requested element.
558 LowerLOAD(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
559 LoadSDNode *LN = cast<LoadSDNode>(Op);
560 SDOperand the_chain = LN->getChain();
561 MVT::ValueType VT = LN->getMemoryVT();
562 MVT::ValueType OpVT = Op.Val->getValueType(0);
563 ISD::LoadExtType ExtType = LN->getExtensionType();
564 unsigned alignment = LN->getAlignment();
567 switch (LN->getAddressingMode()) {
568 case ISD::UNINDEXED: {
572 AlignedLoad(Op, DAG, ST, LN,alignment, offset, rotamt, VT, was16aligned);
577 the_chain = result.getValue(1);
578 // Rotate the chunk if necessary
581 if (rotamt != 0 || !was16aligned) {
582 SDVTList vecvts = DAG.getVTList(MVT::v16i8, MVT::Other);
587 Ops[2] = DAG.getConstant(rotamt, MVT::i16);
589 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
590 LoadSDNode *LN1 = cast<LoadSDNode>(result);
591 Ops[2] = DAG.getNode(ISD::ADD, PtrVT, LN1->getBasePtr(),
592 DAG.getConstant(rotamt, PtrVT));
595 result = DAG.getNode(SPUISD::ROTBYTES_LEFT_CHAINED, vecvts, Ops, 3);
596 the_chain = result.getValue(1);
599 if (VT == OpVT || ExtType == ISD::EXTLOAD) {
601 MVT::ValueType vecVT = MVT::v16i8;
603 // Convert the loaded v16i8 vector to the appropriate vector type
604 // specified by the operand:
607 vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
609 vecVT = MVT::getVectorType(OpVT, (128 / MVT::getSizeInBits(OpVT)));
612 Ops[1] = DAG.getNode(ISD::BIT_CONVERT, vecVT, result);
613 scalarvts = DAG.getVTList((OpVT == VT ? VT : OpVT), MVT::Other);
614 result = DAG.getNode(SPUISD::EXTRACT_ELT0_CHAINED, scalarvts, Ops, 2);
615 the_chain = result.getValue(1);
617 // Handle the sign and zero-extending loads for i1 and i8:
620 if (ExtType == ISD::SEXTLOAD) {
621 NewOpC = (OpVT == MVT::i1
622 ? SPUISD::EXTRACT_I1_SEXT
623 : SPUISD::EXTRACT_I8_SEXT);
625 assert(ExtType == ISD::ZEXTLOAD);
626 NewOpC = (OpVT == MVT::i1
627 ? SPUISD::EXTRACT_I1_ZEXT
628 : SPUISD::EXTRACT_I8_ZEXT);
631 result = DAG.getNode(NewOpC, OpVT, result);
634 SDVTList retvts = DAG.getVTList(OpVT, MVT::Other);
635 SDOperand retops[2] = {
640 result = DAG.getNode(SPUISD::LDRESULT, retvts,
641 retops, sizeof(retops) / sizeof(retops[0]));
648 case ISD::LAST_INDEXED_MODE:
649 cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
651 cerr << (unsigned) LN->getAddressingMode() << "\n";
659 /// Custom lower stores for CellSPU
661 All CellSPU stores are aligned to 16-byte boundaries, so for elements
662 within a 16-byte block, we have to generate a shuffle to insert the
663 requested element into its place, then store the resulting block.
666 LowerSTORE(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
667 StoreSDNode *SN = cast<StoreSDNode>(Op);
668 SDOperand Value = SN->getValue();
669 MVT::ValueType VT = Value.getValueType();
670 MVT::ValueType StVT = (!SN->isTruncatingStore() ? VT : SN->getMemoryVT());
671 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
672 unsigned alignment = SN->getAlignment();
674 switch (SN->getAddressingMode()) {
675 case ISD::UNINDEXED: {
676 int chunk_offset, slot_offset;
679 // The vector type we really want to load from the 16-byte chunk, except
680 // in the case of MVT::i1, which has to be v16i8.
681 unsigned vecVT, stVecVT = MVT::v16i8;
684 stVecVT = MVT::getVectorType(StVT, (128 / MVT::getSizeInBits(StVT)));
685 vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
687 SDOperand alignLoadVec =
688 AlignedLoad(Op, DAG, ST, SN, alignment,
689 chunk_offset, slot_offset, VT, was16aligned);
691 if (alignLoadVec.Val == 0)
694 LoadSDNode *LN = cast<LoadSDNode>(alignLoadVec);
695 SDOperand basePtr = LN->getBasePtr();
696 SDOperand the_chain = alignLoadVec.getValue(1);
697 SDOperand theValue = SN->getValue();
701 && (theValue.getOpcode() == ISD::AssertZext
702 || theValue.getOpcode() == ISD::AssertSext)) {
703 // Drill down and get the value for zero- and sign-extended
705 theValue = theValue.getOperand(0);
710 SDOperand insertEltOffs = DAG.getConstant(chunk_offset, PtrVT);
711 SDOperand insertEltPtr;
712 SDOperand insertEltOp;
714 // If the base pointer is already a D-form address, then just create
715 // a new D-form address with a slot offset and the orignal base pointer.
716 // Otherwise generate a D-form address with the slot offset relative
717 // to the stack pointer, which is always aligned.
718 DEBUG(cerr << "CellSPU LowerSTORE: basePtr = ");
719 DEBUG(basePtr.Val->dump(&DAG));
722 if (basePtr.getOpcode() == SPUISD::IndirectAddr ||
723 (basePtr.getOpcode() == ISD::ADD
724 && basePtr.getOperand(0).getOpcode() == SPUISD::IndirectAddr)) {
725 insertEltPtr = basePtr;
727 insertEltPtr = DAG.getNode(ISD::ADD, PtrVT, basePtr, insertEltOffs);
730 insertEltOp = DAG.getNode(SPUISD::INSERT_MASK, stVecVT, insertEltPtr);
731 result = DAG.getNode(SPUISD::SHUFB, vecVT,
732 DAG.getNode(ISD::SCALAR_TO_VECTOR, vecVT, theValue),
734 DAG.getNode(ISD::BIT_CONVERT, vecVT, insertEltOp));
736 result = DAG.getStore(the_chain, result, basePtr,
737 LN->getSrcValue(), LN->getSrcValueOffset(),
738 LN->isVolatile(), LN->getAlignment());
747 case ISD::LAST_INDEXED_MODE:
748 cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
750 cerr << (unsigned) SN->getAddressingMode() << "\n";
758 /// Generate the address of a constant pool entry.
760 LowerConstantPool(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
761 MVT::ValueType PtrVT = Op.getValueType();
762 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
763 Constant *C = CP->getConstVal();
764 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
765 SDOperand Zero = DAG.getConstant(0, PtrVT);
766 const TargetMachine &TM = DAG.getTarget();
768 if (TM.getRelocationModel() == Reloc::Static) {
769 if (!ST->usingLargeMem()) {
770 // Just return the SDOperand with the constant pool address in it.
771 return DAG.getNode(SPUISD::AFormAddr, PtrVT, CPI, Zero);
773 SDOperand Hi = DAG.getNode(SPUISD::Hi, PtrVT, CPI, Zero);
774 SDOperand Lo = DAG.getNode(SPUISD::Lo, PtrVT, CPI, Zero);
775 return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
780 "LowerConstantPool: Relocation model other than static not supported.");
785 LowerJumpTable(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
786 MVT::ValueType PtrVT = Op.getValueType();
787 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
788 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
789 SDOperand Zero = DAG.getConstant(0, PtrVT);
790 const TargetMachine &TM = DAG.getTarget();
792 if (TM.getRelocationModel() == Reloc::Static) {
793 if (!ST->usingLargeMem()) {
794 return DAG.getNode(SPUISD::AFormAddr, PtrVT, JTI, Zero);
796 SDOperand Hi = DAG.getNode(SPUISD::Hi, PtrVT, JTI, Zero);
797 SDOperand Lo = DAG.getNode(SPUISD::Lo, PtrVT, JTI, Zero);
798 return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
803 "LowerJumpTable: Relocation model other than static not supported.");
808 LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
809 MVT::ValueType PtrVT = Op.getValueType();
810 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
811 GlobalValue *GV = GSDN->getGlobal();
812 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
813 const TargetMachine &TM = DAG.getTarget();
814 SDOperand Zero = DAG.getConstant(0, PtrVT);
816 if (TM.getRelocationModel() == Reloc::Static) {
817 if (!ST->usingLargeMem()) {
818 return DAG.getNode(SPUISD::AFormAddr, PtrVT, GA, Zero);
820 SDOperand Hi = DAG.getNode(SPUISD::Hi, PtrVT, GA, Zero);
821 SDOperand Lo = DAG.getNode(SPUISD::Lo, PtrVT, GA, Zero);
822 return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
825 cerr << "LowerGlobalAddress: Relocation model other than static not "
834 //! Custom lower i64 integer constants
836 This code inserts all of the necessary juggling that needs to occur to load
837 a 64-bit constant into a register.
840 LowerConstant(SDOperand Op, SelectionDAG &DAG) {
841 unsigned VT = Op.getValueType();
842 ConstantSDNode *CN = cast<ConstantSDNode>(Op.Val);
844 if (VT == MVT::i64) {
845 SDOperand T = DAG.getConstant(CN->getValue(), MVT::i64);
846 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
847 DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
849 cerr << "LowerConstant: unhandled constant type "
850 << MVT::getValueTypeString(VT)
859 //! Custom lower double precision floating point constants
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::f64) {
869 uint64_t dbits = DoubleToBits(FP->getValueAPF().convertToDouble());
870 return DAG.getNode(ISD::BIT_CONVERT, VT,
871 LowerConstant(DAG.getConstant(dbits, MVT::i64), DAG));
877 //! Lower MVT::i1, MVT::i8 brcond to a promoted type (MVT::i32, MVT::i16)
879 LowerBRCOND(SDOperand Op, SelectionDAG &DAG)
881 SDOperand Cond = Op.getOperand(1);
882 MVT::ValueType CondVT = Cond.getValueType();
883 MVT::ValueType CondNVT;
885 if (CondVT == MVT::i1 || CondVT == MVT::i8) {
886 CondNVT = (CondVT == MVT::i1 ? MVT::i32 : MVT::i16);
887 return DAG.getNode(ISD::BRCOND, Op.getValueType(),
889 DAG.getNode(ISD::ZERO_EXTEND, CondNVT, Op.getOperand(1)),
892 return SDOperand(); // Unchanged
896 LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG, int &VarArgsFrameIndex)
898 MachineFunction &MF = DAG.getMachineFunction();
899 MachineFrameInfo *MFI = MF.getFrameInfo();
900 MachineRegisterInfo &RegInfo = MF.getRegInfo();
901 SmallVector<SDOperand, 8> ArgValues;
902 SDOperand Root = Op.getOperand(0);
903 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
905 const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
906 const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
908 unsigned ArgOffset = SPUFrameInfo::minStackSize();
909 unsigned ArgRegIdx = 0;
910 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
912 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
914 // Add DAG nodes to load the arguments or copy them out of registers.
915 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
917 bool needsLoad = false;
918 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
919 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
923 cerr << "LowerFORMAL_ARGUMENTS Unhandled argument type: "
924 << MVT::getValueTypeString(ObjectVT)
929 if (!isVarArg && ArgRegIdx < NumArgRegs) {
930 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R8CRegClass);
931 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
932 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i8);
939 if (!isVarArg && ArgRegIdx < NumArgRegs) {
940 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
941 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
942 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i16);
949 if (!isVarArg && ArgRegIdx < NumArgRegs) {
950 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
951 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
952 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
959 if (!isVarArg && ArgRegIdx < NumArgRegs) {
960 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64CRegClass);
961 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
962 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
969 if (!isVarArg && ArgRegIdx < NumArgRegs) {
970 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
971 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
972 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f32);
979 if (!isVarArg && ArgRegIdx < NumArgRegs) {
980 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64FPRegClass);
981 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
982 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f64);
994 if (!isVarArg && ArgRegIdx < NumArgRegs) {
995 unsigned VReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
996 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
997 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1005 // We need to load the argument to a virtual register if we determined above
1006 // that we ran out of physical registers of the appropriate type
1008 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1009 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1010 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1011 ArgOffset += StackSlotSize;
1014 ArgValues.push_back(ArgVal);
1017 // If the function takes variable number of arguments, make a frame index for
1018 // the start of the first vararg value... for expansion of llvm.va_start.
1020 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1022 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1023 // If this function is vararg, store any remaining integer argument regs to
1024 // their spots on the stack so that they may be loaded by deferencing the
1025 // result of va_next.
1026 SmallVector<SDOperand, 8> MemOps;
1027 for (; ArgRegIdx != NumArgRegs; ++ArgRegIdx) {
1028 unsigned VReg = RegInfo.createVirtualRegister(&SPU::GPRCRegClass);
1029 RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
1030 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1031 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1032 MemOps.push_back(Store);
1033 // Increment the address by four for the next argument to store
1034 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1035 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1037 if (!MemOps.empty())
1038 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1041 ArgValues.push_back(Root);
1043 // Return the new list of results.
1044 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1045 Op.Val->value_end());
1046 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1049 /// isLSAAddress - Return the immediate to use if the specified
1050 /// value is representable as a LSA address.
1051 static SDNode *isLSAAddress(SDOperand Op, SelectionDAG &DAG) {
1052 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1055 int Addr = C->getValue();
1056 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1057 (Addr << 14 >> 14) != Addr)
1058 return 0; // Top 14 bits have to be sext of immediate.
1060 return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).Val;
1065 LowerCALL(SDOperand Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1066 SDOperand Chain = Op.getOperand(0);
1068 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1069 bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
1071 SDOperand Callee = Op.getOperand(4);
1072 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1073 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
1074 const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
1075 const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
1077 // Handy pointer type
1078 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1080 // Accumulate how many bytes are to be pushed on the stack, including the
1081 // linkage area, and parameter passing area. According to the SPU ABI,
1082 // we minimally need space for [LR] and [SP]
1083 unsigned NumStackBytes = SPUFrameInfo::minStackSize();
1085 // Set up a copy of the stack pointer for use loading and storing any
1086 // arguments that may not fit in the registers available for argument
1088 SDOperand StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
1090 // Figure out which arguments are going to go in registers, and which in
1092 unsigned ArgOffset = SPUFrameInfo::minStackSize(); // Just below [LR]
1093 unsigned ArgRegIdx = 0;
1095 // Keep track of registers passing arguments
1096 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1097 // And the arguments passed on the stack
1098 SmallVector<SDOperand, 8> MemOpChains;
1100 for (unsigned i = 0; i != NumOps; ++i) {
1101 SDOperand Arg = Op.getOperand(5+2*i);
1103 // PtrOff will be used to store the current argument to the stack if a
1104 // register cannot be found for it.
1105 SDOperand PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1106 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1108 switch (Arg.getValueType()) {
1109 default: assert(0 && "Unexpected ValueType for argument!");
1113 if (ArgRegIdx != NumArgRegs) {
1114 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1116 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1117 ArgOffset += StackSlotSize;
1122 if (ArgRegIdx != NumArgRegs) {
1123 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1125 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1126 ArgOffset += StackSlotSize;
1133 if (ArgRegIdx != NumArgRegs) {
1134 RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
1136 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1137 ArgOffset += StackSlotSize;
1143 // Update number of stack bytes actually used, insert a call sequence start
1144 NumStackBytes = (ArgOffset - SPUFrameInfo::minStackSize());
1145 Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumStackBytes, PtrVT));
1147 if (!MemOpChains.empty()) {
1148 // Adjust the stack pointer for the stack arguments.
1149 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1150 &MemOpChains[0], MemOpChains.size());
1153 // Build a sequence of copy-to-reg nodes chained together with token chain
1154 // and flag operands which copy the outgoing args into the appropriate regs.
1156 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1157 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1159 InFlag = Chain.getValue(1);
1162 std::vector<MVT::ValueType> NodeTys;
1163 NodeTys.push_back(MVT::Other); // Returns a chain
1164 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1166 SmallVector<SDOperand, 8> Ops;
1167 unsigned CallOpc = SPUISD::CALL;
1169 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1170 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1171 // node so that legalize doesn't hack it.
1172 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1173 GlobalValue *GV = G->getGlobal();
1174 unsigned CalleeVT = Callee.getValueType();
1175 SDOperand Zero = DAG.getConstant(0, PtrVT);
1176 SDOperand GA = DAG.getTargetGlobalAddress(GV, CalleeVT);
1178 if (!ST->usingLargeMem()) {
1179 // Turn calls to targets that are defined (i.e., have bodies) into BRSL
1180 // style calls, otherwise, external symbols are BRASL calls. This assumes
1181 // that declared/defined symbols are in the same compilation unit and can
1182 // be reached through PC-relative jumps.
1185 // This may be an unsafe assumption for JIT and really large compilation
1187 if (GV->isDeclaration()) {
1188 Callee = DAG.getNode(SPUISD::AFormAddr, CalleeVT, GA, Zero);
1190 Callee = DAG.getNode(SPUISD::PCRelAddr, CalleeVT, GA, Zero);
1193 // "Large memory" mode: Turn all calls into indirect calls with a X-form
1195 Callee = DAG.getNode(SPUISD::IndirectAddr, PtrVT, GA, Zero);
1197 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1198 Callee = DAG.getExternalSymbol(S->getSymbol(), Callee.getValueType());
1199 else if (SDNode *Dest = isLSAAddress(Callee, DAG)) {
1200 // If this is an absolute destination address that appears to be a legal
1201 // local store address, use the munged value.
1202 Callee = SDOperand(Dest, 0);
1205 Ops.push_back(Chain);
1206 Ops.push_back(Callee);
1208 // Add argument registers to the end of the list so that they are known live
1210 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1211 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1212 RegsToPass[i].second.getValueType()));
1215 Ops.push_back(InFlag);
1216 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1217 InFlag = Chain.getValue(1);
1219 Chain = DAG.getCALLSEQ_END(Chain,
1220 DAG.getConstant(NumStackBytes, PtrVT),
1221 DAG.getConstant(0, PtrVT),
1223 if (Op.Val->getValueType(0) != MVT::Other)
1224 InFlag = Chain.getValue(1);
1226 SDOperand ResultVals[3];
1227 unsigned NumResults = 0;
1230 // If the call has results, copy the values out of the ret val registers.
1231 switch (Op.Val->getValueType(0)) {
1232 default: assert(0 && "Unexpected ret value!");
1233 case MVT::Other: break;
1235 if (Op.Val->getValueType(1) == MVT::i32) {
1236 Chain = DAG.getCopyFromReg(Chain, SPU::R4, MVT::i32, InFlag).getValue(1);
1237 ResultVals[0] = Chain.getValue(0);
1238 Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i32,
1239 Chain.getValue(2)).getValue(1);
1240 ResultVals[1] = Chain.getValue(0);
1242 NodeTys.push_back(MVT::i32);
1244 Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i32, InFlag).getValue(1);
1245 ResultVals[0] = Chain.getValue(0);
1248 NodeTys.push_back(MVT::i32);
1251 Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i64, InFlag).getValue(1);
1252 ResultVals[0] = Chain.getValue(0);
1254 NodeTys.push_back(MVT::i64);
1258 Chain = DAG.getCopyFromReg(Chain, SPU::R3, Op.Val->getValueType(0),
1259 InFlag).getValue(1);
1260 ResultVals[0] = Chain.getValue(0);
1262 NodeTys.push_back(Op.Val->getValueType(0));
1269 Chain = DAG.getCopyFromReg(Chain, SPU::R3, Op.Val->getValueType(0),
1270 InFlag).getValue(1);
1271 ResultVals[0] = Chain.getValue(0);
1273 NodeTys.push_back(Op.Val->getValueType(0));
1277 NodeTys.push_back(MVT::Other);
1279 // If the function returns void, just return the chain.
1280 if (NumResults == 0)
1283 // Otherwise, merge everything together with a MERGE_VALUES node.
1284 ResultVals[NumResults++] = Chain;
1285 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1286 ResultVals, NumResults);
1287 return Res.getValue(Op.ResNo);
1291 LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1292 SmallVector<CCValAssign, 16> RVLocs;
1293 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1294 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1295 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1296 CCInfo.AnalyzeReturn(Op.Val, RetCC_SPU);
1298 // If this is the first return lowered for this function, add the regs to the
1299 // liveout set for the function.
1300 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1301 for (unsigned i = 0; i != RVLocs.size(); ++i)
1302 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1305 SDOperand Chain = Op.getOperand(0);
1308 // Copy the result values into the output registers.
1309 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1310 CCValAssign &VA = RVLocs[i];
1311 assert(VA.isRegLoc() && "Can only return in registers!");
1312 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1313 Flag = Chain.getValue(1);
1317 return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain, Flag);
1319 return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain);
1323 //===----------------------------------------------------------------------===//
1324 // Vector related lowering:
1325 //===----------------------------------------------------------------------===//
1327 static ConstantSDNode *
1328 getVecImm(SDNode *N) {
1329 SDOperand OpVal(0, 0);
1331 // Check to see if this buildvec has a single non-undef value in its elements.
1332 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1333 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1335 OpVal = N->getOperand(i);
1336 else if (OpVal != N->getOperand(i))
1340 if (OpVal.Val != 0) {
1341 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1346 return 0; // All UNDEF: use implicit def.; not Constant node
1349 /// get_vec_i18imm - Test if this vector is a vector filled with the same value
1350 /// and the value fits into an unsigned 18-bit constant, and if so, return the
1352 SDOperand SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
1353 MVT::ValueType ValueType) {
1354 if (ConstantSDNode *CN = getVecImm(N)) {
1355 uint64_t Value = CN->getValue();
1356 if (ValueType == MVT::i64) {
1357 uint64_t UValue = CN->getValue();
1358 uint32_t upper = uint32_t(UValue >> 32);
1359 uint32_t lower = uint32_t(UValue);
1362 Value = Value >> 32;
1364 if (Value <= 0x3ffff)
1365 return DAG.getConstant(Value, ValueType);
1371 /// get_vec_i16imm - Test if this vector is a vector filled with the same value
1372 /// and the value fits into a signed 16-bit constant, and if so, return the
1374 SDOperand SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
1375 MVT::ValueType ValueType) {
1376 if (ConstantSDNode *CN = getVecImm(N)) {
1377 int64_t Value = CN->getSignExtended();
1378 if (ValueType == MVT::i64) {
1379 uint64_t UValue = CN->getValue();
1380 uint32_t upper = uint32_t(UValue >> 32);
1381 uint32_t lower = uint32_t(UValue);
1384 Value = Value >> 32;
1386 if (Value >= -(1 << 15) && Value <= ((1 << 15) - 1)) {
1387 return DAG.getConstant(Value, ValueType);
1394 /// get_vec_i10imm - Test if this vector is a vector filled with the same value
1395 /// and the value fits into a signed 10-bit constant, and if so, return the
1397 SDOperand SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
1398 MVT::ValueType ValueType) {
1399 if (ConstantSDNode *CN = getVecImm(N)) {
1400 int64_t Value = CN->getSignExtended();
1401 if (ValueType == MVT::i64) {
1402 uint64_t UValue = CN->getValue();
1403 uint32_t upper = uint32_t(UValue >> 32);
1404 uint32_t lower = uint32_t(UValue);
1407 Value = Value >> 32;
1409 if (isS10Constant(Value))
1410 return DAG.getConstant(Value, ValueType);
1416 /// get_vec_i8imm - Test if this vector is a vector filled with the same value
1417 /// and the value fits into a signed 8-bit constant, and if so, return the
1420 /// @note: The incoming vector is v16i8 because that's the only way we can load
1421 /// constant vectors. Thus, we test to see if the upper and lower bytes are the
1423 SDOperand SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
1424 MVT::ValueType ValueType) {
1425 if (ConstantSDNode *CN = getVecImm(N)) {
1426 int Value = (int) CN->getValue();
1427 if (ValueType == MVT::i16
1428 && Value <= 0xffff /* truncated from uint64_t */
1429 && ((short) Value >> 8) == ((short) Value & 0xff))
1430 return DAG.getConstant(Value & 0xff, ValueType);
1431 else if (ValueType == MVT::i8
1432 && (Value & 0xff) == Value)
1433 return DAG.getConstant(Value, ValueType);
1439 /// get_ILHUvec_imm - Test if this vector is a vector filled with the same value
1440 /// and the value fits into a signed 16-bit constant, and if so, return the
1442 SDOperand SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
1443 MVT::ValueType ValueType) {
1444 if (ConstantSDNode *CN = getVecImm(N)) {
1445 uint64_t Value = CN->getValue();
1446 if ((ValueType == MVT::i32
1447 && ((unsigned) Value & 0xffff0000) == (unsigned) Value)
1448 || (ValueType == MVT::i64 && (Value & 0xffff0000) == Value))
1449 return DAG.getConstant(Value >> 16, ValueType);
1455 /// get_v4i32_imm - Catch-all for general 32-bit constant vectors
1456 SDOperand SPU::get_v4i32_imm(SDNode *N, SelectionDAG &DAG) {
1457 if (ConstantSDNode *CN = getVecImm(N)) {
1458 return DAG.getConstant((unsigned) CN->getValue(), MVT::i32);
1464 /// get_v4i32_imm - Catch-all for general 64-bit constant vectors
1465 SDOperand SPU::get_v2i64_imm(SDNode *N, SelectionDAG &DAG) {
1466 if (ConstantSDNode *CN = getVecImm(N)) {
1467 return DAG.getConstant((unsigned) CN->getValue(), MVT::i64);
1473 // If this is a vector of constants or undefs, get the bits. A bit in
1474 // UndefBits is set if the corresponding element of the vector is an
1475 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1476 // zero. Return true if this is not an array of constants, false if it is.
1478 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
1479 uint64_t UndefBits[2]) {
1480 // Start with zero'd results.
1481 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
1483 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
1484 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
1485 SDOperand OpVal = BV->getOperand(i);
1487 unsigned PartNo = i >= e/2; // In the upper 128 bits?
1488 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
1490 uint64_t EltBits = 0;
1491 if (OpVal.getOpcode() == ISD::UNDEF) {
1492 uint64_t EltUndefBits = ~0ULL >> (64-EltBitSize);
1493 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
1495 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1496 EltBits = CN->getValue() & (~0ULL >> (64-EltBitSize));
1497 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
1498 const APFloat &apf = CN->getValueAPF();
1499 EltBits = (CN->getValueType(0) == MVT::f32
1500 ? FloatToBits(apf.convertToFloat())
1501 : DoubleToBits(apf.convertToDouble()));
1503 // Nonconstant element.
1507 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
1510 //printf("%llx %llx %llx %llx\n",
1511 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
1515 /// If this is a splat (repetition) of a value across the whole vector, return
1516 /// the smallest size that splats it. For example, "0x01010101010101..." is a
1517 /// splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
1518 /// SplatSize = 1 byte.
1519 static bool isConstantSplat(const uint64_t Bits128[2],
1520 const uint64_t Undef128[2],
1522 uint64_t &SplatBits, uint64_t &SplatUndef,
1524 // Don't let undefs prevent splats from matching. See if the top 64-bits are
1525 // the same as the lower 64-bits, ignoring undefs.
1526 uint64_t Bits64 = Bits128[0] | Bits128[1];
1527 uint64_t Undef64 = Undef128[0] & Undef128[1];
1528 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
1529 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
1530 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
1531 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
1533 if ((Bits128[0] & ~Undef128[1]) == (Bits128[1] & ~Undef128[0])) {
1534 if (MinSplatBits < 64) {
1536 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
1538 if ((Bits64 & (~Undef64 >> 32)) == ((Bits64 >> 32) & ~Undef64)) {
1539 if (MinSplatBits < 32) {
1541 // If the top 16-bits are different than the lower 16-bits, ignoring
1542 // undefs, we have an i32 splat.
1543 if ((Bits32 & (~Undef32 >> 16)) == ((Bits32 >> 16) & ~Undef32)) {
1544 if (MinSplatBits < 16) {
1545 // If the top 8-bits are different than the lower 8-bits, ignoring
1546 // undefs, we have an i16 splat.
1547 if ((Bits16 & (uint16_t(~Undef16) >> 8)) == ((Bits16 >> 8) & ~Undef16)) {
1548 // Otherwise, we have an 8-bit splat.
1549 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
1550 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
1556 SplatUndef = Undef16;
1563 SplatUndef = Undef32;
1569 SplatBits = Bits128[0];
1570 SplatUndef = Undef128[0];
1576 return false; // Can't be a splat if two pieces don't match.
1579 // If this is a case we can't handle, return null and let the default
1580 // expansion code take care of it. If we CAN select this case, and if it
1581 // selects to a single instruction, return Op. Otherwise, if we can codegen
1582 // this case more efficiently than a constant pool load, lower it to the
1583 // sequence of ops that should be used.
1584 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1585 MVT::ValueType VT = Op.getValueType();
1586 // If this is a vector of constants or undefs, get the bits. A bit in
1587 // UndefBits is set if the corresponding element of the vector is an
1588 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1590 uint64_t VectorBits[2];
1591 uint64_t UndefBits[2];
1592 uint64_t SplatBits, SplatUndef;
1594 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits)
1595 || !isConstantSplat(VectorBits, UndefBits,
1596 MVT::getSizeInBits(MVT::getVectorElementType(VT)),
1597 SplatBits, SplatUndef, SplatSize))
1598 return SDOperand(); // Not a constant vector, not a splat.
1603 uint32_t Value32 = SplatBits;
1604 assert(SplatSize == 4
1605 && "LowerBUILD_VECTOR: Unexpected floating point vector element.");
1606 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1607 SDOperand T = DAG.getConstant(Value32, MVT::i32);
1608 return DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32,
1609 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, T, T, T, T));
1613 uint64_t f64val = SplatBits;
1614 assert(SplatSize == 8
1615 && "LowerBUILD_VECTOR: 64-bit float vector element: unexpected size.");
1616 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1617 SDOperand T = DAG.getConstant(f64val, MVT::i64);
1618 return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64,
1619 DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
1623 // 8-bit constants have to be expanded to 16-bits
1624 unsigned short Value16 = SplatBits | (SplatBits << 8);
1626 for (int i = 0; i < 8; ++i)
1627 Ops[i] = DAG.getConstant(Value16, MVT::i16);
1628 return DAG.getNode(ISD::BIT_CONVERT, VT,
1629 DAG.getNode(ISD::BUILD_VECTOR, MVT::v8i16, Ops, 8));
1632 unsigned short Value16;
1634 Value16 = (unsigned short) (SplatBits & 0xffff);
1636 Value16 = (unsigned short) (SplatBits | (SplatBits << 8));
1637 SDOperand T = DAG.getConstant(Value16, MVT::getVectorElementType(VT));
1639 for (int i = 0; i < 8; ++i) Ops[i] = T;
1640 return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops, 8);
1643 unsigned int Value = SplatBits;
1644 SDOperand T = DAG.getConstant(Value, MVT::getVectorElementType(VT));
1645 return DAG.getNode(ISD::BUILD_VECTOR, VT, T, T, T, T);
1648 uint64_t val = SplatBits;
1649 uint32_t upper = uint32_t(val >> 32);
1650 uint32_t lower = uint32_t(val);
1652 if (upper == lower) {
1653 // Magic constant that can be matched by IL, ILA, et. al.
1654 SDOperand Val = DAG.getTargetConstant(val, MVT::i64);
1655 return DAG.getNode(ISD::BUILD_VECTOR, VT, Val, Val);
1659 SmallVector<SDOperand, 16> ShufBytes;
1661 bool upper_special, lower_special;
1663 // NOTE: This code creates common-case shuffle masks that can be easily
1664 // detected as common expressions. It is not attempting to create highly
1665 // specialized masks to replace any and all 0's, 0xff's and 0x80's.
1667 // Detect if the upper or lower half is a special shuffle mask pattern:
1668 upper_special = (upper == 0 || upper == 0xffffffff || upper == 0x80000000);
1669 lower_special = (lower == 0 || lower == 0xffffffff || lower == 0x80000000);
1671 // Create lower vector if not a special pattern
1672 if (!lower_special) {
1673 SDOperand LO32C = DAG.getConstant(lower, MVT::i32);
1674 LO32 = DAG.getNode(ISD::BIT_CONVERT, VT,
1675 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1676 LO32C, LO32C, LO32C, LO32C));
1679 // Create upper vector if not a special pattern
1680 if (!upper_special) {
1681 SDOperand HI32C = DAG.getConstant(upper, MVT::i32);
1682 HI32 = DAG.getNode(ISD::BIT_CONVERT, VT,
1683 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1684 HI32C, HI32C, HI32C, HI32C));
1687 // If either upper or lower are special, then the two input operands are
1688 // the same (basically, one of them is a "don't care")
1693 if (lower_special && upper_special) {
1694 // Unhappy situation... both upper and lower are special, so punt with
1695 // a target constant:
1696 SDOperand Zero = DAG.getConstant(0, MVT::i32);
1697 HI32 = LO32 = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Zero, Zero,
1701 for (int i = 0; i < 4; ++i) {
1702 for (int j = 0; j < 4; ++j) {
1704 bool process_upper, process_lower;
1707 process_upper = (upper_special && (i & 1) == 0);
1708 process_lower = (lower_special && (i & 1) == 1);
1710 if (process_upper || process_lower) {
1711 if ((process_upper && upper == 0)
1712 || (process_lower && lower == 0))
1714 else if ((process_upper && upper == 0xffffffff)
1715 || (process_lower && lower == 0xffffffff))
1717 else if ((process_upper && upper == 0x80000000)
1718 || (process_lower && lower == 0x80000000))
1719 val = (j == 0 ? 0xe0 : 0x80);
1721 val = i * 4 + j + ((i & 1) * 16);
1723 ShufBytes.push_back(DAG.getConstant(val, MVT::i8));
1727 return DAG.getNode(SPUISD::SHUFB, VT, HI32, LO32,
1728 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
1729 &ShufBytes[0], ShufBytes.size()));
1737 /// LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3) to something on
1738 /// which the Cell can operate. The code inspects V3 to ascertain whether the
1739 /// permutation vector, V3, is monotonically increasing with one "exception"
1740 /// element, e.g., (0, 1, _, 3). If this is the case, then generate a
1741 /// INSERT_MASK synthetic instruction. Otherwise, spill V3 to the constant pool.
1742 /// In either case, the net result is going to eventually invoke SHUFB to
1743 /// permute/shuffle the bytes from V1 and V2.
1745 /// INSERT_MASK is eventually selected as one of the C*D instructions, generate
1746 /// control word for byte/halfword/word insertion. This takes care of a single
1747 /// element move from V2 into V1.
1749 /// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
1750 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
1751 SDOperand V1 = Op.getOperand(0);
1752 SDOperand V2 = Op.getOperand(1);
1753 SDOperand PermMask = Op.getOperand(2);
1755 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1757 // If we have a single element being moved from V1 to V2, this can be handled
1758 // using the C*[DX] compute mask instructions, but the vector elements have
1759 // to be monotonically increasing with one exception element.
1760 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
1761 unsigned EltsFromV2 = 0;
1763 unsigned V2EltIdx0 = 0;
1764 unsigned CurrElt = 0;
1765 bool monotonic = true;
1766 if (EltVT == MVT::i8)
1768 else if (EltVT == MVT::i16)
1770 else if (EltVT == MVT::i32)
1773 assert(0 && "Unhandled vector type in LowerVECTOR_SHUFFLE");
1775 for (unsigned i = 0, e = PermMask.getNumOperands();
1776 EltsFromV2 <= 1 && monotonic && i != e;
1779 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1782 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1784 if (SrcElt >= V2EltIdx0) {
1786 V2Elt = (V2EltIdx0 - SrcElt) << 2;
1787 } else if (CurrElt != SrcElt) {
1794 if (EltsFromV2 == 1 && monotonic) {
1795 // Compute mask and shuffle
1796 MachineFunction &MF = DAG.getMachineFunction();
1797 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1798 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
1799 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1800 // Initialize temporary register to 0
1801 SDOperand InitTempReg =
1802 DAG.getCopyToReg(DAG.getEntryNode(), VReg, DAG.getConstant(0, PtrVT));
1803 // Copy register's contents as index in INSERT_MASK:
1804 SDOperand ShufMaskOp =
1805 DAG.getNode(SPUISD::INSERT_MASK, V1.getValueType(),
1806 DAG.getTargetConstant(V2Elt, MVT::i32),
1807 DAG.getCopyFromReg(InitTempReg, VReg, PtrVT));
1808 // Use shuffle mask in SHUFB synthetic instruction:
1809 return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V2, V1, ShufMaskOp);
1811 // Convert the SHUFFLE_VECTOR mask's input element units to the actual bytes.
1812 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
1814 SmallVector<SDOperand, 16> ResultMask;
1815 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
1817 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1820 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1822 for (unsigned j = 0; j < BytesPerElement; ++j) {
1823 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
1828 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
1829 &ResultMask[0], ResultMask.size());
1830 return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V1, V2, VPermMask);
1834 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1835 SDOperand Op0 = Op.getOperand(0); // Op0 = the scalar
1837 if (Op0.Val->getOpcode() == ISD::Constant) {
1838 // For a constant, build the appropriate constant vector, which will
1839 // eventually simplify to a vector register load.
1841 ConstantSDNode *CN = cast<ConstantSDNode>(Op0.Val);
1842 SmallVector<SDOperand, 16> ConstVecValues;
1846 // Create a constant vector:
1847 switch (Op.getValueType()) {
1848 default: assert(0 && "Unexpected constant value type in "
1849 "LowerSCALAR_TO_VECTOR");
1850 case MVT::v16i8: n_copies = 16; VT = MVT::i8; break;
1851 case MVT::v8i16: n_copies = 8; VT = MVT::i16; break;
1852 case MVT::v4i32: n_copies = 4; VT = MVT::i32; break;
1853 case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
1854 case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
1855 case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
1858 SDOperand CValue = DAG.getConstant(CN->getValue(), VT);
1859 for (size_t j = 0; j < n_copies; ++j)
1860 ConstVecValues.push_back(CValue);
1862 return DAG.getNode(ISD::BUILD_VECTOR, Op.getValueType(),
1863 &ConstVecValues[0], ConstVecValues.size());
1865 // Otherwise, copy the value from one register to another:
1866 switch (Op0.getValueType()) {
1867 default: assert(0 && "Unexpected value type in LowerSCALAR_TO_VECTOR");
1874 return DAG.getNode(SPUISD::PROMOTE_SCALAR, Op.getValueType(), Op0, Op0);
1881 static SDOperand LowerVectorMUL(SDOperand Op, SelectionDAG &DAG) {
1882 switch (Op.getValueType()) {
1884 SDOperand rA = Op.getOperand(0);
1885 SDOperand rB = Op.getOperand(1);
1886 SDOperand HiProd1 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rA, rB);
1887 SDOperand HiProd2 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rB, rA);
1888 SDOperand LoProd = DAG.getNode(SPUISD::MPYU, MVT::v4i32, rA, rB);
1889 SDOperand Residual1 = DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd1);
1891 return DAG.getNode(ISD::ADD, MVT::v4i32, Residual1, HiProd2);
1895 // Multiply two v8i16 vectors (pipeline friendly version):
1896 // a) multiply lower halves, mask off upper 16-bit of 32-bit product
1897 // b) multiply upper halves, rotate left by 16 bits (inserts 16 lower zeroes)
1898 // c) Use SELB to select upper and lower halves from the intermediate results
1900 // NOTE: We really want to move the FSMBI to earlier to actually get the
1901 // dual-issue. This code does manage to do this, even if it's a little on
1904 MachineFunction &MF = DAG.getMachineFunction();
1905 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1906 SDOperand Chain = Op.getOperand(0);
1907 SDOperand rA = Op.getOperand(0);
1908 SDOperand rB = Op.getOperand(1);
1909 unsigned FSMBIreg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1910 unsigned HiProdReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
1913 DAG.getCopyToReg(Chain, FSMBIreg,
1914 DAG.getNode(SPUISD::FSMBI, MVT::v8i16,
1915 DAG.getConstant(0xcccc, MVT::i32)));
1918 DAG.getCopyToReg(FSMBOp, HiProdReg,
1919 DAG.getNode(SPUISD::MPYHH, MVT::v8i16, rA, rB));
1921 SDOperand HHProd_v4i32 =
1922 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
1923 DAG.getCopyFromReg(HHProd, HiProdReg, MVT::v4i32));
1925 return DAG.getNode(SPUISD::SELB, MVT::v8i16,
1926 DAG.getNode(SPUISD::MPY, MVT::v8i16, rA, rB),
1927 DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
1928 DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32,
1930 DAG.getConstant(16, MVT::i16))),
1931 DAG.getCopyFromReg(FSMBOp, FSMBIreg, MVT::v4i32));
1934 // This M00sE is N@stI! (apologies to Monty Python)
1936 // SPU doesn't know how to do any 8-bit multiplication, so the solution
1937 // is to break it all apart, sign extend, and reassemble the various
1938 // intermediate products.
1940 SDOperand rA = Op.getOperand(0);
1941 SDOperand rB = Op.getOperand(1);
1942 SDOperand c8 = DAG.getConstant(8, MVT::i32);
1943 SDOperand c16 = DAG.getConstant(16, MVT::i32);
1946 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1947 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rA),
1948 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rB));
1950 SDOperand rALH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rA, c8);
1952 SDOperand rBLH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rB, c8);
1955 DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16,
1956 DAG.getNode(SPUISD::MPY, MVT::v8i16, rALH, rBLH), c8);
1958 SDOperand FSMBmask = DAG.getNode(SPUISD::FSMBI, MVT::v8i16,
1959 DAG.getConstant(0x2222, MVT::i32));
1961 SDOperand LoProdParts =
1962 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
1963 DAG.getNode(SPUISD::SELB, MVT::v8i16,
1964 LLProd, LHProd, FSMBmask));
1966 SDOperand LoProdMask = DAG.getConstant(0xffff, MVT::i32);
1969 DAG.getNode(ISD::AND, MVT::v4i32,
1971 DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1972 LoProdMask, LoProdMask,
1973 LoProdMask, LoProdMask));
1976 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
1977 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rA), c16);
1980 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
1981 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rB), c16);
1984 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1985 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rAH),
1986 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rBH));
1988 SDOperand HHProd_1 =
1989 DAG.getNode(SPUISD::MPY, MVT::v8i16,
1990 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
1991 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rAH, c8)),
1992 DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
1993 DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rBH, c8)));
1996 DAG.getNode(SPUISD::SELB, MVT::v8i16,
1998 DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16, HHProd_1, c8),
2002 DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32, HHProd, c16);
2004 return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8,
2005 DAG.getNode(ISD::OR, MVT::v4i32,
2010 cerr << "CellSPU: Unknown vector multiplication, got "
2011 << MVT::getValueTypeString(Op.getValueType())
2020 static SDOperand LowerFDIVf32(SDOperand Op, SelectionDAG &DAG) {
2021 MachineFunction &MF = DAG.getMachineFunction();
2022 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2024 SDOperand A = Op.getOperand(0);
2025 SDOperand B = Op.getOperand(1);
2026 unsigned VT = Op.getValueType();
2028 unsigned VRegBR, VRegC;
2030 if (VT == MVT::f32) {
2031 VRegBR = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
2032 VRegC = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
2034 VRegBR = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
2035 VRegC = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
2037 // TODO: make sure we're feeding FPInterp the right arguments
2038 // Right now: fi B, frest(B)
2041 // (Floating Interpolate (FP Reciprocal Estimate B))
2043 DAG.getCopyToReg(DAG.getEntryNode(), VRegBR,
2044 DAG.getNode(SPUISD::FPInterp, VT, B,
2045 DAG.getNode(SPUISD::FPRecipEst, VT, B)));
2047 // Computes A * BRcpl and stores in a temporary register
2049 DAG.getCopyToReg(BRcpl, VRegC,
2050 DAG.getNode(ISD::FMUL, VT, A,
2051 DAG.getCopyFromReg(BRcpl, VRegBR, VT)));
2052 // What's the Chain variable do? It's magic!
2053 // TODO: set Chain = Op(0).getEntryNode()
2055 return DAG.getNode(ISD::FADD, VT,
2056 DAG.getCopyFromReg(AxBRcpl, VRegC, VT),
2057 DAG.getNode(ISD::FMUL, VT,
2058 DAG.getCopyFromReg(AxBRcpl, VRegBR, VT),
2059 DAG.getNode(ISD::FSUB, VT, A,
2060 DAG.getNode(ISD::FMUL, VT, B,
2061 DAG.getCopyFromReg(AxBRcpl, VRegC, VT)))));
2064 static SDOperand LowerEXTRACT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2065 unsigned VT = Op.getValueType();
2066 SDOperand N = Op.getOperand(0);
2067 SDOperand Elt = Op.getOperand(1);
2068 SDOperand ShufMask[16];
2069 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt);
2071 assert(C != 0 && "LowerEXTRACT_VECTOR_ELT expecting constant SDNode");
2073 int EltNo = (int) C->getValue();
2076 if (VT == MVT::i8 && EltNo >= 16)
2077 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i8 extraction slot > 15");
2078 else if (VT == MVT::i16 && EltNo >= 8)
2079 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i16 extraction slot > 7");
2080 else if (VT == MVT::i32 && EltNo >= 4)
2081 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i32 extraction slot > 4");
2082 else if (VT == MVT::i64 && EltNo >= 2)
2083 assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i64 extraction slot > 2");
2085 if (EltNo == 0 && (VT == MVT::i32 || VT == MVT::i64)) {
2086 // i32 and i64: Element 0 is the preferred slot
2087 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT, N);
2090 // Need to generate shuffle mask and extract:
2091 int prefslot_begin = -1, prefslot_end = -1;
2092 int elt_byte = EltNo * MVT::getSizeInBits(VT) / 8;
2096 prefslot_begin = prefslot_end = 3;
2100 prefslot_begin = 2; prefslot_end = 3;
2104 prefslot_begin = 0; prefslot_end = 3;
2108 prefslot_begin = 0; prefslot_end = 7;
2113 assert(prefslot_begin != -1 && prefslot_end != -1 &&
2114 "LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
2116 for (int i = 0; i < 16; ++i) {
2117 // zero fill uppper part of preferred slot, don't care about the
2119 unsigned int mask_val;
2121 if (i <= prefslot_end) {
2123 ((i < prefslot_begin)
2125 : elt_byte + (i - prefslot_begin));
2127 ShufMask[i] = DAG.getConstant(mask_val, MVT::i8);
2129 ShufMask[i] = ShufMask[i % (prefslot_end + 1)];
2132 SDOperand ShufMaskVec =
2133 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2135 sizeof(ShufMask) / sizeof(ShufMask[0]));
2137 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
2138 DAG.getNode(SPUISD::SHUFB, N.getValueType(),
2139 N, N, ShufMaskVec));
2143 static SDOperand LowerINSERT_VECTOR_ELT(SDOperand Op, SelectionDAG &DAG) {
2144 SDOperand VecOp = Op.getOperand(0);
2145 SDOperand ValOp = Op.getOperand(1);
2146 SDOperand IdxOp = Op.getOperand(2);
2147 MVT::ValueType VT = Op.getValueType();
2149 ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
2150 assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
2152 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2153 // Use $2 because it's always 16-byte aligned and it's available:
2154 SDOperand PtrBase = DAG.getRegister(SPU::R2, PtrVT);
2157 DAG.getNode(SPUISD::SHUFB, VT,
2158 DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, ValOp),
2160 DAG.getNode(SPUISD::INSERT_MASK, VT,
2161 DAG.getNode(ISD::ADD, PtrVT,
2163 DAG.getConstant(CN->getValue(),
2169 static SDOperand LowerI8Math(SDOperand Op, SelectionDAG &DAG, unsigned Opc)
2171 SDOperand N0 = Op.getOperand(0); // Everything has at least one operand
2173 assert(Op.getValueType() == MVT::i8);
2176 assert(0 && "Unhandled i8 math operator");
2180 // 8-bit subtraction: Promote the arguments up to 16-bits and truncate
2182 SDOperand N1 = Op.getOperand(1);
2183 N0 = (N0.getOpcode() != ISD::Constant
2184 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2185 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2186 N1 = (N1.getOpcode() != ISD::Constant
2187 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N1)
2188 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2189 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2190 DAG.getNode(Opc, MVT::i16, N0, N1));
2194 SDOperand N1 = Op.getOperand(1);
2196 N0 = (N0.getOpcode() != ISD::Constant
2197 ? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
2198 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2199 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::ZERO_EXTEND : ISD::TRUNCATE);
2200 N1 = (N1.getOpcode() != ISD::Constant
2201 ? DAG.getNode(N1Opc, MVT::i16, N1)
2202 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2203 SDOperand ExpandArg =
2204 DAG.getNode(ISD::OR, MVT::i16, N0,
2205 DAG.getNode(ISD::SHL, MVT::i16,
2206 N0, DAG.getConstant(8, MVT::i16)));
2207 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2208 DAG.getNode(Opc, MVT::i16, ExpandArg, N1));
2212 SDOperand N1 = Op.getOperand(1);
2214 N0 = (N0.getOpcode() != ISD::Constant
2215 ? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
2216 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2217 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::ZERO_EXTEND : ISD::TRUNCATE);
2218 N1 = (N1.getOpcode() != ISD::Constant
2219 ? DAG.getNode(N1Opc, MVT::i16, N1)
2220 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2221 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2222 DAG.getNode(Opc, MVT::i16, N0, N1));
2225 SDOperand N1 = Op.getOperand(1);
2227 N0 = (N0.getOpcode() != ISD::Constant
2228 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2229 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2230 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::SIGN_EXTEND : ISD::TRUNCATE);
2231 N1 = (N1.getOpcode() != ISD::Constant
2232 ? DAG.getNode(N1Opc, MVT::i16, N1)
2233 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2234 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2235 DAG.getNode(Opc, MVT::i16, N0, N1));
2238 SDOperand N1 = Op.getOperand(1);
2240 N0 = (N0.getOpcode() != ISD::Constant
2241 ? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
2242 : DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
2243 N1Opc = (N1.getValueType() < MVT::i16 ? ISD::SIGN_EXTEND : ISD::TRUNCATE);
2244 N1 = (N1.getOpcode() != ISD::Constant
2245 ? DAG.getNode(N1Opc, MVT::i16, N1)
2246 : DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
2247 return DAG.getNode(ISD::TRUNCATE, MVT::i8,
2248 DAG.getNode(Opc, MVT::i16, N0, N1));
2256 static SDOperand LowerI64Math(SDOperand Op, SelectionDAG &DAG, unsigned Opc)
2258 MVT::ValueType VT = Op.getValueType();
2260 MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
2262 SDOperand Op0 = Op.getOperand(0);
2265 case ISD::ZERO_EXTEND:
2266 case ISD::SIGN_EXTEND:
2267 case ISD::ANY_EXTEND: {
2268 MVT::ValueType Op0VT = Op0.getValueType();
2270 MVT::getVectorType(Op0VT, (128 / MVT::getSizeInBits(Op0VT)));
2272 assert(Op0VT == MVT::i32
2273 && "CellSPU: Zero/sign extending something other than i32");
2275 unsigned NewOpc = (Opc == ISD::SIGN_EXTEND
2276 ? SPUISD::ROTBYTES_RIGHT_S
2277 : SPUISD::ROTQUAD_RZ_BYTES);
2278 SDOperand PromoteScalar =
2279 DAG.getNode(SPUISD::PROMOTE_SCALAR, Op0VecVT, Op0);
2281 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
2282 DAG.getNode(ISD::BIT_CONVERT, VecVT,
2283 DAG.getNode(NewOpc, Op0VecVT,
2285 DAG.getConstant(4, MVT::i32))));
2289 SDOperand ShiftAmt = Op.getOperand(1);
2290 unsigned ShiftAmtVT = unsigned(ShiftAmt.getValueType());
2291 SDOperand Op0Vec = DAG.getNode(SPUISD::PROMOTE_SCALAR, VecVT, Op0);
2292 SDOperand MaskLower =
2293 DAG.getNode(SPUISD::SELB, VecVT,
2295 DAG.getConstant(0, VecVT),
2296 DAG.getNode(SPUISD::FSMBI, VecVT,
2297 DAG.getConstant(0xff00ULL, MVT::i16)));
2298 SDOperand ShiftAmtBytes =
2299 DAG.getNode(ISD::SRL, ShiftAmtVT,
2301 DAG.getConstant(3, ShiftAmtVT));
2302 SDOperand ShiftAmtBits =
2303 DAG.getNode(ISD::AND, ShiftAmtVT,
2305 DAG.getConstant(7, ShiftAmtVT));
2307 return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
2308 DAG.getNode(SPUISD::SHLQUAD_L_BITS, VecVT,
2309 DAG.getNode(SPUISD::SHLQUAD_L_BYTES, VecVT,
2310 MaskLower, ShiftAmtBytes),
2315 unsigned VT = unsigned(Op.getValueType());
2316 SDOperand ShiftAmt = Op.getOperand(1);
2317 unsigned ShiftAmtVT = unsigned(ShiftAmt.getValueType());
2318 SDOperand ShiftAmtBytes =
2319 DAG.getNode(ISD::SRL, ShiftAmtVT,
2321 DAG.getConstant(3, ShiftAmtVT));
2322 SDOperand ShiftAmtBits =
2323 DAG.getNode(ISD::AND, ShiftAmtVT,
2325 DAG.getConstant(7, ShiftAmtVT));
2327 return DAG.getNode(SPUISD::ROTQUAD_RZ_BITS, VT,
2328 DAG.getNode(SPUISD::ROTQUAD_RZ_BYTES, VT,
2329 Op0, ShiftAmtBytes),
2337 //! Lower byte immediate operations for v16i8 vectors:
2339 LowerByteImmed(SDOperand Op, SelectionDAG &DAG) {
2342 MVT::ValueType VT = Op.getValueType();
2344 ConstVec = Op.getOperand(0);
2345 Arg = Op.getOperand(1);
2346 if (ConstVec.Val->getOpcode() != ISD::BUILD_VECTOR) {
2347 if (ConstVec.Val->getOpcode() == ISD::BIT_CONVERT) {
2348 ConstVec = ConstVec.getOperand(0);
2350 ConstVec = Op.getOperand(1);
2351 Arg = Op.getOperand(0);
2352 if (ConstVec.Val->getOpcode() == ISD::BIT_CONVERT) {
2353 ConstVec = ConstVec.getOperand(0);
2358 if (ConstVec.Val->getOpcode() == ISD::BUILD_VECTOR) {
2359 uint64_t VectorBits[2];
2360 uint64_t UndefBits[2];
2361 uint64_t SplatBits, SplatUndef;
2364 if (!GetConstantBuildVectorBits(ConstVec.Val, VectorBits, UndefBits)
2365 && isConstantSplat(VectorBits, UndefBits,
2366 MVT::getSizeInBits(MVT::getVectorElementType(VT)),
2367 SplatBits, SplatUndef, SplatSize)) {
2368 SDOperand tcVec[16];
2369 SDOperand tc = DAG.getTargetConstant(SplatBits & 0xff, MVT::i8);
2370 const size_t tcVecSize = sizeof(tcVec) / sizeof(tcVec[0]);
2372 // Turn the BUILD_VECTOR into a set of target constants:
2373 for (size_t i = 0; i < tcVecSize; ++i)
2376 return DAG.getNode(Op.Val->getOpcode(), VT, Arg,
2377 DAG.getNode(ISD::BUILD_VECTOR, VT, tcVec, tcVecSize));
2384 //! Lower i32 multiplication
2385 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG, unsigned VT,
2389 cerr << "CellSPU: Unknown LowerMUL value type, got "
2390 << MVT::getValueTypeString(Op.getValueType())
2396 SDOperand rA = Op.getOperand(0);
2397 SDOperand rB = Op.getOperand(1);
2399 return DAG.getNode(ISD::ADD, MVT::i32,
2400 DAG.getNode(ISD::ADD, MVT::i32,
2401 DAG.getNode(SPUISD::MPYH, MVT::i32, rA, rB),
2402 DAG.getNode(SPUISD::MPYH, MVT::i32, rB, rA)),
2403 DAG.getNode(SPUISD::MPYU, MVT::i32, rA, rB));
2410 //! Custom lowering for CTPOP (count population)
2412 Custom lowering code that counts the number ones in the input
2413 operand. SPU has such an instruction, but it counts the number of
2414 ones per byte, which then have to be accumulated.
2416 static SDOperand LowerCTPOP(SDOperand Op, SelectionDAG &DAG) {
2417 unsigned VT = Op.getValueType();
2418 unsigned vecVT = MVT::getVectorType(VT, (128 / MVT::getSizeInBits(VT)));
2422 SDOperand N = Op.getOperand(0);
2423 SDOperand Elt0 = DAG.getConstant(0, MVT::i32);
2425 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2426 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2428 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i8, CNTB, Elt0);
2432 MachineFunction &MF = DAG.getMachineFunction();
2433 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2435 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
2437 SDOperand N = Op.getOperand(0);
2438 SDOperand Elt0 = DAG.getConstant(0, MVT::i16);
2439 SDOperand Mask0 = DAG.getConstant(0x0f, MVT::i16);
2440 SDOperand Shift1 = DAG.getConstant(8, MVT::i16);
2442 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2443 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2445 // CNTB_result becomes the chain to which all of the virtual registers
2446 // CNTB_reg, SUM1_reg become associated:
2447 SDOperand CNTB_result =
2448 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, CNTB, Elt0);
2450 SDOperand CNTB_rescopy =
2451 DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
2453 SDOperand Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i16);
2455 return DAG.getNode(ISD::AND, MVT::i16,
2456 DAG.getNode(ISD::ADD, MVT::i16,
2457 DAG.getNode(ISD::SRL, MVT::i16,
2464 MachineFunction &MF = DAG.getMachineFunction();
2465 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2467 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2468 unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2470 SDOperand N = Op.getOperand(0);
2471 SDOperand Elt0 = DAG.getConstant(0, MVT::i32);
2472 SDOperand Mask0 = DAG.getConstant(0xff, MVT::i32);
2473 SDOperand Shift1 = DAG.getConstant(16, MVT::i32);
2474 SDOperand Shift2 = DAG.getConstant(8, MVT::i32);
2476 SDOperand Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
2477 SDOperand CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
2479 // CNTB_result becomes the chain to which all of the virtual registers
2480 // CNTB_reg, SUM1_reg become associated:
2481 SDOperand CNTB_result =
2482 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, CNTB, Elt0);
2484 SDOperand CNTB_rescopy =
2485 DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
2488 DAG.getNode(ISD::SRL, MVT::i32,
2489 DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32), Shift1);
2492 DAG.getNode(ISD::ADD, MVT::i32,
2493 Comp1, DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32));
2495 SDOperand Sum1_rescopy =
2496 DAG.getCopyToReg(CNTB_result, SUM1_reg, Sum1);
2499 DAG.getNode(ISD::SRL, MVT::i32,
2500 DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32),
2503 DAG.getNode(ISD::ADD, MVT::i32, Comp2,
2504 DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32));
2506 return DAG.getNode(ISD::AND, MVT::i32, Sum2, Mask0);
2516 /// LowerOperation - Provide custom lowering hooks for some operations.
2519 SPUTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG)
2521 unsigned Opc = (unsigned) Op.getOpcode();
2522 unsigned VT = (unsigned) Op.getValueType();
2526 cerr << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
2527 cerr << "Op.getOpcode() = " << Opc << "\n";
2528 cerr << "*Op.Val:\n";
2535 return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
2537 return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
2538 case ISD::ConstantPool:
2539 return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
2540 case ISD::GlobalAddress:
2541 return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
2542 case ISD::JumpTable:
2543 return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
2545 return LowerConstant(Op, DAG);
2546 case ISD::ConstantFP:
2547 return LowerConstantFP(Op, DAG);
2549 return LowerBRCOND(Op, DAG);
2550 case ISD::FORMAL_ARGUMENTS:
2551 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex);
2553 return LowerCALL(Op, DAG, SPUTM.getSubtargetImpl());
2555 return LowerRET(Op, DAG, getTargetMachine());
2558 // i8, i64 math ops:
2559 case ISD::ZERO_EXTEND:
2560 case ISD::SIGN_EXTEND:
2561 case ISD::ANY_EXTEND:
2569 return LowerI8Math(Op, DAG, Opc);
2570 else if (VT == MVT::i64)
2571 return LowerI64Math(Op, DAG, Opc);
2574 // Vector-related lowering.
2575 case ISD::BUILD_VECTOR:
2576 return LowerBUILD_VECTOR(Op, DAG);
2577 case ISD::SCALAR_TO_VECTOR:
2578 return LowerSCALAR_TO_VECTOR(Op, DAG);
2579 case ISD::VECTOR_SHUFFLE:
2580 return LowerVECTOR_SHUFFLE(Op, DAG);
2581 case ISD::EXTRACT_VECTOR_ELT:
2582 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2583 case ISD::INSERT_VECTOR_ELT:
2584 return LowerINSERT_VECTOR_ELT(Op, DAG);
2586 // Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
2590 return LowerByteImmed(Op, DAG);
2592 // Vector and i8 multiply:
2594 if (MVT::isVector(VT))
2595 return LowerVectorMUL(Op, DAG);
2596 else if (VT == MVT::i8)
2597 return LowerI8Math(Op, DAG, Opc);
2599 return LowerMUL(Op, DAG, VT, Opc);
2602 if (VT == MVT::f32 || VT == MVT::v4f32)
2603 return LowerFDIVf32(Op, DAG);
2604 // else if (Op.getValueType() == MVT::f64)
2605 // return LowerFDIVf64(Op, DAG);
2607 assert(0 && "Calling FDIV on unsupported MVT");
2610 return LowerCTPOP(Op, DAG);
2616 //===----------------------------------------------------------------------===//
2617 // Target Optimization Hooks
2618 //===----------------------------------------------------------------------===//
2621 SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
2624 TargetMachine &TM = getTargetMachine();
2626 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
2627 SelectionDAG &DAG = DCI.DAG;
2628 SDOperand Op0 = N->getOperand(0); // everything has at least one operand
2629 SDOperand Result; // Initially, NULL result
2631 switch (N->getOpcode()) {
2634 SDOperand Op1 = N->getOperand(1);
2636 if ((Op1.getOpcode() == ISD::Constant
2637 || Op1.getOpcode() == ISD::TargetConstant)
2638 && Op0.getOpcode() == SPUISD::IndirectAddr) {
2639 SDOperand Op01 = Op0.getOperand(1);
2640 if (Op01.getOpcode() == ISD::Constant
2641 || Op01.getOpcode() == ISD::TargetConstant) {
2642 // (add <const>, (SPUindirect <arg>, <const>)) ->
2643 // (SPUindirect <arg>, <const + const>)
2644 ConstantSDNode *CN0 = cast<ConstantSDNode>(Op1);
2645 ConstantSDNode *CN1 = cast<ConstantSDNode>(Op01);
2646 SDOperand combinedConst =
2647 DAG.getConstant(CN0->getValue() + CN1->getValue(),
2648 Op0.getValueType());
2650 DEBUG(cerr << "Replace: (add " << CN0->getValue() << ", "
2651 << "(SPUindirect <arg>, " << CN1->getValue() << "))\n");
2652 DEBUG(cerr << "With: (SPUindirect <arg>, "
2653 << CN0->getValue() + CN1->getValue() << ")\n");
2654 return DAG.getNode(SPUISD::IndirectAddr, Op0.getValueType(),
2655 Op0.getOperand(0), combinedConst);
2657 } else if ((Op0.getOpcode() == ISD::Constant
2658 || Op0.getOpcode() == ISD::TargetConstant)
2659 && Op1.getOpcode() == SPUISD::IndirectAddr) {
2660 SDOperand Op11 = Op1.getOperand(1);
2661 if (Op11.getOpcode() == ISD::Constant
2662 || Op11.getOpcode() == ISD::TargetConstant) {
2663 // (add (SPUindirect <arg>, <const>), <const>) ->
2664 // (SPUindirect <arg>, <const + const>)
2665 ConstantSDNode *CN0 = cast<ConstantSDNode>(Op0);
2666 ConstantSDNode *CN1 = cast<ConstantSDNode>(Op11);
2667 SDOperand combinedConst =
2668 DAG.getConstant(CN0->getValue() + CN1->getValue(),
2669 Op0.getValueType());
2671 DEBUG(cerr << "Replace: (add " << CN0->getValue() << ", "
2672 << "(SPUindirect <arg>, " << CN1->getValue() << "))\n");
2673 DEBUG(cerr << "With: (SPUindirect <arg>, "
2674 << CN0->getValue() + CN1->getValue() << ")\n");
2676 return DAG.getNode(SPUISD::IndirectAddr, Op1.getValueType(),
2677 Op1.getOperand(0), combinedConst);
2682 case ISD::SIGN_EXTEND:
2683 case ISD::ZERO_EXTEND:
2684 case ISD::ANY_EXTEND: {
2685 if (Op0.getOpcode() == SPUISD::EXTRACT_ELT0 &&
2686 N->getValueType(0) == Op0.getValueType()) {
2687 // (any_extend (SPUextract_elt0 <arg>)) ->
2688 // (SPUextract_elt0 <arg>)
2689 // Types must match, however...
2690 DEBUG(cerr << "Replace: ");
2691 DEBUG(N->dump(&DAG));
2692 DEBUG(cerr << "\nWith: ");
2693 DEBUG(Op0.Val->dump(&DAG));
2694 DEBUG(cerr << "\n");
2700 case SPUISD::IndirectAddr: {
2701 if (!ST->usingLargeMem() && Op0.getOpcode() == SPUISD::AFormAddr) {
2702 ConstantSDNode *CN = cast<ConstantSDNode>(N->getOperand(1));
2703 if (CN->getValue() == 0) {
2704 // (SPUindirect (SPUaform <addr>, 0), 0) ->
2705 // (SPUaform <addr>, 0)
2707 DEBUG(cerr << "Replace: ");
2708 DEBUG(N->dump(&DAG));
2709 DEBUG(cerr << "\nWith: ");
2710 DEBUG(Op0.Val->dump(&DAG));
2711 DEBUG(cerr << "\n");
2718 case SPUISD::SHLQUAD_L_BITS:
2719 case SPUISD::SHLQUAD_L_BYTES:
2720 case SPUISD::VEC_SHL:
2721 case SPUISD::VEC_SRL:
2722 case SPUISD::VEC_SRA:
2723 case SPUISD::ROTQUAD_RZ_BYTES:
2724 case SPUISD::ROTQUAD_RZ_BITS: {
2725 SDOperand Op1 = N->getOperand(1);
2727 if (isa<ConstantSDNode>(Op1)) {
2728 // Kill degenerate vector shifts:
2729 ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
2731 if (CN->getValue() == 0) {
2737 case SPUISD::PROMOTE_SCALAR: {
2738 switch (Op0.getOpcode()) {
2741 case ISD::ANY_EXTEND:
2742 case ISD::ZERO_EXTEND:
2743 case ISD::SIGN_EXTEND: {
2744 // (SPUpromote_scalar (any|sign|zero_extend (SPUextract_elt0 <arg>))) ->
2746 // but only if the SPUpromote_scalar and <arg> types match.
2747 SDOperand Op00 = Op0.getOperand(0);
2748 if (Op00.getOpcode() == SPUISD::EXTRACT_ELT0) {
2749 SDOperand Op000 = Op00.getOperand(0);
2750 if (Op000.getValueType() == N->getValueType(0)) {
2756 case SPUISD::EXTRACT_ELT0: {
2757 // (SPUpromote_scalar (SPUextract_elt0 <arg>)) ->
2759 Result = Op0.getOperand(0);
2766 // Otherwise, return unchanged.
2769 DEBUG(cerr << "\nReplace.SPU: ");
2770 DEBUG(N->dump(&DAG));
2771 DEBUG(cerr << "\nWith: ");
2772 DEBUG(Result.Val->dump(&DAG));
2773 DEBUG(cerr << "\n");
2780 //===----------------------------------------------------------------------===//
2781 // Inline Assembly Support
2782 //===----------------------------------------------------------------------===//
2784 /// getConstraintType - Given a constraint letter, return the type of
2785 /// constraint it is for this target.
2786 SPUTargetLowering::ConstraintType
2787 SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
2788 if (ConstraintLetter.size() == 1) {
2789 switch (ConstraintLetter[0]) {
2796 return C_RegisterClass;
2799 return TargetLowering::getConstraintType(ConstraintLetter);
2802 std::pair<unsigned, const TargetRegisterClass*>
2803 SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
2804 MVT::ValueType VT) const
2806 if (Constraint.size() == 1) {
2807 // GCC RS6000 Constraint Letters
2808 switch (Constraint[0]) {
2812 return std::make_pair(0U, SPU::R64CRegisterClass);
2813 return std::make_pair(0U, SPU::R32CRegisterClass);
2816 return std::make_pair(0U, SPU::R32FPRegisterClass);
2817 else if (VT == MVT::f64)
2818 return std::make_pair(0U, SPU::R64FPRegisterClass);
2821 return std::make_pair(0U, SPU::GPRCRegisterClass);
2825 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
2828 //! Compute used/known bits for a SPU operand
2830 SPUTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
2834 const SelectionDAG &DAG,
2835 unsigned Depth ) const {
2836 const uint64_t uint64_sizebits = sizeof(uint64_t) * 8;
2838 switch (Op.getOpcode()) {
2840 // KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
2850 case SPUISD::PROMOTE_SCALAR: {
2851 SDOperand Op0 = Op.getOperand(0);
2852 uint64_t InMask = MVT::getIntVTBitMask(Op0.getValueType());
2853 KnownZero |= APInt(uint64_sizebits, ~InMask, false);
2854 KnownOne |= APInt(uint64_sizebits, InMask, false);
2858 case SPUISD::LDRESULT:
2859 case SPUISD::EXTRACT_ELT0:
2860 case SPUISD::EXTRACT_ELT0_CHAINED: {
2861 uint64_t InMask = MVT::getIntVTBitMask(Op.getValueType());
2862 KnownZero |= APInt(uint64_sizebits, ~InMask, false);
2863 KnownOne |= APInt(uint64_sizebits, InMask, false);
2868 case EXTRACT_I1_ZEXT:
2869 case EXTRACT_I1_SEXT:
2870 case EXTRACT_I8_ZEXT:
2871 case EXTRACT_I8_SEXT:
2876 case SHLQUAD_L_BITS:
2877 case SHLQUAD_L_BYTES:
2883 case ROTQUAD_RZ_BYTES:
2884 case ROTQUAD_RZ_BITS:
2885 case ROTBYTES_RIGHT_S:
2887 case ROTBYTES_LEFT_CHAINED:
2898 // LowerAsmOperandForConstraint
2900 SPUTargetLowering::LowerAsmOperandForConstraint(SDOperand Op,
2901 char ConstraintLetter,
2902 std::vector<SDOperand> &Ops,
2903 SelectionDAG &DAG) {
2904 // Default, for the time being, to the base class handler
2905 TargetLowering::LowerAsmOperandForConstraint(Op, ConstraintLetter, Ops, DAG);
2908 /// isLegalAddressImmediate - Return true if the integer value can be used
2909 /// as the offset of the target addressing mode.
2910 bool SPUTargetLowering::isLegalAddressImmediate(int64_t V, const Type *Ty) const {
2911 // SPU's addresses are 256K:
2912 return (V > -(1 << 18) && V < (1 << 18) - 1);
2915 bool SPUTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {