2 //===-- 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 "SPUFrameInfo.h"
18 #include "SPUMachineFunction.h"
19 #include "llvm/Constants.h"
20 #include "llvm/Function.h"
21 #include "llvm/Intrinsics.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Type.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/SelectionDAG.h"
30 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/ADT/VectorExtras.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/raw_ostream.h"
41 // Used in getTargetNodeName() below
43 std::map<unsigned, const char *> node_names;
45 // Byte offset of the preferred slot (counted from the MSB)
46 int prefslotOffset(EVT VT) {
48 if (VT==MVT::i1) retval=3;
49 if (VT==MVT::i8) retval=3;
50 if (VT==MVT::i16) retval=2;
55 //! Expand a library call into an actual call DAG node
58 This code is taken from SelectionDAGLegalize, since it is not exposed as
59 part of the LLVM SelectionDAG API.
63 ExpandLibCall(RTLIB::Libcall LC, SDValue Op, SelectionDAG &DAG,
64 bool isSigned, SDValue &Hi, const SPUTargetLowering &TLI) {
65 // The input chain to this libcall is the entry node of the function.
66 // Legalizing the call will automatically add the previous call to the
68 SDValue InChain = DAG.getEntryNode();
70 TargetLowering::ArgListTy Args;
71 TargetLowering::ArgListEntry Entry;
72 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
73 EVT ArgVT = Op.getOperand(i).getValueType();
74 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
75 Entry.Node = Op.getOperand(i);
77 Entry.isSExt = isSigned;
78 Entry.isZExt = !isSigned;
79 Args.push_back(Entry);
81 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
84 // Splice the libcall in wherever FindInputOutputChains tells us to.
86 Op.getNode()->getValueType(0).getTypeForEVT(*DAG.getContext());
87 std::pair<SDValue, SDValue> CallInfo =
88 TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
89 0, TLI.getLibcallCallingConv(LC), false,
90 /*isReturnValueUsed=*/true,
91 Callee, Args, DAG, Op.getDebugLoc());
93 return CallInfo.first;
97 SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
98 : TargetLowering(TM, new TargetLoweringObjectFileELF()),
101 // Use _setjmp/_longjmp instead of setjmp/longjmp.
102 setUseUnderscoreSetJmp(true);
103 setUseUnderscoreLongJmp(true);
105 // Set RTLIB libcall names as used by SPU:
106 setLibcallName(RTLIB::DIV_F64, "__fast_divdf3");
108 // Set up the SPU's register classes:
109 addRegisterClass(MVT::i8, SPU::R8CRegisterClass);
110 addRegisterClass(MVT::i16, SPU::R16CRegisterClass);
111 addRegisterClass(MVT::i32, SPU::R32CRegisterClass);
112 addRegisterClass(MVT::i64, SPU::R64CRegisterClass);
113 addRegisterClass(MVT::f32, SPU::R32FPRegisterClass);
114 addRegisterClass(MVT::f64, SPU::R64FPRegisterClass);
115 addRegisterClass(MVT::i128, SPU::GPRCRegisterClass);
117 // SPU has no sign or zero extended loads for i1, i8, i16:
118 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
119 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
122 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
123 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
125 setTruncStoreAction(MVT::i128, MVT::i64, Expand);
126 setTruncStoreAction(MVT::i128, MVT::i32, Expand);
127 setTruncStoreAction(MVT::i128, MVT::i16, Expand);
128 setTruncStoreAction(MVT::i128, MVT::i8, Expand);
130 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
132 // SPU constant load actions are custom lowered:
133 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
134 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
136 // SPU's loads and stores have to be custom lowered:
137 for (unsigned sctype = (unsigned) MVT::i8; sctype < (unsigned) MVT::i128;
139 MVT::SimpleValueType VT = (MVT::SimpleValueType)sctype;
141 setOperationAction(ISD::LOAD, VT, Custom);
142 setOperationAction(ISD::STORE, VT, Custom);
143 setLoadExtAction(ISD::EXTLOAD, VT, Custom);
144 setLoadExtAction(ISD::ZEXTLOAD, VT, Custom);
145 setLoadExtAction(ISD::SEXTLOAD, VT, Custom);
147 for (unsigned stype = sctype - 1; stype >= (unsigned) MVT::i8; --stype) {
148 MVT::SimpleValueType StoreVT = (MVT::SimpleValueType) stype;
149 setTruncStoreAction(VT, StoreVT, Expand);
153 for (unsigned sctype = (unsigned) MVT::f32; sctype < (unsigned) MVT::f64;
155 MVT::SimpleValueType VT = (MVT::SimpleValueType) sctype;
157 setOperationAction(ISD::LOAD, VT, Custom);
158 setOperationAction(ISD::STORE, VT, Custom);
160 for (unsigned stype = sctype - 1; stype >= (unsigned) MVT::f32; --stype) {
161 MVT::SimpleValueType StoreVT = (MVT::SimpleValueType) stype;
162 setTruncStoreAction(VT, StoreVT, Expand);
166 // Expand the jumptable branches
167 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
168 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
170 // Custom lower SELECT_CC for most cases, but expand by default
171 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
172 setOperationAction(ISD::SELECT_CC, MVT::i8, Custom);
173 setOperationAction(ISD::SELECT_CC, MVT::i16, Custom);
174 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
175 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
177 // SPU has no intrinsics for these particular operations:
178 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
180 // SPU has no division/remainder instructions
181 setOperationAction(ISD::SREM, MVT::i8, Expand);
182 setOperationAction(ISD::UREM, MVT::i8, Expand);
183 setOperationAction(ISD::SDIV, MVT::i8, Expand);
184 setOperationAction(ISD::UDIV, MVT::i8, Expand);
185 setOperationAction(ISD::SDIVREM, MVT::i8, Expand);
186 setOperationAction(ISD::UDIVREM, MVT::i8, Expand);
187 setOperationAction(ISD::SREM, MVT::i16, Expand);
188 setOperationAction(ISD::UREM, MVT::i16, Expand);
189 setOperationAction(ISD::SDIV, MVT::i16, Expand);
190 setOperationAction(ISD::UDIV, MVT::i16, Expand);
191 setOperationAction(ISD::SDIVREM, MVT::i16, Expand);
192 setOperationAction(ISD::UDIVREM, MVT::i16, Expand);
193 setOperationAction(ISD::SREM, MVT::i32, Expand);
194 setOperationAction(ISD::UREM, MVT::i32, Expand);
195 setOperationAction(ISD::SDIV, MVT::i32, Expand);
196 setOperationAction(ISD::UDIV, MVT::i32, Expand);
197 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
198 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
199 setOperationAction(ISD::SREM, MVT::i64, Expand);
200 setOperationAction(ISD::UREM, MVT::i64, Expand);
201 setOperationAction(ISD::SDIV, MVT::i64, Expand);
202 setOperationAction(ISD::UDIV, MVT::i64, Expand);
203 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
204 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
205 setOperationAction(ISD::SREM, MVT::i128, Expand);
206 setOperationAction(ISD::UREM, MVT::i128, Expand);
207 setOperationAction(ISD::SDIV, MVT::i128, Expand);
208 setOperationAction(ISD::UDIV, MVT::i128, Expand);
209 setOperationAction(ISD::SDIVREM, MVT::i128, Expand);
210 setOperationAction(ISD::UDIVREM, MVT::i128, Expand);
212 // We don't support sin/cos/sqrt/fmod
213 setOperationAction(ISD::FSIN , MVT::f64, Expand);
214 setOperationAction(ISD::FCOS , MVT::f64, Expand);
215 setOperationAction(ISD::FREM , MVT::f64, Expand);
216 setOperationAction(ISD::FSIN , MVT::f32, Expand);
217 setOperationAction(ISD::FCOS , MVT::f32, Expand);
218 setOperationAction(ISD::FREM , MVT::f32, Expand);
220 // Expand fsqrt to the appropriate libcall (NOTE: should use h/w fsqrt
222 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
223 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
225 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
226 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
228 // SPU can do rotate right and left, so legalize it... but customize for i8
229 // because instructions don't exist.
231 // FIXME: Change from "expand" to appropriate type once ROTR is supported in
233 setOperationAction(ISD::ROTR, MVT::i32, Expand /*Legal*/);
234 setOperationAction(ISD::ROTR, MVT::i16, Expand /*Legal*/);
235 setOperationAction(ISD::ROTR, MVT::i8, Expand /*Custom*/);
237 setOperationAction(ISD::ROTL, MVT::i32, Legal);
238 setOperationAction(ISD::ROTL, MVT::i16, Legal);
239 setOperationAction(ISD::ROTL, MVT::i8, Custom);
241 // SPU has no native version of shift left/right for i8
242 setOperationAction(ISD::SHL, MVT::i8, Custom);
243 setOperationAction(ISD::SRL, MVT::i8, Custom);
244 setOperationAction(ISD::SRA, MVT::i8, Custom);
246 // Make these operations legal and handle them during instruction selection:
247 setOperationAction(ISD::SHL, MVT::i64, Legal);
248 setOperationAction(ISD::SRL, MVT::i64, Legal);
249 setOperationAction(ISD::SRA, MVT::i64, Legal);
251 // Custom lower i8, i32 and i64 multiplications
252 setOperationAction(ISD::MUL, MVT::i8, Custom);
253 setOperationAction(ISD::MUL, MVT::i32, Legal);
254 setOperationAction(ISD::MUL, MVT::i64, Legal);
256 // Expand double-width multiplication
257 // FIXME: It would probably be reasonable to support some of these operations
258 setOperationAction(ISD::UMUL_LOHI, MVT::i8, Expand);
259 setOperationAction(ISD::SMUL_LOHI, MVT::i8, Expand);
260 setOperationAction(ISD::MULHU, MVT::i8, Expand);
261 setOperationAction(ISD::MULHS, MVT::i8, Expand);
262 setOperationAction(ISD::UMUL_LOHI, MVT::i16, Expand);
263 setOperationAction(ISD::SMUL_LOHI, MVT::i16, Expand);
264 setOperationAction(ISD::MULHU, MVT::i16, Expand);
265 setOperationAction(ISD::MULHS, MVT::i16, Expand);
266 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
267 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
268 setOperationAction(ISD::MULHU, MVT::i32, Expand);
269 setOperationAction(ISD::MULHS, MVT::i32, Expand);
270 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
271 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
272 setOperationAction(ISD::MULHU, MVT::i64, Expand);
273 setOperationAction(ISD::MULHS, MVT::i64, Expand);
275 // Need to custom handle (some) common i8, i64 math ops
276 setOperationAction(ISD::ADD, MVT::i8, Custom);
277 setOperationAction(ISD::ADD, MVT::i64, Legal);
278 setOperationAction(ISD::SUB, MVT::i8, Custom);
279 setOperationAction(ISD::SUB, MVT::i64, Legal);
281 // SPU does not have BSWAP. It does have i32 support CTLZ.
282 // CTPOP has to be custom lowered.
283 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
284 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
286 setOperationAction(ISD::CTPOP, MVT::i8, Custom);
287 setOperationAction(ISD::CTPOP, MVT::i16, Custom);
288 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
289 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
290 setOperationAction(ISD::CTPOP, MVT::i128, Expand);
292 setOperationAction(ISD::CTTZ , MVT::i8, Expand);
293 setOperationAction(ISD::CTTZ , MVT::i16, Expand);
294 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
295 setOperationAction(ISD::CTTZ , MVT::i64, Expand);
296 setOperationAction(ISD::CTTZ , MVT::i128, Expand);
298 setOperationAction(ISD::CTLZ , MVT::i8, Promote);
299 setOperationAction(ISD::CTLZ , MVT::i16, Promote);
300 setOperationAction(ISD::CTLZ , MVT::i32, Legal);
301 setOperationAction(ISD::CTLZ , MVT::i64, Expand);
302 setOperationAction(ISD::CTLZ , MVT::i128, Expand);
304 // SPU has a version of select that implements (a&~c)|(b&c), just like
305 // select ought to work:
306 setOperationAction(ISD::SELECT, MVT::i8, Legal);
307 setOperationAction(ISD::SELECT, MVT::i16, Legal);
308 setOperationAction(ISD::SELECT, MVT::i32, Legal);
309 setOperationAction(ISD::SELECT, MVT::i64, Legal);
311 setOperationAction(ISD::SETCC, MVT::i8, Legal);
312 setOperationAction(ISD::SETCC, MVT::i16, Legal);
313 setOperationAction(ISD::SETCC, MVT::i32, Legal);
314 setOperationAction(ISD::SETCC, MVT::i64, Legal);
315 setOperationAction(ISD::SETCC, MVT::f64, Custom);
317 // Custom lower i128 -> i64 truncates
318 setOperationAction(ISD::TRUNCATE, MVT::i64, Custom);
320 // Custom lower i32/i64 -> i128 sign extend
321 setOperationAction(ISD::SIGN_EXTEND, MVT::i128, Custom);
323 setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote);
324 setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote);
325 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
326 setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
327 // SPU has a legal FP -> signed INT instruction for f32, but for f64, need
328 // to expand to a libcall, hence the custom lowering:
329 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
330 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
331 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
332 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
333 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Expand);
334 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Expand);
336 // FDIV on SPU requires custom lowering
337 setOperationAction(ISD::FDIV, MVT::f64, Expand); // to libcall
339 // SPU has [U|S]INT_TO_FP for f32->i32, but not for f64->i32, f64->i64:
340 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
341 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
342 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
343 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
344 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
345 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
346 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
347 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
349 setOperationAction(ISD::BITCAST, MVT::i32, Legal);
350 setOperationAction(ISD::BITCAST, MVT::f32, Legal);
351 setOperationAction(ISD::BITCAST, MVT::i64, Legal);
352 setOperationAction(ISD::BITCAST, MVT::f64, Legal);
354 // We cannot sextinreg(i1). Expand to shifts.
355 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
357 // We want to legalize GlobalAddress and ConstantPool nodes into the
358 // appropriate instructions to materialize the address.
359 for (unsigned sctype = (unsigned) MVT::i8; sctype < (unsigned) MVT::f128;
361 MVT::SimpleValueType VT = (MVT::SimpleValueType)sctype;
363 setOperationAction(ISD::GlobalAddress, VT, Custom);
364 setOperationAction(ISD::ConstantPool, VT, Custom);
365 setOperationAction(ISD::JumpTable, VT, Custom);
368 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
369 setOperationAction(ISD::VASTART , MVT::Other, Custom);
371 // Use the default implementation.
372 setOperationAction(ISD::VAARG , MVT::Other, Expand);
373 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
374 setOperationAction(ISD::VAEND , MVT::Other, Expand);
375 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
376 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
377 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
378 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Expand);
380 // Cell SPU has instructions for converting between i64 and fp.
381 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
382 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
384 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
385 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
387 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
388 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
390 // First set operation action for all vector types to expand. Then we
391 // will selectively turn on ones that can be effectively codegen'd.
392 addRegisterClass(MVT::v16i8, SPU::VECREGRegisterClass);
393 addRegisterClass(MVT::v8i16, SPU::VECREGRegisterClass);
394 addRegisterClass(MVT::v4i32, SPU::VECREGRegisterClass);
395 addRegisterClass(MVT::v2i64, SPU::VECREGRegisterClass);
396 addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
397 addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
399 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
400 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
401 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
403 // add/sub are legal for all supported vector VT's.
404 setOperationAction(ISD::ADD, VT, Legal);
405 setOperationAction(ISD::SUB, VT, Legal);
406 // mul has to be custom lowered.
407 setOperationAction(ISD::MUL, VT, Legal);
409 setOperationAction(ISD::AND, VT, Legal);
410 setOperationAction(ISD::OR, VT, Legal);
411 setOperationAction(ISD::XOR, VT, Legal);
412 setOperationAction(ISD::LOAD, VT, Custom);
413 setOperationAction(ISD::SELECT, VT, Legal);
414 setOperationAction(ISD::STORE, VT, Custom);
416 // These operations need to be expanded:
417 setOperationAction(ISD::SDIV, VT, Expand);
418 setOperationAction(ISD::SREM, VT, Expand);
419 setOperationAction(ISD::UDIV, VT, Expand);
420 setOperationAction(ISD::UREM, VT, Expand);
422 // Custom lower build_vector, constant pool spills, insert and
423 // extract vector elements:
424 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
425 setOperationAction(ISD::ConstantPool, VT, Custom);
426 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
427 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
428 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
429 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
432 setOperationAction(ISD::AND, MVT::v16i8, Custom);
433 setOperationAction(ISD::OR, MVT::v16i8, Custom);
434 setOperationAction(ISD::XOR, MVT::v16i8, Custom);
435 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
437 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
439 setShiftAmountType(MVT::i32);
440 setBooleanContents(ZeroOrNegativeOneBooleanContent);
442 setStackPointerRegisterToSaveRestore(SPU::R1);
444 // We have target-specific dag combine patterns for the following nodes:
445 setTargetDAGCombine(ISD::ADD);
446 setTargetDAGCombine(ISD::ZERO_EXTEND);
447 setTargetDAGCombine(ISD::SIGN_EXTEND);
448 setTargetDAGCombine(ISD::ANY_EXTEND);
450 computeRegisterProperties();
452 // Set pre-RA register scheduler default to BURR, which produces slightly
453 // better code than the default (could also be TDRR, but TargetLowering.h
454 // needs a mod to support that model):
455 setSchedulingPreference(Sched::RegPressure);
459 SPUTargetLowering::getTargetNodeName(unsigned Opcode) const
461 if (node_names.empty()) {
462 node_names[(unsigned) SPUISD::RET_FLAG] = "SPUISD::RET_FLAG";
463 node_names[(unsigned) SPUISD::Hi] = "SPUISD::Hi";
464 node_names[(unsigned) SPUISD::Lo] = "SPUISD::Lo";
465 node_names[(unsigned) SPUISD::PCRelAddr] = "SPUISD::PCRelAddr";
466 node_names[(unsigned) SPUISD::AFormAddr] = "SPUISD::AFormAddr";
467 node_names[(unsigned) SPUISD::IndirectAddr] = "SPUISD::IndirectAddr";
468 node_names[(unsigned) SPUISD::LDRESULT] = "SPUISD::LDRESULT";
469 node_names[(unsigned) SPUISD::CALL] = "SPUISD::CALL";
470 node_names[(unsigned) SPUISD::SHUFB] = "SPUISD::SHUFB";
471 node_names[(unsigned) SPUISD::SHUFFLE_MASK] = "SPUISD::SHUFFLE_MASK";
472 node_names[(unsigned) SPUISD::CNTB] = "SPUISD::CNTB";
473 node_names[(unsigned) SPUISD::PREFSLOT2VEC] = "SPUISD::PREFSLOT2VEC";
474 node_names[(unsigned) SPUISD::VEC2PREFSLOT] = "SPUISD::VEC2PREFSLOT";
475 node_names[(unsigned) SPUISD::SHL_BITS] = "SPUISD::SHL_BITS";
476 node_names[(unsigned) SPUISD::SHL_BYTES] = "SPUISD::SHL_BYTES";
477 node_names[(unsigned) SPUISD::VEC_ROTL] = "SPUISD::VEC_ROTL";
478 node_names[(unsigned) SPUISD::VEC_ROTR] = "SPUISD::VEC_ROTR";
479 node_names[(unsigned) SPUISD::ROTBYTES_LEFT] = "SPUISD::ROTBYTES_LEFT";
480 node_names[(unsigned) SPUISD::ROTBYTES_LEFT_BITS] =
481 "SPUISD::ROTBYTES_LEFT_BITS";
482 node_names[(unsigned) SPUISD::SELECT_MASK] = "SPUISD::SELECT_MASK";
483 node_names[(unsigned) SPUISD::SELB] = "SPUISD::SELB";
484 node_names[(unsigned) SPUISD::ADD64_MARKER] = "SPUISD::ADD64_MARKER";
485 node_names[(unsigned) SPUISD::SUB64_MARKER] = "SPUISD::SUB64_MARKER";
486 node_names[(unsigned) SPUISD::MUL64_MARKER] = "SPUISD::MUL64_MARKER";
489 std::map<unsigned, const char *>::iterator i = node_names.find(Opcode);
491 return ((i != node_names.end()) ? i->second : 0);
494 /// getFunctionAlignment - Return the Log2 alignment of this function.
495 unsigned SPUTargetLowering::getFunctionAlignment(const Function *) const {
499 //===----------------------------------------------------------------------===//
500 // Return the Cell SPU's SETCC result type
501 //===----------------------------------------------------------------------===//
503 MVT::SimpleValueType SPUTargetLowering::getSetCCResultType(EVT VT) const {
504 // i16 and i32 are valid SETCC result types
505 return ((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) ?
506 VT.getSimpleVT().SimpleTy :
510 //===----------------------------------------------------------------------===//
511 // Calling convention code:
512 //===----------------------------------------------------------------------===//
514 #include "SPUGenCallingConv.inc"
516 //===----------------------------------------------------------------------===//
517 // LowerOperation implementation
518 //===----------------------------------------------------------------------===//
520 /// Custom lower loads for CellSPU
522 All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
523 within a 16-byte block, we have to rotate to extract the requested element.
525 For extending loads, we also want to ensure that the following sequence is
526 emitted, e.g. for MVT::f32 extending load to MVT::f64:
530 %2 v16i8,ch = rotate %1
531 %3 v4f8, ch = bitconvert %2
532 %4 f32 = vec2perfslot %3
533 %5 f64 = fp_extend %4
537 LowerLOAD(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
538 LoadSDNode *LN = cast<LoadSDNode>(Op);
539 SDValue the_chain = LN->getChain();
540 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
541 EVT InVT = LN->getMemoryVT();
542 EVT OutVT = Op.getValueType();
543 ISD::LoadExtType ExtType = LN->getExtensionType();
544 unsigned alignment = LN->getAlignment();
545 int pso = prefslotOffset(InVT);
546 DebugLoc dl = Op.getDebugLoc();
547 EVT vecVT = InVT.isVector()? InVT: EVT::getVectorVT(*DAG.getContext(), InVT,
548 (128 / InVT.getSizeInBits()));
551 assert( LN->getAddressingMode() == ISD::UNINDEXED
552 && "we should get only UNINDEXED adresses");
553 // clean aligned loads can be selected as-is
554 if (InVT.getSizeInBits() == 128 && alignment == 16)
557 // Get pointerinfos to the memory chunk(s) that contain the data to load
558 uint64_t mpi_offset = LN->getPointerInfo().Offset;
559 mpi_offset -= mpi_offset%16;
560 MachinePointerInfo lowMemPtr(LN->getPointerInfo().V, mpi_offset);
561 MachinePointerInfo highMemPtr(LN->getPointerInfo().V, mpi_offset+16);
564 SDValue basePtr = LN->getBasePtr();
567 if (alignment == 16) {
570 // Special cases for a known aligned load to simplify the base pointer
571 // and the rotation amount:
572 if (basePtr.getOpcode() == ISD::ADD
573 && (CN = dyn_cast<ConstantSDNode > (basePtr.getOperand(1))) != 0) {
574 // Known offset into basePtr
575 int64_t offset = CN->getSExtValue();
576 int64_t rotamt = int64_t((offset & 0xf) - pso);
581 rotate = DAG.getConstant(rotamt, MVT::i16);
583 // Simplify the base pointer for this case:
584 basePtr = basePtr.getOperand(0);
585 if ((offset & ~0xf) > 0) {
586 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
588 DAG.getConstant((offset & ~0xf), PtrVT));
590 } else if ((basePtr.getOpcode() == SPUISD::AFormAddr)
591 || (basePtr.getOpcode() == SPUISD::IndirectAddr
592 && basePtr.getOperand(0).getOpcode() == SPUISD::Hi
593 && basePtr.getOperand(1).getOpcode() == SPUISD::Lo)) {
594 // Plain aligned a-form address: rotate into preferred slot
595 // Same for (SPUindirect (SPUhi ...), (SPUlo ...))
596 int64_t rotamt = -pso;
599 rotate = DAG.getConstant(rotamt, MVT::i16);
601 // Offset the rotate amount by the basePtr and the preferred slot
603 int64_t rotamt = -pso;
606 rotate = DAG.getNode(ISD::ADD, dl, PtrVT,
608 DAG.getConstant(rotamt, PtrVT));
611 // Unaligned load: must be more pessimistic about addressing modes:
612 if (basePtr.getOpcode() == ISD::ADD) {
613 MachineFunction &MF = DAG.getMachineFunction();
614 MachineRegisterInfo &RegInfo = MF.getRegInfo();
615 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
618 SDValue Op0 = basePtr.getOperand(0);
619 SDValue Op1 = basePtr.getOperand(1);
621 if (isa<ConstantSDNode>(Op1)) {
622 // Convert the (add <ptr>, <const>) to an indirect address contained
623 // in a register. Note that this is done because we need to avoid
624 // creating a 0(reg) d-form address due to the SPU's block loads.
625 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
626 the_chain = DAG.getCopyToReg(the_chain, dl, VReg, basePtr, Flag);
627 basePtr = DAG.getCopyFromReg(the_chain, dl, VReg, PtrVT);
629 // Convert the (add <arg1>, <arg2>) to an indirect address, which
630 // will likely be lowered as a reg(reg) x-form address.
631 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
634 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
636 DAG.getConstant(0, PtrVT));
639 // Offset the rotate amount by the basePtr and the preferred slot
641 rotate = DAG.getNode(ISD::ADD, dl, PtrVT,
643 DAG.getConstant(-pso, PtrVT));
646 // Do the load as a i128 to allow possible shifting
647 SDValue low = DAG.getLoad(MVT::i128, dl, the_chain, basePtr,
649 LN->isVolatile(), LN->isNonTemporal(), 16);
651 // When the size is not greater than alignment we get all data with just
653 if (alignment >= InVT.getSizeInBits()/8) {
655 the_chain = low.getValue(1);
657 // Rotate into the preferred slot:
658 result = DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, MVT::i128,
659 low.getValue(0), rotate);
661 // Convert the loaded v16i8 vector to the appropriate vector type
662 // specified by the operand:
663 EVT vecVT = EVT::getVectorVT(*DAG.getContext(),
664 InVT, (128 / InVT.getSizeInBits()));
665 result = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, InVT,
666 DAG.getNode(ISD::BITCAST, dl, vecVT, result));
668 // When alignment is less than the size, we might need (known only at
669 // run-time) two loads
670 // TODO: if the memory address is composed only from constants, we have
671 // extra kowledge, and might avoid the second load
673 // storage position offset from lower 16 byte aligned memory chunk
674 SDValue offset = DAG.getNode(ISD::AND, dl, MVT::i32,
675 basePtr, DAG.getConstant( 0xf, MVT::i32 ) );
677 SDValue offset_compl = DAG.getNode(ISD::SUB, dl, MVT::i32,
678 DAG.getConstant( 16, MVT::i32),
680 // get a registerfull of ones. (this implementation is a workaround: LLVM
681 // cannot handle 128 bit signed int constants)
682 SDValue ones = DAG.getConstant(-1, MVT::v4i32 );
683 ones = DAG.getNode(ISD::BITCAST, dl, MVT::i128, ones);
685 SDValue high = DAG.getLoad(MVT::i128, dl, the_chain,
686 DAG.getNode(ISD::ADD, dl, PtrVT,
688 DAG.getConstant(16, PtrVT)),
690 LN->isVolatile(), LN->isNonTemporal(), 16);
692 the_chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(1),
695 // Shift the (possible) high part right to compensate the misalignemnt.
696 // if there is no highpart (i.e. value is i64 and offset is 4), this
697 // will zero out the high value.
698 high = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, high,
699 DAG.getNode(ISD::SUB, dl, MVT::i32,
700 DAG.getConstant( 16, MVT::i32),
704 // Shift the low similarily
705 // TODO: add SPUISD::SHL_BYTES
706 low = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, low, offset );
708 // Merge the two parts
709 result = DAG.getNode(ISD::BITCAST, dl, vecVT,
710 DAG.getNode(ISD::OR, dl, MVT::i128, low, high));
712 if (!InVT.isVector()) {
713 result = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, InVT, result );
717 // Handle extending loads by extending the scalar result:
718 if (ExtType == ISD::SEXTLOAD) {
719 result = DAG.getNode(ISD::SIGN_EXTEND, dl, OutVT, result);
720 } else if (ExtType == ISD::ZEXTLOAD) {
721 result = DAG.getNode(ISD::ZERO_EXTEND, dl, OutVT, result);
722 } else if (ExtType == ISD::EXTLOAD) {
723 unsigned NewOpc = ISD::ANY_EXTEND;
725 if (OutVT.isFloatingPoint())
726 NewOpc = ISD::FP_EXTEND;
728 result = DAG.getNode(NewOpc, dl, OutVT, result);
731 SDVTList retvts = DAG.getVTList(OutVT, MVT::Other);
732 SDValue retops[2] = {
737 result = DAG.getNode(SPUISD::LDRESULT, dl, retvts,
738 retops, sizeof(retops) / sizeof(retops[0]));
742 /// Custom lower stores for CellSPU
744 All CellSPU stores are aligned to 16-byte boundaries, so for elements
745 within a 16-byte block, we have to generate a shuffle to insert the
746 requested element into its place, then store the resulting block.
749 LowerSTORE(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
750 StoreSDNode *SN = cast<StoreSDNode>(Op);
751 SDValue Value = SN->getValue();
752 EVT VT = Value.getValueType();
753 EVT StVT = (!SN->isTruncatingStore() ? VT : SN->getMemoryVT());
754 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
755 DebugLoc dl = Op.getDebugLoc();
756 unsigned alignment = SN->getAlignment();
758 EVT vecVT = StVT.isVector()? StVT: EVT::getVectorVT(*DAG.getContext(), StVT,
759 (128 / StVT.getSizeInBits()));
760 // Get pointerinfos to the memory chunk(s) that contain the data to load
761 uint64_t mpi_offset = SN->getPointerInfo().Offset;
762 mpi_offset -= mpi_offset%16;
763 MachinePointerInfo lowMemPtr(SN->getPointerInfo().V, mpi_offset);
764 MachinePointerInfo highMemPtr(SN->getPointerInfo().V, mpi_offset+16);
768 assert( SN->getAddressingMode() == ISD::UNINDEXED
769 && "we should get only UNINDEXED adresses");
770 // clean aligned loads can be selected as-is
771 if (StVT.getSizeInBits() == 128 && alignment == 16)
774 SDValue alignLoadVec;
775 SDValue basePtr = SN->getBasePtr();
776 SDValue the_chain = SN->getChain();
777 SDValue insertEltOffs;
779 if (alignment == 16) {
781 // Special cases for a known aligned load to simplify the base pointer
782 // and insertion byte:
783 if (basePtr.getOpcode() == ISD::ADD
784 && (CN = dyn_cast<ConstantSDNode>(basePtr.getOperand(1))) != 0) {
785 // Known offset into basePtr
786 int64_t offset = CN->getSExtValue();
788 // Simplify the base pointer for this case:
789 basePtr = basePtr.getOperand(0);
790 insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
792 DAG.getConstant((offset & 0xf), PtrVT));
794 if ((offset & ~0xf) > 0) {
795 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
797 DAG.getConstant((offset & ~0xf), PtrVT));
800 // Otherwise, assume it's at byte 0 of basePtr
801 insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
803 DAG.getConstant(0, PtrVT));
804 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
806 DAG.getConstant(0, PtrVT));
809 // Unaligned load: must be more pessimistic about addressing modes:
810 if (basePtr.getOpcode() == ISD::ADD) {
811 MachineFunction &MF = DAG.getMachineFunction();
812 MachineRegisterInfo &RegInfo = MF.getRegInfo();
813 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
816 SDValue Op0 = basePtr.getOperand(0);
817 SDValue Op1 = basePtr.getOperand(1);
819 if (isa<ConstantSDNode>(Op1)) {
820 // Convert the (add <ptr>, <const>) to an indirect address contained
821 // in a register. Note that this is done because we need to avoid
822 // creating a 0(reg) d-form address due to the SPU's block loads.
823 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
824 the_chain = DAG.getCopyToReg(the_chain, dl, VReg, basePtr, Flag);
825 basePtr = DAG.getCopyFromReg(the_chain, dl, VReg, PtrVT);
827 // Convert the (add <arg1>, <arg2>) to an indirect address, which
828 // will likely be lowered as a reg(reg) x-form address.
829 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
832 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
834 DAG.getConstant(0, PtrVT));
837 // Insertion point is solely determined by basePtr's contents
838 insertEltOffs = DAG.getNode(ISD::ADD, dl, PtrVT,
840 DAG.getConstant(0, PtrVT));
843 // Load the lower part of the memory to which to store.
844 SDValue low = DAG.getLoad(vecVT, dl, the_chain, basePtr,
845 lowMemPtr, SN->isVolatile(), SN->isNonTemporal(), 16);
847 // if we don't need to store over the 16 byte boundary, one store suffices
848 if (alignment >= StVT.getSizeInBits()/8) {
850 the_chain = low.getValue(1);
852 LoadSDNode *LN = cast<LoadSDNode>(low);
853 SDValue theValue = SN->getValue();
856 && (theValue.getOpcode() == ISD::AssertZext
857 || theValue.getOpcode() == ISD::AssertSext)) {
858 // Drill down and get the value for zero- and sign-extended
860 theValue = theValue.getOperand(0);
863 // If the base pointer is already a D-form address, then just create
864 // a new D-form address with a slot offset and the orignal base pointer.
865 // Otherwise generate a D-form address with the slot offset relative
866 // to the stack pointer, which is always aligned.
868 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
869 errs() << "CellSPU LowerSTORE: basePtr = ";
870 basePtr.getNode()->dump(&DAG);
875 SDValue insertEltOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl, vecVT,
877 SDValue vectorizeOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, vecVT,
880 result = DAG.getNode(SPUISD::SHUFB, dl, vecVT,
882 DAG.getNode(ISD::BITCAST, dl,
883 MVT::v4i32, insertEltOp));
885 result = DAG.getStore(the_chain, dl, result, basePtr,
887 LN->isVolatile(), LN->isNonTemporal(),
891 // do the store when it might cross the 16 byte memory access boundary.
893 // TODO issue a warning if SN->isVolatile()== true? This is likely not
894 // what the user wanted.
896 // address offset from nearest lower 16byte alinged address
897 SDValue offset = DAG.getNode(ISD::AND, dl, MVT::i32,
899 DAG.getConstant(0xf, MVT::i32));
901 SDValue offset_compl = DAG.getNode(ISD::SUB, dl, MVT::i32,
902 DAG.getConstant( 16, MVT::i32),
904 SDValue hi_shift = DAG.getNode(ISD::SUB, dl, MVT::i32,
905 DAG.getConstant( VT.getSizeInBits()/8,
908 // 16 - sizeof(Value)
909 SDValue surplus = DAG.getNode(ISD::SUB, dl, MVT::i32,
910 DAG.getConstant( 16, MVT::i32),
911 DAG.getConstant( VT.getSizeInBits()/8,
913 // get a registerfull of ones
914 SDValue ones = DAG.getConstant(-1, MVT::v4i32);
915 ones = DAG.getNode(ISD::BITCAST, dl, MVT::i128, ones);
917 // Create the 128 bit masks that have ones where the data to store is
919 SDValue lowmask, himask;
920 // if the value to store don't fill up the an entire 128 bits, zero
921 // out the last bits of the mask so that only the value we want to store
923 // this is e.g. in the case of store i32, align 2
925 Value = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, Value);
926 lowmask = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, ones, surplus);
927 lowmask = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, lowmask,
929 Value = DAG.getNode(ISD::BITCAST, dl, MVT::i128, Value);
930 Value = DAG.getNode(ISD::AND, dl, MVT::i128, Value, lowmask);
935 Value = DAG.getNode(ISD::BITCAST, dl, MVT::i128, Value);
937 // this will zero, if there are no data that goes to the high quad
938 himask = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, lowmask,
940 lowmask = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, lowmask,
943 // Load in the old data and zero out the parts that will be overwritten with
944 // the new data to store.
945 SDValue hi = DAG.getLoad(MVT::i128, dl, the_chain,
946 DAG.getNode(ISD::ADD, dl, PtrVT, basePtr,
947 DAG.getConstant( 16, PtrVT)),
949 SN->isVolatile(), SN->isNonTemporal(), 16);
950 the_chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(1),
953 low = DAG.getNode(ISD::AND, dl, MVT::i128,
954 DAG.getNode( ISD::BITCAST, dl, MVT::i128, low),
955 DAG.getNode( ISD::XOR, dl, MVT::i128, lowmask, ones));
956 hi = DAG.getNode(ISD::AND, dl, MVT::i128,
957 DAG.getNode( ISD::BITCAST, dl, MVT::i128, hi),
958 DAG.getNode( ISD::XOR, dl, MVT::i128, himask, ones));
960 // Shift the Value to store into place. rlow contains the parts that go to
961 // the lower memory chunk, rhi has the parts that go to the upper one.
962 SDValue rlow = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, Value, offset);
963 rlow = DAG.getNode(ISD::AND, dl, MVT::i128, rlow, lowmask);
964 SDValue rhi = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, Value,
967 // Merge the old data and the new data and store the results
968 // Need to convert vectors here to integer as 'OR'ing floats assert
969 rlow = DAG.getNode(ISD::OR, dl, MVT::i128,
970 DAG.getNode(ISD::BITCAST, dl, MVT::i128, low),
971 DAG.getNode(ISD::BITCAST, dl, MVT::i128, rlow));
972 rhi = DAG.getNode(ISD::OR, dl, MVT::i128,
973 DAG.getNode(ISD::BITCAST, dl, MVT::i128, hi),
974 DAG.getNode(ISD::BITCAST, dl, MVT::i128, rhi));
976 low = DAG.getStore(the_chain, dl, rlow, basePtr,
978 SN->isVolatile(), SN->isNonTemporal(), 16);
979 hi = DAG.getStore(the_chain, dl, rhi,
980 DAG.getNode(ISD::ADD, dl, PtrVT, basePtr,
981 DAG.getConstant( 16, PtrVT)),
983 SN->isVolatile(), SN->isNonTemporal(), 16);
984 result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(0),
991 //! Generate the address of a constant pool entry.
993 LowerConstantPool(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
994 EVT PtrVT = Op.getValueType();
995 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
996 const Constant *C = CP->getConstVal();
997 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
998 SDValue Zero = DAG.getConstant(0, PtrVT);
999 const TargetMachine &TM = DAG.getTarget();
1000 // FIXME there is no actual debug info here
1001 DebugLoc dl = Op.getDebugLoc();
1003 if (TM.getRelocationModel() == Reloc::Static) {
1004 if (!ST->usingLargeMem()) {
1005 // Just return the SDValue with the constant pool address in it.
1006 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, CPI, Zero);
1008 SDValue Hi = DAG.getNode(SPUISD::Hi, dl, PtrVT, CPI, Zero);
1009 SDValue Lo = DAG.getNode(SPUISD::Lo, dl, PtrVT, CPI, Zero);
1010 return DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Hi, Lo);
1014 llvm_unreachable("LowerConstantPool: Relocation model other than static"
1019 //! Alternate entry point for generating the address of a constant pool entry
1021 SPU::LowerConstantPool(SDValue Op, SelectionDAG &DAG, const SPUTargetMachine &TM) {
1022 return ::LowerConstantPool(Op, DAG, TM.getSubtargetImpl());
1026 LowerJumpTable(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1027 EVT PtrVT = Op.getValueType();
1028 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1029 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1030 SDValue Zero = DAG.getConstant(0, PtrVT);
1031 const TargetMachine &TM = DAG.getTarget();
1032 // FIXME there is no actual debug info here
1033 DebugLoc dl = Op.getDebugLoc();
1035 if (TM.getRelocationModel() == Reloc::Static) {
1036 if (!ST->usingLargeMem()) {
1037 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, JTI, Zero);
1039 SDValue Hi = DAG.getNode(SPUISD::Hi, dl, PtrVT, JTI, Zero);
1040 SDValue Lo = DAG.getNode(SPUISD::Lo, dl, PtrVT, JTI, Zero);
1041 return DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Hi, Lo);
1045 llvm_unreachable("LowerJumpTable: Relocation model other than static"
1051 LowerGlobalAddress(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1052 EVT PtrVT = Op.getValueType();
1053 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1054 const GlobalValue *GV = GSDN->getGlobal();
1055 SDValue GA = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
1056 PtrVT, GSDN->getOffset());
1057 const TargetMachine &TM = DAG.getTarget();
1058 SDValue Zero = DAG.getConstant(0, PtrVT);
1059 // FIXME there is no actual debug info here
1060 DebugLoc dl = Op.getDebugLoc();
1062 if (TM.getRelocationModel() == Reloc::Static) {
1063 if (!ST->usingLargeMem()) {
1064 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, GA, Zero);
1066 SDValue Hi = DAG.getNode(SPUISD::Hi, dl, PtrVT, GA, Zero);
1067 SDValue Lo = DAG.getNode(SPUISD::Lo, dl, PtrVT, GA, Zero);
1068 return DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Hi, Lo);
1071 report_fatal_error("LowerGlobalAddress: Relocation model other than static"
1079 //! Custom lower double precision floating point constants
1081 LowerConstantFP(SDValue Op, SelectionDAG &DAG) {
1082 EVT VT = Op.getValueType();
1083 // FIXME there is no actual debug info here
1084 DebugLoc dl = Op.getDebugLoc();
1086 if (VT == MVT::f64) {
1087 ConstantFPSDNode *FP = cast<ConstantFPSDNode>(Op.getNode());
1090 "LowerConstantFP: Node is not ConstantFPSDNode");
1092 uint64_t dbits = DoubleToBits(FP->getValueAPF().convertToDouble());
1093 SDValue T = DAG.getConstant(dbits, MVT::i64);
1094 SDValue Tvec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, T, T);
1095 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
1096 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Tvec));
1103 SPUTargetLowering::LowerFormalArguments(SDValue Chain,
1104 CallingConv::ID CallConv, bool isVarArg,
1105 const SmallVectorImpl<ISD::InputArg>
1107 DebugLoc dl, SelectionDAG &DAG,
1108 SmallVectorImpl<SDValue> &InVals)
1111 MachineFunction &MF = DAG.getMachineFunction();
1112 MachineFrameInfo *MFI = MF.getFrameInfo();
1113 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1114 SPUFunctionInfo *FuncInfo = MF.getInfo<SPUFunctionInfo>();
1116 unsigned ArgOffset = SPUFrameInfo::minStackSize();
1117 unsigned ArgRegIdx = 0;
1118 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
1120 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1122 SmallVector<CCValAssign, 16> ArgLocs;
1123 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1125 // FIXME: allow for other calling conventions
1126 CCInfo.AnalyzeFormalArguments(Ins, CCC_SPU);
1128 // Add DAG nodes to load the arguments or copy them out of registers.
1129 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1130 EVT ObjectVT = Ins[ArgNo].VT;
1131 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1133 CCValAssign &VA = ArgLocs[ArgNo];
1135 if (VA.isRegLoc()) {
1136 const TargetRegisterClass *ArgRegClass;
1138 switch (ObjectVT.getSimpleVT().SimpleTy) {
1140 report_fatal_error("LowerFormalArguments Unhandled argument type: " +
1141 Twine(ObjectVT.getEVTString()));
1143 ArgRegClass = &SPU::R8CRegClass;
1146 ArgRegClass = &SPU::R16CRegClass;
1149 ArgRegClass = &SPU::R32CRegClass;
1152 ArgRegClass = &SPU::R64CRegClass;
1155 ArgRegClass = &SPU::GPRCRegClass;
1158 ArgRegClass = &SPU::R32FPRegClass;
1161 ArgRegClass = &SPU::R64FPRegClass;
1169 ArgRegClass = &SPU::VECREGRegClass;
1173 unsigned VReg = RegInfo.createVirtualRegister(ArgRegClass);
1174 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1175 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1178 // We need to load the argument to a virtual register if we determined
1179 // above that we ran out of physical registers of the appropriate type
1180 // or we're forced to do vararg
1181 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset, true);
1182 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1183 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
1185 ArgOffset += StackSlotSize;
1188 InVals.push_back(ArgVal);
1190 Chain = ArgVal.getOperand(0);
1195 // FIXME: we should be able to query the argument registers from
1196 // tablegen generated code.
1197 static const unsigned ArgRegs[] = {
1198 SPU::R3, SPU::R4, SPU::R5, SPU::R6, SPU::R7, SPU::R8, SPU::R9,
1199 SPU::R10, SPU::R11, SPU::R12, SPU::R13, SPU::R14, SPU::R15, SPU::R16,
1200 SPU::R17, SPU::R18, SPU::R19, SPU::R20, SPU::R21, SPU::R22, SPU::R23,
1201 SPU::R24, SPU::R25, SPU::R26, SPU::R27, SPU::R28, SPU::R29, SPU::R30,
1202 SPU::R31, SPU::R32, SPU::R33, SPU::R34, SPU::R35, SPU::R36, SPU::R37,
1203 SPU::R38, SPU::R39, SPU::R40, SPU::R41, SPU::R42, SPU::R43, SPU::R44,
1204 SPU::R45, SPU::R46, SPU::R47, SPU::R48, SPU::R49, SPU::R50, SPU::R51,
1205 SPU::R52, SPU::R53, SPU::R54, SPU::R55, SPU::R56, SPU::R57, SPU::R58,
1206 SPU::R59, SPU::R60, SPU::R61, SPU::R62, SPU::R63, SPU::R64, SPU::R65,
1207 SPU::R66, SPU::R67, SPU::R68, SPU::R69, SPU::R70, SPU::R71, SPU::R72,
1208 SPU::R73, SPU::R74, SPU::R75, SPU::R76, SPU::R77, SPU::R78, SPU::R79
1210 // size of ArgRegs array
1211 unsigned NumArgRegs = 77;
1213 // We will spill (79-3)+1 registers to the stack
1214 SmallVector<SDValue, 79-3+1> MemOps;
1216 // Create the frame slot
1217 for (; ArgRegIdx != NumArgRegs; ++ArgRegIdx) {
1218 FuncInfo->setVarArgsFrameIndex(
1219 MFI->CreateFixedObject(StackSlotSize, ArgOffset, true));
1220 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1221 unsigned VReg = MF.addLiveIn(ArgRegs[ArgRegIdx], &SPU::R32CRegClass);
1222 SDValue ArgVal = DAG.getRegister(VReg, MVT::v16i8);
1223 SDValue Store = DAG.getStore(Chain, dl, ArgVal, FIN, MachinePointerInfo(),
1225 Chain = Store.getOperand(0);
1226 MemOps.push_back(Store);
1228 // Increment address by stack slot size for the next stored argument
1229 ArgOffset += StackSlotSize;
1231 if (!MemOps.empty())
1232 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1233 &MemOps[0], MemOps.size());
1239 /// isLSAAddress - Return the immediate to use if the specified
1240 /// value is representable as a LSA address.
1241 static SDNode *isLSAAddress(SDValue Op, SelectionDAG &DAG) {
1242 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1245 int Addr = C->getZExtValue();
1246 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1247 (Addr << 14 >> 14) != Addr)
1248 return 0; // Top 14 bits have to be sext of immediate.
1250 return DAG.getConstant((int)C->getZExtValue() >> 2, MVT::i32).getNode();
1254 SPUTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1255 CallingConv::ID CallConv, bool isVarArg,
1257 const SmallVectorImpl<ISD::OutputArg> &Outs,
1258 const SmallVectorImpl<SDValue> &OutVals,
1259 const SmallVectorImpl<ISD::InputArg> &Ins,
1260 DebugLoc dl, SelectionDAG &DAG,
1261 SmallVectorImpl<SDValue> &InVals) const {
1262 // CellSPU target does not yet support tail call optimization.
1265 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
1266 unsigned NumOps = Outs.size();
1267 unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
1269 SmallVector<CCValAssign, 16> ArgLocs;
1270 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1272 // FIXME: allow for other calling conventions
1273 CCInfo.AnalyzeCallOperands(Outs, CCC_SPU);
1275 const unsigned NumArgRegs = ArgLocs.size();
1278 // Handy pointer type
1279 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1281 // Set up a copy of the stack pointer for use loading and storing any
1282 // arguments that may not fit in the registers available for argument
1284 SDValue StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
1286 // Figure out which arguments are going to go in registers, and which in
1288 unsigned ArgOffset = SPUFrameInfo::minStackSize(); // Just below [LR]
1289 unsigned ArgRegIdx = 0;
1291 // Keep track of registers passing arguments
1292 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
1293 // And the arguments passed on the stack
1294 SmallVector<SDValue, 8> MemOpChains;
1296 for (; ArgRegIdx != NumOps; ++ArgRegIdx) {
1297 SDValue Arg = OutVals[ArgRegIdx];
1298 CCValAssign &VA = ArgLocs[ArgRegIdx];
1300 // PtrOff will be used to store the current argument to the stack if a
1301 // register cannot be found for it.
1302 SDValue PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1303 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
1305 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
1306 default: llvm_unreachable("Unexpected ValueType for argument!");
1320 if (ArgRegIdx != NumArgRegs) {
1321 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1323 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
1324 MachinePointerInfo(),
1326 ArgOffset += StackSlotSize;
1332 // Accumulate how many bytes are to be pushed on the stack, including the
1333 // linkage area, and parameter passing area. According to the SPU ABI,
1334 // we minimally need space for [LR] and [SP].
1335 unsigned NumStackBytes = ArgOffset - SPUFrameInfo::minStackSize();
1337 // Insert a call sequence start
1338 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumStackBytes,
1341 if (!MemOpChains.empty()) {
1342 // Adjust the stack pointer for the stack arguments.
1343 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1344 &MemOpChains[0], MemOpChains.size());
1347 // Build a sequence of copy-to-reg nodes chained together with token chain
1348 // and flag operands which copy the outgoing args into the appropriate regs.
1350 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1351 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1352 RegsToPass[i].second, InFlag);
1353 InFlag = Chain.getValue(1);
1356 SmallVector<SDValue, 8> Ops;
1357 unsigned CallOpc = SPUISD::CALL;
1359 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1360 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1361 // node so that legalize doesn't hack it.
1362 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1363 const GlobalValue *GV = G->getGlobal();
1364 EVT CalleeVT = Callee.getValueType();
1365 SDValue Zero = DAG.getConstant(0, PtrVT);
1366 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, CalleeVT);
1368 if (!ST->usingLargeMem()) {
1369 // Turn calls to targets that are defined (i.e., have bodies) into BRSL
1370 // style calls, otherwise, external symbols are BRASL calls. This assumes
1371 // that declared/defined symbols are in the same compilation unit and can
1372 // be reached through PC-relative jumps.
1375 // This may be an unsafe assumption for JIT and really large compilation
1377 if (GV->isDeclaration()) {
1378 Callee = DAG.getNode(SPUISD::AFormAddr, dl, CalleeVT, GA, Zero);
1380 Callee = DAG.getNode(SPUISD::PCRelAddr, dl, CalleeVT, GA, Zero);
1383 // "Large memory" mode: Turn all calls into indirect calls with a X-form
1385 Callee = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, GA, Zero);
1387 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1388 EVT CalleeVT = Callee.getValueType();
1389 SDValue Zero = DAG.getConstant(0, PtrVT);
1390 SDValue ExtSym = DAG.getTargetExternalSymbol(S->getSymbol(),
1391 Callee.getValueType());
1393 if (!ST->usingLargeMem()) {
1394 Callee = DAG.getNode(SPUISD::AFormAddr, dl, CalleeVT, ExtSym, Zero);
1396 Callee = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, ExtSym, Zero);
1398 } else if (SDNode *Dest = isLSAAddress(Callee, DAG)) {
1399 // If this is an absolute destination address that appears to be a legal
1400 // local store address, use the munged value.
1401 Callee = SDValue(Dest, 0);
1404 Ops.push_back(Chain);
1405 Ops.push_back(Callee);
1407 // Add argument registers to the end of the list so that they are known live
1409 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1410 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1411 RegsToPass[i].second.getValueType()));
1413 if (InFlag.getNode())
1414 Ops.push_back(InFlag);
1415 // Returns a chain and a flag for retval copy to use.
1416 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag),
1417 &Ops[0], Ops.size());
1418 InFlag = Chain.getValue(1);
1420 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumStackBytes, true),
1421 DAG.getIntPtrConstant(0, true), InFlag);
1423 InFlag = Chain.getValue(1);
1425 // If the function returns void, just return the chain.
1429 // Now handle the return value(s)
1430 SmallVector<CCValAssign, 16> RVLocs;
1431 CCState CCRetInfo(CallConv, isVarArg, getTargetMachine(),
1432 RVLocs, *DAG.getContext());
1433 CCRetInfo.AnalyzeCallResult(Ins, CCC_SPU);
1436 // If the call has results, copy the values out of the ret val registers.
1437 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1438 CCValAssign VA = RVLocs[i];
1440 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1442 Chain = Val.getValue(1);
1443 InFlag = Val.getValue(2);
1444 InVals.push_back(Val);
1451 SPUTargetLowering::LowerReturn(SDValue Chain,
1452 CallingConv::ID CallConv, bool isVarArg,
1453 const SmallVectorImpl<ISD::OutputArg> &Outs,
1454 const SmallVectorImpl<SDValue> &OutVals,
1455 DebugLoc dl, SelectionDAG &DAG) const {
1457 SmallVector<CCValAssign, 16> RVLocs;
1458 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1459 RVLocs, *DAG.getContext());
1460 CCInfo.AnalyzeReturn(Outs, RetCC_SPU);
1462 // If this is the first return lowered for this function, add the regs to the
1463 // liveout set for the function.
1464 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1465 for (unsigned i = 0; i != RVLocs.size(); ++i)
1466 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1471 // Copy the result values into the output registers.
1472 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1473 CCValAssign &VA = RVLocs[i];
1474 assert(VA.isRegLoc() && "Can only return in registers!");
1475 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1477 Flag = Chain.getValue(1);
1481 return DAG.getNode(SPUISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1483 return DAG.getNode(SPUISD::RET_FLAG, dl, MVT::Other, Chain);
1487 //===----------------------------------------------------------------------===//
1488 // Vector related lowering:
1489 //===----------------------------------------------------------------------===//
1491 static ConstantSDNode *
1492 getVecImm(SDNode *N) {
1493 SDValue OpVal(0, 0);
1495 // Check to see if this buildvec has a single non-undef value in its elements.
1496 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1497 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1498 if (OpVal.getNode() == 0)
1499 OpVal = N->getOperand(i);
1500 else if (OpVal != N->getOperand(i))
1504 if (OpVal.getNode() != 0) {
1505 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1513 /// get_vec_i18imm - Test if this vector is a vector filled with the same value
1514 /// and the value fits into an unsigned 18-bit constant, and if so, return the
1516 SDValue SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
1518 if (ConstantSDNode *CN = getVecImm(N)) {
1519 uint64_t Value = CN->getZExtValue();
1520 if (ValueType == MVT::i64) {
1521 uint64_t UValue = CN->getZExtValue();
1522 uint32_t upper = uint32_t(UValue >> 32);
1523 uint32_t lower = uint32_t(UValue);
1526 Value = Value >> 32;
1528 if (Value <= 0x3ffff)
1529 return DAG.getTargetConstant(Value, ValueType);
1535 /// get_vec_i16imm - Test if this vector is a vector filled with the same value
1536 /// and the value fits into a signed 16-bit constant, and if so, return the
1538 SDValue SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
1540 if (ConstantSDNode *CN = getVecImm(N)) {
1541 int64_t Value = CN->getSExtValue();
1542 if (ValueType == MVT::i64) {
1543 uint64_t UValue = CN->getZExtValue();
1544 uint32_t upper = uint32_t(UValue >> 32);
1545 uint32_t lower = uint32_t(UValue);
1548 Value = Value >> 32;
1550 if (Value >= -(1 << 15) && Value <= ((1 << 15) - 1)) {
1551 return DAG.getTargetConstant(Value, ValueType);
1558 /// get_vec_i10imm - Test if this vector is a vector filled with the same value
1559 /// and the value fits into a signed 10-bit constant, and if so, return the
1561 SDValue SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
1563 if (ConstantSDNode *CN = getVecImm(N)) {
1564 int64_t Value = CN->getSExtValue();
1565 if (ValueType == MVT::i64) {
1566 uint64_t UValue = CN->getZExtValue();
1567 uint32_t upper = uint32_t(UValue >> 32);
1568 uint32_t lower = uint32_t(UValue);
1571 Value = Value >> 32;
1573 if (isInt<10>(Value))
1574 return DAG.getTargetConstant(Value, ValueType);
1580 /// get_vec_i8imm - Test if this vector is a vector filled with the same value
1581 /// and the value fits into a signed 8-bit constant, and if so, return the
1584 /// @note: The incoming vector is v16i8 because that's the only way we can load
1585 /// constant vectors. Thus, we test to see if the upper and lower bytes are the
1587 SDValue SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
1589 if (ConstantSDNode *CN = getVecImm(N)) {
1590 int Value = (int) CN->getZExtValue();
1591 if (ValueType == MVT::i16
1592 && Value <= 0xffff /* truncated from uint64_t */
1593 && ((short) Value >> 8) == ((short) Value & 0xff))
1594 return DAG.getTargetConstant(Value & 0xff, ValueType);
1595 else if (ValueType == MVT::i8
1596 && (Value & 0xff) == Value)
1597 return DAG.getTargetConstant(Value, ValueType);
1603 /// get_ILHUvec_imm - Test if this vector is a vector filled with the same value
1604 /// and the value fits into a signed 16-bit constant, and if so, return the
1606 SDValue SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
1608 if (ConstantSDNode *CN = getVecImm(N)) {
1609 uint64_t Value = CN->getZExtValue();
1610 if ((ValueType == MVT::i32
1611 && ((unsigned) Value & 0xffff0000) == (unsigned) Value)
1612 || (ValueType == MVT::i64 && (Value & 0xffff0000) == Value))
1613 return DAG.getTargetConstant(Value >> 16, ValueType);
1619 /// get_v4i32_imm - Catch-all for general 32-bit constant vectors
1620 SDValue SPU::get_v4i32_imm(SDNode *N, SelectionDAG &DAG) {
1621 if (ConstantSDNode *CN = getVecImm(N)) {
1622 return DAG.getTargetConstant((unsigned) CN->getZExtValue(), MVT::i32);
1628 /// get_v4i32_imm - Catch-all for general 64-bit constant vectors
1629 SDValue SPU::get_v2i64_imm(SDNode *N, SelectionDAG &DAG) {
1630 if (ConstantSDNode *CN = getVecImm(N)) {
1631 return DAG.getTargetConstant((unsigned) CN->getZExtValue(), MVT::i64);
1637 //! Lower a BUILD_VECTOR instruction creatively:
1639 LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
1640 EVT VT = Op.getValueType();
1641 EVT EltVT = VT.getVectorElementType();
1642 DebugLoc dl = Op.getDebugLoc();
1643 BuildVectorSDNode *BCN = dyn_cast<BuildVectorSDNode>(Op.getNode());
1644 assert(BCN != 0 && "Expected BuildVectorSDNode in SPU LowerBUILD_VECTOR");
1645 unsigned minSplatBits = EltVT.getSizeInBits();
1647 if (minSplatBits < 16)
1650 APInt APSplatBits, APSplatUndef;
1651 unsigned SplatBitSize;
1654 if (!BCN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
1655 HasAnyUndefs, minSplatBits)
1656 || minSplatBits < SplatBitSize)
1657 return SDValue(); // Wasn't a constant vector or splat exceeded min
1659 uint64_t SplatBits = APSplatBits.getZExtValue();
1661 switch (VT.getSimpleVT().SimpleTy) {
1663 report_fatal_error("CellSPU: Unhandled VT in LowerBUILD_VECTOR, VT = " +
1664 Twine(VT.getEVTString()));
1667 uint32_t Value32 = uint32_t(SplatBits);
1668 assert(SplatBitSize == 32
1669 && "LowerBUILD_VECTOR: Unexpected floating point vector element.");
1670 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1671 SDValue T = DAG.getConstant(Value32, MVT::i32);
1672 return DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,
1673 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, T,T,T,T));
1677 uint64_t f64val = uint64_t(SplatBits);
1678 assert(SplatBitSize == 64
1679 && "LowerBUILD_VECTOR: 64-bit float vector size > 8 bytes.");
1680 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1681 SDValue T = DAG.getConstant(f64val, MVT::i64);
1682 return DAG.getNode(ISD::BITCAST, dl, MVT::v2f64,
1683 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, T, T));
1687 // 8-bit constants have to be expanded to 16-bits
1688 unsigned short Value16 = SplatBits /* | (SplatBits << 8) */;
1689 SmallVector<SDValue, 8> Ops;
1691 Ops.assign(8, DAG.getConstant(Value16, MVT::i16));
1692 return DAG.getNode(ISD::BITCAST, dl, VT,
1693 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i16, &Ops[0], Ops.size()));
1696 unsigned short Value16 = SplatBits;
1697 SDValue T = DAG.getConstant(Value16, EltVT);
1698 SmallVector<SDValue, 8> Ops;
1701 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
1704 SDValue T = DAG.getConstant(unsigned(SplatBits), VT.getVectorElementType());
1705 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, T, T, T, T);
1708 return SPU::LowerV2I64Splat(VT, DAG, SplatBits, dl);
1718 SPU::LowerV2I64Splat(EVT OpVT, SelectionDAG& DAG, uint64_t SplatVal,
1720 uint32_t upper = uint32_t(SplatVal >> 32);
1721 uint32_t lower = uint32_t(SplatVal);
1723 if (upper == lower) {
1724 // Magic constant that can be matched by IL, ILA, et. al.
1725 SDValue Val = DAG.getTargetConstant(upper, MVT::i32);
1726 return DAG.getNode(ISD::BITCAST, dl, OpVT,
1727 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1728 Val, Val, Val, Val));
1730 bool upper_special, lower_special;
1732 // NOTE: This code creates common-case shuffle masks that can be easily
1733 // detected as common expressions. It is not attempting to create highly
1734 // specialized masks to replace any and all 0's, 0xff's and 0x80's.
1736 // Detect if the upper or lower half is a special shuffle mask pattern:
1737 upper_special = (upper == 0 || upper == 0xffffffff || upper == 0x80000000);
1738 lower_special = (lower == 0 || lower == 0xffffffff || lower == 0x80000000);
1740 // Both upper and lower are special, lower to a constant pool load:
1741 if (lower_special && upper_special) {
1742 SDValue SplatValCN = DAG.getConstant(SplatVal, MVT::i64);
1743 return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64,
1744 SplatValCN, SplatValCN);
1749 SmallVector<SDValue, 16> ShufBytes;
1752 // Create lower vector if not a special pattern
1753 if (!lower_special) {
1754 SDValue LO32C = DAG.getConstant(lower, MVT::i32);
1755 LO32 = DAG.getNode(ISD::BITCAST, dl, OpVT,
1756 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1757 LO32C, LO32C, LO32C, LO32C));
1760 // Create upper vector if not a special pattern
1761 if (!upper_special) {
1762 SDValue HI32C = DAG.getConstant(upper, MVT::i32);
1763 HI32 = DAG.getNode(ISD::BITCAST, dl, OpVT,
1764 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1765 HI32C, HI32C, HI32C, HI32C));
1768 // If either upper or lower are special, then the two input operands are
1769 // the same (basically, one of them is a "don't care")
1775 for (int i = 0; i < 4; ++i) {
1777 for (int j = 0; j < 4; ++j) {
1779 bool process_upper, process_lower;
1781 process_upper = (upper_special && (i & 1) == 0);
1782 process_lower = (lower_special && (i & 1) == 1);
1784 if (process_upper || process_lower) {
1785 if ((process_upper && upper == 0)
1786 || (process_lower && lower == 0))
1788 else if ((process_upper && upper == 0xffffffff)
1789 || (process_lower && lower == 0xffffffff))
1791 else if ((process_upper && upper == 0x80000000)
1792 || (process_lower && lower == 0x80000000))
1793 val |= (j == 0 ? 0xe0 : 0x80);
1795 val |= i * 4 + j + ((i & 1) * 16);
1798 ShufBytes.push_back(DAG.getConstant(val, MVT::i32));
1801 return DAG.getNode(SPUISD::SHUFB, dl, OpVT, HI32, LO32,
1802 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1803 &ShufBytes[0], ShufBytes.size()));
1807 /// LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3) to something on
1808 /// which the Cell can operate. The code inspects V3 to ascertain whether the
1809 /// permutation vector, V3, is monotonically increasing with one "exception"
1810 /// element, e.g., (0, 1, _, 3). If this is the case, then generate a
1811 /// SHUFFLE_MASK synthetic instruction. Otherwise, spill V3 to the constant pool.
1812 /// In either case, the net result is going to eventually invoke SHUFB to
1813 /// permute/shuffle the bytes from V1 and V2.
1815 /// SHUFFLE_MASK is eventually selected as one of the C*D instructions, generate
1816 /// control word for byte/halfword/word insertion. This takes care of a single
1817 /// element move from V2 into V1.
1819 /// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
1820 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
1821 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
1822 SDValue V1 = Op.getOperand(0);
1823 SDValue V2 = Op.getOperand(1);
1824 DebugLoc dl = Op.getDebugLoc();
1826 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1828 // If we have a single element being moved from V1 to V2, this can be handled
1829 // using the C*[DX] compute mask instructions, but the vector elements have
1830 // to be monotonically increasing with one exception element, and the source
1831 // slot of the element to move must be the same as the destination.
1832 EVT VecVT = V1.getValueType();
1833 EVT EltVT = VecVT.getVectorElementType();
1834 unsigned EltsFromV2 = 0;
1835 unsigned V2EltOffset = 0;
1836 unsigned V2EltIdx0 = 0;
1837 unsigned CurrElt = 0;
1838 unsigned MaxElts = VecVT.getVectorNumElements();
1839 unsigned PrevElt = 0;
1840 bool monotonic = true;
1843 EVT maskVT; // which of the c?d instructions to use
1845 if (EltVT == MVT::i8) {
1847 maskVT = MVT::v16i8;
1848 } else if (EltVT == MVT::i16) {
1850 maskVT = MVT::v8i16;
1851 } else if (EltVT == MVT::i32 || EltVT == MVT::f32) {
1853 maskVT = MVT::v4i32;
1854 } else if (EltVT == MVT::i64 || EltVT == MVT::f64) {
1856 maskVT = MVT::v2i64;
1858 llvm_unreachable("Unhandled vector type in LowerVECTOR_SHUFFLE");
1860 for (unsigned i = 0; i != MaxElts; ++i) {
1861 if (SVN->getMaskElt(i) < 0)
1864 unsigned SrcElt = SVN->getMaskElt(i);
1867 if (SrcElt >= V2EltIdx0) {
1868 // TODO: optimize for the monotonic case when several consecutive
1869 // elements are taken form V2. Do we ever get such a case?
1870 if (EltsFromV2 == 0 && CurrElt == (SrcElt - V2EltIdx0))
1871 V2EltOffset = (SrcElt - V2EltIdx0) * (EltVT.getSizeInBits()/8);
1875 } else if (CurrElt != SrcElt) {
1883 if (PrevElt > 0 && SrcElt < MaxElts) {
1884 if ((PrevElt == SrcElt - 1)
1885 || (PrevElt == MaxElts - 1 && SrcElt == 0)) {
1890 } else if (i == 0 || (PrevElt==0 && SrcElt==1)) {
1891 // First time or after a "wrap around"
1895 // This isn't a rotation, takes elements from vector 2
1901 if (EltsFromV2 == 1 && monotonic) {
1902 // Compute mask and shuffle
1903 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1905 // As SHUFFLE_MASK becomes a c?d instruction, feed it an address
1906 // R1 ($sp) is used here only as it is guaranteed to have last bits zero
1907 SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
1908 DAG.getRegister(SPU::R1, PtrVT),
1909 DAG.getConstant(V2EltOffset, MVT::i32));
1910 SDValue ShufMaskOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl,
1913 // Use shuffle mask in SHUFB synthetic instruction:
1914 return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V2, V1,
1916 } else if (rotate) {
1919 rotamt *= EltVT.getSizeInBits()/8;
1920 return DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, V1.getValueType(),
1921 V1, DAG.getConstant(rotamt, MVT::i16));
1923 // Convert the SHUFFLE_VECTOR mask's input element units to the
1925 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
1927 SmallVector<SDValue, 16> ResultMask;
1928 for (unsigned i = 0, e = MaxElts; i != e; ++i) {
1929 unsigned SrcElt = SVN->getMaskElt(i) < 0 ? 0 : SVN->getMaskElt(i);
1931 for (unsigned j = 0; j < BytesPerElement; ++j)
1932 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,MVT::i8));
1934 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
1935 &ResultMask[0], ResultMask.size());
1936 return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V1, V2, VPermMask);
1940 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
1941 SDValue Op0 = Op.getOperand(0); // Op0 = the scalar
1942 DebugLoc dl = Op.getDebugLoc();
1944 if (Op0.getNode()->getOpcode() == ISD::Constant) {
1945 // For a constant, build the appropriate constant vector, which will
1946 // eventually simplify to a vector register load.
1948 ConstantSDNode *CN = cast<ConstantSDNode>(Op0.getNode());
1949 SmallVector<SDValue, 16> ConstVecValues;
1953 // Create a constant vector:
1954 switch (Op.getValueType().getSimpleVT().SimpleTy) {
1955 default: llvm_unreachable("Unexpected constant value type in "
1956 "LowerSCALAR_TO_VECTOR");
1957 case MVT::v16i8: n_copies = 16; VT = MVT::i8; break;
1958 case MVT::v8i16: n_copies = 8; VT = MVT::i16; break;
1959 case MVT::v4i32: n_copies = 4; VT = MVT::i32; break;
1960 case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
1961 case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
1962 case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
1965 SDValue CValue = DAG.getConstant(CN->getZExtValue(), VT);
1966 for (size_t j = 0; j < n_copies; ++j)
1967 ConstVecValues.push_back(CValue);
1969 return DAG.getNode(ISD::BUILD_VECTOR, dl, Op.getValueType(),
1970 &ConstVecValues[0], ConstVecValues.size());
1972 // Otherwise, copy the value from one register to another:
1973 switch (Op0.getValueType().getSimpleVT().SimpleTy) {
1974 default: llvm_unreachable("Unexpected value type in LowerSCALAR_TO_VECTOR");
1981 return DAG.getNode(SPUISD::PREFSLOT2VEC, dl, Op.getValueType(), Op0, Op0);
1988 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
1989 EVT VT = Op.getValueType();
1990 SDValue N = Op.getOperand(0);
1991 SDValue Elt = Op.getOperand(1);
1992 DebugLoc dl = Op.getDebugLoc();
1995 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1996 // Constant argument:
1997 int EltNo = (int) C->getZExtValue();
2000 if (VT == MVT::i8 && EltNo >= 16)
2001 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i8 extraction slot > 15");
2002 else if (VT == MVT::i16 && EltNo >= 8)
2003 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i16 extraction slot > 7");
2004 else if (VT == MVT::i32 && EltNo >= 4)
2005 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i32 extraction slot > 4");
2006 else if (VT == MVT::i64 && EltNo >= 2)
2007 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i64 extraction slot > 2");
2009 if (EltNo == 0 && (VT == MVT::i32 || VT == MVT::i64)) {
2010 // i32 and i64: Element 0 is the preferred slot
2011 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT, N);
2014 // Need to generate shuffle mask and extract:
2015 int prefslot_begin = -1, prefslot_end = -1;
2016 int elt_byte = EltNo * VT.getSizeInBits() / 8;
2018 switch (VT.getSimpleVT().SimpleTy) {
2020 assert(false && "Invalid value type!");
2022 prefslot_begin = prefslot_end = 3;
2026 prefslot_begin = 2; prefslot_end = 3;
2031 prefslot_begin = 0; prefslot_end = 3;
2036 prefslot_begin = 0; prefslot_end = 7;
2041 assert(prefslot_begin != -1 && prefslot_end != -1 &&
2042 "LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
2044 unsigned int ShufBytes[16] = {
2045 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
2047 for (int i = 0; i < 16; ++i) {
2048 // zero fill uppper part of preferred slot, don't care about the
2050 unsigned int mask_val;
2051 if (i <= prefslot_end) {
2053 ((i < prefslot_begin)
2055 : elt_byte + (i - prefslot_begin));
2057 ShufBytes[i] = mask_val;
2059 ShufBytes[i] = ShufBytes[i % (prefslot_end + 1)];
2062 SDValue ShufMask[4];
2063 for (unsigned i = 0; i < sizeof(ShufMask)/sizeof(ShufMask[0]); ++i) {
2064 unsigned bidx = i * 4;
2065 unsigned int bits = ((ShufBytes[bidx] << 24) |
2066 (ShufBytes[bidx+1] << 16) |
2067 (ShufBytes[bidx+2] << 8) |
2069 ShufMask[i] = DAG.getConstant(bits, MVT::i32);
2072 SDValue ShufMaskVec =
2073 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2074 &ShufMask[0], sizeof(ShufMask)/sizeof(ShufMask[0]));
2076 retval = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
2077 DAG.getNode(SPUISD::SHUFB, dl, N.getValueType(),
2078 N, N, ShufMaskVec));
2080 // Variable index: Rotate the requested element into slot 0, then replicate
2081 // slot 0 across the vector
2082 EVT VecVT = N.getValueType();
2083 if (!VecVT.isSimple() || !VecVT.isVector()) {
2084 report_fatal_error("LowerEXTRACT_VECTOR_ELT: Must have a simple, 128-bit"
2088 // Make life easier by making sure the index is zero-extended to i32
2089 if (Elt.getValueType() != MVT::i32)
2090 Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Elt);
2092 // Scale the index to a bit/byte shift quantity
2094 APInt(32, uint64_t(16 / N.getValueType().getVectorNumElements()), false);
2095 unsigned scaleShift = scaleFactor.logBase2();
2098 if (scaleShift > 0) {
2099 // Scale the shift factor:
2100 Elt = DAG.getNode(ISD::SHL, dl, MVT::i32, Elt,
2101 DAG.getConstant(scaleShift, MVT::i32));
2104 vecShift = DAG.getNode(SPUISD::SHL_BYTES, dl, VecVT, N, Elt);
2106 // Replicate the bytes starting at byte 0 across the entire vector (for
2107 // consistency with the notion of a unified register set)
2110 switch (VT.getSimpleVT().SimpleTy) {
2112 report_fatal_error("LowerEXTRACT_VECTOR_ELT(varable): Unhandled vector"
2116 SDValue factor = DAG.getConstant(0x00000000, MVT::i32);
2117 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2118 factor, factor, factor, factor);
2122 SDValue factor = DAG.getConstant(0x00010001, MVT::i32);
2123 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2124 factor, factor, factor, factor);
2129 SDValue factor = DAG.getConstant(0x00010203, MVT::i32);
2130 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2131 factor, factor, factor, factor);
2136 SDValue loFactor = DAG.getConstant(0x00010203, MVT::i32);
2137 SDValue hiFactor = DAG.getConstant(0x04050607, MVT::i32);
2138 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2139 loFactor, hiFactor, loFactor, hiFactor);
2144 retval = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
2145 DAG.getNode(SPUISD::SHUFB, dl, VecVT,
2146 vecShift, vecShift, replicate));
2152 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2153 SDValue VecOp = Op.getOperand(0);
2154 SDValue ValOp = Op.getOperand(1);
2155 SDValue IdxOp = Op.getOperand(2);
2156 DebugLoc dl = Op.getDebugLoc();
2157 EVT VT = Op.getValueType();
2158 EVT eltVT = ValOp.getValueType();
2160 // use 0 when the lane to insert to is 'undef'
2162 if (IdxOp.getOpcode() != ISD::UNDEF) {
2163 ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
2164 assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
2165 Offset = (CN->getSExtValue()) * eltVT.getSizeInBits()/8;
2168 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2169 // Use $sp ($1) because it's always 16-byte aligned and it's available:
2170 SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
2171 DAG.getRegister(SPU::R1, PtrVT),
2172 DAG.getConstant(Offset, PtrVT));
2173 // widen the mask when dealing with half vectors
2174 EVT maskVT = EVT::getVectorVT(*(DAG.getContext()), VT.getVectorElementType(),
2175 128/ VT.getVectorElementType().getSizeInBits());
2176 SDValue ShufMask = DAG.getNode(SPUISD::SHUFFLE_MASK, dl, maskVT, Pointer);
2179 DAG.getNode(SPUISD::SHUFB, dl, VT,
2180 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, ValOp),
2182 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, ShufMask));
2187 static SDValue LowerI8Math(SDValue Op, SelectionDAG &DAG, unsigned Opc,
2188 const TargetLowering &TLI)
2190 SDValue N0 = Op.getOperand(0); // Everything has at least one operand
2191 DebugLoc dl = Op.getDebugLoc();
2192 EVT ShiftVT = TLI.getShiftAmountTy();
2194 assert(Op.getValueType() == MVT::i8);
2197 llvm_unreachable("Unhandled i8 math operator");
2201 // 8-bit addition: Promote the arguments up to 16-bits and truncate
2203 SDValue N1 = Op.getOperand(1);
2204 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2205 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N1);
2206 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2207 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2212 // 8-bit subtraction: Promote the arguments up to 16-bits and truncate
2214 SDValue N1 = Op.getOperand(1);
2215 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2216 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N1);
2217 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2218 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2222 SDValue N1 = Op.getOperand(1);
2223 EVT N1VT = N1.getValueType();
2225 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, N0);
2226 if (!N1VT.bitsEq(ShiftVT)) {
2227 unsigned N1Opc = N1.getValueType().bitsLT(ShiftVT)
2230 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2233 // Replicate lower 8-bits into upper 8:
2235 DAG.getNode(ISD::OR, dl, MVT::i16, N0,
2236 DAG.getNode(ISD::SHL, dl, MVT::i16,
2237 N0, DAG.getConstant(8, MVT::i32)));
2239 // Truncate back down to i8
2240 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2241 DAG.getNode(Opc, dl, MVT::i16, ExpandArg, N1));
2245 SDValue N1 = Op.getOperand(1);
2246 EVT N1VT = N1.getValueType();
2248 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, N0);
2249 if (!N1VT.bitsEq(ShiftVT)) {
2250 unsigned N1Opc = ISD::ZERO_EXTEND;
2252 if (N1.getValueType().bitsGT(ShiftVT))
2253 N1Opc = ISD::TRUNCATE;
2255 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2258 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2259 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2262 SDValue N1 = Op.getOperand(1);
2263 EVT N1VT = N1.getValueType();
2265 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2266 if (!N1VT.bitsEq(ShiftVT)) {
2267 unsigned N1Opc = ISD::SIGN_EXTEND;
2269 if (N1VT.bitsGT(ShiftVT))
2270 N1Opc = ISD::TRUNCATE;
2271 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2274 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2275 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2278 SDValue N1 = Op.getOperand(1);
2280 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2281 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N1);
2282 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2283 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2291 //! Lower byte immediate operations for v16i8 vectors:
2293 LowerByteImmed(SDValue Op, SelectionDAG &DAG) {
2296 EVT VT = Op.getValueType();
2297 DebugLoc dl = Op.getDebugLoc();
2299 ConstVec = Op.getOperand(0);
2300 Arg = Op.getOperand(1);
2301 if (ConstVec.getNode()->getOpcode() != ISD::BUILD_VECTOR) {
2302 if (ConstVec.getNode()->getOpcode() == ISD::BITCAST) {
2303 ConstVec = ConstVec.getOperand(0);
2305 ConstVec = Op.getOperand(1);
2306 Arg = Op.getOperand(0);
2307 if (ConstVec.getNode()->getOpcode() == ISD::BITCAST) {
2308 ConstVec = ConstVec.getOperand(0);
2313 if (ConstVec.getNode()->getOpcode() == ISD::BUILD_VECTOR) {
2314 BuildVectorSDNode *BCN = dyn_cast<BuildVectorSDNode>(ConstVec.getNode());
2315 assert(BCN != 0 && "Expected BuildVectorSDNode in SPU LowerByteImmed");
2317 APInt APSplatBits, APSplatUndef;
2318 unsigned SplatBitSize;
2320 unsigned minSplatBits = VT.getVectorElementType().getSizeInBits();
2322 if (BCN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
2323 HasAnyUndefs, minSplatBits)
2324 && minSplatBits <= SplatBitSize) {
2325 uint64_t SplatBits = APSplatBits.getZExtValue();
2326 SDValue tc = DAG.getTargetConstant(SplatBits & 0xff, MVT::i8);
2328 SmallVector<SDValue, 16> tcVec;
2329 tcVec.assign(16, tc);
2330 return DAG.getNode(Op.getNode()->getOpcode(), dl, VT, Arg,
2331 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &tcVec[0], tcVec.size()));
2335 // These operations (AND, OR, XOR) are legal, they just couldn't be custom
2336 // lowered. Return the operation, rather than a null SDValue.
2340 //! Custom lowering for CTPOP (count population)
2342 Custom lowering code that counts the number ones in the input
2343 operand. SPU has such an instruction, but it counts the number of
2344 ones per byte, which then have to be accumulated.
2346 static SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) {
2347 EVT VT = Op.getValueType();
2348 EVT vecVT = EVT::getVectorVT(*DAG.getContext(),
2349 VT, (128 / VT.getSizeInBits()));
2350 DebugLoc dl = Op.getDebugLoc();
2352 switch (VT.getSimpleVT().SimpleTy) {
2354 assert(false && "Invalid value type!");
2356 SDValue N = Op.getOperand(0);
2357 SDValue Elt0 = DAG.getConstant(0, MVT::i32);
2359 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2360 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2362 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i8, CNTB, Elt0);
2366 MachineFunction &MF = DAG.getMachineFunction();
2367 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2369 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
2371 SDValue N = Op.getOperand(0);
2372 SDValue Elt0 = DAG.getConstant(0, MVT::i16);
2373 SDValue Mask0 = DAG.getConstant(0x0f, MVT::i16);
2374 SDValue Shift1 = DAG.getConstant(8, MVT::i32);
2376 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2377 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2379 // CNTB_result becomes the chain to which all of the virtual registers
2380 // CNTB_reg, SUM1_reg become associated:
2381 SDValue CNTB_result =
2382 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, CNTB, Elt0);
2384 SDValue CNTB_rescopy =
2385 DAG.getCopyToReg(CNTB_result, dl, CNTB_reg, CNTB_result);
2387 SDValue Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i16);
2389 return DAG.getNode(ISD::AND, dl, MVT::i16,
2390 DAG.getNode(ISD::ADD, dl, MVT::i16,
2391 DAG.getNode(ISD::SRL, dl, MVT::i16,
2398 MachineFunction &MF = DAG.getMachineFunction();
2399 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2401 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2402 unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2404 SDValue N = Op.getOperand(0);
2405 SDValue Elt0 = DAG.getConstant(0, MVT::i32);
2406 SDValue Mask0 = DAG.getConstant(0xff, MVT::i32);
2407 SDValue Shift1 = DAG.getConstant(16, MVT::i32);
2408 SDValue Shift2 = DAG.getConstant(8, MVT::i32);
2410 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2411 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2413 // CNTB_result becomes the chain to which all of the virtual registers
2414 // CNTB_reg, SUM1_reg become associated:
2415 SDValue CNTB_result =
2416 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, CNTB, Elt0);
2418 SDValue CNTB_rescopy =
2419 DAG.getCopyToReg(CNTB_result, dl, CNTB_reg, CNTB_result);
2422 DAG.getNode(ISD::SRL, dl, MVT::i32,
2423 DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i32),
2427 DAG.getNode(ISD::ADD, dl, MVT::i32, Comp1,
2428 DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i32));
2430 SDValue Sum1_rescopy =
2431 DAG.getCopyToReg(CNTB_result, dl, SUM1_reg, Sum1);
2434 DAG.getNode(ISD::SRL, dl, MVT::i32,
2435 DAG.getCopyFromReg(Sum1_rescopy, dl, SUM1_reg, MVT::i32),
2438 DAG.getNode(ISD::ADD, dl, MVT::i32, Comp2,
2439 DAG.getCopyFromReg(Sum1_rescopy, dl, SUM1_reg, MVT::i32));
2441 return DAG.getNode(ISD::AND, dl, MVT::i32, Sum2, Mask0);
2451 //! Lower ISD::FP_TO_SINT, ISD::FP_TO_UINT for i32
2453 f32->i32 passes through unchanged, whereas f64->i32 expands to a libcall.
2454 All conversions to i64 are expanded to a libcall.
2456 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2457 const SPUTargetLowering &TLI) {
2458 EVT OpVT = Op.getValueType();
2459 SDValue Op0 = Op.getOperand(0);
2460 EVT Op0VT = Op0.getValueType();
2462 if ((OpVT == MVT::i32 && Op0VT == MVT::f64)
2463 || OpVT == MVT::i64) {
2464 // Convert f32 / f64 to i32 / i64 via libcall.
2466 (Op.getOpcode() == ISD::FP_TO_SINT)
2467 ? RTLIB::getFPTOSINT(Op0VT, OpVT)
2468 : RTLIB::getFPTOUINT(Op0VT, OpVT);
2469 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpectd fp-to-int conversion!");
2471 return ExpandLibCall(LC, Op, DAG, false, Dummy, TLI);
2477 //! Lower ISD::SINT_TO_FP, ISD::UINT_TO_FP for i32
2479 i32->f32 passes through unchanged, whereas i32->f64 is expanded to a libcall.
2480 All conversions from i64 are expanded to a libcall.
2482 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2483 const SPUTargetLowering &TLI) {
2484 EVT OpVT = Op.getValueType();
2485 SDValue Op0 = Op.getOperand(0);
2486 EVT Op0VT = Op0.getValueType();
2488 if ((OpVT == MVT::f64 && Op0VT == MVT::i32)
2489 || Op0VT == MVT::i64) {
2490 // Convert i32, i64 to f64 via libcall:
2492 (Op.getOpcode() == ISD::SINT_TO_FP)
2493 ? RTLIB::getSINTTOFP(Op0VT, OpVT)
2494 : RTLIB::getUINTTOFP(Op0VT, OpVT);
2495 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpectd int-to-fp conversion!");
2497 return ExpandLibCall(LC, Op, DAG, false, Dummy, TLI);
2503 //! Lower ISD::SETCC
2505 This handles MVT::f64 (double floating point) condition lowering
2507 static SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG,
2508 const TargetLowering &TLI) {
2509 CondCodeSDNode *CC = dyn_cast<CondCodeSDNode>(Op.getOperand(2));
2510 DebugLoc dl = Op.getDebugLoc();
2511 assert(CC != 0 && "LowerSETCC: CondCodeSDNode should not be null here!\n");
2513 SDValue lhs = Op.getOperand(0);
2514 SDValue rhs = Op.getOperand(1);
2515 EVT lhsVT = lhs.getValueType();
2516 assert(lhsVT == MVT::f64 && "LowerSETCC: type other than MVT::64\n");
2518 EVT ccResultVT = TLI.getSetCCResultType(lhs.getValueType());
2519 APInt ccResultOnes = APInt::getAllOnesValue(ccResultVT.getSizeInBits());
2520 EVT IntVT(MVT::i64);
2522 // Take advantage of the fact that (truncate (sra arg, 32)) is efficiently
2523 // selected to a NOP:
2524 SDValue i64lhs = DAG.getNode(ISD::BITCAST, dl, IntVT, lhs);
2526 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
2527 DAG.getNode(ISD::SRL, dl, IntVT,
2528 i64lhs, DAG.getConstant(32, MVT::i32)));
2529 SDValue lhsHi32abs =
2530 DAG.getNode(ISD::AND, dl, MVT::i32,
2531 lhsHi32, DAG.getConstant(0x7fffffff, MVT::i32));
2533 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, i64lhs);
2535 // SETO and SETUO only use the lhs operand:
2536 if (CC->get() == ISD::SETO) {
2537 // Evaluates to true if Op0 is not [SQ]NaN - lowers to the inverse of
2539 APInt ccResultAllOnes = APInt::getAllOnesValue(ccResultVT.getSizeInBits());
2540 return DAG.getNode(ISD::XOR, dl, ccResultVT,
2541 DAG.getSetCC(dl, ccResultVT,
2542 lhs, DAG.getConstantFP(0.0, lhsVT),
2544 DAG.getConstant(ccResultAllOnes, ccResultVT));
2545 } else if (CC->get() == ISD::SETUO) {
2546 // Evaluates to true if Op0 is [SQ]NaN
2547 return DAG.getNode(ISD::AND, dl, ccResultVT,
2548 DAG.getSetCC(dl, ccResultVT,
2550 DAG.getConstant(0x7ff00000, MVT::i32),
2552 DAG.getSetCC(dl, ccResultVT,
2554 DAG.getConstant(0, MVT::i32),
2558 SDValue i64rhs = DAG.getNode(ISD::BITCAST, dl, IntVT, rhs);
2560 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
2561 DAG.getNode(ISD::SRL, dl, IntVT,
2562 i64rhs, DAG.getConstant(32, MVT::i32)));
2564 // If a value is negative, subtract from the sign magnitude constant:
2565 SDValue signMag2TC = DAG.getConstant(0x8000000000000000ULL, IntVT);
2567 // Convert the sign-magnitude representation into 2's complement:
2568 SDValue lhsSelectMask = DAG.getNode(ISD::SRA, dl, ccResultVT,
2569 lhsHi32, DAG.getConstant(31, MVT::i32));
2570 SDValue lhsSignMag2TC = DAG.getNode(ISD::SUB, dl, IntVT, signMag2TC, i64lhs);
2572 DAG.getNode(ISD::SELECT, dl, IntVT,
2573 lhsSelectMask, lhsSignMag2TC, i64lhs);
2575 SDValue rhsSelectMask = DAG.getNode(ISD::SRA, dl, ccResultVT,
2576 rhsHi32, DAG.getConstant(31, MVT::i32));
2577 SDValue rhsSignMag2TC = DAG.getNode(ISD::SUB, dl, IntVT, signMag2TC, i64rhs);
2579 DAG.getNode(ISD::SELECT, dl, IntVT,
2580 rhsSelectMask, rhsSignMag2TC, i64rhs);
2584 switch (CC->get()) {
2587 compareOp = ISD::SETEQ; break;
2590 compareOp = ISD::SETGT; break;
2593 compareOp = ISD::SETGE; break;
2596 compareOp = ISD::SETLT; break;
2599 compareOp = ISD::SETLE; break;
2602 compareOp = ISD::SETNE; break;
2604 report_fatal_error("CellSPU ISel Select: unimplemented f64 condition");
2608 DAG.getSetCC(dl, ccResultVT, lhsSelect, rhsSelect,
2609 (ISD::CondCode) compareOp);
2611 if ((CC->get() & 0x8) == 0) {
2612 // Ordered comparison:
2613 SDValue lhsNaN = DAG.getSetCC(dl, ccResultVT,
2614 lhs, DAG.getConstantFP(0.0, MVT::f64),
2616 SDValue rhsNaN = DAG.getSetCC(dl, ccResultVT,
2617 rhs, DAG.getConstantFP(0.0, MVT::f64),
2619 SDValue ordered = DAG.getNode(ISD::AND, dl, ccResultVT, lhsNaN, rhsNaN);
2621 result = DAG.getNode(ISD::AND, dl, ccResultVT, ordered, result);
2627 //! Lower ISD::SELECT_CC
2629 ISD::SELECT_CC can (generally) be implemented directly on the SPU using the
2632 \note Need to revisit this in the future: if the code path through the true
2633 and false value computations is longer than the latency of a branch (6
2634 cycles), then it would be more advantageous to branch and insert a new basic
2635 block and branch on the condition. However, this code does not make that
2636 assumption, given the simplisitc uses so far.
2639 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
2640 const TargetLowering &TLI) {
2641 EVT VT = Op.getValueType();
2642 SDValue lhs = Op.getOperand(0);
2643 SDValue rhs = Op.getOperand(1);
2644 SDValue trueval = Op.getOperand(2);
2645 SDValue falseval = Op.getOperand(3);
2646 SDValue condition = Op.getOperand(4);
2647 DebugLoc dl = Op.getDebugLoc();
2649 // NOTE: SELB's arguments: $rA, $rB, $mask
2651 // SELB selects bits from $rA where bits in $mask are 0, bits from $rB
2652 // where bits in $mask are 1. CCond will be inverted, having 1s where the
2653 // condition was true and 0s where the condition was false. Hence, the
2654 // arguments to SELB get reversed.
2656 // Note: Really should be ISD::SELECT instead of SPUISD::SELB, but LLVM's
2657 // legalizer insists on combining SETCC/SELECT into SELECT_CC, so we end up
2658 // with another "cannot select select_cc" assert:
2660 SDValue compare = DAG.getNode(ISD::SETCC, dl,
2661 TLI.getSetCCResultType(Op.getValueType()),
2662 lhs, rhs, condition);
2663 return DAG.getNode(SPUISD::SELB, dl, VT, falseval, trueval, compare);
2666 //! Custom lower ISD::TRUNCATE
2667 static SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG)
2669 // Type to truncate to
2670 EVT VT = Op.getValueType();
2671 MVT simpleVT = VT.getSimpleVT();
2672 EVT VecVT = EVT::getVectorVT(*DAG.getContext(),
2673 VT, (128 / VT.getSizeInBits()));
2674 DebugLoc dl = Op.getDebugLoc();
2676 // Type to truncate from
2677 SDValue Op0 = Op.getOperand(0);
2678 EVT Op0VT = Op0.getValueType();
2680 if (Op0VT == MVT::i128 && simpleVT == MVT::i64) {
2681 // Create shuffle mask, least significant doubleword of quadword
2682 unsigned maskHigh = 0x08090a0b;
2683 unsigned maskLow = 0x0c0d0e0f;
2684 // Use a shuffle to perform the truncation
2685 SDValue shufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2686 DAG.getConstant(maskHigh, MVT::i32),
2687 DAG.getConstant(maskLow, MVT::i32),
2688 DAG.getConstant(maskHigh, MVT::i32),
2689 DAG.getConstant(maskLow, MVT::i32));
2691 SDValue truncShuffle = DAG.getNode(SPUISD::SHUFB, dl, VecVT,
2692 Op0, Op0, shufMask);
2694 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT, truncShuffle);
2697 return SDValue(); // Leave the truncate unmolested
2701 * Emit the instruction sequence for i64/i32 -> i128 sign extend. The basic
2702 * algorithm is to duplicate the sign bit using rotmai to generate at
2703 * least one byte full of sign bits. Then propagate the "sign-byte" into
2704 * the leftmost words and the i64/i32 into the rightmost words using shufb.
2706 * @param Op The sext operand
2707 * @param DAG The current DAG
2708 * @return The SDValue with the entire instruction sequence
2710 static SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG)
2712 DebugLoc dl = Op.getDebugLoc();
2714 // Type to extend to
2715 MVT OpVT = Op.getValueType().getSimpleVT();
2717 // Type to extend from
2718 SDValue Op0 = Op.getOperand(0);
2719 MVT Op0VT = Op0.getValueType().getSimpleVT();
2721 // The type to extend to needs to be a i128 and
2722 // the type to extend from needs to be i64 or i32.
2723 assert((OpVT == MVT::i128 && (Op0VT == MVT::i64 || Op0VT == MVT::i32)) &&
2724 "LowerSIGN_EXTEND: input and/or output operand have wrong size");
2726 // Create shuffle mask
2727 unsigned mask1 = 0x10101010; // byte 0 - 3 and 4 - 7
2728 unsigned mask2 = Op0VT == MVT::i64 ? 0x00010203 : 0x10101010; // byte 8 - 11
2729 unsigned mask3 = Op0VT == MVT::i64 ? 0x04050607 : 0x00010203; // byte 12 - 15
2730 SDValue shufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2731 DAG.getConstant(mask1, MVT::i32),
2732 DAG.getConstant(mask1, MVT::i32),
2733 DAG.getConstant(mask2, MVT::i32),
2734 DAG.getConstant(mask3, MVT::i32));
2736 // Word wise arithmetic right shift to generate at least one byte
2737 // that contains sign bits.
2738 MVT mvt = Op0VT == MVT::i64 ? MVT::v2i64 : MVT::v4i32;
2739 SDValue sraVal = DAG.getNode(ISD::SRA,
2742 DAG.getNode(SPUISD::PREFSLOT2VEC, dl, mvt, Op0, Op0),
2743 DAG.getConstant(31, MVT::i32));
2745 // reinterpret as a i128 (SHUFB requires it). This gets lowered away.
2746 SDValue extended = SDValue(DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
2748 DAG.getTargetConstant(
2749 SPU::GPRCRegClass.getID(),
2751 // Shuffle bytes - Copy the sign bits into the upper 64 bits
2752 // and the input value into the lower 64 bits.
2753 SDValue extShuffle = DAG.getNode(SPUISD::SHUFB, dl, mvt,
2754 extended, sraVal, shufMask);
2755 return DAG.getNode(ISD::BITCAST, dl, MVT::i128, extShuffle);
2758 //! Custom (target-specific) lowering entry point
2760 This is where LLVM's DAG selection process calls to do target-specific
2764 SPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
2766 unsigned Opc = (unsigned) Op.getOpcode();
2767 EVT VT = Op.getValueType();
2772 errs() << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
2773 errs() << "Op.getOpcode() = " << Opc << "\n";
2774 errs() << "*Op.getNode():\n";
2775 Op.getNode()->dump();
2777 llvm_unreachable(0);
2783 return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
2785 return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
2786 case ISD::ConstantPool:
2787 return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
2788 case ISD::GlobalAddress:
2789 return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
2790 case ISD::JumpTable:
2791 return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
2792 case ISD::ConstantFP:
2793 return LowerConstantFP(Op, DAG);
2795 // i8, i64 math ops:
2804 return LowerI8Math(Op, DAG, Opc, *this);
2808 case ISD::FP_TO_SINT:
2809 case ISD::FP_TO_UINT:
2810 return LowerFP_TO_INT(Op, DAG, *this);
2812 case ISD::SINT_TO_FP:
2813 case ISD::UINT_TO_FP:
2814 return LowerINT_TO_FP(Op, DAG, *this);
2816 // Vector-related lowering.
2817 case ISD::BUILD_VECTOR:
2818 return LowerBUILD_VECTOR(Op, DAG);
2819 case ISD::SCALAR_TO_VECTOR:
2820 return LowerSCALAR_TO_VECTOR(Op, DAG);
2821 case ISD::VECTOR_SHUFFLE:
2822 return LowerVECTOR_SHUFFLE(Op, DAG);
2823 case ISD::EXTRACT_VECTOR_ELT:
2824 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2825 case ISD::INSERT_VECTOR_ELT:
2826 return LowerINSERT_VECTOR_ELT(Op, DAG);
2828 // Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
2832 return LowerByteImmed(Op, DAG);
2834 // Vector and i8 multiply:
2837 return LowerI8Math(Op, DAG, Opc, *this);
2840 return LowerCTPOP(Op, DAG);
2842 case ISD::SELECT_CC:
2843 return LowerSELECT_CC(Op, DAG, *this);
2846 return LowerSETCC(Op, DAG, *this);
2849 return LowerTRUNCATE(Op, DAG);
2851 case ISD::SIGN_EXTEND:
2852 return LowerSIGN_EXTEND(Op, DAG);
2858 void SPUTargetLowering::ReplaceNodeResults(SDNode *N,
2859 SmallVectorImpl<SDValue>&Results,
2860 SelectionDAG &DAG) const
2863 unsigned Opc = (unsigned) N->getOpcode();
2864 EVT OpVT = N->getValueType(0);
2868 errs() << "SPUTargetLowering::ReplaceNodeResults(): need to fix this!\n";
2869 errs() << "Op.getOpcode() = " << Opc << "\n";
2870 errs() << "*Op.getNode():\n";
2878 /* Otherwise, return unchanged */
2881 //===----------------------------------------------------------------------===//
2882 // Target Optimization Hooks
2883 //===----------------------------------------------------------------------===//
2886 SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
2889 TargetMachine &TM = getTargetMachine();
2891 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
2892 SelectionDAG &DAG = DCI.DAG;
2893 SDValue Op0 = N->getOperand(0); // everything has at least one operand
2894 EVT NodeVT = N->getValueType(0); // The node's value type
2895 EVT Op0VT = Op0.getValueType(); // The first operand's result
2896 SDValue Result; // Initially, empty result
2897 DebugLoc dl = N->getDebugLoc();
2899 switch (N->getOpcode()) {
2902 SDValue Op1 = N->getOperand(1);
2904 if (Op0.getOpcode() == SPUISD::IndirectAddr
2905 || Op1.getOpcode() == SPUISD::IndirectAddr) {
2906 // Normalize the operands to reduce repeated code
2907 SDValue IndirectArg = Op0, AddArg = Op1;
2909 if (Op1.getOpcode() == SPUISD::IndirectAddr) {
2914 if (isa<ConstantSDNode>(AddArg)) {
2915 ConstantSDNode *CN0 = cast<ConstantSDNode > (AddArg);
2916 SDValue IndOp1 = IndirectArg.getOperand(1);
2918 if (CN0->isNullValue()) {
2919 // (add (SPUindirect <arg>, <arg>), 0) ->
2920 // (SPUindirect <arg>, <arg>)
2922 #if !defined(NDEBUG)
2923 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2925 << "Replace: (add (SPUindirect <arg>, <arg>), 0)\n"
2926 << "With: (SPUindirect <arg>, <arg>)\n";
2931 } else if (isa<ConstantSDNode>(IndOp1)) {
2932 // (add (SPUindirect <arg>, <const>), <const>) ->
2933 // (SPUindirect <arg>, <const + const>)
2934 ConstantSDNode *CN1 = cast<ConstantSDNode > (IndOp1);
2935 int64_t combinedConst = CN0->getSExtValue() + CN1->getSExtValue();
2936 SDValue combinedValue = DAG.getConstant(combinedConst, Op0VT);
2938 #if !defined(NDEBUG)
2939 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2941 << "Replace: (add (SPUindirect <arg>, " << CN1->getSExtValue()
2942 << "), " << CN0->getSExtValue() << ")\n"
2943 << "With: (SPUindirect <arg>, "
2944 << combinedConst << ")\n";
2948 return DAG.getNode(SPUISD::IndirectAddr, dl, Op0VT,
2949 IndirectArg, combinedValue);
2955 case ISD::SIGN_EXTEND:
2956 case ISD::ZERO_EXTEND:
2957 case ISD::ANY_EXTEND: {
2958 if (Op0.getOpcode() == SPUISD::VEC2PREFSLOT && NodeVT == Op0VT) {
2959 // (any_extend (SPUextract_elt0 <arg>)) ->
2960 // (SPUextract_elt0 <arg>)
2961 // Types must match, however...
2962 #if !defined(NDEBUG)
2963 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2964 errs() << "\nReplace: ";
2966 errs() << "\nWith: ";
2967 Op0.getNode()->dump(&DAG);
2976 case SPUISD::IndirectAddr: {
2977 if (!ST->usingLargeMem() && Op0.getOpcode() == SPUISD::AFormAddr) {
2978 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1));
2979 if (CN != 0 && CN->isNullValue()) {
2980 // (SPUindirect (SPUaform <addr>, 0), 0) ->
2981 // (SPUaform <addr>, 0)
2983 DEBUG(errs() << "Replace: ");
2984 DEBUG(N->dump(&DAG));
2985 DEBUG(errs() << "\nWith: ");
2986 DEBUG(Op0.getNode()->dump(&DAG));
2987 DEBUG(errs() << "\n");
2991 } else if (Op0.getOpcode() == ISD::ADD) {
2992 SDValue Op1 = N->getOperand(1);
2993 if (ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(Op1)) {
2994 // (SPUindirect (add <arg>, <arg>), 0) ->
2995 // (SPUindirect <arg>, <arg>)
2996 if (CN1->isNullValue()) {
2998 #if !defined(NDEBUG)
2999 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
3001 << "Replace: (SPUindirect (add <arg>, <arg>), 0)\n"
3002 << "With: (SPUindirect <arg>, <arg>)\n";
3006 return DAG.getNode(SPUISD::IndirectAddr, dl, Op0VT,
3007 Op0.getOperand(0), Op0.getOperand(1));
3013 case SPUISD::SHL_BITS:
3014 case SPUISD::SHL_BYTES:
3015 case SPUISD::ROTBYTES_LEFT: {
3016 SDValue Op1 = N->getOperand(1);
3018 // Kill degenerate vector shifts:
3019 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
3020 if (CN->isNullValue()) {
3026 case SPUISD::PREFSLOT2VEC: {
3027 switch (Op0.getOpcode()) {
3030 case ISD::ANY_EXTEND:
3031 case ISD::ZERO_EXTEND:
3032 case ISD::SIGN_EXTEND: {
3033 // (SPUprefslot2vec (any|zero|sign_extend (SPUvec2prefslot <arg>))) ->
3035 // but only if the SPUprefslot2vec and <arg> types match.
3036 SDValue Op00 = Op0.getOperand(0);
3037 if (Op00.getOpcode() == SPUISD::VEC2PREFSLOT) {
3038 SDValue Op000 = Op00.getOperand(0);
3039 if (Op000.getValueType() == NodeVT) {
3045 case SPUISD::VEC2PREFSLOT: {
3046 // (SPUprefslot2vec (SPUvec2prefslot <arg>)) ->
3048 Result = Op0.getOperand(0);
3056 // Otherwise, return unchanged.
3058 if (Result.getNode()) {
3059 DEBUG(errs() << "\nReplace.SPU: ");
3060 DEBUG(N->dump(&DAG));
3061 DEBUG(errs() << "\nWith: ");
3062 DEBUG(Result.getNode()->dump(&DAG));
3063 DEBUG(errs() << "\n");
3070 //===----------------------------------------------------------------------===//
3071 // Inline Assembly Support
3072 //===----------------------------------------------------------------------===//
3074 /// getConstraintType - Given a constraint letter, return the type of
3075 /// constraint it is for this target.
3076 SPUTargetLowering::ConstraintType
3077 SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
3078 if (ConstraintLetter.size() == 1) {
3079 switch (ConstraintLetter[0]) {
3086 return C_RegisterClass;
3089 return TargetLowering::getConstraintType(ConstraintLetter);
3092 /// Examine constraint type and operand type and determine a weight value.
3093 /// This object must already have been set up with the operand type
3094 /// and the current alternative constraint selected.
3095 TargetLowering::ConstraintWeight
3096 SPUTargetLowering::getSingleConstraintMatchWeight(
3097 AsmOperandInfo &info, const char *constraint) const {
3098 ConstraintWeight weight = CW_Invalid;
3099 Value *CallOperandVal = info.CallOperandVal;
3100 // If we don't have a value, we can't do a match,
3101 // but allow it at the lowest weight.
3102 if (CallOperandVal == NULL)
3104 // Look at the constraint type.
3105 switch (*constraint) {
3107 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3109 //FIXME: Seems like the supported constraint letters were just copied
3110 // from PPC, as the following doesn't correspond to the GCC docs.
3111 // I'm leaving it so until someone adds the corresponding lowering support.
3118 weight = CW_Register;
3124 std::pair<unsigned, const TargetRegisterClass*>
3125 SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3128 if (Constraint.size() == 1) {
3129 // GCC RS6000 Constraint Letters
3130 switch (Constraint[0]) {
3134 return std::make_pair(0U, SPU::R64CRegisterClass);
3135 return std::make_pair(0U, SPU::R32CRegisterClass);
3138 return std::make_pair(0U, SPU::R32FPRegisterClass);
3139 else if (VT == MVT::f64)
3140 return std::make_pair(0U, SPU::R64FPRegisterClass);
3143 return std::make_pair(0U, SPU::GPRCRegisterClass);
3147 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3150 //! Compute used/known bits for a SPU operand
3152 SPUTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3156 const SelectionDAG &DAG,
3157 unsigned Depth ) const {
3159 const uint64_t uint64_sizebits = sizeof(uint64_t) * CHAR_BIT;
3161 switch (Op.getOpcode()) {
3163 // KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3169 case SPUISD::PREFSLOT2VEC:
3170 case SPUISD::LDRESULT:
3171 case SPUISD::VEC2PREFSLOT:
3172 case SPUISD::SHLQUAD_L_BITS:
3173 case SPUISD::SHLQUAD_L_BYTES:
3174 case SPUISD::VEC_ROTL:
3175 case SPUISD::VEC_ROTR:
3176 case SPUISD::ROTBYTES_LEFT:
3177 case SPUISD::SELECT_MASK:
3184 SPUTargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
3185 unsigned Depth) const {
3186 switch (Op.getOpcode()) {
3191 EVT VT = Op.getValueType();
3193 if (VT != MVT::i8 && VT != MVT::i16 && VT != MVT::i32) {
3196 return VT.getSizeInBits();
3201 // LowerAsmOperandForConstraint
3203 SPUTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3204 char ConstraintLetter,
3205 std::vector<SDValue> &Ops,
3206 SelectionDAG &DAG) const {
3207 // Default, for the time being, to the base class handler
3208 TargetLowering::LowerAsmOperandForConstraint(Op, ConstraintLetter, Ops, DAG);
3211 /// isLegalAddressImmediate - Return true if the integer value can be used
3212 /// as the offset of the target addressing mode.
3213 bool SPUTargetLowering::isLegalAddressImmediate(int64_t V,
3214 const Type *Ty) const {
3215 // SPU's addresses are 256K:
3216 return (V > -(1 << 18) && V < (1 << 18) - 1);
3219 bool SPUTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3224 SPUTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3225 // The SPU target isn't yet aware of offsets.
3229 // can we compare to Imm without writing it into a register?
3230 bool SPUTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
3231 //ceqi, cgti, etc. all take s10 operand
3232 return isInt<10>(Imm);
3236 SPUTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3237 const Type * ) const{
3239 // A-form: 18bit absolute address.
3240 if (AM.BaseGV && !AM.HasBaseReg && AM.Scale == 0 && AM.BaseOffs == 0)
3243 // D-form: reg + 14bit offset
3244 if (AM.BaseGV ==0 && AM.HasBaseReg && AM.Scale == 0 && isInt<14>(AM.BaseOffs))
3248 if (AM.BaseGV == 0 && AM.HasBaseReg && AM.Scale == 1 && AM.BaseOffs ==0)