?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Use %ull here.
[llvm/stm8.git] / lib / Target / CellSPU / SPUISelLowering.cpp
blob743a4d7a0f78c5eccb843d1f47670e61b8637177
1 //===-- SPUISelLowering.cpp - Cell SPU DAG Lowering Implementation --------===//
2 // The LLVM Compiler Infrastructure
3 //
4 // This file is distributed under the University of Illinois Open Source
5 // License. See LICENSE.TXT for details.
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the SPUTargetLowering class.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "SPURegisterNames.h"
14 #include "SPUISelLowering.h"
15 #include "SPUTargetMachine.h"
16 #include "SPUFrameLowering.h"
17 #include "SPUMachineFunction.h"
18 #include "llvm/Constants.h"
19 #include "llvm/Function.h"
20 #include "llvm/Intrinsics.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Type.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/Target/TargetOptions.h"
31 #include "llvm/ADT/VectorExtras.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include <map>
37
38 using namespace llvm;
39
40 // Used in getTargetNodeName() below
41 namespace {
42 std::map<unsigned, const char *> node_names;
43
44 // Byte offset of the preferred slot (counted from the MSB)
45 int prefslotOffset(EVT VT) {
46 int retval=0;
47 if (VT==MVT::i1) retval=3;
48 if (VT==MVT::i8) retval=3;
49 if (VT==MVT::i16) retval=2;
50
51 return retval;
52 }
53
54 //! Expand a library call into an actual call DAG node
55 /*!
56 \note
57 This code is taken from SelectionDAGLegalize, since it is not exposed as
58 part of the LLVM SelectionDAG API.
59 */
60
61 SDValue
62 ExpandLibCall(RTLIB::Libcall LC, SDValue Op, SelectionDAG &DAG,
63 bool isSigned, SDValue &Hi, const SPUTargetLowering &TLI) {
64 // The input chain to this libcall is the entry node of the function.
65 // Legalizing the call will automatically add the previous call to the
66 // dependence.
67 SDValue InChain = DAG.getEntryNode();
68
69 TargetLowering::ArgListTy Args;
70 TargetLowering::ArgListEntry Entry;
71 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
72 EVT ArgVT = Op.getOperand(i).getValueType();
73 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
74 Entry.Node = Op.getOperand(i);
75 Entry.Ty = ArgTy;
76 Entry.isSExt = isSigned;
77 Entry.isZExt = !isSigned;
78 Args.push_back(Entry);
79 }
80 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
81 TLI.getPointerTy());
82
83 // Splice the libcall in wherever FindInputOutputChains tells us to.
84 const Type *RetTy =
85 Op.getNode()->getValueType(0).getTypeForEVT(*DAG.getContext());
86 std::pair<SDValue, SDValue> CallInfo =
87 TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
88 0, TLI.getLibcallCallingConv(LC), false,
89 /*isReturnValueUsed=*/true,
90 Callee, Args, DAG, Op.getDebugLoc());
91
92 return CallInfo.first;
93 }
94 }
95
96 SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
97 : TargetLowering(TM, new TargetLoweringObjectFileELF()),
98 SPUTM(TM) {
99
100 // Use _setjmp/_longjmp instead of setjmp/longjmp.
101 setUseUnderscoreSetJmp(true);
102 setUseUnderscoreLongJmp(true);
103
104 // Set RTLIB libcall names as used by SPU:
105 setLibcallName(RTLIB::DIV_F64, "__fast_divdf3");
106
107 // Set up the SPU's register classes:
108 addRegisterClass(MVT::i8, SPU::R8CRegisterClass);
109 addRegisterClass(MVT::i16, SPU::R16CRegisterClass);
110 addRegisterClass(MVT::i32, SPU::R32CRegisterClass);
111 addRegisterClass(MVT::i64, SPU::R64CRegisterClass);
112 addRegisterClass(MVT::f32, SPU::R32FPRegisterClass);
113 addRegisterClass(MVT::f64, SPU::R64FPRegisterClass);
114 addRegisterClass(MVT::i128, SPU::GPRCRegisterClass);
115
116 // SPU has no sign or zero extended loads for i1, i8, i16:
117 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
119 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
120
121 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
122 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
123
124 setTruncStoreAction(MVT::i128, MVT::i64, Expand);
125 setTruncStoreAction(MVT::i128, MVT::i32, Expand);
126 setTruncStoreAction(MVT::i128, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i128, MVT::i8, Expand);
128
129 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
130
131 // SPU constant load actions are custom lowered:
132 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
133 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
134
135 // SPU's loads and stores have to be custom lowered:
136 for (unsigned sctype = (unsigned) MVT::i8; sctype < (unsigned) MVT::i128;
137 ++sctype) {
138 MVT::SimpleValueType VT = (MVT::SimpleValueType)sctype;
139
140 setOperationAction(ISD::LOAD, VT, Custom);
141 setOperationAction(ISD::STORE, VT, Custom);
142 setLoadExtAction(ISD::EXTLOAD, VT, Custom);
143 setLoadExtAction(ISD::ZEXTLOAD, VT, Custom);
144 setLoadExtAction(ISD::SEXTLOAD, VT, Custom);
145
146 for (unsigned stype = sctype - 1; stype >= (unsigned) MVT::i8; --stype) {
147 MVT::SimpleValueType StoreVT = (MVT::SimpleValueType) stype;
148 setTruncStoreAction(VT, StoreVT, Expand);
149 }
150 }
151
152 for (unsigned sctype = (unsigned) MVT::f32; sctype < (unsigned) MVT::f64;
153 ++sctype) {
154 MVT::SimpleValueType VT = (MVT::SimpleValueType) sctype;
155
156 setOperationAction(ISD::LOAD, VT, Custom);
157 setOperationAction(ISD::STORE, VT, Custom);
158
159 for (unsigned stype = sctype - 1; stype >= (unsigned) MVT::f32; --stype) {
160 MVT::SimpleValueType StoreVT = (MVT::SimpleValueType) stype;
161 setTruncStoreAction(VT, StoreVT, Expand);
162 }
163 }
164
165 // Expand the jumptable branches
166 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
167 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
168
169 // Custom lower SELECT_CC for most cases, but expand by default
170 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
171 setOperationAction(ISD::SELECT_CC, MVT::i8, Custom);
172 setOperationAction(ISD::SELECT_CC, MVT::i16, Custom);
173 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
174 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
175
176 // SPU has no intrinsics for these particular operations:
177 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
178
179 // SPU has no division/remainder instructions
180 setOperationAction(ISD::SREM, MVT::i8, Expand);
181 setOperationAction(ISD::UREM, MVT::i8, Expand);
182 setOperationAction(ISD::SDIV, MVT::i8, Expand);
183 setOperationAction(ISD::UDIV, MVT::i8, Expand);
184 setOperationAction(ISD::SDIVREM, MVT::i8, Expand);
185 setOperationAction(ISD::UDIVREM, MVT::i8, Expand);
186 setOperationAction(ISD::SREM, MVT::i16, Expand);
187 setOperationAction(ISD::UREM, MVT::i16, Expand);
188 setOperationAction(ISD::SDIV, MVT::i16, Expand);
189 setOperationAction(ISD::UDIV, MVT::i16, Expand);
190 setOperationAction(ISD::SDIVREM, MVT::i16, Expand);
191 setOperationAction(ISD::UDIVREM, MVT::i16, Expand);
192 setOperationAction(ISD::SREM, MVT::i32, Expand);
193 setOperationAction(ISD::UREM, MVT::i32, Expand);
194 setOperationAction(ISD::SDIV, MVT::i32, Expand);
195 setOperationAction(ISD::UDIV, MVT::i32, Expand);
196 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
197 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
198 setOperationAction(ISD::SREM, MVT::i64, Expand);
199 setOperationAction(ISD::UREM, MVT::i64, Expand);
200 setOperationAction(ISD::SDIV, MVT::i64, Expand);
201 setOperationAction(ISD::UDIV, MVT::i64, Expand);
202 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
203 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
204 setOperationAction(ISD::SREM, MVT::i128, Expand);
205 setOperationAction(ISD::UREM, MVT::i128, Expand);
206 setOperationAction(ISD::SDIV, MVT::i128, Expand);
207 setOperationAction(ISD::UDIV, MVT::i128, Expand);
208 setOperationAction(ISD::SDIVREM, MVT::i128, Expand);
209 setOperationAction(ISD::UDIVREM, MVT::i128, Expand);
210
211 // We don't support sin/cos/sqrt/fmod
212 setOperationAction(ISD::FSIN , MVT::f64, Expand);
213 setOperationAction(ISD::FCOS , MVT::f64, Expand);
214 setOperationAction(ISD::FREM , MVT::f64, Expand);
215 setOperationAction(ISD::FSIN , MVT::f32, Expand);
216 setOperationAction(ISD::FCOS , MVT::f32, Expand);
217 setOperationAction(ISD::FREM , MVT::f32, Expand);
218
219 // Expand fsqrt to the appropriate libcall (NOTE: should use h/w fsqrt
220 // for f32!)
221 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
222 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
223
224 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
225 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
226
227 // SPU can do rotate right and left, so legalize it... but customize for i8
228 // because instructions don't exist.
229
230 // FIXME: Change from "expand" to appropriate type once ROTR is supported in
231 // .td files.
232 setOperationAction(ISD::ROTR, MVT::i32, Expand /*Legal*/);
233 setOperationAction(ISD::ROTR, MVT::i16, Expand /*Legal*/);
234 setOperationAction(ISD::ROTR, MVT::i8, Expand /*Custom*/);
235
236 setOperationAction(ISD::ROTL, MVT::i32, Legal);
237 setOperationAction(ISD::ROTL, MVT::i16, Legal);
238 setOperationAction(ISD::ROTL, MVT::i8, Custom);
239
240 // SPU has no native version of shift left/right for i8
241 setOperationAction(ISD::SHL, MVT::i8, Custom);
242 setOperationAction(ISD::SRL, MVT::i8, Custom);
243 setOperationAction(ISD::SRA, MVT::i8, Custom);
244
245 // Make these operations legal and handle them during instruction selection:
246 setOperationAction(ISD::SHL, MVT::i64, Legal);
247 setOperationAction(ISD::SRL, MVT::i64, Legal);
248 setOperationAction(ISD::SRA, MVT::i64, Legal);
249
250 // Custom lower i8, i32 and i64 multiplications
251 setOperationAction(ISD::MUL, MVT::i8, Custom);
252 setOperationAction(ISD::MUL, MVT::i32, Legal);
253 setOperationAction(ISD::MUL, MVT::i64, Legal);
254
255 // Expand double-width multiplication
256 // FIXME: It would probably be reasonable to support some of these operations
257 setOperationAction(ISD::UMUL_LOHI, MVT::i8, Expand);
258 setOperationAction(ISD::SMUL_LOHI, MVT::i8, Expand);
259 setOperationAction(ISD::MULHU, MVT::i8, Expand);
260 setOperationAction(ISD::MULHS, MVT::i8, Expand);
261 setOperationAction(ISD::UMUL_LOHI, MVT::i16, Expand);
262 setOperationAction(ISD::SMUL_LOHI, MVT::i16, Expand);
263 setOperationAction(ISD::MULHU, MVT::i16, Expand);
264 setOperationAction(ISD::MULHS, MVT::i16, Expand);
265 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
266 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
267 setOperationAction(ISD::MULHU, MVT::i32, Expand);
268 setOperationAction(ISD::MULHS, MVT::i32, Expand);
269 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
270 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
271 setOperationAction(ISD::MULHU, MVT::i64, Expand);
272 setOperationAction(ISD::MULHS, MVT::i64, Expand);
273
274 // Need to custom handle (some) common i8, i64 math ops
275 setOperationAction(ISD::ADD, MVT::i8, Custom);
276 setOperationAction(ISD::ADD, MVT::i64, Legal);
277 setOperationAction(ISD::SUB, MVT::i8, Custom);
278 setOperationAction(ISD::SUB, MVT::i64, Legal);
279
280 // SPU does not have BSWAP. It does have i32 support CTLZ.
281 // CTPOP has to be custom lowered.
282 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
283 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
284
285 setOperationAction(ISD::CTPOP, MVT::i8, Custom);
286 setOperationAction(ISD::CTPOP, MVT::i16, Custom);
287 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
288 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
289 setOperationAction(ISD::CTPOP, MVT::i128, Expand);
290
291 setOperationAction(ISD::CTTZ , MVT::i8, Expand);
292 setOperationAction(ISD::CTTZ , MVT::i16, Expand);
293 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
294 setOperationAction(ISD::CTTZ , MVT::i64, Expand);
295 setOperationAction(ISD::CTTZ , MVT::i128, Expand);
296
297 setOperationAction(ISD::CTLZ , MVT::i8, Promote);
298 setOperationAction(ISD::CTLZ , MVT::i16, Promote);
299 setOperationAction(ISD::CTLZ , MVT::i32, Legal);
300 setOperationAction(ISD::CTLZ , MVT::i64, Expand);
301 setOperationAction(ISD::CTLZ , MVT::i128, Expand);
302
303 // SPU has a version of select that implements (a&~c)|(b&c), just like
304 // select ought to work:
305 setOperationAction(ISD::SELECT, MVT::i8, Legal);
306 setOperationAction(ISD::SELECT, MVT::i16, Legal);
307 setOperationAction(ISD::SELECT, MVT::i32, Legal);
308 setOperationAction(ISD::SELECT, MVT::i64, Legal);
309
310 setOperationAction(ISD::SETCC, MVT::i8, Legal);
311 setOperationAction(ISD::SETCC, MVT::i16, Legal);
312 setOperationAction(ISD::SETCC, MVT::i32, Legal);
313 setOperationAction(ISD::SETCC, MVT::i64, Legal);
314 setOperationAction(ISD::SETCC, MVT::f64, Custom);
315
316 // Custom lower i128 -> i64 truncates
317 setOperationAction(ISD::TRUNCATE, MVT::i64, Custom);
318
319 // Custom lower i32/i64 -> i128 sign extend
320 setOperationAction(ISD::SIGN_EXTEND, MVT::i128, Custom);
321
322 setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote);
323 setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote);
324 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
325 setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
326 // SPU has a legal FP -> signed INT instruction for f32, but for f64, need
327 // to expand to a libcall, hence the custom lowering:
328 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
329 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
330 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
331 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
332 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Expand);
333 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Expand);
334
335 // FDIV on SPU requires custom lowering
336 setOperationAction(ISD::FDIV, MVT::f64, Expand); // to libcall
337
338 // SPU has [U|S]INT_TO_FP for f32->i32, but not for f64->i32, f64->i64:
339 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
340 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
341 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
342 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
343 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
344 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
345 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
346 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
347
348 setOperationAction(ISD::BITCAST, MVT::i32, Legal);
349 setOperationAction(ISD::BITCAST, MVT::f32, Legal);
350 setOperationAction(ISD::BITCAST, MVT::i64, Legal);
351 setOperationAction(ISD::BITCAST, MVT::f64, Legal);
352
353 // We cannot sextinreg(i1). Expand to shifts.
354 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
355
356 // We want to legalize GlobalAddress and ConstantPool nodes into the
357 // appropriate instructions to materialize the address.
358 for (unsigned sctype = (unsigned) MVT::i8; sctype < (unsigned) MVT::f128;
359 ++sctype) {
360 MVT::SimpleValueType VT = (MVT::SimpleValueType)sctype;
361
362 setOperationAction(ISD::GlobalAddress, VT, Custom);
363 setOperationAction(ISD::ConstantPool, VT, Custom);
364 setOperationAction(ISD::JumpTable, VT, Custom);
365 }
366
367 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
368 setOperationAction(ISD::VASTART , MVT::Other, Custom);
369
370 // Use the default implementation.
371 setOperationAction(ISD::VAARG , MVT::Other, Expand);
372 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
373 setOperationAction(ISD::VAEND , MVT::Other, Expand);
374 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
375 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
376 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
377 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Expand);
378
379 // Cell SPU has instructions for converting between i64 and fp.
380 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
381 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
382
383 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
384 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
385
386 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
387 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
388
389 // First set operation action for all vector types to expand. Then we
390 // will selectively turn on ones that can be effectively codegen'd.
391 addRegisterClass(MVT::v16i8, SPU::VECREGRegisterClass);
392 addRegisterClass(MVT::v8i16, SPU::VECREGRegisterClass);
393 addRegisterClass(MVT::v4i32, SPU::VECREGRegisterClass);
394 addRegisterClass(MVT::v2i64, SPU::VECREGRegisterClass);
395 addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
396 addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
397
398 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
399 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
400 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
401
402 // add/sub are legal for all supported vector VT's.
403 setOperationAction(ISD::ADD, VT, Legal);
404 setOperationAction(ISD::SUB, VT, Legal);
405 // mul has to be custom lowered.
406 setOperationAction(ISD::MUL, VT, Legal);
407
408 setOperationAction(ISD::AND, VT, Legal);
409 setOperationAction(ISD::OR, VT, Legal);
410 setOperationAction(ISD::XOR, VT, Legal);
411 setOperationAction(ISD::LOAD, VT, Custom);
412 setOperationAction(ISD::SELECT, VT, Legal);
413 setOperationAction(ISD::STORE, VT, Custom);
414
415 // These operations need to be expanded:
416 setOperationAction(ISD::SDIV, VT, Expand);
417 setOperationAction(ISD::SREM, VT, Expand);
418 setOperationAction(ISD::UDIV, VT, Expand);
419 setOperationAction(ISD::UREM, VT, Expand);
420
421 // Custom lower build_vector, constant pool spills, insert and
422 // extract vector elements:
423 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
424 setOperationAction(ISD::ConstantPool, VT, Custom);
425 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
426 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
427 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
428 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
429 }
430
431 setOperationAction(ISD::AND, MVT::v16i8, Custom);
432 setOperationAction(ISD::OR, MVT::v16i8, Custom);
433 setOperationAction(ISD::XOR, MVT::v16i8, Custom);
434 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
435
436 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
437
438 setBooleanContents(ZeroOrNegativeOneBooleanContent);
439
440 setStackPointerRegisterToSaveRestore(SPU::R1);
441
442 // We have target-specific dag combine patterns for the following nodes:
443 setTargetDAGCombine(ISD::ADD);
444 setTargetDAGCombine(ISD::ZERO_EXTEND);
445 setTargetDAGCombine(ISD::SIGN_EXTEND);
446 setTargetDAGCombine(ISD::ANY_EXTEND);
447
448 computeRegisterProperties();
449
450 // Set pre-RA register scheduler default to BURR, which produces slightly
451 // better code than the default (could also be TDRR, but TargetLowering.h
452 // needs a mod to support that model):
453 setSchedulingPreference(Sched::RegPressure);
454 }
455
456 const char *
457 SPUTargetLowering::getTargetNodeName(unsigned Opcode) const
458 {
459 if (node_names.empty()) {
460 node_names[(unsigned) SPUISD::RET_FLAG] = "SPUISD::RET_FLAG";
461 node_names[(unsigned) SPUISD::Hi] = "SPUISD::Hi";
462 node_names[(unsigned) SPUISD::Lo] = "SPUISD::Lo";
463 node_names[(unsigned) SPUISD::PCRelAddr] = "SPUISD::PCRelAddr";
464 node_names[(unsigned) SPUISD::AFormAddr] = "SPUISD::AFormAddr";
465 node_names[(unsigned) SPUISD::IndirectAddr] = "SPUISD::IndirectAddr";
466 node_names[(unsigned) SPUISD::LDRESULT] = "SPUISD::LDRESULT";
467 node_names[(unsigned) SPUISD::CALL] = "SPUISD::CALL";
468 node_names[(unsigned) SPUISD::SHUFB] = "SPUISD::SHUFB";
469 node_names[(unsigned) SPUISD::SHUFFLE_MASK] = "SPUISD::SHUFFLE_MASK";
470 node_names[(unsigned) SPUISD::CNTB] = "SPUISD::CNTB";
471 node_names[(unsigned) SPUISD::PREFSLOT2VEC] = "SPUISD::PREFSLOT2VEC";
472 node_names[(unsigned) SPUISD::VEC2PREFSLOT] = "SPUISD::VEC2PREFSLOT";
473 node_names[(unsigned) SPUISD::SHL_BITS] = "SPUISD::SHL_BITS";
474 node_names[(unsigned) SPUISD::SHL_BYTES] = "SPUISD::SHL_BYTES";
475 node_names[(unsigned) SPUISD::VEC_ROTL] = "SPUISD::VEC_ROTL";
476 node_names[(unsigned) SPUISD::VEC_ROTR] = "SPUISD::VEC_ROTR";
477 node_names[(unsigned) SPUISD::ROTBYTES_LEFT] = "SPUISD::ROTBYTES_LEFT";
478 node_names[(unsigned) SPUISD::ROTBYTES_LEFT_BITS] =
479 "SPUISD::ROTBYTES_LEFT_BITS";
480 node_names[(unsigned) SPUISD::SELECT_MASK] = "SPUISD::SELECT_MASK";
481 node_names[(unsigned) SPUISD::SELB] = "SPUISD::SELB";
482 node_names[(unsigned) SPUISD::ADD64_MARKER] = "SPUISD::ADD64_MARKER";
483 node_names[(unsigned) SPUISD::SUB64_MARKER] = "SPUISD::SUB64_MARKER";
484 node_names[(unsigned) SPUISD::MUL64_MARKER] = "SPUISD::MUL64_MARKER";
485 }
486
487 std::map<unsigned, const char *>::iterator i = node_names.find(Opcode);
488
489 return ((i != node_names.end()) ? i->second : 0);
490 }
491
492 /// getFunctionAlignment - Return the Log2 alignment of this function.
493 unsigned SPUTargetLowering::getFunctionAlignment(const Function *) const {
494 return 3;
495 }
496
497 //===----------------------------------------------------------------------===//
498 // Return the Cell SPU's SETCC result type
499 //===----------------------------------------------------------------------===//
500
501 MVT::SimpleValueType SPUTargetLowering::getSetCCResultType(EVT VT) const {
502 // i8, i16 and i32 are valid SETCC result types
503 MVT::SimpleValueType retval;
504
505 switch(VT.getSimpleVT().SimpleTy){
506 case MVT::i1:
507 case MVT::i8:
508 retval = MVT::i8; break;
509 case MVT::i16:
510 retval = MVT::i16; break;
511 case MVT::i32:
512 default:
513 retval = MVT::i32;
514 }
515 return retval;
516 }
517
518 //===----------------------------------------------------------------------===//
519 // Calling convention code:
520 //===----------------------------------------------------------------------===//
521
522 #include "SPUGenCallingConv.inc"
523
524 //===----------------------------------------------------------------------===//
525 // LowerOperation implementation
526 //===----------------------------------------------------------------------===//
527
528 /// Custom lower loads for CellSPU
529 /*!
530 All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
531 within a 16-byte block, we have to rotate to extract the requested element.
532
533 For extending loads, we also want to ensure that the following sequence is
534 emitted, e.g. for MVT::f32 extending load to MVT::f64:
535
536 \verbatim
537 %1 v16i8,ch = load
538 %2 v16i8,ch = rotate %1
539 %3 v4f8, ch = bitconvert %2
540 %4 f32 = vec2perfslot %3
541 %5 f64 = fp_extend %4
542 \endverbatim
543 */
544 static SDValue
545 LowerLOAD(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
546 LoadSDNode *LN = cast<LoadSDNode>(Op);
547 SDValue the_chain = LN->getChain();
548 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
549 EVT InVT = LN->getMemoryVT();
550 EVT OutVT = Op.getValueType();
551 ISD::LoadExtType ExtType = LN->getExtensionType();
552 unsigned alignment = LN->getAlignment();
553 int pso = prefslotOffset(InVT);
554 DebugLoc dl = Op.getDebugLoc();
555 EVT vecVT = InVT.isVector()? InVT: EVT::getVectorVT(*DAG.getContext(), InVT,
556 (128 / InVT.getSizeInBits()));
557
558 // two sanity checks
559 assert( LN->getAddressingMode() == ISD::UNINDEXED
560 && "we should get only UNINDEXED adresses");
561 // clean aligned loads can be selected as-is
562 if (InVT.getSizeInBits() == 128 && (alignment%16) == 0)
563 return SDValue();
564
565 // Get pointerinfos to the memory chunk(s) that contain the data to load
566 uint64_t mpi_offset = LN->getPointerInfo().Offset;
567 mpi_offset -= mpi_offset%16;
568 MachinePointerInfo lowMemPtr(LN->getPointerInfo().V, mpi_offset);
569 MachinePointerInfo highMemPtr(LN->getPointerInfo().V, mpi_offset+16);
570
571 SDValue result;
572 SDValue basePtr = LN->getBasePtr();
573 SDValue rotate;
574
575 if ((alignment%16) == 0) {
576 ConstantSDNode *CN;
577
578 // Special cases for a known aligned load to simplify the base pointer
579 // and the rotation amount:
580 if (basePtr.getOpcode() == ISD::ADD
581 && (CN = dyn_cast<ConstantSDNode > (basePtr.getOperand(1))) != 0) {
582 // Known offset into basePtr
583 int64_t offset = CN->getSExtValue();
584 int64_t rotamt = int64_t((offset & 0xf) - pso);
585
586 if (rotamt < 0)
587 rotamt += 16;
588
589 rotate = DAG.getConstant(rotamt, MVT::i16);
590
591 // Simplify the base pointer for this case:
592 basePtr = basePtr.getOperand(0);
593 if ((offset & ~0xf) > 0) {
594 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
595 basePtr,
596 DAG.getConstant((offset & ~0xf), PtrVT));
597 }
598 } else if ((basePtr.getOpcode() == SPUISD::AFormAddr)
599 || (basePtr.getOpcode() == SPUISD::IndirectAddr
600 && basePtr.getOperand(0).getOpcode() == SPUISD::Hi
601 && basePtr.getOperand(1).getOpcode() == SPUISD::Lo)) {
602 // Plain aligned a-form address: rotate into preferred slot
603 // Same for (SPUindirect (SPUhi ...), (SPUlo ...))
604 int64_t rotamt = -pso;
605 if (rotamt < 0)
606 rotamt += 16;
607 rotate = DAG.getConstant(rotamt, MVT::i16);
608 } else {
609 // Offset the rotate amount by the basePtr and the preferred slot
610 // byte offset
611 int64_t rotamt = -pso;
612 if (rotamt < 0)
613 rotamt += 16;
614 rotate = DAG.getNode(ISD::ADD, dl, PtrVT,
615 basePtr,
616 DAG.getConstant(rotamt, PtrVT));
617 }
618 } else {
619 // Unaligned load: must be more pessimistic about addressing modes:
620 if (basePtr.getOpcode() == ISD::ADD) {
621 MachineFunction &MF = DAG.getMachineFunction();
622 MachineRegisterInfo &RegInfo = MF.getRegInfo();
623 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
624 SDValue Flag;
625
626 SDValue Op0 = basePtr.getOperand(0);
627 SDValue Op1 = basePtr.getOperand(1);
628
629 if (isa<ConstantSDNode>(Op1)) {
630 // Convert the (add <ptr>, <const>) to an indirect address contained
631 // in a register. Note that this is done because we need to avoid
632 // creating a 0(reg) d-form address due to the SPU's block loads.
633 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
634 the_chain = DAG.getCopyToReg(the_chain, dl, VReg, basePtr, Flag);
635 basePtr = DAG.getCopyFromReg(the_chain, dl, VReg, PtrVT);
636 } else {
637 // Convert the (add <arg1>, <arg2>) to an indirect address, which
638 // will likely be lowered as a reg(reg) x-form address.
639 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
640 }
641 } else {
642 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
643 basePtr,
644 DAG.getConstant(0, PtrVT));
645 }
646
647 // Offset the rotate amount by the basePtr and the preferred slot
648 // byte offset
649 rotate = DAG.getNode(ISD::ADD, dl, PtrVT,
650 basePtr,
651 DAG.getConstant(-pso, PtrVT));
652 }
653
654 // Do the load as a i128 to allow possible shifting
655 SDValue low = DAG.getLoad(MVT::i128, dl, the_chain, basePtr,
656 lowMemPtr,
657 LN->isVolatile(), LN->isNonTemporal(), 16);
658
659 // When the size is not greater than alignment we get all data with just
660 // one load
661 if (alignment >= InVT.getSizeInBits()/8) {
662 // Update the chain
663 the_chain = low.getValue(1);
664
665 // Rotate into the preferred slot:
666 result = DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, MVT::i128,
667 low.getValue(0), rotate);
668
669 // Convert the loaded v16i8 vector to the appropriate vector type
670 // specified by the operand:
671 EVT vecVT = EVT::getVectorVT(*DAG.getContext(),
672 InVT, (128 / InVT.getSizeInBits()));
673 result = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, InVT,
674 DAG.getNode(ISD::BITCAST, dl, vecVT, result));
675 }
676 // When alignment is less than the size, we might need (known only at
677 // run-time) two loads
678 // TODO: if the memory address is composed only from constants, we have
679 // extra kowledge, and might avoid the second load
680 else {
681 // storage position offset from lower 16 byte aligned memory chunk
682 SDValue offset = DAG.getNode(ISD::AND, dl, MVT::i32,
683 basePtr, DAG.getConstant( 0xf, MVT::i32 ) );
684 // get a registerfull of ones. (this implementation is a workaround: LLVM
685 // cannot handle 128 bit signed int constants)
686 SDValue ones = DAG.getConstant(-1, MVT::v4i32 );
687 ones = DAG.getNode(ISD::BITCAST, dl, MVT::i128, ones);
688
689 SDValue high = DAG.getLoad(MVT::i128, dl, the_chain,
690 DAG.getNode(ISD::ADD, dl, PtrVT,
691 basePtr,
692 DAG.getConstant(16, PtrVT)),
693 highMemPtr,
694 LN->isVolatile(), LN->isNonTemporal(), 16);
695
696 the_chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(1),
697 high.getValue(1));
698
699 // Shift the (possible) high part right to compensate the misalignemnt.
700 // if there is no highpart (i.e. value is i64 and offset is 4), this
701 // will zero out the high value.
702 high = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, high,
703 DAG.getNode(ISD::SUB, dl, MVT::i32,
704 DAG.getConstant( 16, MVT::i32),
705 offset
706 ));
707
708 // Shift the low similarily
709 // TODO: add SPUISD::SHL_BYTES
710 low = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, low, offset );
711
712 // Merge the two parts
713 result = DAG.getNode(ISD::BITCAST, dl, vecVT,
714 DAG.getNode(ISD::OR, dl, MVT::i128, low, high));
715
716 if (!InVT.isVector()) {
717 result = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, InVT, result );
718 }
719
720 }
721 // Handle extending loads by extending the scalar result:
722 if (ExtType == ISD::SEXTLOAD) {
723 result = DAG.getNode(ISD::SIGN_EXTEND, dl, OutVT, result);
724 } else if (ExtType == ISD::ZEXTLOAD) {
725 result = DAG.getNode(ISD::ZERO_EXTEND, dl, OutVT, result);
726 } else if (ExtType == ISD::EXTLOAD) {
727 unsigned NewOpc = ISD::ANY_EXTEND;
728
729 if (OutVT.isFloatingPoint())
730 NewOpc = ISD::FP_EXTEND;
731
732 result = DAG.getNode(NewOpc, dl, OutVT, result);
733 }
734
735 SDVTList retvts = DAG.getVTList(OutVT, MVT::Other);
736 SDValue retops[2] = {
737 result,
738 the_chain
739 };
740
741 result = DAG.getNode(SPUISD::LDRESULT, dl, retvts,
742 retops, sizeof(retops) / sizeof(retops[0]));
743 return result;
744 }
745
746 /// Custom lower stores for CellSPU
747 /*!
748 All CellSPU stores are aligned to 16-byte boundaries, so for elements
749 within a 16-byte block, we have to generate a shuffle to insert the
750 requested element into its place, then store the resulting block.
751 */
752 static SDValue
753 LowerSTORE(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
754 StoreSDNode *SN = cast<StoreSDNode>(Op);
755 SDValue Value = SN->getValue();
756 EVT VT = Value.getValueType();
757 EVT StVT = (!SN->isTruncatingStore() ? VT : SN->getMemoryVT());
758 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
759 DebugLoc dl = Op.getDebugLoc();
760 unsigned alignment = SN->getAlignment();
761 SDValue result;
762 EVT vecVT = StVT.isVector()? StVT: EVT::getVectorVT(*DAG.getContext(), StVT,
763 (128 / StVT.getSizeInBits()));
764 // Get pointerinfos to the memory chunk(s) that contain the data to load
765 uint64_t mpi_offset = SN->getPointerInfo().Offset;
766 mpi_offset -= mpi_offset%16;
767 MachinePointerInfo lowMemPtr(SN->getPointerInfo().V, mpi_offset);
768 MachinePointerInfo highMemPtr(SN->getPointerInfo().V, mpi_offset+16);
769
770
771 // two sanity checks
772 assert( SN->getAddressingMode() == ISD::UNINDEXED
773 && "we should get only UNINDEXED adresses");
774 // clean aligned loads can be selected as-is
775 if (StVT.getSizeInBits() == 128 && (alignment%16) == 0)
776 return SDValue();
777
778 SDValue alignLoadVec;
779 SDValue basePtr = SN->getBasePtr();
780 SDValue the_chain = SN->getChain();
781 SDValue insertEltOffs;
782
783 if ((alignment%16) == 0) {
784 ConstantSDNode *CN;
785 // Special cases for a known aligned load to simplify the base pointer
786 // and insertion byte:
787 if (basePtr.getOpcode() == ISD::ADD
788 && (CN = dyn_cast<ConstantSDNode>(basePtr.getOperand(1))) != 0) {
789 // Known offset into basePtr
790 int64_t offset = CN->getSExtValue();
791
792 // Simplify the base pointer for this case:
793 basePtr = basePtr.getOperand(0);
794 insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
795 basePtr,
796 DAG.getConstant((offset & 0xf), PtrVT));
797
798 if ((offset & ~0xf) > 0) {
799 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
800 basePtr,
801 DAG.getConstant((offset & ~0xf), PtrVT));
802 }
803 } else {
804 // Otherwise, assume it's at byte 0 of basePtr
805 insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
806 basePtr,
807 DAG.getConstant(0, PtrVT));
808 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
809 basePtr,
810 DAG.getConstant(0, PtrVT));
811 }
812 } else {
813 // Unaligned load: must be more pessimistic about addressing modes:
814 if (basePtr.getOpcode() == ISD::ADD) {
815 MachineFunction &MF = DAG.getMachineFunction();
816 MachineRegisterInfo &RegInfo = MF.getRegInfo();
817 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
818 SDValue Flag;
819
820 SDValue Op0 = basePtr.getOperand(0);
821 SDValue Op1 = basePtr.getOperand(1);
822
823 if (isa<ConstantSDNode>(Op1)) {
824 // Convert the (add <ptr>, <const>) to an indirect address contained
825 // in a register. Note that this is done because we need to avoid
826 // creating a 0(reg) d-form address due to the SPU's block loads.
827 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
828 the_chain = DAG.getCopyToReg(the_chain, dl, VReg, basePtr, Flag);
829 basePtr = DAG.getCopyFromReg(the_chain, dl, VReg, PtrVT);
830 } else {
831 // Convert the (add <arg1>, <arg2>) to an indirect address, which
832 // will likely be lowered as a reg(reg) x-form address.
833 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
834 }
835 } else {
836 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
837 basePtr,
838 DAG.getConstant(0, PtrVT));
839 }
840
841 // Insertion point is solely determined by basePtr's contents
842 insertEltOffs = DAG.getNode(ISD::ADD, dl, PtrVT,
843 basePtr,
844 DAG.getConstant(0, PtrVT));
845 }
846
847 // Load the lower part of the memory to which to store.
848 SDValue low = DAG.getLoad(vecVT, dl, the_chain, basePtr,
849 lowMemPtr, SN->isVolatile(), SN->isNonTemporal(), 16);
850
851 // if we don't need to store over the 16 byte boundary, one store suffices
852 if (alignment >= StVT.getSizeInBits()/8) {
853 // Update the chain
854 the_chain = low.getValue(1);
855
856 LoadSDNode *LN = cast<LoadSDNode>(low);
857 SDValue theValue = SN->getValue();
858
859 if (StVT != VT
860 && (theValue.getOpcode() == ISD::AssertZext
861 || theValue.getOpcode() == ISD::AssertSext)) {
862 // Drill down and get the value for zero- and sign-extended
863 // quantities
864 theValue = theValue.getOperand(0);
865 }
866
867 // If the base pointer is already a D-form address, then just create
868 // a new D-form address with a slot offset and the orignal base pointer.
869 // Otherwise generate a D-form address with the slot offset relative
870 // to the stack pointer, which is always aligned.
871 #if !defined(NDEBUG)
872 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
873 errs() << "CellSPU LowerSTORE: basePtr = ";
874 basePtr.getNode()->dump(&DAG);
875 errs() << "\n";
876 }
877 #endif
878
879 SDValue insertEltOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl, vecVT,
880 insertEltOffs);
881 SDValue vectorizeOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, vecVT,
882 theValue);
883
884 result = DAG.getNode(SPUISD::SHUFB, dl, vecVT,
885 vectorizeOp, low,
886 DAG.getNode(ISD::BITCAST, dl,
887 MVT::v4i32, insertEltOp));
888
889 result = DAG.getStore(the_chain, dl, result, basePtr,
890 lowMemPtr,
891 LN->isVolatile(), LN->isNonTemporal(),
892 16);
893
894 }
895 // do the store when it might cross the 16 byte memory access boundary.
896 else {
897 // TODO issue a warning if SN->isVolatile()== true? This is likely not
898 // what the user wanted.
899
900 // address offset from nearest lower 16byte alinged address
901 SDValue offset = DAG.getNode(ISD::AND, dl, MVT::i32,
902 SN->getBasePtr(),
903 DAG.getConstant(0xf, MVT::i32));
904 // 16 - offset
905 SDValue offset_compl = DAG.getNode(ISD::SUB, dl, MVT::i32,
906 DAG.getConstant( 16, MVT::i32),
907 offset);
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,
912 MVT::i32));
913 // get a registerfull of ones
914 SDValue ones = DAG.getConstant(-1, MVT::v4i32);
915 ones = DAG.getNode(ISD::BITCAST, dl, MVT::i128, ones);
916
917 // Create the 128 bit masks that have ones where the data to store is
918 // located.
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
922 // is masked.
923 // this is e.g. in the case of store i32, align 2
924 if (!VT.isVector()){
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,
928 surplus);
929 Value = DAG.getNode(ISD::BITCAST, dl, MVT::i128, Value);
930 Value = DAG.getNode(ISD::AND, dl, MVT::i128, Value, lowmask);
931
932 }
933 else {
934 lowmask = ones;
935 Value = DAG.getNode(ISD::BITCAST, dl, MVT::i128, Value);
936 }
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,
939 offset_compl);
940 lowmask = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, lowmask,
941 offset);
942
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)),
948 highMemPtr,
949 SN->isVolatile(), SN->isNonTemporal(), 16);
950 the_chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(1),
951 hi.getValue(1));
952
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));
959
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,
965 offset_compl);
966
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));
975
976 low = DAG.getStore(the_chain, dl, rlow, basePtr,
977 lowMemPtr,
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)),
982 highMemPtr,
983 SN->isVolatile(), SN->isNonTemporal(), 16);
984 result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(0),
985 hi.getValue(0));
986 }
987
988 return result;
989 }
990
991 //! Generate the address of a constant pool entry.
992 static SDValue
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();
1002
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);
1007 } else {
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);
1011 }
1012 }
1013
1014 llvm_unreachable("LowerConstantPool: Relocation model other than static"
1015 " not supported.");
1016 return SDValue();
1017 }
1018
1019 //! Alternate entry point for generating the address of a constant pool entry
1020 SDValue
1021 SPU::LowerConstantPool(SDValue Op, SelectionDAG &DAG, const SPUTargetMachine &TM) {
1022 return ::LowerConstantPool(Op, DAG, TM.getSubtargetImpl());
1023 }
1024
1025 static SDValue
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();
1034
1035 if (TM.getRelocationModel() == Reloc::Static) {
1036 if (!ST->usingLargeMem()) {
1037 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, JTI, Zero);
1038 } else {
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);
1042 }
1043 }
1044
1045 llvm_unreachable("LowerJumpTable: Relocation model other than static"
1046 " not supported.");
1047 return SDValue();
1048 }
1049
1050 static SDValue
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();
1061
1062 if (TM.getRelocationModel() == Reloc::Static) {
1063 if (!ST->usingLargeMem()) {
1064 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, GA, Zero);
1065 } else {
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);
1069 }
1070 } else {
1071 report_fatal_error("LowerGlobalAddress: Relocation model other than static"
1072 "not supported.");
1073 /*NOTREACHED*/
1074 }
1075
1076 return SDValue();
1077 }
1078
1079 //! Custom lower double precision floating point constants
1080 static SDValue
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();
1085
1086 if (VT == MVT::f64) {
1087 ConstantFPSDNode *FP = cast<ConstantFPSDNode>(Op.getNode());
1088
1089 assert((FP != 0) &&
1090 "LowerConstantFP: Node is not ConstantFPSDNode");
1091
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));
1097 }
1098
1099 return SDValue();
1100 }
1101
1102 SDValue
1103 SPUTargetLowering::LowerFormalArguments(SDValue Chain,
1104 CallingConv::ID CallConv, bool isVarArg,
1105 const SmallVectorImpl<ISD::InputArg>
1106 &Ins,
1107 DebugLoc dl, SelectionDAG &DAG,
1108 SmallVectorImpl<SDValue> &InVals)
1109 const {
1110
1111 MachineFunction &MF = DAG.getMachineFunction();
1112 MachineFrameInfo *MFI = MF.getFrameInfo();
1113 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1114 SPUFunctionInfo *FuncInfo = MF.getInfo<SPUFunctionInfo>();
1115
1116 unsigned ArgOffset = SPUFrameLowering::minStackSize();
1117 unsigned ArgRegIdx = 0;
1118 unsigned StackSlotSize = SPUFrameLowering::stackSlotSize();
1119
1120 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1121
1122 SmallVector<CCValAssign, 16> ArgLocs;
1123 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1124 *DAG.getContext());
1125 // FIXME: allow for other calling conventions
1126 CCInfo.AnalyzeFormalArguments(Ins, CCC_SPU);
1127
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;
1132 SDValue ArgVal;
1133 CCValAssign &VA = ArgLocs[ArgNo];
1134
1135 if (VA.isRegLoc()) {
1136 const TargetRegisterClass *ArgRegClass;
1137
1138 switch (ObjectVT.getSimpleVT().SimpleTy) {
1139 default:
1140 report_fatal_error("LowerFormalArguments Unhandled argument type: " +
1141 Twine(ObjectVT.getEVTString()));
1142 case MVT::i8:
1143 ArgRegClass = &SPU::R8CRegClass;
1144 break;
1145 case MVT::i16:
1146 ArgRegClass = &SPU::R16CRegClass;
1147 break;
1148 case MVT::i32:
1149 ArgRegClass = &SPU::R32CRegClass;
1150 break;
1151 case MVT::i64:
1152 ArgRegClass = &SPU::R64CRegClass;
1153 break;
1154 case MVT::i128:
1155 ArgRegClass = &SPU::GPRCRegClass;
1156 break;
1157 case MVT::f32:
1158 ArgRegClass = &SPU::R32FPRegClass;
1159 break;
1160 case MVT::f64:
1161 ArgRegClass = &SPU::R64FPRegClass;
1162 break;
1163 case MVT::v2f64:
1164 case MVT::v4f32:
1165 case MVT::v2i64:
1166 case MVT::v4i32:
1167 case MVT::v8i16:
1168 case MVT::v16i8:
1169 ArgRegClass = &SPU::VECREGRegClass;
1170 break;
1171 }
1172
1173 unsigned VReg = RegInfo.createVirtualRegister(ArgRegClass);
1174 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1175 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1176 ++ArgRegIdx;
1177 } else {
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(),
1184 false, false, 0);
1185 ArgOffset += StackSlotSize;
1186 }
1187
1188 InVals.push_back(ArgVal);
1189 // Update the chain
1190 Chain = ArgVal.getOperand(0);
1191 }
1192
1193 // vararg handling:
1194 if (isVarArg) {
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
1209 };
1210 // size of ArgRegs array
1211 unsigned NumArgRegs = 77;
1212
1213 // We will spill (79-3)+1 registers to the stack
1214 SmallVector<SDValue, 79-3+1> MemOps;
1215
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(),
1224 false, false, 0);
1225 Chain = Store.getOperand(0);
1226 MemOps.push_back(Store);
1227
1228 // Increment address by stack slot size for the next stored argument
1229 ArgOffset += StackSlotSize;
1230 }
1231 if (!MemOps.empty())
1232 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1233 &MemOps[0], MemOps.size());
1234 }
1235
1236 return Chain;
1237 }
1238
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);
1243 if (!C) return 0;
1244
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.
1249
1250 return DAG.getConstant((int)C->getZExtValue() >> 2, MVT::i32).getNode();
1251 }
1252
1253 SDValue
1254 SPUTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1255 CallingConv::ID CallConv, bool isVarArg,
1256 bool &isTailCall,
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.
1263 isTailCall = false;
1264
1265 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
1266 unsigned NumOps = Outs.size();
1267 unsigned StackSlotSize = SPUFrameLowering::stackSlotSize();
1268
1269 SmallVector<CCValAssign, 16> ArgLocs;
1270 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs,
1271 *DAG.getContext());
1272 // FIXME: allow for other calling conventions
1273 CCInfo.AnalyzeCallOperands(Outs, CCC_SPU);
1274
1275 const unsigned NumArgRegs = ArgLocs.size();
1276
1277
1278 // Handy pointer type
1279 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1280
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
1283 // passing.
1284 SDValue StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
1285
1286 // Figure out which arguments are going to go in registers, and which in
1287 // memory.
1288 unsigned ArgOffset = SPUFrameLowering::minStackSize(); // Just below [LR]
1289 unsigned ArgRegIdx = 0;
1290
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;
1295
1296 for (; ArgRegIdx != NumOps; ++ArgRegIdx) {
1297 SDValue Arg = OutVals[ArgRegIdx];
1298 CCValAssign &VA = ArgLocs[ArgRegIdx];
1299
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);
1304
1305 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
1306 default: llvm_unreachable("Unexpected ValueType for argument!");
1307 case MVT::i8:
1308 case MVT::i16:
1309 case MVT::i32:
1310 case MVT::i64:
1311 case MVT::i128:
1312 case MVT::f32:
1313 case MVT::f64:
1314 case MVT::v2i64:
1315 case MVT::v2f64:
1316 case MVT::v4f32:
1317 case MVT::v4i32:
1318 case MVT::v8i16:
1319 case MVT::v16i8:
1320 if (ArgRegIdx != NumArgRegs) {
1321 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1322 } else {
1323 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
1324 MachinePointerInfo(),
1325 false, false, 0));
1326 ArgOffset += StackSlotSize;
1327 }
1328 break;
1329 }
1330 }
1331
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 - SPUFrameLowering::minStackSize();
1336
1337 // Insert a call sequence start
1338 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumStackBytes,
1339 true));
1340
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());
1345 }
1346
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.
1349 SDValue InFlag;
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);
1354 }
1355
1356 SmallVector<SDValue, 8> Ops;
1357 unsigned CallOpc = SPUISD::CALL;
1358
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);
1367
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.
1373 //
1374 // NOTE:
1375 // This may be an unsafe assumption for JIT and really large compilation
1376 // units.
1377 if (GV->isDeclaration()) {
1378 Callee = DAG.getNode(SPUISD::AFormAddr, dl, CalleeVT, GA, Zero);
1379 } else {
1380 Callee = DAG.getNode(SPUISD::PCRelAddr, dl, CalleeVT, GA, Zero);
1381 }
1382 } else {
1383 // "Large memory" mode: Turn all calls into indirect calls with a X-form
1384 // address pairs:
1385 Callee = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, GA, Zero);
1386 }
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());
1392
1393 if (!ST->usingLargeMem()) {
1394 Callee = DAG.getNode(SPUISD::AFormAddr, dl, CalleeVT, ExtSym, Zero);
1395 } else {
1396 Callee = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, ExtSym, Zero);
1397 }
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);
1402 }
1403
1404 Ops.push_back(Chain);
1405 Ops.push_back(Callee);
1406
1407 // Add argument registers to the end of the list so that they are known live
1408 // into the call.
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()));
1412
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::Glue),
1417 &Ops[0], Ops.size());
1418 InFlag = Chain.getValue(1);
1419
1420 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumStackBytes, true),
1421 DAG.getIntPtrConstant(0, true), InFlag);
1422 if (!Ins.empty())
1423 InFlag = Chain.getValue(1);
1424
1425 // If the function returns void, just return the chain.
1426 if (Ins.empty())
1427 return Chain;
1428
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);
1434
1435
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];
1439
1440 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1441 InFlag);
1442 Chain = Val.getValue(1);
1443 InFlag = Val.getValue(2);
1444 InVals.push_back(Val);
1445 }
1446
1447 return Chain;
1448 }
1449
1450 SDValue
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 {
1456
1457 SmallVector<CCValAssign, 16> RVLocs;
1458 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1459 RVLocs, *DAG.getContext());
1460 CCInfo.AnalyzeReturn(Outs, RetCC_SPU);
1461
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());
1467 }
1468
1469 SDValue Flag;
1470
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(),
1476 OutVals[i], Flag);
1477 Flag = Chain.getValue(1);
1478 }
1479
1480 if (Flag.getNode())
1481 return DAG.getNode(SPUISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1482 else
1483 return DAG.getNode(SPUISD::RET_FLAG, dl, MVT::Other, Chain);
1484 }
1485
1486
1487 //===----------------------------------------------------------------------===//
1488 // Vector related lowering:
1489 //===----------------------------------------------------------------------===//
1490
1491 static ConstantSDNode *
1492 getVecImm(SDNode *N) {
1493 SDValue OpVal(0, 0);
1494
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))
1501 return 0;
1502 }
1503
1504 if (OpVal.getNode() != 0) {
1505 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1506 return CN;
1507 }
1508 }
1509
1510 return 0;
1511 }
1512
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
1515 /// constant
1516 SDValue SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
1517 EVT ValueType) {
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);
1524 if (upper != lower)
1525 return SDValue();
1526 Value = Value >> 32;
1527 }
1528 if (Value <= 0x3ffff)
1529 return DAG.getTargetConstant(Value, ValueType);
1530 }
1531
1532 return SDValue();
1533 }
1534
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
1537 /// constant
1538 SDValue SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
1539 EVT ValueType) {
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);
1546 if (upper != lower)
1547 return SDValue();
1548 Value = Value >> 32;
1549 }
1550 if (Value >= -(1 << 15) && Value <= ((1 << 15) - 1)) {
1551 return DAG.getTargetConstant(Value, ValueType);
1552 }
1553 }
1554
1555 return SDValue();
1556 }
1557
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
1560 /// constant
1561 SDValue SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
1562 EVT ValueType) {
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);
1569 if (upper != lower)
1570 return SDValue();
1571 Value = Value >> 32;
1572 }
1573 if (isInt<10>(Value))
1574 return DAG.getTargetConstant(Value, ValueType);
1575 }
1576
1577 return SDValue();
1578 }
1579
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
1582 /// constant.
1583 ///
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
1586 /// same value.
1587 SDValue SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
1588 EVT ValueType) {
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);
1598 }
1599
1600 return SDValue();
1601 }
1602
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
1605 /// constant
1606 SDValue SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
1607 EVT ValueType) {
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);
1614 }
1615
1616 return SDValue();
1617 }
1618
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);
1623 }
1624
1625 return SDValue();
1626 }
1627
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);
1632 }
1633
1634 return SDValue();
1635 }
1636
1637 //! Lower a BUILD_VECTOR instruction creatively:
1638 static SDValue
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();
1646
1647 if (minSplatBits < 16)
1648 minSplatBits = 16;
1649
1650 APInt APSplatBits, APSplatUndef;
1651 unsigned SplatBitSize;
1652 bool HasAnyUndefs;
1653
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
1658
1659 uint64_t SplatBits = APSplatBits.getZExtValue();
1660
1661 switch (VT.getSimpleVT().SimpleTy) {
1662 default:
1663 report_fatal_error("CellSPU: Unhandled VT in LowerBUILD_VECTOR, VT = " +
1664 Twine(VT.getEVTString()));
1665 /*NOTREACHED*/
1666 case MVT::v4f32: {
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));
1674 break;
1675 }
1676 case MVT::v2f64: {
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));
1684 break;
1685 }
1686 case MVT::v16i8: {
1687 // 8-bit constants have to be expanded to 16-bits
1688 unsigned short Value16 = SplatBits /* | (SplatBits << 8) */;
1689 SmallVector<SDValue, 8> Ops;
1690
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()));
1694 }
1695 case MVT::v8i16: {
1696 unsigned short Value16 = SplatBits;
1697 SDValue T = DAG.getConstant(Value16, EltVT);
1698 SmallVector<SDValue, 8> Ops;
1699
1700 Ops.assign(8, T);
1701 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
1702 }
1703 case MVT::v4i32: {
1704 SDValue T = DAG.getConstant(unsigned(SplatBits), VT.getVectorElementType());
1705 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, T, T, T, T);
1706 }
1707 case MVT::v2i64: {
1708 return SPU::LowerV2I64Splat(VT, DAG, SplatBits, dl);
1709 }
1710 }
1711
1712 return SDValue();
1713 }
1714
1715 /*!
1716 */
1717 SDValue
1718 SPU::LowerV2I64Splat(EVT OpVT, SelectionDAG& DAG, uint64_t SplatVal,
1719 DebugLoc dl) {
1720 uint32_t upper = uint32_t(SplatVal >> 32);
1721 uint32_t lower = uint32_t(SplatVal);
1722
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));
1729 } else {
1730 bool upper_special, lower_special;
1731
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.
1735
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);
1739
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);
1745 }
1746
1747 SDValue LO32;
1748 SDValue HI32;
1749 SmallVector<SDValue, 16> ShufBytes;
1750 SDValue Result;
1751
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));
1758 }
1759
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));
1766 }
1767
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")
1770 if (lower_special)
1771 LO32 = HI32;
1772 if (upper_special)
1773 HI32 = LO32;
1774
1775 for (int i = 0; i < 4; ++i) {
1776 uint64_t val = 0;
1777 for (int j = 0; j < 4; ++j) {
1778 SDValue V;
1779 bool process_upper, process_lower;
1780 val <<= 8;
1781 process_upper = (upper_special && (i & 1) == 0);
1782 process_lower = (lower_special && (i & 1) == 1);
1783
1784 if (process_upper || process_lower) {
1785 if ((process_upper && upper == 0)
1786 || (process_lower && lower == 0))
1787 val |= 0x80;
1788 else if ((process_upper && upper == 0xffffffff)
1789 || (process_lower && lower == 0xffffffff))
1790 val |= 0xc0;
1791 else if ((process_upper && upper == 0x80000000)
1792 || (process_lower && lower == 0x80000000))
1793 val |= (j == 0 ? 0xe0 : 0x80);
1794 } else
1795 val |= i * 4 + j + ((i & 1) * 16);
1796 }
1797
1798 ShufBytes.push_back(DAG.getConstant(val, MVT::i32));
1799 }
1800
1801 return DAG.getNode(SPUISD::SHUFB, dl, OpVT, HI32, LO32,
1802 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1803 &ShufBytes[0], ShufBytes.size()));
1804 }
1805 }
1806
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.
1814 /// \note
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.
1818 /// \note
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();
1825
1826 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1827
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;
1841 bool rotate = true;
1842 int rotamt=0;
1843 EVT maskVT; // which of the c?d instructions to use
1844
1845 if (EltVT == MVT::i8) {
1846 V2EltIdx0 = 16;
1847 maskVT = MVT::v16i8;
1848 } else if (EltVT == MVT::i16) {
1849 V2EltIdx0 = 8;
1850 maskVT = MVT::v8i16;
1851 } else if (EltVT == MVT::i32 || EltVT == MVT::f32) {
1852 V2EltIdx0 = 4;
1853 maskVT = MVT::v4i32;
1854 } else if (EltVT == MVT::i64 || EltVT == MVT::f64) {
1855 V2EltIdx0 = 2;
1856 maskVT = MVT::v2i64;
1857 } else
1858 llvm_unreachable("Unhandled vector type in LowerVECTOR_SHUFFLE");
1859
1860 for (unsigned i = 0; i != MaxElts; ++i) {
1861 if (SVN->getMaskElt(i) < 0)
1862 continue;
1863
1864 unsigned SrcElt = SVN->getMaskElt(i);
1865
1866 if (monotonic) {
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);
1872 else
1873 monotonic = false;
1874 ++EltsFromV2;
1875 } else if (CurrElt != SrcElt) {
1876 monotonic = false;
1877 }
1878
1879 ++CurrElt;
1880 }
1881
1882 if (rotate) {
1883 if (PrevElt > 0 && SrcElt < MaxElts) {
1884 if ((PrevElt == SrcElt - 1)
1885 || (PrevElt == MaxElts - 1 && SrcElt == 0)) {
1886 PrevElt = SrcElt;
1887 } else {
1888 rotate = false;
1889 }
1890 } else if (i == 0 || (PrevElt==0 && SrcElt==1)) {
1891 // First time or after a "wrap around"
1892 rotamt = SrcElt-i;
1893 PrevElt = SrcElt;
1894 } else {
1895 // This isn't a rotation, takes elements from vector 2
1896 rotate = false;
1897 }
1898 }
1899 }
1900
1901 if (EltsFromV2 == 1 && monotonic) {
1902 // Compute mask and shuffle
1903 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1904
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,
1911 maskVT, Pointer);
1912
1913 // Use shuffle mask in SHUFB synthetic instruction:
1914 return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V2, V1,
1915 ShufMaskOp);
1916 } else if (rotate) {
1917 if (rotamt < 0)
1918 rotamt +=MaxElts;
1919 rotamt *= EltVT.getSizeInBits()/8;
1920 return DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, V1.getValueType(),
1921 V1, DAG.getConstant(rotamt, MVT::i16));
1922 } else {
1923 // Convert the SHUFFLE_VECTOR mask's input element units to the
1924 // actual bytes.
1925 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
1926
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);
1930
1931 for (unsigned j = 0; j < BytesPerElement; ++j)
1932 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,MVT::i8));
1933 }
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);
1937 }
1938 }
1939
1940 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
1941 SDValue Op0 = Op.getOperand(0); // Op0 = the scalar
1942 DebugLoc dl = Op.getDebugLoc();
1943
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.
1947
1948 ConstantSDNode *CN = cast<ConstantSDNode>(Op0.getNode());
1949 SmallVector<SDValue, 16> ConstVecValues;
1950 EVT VT;
1951 size_t n_copies;
1952
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;
1963 }
1964
1965 SDValue CValue = DAG.getConstant(CN->getZExtValue(), VT);
1966 for (size_t j = 0; j < n_copies; ++j)
1967 ConstVecValues.push_back(CValue);
1968
1969 return DAG.getNode(ISD::BUILD_VECTOR, dl, Op.getValueType(),
1970 &ConstVecValues[0], ConstVecValues.size());
1971 } else {
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");
1975 case MVT::i8:
1976 case MVT::i16:
1977 case MVT::i32:
1978 case MVT::i64:
1979 case MVT::f32:
1980 case MVT::f64:
1981 return DAG.getNode(SPUISD::PREFSLOT2VEC, dl, Op.getValueType(), Op0, Op0);
1982 }
1983 }
1984
1985 return SDValue();
1986 }
1987
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();
1993 SDValue retval;
1994
1995 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
1996 // Constant argument:
1997 int EltNo = (int) C->getZExtValue();
1998
1999 // sanity checks:
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");
2008
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);
2012 }
2013
2014 // Need to generate shuffle mask and extract:
2015 int prefslot_begin = -1, prefslot_end = -1;
2016 int elt_byte = EltNo * VT.getSizeInBits() / 8;
2017
2018 switch (VT.getSimpleVT().SimpleTy) {
2019 default:
2020 assert(false && "Invalid value type!");
2021 case MVT::i8: {
2022 prefslot_begin = prefslot_end = 3;
2023 break;
2024 }
2025 case MVT::i16: {
2026 prefslot_begin = 2; prefslot_end = 3;
2027 break;
2028 }
2029 case MVT::i32:
2030 case MVT::f32: {
2031 prefslot_begin = 0; prefslot_end = 3;
2032 break;
2033 }
2034 case MVT::i64:
2035 case MVT::f64: {
2036 prefslot_begin = 0; prefslot_end = 7;
2037 break;
2038 }
2039 }
2040
2041 assert(prefslot_begin != -1 && prefslot_end != -1 &&
2042 "LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
2043
2044 unsigned int ShufBytes[16] = {
2045 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
2046 };
2047 for (int i = 0; i < 16; ++i) {
2048 // zero fill uppper part of preferred slot, don't care about the
2049 // other slots:
2050 unsigned int mask_val;
2051 if (i <= prefslot_end) {
2052 mask_val =
2053 ((i < prefslot_begin)
2054 ? 0x80
2055 : elt_byte + (i - prefslot_begin));
2056
2057 ShufBytes[i] = mask_val;
2058 } else
2059 ShufBytes[i] = ShufBytes[i % (prefslot_end + 1)];
2060 }
2061
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) |
2068 ShufBytes[bidx+3]);
2069 ShufMask[i] = DAG.getConstant(bits, MVT::i32);
2070 }
2071
2072 SDValue ShufMaskVec =
2073 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2074 &ShufMask[0], sizeof(ShufMask)/sizeof(ShufMask[0]));
2075
2076 retval = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
2077 DAG.getNode(SPUISD::SHUFB, dl, N.getValueType(),
2078 N, N, ShufMaskVec));
2079 } else {
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"
2085 "vector type!");
2086 }
2087
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);
2091
2092 // Scale the index to a bit/byte shift quantity
2093 APInt scaleFactor =
2094 APInt(32, uint64_t(16 / N.getValueType().getVectorNumElements()), false);
2095 unsigned scaleShift = scaleFactor.logBase2();
2096 SDValue vecShift;
2097
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));
2102 }
2103
2104 vecShift = DAG.getNode(SPUISD::SHL_BYTES, dl, VecVT, N, Elt);
2105
2106 // Replicate the bytes starting at byte 0 across the entire vector (for
2107 // consistency with the notion of a unified register set)
2108 SDValue replicate;
2109
2110 switch (VT.getSimpleVT().SimpleTy) {
2111 default:
2112 report_fatal_error("LowerEXTRACT_VECTOR_ELT(varable): Unhandled vector"
2113 "type");
2114 /*NOTREACHED*/
2115 case MVT::i8: {
2116 SDValue factor = DAG.getConstant(0x00000000, MVT::i32);
2117 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2118 factor, factor, factor, factor);
2119 break;
2120 }
2121 case MVT::i16: {
2122 SDValue factor = DAG.getConstant(0x00010001, MVT::i32);
2123 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2124 factor, factor, factor, factor);
2125 break;
2126 }
2127 case MVT::i32:
2128 case MVT::f32: {
2129 SDValue factor = DAG.getConstant(0x00010203, MVT::i32);
2130 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2131 factor, factor, factor, factor);
2132 break;
2133 }
2134 case MVT::i64:
2135 case MVT::f64: {
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);
2140 break;
2141 }
2142 }
2143
2144 retval = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
2145 DAG.getNode(SPUISD::SHUFB, dl, VecVT,
2146 vecShift, vecShift, replicate));
2147 }
2148
2149 return retval;
2150 }
2151
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();
2159
2160 // use 0 when the lane to insert to is 'undef'
2161 int64_t Offset=0;
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;
2166 }
2167
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);
2177
2178 SDValue result =
2179 DAG.getNode(SPUISD::SHUFB, dl, VT,
2180 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, ValOp),
2181 VecOp,
2182 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, ShufMask));
2183
2184 return result;
2185 }
2186
2187 static SDValue LowerI8Math(SDValue Op, SelectionDAG &DAG, unsigned Opc,
2188 const TargetLowering &TLI)
2189 {
2190 SDValue N0 = Op.getOperand(0); // Everything has at least one operand
2191 DebugLoc dl = Op.getDebugLoc();
2192 EVT ShiftVT = TLI.getShiftAmountTy(N0.getValueType());
2193
2194 assert(Op.getValueType() == MVT::i8);
2195 switch (Opc) {
2196 default:
2197 llvm_unreachable("Unhandled i8 math operator");
2198 /*NOTREACHED*/
2199 break;
2200 case ISD::ADD: {
2201 // 8-bit addition: Promote the arguments up to 16-bits and truncate
2202 // the result:
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));
2208
2209 }
2210
2211 case ISD::SUB: {
2212 // 8-bit subtraction: Promote the arguments up to 16-bits and truncate
2213 // the result:
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));
2219 }
2220 case ISD::ROTR:
2221 case ISD::ROTL: {
2222 SDValue N1 = Op.getOperand(1);
2223 EVT N1VT = N1.getValueType();
2224
2225 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, N0);
2226 if (!N1VT.bitsEq(ShiftVT)) {
2227 unsigned N1Opc = N1.getValueType().bitsLT(ShiftVT)
2228 ? ISD::ZERO_EXTEND
2229 : ISD::TRUNCATE;
2230 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2231 }
2232
2233 // Replicate lower 8-bits into upper 8:
2234 SDValue ExpandArg =
2235 DAG.getNode(ISD::OR, dl, MVT::i16, N0,
2236 DAG.getNode(ISD::SHL, dl, MVT::i16,
2237 N0, DAG.getConstant(8, MVT::i32)));
2238
2239 // Truncate back down to i8
2240 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2241 DAG.getNode(Opc, dl, MVT::i16, ExpandArg, N1));
2242 }
2243 case ISD::SRL:
2244 case ISD::SHL: {
2245 SDValue N1 = Op.getOperand(1);
2246 EVT N1VT = N1.getValueType();
2247
2248 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, N0);
2249 if (!N1VT.bitsEq(ShiftVT)) {
2250 unsigned N1Opc = ISD::ZERO_EXTEND;
2251
2252 if (N1.getValueType().bitsGT(ShiftVT))
2253 N1Opc = ISD::TRUNCATE;
2254
2255 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2256 }
2257
2258 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2259 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2260 }
2261 case ISD::SRA: {
2262 SDValue N1 = Op.getOperand(1);
2263 EVT N1VT = N1.getValueType();
2264
2265 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2266 if (!N1VT.bitsEq(ShiftVT)) {
2267 unsigned N1Opc = ISD::SIGN_EXTEND;
2268
2269 if (N1VT.bitsGT(ShiftVT))
2270 N1Opc = ISD::TRUNCATE;
2271 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2272 }
2273
2274 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2275 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2276 }
2277 case ISD::MUL: {
2278 SDValue N1 = Op.getOperand(1);
2279
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));
2284 break;
2285 }
2286 }
2287
2288 return SDValue();
2289 }
2290
2291 //! Lower byte immediate operations for v16i8 vectors:
2292 static SDValue
2293 LowerByteImmed(SDValue Op, SelectionDAG &DAG) {
2294 SDValue ConstVec;
2295 SDValue Arg;
2296 EVT VT = Op.getValueType();
2297 DebugLoc dl = Op.getDebugLoc();
2298
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);
2304 } else {
2305 ConstVec = Op.getOperand(1);
2306 Arg = Op.getOperand(0);
2307 if (ConstVec.getNode()->getOpcode() == ISD::BITCAST) {
2308 ConstVec = ConstVec.getOperand(0);
2309 }
2310 }
2311 }
2312
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");
2316
2317 APInt APSplatBits, APSplatUndef;
2318 unsigned SplatBitSize;
2319 bool HasAnyUndefs;
2320 unsigned minSplatBits = VT.getVectorElementType().getSizeInBits();
2321
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);
2327
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()));
2332 }
2333 }
2334
2335 // These operations (AND, OR, XOR) are legal, they just couldn't be custom
2336 // lowered. Return the operation, rather than a null SDValue.
2337 return Op;
2338 }
2339
2340 //! Custom lowering for CTPOP (count population)
2341 /*!
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.
2345 */
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();
2351
2352 switch (VT.getSimpleVT().SimpleTy) {
2353 default:
2354 assert(false && "Invalid value type!");
2355 case MVT::i8: {
2356 SDValue N = Op.getOperand(0);
2357 SDValue Elt0 = DAG.getConstant(0, MVT::i32);
2358
2359 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2360 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2361
2362 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i8, CNTB, Elt0);
2363 }
2364
2365 case MVT::i16: {
2366 MachineFunction &MF = DAG.getMachineFunction();
2367 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2368
2369 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
2370
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);
2375
2376 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2377 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2378
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);
2383
2384 SDValue CNTB_rescopy =
2385 DAG.getCopyToReg(CNTB_result, dl, CNTB_reg, CNTB_result);
2386
2387 SDValue Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i16);
2388
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,
2392 Tmp1, Shift1),
2393 Tmp1),
2394 Mask0);
2395 }
2396
2397 case MVT::i32: {
2398 MachineFunction &MF = DAG.getMachineFunction();
2399 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2400
2401 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2402 unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2403
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);
2409
2410 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2411 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2412
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);
2417
2418 SDValue CNTB_rescopy =
2419 DAG.getCopyToReg(CNTB_result, dl, CNTB_reg, CNTB_result);
2420
2421 SDValue Comp1 =
2422 DAG.getNode(ISD::SRL, dl, MVT::i32,
2423 DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i32),
2424 Shift1);
2425
2426 SDValue Sum1 =
2427 DAG.getNode(ISD::ADD, dl, MVT::i32, Comp1,
2428 DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i32));
2429
2430 SDValue Sum1_rescopy =
2431 DAG.getCopyToReg(CNTB_result, dl, SUM1_reg, Sum1);
2432
2433 SDValue Comp2 =
2434 DAG.getNode(ISD::SRL, dl, MVT::i32,
2435 DAG.getCopyFromReg(Sum1_rescopy, dl, SUM1_reg, MVT::i32),
2436 Shift2);
2437 SDValue Sum2 =
2438 DAG.getNode(ISD::ADD, dl, MVT::i32, Comp2,
2439 DAG.getCopyFromReg(Sum1_rescopy, dl, SUM1_reg, MVT::i32));
2440
2441 return DAG.getNode(ISD::AND, dl, MVT::i32, Sum2, Mask0);
2442 }
2443
2444 case MVT::i64:
2445 break;
2446 }
2447
2448 return SDValue();
2449 }
2450
2451 //! Lower ISD::FP_TO_SINT, ISD::FP_TO_UINT for i32
2452 /*!
2453 f32->i32 passes through unchanged, whereas f64->i32 expands to a libcall.
2454 All conversions to i64 are expanded to a libcall.
2455 */
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();
2461
2462 if ((OpVT == MVT::i32 && Op0VT == MVT::f64)
2463 || OpVT == MVT::i64) {
2464 // Convert f32 / f64 to i32 / i64 via libcall.
2465 RTLIB::Libcall LC =
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!");
2470 SDValue Dummy;
2471 return ExpandLibCall(LC, Op, DAG, false, Dummy, TLI);
2472 }
2473
2474 return Op;
2475 }
2476
2477 //! Lower ISD::SINT_TO_FP, ISD::UINT_TO_FP for i32
2478 /*!
2479 i32->f32 passes through unchanged, whereas i32->f64 is expanded to a libcall.
2480 All conversions from i64 are expanded to a libcall.
2481 */
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();
2487
2488 if ((OpVT == MVT::f64 && Op0VT == MVT::i32)
2489 || Op0VT == MVT::i64) {
2490 // Convert i32, i64 to f64 via libcall:
2491 RTLIB::Libcall LC =
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!");
2496 SDValue Dummy;
2497 return ExpandLibCall(LC, Op, DAG, false, Dummy, TLI);
2498 }
2499
2500 return Op;
2501 }
2502
2503 //! Lower ISD::SETCC
2504 /*!
2505 This handles MVT::f64 (double floating point) condition lowering
2506 */
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");
2512
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");
2517
2518 EVT ccResultVT = TLI.getSetCCResultType(lhs.getValueType());
2519 APInt ccResultOnes = APInt::getAllOnesValue(ccResultVT.getSizeInBits());
2520 EVT IntVT(MVT::i64);
2521
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);
2525 SDValue lhsHi32 =
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));
2532 SDValue lhsLo32 =
2533 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, i64lhs);
2534
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
2538 // SETUO
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),
2543 ISD::SETUO),
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,
2549 lhsHi32abs,
2550 DAG.getConstant(0x7ff00000, MVT::i32),
2551 ISD::SETGE),
2552 DAG.getSetCC(dl, ccResultVT,
2553 lhsLo32,
2554 DAG.getConstant(0, MVT::i32),
2555 ISD::SETGT));
2556 }
2557
2558 SDValue i64rhs = DAG.getNode(ISD::BITCAST, dl, IntVT, rhs);
2559 SDValue rhsHi32 =
2560 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
2561 DAG.getNode(ISD::SRL, dl, IntVT,
2562 i64rhs, DAG.getConstant(32, MVT::i32)));
2563
2564 // If a value is negative, subtract from the sign magnitude constant:
2565 SDValue signMag2TC = DAG.getConstant(0x8000000000000000ULL, IntVT);
2566
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);
2571 SDValue lhsSelect =
2572 DAG.getNode(ISD::SELECT, dl, IntVT,
2573 lhsSelectMask, lhsSignMag2TC, i64lhs);
2574
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);
2578 SDValue rhsSelect =
2579 DAG.getNode(ISD::SELECT, dl, IntVT,
2580 rhsSelectMask, rhsSignMag2TC, i64rhs);
2581
2582 unsigned compareOp;
2583
2584 switch (CC->get()) {
2585 case ISD::SETOEQ:
2586 case ISD::SETUEQ:
2587 compareOp = ISD::SETEQ; break;
2588 case ISD::SETOGT:
2589 case ISD::SETUGT:
2590 compareOp = ISD::SETGT; break;
2591 case ISD::SETOGE:
2592 case ISD::SETUGE:
2593 compareOp = ISD::SETGE; break;
2594 case ISD::SETOLT:
2595 case ISD::SETULT:
2596 compareOp = ISD::SETLT; break;
2597 case ISD::SETOLE:
2598 case ISD::SETULE:
2599 compareOp = ISD::SETLE; break;
2600 case ISD::SETUNE:
2601 case ISD::SETONE:
2602 compareOp = ISD::SETNE; break;
2603 default:
2604 report_fatal_error("CellSPU ISel Select: unimplemented f64 condition");
2605 }
2606
2607 SDValue result =
2608 DAG.getSetCC(dl, ccResultVT, lhsSelect, rhsSelect,
2609 (ISD::CondCode) compareOp);
2610
2611 if ((CC->get() & 0x8) == 0) {
2612 // Ordered comparison:
2613 SDValue lhsNaN = DAG.getSetCC(dl, ccResultVT,
2614 lhs, DAG.getConstantFP(0.0, MVT::f64),
2615 ISD::SETO);
2616 SDValue rhsNaN = DAG.getSetCC(dl, ccResultVT,
2617 rhs, DAG.getConstantFP(0.0, MVT::f64),
2618 ISD::SETO);
2619 SDValue ordered = DAG.getNode(ISD::AND, dl, ccResultVT, lhsNaN, rhsNaN);
2620
2621 result = DAG.getNode(ISD::AND, dl, ccResultVT, ordered, result);
2622 }
2623
2624 return result;
2625 }
2626
2627 //! Lower ISD::SELECT_CC
2628 /*!
2629 ISD::SELECT_CC can (generally) be implemented directly on the SPU using the
2630 SELB instruction.
2631
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.
2637 */
2638
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();
2648
2649 // NOTE: SELB's arguments: $rA, $rB, $mask
2650 //
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.
2655
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:
2659
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);
2664 }
2665
2666 //! Custom lower ISD::TRUNCATE
2667 static SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG)
2668 {
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();
2675
2676 // Type to truncate from
2677 SDValue Op0 = Op.getOperand(0);
2678 EVT Op0VT = Op0.getValueType();
2679
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));
2690
2691 SDValue truncShuffle = DAG.getNode(SPUISD::SHUFB, dl, VecVT,
2692 Op0, Op0, shufMask);
2693
2694 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT, truncShuffle);
2695 }
2696
2697 return SDValue(); // Leave the truncate unmolested
2698 }
2699
2700 /*!
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.
2705 *
2706 * @param Op The sext operand
2707 * @param DAG The current DAG
2708 * @return The SDValue with the entire instruction sequence
2709 */
2710 static SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG)
2711 {
2712 DebugLoc dl = Op.getDebugLoc();
2713
2714 // Type to extend to
2715 MVT OpVT = Op.getValueType().getSimpleVT();
2716
2717 // Type to extend from
2718 SDValue Op0 = Op.getOperand(0);
2719 MVT Op0VT = Op0.getValueType().getSimpleVT();
2720
2721 // extend i8 & i16 via i32
2722 if (Op0VT == MVT::i8 || Op0VT == MVT::i16) {
2723 Op0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, Op0);
2724 Op0VT = MVT::i32;
2725 }
2726
2727 // The type to extend to needs to be a i128 and
2728 // the type to extend from needs to be i64 or i32.
2729 assert((OpVT == MVT::i128 && (Op0VT == MVT::i64 || Op0VT == MVT::i32)) &&
2730 "LowerSIGN_EXTEND: input and/or output operand have wrong size");
2731
2732 // Create shuffle mask
2733 unsigned mask1 = 0x10101010; // byte 0 - 3 and 4 - 7
2734 unsigned mask2 = Op0VT == MVT::i64 ? 0x00010203 : 0x10101010; // byte 8 - 11
2735 unsigned mask3 = Op0VT == MVT::i64 ? 0x04050607 : 0x00010203; // byte 12 - 15
2736 SDValue shufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2737 DAG.getConstant(mask1, MVT::i32),
2738 DAG.getConstant(mask1, MVT::i32),
2739 DAG.getConstant(mask2, MVT::i32),
2740 DAG.getConstant(mask3, MVT::i32));
2741
2742 // Word wise arithmetic right shift to generate at least one byte
2743 // that contains sign bits.
2744 MVT mvt = Op0VT == MVT::i64 ? MVT::v2i64 : MVT::v4i32;
2745 SDValue sraVal = DAG.getNode(ISD::SRA,
2746 dl,
2747 mvt,
2748 DAG.getNode(SPUISD::PREFSLOT2VEC, dl, mvt, Op0, Op0),
2749 DAG.getConstant(31, MVT::i32));
2750
2751 // reinterpret as a i128 (SHUFB requires it). This gets lowered away.
2752 SDValue extended = SDValue(DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
2753 dl, Op0VT, Op0,
2754 DAG.getTargetConstant(
2755 SPU::GPRCRegClass.getID(),
2756 MVT::i32)), 0);
2757 // Shuffle bytes - Copy the sign bits into the upper 64 bits
2758 // and the input value into the lower 64 bits.
2759 SDValue extShuffle = DAG.getNode(SPUISD::SHUFB, dl, mvt,
2760 extended, sraVal, shufMask);
2761 return DAG.getNode(ISD::BITCAST, dl, MVT::i128, extShuffle);
2762 }
2763
2764 //! Custom (target-specific) lowering entry point
2765 /*!
2766 This is where LLVM's DAG selection process calls to do target-specific
2767 lowering of nodes.
2768 */
2769 SDValue
2770 SPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
2771 {
2772 unsigned Opc = (unsigned) Op.getOpcode();
2773 EVT VT = Op.getValueType();
2774
2775 switch (Opc) {
2776 default: {
2777 #ifndef NDEBUG
2778 errs() << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
2779 errs() << "Op.getOpcode() = " << Opc << "\n";
2780 errs() << "*Op.getNode():\n";
2781 Op.getNode()->dump();
2782 #endif
2783 llvm_unreachable(0);
2784 }
2785 case ISD::LOAD:
2786 case ISD::EXTLOAD:
2787 case ISD::SEXTLOAD:
2788 case ISD::ZEXTLOAD:
2789 return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
2790 case ISD::STORE:
2791 return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
2792 case ISD::ConstantPool:
2793 return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
2794 case ISD::GlobalAddress:
2795 return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
2796 case ISD::JumpTable:
2797 return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
2798 case ISD::ConstantFP:
2799 return LowerConstantFP(Op, DAG);
2800
2801 // i8, i64 math ops:
2802 case ISD::ADD:
2803 case ISD::SUB:
2804 case ISD::ROTR:
2805 case ISD::ROTL:
2806 case ISD::SRL:
2807 case ISD::SHL:
2808 case ISD::SRA: {
2809 if (VT == MVT::i8)
2810 return LowerI8Math(Op, DAG, Opc, *this);
2811 break;
2812 }
2813
2814 case ISD::FP_TO_SINT:
2815 case ISD::FP_TO_UINT:
2816 return LowerFP_TO_INT(Op, DAG, *this);
2817
2818 case ISD::SINT_TO_FP:
2819 case ISD::UINT_TO_FP:
2820 return LowerINT_TO_FP(Op, DAG, *this);
2821
2822 // Vector-related lowering.
2823 case ISD::BUILD_VECTOR:
2824 return LowerBUILD_VECTOR(Op, DAG);
2825 case ISD::SCALAR_TO_VECTOR:
2826 return LowerSCALAR_TO_VECTOR(Op, DAG);
2827 case ISD::VECTOR_SHUFFLE:
2828 return LowerVECTOR_SHUFFLE(Op, DAG);
2829 case ISD::EXTRACT_VECTOR_ELT:
2830 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2831 case ISD::INSERT_VECTOR_ELT:
2832 return LowerINSERT_VECTOR_ELT(Op, DAG);
2833
2834 // Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
2835 case ISD::AND:
2836 case ISD::OR:
2837 case ISD::XOR:
2838 return LowerByteImmed(Op, DAG);
2839
2840 // Vector and i8 multiply:
2841 case ISD::MUL:
2842 if (VT == MVT::i8)
2843 return LowerI8Math(Op, DAG, Opc, *this);
2844
2845 case ISD::CTPOP:
2846 return LowerCTPOP(Op, DAG);
2847
2848 case ISD::SELECT_CC:
2849 return LowerSELECT_CC(Op, DAG, *this);
2850
2851 case ISD::SETCC:
2852 return LowerSETCC(Op, DAG, *this);
2853
2854 case ISD::TRUNCATE:
2855 return LowerTRUNCATE(Op, DAG);
2856
2857 case ISD::SIGN_EXTEND:
2858 return LowerSIGN_EXTEND(Op, DAG);
2859 }
2860
2861 return SDValue();
2862 }
2863
2864 void SPUTargetLowering::ReplaceNodeResults(SDNode *N,
2865 SmallVectorImpl<SDValue>&Results,
2866 SelectionDAG &DAG) const
2867 {
2868 #if 0
2869 unsigned Opc = (unsigned) N->getOpcode();
2870 EVT OpVT = N->getValueType(0);
2871
2872 switch (Opc) {
2873 default: {
2874 errs() << "SPUTargetLowering::ReplaceNodeResults(): need to fix this!\n";
2875 errs() << "Op.getOpcode() = " << Opc << "\n";
2876 errs() << "*Op.getNode():\n";
2877 N->dump();
2878 abort();
2879 /*NOTREACHED*/
2880 }
2881 }
2882 #endif
2883
2884 /* Otherwise, return unchanged */
2885 }
2886
2887 //===----------------------------------------------------------------------===//
2888 // Target Optimization Hooks
2889 //===----------------------------------------------------------------------===//
2890
2891 SDValue
2892 SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
2893 {
2894 #if 0
2895 TargetMachine &TM = getTargetMachine();
2896 #endif
2897 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
2898 SelectionDAG &DAG = DCI.DAG;
2899 SDValue Op0 = N->getOperand(0); // everything has at least one operand
2900 EVT NodeVT = N->getValueType(0); // The node's value type
2901 EVT Op0VT = Op0.getValueType(); // The first operand's result
2902 SDValue Result; // Initially, empty result
2903 DebugLoc dl = N->getDebugLoc();
2904
2905 switch (N->getOpcode()) {
2906 default: break;
2907 case ISD::ADD: {
2908 SDValue Op1 = N->getOperand(1);
2909
2910 if (Op0.getOpcode() == SPUISD::IndirectAddr
2911 || Op1.getOpcode() == SPUISD::IndirectAddr) {
2912 // Normalize the operands to reduce repeated code
2913 SDValue IndirectArg = Op0, AddArg = Op1;
2914
2915 if (Op1.getOpcode() == SPUISD::IndirectAddr) {
2916 IndirectArg = Op1;
2917 AddArg = Op0;
2918 }
2919
2920 if (isa<ConstantSDNode>(AddArg)) {
2921 ConstantSDNode *CN0 = cast<ConstantSDNode > (AddArg);
2922 SDValue IndOp1 = IndirectArg.getOperand(1);
2923
2924 if (CN0->isNullValue()) {
2925 // (add (SPUindirect <arg>, <arg>), 0) ->
2926 // (SPUindirect <arg>, <arg>)
2927
2928 #if !defined(NDEBUG)
2929 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2930 errs() << "\n"
2931 << "Replace: (add (SPUindirect <arg>, <arg>), 0)\n"
2932 << "With: (SPUindirect <arg>, <arg>)\n";
2933 }
2934 #endif
2935
2936 return IndirectArg;
2937 } else if (isa<ConstantSDNode>(IndOp1)) {
2938 // (add (SPUindirect <arg>, <const>), <const>) ->
2939 // (SPUindirect <arg>, <const + const>)
2940 ConstantSDNode *CN1 = cast<ConstantSDNode > (IndOp1);
2941 int64_t combinedConst = CN0->getSExtValue() + CN1->getSExtValue();
2942 SDValue combinedValue = DAG.getConstant(combinedConst, Op0VT);
2943
2944 #if !defined(NDEBUG)
2945 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2946 errs() << "\n"
2947 << "Replace: (add (SPUindirect <arg>, " << CN1->getSExtValue()
2948 << "), " << CN0->getSExtValue() << ")\n"
2949 << "With: (SPUindirect <arg>, "
2950 << combinedConst << ")\n";
2951 }
2952 #endif
2953
2954 return DAG.getNode(SPUISD::IndirectAddr, dl, Op0VT,
2955 IndirectArg, combinedValue);
2956 }
2957 }
2958 }
2959 break;
2960 }
2961 case ISD::SIGN_EXTEND:
2962 case ISD::ZERO_EXTEND:
2963 case ISD::ANY_EXTEND: {
2964 if (Op0.getOpcode() == SPUISD::VEC2PREFSLOT && NodeVT == Op0VT) {
2965 // (any_extend (SPUextract_elt0 <arg>)) ->
2966 // (SPUextract_elt0 <arg>)
2967 // Types must match, however...
2968 #if !defined(NDEBUG)
2969 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2970 errs() << "\nReplace: ";
2971 N->dump(&DAG);
2972 errs() << "\nWith: ";
2973 Op0.getNode()->dump(&DAG);
2974 errs() << "\n";
2975 }
2976 #endif
2977
2978 return Op0;
2979 }
2980 break;
2981 }
2982 case SPUISD::IndirectAddr: {
2983 if (!ST->usingLargeMem() && Op0.getOpcode() == SPUISD::AFormAddr) {
2984 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1));
2985 if (CN != 0 && CN->isNullValue()) {
2986 // (SPUindirect (SPUaform <addr>, 0), 0) ->
2987 // (SPUaform <addr>, 0)
2988
2989 DEBUG(errs() << "Replace: ");
2990 DEBUG(N->dump(&DAG));
2991 DEBUG(errs() << "\nWith: ");
2992 DEBUG(Op0.getNode()->dump(&DAG));
2993 DEBUG(errs() << "\n");
2994
2995 return Op0;
2996 }
2997 } else if (Op0.getOpcode() == ISD::ADD) {
2998 SDValue Op1 = N->getOperand(1);
2999 if (ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(Op1)) {
3000 // (SPUindirect (add <arg>, <arg>), 0) ->
3001 // (SPUindirect <arg>, <arg>)
3002 if (CN1->isNullValue()) {
3003
3004 #if !defined(NDEBUG)
3005 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
3006 errs() << "\n"
3007 << "Replace: (SPUindirect (add <arg>, <arg>), 0)\n"
3008 << "With: (SPUindirect <arg>, <arg>)\n";
3009 }
3010 #endif
3011
3012 return DAG.getNode(SPUISD::IndirectAddr, dl, Op0VT,
3013 Op0.getOperand(0), Op0.getOperand(1));
3014 }
3015 }
3016 }
3017 break;
3018 }
3019 case SPUISD::SHL_BITS:
3020 case SPUISD::SHL_BYTES:
3021 case SPUISD::ROTBYTES_LEFT: {
3022 SDValue Op1 = N->getOperand(1);
3023
3024 // Kill degenerate vector shifts:
3025 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
3026 if (CN->isNullValue()) {
3027 Result = Op0;
3028 }
3029 }
3030 break;
3031 }
3032 case SPUISD::PREFSLOT2VEC: {
3033 switch (Op0.getOpcode()) {
3034 default:
3035 break;
3036 case ISD::ANY_EXTEND:
3037 case ISD::ZERO_EXTEND:
3038 case ISD::SIGN_EXTEND: {
3039 // (SPUprefslot2vec (any|zero|sign_extend (SPUvec2prefslot <arg>))) ->
3040 // <arg>
3041 // but only if the SPUprefslot2vec and <arg> types match.
3042 SDValue Op00 = Op0.getOperand(0);
3043 if (Op00.getOpcode() == SPUISD::VEC2PREFSLOT) {
3044 SDValue Op000 = Op00.getOperand(0);
3045 if (Op000.getValueType() == NodeVT) {
3046 Result = Op000;
3047 }
3048 }
3049 break;
3050 }
3051 case SPUISD::VEC2PREFSLOT: {
3052 // (SPUprefslot2vec (SPUvec2prefslot <arg>)) ->
3053 // <arg>
3054 Result = Op0.getOperand(0);
3055 break;
3056 }
3057 }
3058 break;
3059 }
3060 }
3061
3062 // Otherwise, return unchanged.
3063 #ifndef NDEBUG
3064 if (Result.getNode()) {
3065 DEBUG(errs() << "\nReplace.SPU: ");
3066 DEBUG(N->dump(&DAG));
3067 DEBUG(errs() << "\nWith: ");
3068 DEBUG(Result.getNode()->dump(&DAG));
3069 DEBUG(errs() << "\n");
3070 }
3071 #endif
3072
3073 return Result;
3074 }
3075
3076 //===----------------------------------------------------------------------===//
3077 // Inline Assembly Support
3078 //===----------------------------------------------------------------------===//
3079
3080 /// getConstraintType - Given a constraint letter, return the type of
3081 /// constraint it is for this target.
3082 SPUTargetLowering::ConstraintType
3083 SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
3084 if (ConstraintLetter.size() == 1) {
3085 switch (ConstraintLetter[0]) {
3086 default: break;
3087 case 'b':
3088 case 'r':
3089 case 'f':
3090 case 'v':
3091 case 'y':
3092 return C_RegisterClass;
3093 }
3094 }
3095 return TargetLowering::getConstraintType(ConstraintLetter);
3096 }
3097
3098 /// Examine constraint type and operand type and determine a weight value.
3099 /// This object must already have been set up with the operand type
3100 /// and the current alternative constraint selected.
3101 TargetLowering::ConstraintWeight
3102 SPUTargetLowering::getSingleConstraintMatchWeight(
3103 AsmOperandInfo &info, const char *constraint) const {
3104 ConstraintWeight weight = CW_Invalid;
3105 Value *CallOperandVal = info.CallOperandVal;
3106 // If we don't have a value, we can't do a match,
3107 // but allow it at the lowest weight.
3108 if (CallOperandVal == NULL)
3109 return CW_Default;
3110 // Look at the constraint type.
3111 switch (*constraint) {
3112 default:
3113 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3114 break;
3115 //FIXME: Seems like the supported constraint letters were just copied
3116 // from PPC, as the following doesn't correspond to the GCC docs.
3117 // I'm leaving it so until someone adds the corresponding lowering support.
3118 case 'b':
3119 case 'r':
3120 case 'f':
3121 case 'd':
3122 case 'v':
3123 case 'y':
3124 weight = CW_Register;
3125 break;
3126 }
3127 return weight;
3128 }
3129
3130 std::pair<unsigned, const TargetRegisterClass*>
3131 SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3132 EVT VT) const
3133 {
3134 if (Constraint.size() == 1) {
3135 // GCC RS6000 Constraint Letters
3136 switch (Constraint[0]) {
3137 case 'b': // R1-R31
3138 case 'r': // R0-R31
3139 if (VT == MVT::i64)
3140 return std::make_pair(0U, SPU::R64CRegisterClass);
3141 return std::make_pair(0U, SPU::R32CRegisterClass);
3142 case 'f':
3143 if (VT == MVT::f32)
3144 return std::make_pair(0U, SPU::R32FPRegisterClass);
3145 else if (VT == MVT::f64)
3146 return std::make_pair(0U, SPU::R64FPRegisterClass);
3147 break;
3148 case 'v':
3149 return std::make_pair(0U, SPU::GPRCRegisterClass);
3150 }
3151 }
3152
3153 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3154 }
3155
3156 //! Compute used/known bits for a SPU operand
3157 void
3158 SPUTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3159 const APInt &Mask,
3160 APInt &KnownZero,
3161 APInt &KnownOne,
3162 const SelectionDAG &DAG,
3163 unsigned Depth ) const {
3164 #if 0
3165 const uint64_t uint64_sizebits = sizeof(uint64_t) * CHAR_BIT;
3166
3167 switch (Op.getOpcode()) {
3168 default:
3169 // KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3170 break;
3171 case CALL:
3172 case SHUFB:
3173 case SHUFFLE_MASK:
3174 case CNTB:
3175 case SPUISD::PREFSLOT2VEC:
3176 case SPUISD::LDRESULT:
3177 case SPUISD::VEC2PREFSLOT:
3178 case SPUISD::SHLQUAD_L_BITS:
3179 case SPUISD::SHLQUAD_L_BYTES:
3180 case SPUISD::VEC_ROTL:
3181 case SPUISD::VEC_ROTR:
3182 case SPUISD::ROTBYTES_LEFT:
3183 case SPUISD::SELECT_MASK:
3184 case SPUISD::SELB:
3185 }
3186 #endif
3187 }
3188
3189 unsigned
3190 SPUTargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
3191 unsigned Depth) const {
3192 switch (Op.getOpcode()) {
3193 default:
3194 return 1;
3195
3196 case ISD::SETCC: {
3197 EVT VT = Op.getValueType();
3198
3199 if (VT != MVT::i8 && VT != MVT::i16 && VT != MVT::i32) {
3200 VT = MVT::i32;
3201 }
3202 return VT.getSizeInBits();
3203 }
3204 }
3205 }
3206
3207 // LowerAsmOperandForConstraint
3208 void
3209 SPUTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3210 char ConstraintLetter,
3211 std::vector<SDValue> &Ops,
3212 SelectionDAG &DAG) const {
3213 // Default, for the time being, to the base class handler
3214 TargetLowering::LowerAsmOperandForConstraint(Op, ConstraintLetter, Ops, DAG);
3215 }
3216
3217 /// isLegalAddressImmediate - Return true if the integer value can be used
3218 /// as the offset of the target addressing mode.
3219 bool SPUTargetLowering::isLegalAddressImmediate(int64_t V,
3220 const Type *Ty) const {
3221 // SPU's addresses are 256K:
3222 return (V > -(1 << 18) && V < (1 << 18) - 1);
3223 }
3224
3225 bool SPUTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3226 return false;
3227 }
3228
3229 bool
3230 SPUTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3231 // The SPU target isn't yet aware of offsets.
3232 return false;
3233 }
3234
3235 // can we compare to Imm without writing it into a register?
3236 bool SPUTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
3237 //ceqi, cgti, etc. all take s10 operand
3238 return isInt<10>(Imm);
3239 }
3240
3241 bool
3242 SPUTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3243 const Type * ) const{
3244
3245 // A-form: 18bit absolute address.
3246 if (AM.BaseGV && !AM.HasBaseReg && AM.Scale == 0 && AM.BaseOffs == 0)
3247 return true;
3248
3249 // D-form: reg + 14bit offset
3250 if (AM.BaseGV ==0 && AM.HasBaseReg && AM.Scale == 0 && isInt<14>(AM.BaseOffs))
3251 return true;
3252
3253 // X-form: reg+reg
3254 if (AM.BaseGV == 0 && AM.HasBaseReg && AM.Scale == 1 && AM.BaseOffs ==0)
3255 return true;
3256
3257 return false;
3258 }
STM8 backend for LLVM