1 //===-- SPUISelLowering.cpp - Cell SPU DAG Lowering Implementation --------===//
2 // The LLVM Compiler Infrastructure
4 // This file is distributed under the University of Illinois Open Source
5 // License. See LICENSE.TXT for details.
7 //===----------------------------------------------------------------------===//
9 // This file implements the SPUTargetLowering class.
11 //===----------------------------------------------------------------------===//
13 #include "SPUISelLowering.h"
14 #include "SPUTargetMachine.h"
15 #include "SPUFrameLowering.h"
16 #include "SPUMachineFunction.h"
17 #include "llvm/Constants.h"
18 #include "llvm/Function.h"
19 #include "llvm/Intrinsics.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Type.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
38 // Byte offset of the preferred slot (counted from the MSB)
39 int prefslotOffset(EVT VT) {
41 if (VT==MVT::i1) retval=3;
42 if (VT==MVT::i8) retval=3;
43 if (VT==MVT::i16) retval=2;
48 //! Expand a library call into an actual call DAG node
51 This code is taken from SelectionDAGLegalize, since it is not exposed as
52 part of the LLVM SelectionDAG API.
56 ExpandLibCall(RTLIB::Libcall LC, SDValue Op, SelectionDAG &DAG,
57 bool isSigned, SDValue &Hi, const SPUTargetLowering &TLI) {
58 // The input chain to this libcall is the entry node of the function.
59 // Legalizing the call will automatically add the previous call to the
61 SDValue InChain = DAG.getEntryNode();
63 TargetLowering::ArgListTy Args;
64 TargetLowering::ArgListEntry Entry;
65 for (unsigned i = 0, e = Op.getNumOperands(); i != e; ++i) {
66 EVT ArgVT = Op.getOperand(i).getValueType();
67 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
68 Entry.Node = Op.getOperand(i);
70 Entry.isSExt = isSigned;
71 Entry.isZExt = !isSigned;
72 Args.push_back(Entry);
74 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
77 // Splice the libcall in wherever FindInputOutputChains tells us to.
79 Op.getNode()->getValueType(0).getTypeForEVT(*DAG.getContext());
80 std::pair<SDValue, SDValue> CallInfo =
81 TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false,
82 0, TLI.getLibcallCallingConv(LC),
84 /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
85 Callee, Args, DAG, Op.getDebugLoc());
87 return CallInfo.first;
91 SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
92 : TargetLowering(TM, new TargetLoweringObjectFileELF()),
95 // Use _setjmp/_longjmp instead of setjmp/longjmp.
96 setUseUnderscoreSetJmp(true);
97 setUseUnderscoreLongJmp(true);
99 // Set RTLIB libcall names as used by SPU:
100 setLibcallName(RTLIB::DIV_F64, "__fast_divdf3");
102 // Set up the SPU's register classes:
103 addRegisterClass(MVT::i8, &SPU::R8CRegClass);
104 addRegisterClass(MVT::i16, &SPU::R16CRegClass);
105 addRegisterClass(MVT::i32, &SPU::R32CRegClass);
106 addRegisterClass(MVT::i64, &SPU::R64CRegClass);
107 addRegisterClass(MVT::f32, &SPU::R32FPRegClass);
108 addRegisterClass(MVT::f64, &SPU::R64FPRegClass);
109 addRegisterClass(MVT::i128, &SPU::GPRCRegClass);
111 // SPU has no sign or zero extended loads for i1, i8, i16:
112 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
113 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
114 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
116 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
117 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
119 setTruncStoreAction(MVT::i128, MVT::i64, Expand);
120 setTruncStoreAction(MVT::i128, MVT::i32, Expand);
121 setTruncStoreAction(MVT::i128, MVT::i16, Expand);
122 setTruncStoreAction(MVT::i128, MVT::i8, Expand);
124 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
126 // SPU constant load actions are custom lowered:
127 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
128 setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
130 // SPU's loads and stores have to be custom lowered:
131 for (unsigned sctype = (unsigned) MVT::i8; sctype < (unsigned) MVT::i128;
133 MVT::SimpleValueType VT = (MVT::SimpleValueType)sctype;
135 setOperationAction(ISD::LOAD, VT, Custom);
136 setOperationAction(ISD::STORE, VT, Custom);
137 setLoadExtAction(ISD::EXTLOAD, VT, Custom);
138 setLoadExtAction(ISD::ZEXTLOAD, VT, Custom);
139 setLoadExtAction(ISD::SEXTLOAD, VT, Custom);
141 for (unsigned stype = sctype - 1; stype >= (unsigned) MVT::i8; --stype) {
142 MVT::SimpleValueType StoreVT = (MVT::SimpleValueType) stype;
143 setTruncStoreAction(VT, StoreVT, Expand);
147 for (unsigned sctype = (unsigned) MVT::f32; sctype < (unsigned) MVT::f64;
149 MVT::SimpleValueType VT = (MVT::SimpleValueType) sctype;
151 setOperationAction(ISD::LOAD, VT, Custom);
152 setOperationAction(ISD::STORE, VT, Custom);
154 for (unsigned stype = sctype - 1; stype >= (unsigned) MVT::f32; --stype) {
155 MVT::SimpleValueType StoreVT = (MVT::SimpleValueType) stype;
156 setTruncStoreAction(VT, StoreVT, Expand);
160 // Expand the jumptable branches
161 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
162 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
164 // Custom lower SELECT_CC for most cases, but expand by default
165 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
166 setOperationAction(ISD::SELECT_CC, MVT::i8, Custom);
167 setOperationAction(ISD::SELECT_CC, MVT::i16, Custom);
168 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
169 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
171 // SPU has no intrinsics for these particular operations:
172 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
173 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
175 // SPU has no division/remainder instructions
176 setOperationAction(ISD::SREM, MVT::i8, Expand);
177 setOperationAction(ISD::UREM, MVT::i8, Expand);
178 setOperationAction(ISD::SDIV, MVT::i8, Expand);
179 setOperationAction(ISD::UDIV, MVT::i8, Expand);
180 setOperationAction(ISD::SDIVREM, MVT::i8, Expand);
181 setOperationAction(ISD::UDIVREM, MVT::i8, Expand);
182 setOperationAction(ISD::SREM, MVT::i16, Expand);
183 setOperationAction(ISD::UREM, MVT::i16, Expand);
184 setOperationAction(ISD::SDIV, MVT::i16, Expand);
185 setOperationAction(ISD::UDIV, MVT::i16, Expand);
186 setOperationAction(ISD::SDIVREM, MVT::i16, Expand);
187 setOperationAction(ISD::UDIVREM, MVT::i16, Expand);
188 setOperationAction(ISD::SREM, MVT::i32, Expand);
189 setOperationAction(ISD::UREM, MVT::i32, Expand);
190 setOperationAction(ISD::SDIV, MVT::i32, Expand);
191 setOperationAction(ISD::UDIV, MVT::i32, Expand);
192 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
193 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
194 setOperationAction(ISD::SREM, MVT::i64, Expand);
195 setOperationAction(ISD::UREM, MVT::i64, Expand);
196 setOperationAction(ISD::SDIV, MVT::i64, Expand);
197 setOperationAction(ISD::UDIV, MVT::i64, Expand);
198 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
199 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
200 setOperationAction(ISD::SREM, MVT::i128, Expand);
201 setOperationAction(ISD::UREM, MVT::i128, Expand);
202 setOperationAction(ISD::SDIV, MVT::i128, Expand);
203 setOperationAction(ISD::UDIV, MVT::i128, Expand);
204 setOperationAction(ISD::SDIVREM, MVT::i128, Expand);
205 setOperationAction(ISD::UDIVREM, MVT::i128, Expand);
207 // We don't support sin/cos/sqrt/fmod
208 setOperationAction(ISD::FSIN , MVT::f64, Expand);
209 setOperationAction(ISD::FCOS , MVT::f64, Expand);
210 setOperationAction(ISD::FREM , MVT::f64, Expand);
211 setOperationAction(ISD::FSIN , MVT::f32, Expand);
212 setOperationAction(ISD::FCOS , MVT::f32, Expand);
213 setOperationAction(ISD::FREM , MVT::f32, Expand);
215 // Expand fsqrt to the appropriate libcall (NOTE: should use h/w fsqrt
217 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
218 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
220 setOperationAction(ISD::FMA, MVT::f64, Expand);
221 setOperationAction(ISD::FMA, MVT::f32, Expand);
223 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
224 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
226 // SPU can do rotate right and left, so legalize it... but customize for i8
227 // because instructions don't exist.
229 // FIXME: Change from "expand" to appropriate type once ROTR is supported in
231 setOperationAction(ISD::ROTR, MVT::i32, Expand /*Legal*/);
232 setOperationAction(ISD::ROTR, MVT::i16, Expand /*Legal*/);
233 setOperationAction(ISD::ROTR, MVT::i8, Expand /*Custom*/);
235 setOperationAction(ISD::ROTL, MVT::i32, Legal);
236 setOperationAction(ISD::ROTL, MVT::i16, Legal);
237 setOperationAction(ISD::ROTL, MVT::i8, Custom);
239 // SPU has no native version of shift left/right for i8
240 setOperationAction(ISD::SHL, MVT::i8, Custom);
241 setOperationAction(ISD::SRL, MVT::i8, Custom);
242 setOperationAction(ISD::SRA, MVT::i8, Custom);
244 // Make these operations legal and handle them during instruction selection:
245 setOperationAction(ISD::SHL, MVT::i64, Legal);
246 setOperationAction(ISD::SRL, MVT::i64, Legal);
247 setOperationAction(ISD::SRA, MVT::i64, Legal);
249 // Custom lower i8, i32 and i64 multiplications
250 setOperationAction(ISD::MUL, MVT::i8, Custom);
251 setOperationAction(ISD::MUL, MVT::i32, Legal);
252 setOperationAction(ISD::MUL, MVT::i64, Legal);
254 // Expand double-width multiplication
255 // FIXME: It would probably be reasonable to support some of these operations
256 setOperationAction(ISD::UMUL_LOHI, MVT::i8, Expand);
257 setOperationAction(ISD::SMUL_LOHI, MVT::i8, Expand);
258 setOperationAction(ISD::MULHU, MVT::i8, Expand);
259 setOperationAction(ISD::MULHS, MVT::i8, Expand);
260 setOperationAction(ISD::UMUL_LOHI, MVT::i16, Expand);
261 setOperationAction(ISD::SMUL_LOHI, MVT::i16, Expand);
262 setOperationAction(ISD::MULHU, MVT::i16, Expand);
263 setOperationAction(ISD::MULHS, MVT::i16, Expand);
264 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
265 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
266 setOperationAction(ISD::MULHU, MVT::i32, Expand);
267 setOperationAction(ISD::MULHS, MVT::i32, Expand);
268 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
269 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
270 setOperationAction(ISD::MULHU, MVT::i64, Expand);
271 setOperationAction(ISD::MULHS, MVT::i64, Expand);
273 // Need to custom handle (some) common i8, i64 math ops
274 setOperationAction(ISD::ADD, MVT::i8, Custom);
275 setOperationAction(ISD::ADD, MVT::i64, Legal);
276 setOperationAction(ISD::SUB, MVT::i8, Custom);
277 setOperationAction(ISD::SUB, MVT::i64, Legal);
279 // SPU does not have BSWAP. It does have i32 support CTLZ.
280 // CTPOP has to be custom lowered.
281 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
282 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
284 setOperationAction(ISD::CTPOP, MVT::i8, Custom);
285 setOperationAction(ISD::CTPOP, MVT::i16, Custom);
286 setOperationAction(ISD::CTPOP, MVT::i32, Custom);
287 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
288 setOperationAction(ISD::CTPOP, MVT::i128, Expand);
290 setOperationAction(ISD::CTTZ , MVT::i8, Expand);
291 setOperationAction(ISD::CTTZ , MVT::i16, Expand);
292 setOperationAction(ISD::CTTZ , MVT::i32, Expand);
293 setOperationAction(ISD::CTTZ , MVT::i64, Expand);
294 setOperationAction(ISD::CTTZ , MVT::i128, Expand);
295 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i8, Expand);
296 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Expand);
297 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
298 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
299 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i128, Expand);
301 setOperationAction(ISD::CTLZ , MVT::i8, Promote);
302 setOperationAction(ISD::CTLZ , MVT::i16, Promote);
303 setOperationAction(ISD::CTLZ , MVT::i32, Legal);
304 setOperationAction(ISD::CTLZ , MVT::i64, Expand);
305 setOperationAction(ISD::CTLZ , MVT::i128, Expand);
306 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8, Expand);
307 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Expand);
308 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
309 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
310 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i128, Expand);
312 // SPU has a version of select that implements (a&~c)|(b&c), just like
313 // select ought to work:
314 setOperationAction(ISD::SELECT, MVT::i8, Legal);
315 setOperationAction(ISD::SELECT, MVT::i16, Legal);
316 setOperationAction(ISD::SELECT, MVT::i32, Legal);
317 setOperationAction(ISD::SELECT, MVT::i64, Legal);
319 setOperationAction(ISD::SETCC, MVT::i8, Legal);
320 setOperationAction(ISD::SETCC, MVT::i16, Legal);
321 setOperationAction(ISD::SETCC, MVT::i32, Legal);
322 setOperationAction(ISD::SETCC, MVT::i64, Legal);
323 setOperationAction(ISD::SETCC, MVT::f64, Custom);
325 // Custom lower i128 -> i64 truncates
326 setOperationAction(ISD::TRUNCATE, MVT::i64, Custom);
328 // Custom lower i32/i64 -> i128 sign extend
329 setOperationAction(ISD::SIGN_EXTEND, MVT::i128, Custom);
331 setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote);
332 setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote);
333 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
334 setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
335 // SPU has a legal FP -> signed INT instruction for f32, but for f64, need
336 // to expand to a libcall, hence the custom lowering:
337 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
338 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
339 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
340 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
341 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Expand);
342 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Expand);
344 // FDIV on SPU requires custom lowering
345 setOperationAction(ISD::FDIV, MVT::f64, Expand); // to libcall
347 // SPU has [U|S]INT_TO_FP for f32->i32, but not for f64->i32, f64->i64:
348 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
349 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
350 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
351 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
352 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
353 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
354 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
355 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
357 setOperationAction(ISD::BITCAST, MVT::i32, Legal);
358 setOperationAction(ISD::BITCAST, MVT::f32, Legal);
359 setOperationAction(ISD::BITCAST, MVT::i64, Legal);
360 setOperationAction(ISD::BITCAST, MVT::f64, Legal);
362 // We cannot sextinreg(i1). Expand to shifts.
363 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
365 // We want to legalize GlobalAddress and ConstantPool nodes into the
366 // appropriate instructions to materialize the address.
367 for (unsigned sctype = (unsigned) MVT::i8; sctype < (unsigned) MVT::f128;
369 MVT::SimpleValueType VT = (MVT::SimpleValueType)sctype;
371 setOperationAction(ISD::GlobalAddress, VT, Custom);
372 setOperationAction(ISD::ConstantPool, VT, Custom);
373 setOperationAction(ISD::JumpTable, VT, Custom);
376 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
377 setOperationAction(ISD::VASTART , MVT::Other, Custom);
379 // Use the default implementation.
380 setOperationAction(ISD::VAARG , MVT::Other, Expand);
381 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
382 setOperationAction(ISD::VAEND , MVT::Other, Expand);
383 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
384 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
385 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
386 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Expand);
388 // Cell SPU has instructions for converting between i64 and fp.
389 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
390 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
392 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
393 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
395 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
396 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
398 // First set operation action for all vector types to expand. Then we
399 // will selectively turn on ones that can be effectively codegen'd.
400 addRegisterClass(MVT::v16i8, &SPU::VECREGRegClass);
401 addRegisterClass(MVT::v8i16, &SPU::VECREGRegClass);
402 addRegisterClass(MVT::v4i32, &SPU::VECREGRegClass);
403 addRegisterClass(MVT::v2i64, &SPU::VECREGRegClass);
404 addRegisterClass(MVT::v4f32, &SPU::VECREGRegClass);
405 addRegisterClass(MVT::v2f64, &SPU::VECREGRegClass);
407 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
408 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
409 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
411 // Set operation actions to legal types only.
412 if (!isTypeLegal(VT)) continue;
414 // add/sub are legal for all supported vector VT's.
415 setOperationAction(ISD::ADD, VT, Legal);
416 setOperationAction(ISD::SUB, VT, Legal);
417 // mul has to be custom lowered.
418 setOperationAction(ISD::MUL, VT, Legal);
420 setOperationAction(ISD::AND, VT, Legal);
421 setOperationAction(ISD::OR, VT, Legal);
422 setOperationAction(ISD::XOR, VT, Legal);
423 setOperationAction(ISD::LOAD, VT, Custom);
424 setOperationAction(ISD::SELECT, VT, Legal);
425 setOperationAction(ISD::STORE, VT, Custom);
427 // These operations need to be expanded:
428 setOperationAction(ISD::SDIV, VT, Expand);
429 setOperationAction(ISD::SREM, VT, Expand);
430 setOperationAction(ISD::UDIV, VT, Expand);
431 setOperationAction(ISD::UREM, VT, Expand);
433 // Expand all trunc stores
434 for (unsigned j = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
435 j <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++j) {
436 MVT::SimpleValueType TargetVT = (MVT::SimpleValueType)j;
437 setTruncStoreAction(VT, TargetVT, Expand);
440 // Custom lower build_vector, constant pool spills, insert and
441 // extract vector elements:
442 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
443 setOperationAction(ISD::ConstantPool, VT, Custom);
444 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
445 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
446 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
447 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
450 setOperationAction(ISD::SHL, MVT::v2i64, Expand);
452 setOperationAction(ISD::AND, MVT::v16i8, Custom);
453 setOperationAction(ISD::OR, MVT::v16i8, Custom);
454 setOperationAction(ISD::XOR, MVT::v16i8, Custom);
455 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
457 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
459 setBooleanContents(ZeroOrNegativeOneBooleanContent);
460 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); // FIXME: Is this correct?
462 setStackPointerRegisterToSaveRestore(SPU::R1);
464 // We have target-specific dag combine patterns for the following nodes:
465 setTargetDAGCombine(ISD::ADD);
466 setTargetDAGCombine(ISD::ZERO_EXTEND);
467 setTargetDAGCombine(ISD::SIGN_EXTEND);
468 setTargetDAGCombine(ISD::ANY_EXTEND);
470 setMinFunctionAlignment(3);
472 computeRegisterProperties();
474 // Set pre-RA register scheduler default to BURR, which produces slightly
475 // better code than the default (could also be TDRR, but TargetLowering.h
476 // needs a mod to support that model):
477 setSchedulingPreference(Sched::RegPressure);
480 const char *SPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
483 case SPUISD::RET_FLAG: return "SPUISD::RET_FLAG";
484 case SPUISD::Hi: return "SPUISD::Hi";
485 case SPUISD::Lo: return "SPUISD::Lo";
486 case SPUISD::PCRelAddr: return "SPUISD::PCRelAddr";
487 case SPUISD::AFormAddr: return "SPUISD::AFormAddr";
488 case SPUISD::IndirectAddr: return "SPUISD::IndirectAddr";
489 case SPUISD::LDRESULT: return "SPUISD::LDRESULT";
490 case SPUISD::CALL: return "SPUISD::CALL";
491 case SPUISD::SHUFB: return "SPUISD::SHUFB";
492 case SPUISD::SHUFFLE_MASK: return "SPUISD::SHUFFLE_MASK";
493 case SPUISD::CNTB: return "SPUISD::CNTB";
494 case SPUISD::PREFSLOT2VEC: return "SPUISD::PREFSLOT2VEC";
495 case SPUISD::VEC2PREFSLOT: return "SPUISD::VEC2PREFSLOT";
496 case SPUISD::SHL_BITS: return "SPUISD::SHL_BITS";
497 case SPUISD::SHL_BYTES: return "SPUISD::SHL_BYTES";
498 case SPUISD::VEC_ROTL: return "SPUISD::VEC_ROTL";
499 case SPUISD::VEC_ROTR: return "SPUISD::VEC_ROTR";
500 case SPUISD::ROTBYTES_LEFT: return "SPUISD::ROTBYTES_LEFT";
501 case SPUISD::ROTBYTES_LEFT_BITS: return "SPUISD::ROTBYTES_LEFT_BITS";
502 case SPUISD::SELECT_MASK: return "SPUISD::SELECT_MASK";
503 case SPUISD::SELB: return "SPUISD::SELB";
504 case SPUISD::ADD64_MARKER: return "SPUISD::ADD64_MARKER";
505 case SPUISD::SUB64_MARKER: return "SPUISD::SUB64_MARKER";
506 case SPUISD::MUL64_MARKER: return "SPUISD::MUL64_MARKER";
510 //===----------------------------------------------------------------------===//
511 // Return the Cell SPU's SETCC result type
512 //===----------------------------------------------------------------------===//
514 EVT SPUTargetLowering::getSetCCResultType(EVT VT) const {
515 // i8, i16 and i32 are valid SETCC result types
516 MVT::SimpleValueType retval;
518 switch(VT.getSimpleVT().SimpleTy){
521 retval = MVT::i8; break;
523 retval = MVT::i16; break;
531 //===----------------------------------------------------------------------===//
532 // Calling convention code:
533 //===----------------------------------------------------------------------===//
535 #include "SPUGenCallingConv.inc"
537 //===----------------------------------------------------------------------===//
538 // LowerOperation implementation
539 //===----------------------------------------------------------------------===//
541 /// Custom lower loads for CellSPU
543 All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
544 within a 16-byte block, we have to rotate to extract the requested element.
546 For extending loads, we also want to ensure that the following sequence is
547 emitted, e.g. for MVT::f32 extending load to MVT::f64:
551 %2 v16i8,ch = rotate %1
552 %3 v4f8, ch = bitconvert %2
553 %4 f32 = vec2perfslot %3
554 %5 f64 = fp_extend %4
558 LowerLOAD(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
559 LoadSDNode *LN = cast<LoadSDNode>(Op);
560 SDValue the_chain = LN->getChain();
561 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
562 EVT InVT = LN->getMemoryVT();
563 EVT OutVT = Op.getValueType();
564 ISD::LoadExtType ExtType = LN->getExtensionType();
565 unsigned alignment = LN->getAlignment();
566 int pso = prefslotOffset(InVT);
567 DebugLoc dl = Op.getDebugLoc();
568 EVT vecVT = InVT.isVector()? InVT: EVT::getVectorVT(*DAG.getContext(), InVT,
569 (128 / InVT.getSizeInBits()));
572 assert( LN->getAddressingMode() == ISD::UNINDEXED
573 && "we should get only UNINDEXED adresses");
574 // clean aligned loads can be selected as-is
575 if (InVT.getSizeInBits() == 128 && (alignment%16) == 0)
578 // Get pointerinfos to the memory chunk(s) that contain the data to load
579 uint64_t mpi_offset = LN->getPointerInfo().Offset;
580 mpi_offset -= mpi_offset%16;
581 MachinePointerInfo lowMemPtr(LN->getPointerInfo().V, mpi_offset);
582 MachinePointerInfo highMemPtr(LN->getPointerInfo().V, mpi_offset+16);
585 SDValue basePtr = LN->getBasePtr();
588 if ((alignment%16) == 0) {
591 // Special cases for a known aligned load to simplify the base pointer
592 // and the rotation amount:
593 if (basePtr.getOpcode() == ISD::ADD
594 && (CN = dyn_cast<ConstantSDNode > (basePtr.getOperand(1))) != 0) {
595 // Known offset into basePtr
596 int64_t offset = CN->getSExtValue();
597 int64_t rotamt = int64_t((offset & 0xf) - pso);
602 rotate = DAG.getConstant(rotamt, MVT::i16);
604 // Simplify the base pointer for this case:
605 basePtr = basePtr.getOperand(0);
606 if ((offset & ~0xf) > 0) {
607 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
609 DAG.getConstant((offset & ~0xf), PtrVT));
611 } else if ((basePtr.getOpcode() == SPUISD::AFormAddr)
612 || (basePtr.getOpcode() == SPUISD::IndirectAddr
613 && basePtr.getOperand(0).getOpcode() == SPUISD::Hi
614 && basePtr.getOperand(1).getOpcode() == SPUISD::Lo)) {
615 // Plain aligned a-form address: rotate into preferred slot
616 // Same for (SPUindirect (SPUhi ...), (SPUlo ...))
617 int64_t rotamt = -pso;
620 rotate = DAG.getConstant(rotamt, MVT::i16);
622 // Offset the rotate amount by the basePtr and the preferred slot
624 int64_t rotamt = -pso;
627 rotate = DAG.getNode(ISD::ADD, dl, PtrVT,
629 DAG.getConstant(rotamt, PtrVT));
632 // Unaligned load: must be more pessimistic about addressing modes:
633 if (basePtr.getOpcode() == ISD::ADD) {
634 MachineFunction &MF = DAG.getMachineFunction();
635 MachineRegisterInfo &RegInfo = MF.getRegInfo();
636 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
639 SDValue Op0 = basePtr.getOperand(0);
640 SDValue Op1 = basePtr.getOperand(1);
642 if (isa<ConstantSDNode>(Op1)) {
643 // Convert the (add <ptr>, <const>) to an indirect address contained
644 // in a register. Note that this is done because we need to avoid
645 // creating a 0(reg) d-form address due to the SPU's block loads.
646 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
647 the_chain = DAG.getCopyToReg(the_chain, dl, VReg, basePtr, Flag);
648 basePtr = DAG.getCopyFromReg(the_chain, dl, VReg, PtrVT);
650 // Convert the (add <arg1>, <arg2>) to an indirect address, which
651 // will likely be lowered as a reg(reg) x-form address.
652 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
655 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
657 DAG.getConstant(0, PtrVT));
660 // Offset the rotate amount by the basePtr and the preferred slot
662 rotate = DAG.getNode(ISD::ADD, dl, PtrVT,
664 DAG.getConstant(-pso, PtrVT));
667 // Do the load as a i128 to allow possible shifting
668 SDValue low = DAG.getLoad(MVT::i128, dl, the_chain, basePtr,
670 LN->isVolatile(), LN->isNonTemporal(), false, 16);
672 // When the size is not greater than alignment we get all data with just
674 if (alignment >= InVT.getSizeInBits()/8) {
676 the_chain = low.getValue(1);
678 // Rotate into the preferred slot:
679 result = DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, MVT::i128,
680 low.getValue(0), rotate);
682 // Convert the loaded v16i8 vector to the appropriate vector type
683 // specified by the operand:
684 EVT vecVT = EVT::getVectorVT(*DAG.getContext(),
685 InVT, (128 / InVT.getSizeInBits()));
686 result = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, InVT,
687 DAG.getNode(ISD::BITCAST, dl, vecVT, result));
689 // When alignment is less than the size, we might need (known only at
690 // run-time) two loads
691 // TODO: if the memory address is composed only from constants, we have
692 // extra kowledge, and might avoid the second load
694 // storage position offset from lower 16 byte aligned memory chunk
695 SDValue offset = DAG.getNode(ISD::AND, dl, MVT::i32,
696 basePtr, DAG.getConstant( 0xf, MVT::i32 ) );
697 // get a registerfull of ones. (this implementation is a workaround: LLVM
698 // cannot handle 128 bit signed int constants)
699 SDValue ones = DAG.getConstant(-1, MVT::v4i32 );
700 ones = DAG.getNode(ISD::BITCAST, dl, MVT::i128, ones);
702 SDValue high = DAG.getLoad(MVT::i128, dl, the_chain,
703 DAG.getNode(ISD::ADD, dl, PtrVT,
705 DAG.getConstant(16, PtrVT)),
707 LN->isVolatile(), LN->isNonTemporal(), false,
710 the_chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(1),
713 // Shift the (possible) high part right to compensate the misalignemnt.
714 // if there is no highpart (i.e. value is i64 and offset is 4), this
715 // will zero out the high value.
716 high = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, high,
717 DAG.getNode(ISD::SUB, dl, MVT::i32,
718 DAG.getConstant( 16, MVT::i32),
722 // Shift the low similarly
723 // TODO: add SPUISD::SHL_BYTES
724 low = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, low, offset );
726 // Merge the two parts
727 result = DAG.getNode(ISD::BITCAST, dl, vecVT,
728 DAG.getNode(ISD::OR, dl, MVT::i128, low, high));
730 if (!InVT.isVector()) {
731 result = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, InVT, result );
735 // Handle extending loads by extending the scalar result:
736 if (ExtType == ISD::SEXTLOAD) {
737 result = DAG.getNode(ISD::SIGN_EXTEND, dl, OutVT, result);
738 } else if (ExtType == ISD::ZEXTLOAD) {
739 result = DAG.getNode(ISD::ZERO_EXTEND, dl, OutVT, result);
740 } else if (ExtType == ISD::EXTLOAD) {
741 unsigned NewOpc = ISD::ANY_EXTEND;
743 if (OutVT.isFloatingPoint())
744 NewOpc = ISD::FP_EXTEND;
746 result = DAG.getNode(NewOpc, dl, OutVT, result);
749 SDVTList retvts = DAG.getVTList(OutVT, MVT::Other);
750 SDValue retops[2] = {
755 result = DAG.getNode(SPUISD::LDRESULT, dl, retvts,
756 retops, sizeof(retops) / sizeof(retops[0]));
760 /// Custom lower stores for CellSPU
762 All CellSPU stores are aligned to 16-byte boundaries, so for elements
763 within a 16-byte block, we have to generate a shuffle to insert the
764 requested element into its place, then store the resulting block.
767 LowerSTORE(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
768 StoreSDNode *SN = cast<StoreSDNode>(Op);
769 SDValue Value = SN->getValue();
770 EVT VT = Value.getValueType();
771 EVT StVT = (!SN->isTruncatingStore() ? VT : SN->getMemoryVT());
772 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
773 DebugLoc dl = Op.getDebugLoc();
774 unsigned alignment = SN->getAlignment();
776 EVT vecVT = StVT.isVector()? StVT: EVT::getVectorVT(*DAG.getContext(), StVT,
777 (128 / StVT.getSizeInBits()));
778 // Get pointerinfos to the memory chunk(s) that contain the data to load
779 uint64_t mpi_offset = SN->getPointerInfo().Offset;
780 mpi_offset -= mpi_offset%16;
781 MachinePointerInfo lowMemPtr(SN->getPointerInfo().V, mpi_offset);
782 MachinePointerInfo highMemPtr(SN->getPointerInfo().V, mpi_offset+16);
786 assert( SN->getAddressingMode() == ISD::UNINDEXED
787 && "we should get only UNINDEXED adresses");
788 // clean aligned loads can be selected as-is
789 if (StVT.getSizeInBits() == 128 && (alignment%16) == 0)
792 SDValue alignLoadVec;
793 SDValue basePtr = SN->getBasePtr();
794 SDValue the_chain = SN->getChain();
795 SDValue insertEltOffs;
797 if ((alignment%16) == 0) {
799 // Special cases for a known aligned load to simplify the base pointer
800 // and insertion byte:
801 if (basePtr.getOpcode() == ISD::ADD
802 && (CN = dyn_cast<ConstantSDNode>(basePtr.getOperand(1))) != 0) {
803 // Known offset into basePtr
804 int64_t offset = CN->getSExtValue();
806 // Simplify the base pointer for this case:
807 basePtr = basePtr.getOperand(0);
808 insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
810 DAG.getConstant((offset & 0xf), PtrVT));
812 if ((offset & ~0xf) > 0) {
813 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
815 DAG.getConstant((offset & ~0xf), PtrVT));
818 // Otherwise, assume it's at byte 0 of basePtr
819 insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
821 DAG.getConstant(0, PtrVT));
822 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
824 DAG.getConstant(0, PtrVT));
827 // Unaligned load: must be more pessimistic about addressing modes:
828 if (basePtr.getOpcode() == ISD::ADD) {
829 MachineFunction &MF = DAG.getMachineFunction();
830 MachineRegisterInfo &RegInfo = MF.getRegInfo();
831 unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
834 SDValue Op0 = basePtr.getOperand(0);
835 SDValue Op1 = basePtr.getOperand(1);
837 if (isa<ConstantSDNode>(Op1)) {
838 // Convert the (add <ptr>, <const>) to an indirect address contained
839 // in a register. Note that this is done because we need to avoid
840 // creating a 0(reg) d-form address due to the SPU's block loads.
841 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
842 the_chain = DAG.getCopyToReg(the_chain, dl, VReg, basePtr, Flag);
843 basePtr = DAG.getCopyFromReg(the_chain, dl, VReg, PtrVT);
845 // Convert the (add <arg1>, <arg2>) to an indirect address, which
846 // will likely be lowered as a reg(reg) x-form address.
847 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Op0, Op1);
850 basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
852 DAG.getConstant(0, PtrVT));
855 // Insertion point is solely determined by basePtr's contents
856 insertEltOffs = DAG.getNode(ISD::ADD, dl, PtrVT,
858 DAG.getConstant(0, PtrVT));
861 // Load the lower part of the memory to which to store.
862 SDValue low = DAG.getLoad(vecVT, dl, the_chain, basePtr,
863 lowMemPtr, SN->isVolatile(), SN->isNonTemporal(),
866 // if we don't need to store over the 16 byte boundary, one store suffices
867 if (alignment >= StVT.getSizeInBits()/8) {
869 the_chain = low.getValue(1);
871 LoadSDNode *LN = cast<LoadSDNode>(low);
872 SDValue theValue = SN->getValue();
875 && (theValue.getOpcode() == ISD::AssertZext
876 || theValue.getOpcode() == ISD::AssertSext)) {
877 // Drill down and get the value for zero- and sign-extended
879 theValue = theValue.getOperand(0);
882 // If the base pointer is already a D-form address, then just create
883 // a new D-form address with a slot offset and the orignal base pointer.
884 // Otherwise generate a D-form address with the slot offset relative
885 // to the stack pointer, which is always aligned.
887 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
888 errs() << "CellSPU LowerSTORE: basePtr = ";
889 basePtr.getNode()->dump(&DAG);
894 SDValue insertEltOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl, vecVT,
896 SDValue vectorizeOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, vecVT,
899 result = DAG.getNode(SPUISD::SHUFB, dl, vecVT,
901 DAG.getNode(ISD::BITCAST, dl,
902 MVT::v4i32, insertEltOp));
904 result = DAG.getStore(the_chain, dl, result, basePtr,
906 LN->isVolatile(), LN->isNonTemporal(),
910 // do the store when it might cross the 16 byte memory access boundary.
912 // TODO issue a warning if SN->isVolatile()== true? This is likely not
913 // what the user wanted.
915 // address offset from nearest lower 16byte alinged address
916 SDValue offset = DAG.getNode(ISD::AND, dl, MVT::i32,
918 DAG.getConstant(0xf, MVT::i32));
920 SDValue offset_compl = DAG.getNode(ISD::SUB, dl, MVT::i32,
921 DAG.getConstant( 16, MVT::i32),
923 // 16 - sizeof(Value)
924 SDValue surplus = DAG.getNode(ISD::SUB, dl, MVT::i32,
925 DAG.getConstant( 16, MVT::i32),
926 DAG.getConstant( VT.getSizeInBits()/8,
928 // get a registerfull of ones
929 SDValue ones = DAG.getConstant(-1, MVT::v4i32);
930 ones = DAG.getNode(ISD::BITCAST, dl, MVT::i128, ones);
932 // Create the 128 bit masks that have ones where the data to store is
934 SDValue lowmask, himask;
935 // if the value to store don't fill up the an entire 128 bits, zero
936 // out the last bits of the mask so that only the value we want to store
938 // this is e.g. in the case of store i32, align 2
940 Value = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, Value);
941 lowmask = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, ones, surplus);
942 lowmask = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, lowmask,
944 Value = DAG.getNode(ISD::BITCAST, dl, MVT::i128, Value);
945 Value = DAG.getNode(ISD::AND, dl, MVT::i128, Value, lowmask);
950 Value = DAG.getNode(ISD::BITCAST, dl, MVT::i128, Value);
952 // this will zero, if there are no data that goes to the high quad
953 himask = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, lowmask,
955 lowmask = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, lowmask,
958 // Load in the old data and zero out the parts that will be overwritten with
959 // the new data to store.
960 SDValue hi = DAG.getLoad(MVT::i128, dl, the_chain,
961 DAG.getNode(ISD::ADD, dl, PtrVT, basePtr,
962 DAG.getConstant( 16, PtrVT)),
964 SN->isVolatile(), SN->isNonTemporal(),
966 the_chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(1),
969 low = DAG.getNode(ISD::AND, dl, MVT::i128,
970 DAG.getNode( ISD::BITCAST, dl, MVT::i128, low),
971 DAG.getNode( ISD::XOR, dl, MVT::i128, lowmask, ones));
972 hi = DAG.getNode(ISD::AND, dl, MVT::i128,
973 DAG.getNode( ISD::BITCAST, dl, MVT::i128, hi),
974 DAG.getNode( ISD::XOR, dl, MVT::i128, himask, ones));
976 // Shift the Value to store into place. rlow contains the parts that go to
977 // the lower memory chunk, rhi has the parts that go to the upper one.
978 SDValue rlow = DAG.getNode(SPUISD::SRL_BYTES, dl, MVT::i128, Value, offset);
979 rlow = DAG.getNode(ISD::AND, dl, MVT::i128, rlow, lowmask);
980 SDValue rhi = DAG.getNode(SPUISD::SHL_BYTES, dl, MVT::i128, Value,
983 // Merge the old data and the new data and store the results
984 // Need to convert vectors here to integer as 'OR'ing floats assert
985 rlow = DAG.getNode(ISD::OR, dl, MVT::i128,
986 DAG.getNode(ISD::BITCAST, dl, MVT::i128, low),
987 DAG.getNode(ISD::BITCAST, dl, MVT::i128, rlow));
988 rhi = DAG.getNode(ISD::OR, dl, MVT::i128,
989 DAG.getNode(ISD::BITCAST, dl, MVT::i128, hi),
990 DAG.getNode(ISD::BITCAST, dl, MVT::i128, rhi));
992 low = DAG.getStore(the_chain, dl, rlow, basePtr,
994 SN->isVolatile(), SN->isNonTemporal(), 16);
995 hi = DAG.getStore(the_chain, dl, rhi,
996 DAG.getNode(ISD::ADD, dl, PtrVT, basePtr,
997 DAG.getConstant( 16, PtrVT)),
999 SN->isVolatile(), SN->isNonTemporal(), 16);
1000 result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, low.getValue(0),
1007 //! Generate the address of a constant pool entry.
1009 LowerConstantPool(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1010 EVT PtrVT = Op.getValueType();
1011 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1012 const Constant *C = CP->getConstVal();
1013 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1014 SDValue Zero = DAG.getConstant(0, PtrVT);
1015 const TargetMachine &TM = DAG.getTarget();
1016 // FIXME there is no actual debug info here
1017 DebugLoc dl = Op.getDebugLoc();
1019 if (TM.getRelocationModel() == Reloc::Static) {
1020 if (!ST->usingLargeMem()) {
1021 // Just return the SDValue with the constant pool address in it.
1022 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, CPI, Zero);
1024 SDValue Hi = DAG.getNode(SPUISD::Hi, dl, PtrVT, CPI, Zero);
1025 SDValue Lo = DAG.getNode(SPUISD::Lo, dl, PtrVT, CPI, Zero);
1026 return DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Hi, Lo);
1030 llvm_unreachable("LowerConstantPool: Relocation model other than static"
1034 //! Alternate entry point for generating the address of a constant pool entry
1036 SPU::LowerConstantPool(SDValue Op, SelectionDAG &DAG, const SPUTargetMachine &TM) {
1037 return ::LowerConstantPool(Op, DAG, TM.getSubtargetImpl());
1041 LowerJumpTable(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1042 EVT PtrVT = Op.getValueType();
1043 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1044 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1045 SDValue Zero = DAG.getConstant(0, PtrVT);
1046 const TargetMachine &TM = DAG.getTarget();
1047 // FIXME there is no actual debug info here
1048 DebugLoc dl = Op.getDebugLoc();
1050 if (TM.getRelocationModel() == Reloc::Static) {
1051 if (!ST->usingLargeMem()) {
1052 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, JTI, Zero);
1054 SDValue Hi = DAG.getNode(SPUISD::Hi, dl, PtrVT, JTI, Zero);
1055 SDValue Lo = DAG.getNode(SPUISD::Lo, dl, PtrVT, JTI, Zero);
1056 return DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Hi, Lo);
1060 llvm_unreachable("LowerJumpTable: Relocation model other than static"
1065 LowerGlobalAddress(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
1066 EVT PtrVT = Op.getValueType();
1067 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1068 const GlobalValue *GV = GSDN->getGlobal();
1069 SDValue GA = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
1070 PtrVT, GSDN->getOffset());
1071 const TargetMachine &TM = DAG.getTarget();
1072 SDValue Zero = DAG.getConstant(0, PtrVT);
1073 // FIXME there is no actual debug info here
1074 DebugLoc dl = Op.getDebugLoc();
1076 if (TM.getRelocationModel() == Reloc::Static) {
1077 if (!ST->usingLargeMem()) {
1078 return DAG.getNode(SPUISD::AFormAddr, dl, PtrVT, GA, Zero);
1080 SDValue Hi = DAG.getNode(SPUISD::Hi, dl, PtrVT, GA, Zero);
1081 SDValue Lo = DAG.getNode(SPUISD::Lo, dl, PtrVT, GA, Zero);
1082 return DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, Hi, Lo);
1085 report_fatal_error("LowerGlobalAddress: Relocation model other than static"
1091 //! Custom lower double precision floating point constants
1093 LowerConstantFP(SDValue Op, SelectionDAG &DAG) {
1094 EVT VT = Op.getValueType();
1095 // FIXME there is no actual debug info here
1096 DebugLoc dl = Op.getDebugLoc();
1098 if (VT == MVT::f64) {
1099 ConstantFPSDNode *FP = cast<ConstantFPSDNode>(Op.getNode());
1102 "LowerConstantFP: Node is not ConstantFPSDNode");
1104 uint64_t dbits = DoubleToBits(FP->getValueAPF().convertToDouble());
1105 SDValue T = DAG.getConstant(dbits, MVT::i64);
1106 SDValue Tvec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, T, T);
1107 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
1108 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Tvec));
1115 SPUTargetLowering::LowerFormalArguments(SDValue Chain,
1116 CallingConv::ID CallConv, bool isVarArg,
1117 const SmallVectorImpl<ISD::InputArg>
1119 DebugLoc dl, SelectionDAG &DAG,
1120 SmallVectorImpl<SDValue> &InVals)
1123 MachineFunction &MF = DAG.getMachineFunction();
1124 MachineFrameInfo *MFI = MF.getFrameInfo();
1125 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1126 SPUFunctionInfo *FuncInfo = MF.getInfo<SPUFunctionInfo>();
1128 unsigned ArgOffset = SPUFrameLowering::minStackSize();
1129 unsigned ArgRegIdx = 0;
1130 unsigned StackSlotSize = SPUFrameLowering::stackSlotSize();
1132 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1134 SmallVector<CCValAssign, 16> ArgLocs;
1135 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1136 getTargetMachine(), ArgLocs, *DAG.getContext());
1137 // FIXME: allow for other calling conventions
1138 CCInfo.AnalyzeFormalArguments(Ins, CCC_SPU);
1140 // Add DAG nodes to load the arguments or copy them out of registers.
1141 for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) {
1142 EVT ObjectVT = Ins[ArgNo].VT;
1143 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1145 CCValAssign &VA = ArgLocs[ArgNo];
1147 if (VA.isRegLoc()) {
1148 const TargetRegisterClass *ArgRegClass;
1150 switch (ObjectVT.getSimpleVT().SimpleTy) {
1152 report_fatal_error("LowerFormalArguments Unhandled argument type: " +
1153 Twine(ObjectVT.getEVTString()));
1155 ArgRegClass = &SPU::R8CRegClass;
1158 ArgRegClass = &SPU::R16CRegClass;
1161 ArgRegClass = &SPU::R32CRegClass;
1164 ArgRegClass = &SPU::R64CRegClass;
1167 ArgRegClass = &SPU::GPRCRegClass;
1170 ArgRegClass = &SPU::R32FPRegClass;
1173 ArgRegClass = &SPU::R64FPRegClass;
1181 ArgRegClass = &SPU::VECREGRegClass;
1185 unsigned VReg = RegInfo.createVirtualRegister(ArgRegClass);
1186 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1187 ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
1190 // We need to load the argument to a virtual register if we determined
1191 // above that we ran out of physical registers of the appropriate type
1192 // or we're forced to do vararg
1193 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset, true);
1194 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1195 ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
1196 false, false, false, 0);
1197 ArgOffset += StackSlotSize;
1200 InVals.push_back(ArgVal);
1202 Chain = ArgVal.getOperand(0);
1207 // FIXME: we should be able to query the argument registers from
1208 // tablegen generated code.
1209 static const uint16_t ArgRegs[] = {
1210 SPU::R3, SPU::R4, SPU::R5, SPU::R6, SPU::R7, SPU::R8, SPU::R9,
1211 SPU::R10, SPU::R11, SPU::R12, SPU::R13, SPU::R14, SPU::R15, SPU::R16,
1212 SPU::R17, SPU::R18, SPU::R19, SPU::R20, SPU::R21, SPU::R22, SPU::R23,
1213 SPU::R24, SPU::R25, SPU::R26, SPU::R27, SPU::R28, SPU::R29, SPU::R30,
1214 SPU::R31, SPU::R32, SPU::R33, SPU::R34, SPU::R35, SPU::R36, SPU::R37,
1215 SPU::R38, SPU::R39, SPU::R40, SPU::R41, SPU::R42, SPU::R43, SPU::R44,
1216 SPU::R45, SPU::R46, SPU::R47, SPU::R48, SPU::R49, SPU::R50, SPU::R51,
1217 SPU::R52, SPU::R53, SPU::R54, SPU::R55, SPU::R56, SPU::R57, SPU::R58,
1218 SPU::R59, SPU::R60, SPU::R61, SPU::R62, SPU::R63, SPU::R64, SPU::R65,
1219 SPU::R66, SPU::R67, SPU::R68, SPU::R69, SPU::R70, SPU::R71, SPU::R72,
1220 SPU::R73, SPU::R74, SPU::R75, SPU::R76, SPU::R77, SPU::R78, SPU::R79
1222 // size of ArgRegs array
1223 const unsigned NumArgRegs = 77;
1225 // We will spill (79-3)+1 registers to the stack
1226 SmallVector<SDValue, 79-3+1> MemOps;
1228 // Create the frame slot
1229 for (; ArgRegIdx != NumArgRegs; ++ArgRegIdx) {
1230 FuncInfo->setVarArgsFrameIndex(
1231 MFI->CreateFixedObject(StackSlotSize, ArgOffset, true));
1232 SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
1233 unsigned VReg = MF.addLiveIn(ArgRegs[ArgRegIdx], &SPU::VECREGRegClass);
1234 SDValue ArgVal = DAG.getRegister(VReg, MVT::v16i8);
1235 SDValue Store = DAG.getStore(Chain, dl, ArgVal, FIN, MachinePointerInfo(),
1237 Chain = Store.getOperand(0);
1238 MemOps.push_back(Store);
1240 // Increment address by stack slot size for the next stored argument
1241 ArgOffset += StackSlotSize;
1243 if (!MemOps.empty())
1244 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1245 &MemOps[0], MemOps.size());
1251 /// isLSAAddress - Return the immediate to use if the specified
1252 /// value is representable as a LSA address.
1253 static SDNode *isLSAAddress(SDValue Op, SelectionDAG &DAG) {
1254 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1257 int Addr = C->getZExtValue();
1258 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1259 (Addr << 14 >> 14) != Addr)
1260 return 0; // Top 14 bits have to be sext of immediate.
1262 return DAG.getConstant((int)C->getZExtValue() >> 2, MVT::i32).getNode();
1266 SPUTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1267 CallingConv::ID CallConv, bool isVarArg,
1268 bool doesNotRet, bool &isTailCall,
1269 const SmallVectorImpl<ISD::OutputArg> &Outs,
1270 const SmallVectorImpl<SDValue> &OutVals,
1271 const SmallVectorImpl<ISD::InputArg> &Ins,
1272 DebugLoc dl, SelectionDAG &DAG,
1273 SmallVectorImpl<SDValue> &InVals) const {
1274 // CellSPU target does not yet support tail call optimization.
1277 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
1278 unsigned NumOps = Outs.size();
1279 unsigned StackSlotSize = SPUFrameLowering::stackSlotSize();
1281 SmallVector<CCValAssign, 16> ArgLocs;
1282 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1283 getTargetMachine(), ArgLocs, *DAG.getContext());
1284 // FIXME: allow for other calling conventions
1285 CCInfo.AnalyzeCallOperands(Outs, CCC_SPU);
1287 const unsigned NumArgRegs = ArgLocs.size();
1290 // Handy pointer type
1291 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1293 // Set up a copy of the stack pointer for use loading and storing any
1294 // arguments that may not fit in the registers available for argument
1296 SDValue StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
1298 // Figure out which arguments are going to go in registers, and which in
1300 unsigned ArgOffset = SPUFrameLowering::minStackSize(); // Just below [LR]
1301 unsigned ArgRegIdx = 0;
1303 // Keep track of registers passing arguments
1304 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
1305 // And the arguments passed on the stack
1306 SmallVector<SDValue, 8> MemOpChains;
1308 for (; ArgRegIdx != NumOps; ++ArgRegIdx) {
1309 SDValue Arg = OutVals[ArgRegIdx];
1310 CCValAssign &VA = ArgLocs[ArgRegIdx];
1312 // PtrOff will be used to store the current argument to the stack if a
1313 // register cannot be found for it.
1314 SDValue PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1315 PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff);
1317 switch (Arg.getValueType().getSimpleVT().SimpleTy) {
1318 default: llvm_unreachable("Unexpected ValueType for argument!");
1332 if (ArgRegIdx != NumArgRegs) {
1333 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1335 MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
1336 MachinePointerInfo(),
1338 ArgOffset += StackSlotSize;
1344 // Accumulate how many bytes are to be pushed on the stack, including the
1345 // linkage area, and parameter passing area. According to the SPU ABI,
1346 // we minimally need space for [LR] and [SP].
1347 unsigned NumStackBytes = ArgOffset - SPUFrameLowering::minStackSize();
1349 // Insert a call sequence start
1350 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumStackBytes,
1353 if (!MemOpChains.empty()) {
1354 // Adjust the stack pointer for the stack arguments.
1355 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1356 &MemOpChains[0], MemOpChains.size());
1359 // Build a sequence of copy-to-reg nodes chained together with token chain
1360 // and flag operands which copy the outgoing args into the appropriate regs.
1362 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1363 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1364 RegsToPass[i].second, InFlag);
1365 InFlag = Chain.getValue(1);
1368 SmallVector<SDValue, 8> Ops;
1369 unsigned CallOpc = SPUISD::CALL;
1371 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1372 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1373 // node so that legalize doesn't hack it.
1374 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1375 const GlobalValue *GV = G->getGlobal();
1376 EVT CalleeVT = Callee.getValueType();
1377 SDValue Zero = DAG.getConstant(0, PtrVT);
1378 SDValue GA = DAG.getTargetGlobalAddress(GV, dl, CalleeVT);
1380 if (!ST->usingLargeMem()) {
1381 // Turn calls to targets that are defined (i.e., have bodies) into BRSL
1382 // style calls, otherwise, external symbols are BRASL calls. This assumes
1383 // that declared/defined symbols are in the same compilation unit and can
1384 // be reached through PC-relative jumps.
1387 // This may be an unsafe assumption for JIT and really large compilation
1389 if (GV->isDeclaration()) {
1390 Callee = DAG.getNode(SPUISD::AFormAddr, dl, CalleeVT, GA, Zero);
1392 Callee = DAG.getNode(SPUISD::PCRelAddr, dl, CalleeVT, GA, Zero);
1395 // "Large memory" mode: Turn all calls into indirect calls with a X-form
1397 Callee = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, GA, Zero);
1399 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1400 EVT CalleeVT = Callee.getValueType();
1401 SDValue Zero = DAG.getConstant(0, PtrVT);
1402 SDValue ExtSym = DAG.getTargetExternalSymbol(S->getSymbol(),
1403 Callee.getValueType());
1405 if (!ST->usingLargeMem()) {
1406 Callee = DAG.getNode(SPUISD::AFormAddr, dl, CalleeVT, ExtSym, Zero);
1408 Callee = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT, ExtSym, Zero);
1410 } else if (SDNode *Dest = isLSAAddress(Callee, DAG)) {
1411 // If this is an absolute destination address that appears to be a legal
1412 // local store address, use the munged value.
1413 Callee = SDValue(Dest, 0);
1416 Ops.push_back(Chain);
1417 Ops.push_back(Callee);
1419 // Add argument registers to the end of the list so that they are known live
1421 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1422 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1423 RegsToPass[i].second.getValueType()));
1425 if (InFlag.getNode())
1426 Ops.push_back(InFlag);
1427 // Returns a chain and a flag for retval copy to use.
1428 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Glue),
1429 &Ops[0], Ops.size());
1430 InFlag = Chain.getValue(1);
1432 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumStackBytes, true),
1433 DAG.getIntPtrConstant(0, true), InFlag);
1435 InFlag = Chain.getValue(1);
1437 // If the function returns void, just return the chain.
1441 // Now handle the return value(s)
1442 SmallVector<CCValAssign, 16> RVLocs;
1443 CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1444 getTargetMachine(), RVLocs, *DAG.getContext());
1445 CCRetInfo.AnalyzeCallResult(Ins, CCC_SPU);
1448 // If the call has results, copy the values out of the ret val registers.
1449 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1450 CCValAssign VA = RVLocs[i];
1452 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1454 Chain = Val.getValue(1);
1455 InFlag = Val.getValue(2);
1456 InVals.push_back(Val);
1463 SPUTargetLowering::LowerReturn(SDValue Chain,
1464 CallingConv::ID CallConv, bool isVarArg,
1465 const SmallVectorImpl<ISD::OutputArg> &Outs,
1466 const SmallVectorImpl<SDValue> &OutVals,
1467 DebugLoc dl, SelectionDAG &DAG) const {
1469 SmallVector<CCValAssign, 16> RVLocs;
1470 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1471 getTargetMachine(), RVLocs, *DAG.getContext());
1472 CCInfo.AnalyzeReturn(Outs, RetCC_SPU);
1474 // If this is the first return lowered for this function, add the regs to the
1475 // liveout set for the function.
1476 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1477 for (unsigned i = 0; i != RVLocs.size(); ++i)
1478 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1483 // Copy the result values into the output registers.
1484 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1485 CCValAssign &VA = RVLocs[i];
1486 assert(VA.isRegLoc() && "Can only return in registers!");
1487 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1489 Flag = Chain.getValue(1);
1493 return DAG.getNode(SPUISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1495 return DAG.getNode(SPUISD::RET_FLAG, dl, MVT::Other, Chain);
1499 //===----------------------------------------------------------------------===//
1500 // Vector related lowering:
1501 //===----------------------------------------------------------------------===//
1503 static ConstantSDNode *
1504 getVecImm(SDNode *N) {
1505 SDValue OpVal(0, 0);
1507 // Check to see if this buildvec has a single non-undef value in its elements.
1508 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
1509 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1510 if (OpVal.getNode() == 0)
1511 OpVal = N->getOperand(i);
1512 else if (OpVal != N->getOperand(i))
1516 if (OpVal.getNode() != 0) {
1517 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
1525 /// get_vec_i18imm - Test if this vector is a vector filled with the same value
1526 /// and the value fits into an unsigned 18-bit constant, and if so, return the
1528 SDValue SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
1530 if (ConstantSDNode *CN = getVecImm(N)) {
1531 uint64_t Value = CN->getZExtValue();
1532 if (ValueType == MVT::i64) {
1533 uint64_t UValue = CN->getZExtValue();
1534 uint32_t upper = uint32_t(UValue >> 32);
1535 uint32_t lower = uint32_t(UValue);
1538 Value = Value >> 32;
1540 if (Value <= 0x3ffff)
1541 return DAG.getTargetConstant(Value, ValueType);
1547 /// get_vec_i16imm - Test if this vector is a vector filled with the same value
1548 /// and the value fits into a signed 16-bit constant, and if so, return the
1550 SDValue SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
1552 if (ConstantSDNode *CN = getVecImm(N)) {
1553 int64_t Value = CN->getSExtValue();
1554 if (ValueType == MVT::i64) {
1555 uint64_t UValue = CN->getZExtValue();
1556 uint32_t upper = uint32_t(UValue >> 32);
1557 uint32_t lower = uint32_t(UValue);
1560 Value = Value >> 32;
1562 if (Value >= -(1 << 15) && Value <= ((1 << 15) - 1)) {
1563 return DAG.getTargetConstant(Value, ValueType);
1570 /// get_vec_i10imm - Test if this vector is a vector filled with the same value
1571 /// and the value fits into a signed 10-bit constant, and if so, return the
1573 SDValue SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
1575 if (ConstantSDNode *CN = getVecImm(N)) {
1576 int64_t Value = CN->getSExtValue();
1577 if (ValueType == MVT::i64) {
1578 uint64_t UValue = CN->getZExtValue();
1579 uint32_t upper = uint32_t(UValue >> 32);
1580 uint32_t lower = uint32_t(UValue);
1583 Value = Value >> 32;
1585 if (isInt<10>(Value))
1586 return DAG.getTargetConstant(Value, ValueType);
1592 /// get_vec_i8imm - Test if this vector is a vector filled with the same value
1593 /// and the value fits into a signed 8-bit constant, and if so, return the
1596 /// @note: The incoming vector is v16i8 because that's the only way we can load
1597 /// constant vectors. Thus, we test to see if the upper and lower bytes are the
1599 SDValue SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
1601 if (ConstantSDNode *CN = getVecImm(N)) {
1602 int Value = (int) CN->getZExtValue();
1603 if (ValueType == MVT::i16
1604 && Value <= 0xffff /* truncated from uint64_t */
1605 && ((short) Value >> 8) == ((short) Value & 0xff))
1606 return DAG.getTargetConstant(Value & 0xff, ValueType);
1607 else if (ValueType == MVT::i8
1608 && (Value & 0xff) == Value)
1609 return DAG.getTargetConstant(Value, ValueType);
1615 /// get_ILHUvec_imm - Test if this vector is a vector filled with the same value
1616 /// and the value fits into a signed 16-bit constant, and if so, return the
1618 SDValue SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
1620 if (ConstantSDNode *CN = getVecImm(N)) {
1621 uint64_t Value = CN->getZExtValue();
1622 if ((ValueType == MVT::i32
1623 && ((unsigned) Value & 0xffff0000) == (unsigned) Value)
1624 || (ValueType == MVT::i64 && (Value & 0xffff0000) == Value))
1625 return DAG.getTargetConstant(Value >> 16, ValueType);
1631 /// get_v4i32_imm - Catch-all for general 32-bit constant vectors
1632 SDValue SPU::get_v4i32_imm(SDNode *N, SelectionDAG &DAG) {
1633 if (ConstantSDNode *CN = getVecImm(N)) {
1634 return DAG.getTargetConstant((unsigned) CN->getZExtValue(), MVT::i32);
1640 /// get_v4i32_imm - Catch-all for general 64-bit constant vectors
1641 SDValue SPU::get_v2i64_imm(SDNode *N, SelectionDAG &DAG) {
1642 if (ConstantSDNode *CN = getVecImm(N)) {
1643 return DAG.getTargetConstant((unsigned) CN->getZExtValue(), MVT::i64);
1649 //! Lower a BUILD_VECTOR instruction creatively:
1651 LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
1652 EVT VT = Op.getValueType();
1653 EVT EltVT = VT.getVectorElementType();
1654 DebugLoc dl = Op.getDebugLoc();
1655 BuildVectorSDNode *BCN = dyn_cast<BuildVectorSDNode>(Op.getNode());
1656 assert(BCN != 0 && "Expected BuildVectorSDNode in SPU LowerBUILD_VECTOR");
1657 unsigned minSplatBits = EltVT.getSizeInBits();
1659 if (minSplatBits < 16)
1662 APInt APSplatBits, APSplatUndef;
1663 unsigned SplatBitSize;
1666 if (!BCN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
1667 HasAnyUndefs, minSplatBits)
1668 || minSplatBits < SplatBitSize)
1669 return SDValue(); // Wasn't a constant vector or splat exceeded min
1671 uint64_t SplatBits = APSplatBits.getZExtValue();
1673 switch (VT.getSimpleVT().SimpleTy) {
1675 report_fatal_error("CellSPU: Unhandled VT in LowerBUILD_VECTOR, VT = " +
1676 Twine(VT.getEVTString()));
1679 uint32_t Value32 = uint32_t(SplatBits);
1680 assert(SplatBitSize == 32
1681 && "LowerBUILD_VECTOR: Unexpected floating point vector element.");
1682 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1683 SDValue T = DAG.getConstant(Value32, MVT::i32);
1684 return DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,
1685 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, T,T,T,T));
1688 uint64_t f64val = uint64_t(SplatBits);
1689 assert(SplatBitSize == 64
1690 && "LowerBUILD_VECTOR: 64-bit float vector size > 8 bytes.");
1691 // NOTE: pretend the constant is an integer. LLVM won't load FP constants
1692 SDValue T = DAG.getConstant(f64val, MVT::i64);
1693 return DAG.getNode(ISD::BITCAST, dl, MVT::v2f64,
1694 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, T, T));
1697 // 8-bit constants have to be expanded to 16-bits
1698 unsigned short Value16 = SplatBits /* | (SplatBits << 8) */;
1699 SmallVector<SDValue, 8> Ops;
1701 Ops.assign(8, DAG.getConstant(Value16, MVT::i16));
1702 return DAG.getNode(ISD::BITCAST, dl, VT,
1703 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i16, &Ops[0], Ops.size()));
1706 unsigned short Value16 = SplatBits;
1707 SDValue T = DAG.getConstant(Value16, EltVT);
1708 SmallVector<SDValue, 8> Ops;
1711 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Ops[0], Ops.size());
1714 SDValue T = DAG.getConstant(unsigned(SplatBits), VT.getVectorElementType());
1715 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, T, T, T, T);
1718 return SPU::LowerV2I64Splat(VT, DAG, SplatBits, dl);
1726 SPU::LowerV2I64Splat(EVT OpVT, SelectionDAG& DAG, uint64_t SplatVal,
1728 uint32_t upper = uint32_t(SplatVal >> 32);
1729 uint32_t lower = uint32_t(SplatVal);
1731 if (upper == lower) {
1732 // Magic constant that can be matched by IL, ILA, et. al.
1733 SDValue Val = DAG.getTargetConstant(upper, MVT::i32);
1734 return DAG.getNode(ISD::BITCAST, dl, OpVT,
1735 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1736 Val, Val, Val, Val));
1738 bool upper_special, lower_special;
1740 // NOTE: This code creates common-case shuffle masks that can be easily
1741 // detected as common expressions. It is not attempting to create highly
1742 // specialized masks to replace any and all 0's, 0xff's and 0x80's.
1744 // Detect if the upper or lower half is a special shuffle mask pattern:
1745 upper_special = (upper == 0 || upper == 0xffffffff || upper == 0x80000000);
1746 lower_special = (lower == 0 || lower == 0xffffffff || lower == 0x80000000);
1748 // Both upper and lower are special, lower to a constant pool load:
1749 if (lower_special && upper_special) {
1750 SDValue UpperVal = DAG.getConstant(upper, MVT::i32);
1751 SDValue LowerVal = DAG.getConstant(lower, MVT::i32);
1752 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1753 UpperVal, LowerVal, UpperVal, LowerVal);
1754 return DAG.getNode(ISD::BITCAST, dl, OpVT, BV);
1759 SmallVector<SDValue, 16> ShufBytes;
1762 // Create lower vector if not a special pattern
1763 if (!lower_special) {
1764 SDValue LO32C = DAG.getConstant(lower, MVT::i32);
1765 LO32 = DAG.getNode(ISD::BITCAST, dl, OpVT,
1766 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1767 LO32C, LO32C, LO32C, LO32C));
1770 // Create upper vector if not a special pattern
1771 if (!upper_special) {
1772 SDValue HI32C = DAG.getConstant(upper, MVT::i32);
1773 HI32 = DAG.getNode(ISD::BITCAST, dl, OpVT,
1774 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1775 HI32C, HI32C, HI32C, HI32C));
1778 // If either upper or lower are special, then the two input operands are
1779 // the same (basically, one of them is a "don't care")
1785 for (int i = 0; i < 4; ++i) {
1787 for (int j = 0; j < 4; ++j) {
1789 bool process_upper, process_lower;
1791 process_upper = (upper_special && (i & 1) == 0);
1792 process_lower = (lower_special && (i & 1) == 1);
1794 if (process_upper || process_lower) {
1795 if ((process_upper && upper == 0)
1796 || (process_lower && lower == 0))
1798 else if ((process_upper && upper == 0xffffffff)
1799 || (process_lower && lower == 0xffffffff))
1801 else if ((process_upper && upper == 0x80000000)
1802 || (process_lower && lower == 0x80000000))
1803 val |= (j == 0 ? 0xe0 : 0x80);
1805 val |= i * 4 + j + ((i & 1) * 16);
1808 ShufBytes.push_back(DAG.getConstant(val, MVT::i32));
1811 return DAG.getNode(SPUISD::SHUFB, dl, OpVT, HI32, LO32,
1812 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
1813 &ShufBytes[0], ShufBytes.size()));
1817 /// LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3) to something on
1818 /// which the Cell can operate. The code inspects V3 to ascertain whether the
1819 /// permutation vector, V3, is monotonically increasing with one "exception"
1820 /// element, e.g., (0, 1, _, 3). If this is the case, then generate a
1821 /// SHUFFLE_MASK synthetic instruction. Otherwise, spill V3 to the constant pool.
1822 /// In either case, the net result is going to eventually invoke SHUFB to
1823 /// permute/shuffle the bytes from V1 and V2.
1825 /// SHUFFLE_MASK is eventually selected as one of the C*D instructions, generate
1826 /// control word for byte/halfword/word insertion. This takes care of a single
1827 /// element move from V2 into V1.
1829 /// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
1830 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
1831 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
1832 SDValue V1 = Op.getOperand(0);
1833 SDValue V2 = Op.getOperand(1);
1834 DebugLoc dl = Op.getDebugLoc();
1836 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1838 // If we have a single element being moved from V1 to V2, this can be handled
1839 // using the C*[DX] compute mask instructions, but the vector elements have
1840 // to be monotonically increasing with one exception element, and the source
1841 // slot of the element to move must be the same as the destination.
1842 EVT VecVT = V1.getValueType();
1843 EVT EltVT = VecVT.getVectorElementType();
1844 unsigned EltsFromV2 = 0;
1845 unsigned V2EltOffset = 0;
1846 unsigned V2EltIdx0 = 0;
1847 unsigned CurrElt = 0;
1848 unsigned MaxElts = VecVT.getVectorNumElements();
1849 unsigned PrevElt = 0;
1850 bool monotonic = true;
1853 EVT maskVT; // which of the c?d instructions to use
1855 if (EltVT == MVT::i8) {
1857 maskVT = MVT::v16i8;
1858 } else if (EltVT == MVT::i16) {
1860 maskVT = MVT::v8i16;
1861 } else if (EltVT == MVT::i32 || EltVT == MVT::f32) {
1863 maskVT = MVT::v4i32;
1864 } else if (EltVT == MVT::i64 || EltVT == MVT::f64) {
1866 maskVT = MVT::v2i64;
1868 llvm_unreachable("Unhandled vector type in LowerVECTOR_SHUFFLE");
1870 for (unsigned i = 0; i != MaxElts; ++i) {
1871 if (SVN->getMaskElt(i) < 0)
1874 unsigned SrcElt = SVN->getMaskElt(i);
1877 if (SrcElt >= V2EltIdx0) {
1878 // TODO: optimize for the monotonic case when several consecutive
1879 // elements are taken form V2. Do we ever get such a case?
1880 if (EltsFromV2 == 0 && CurrElt == (SrcElt - V2EltIdx0))
1881 V2EltOffset = (SrcElt - V2EltIdx0) * (EltVT.getSizeInBits()/8);
1885 } else if (CurrElt != SrcElt) {
1893 if (PrevElt > 0 && SrcElt < MaxElts) {
1894 if ((PrevElt == SrcElt - 1)
1895 || (PrevElt == MaxElts - 1 && SrcElt == 0)) {
1900 } else if (i == 0 || (PrevElt==0 && SrcElt==1)) {
1901 // First time or after a "wrap around"
1905 // This isn't a rotation, takes elements from vector 2
1911 if (EltsFromV2 == 1 && monotonic) {
1912 // Compute mask and shuffle
1913 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1915 // As SHUFFLE_MASK becomes a c?d instruction, feed it an address
1916 // R1 ($sp) is used here only as it is guaranteed to have last bits zero
1917 SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
1918 DAG.getRegister(SPU::R1, PtrVT),
1919 DAG.getConstant(V2EltOffset, MVT::i32));
1920 SDValue ShufMaskOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl,
1923 // Use shuffle mask in SHUFB synthetic instruction:
1924 return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V2, V1,
1926 } else if (rotate) {
1929 rotamt *= EltVT.getSizeInBits()/8;
1930 return DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, V1.getValueType(),
1931 V1, DAG.getConstant(rotamt, MVT::i16));
1933 // Convert the SHUFFLE_VECTOR mask's input element units to the
1935 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
1937 SmallVector<SDValue, 16> ResultMask;
1938 for (unsigned i = 0, e = MaxElts; i != e; ++i) {
1939 unsigned SrcElt = SVN->getMaskElt(i) < 0 ? 0 : SVN->getMaskElt(i);
1941 for (unsigned j = 0; j < BytesPerElement; ++j)
1942 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,MVT::i8));
1944 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
1945 &ResultMask[0], ResultMask.size());
1946 return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V1, V2, VPermMask);
1950 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
1951 SDValue Op0 = Op.getOperand(0); // Op0 = the scalar
1952 DebugLoc dl = Op.getDebugLoc();
1954 if (Op0.getNode()->getOpcode() == ISD::Constant) {
1955 // For a constant, build the appropriate constant vector, which will
1956 // eventually simplify to a vector register load.
1958 ConstantSDNode *CN = cast<ConstantSDNode>(Op0.getNode());
1959 SmallVector<SDValue, 16> ConstVecValues;
1963 // Create a constant vector:
1964 switch (Op.getValueType().getSimpleVT().SimpleTy) {
1965 default: llvm_unreachable("Unexpected constant value type in "
1966 "LowerSCALAR_TO_VECTOR");
1967 case MVT::v16i8: n_copies = 16; VT = MVT::i8; break;
1968 case MVT::v8i16: n_copies = 8; VT = MVT::i16; break;
1969 case MVT::v4i32: n_copies = 4; VT = MVT::i32; break;
1970 case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
1971 case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
1972 case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
1975 SDValue CValue = DAG.getConstant(CN->getZExtValue(), VT);
1976 for (size_t j = 0; j < n_copies; ++j)
1977 ConstVecValues.push_back(CValue);
1979 return DAG.getNode(ISD::BUILD_VECTOR, dl, Op.getValueType(),
1980 &ConstVecValues[0], ConstVecValues.size());
1982 // Otherwise, copy the value from one register to another:
1983 switch (Op0.getValueType().getSimpleVT().SimpleTy) {
1984 default: llvm_unreachable("Unexpected value type in LowerSCALAR_TO_VECTOR");
1991 return DAG.getNode(SPUISD::PREFSLOT2VEC, dl, Op.getValueType(), Op0, Op0);
1996 static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
1997 EVT VT = Op.getValueType();
1998 SDValue N = Op.getOperand(0);
1999 SDValue Elt = Op.getOperand(1);
2000 DebugLoc dl = Op.getDebugLoc();
2003 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
2004 // Constant argument:
2005 int EltNo = (int) C->getZExtValue();
2008 if (VT == MVT::i8 && EltNo >= 16)
2009 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i8 extraction slot > 15");
2010 else if (VT == MVT::i16 && EltNo >= 8)
2011 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i16 extraction slot > 7");
2012 else if (VT == MVT::i32 && EltNo >= 4)
2013 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i32 extraction slot > 4");
2014 else if (VT == MVT::i64 && EltNo >= 2)
2015 llvm_unreachable("SPU LowerEXTRACT_VECTOR_ELT: i64 extraction slot > 2");
2017 if (EltNo == 0 && (VT == MVT::i32 || VT == MVT::i64)) {
2018 // i32 and i64: Element 0 is the preferred slot
2019 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT, N);
2022 // Need to generate shuffle mask and extract:
2023 int prefslot_begin = -1, prefslot_end = -1;
2024 int elt_byte = EltNo * VT.getSizeInBits() / 8;
2026 switch (VT.getSimpleVT().SimpleTy) {
2027 default: llvm_unreachable("Invalid value type!");
2029 prefslot_begin = prefslot_end = 3;
2033 prefslot_begin = 2; prefslot_end = 3;
2038 prefslot_begin = 0; prefslot_end = 3;
2043 prefslot_begin = 0; prefslot_end = 7;
2048 assert(prefslot_begin != -1 && prefslot_end != -1 &&
2049 "LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
2051 unsigned int ShufBytes[16] = {
2052 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
2054 for (int i = 0; i < 16; ++i) {
2055 // zero fill uppper part of preferred slot, don't care about the
2057 unsigned int mask_val;
2058 if (i <= prefslot_end) {
2060 ((i < prefslot_begin)
2062 : elt_byte + (i - prefslot_begin));
2064 ShufBytes[i] = mask_val;
2066 ShufBytes[i] = ShufBytes[i % (prefslot_end + 1)];
2069 SDValue ShufMask[4];
2070 for (unsigned i = 0; i < sizeof(ShufMask)/sizeof(ShufMask[0]); ++i) {
2071 unsigned bidx = i * 4;
2072 unsigned int bits = ((ShufBytes[bidx] << 24) |
2073 (ShufBytes[bidx+1] << 16) |
2074 (ShufBytes[bidx+2] << 8) |
2076 ShufMask[i] = DAG.getConstant(bits, MVT::i32);
2079 SDValue ShufMaskVec =
2080 DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2081 &ShufMask[0], sizeof(ShufMask)/sizeof(ShufMask[0]));
2083 retval = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
2084 DAG.getNode(SPUISD::SHUFB, dl, N.getValueType(),
2085 N, N, ShufMaskVec));
2087 // Variable index: Rotate the requested element into slot 0, then replicate
2088 // slot 0 across the vector
2089 EVT VecVT = N.getValueType();
2090 if (!VecVT.isSimple() || !VecVT.isVector()) {
2091 report_fatal_error("LowerEXTRACT_VECTOR_ELT: Must have a simple, 128-bit"
2095 // Make life easier by making sure the index is zero-extended to i32
2096 if (Elt.getValueType() != MVT::i32)
2097 Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Elt);
2099 // Scale the index to a bit/byte shift quantity
2101 APInt(32, uint64_t(16 / N.getValueType().getVectorNumElements()), false);
2102 unsigned scaleShift = scaleFactor.logBase2();
2105 if (scaleShift > 0) {
2106 // Scale the shift factor:
2107 Elt = DAG.getNode(ISD::SHL, dl, MVT::i32, Elt,
2108 DAG.getConstant(scaleShift, MVT::i32));
2111 vecShift = DAG.getNode(SPUISD::SHL_BYTES, dl, VecVT, N, Elt);
2113 // Replicate the bytes starting at byte 0 across the entire vector (for
2114 // consistency with the notion of a unified register set)
2117 switch (VT.getSimpleVT().SimpleTy) {
2119 report_fatal_error("LowerEXTRACT_VECTOR_ELT(varable): Unhandled vector"
2123 SDValue factor = DAG.getConstant(0x00000000, MVT::i32);
2124 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2125 factor, factor, factor, factor);
2129 SDValue factor = DAG.getConstant(0x00010001, MVT::i32);
2130 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2131 factor, factor, factor, factor);
2136 SDValue factor = DAG.getConstant(0x00010203, MVT::i32);
2137 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2138 factor, factor, factor, factor);
2143 SDValue loFactor = DAG.getConstant(0x00010203, MVT::i32);
2144 SDValue hiFactor = DAG.getConstant(0x04050607, MVT::i32);
2145 replicate = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2146 loFactor, hiFactor, loFactor, hiFactor);
2151 retval = DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT,
2152 DAG.getNode(SPUISD::SHUFB, dl, VecVT,
2153 vecShift, vecShift, replicate));
2159 static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
2160 SDValue VecOp = Op.getOperand(0);
2161 SDValue ValOp = Op.getOperand(1);
2162 SDValue IdxOp = Op.getOperand(2);
2163 DebugLoc dl = Op.getDebugLoc();
2164 EVT VT = Op.getValueType();
2165 EVT eltVT = ValOp.getValueType();
2167 // use 0 when the lane to insert to is 'undef'
2169 if (IdxOp.getOpcode() != ISD::UNDEF) {
2170 ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
2171 assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
2172 Offset = (CN->getSExtValue()) * eltVT.getSizeInBits()/8;
2175 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2176 // Use $sp ($1) because it's always 16-byte aligned and it's available:
2177 SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
2178 DAG.getRegister(SPU::R1, PtrVT),
2179 DAG.getConstant(Offset, PtrVT));
2180 // widen the mask when dealing with half vectors
2181 EVT maskVT = EVT::getVectorVT(*(DAG.getContext()), VT.getVectorElementType(),
2182 128/ VT.getVectorElementType().getSizeInBits());
2183 SDValue ShufMask = DAG.getNode(SPUISD::SHUFFLE_MASK, dl, maskVT, Pointer);
2186 DAG.getNode(SPUISD::SHUFB, dl, VT,
2187 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, ValOp),
2189 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, ShufMask));
2194 static SDValue LowerI8Math(SDValue Op, SelectionDAG &DAG, unsigned Opc,
2195 const TargetLowering &TLI)
2197 SDValue N0 = Op.getOperand(0); // Everything has at least one operand
2198 DebugLoc dl = Op.getDebugLoc();
2199 EVT ShiftVT = TLI.getShiftAmountTy(N0.getValueType());
2201 assert(Op.getValueType() == MVT::i8);
2204 llvm_unreachable("Unhandled i8 math operator");
2206 // 8-bit addition: Promote the arguments up to 16-bits and truncate
2208 SDValue N1 = Op.getOperand(1);
2209 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2210 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N1);
2211 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2212 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2217 // 8-bit subtraction: Promote the arguments up to 16-bits and truncate
2219 SDValue N1 = Op.getOperand(1);
2220 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2221 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N1);
2222 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2223 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2227 SDValue N1 = Op.getOperand(1);
2228 EVT N1VT = N1.getValueType();
2230 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, N0);
2231 if (!N1VT.bitsEq(ShiftVT)) {
2232 unsigned N1Opc = N1.getValueType().bitsLT(ShiftVT)
2235 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2238 // Replicate lower 8-bits into upper 8:
2240 DAG.getNode(ISD::OR, dl, MVT::i16, N0,
2241 DAG.getNode(ISD::SHL, dl, MVT::i16,
2242 N0, DAG.getConstant(8, MVT::i32)));
2244 // Truncate back down to i8
2245 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2246 DAG.getNode(Opc, dl, MVT::i16, ExpandArg, N1));
2250 SDValue N1 = Op.getOperand(1);
2251 EVT N1VT = N1.getValueType();
2253 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, N0);
2254 if (!N1VT.bitsEq(ShiftVT)) {
2255 unsigned N1Opc = ISD::ZERO_EXTEND;
2257 if (N1.getValueType().bitsGT(ShiftVT))
2258 N1Opc = ISD::TRUNCATE;
2260 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2263 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2264 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2267 SDValue N1 = Op.getOperand(1);
2268 EVT N1VT = N1.getValueType();
2270 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2271 if (!N1VT.bitsEq(ShiftVT)) {
2272 unsigned N1Opc = ISD::SIGN_EXTEND;
2274 if (N1VT.bitsGT(ShiftVT))
2275 N1Opc = ISD::TRUNCATE;
2276 N1 = DAG.getNode(N1Opc, dl, ShiftVT, N1);
2279 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2280 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2283 SDValue N1 = Op.getOperand(1);
2285 N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N0);
2286 N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i16, N1);
2287 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
2288 DAG.getNode(Opc, dl, MVT::i16, N0, N1));
2293 //! Lower byte immediate operations for v16i8 vectors:
2295 LowerByteImmed(SDValue Op, SelectionDAG &DAG) {
2298 EVT VT = Op.getValueType();
2299 DebugLoc dl = Op.getDebugLoc();
2301 ConstVec = Op.getOperand(0);
2302 Arg = Op.getOperand(1);
2303 if (ConstVec.getNode()->getOpcode() != ISD::BUILD_VECTOR) {
2304 if (ConstVec.getNode()->getOpcode() == ISD::BITCAST) {
2305 ConstVec = ConstVec.getOperand(0);
2307 ConstVec = Op.getOperand(1);
2308 Arg = Op.getOperand(0);
2309 if (ConstVec.getNode()->getOpcode() == ISD::BITCAST) {
2310 ConstVec = ConstVec.getOperand(0);
2315 if (ConstVec.getNode()->getOpcode() == ISD::BUILD_VECTOR) {
2316 BuildVectorSDNode *BCN = dyn_cast<BuildVectorSDNode>(ConstVec.getNode());
2317 assert(BCN != 0 && "Expected BuildVectorSDNode in SPU LowerByteImmed");
2319 APInt APSplatBits, APSplatUndef;
2320 unsigned SplatBitSize;
2322 unsigned minSplatBits = VT.getVectorElementType().getSizeInBits();
2324 if (BCN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
2325 HasAnyUndefs, minSplatBits)
2326 && minSplatBits <= SplatBitSize) {
2327 uint64_t SplatBits = APSplatBits.getZExtValue();
2328 SDValue tc = DAG.getTargetConstant(SplatBits & 0xff, MVT::i8);
2330 SmallVector<SDValue, 16> tcVec;
2331 tcVec.assign(16, tc);
2332 return DAG.getNode(Op.getNode()->getOpcode(), dl, VT, Arg,
2333 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &tcVec[0], tcVec.size()));
2337 // These operations (AND, OR, XOR) are legal, they just couldn't be custom
2338 // lowered. Return the operation, rather than a null SDValue.
2342 //! Custom lowering for CTPOP (count population)
2344 Custom lowering code that counts the number ones in the input
2345 operand. SPU has such an instruction, but it counts the number of
2346 ones per byte, which then have to be accumulated.
2348 static SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) {
2349 EVT VT = Op.getValueType();
2350 EVT vecVT = EVT::getVectorVT(*DAG.getContext(),
2351 VT, (128 / VT.getSizeInBits()));
2352 DebugLoc dl = Op.getDebugLoc();
2354 switch (VT.getSimpleVT().SimpleTy) {
2355 default: llvm_unreachable("Invalid value type!");
2357 SDValue N = Op.getOperand(0);
2358 SDValue Elt0 = DAG.getConstant(0, MVT::i32);
2360 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2361 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2363 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i8, CNTB, Elt0);
2367 MachineFunction &MF = DAG.getMachineFunction();
2368 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2370 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
2372 SDValue N = Op.getOperand(0);
2373 SDValue Elt0 = DAG.getConstant(0, MVT::i16);
2374 SDValue Mask0 = DAG.getConstant(0x0f, MVT::i16);
2375 SDValue Shift1 = DAG.getConstant(8, MVT::i32);
2377 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2378 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2380 // CNTB_result becomes the chain to which all of the virtual registers
2381 // CNTB_reg, SUM1_reg become associated:
2382 SDValue CNTB_result =
2383 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, CNTB, Elt0);
2385 SDValue CNTB_rescopy =
2386 DAG.getCopyToReg(CNTB_result, dl, CNTB_reg, CNTB_result);
2388 SDValue Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i16);
2390 return DAG.getNode(ISD::AND, dl, MVT::i16,
2391 DAG.getNode(ISD::ADD, dl, MVT::i16,
2392 DAG.getNode(ISD::SRL, dl, MVT::i16,
2399 MachineFunction &MF = DAG.getMachineFunction();
2400 MachineRegisterInfo &RegInfo = MF.getRegInfo();
2402 unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2403 unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
2405 SDValue N = Op.getOperand(0);
2406 SDValue Elt0 = DAG.getConstant(0, MVT::i32);
2407 SDValue Mask0 = DAG.getConstant(0xff, MVT::i32);
2408 SDValue Shift1 = DAG.getConstant(16, MVT::i32);
2409 SDValue Shift2 = DAG.getConstant(8, MVT::i32);
2411 SDValue Promote = DAG.getNode(SPUISD::PREFSLOT2VEC, dl, vecVT, N, N);
2412 SDValue CNTB = DAG.getNode(SPUISD::CNTB, dl, vecVT, Promote);
2414 // CNTB_result becomes the chain to which all of the virtual registers
2415 // CNTB_reg, SUM1_reg become associated:
2416 SDValue CNTB_result =
2417 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32, CNTB, Elt0);
2419 SDValue CNTB_rescopy =
2420 DAG.getCopyToReg(CNTB_result, dl, CNTB_reg, CNTB_result);
2423 DAG.getNode(ISD::SRL, dl, MVT::i32,
2424 DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i32),
2428 DAG.getNode(ISD::ADD, dl, MVT::i32, Comp1,
2429 DAG.getCopyFromReg(CNTB_rescopy, dl, CNTB_reg, MVT::i32));
2431 SDValue Sum1_rescopy =
2432 DAG.getCopyToReg(CNTB_result, dl, SUM1_reg, Sum1);
2435 DAG.getNode(ISD::SRL, dl, MVT::i32,
2436 DAG.getCopyFromReg(Sum1_rescopy, dl, SUM1_reg, MVT::i32),
2439 DAG.getNode(ISD::ADD, dl, MVT::i32, Comp2,
2440 DAG.getCopyFromReg(Sum1_rescopy, dl, SUM1_reg, MVT::i32));
2442 return DAG.getNode(ISD::AND, dl, MVT::i32, Sum2, Mask0);
2452 //! Lower ISD::FP_TO_SINT, ISD::FP_TO_UINT for i32
2454 f32->i32 passes through unchanged, whereas f64->i32 expands to a libcall.
2455 All conversions to i64 are expanded to a libcall.
2457 static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2458 const SPUTargetLowering &TLI) {
2459 EVT OpVT = Op.getValueType();
2460 SDValue Op0 = Op.getOperand(0);
2461 EVT Op0VT = Op0.getValueType();
2463 if ((OpVT == MVT::i32 && Op0VT == MVT::f64)
2464 || OpVT == MVT::i64) {
2465 // Convert f32 / f64 to i32 / i64 via libcall.
2467 (Op.getOpcode() == ISD::FP_TO_SINT)
2468 ? RTLIB::getFPTOSINT(Op0VT, OpVT)
2469 : RTLIB::getFPTOUINT(Op0VT, OpVT);
2470 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpectd fp-to-int conversion!");
2472 return ExpandLibCall(LC, Op, DAG, false, Dummy, TLI);
2478 //! Lower ISD::SINT_TO_FP, ISD::UINT_TO_FP for i32
2480 i32->f32 passes through unchanged, whereas i32->f64 is expanded to a libcall.
2481 All conversions from i64 are expanded to a libcall.
2483 static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2484 const SPUTargetLowering &TLI) {
2485 EVT OpVT = Op.getValueType();
2486 SDValue Op0 = Op.getOperand(0);
2487 EVT Op0VT = Op0.getValueType();
2489 if ((OpVT == MVT::f64 && Op0VT == MVT::i32)
2490 || Op0VT == MVT::i64) {
2491 // Convert i32, i64 to f64 via libcall:
2493 (Op.getOpcode() == ISD::SINT_TO_FP)
2494 ? RTLIB::getSINTTOFP(Op0VT, OpVT)
2495 : RTLIB::getUINTTOFP(Op0VT, OpVT);
2496 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpectd int-to-fp conversion!");
2498 return ExpandLibCall(LC, Op, DAG, false, Dummy, TLI);
2504 //! Lower ISD::SETCC
2506 This handles MVT::f64 (double floating point) condition lowering
2508 static SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG,
2509 const TargetLowering &TLI) {
2510 CondCodeSDNode *CC = dyn_cast<CondCodeSDNode>(Op.getOperand(2));
2511 DebugLoc dl = Op.getDebugLoc();
2512 assert(CC != 0 && "LowerSETCC: CondCodeSDNode should not be null here!\n");
2514 SDValue lhs = Op.getOperand(0);
2515 SDValue rhs = Op.getOperand(1);
2516 EVT lhsVT = lhs.getValueType();
2517 assert(lhsVT == MVT::f64 && "LowerSETCC: type other than MVT::64\n");
2519 EVT ccResultVT = TLI.getSetCCResultType(lhs.getValueType());
2520 APInt ccResultOnes = APInt::getAllOnesValue(ccResultVT.getSizeInBits());
2521 EVT IntVT(MVT::i64);
2523 // Take advantage of the fact that (truncate (sra arg, 32)) is efficiently
2524 // selected to a NOP:
2525 SDValue i64lhs = DAG.getNode(ISD::BITCAST, dl, IntVT, lhs);
2527 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
2528 DAG.getNode(ISD::SRL, dl, IntVT,
2529 i64lhs, DAG.getConstant(32, MVT::i32)));
2530 SDValue lhsHi32abs =
2531 DAG.getNode(ISD::AND, dl, MVT::i32,
2532 lhsHi32, DAG.getConstant(0x7fffffff, MVT::i32));
2534 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, i64lhs);
2536 // SETO and SETUO only use the lhs operand:
2537 if (CC->get() == ISD::SETO) {
2538 // Evaluates to true if Op0 is not [SQ]NaN - lowers to the inverse of
2540 APInt ccResultAllOnes = APInt::getAllOnesValue(ccResultVT.getSizeInBits());
2541 return DAG.getNode(ISD::XOR, dl, ccResultVT,
2542 DAG.getSetCC(dl, ccResultVT,
2543 lhs, DAG.getConstantFP(0.0, lhsVT),
2545 DAG.getConstant(ccResultAllOnes, ccResultVT));
2546 } else if (CC->get() == ISD::SETUO) {
2547 // Evaluates to true if Op0 is [SQ]NaN
2548 return DAG.getNode(ISD::AND, dl, ccResultVT,
2549 DAG.getSetCC(dl, ccResultVT,
2551 DAG.getConstant(0x7ff00000, MVT::i32),
2553 DAG.getSetCC(dl, ccResultVT,
2555 DAG.getConstant(0, MVT::i32),
2559 SDValue i64rhs = DAG.getNode(ISD::BITCAST, dl, IntVT, rhs);
2561 DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
2562 DAG.getNode(ISD::SRL, dl, IntVT,
2563 i64rhs, DAG.getConstant(32, MVT::i32)));
2565 // If a value is negative, subtract from the sign magnitude constant:
2566 SDValue signMag2TC = DAG.getConstant(0x8000000000000000ULL, IntVT);
2568 // Convert the sign-magnitude representation into 2's complement:
2569 SDValue lhsSelectMask = DAG.getNode(ISD::SRA, dl, ccResultVT,
2570 lhsHi32, DAG.getConstant(31, MVT::i32));
2571 SDValue lhsSignMag2TC = DAG.getNode(ISD::SUB, dl, IntVT, signMag2TC, i64lhs);
2573 DAG.getNode(ISD::SELECT, dl, IntVT,
2574 lhsSelectMask, lhsSignMag2TC, i64lhs);
2576 SDValue rhsSelectMask = DAG.getNode(ISD::SRA, dl, ccResultVT,
2577 rhsHi32, DAG.getConstant(31, MVT::i32));
2578 SDValue rhsSignMag2TC = DAG.getNode(ISD::SUB, dl, IntVT, signMag2TC, i64rhs);
2580 DAG.getNode(ISD::SELECT, dl, IntVT,
2581 rhsSelectMask, rhsSignMag2TC, i64rhs);
2585 switch (CC->get()) {
2588 compareOp = ISD::SETEQ; break;
2591 compareOp = ISD::SETGT; break;
2594 compareOp = ISD::SETGE; break;
2597 compareOp = ISD::SETLT; break;
2600 compareOp = ISD::SETLE; break;
2603 compareOp = ISD::SETNE; break;
2605 report_fatal_error("CellSPU ISel Select: unimplemented f64 condition");
2609 DAG.getSetCC(dl, ccResultVT, lhsSelect, rhsSelect,
2610 (ISD::CondCode) compareOp);
2612 if ((CC->get() & 0x8) == 0) {
2613 // Ordered comparison:
2614 SDValue lhsNaN = DAG.getSetCC(dl, ccResultVT,
2615 lhs, DAG.getConstantFP(0.0, MVT::f64),
2617 SDValue rhsNaN = DAG.getSetCC(dl, ccResultVT,
2618 rhs, DAG.getConstantFP(0.0, MVT::f64),
2620 SDValue ordered = DAG.getNode(ISD::AND, dl, ccResultVT, lhsNaN, rhsNaN);
2622 result = DAG.getNode(ISD::AND, dl, ccResultVT, ordered, result);
2628 //! Lower ISD::SELECT_CC
2630 ISD::SELECT_CC can (generally) be implemented directly on the SPU using the
2633 \note Need to revisit this in the future: if the code path through the true
2634 and false value computations is longer than the latency of a branch (6
2635 cycles), then it would be more advantageous to branch and insert a new basic
2636 block and branch on the condition. However, this code does not make that
2637 assumption, given the simplisitc uses so far.
2640 static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
2641 const TargetLowering &TLI) {
2642 EVT VT = Op.getValueType();
2643 SDValue lhs = Op.getOperand(0);
2644 SDValue rhs = Op.getOperand(1);
2645 SDValue trueval = Op.getOperand(2);
2646 SDValue falseval = Op.getOperand(3);
2647 SDValue condition = Op.getOperand(4);
2648 DebugLoc dl = Op.getDebugLoc();
2650 // NOTE: SELB's arguments: $rA, $rB, $mask
2652 // SELB selects bits from $rA where bits in $mask are 0, bits from $rB
2653 // where bits in $mask are 1. CCond will be inverted, having 1s where the
2654 // condition was true and 0s where the condition was false. Hence, the
2655 // arguments to SELB get reversed.
2657 // Note: Really should be ISD::SELECT instead of SPUISD::SELB, but LLVM's
2658 // legalizer insists on combining SETCC/SELECT into SELECT_CC, so we end up
2659 // with another "cannot select select_cc" assert:
2661 SDValue compare = DAG.getNode(ISD::SETCC, dl,
2662 TLI.getSetCCResultType(Op.getValueType()),
2663 lhs, rhs, condition);
2664 return DAG.getNode(SPUISD::SELB, dl, VT, falseval, trueval, compare);
2667 //! Custom lower ISD::TRUNCATE
2668 static SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG)
2670 // Type to truncate to
2671 EVT VT = Op.getValueType();
2672 MVT simpleVT = VT.getSimpleVT();
2673 EVT VecVT = EVT::getVectorVT(*DAG.getContext(),
2674 VT, (128 / VT.getSizeInBits()));
2675 DebugLoc dl = Op.getDebugLoc();
2677 // Type to truncate from
2678 SDValue Op0 = Op.getOperand(0);
2679 EVT Op0VT = Op0.getValueType();
2681 if (Op0VT == MVT::i128 && simpleVT == MVT::i64) {
2682 // Create shuffle mask, least significant doubleword of quadword
2683 unsigned maskHigh = 0x08090a0b;
2684 unsigned maskLow = 0x0c0d0e0f;
2685 // Use a shuffle to perform the truncation
2686 SDValue shufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2687 DAG.getConstant(maskHigh, MVT::i32),
2688 DAG.getConstant(maskLow, MVT::i32),
2689 DAG.getConstant(maskHigh, MVT::i32),
2690 DAG.getConstant(maskLow, MVT::i32));
2692 SDValue truncShuffle = DAG.getNode(SPUISD::SHUFB, dl, VecVT,
2693 Op0, Op0, shufMask);
2695 return DAG.getNode(SPUISD::VEC2PREFSLOT, dl, VT, truncShuffle);
2698 return SDValue(); // Leave the truncate unmolested
2702 * Emit the instruction sequence for i64/i32 -> i128 sign extend. The basic
2703 * algorithm is to duplicate the sign bit using rotmai to generate at
2704 * least one byte full of sign bits. Then propagate the "sign-byte" into
2705 * the leftmost words and the i64/i32 into the rightmost words using shufb.
2707 * @param Op The sext operand
2708 * @param DAG The current DAG
2709 * @return The SDValue with the entire instruction sequence
2711 static SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG)
2713 DebugLoc dl = Op.getDebugLoc();
2715 // Type to extend to
2716 MVT OpVT = Op.getValueType().getSimpleVT();
2718 // Type to extend from
2719 SDValue Op0 = Op.getOperand(0);
2720 MVT Op0VT = Op0.getValueType().getSimpleVT();
2722 // extend i8 & i16 via i32
2723 if (Op0VT == MVT::i8 || Op0VT == MVT::i16) {
2724 Op0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, Op0);
2728 // The type to extend to needs to be a i128 and
2729 // the type to extend from needs to be i64 or i32.
2730 assert((OpVT == MVT::i128 && (Op0VT == MVT::i64 || Op0VT == MVT::i32)) &&
2731 "LowerSIGN_EXTEND: input and/or output operand have wrong size");
2734 // Create shuffle mask
2735 unsigned mask1 = 0x10101010; // byte 0 - 3 and 4 - 7
2736 unsigned mask2 = Op0VT == MVT::i64 ? 0x00010203 : 0x10101010; // byte 8 - 11
2737 unsigned mask3 = Op0VT == MVT::i64 ? 0x04050607 : 0x00010203; // byte 12 - 15
2738 SDValue shufMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
2739 DAG.getConstant(mask1, MVT::i32),
2740 DAG.getConstant(mask1, MVT::i32),
2741 DAG.getConstant(mask2, MVT::i32),
2742 DAG.getConstant(mask3, MVT::i32));
2744 // Word wise arithmetic right shift to generate at least one byte
2745 // that contains sign bits.
2746 MVT mvt = Op0VT == MVT::i64 ? MVT::v2i64 : MVT::v4i32;
2747 SDValue sraVal = DAG.getNode(ISD::SRA,
2750 DAG.getNode(SPUISD::PREFSLOT2VEC, dl, mvt, Op0, Op0),
2751 DAG.getConstant(31, MVT::i32));
2753 // reinterpret as a i128 (SHUFB requires it). This gets lowered away.
2754 SDValue extended = SDValue(DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
2756 DAG.getTargetConstant(
2757 SPU::GPRCRegClass.getID(),
2759 // Shuffle bytes - Copy the sign bits into the upper 64 bits
2760 // and the input value into the lower 64 bits.
2761 SDValue extShuffle = DAG.getNode(SPUISD::SHUFB, dl, mvt,
2762 extended, sraVal, shufMask);
2763 return DAG.getNode(ISD::BITCAST, dl, MVT::i128, extShuffle);
2766 //! Custom (target-specific) lowering entry point
2768 This is where LLVM's DAG selection process calls to do target-specific
2772 SPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
2774 unsigned Opc = (unsigned) Op.getOpcode();
2775 EVT VT = Op.getValueType();
2780 errs() << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
2781 errs() << "Op.getOpcode() = " << Opc << "\n";
2782 errs() << "*Op.getNode():\n";
2783 Op.getNode()->dump();
2785 llvm_unreachable(0);
2791 return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
2793 return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
2794 case ISD::ConstantPool:
2795 return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
2796 case ISD::GlobalAddress:
2797 return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
2798 case ISD::JumpTable:
2799 return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
2800 case ISD::ConstantFP:
2801 return LowerConstantFP(Op, DAG);
2803 // i8, i64 math ops:
2812 return LowerI8Math(Op, DAG, Opc, *this);
2816 case ISD::FP_TO_SINT:
2817 case ISD::FP_TO_UINT:
2818 return LowerFP_TO_INT(Op, DAG, *this);
2820 case ISD::SINT_TO_FP:
2821 case ISD::UINT_TO_FP:
2822 return LowerINT_TO_FP(Op, DAG, *this);
2824 // Vector-related lowering.
2825 case ISD::BUILD_VECTOR:
2826 return LowerBUILD_VECTOR(Op, DAG);
2827 case ISD::SCALAR_TO_VECTOR:
2828 return LowerSCALAR_TO_VECTOR(Op, DAG);
2829 case ISD::VECTOR_SHUFFLE:
2830 return LowerVECTOR_SHUFFLE(Op, DAG);
2831 case ISD::EXTRACT_VECTOR_ELT:
2832 return LowerEXTRACT_VECTOR_ELT(Op, DAG);
2833 case ISD::INSERT_VECTOR_ELT:
2834 return LowerINSERT_VECTOR_ELT(Op, DAG);
2836 // Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
2840 return LowerByteImmed(Op, DAG);
2842 // Vector and i8 multiply:
2845 return LowerI8Math(Op, DAG, Opc, *this);
2848 return LowerCTPOP(Op, DAG);
2850 case ISD::SELECT_CC:
2851 return LowerSELECT_CC(Op, DAG, *this);
2854 return LowerSETCC(Op, DAG, *this);
2857 return LowerTRUNCATE(Op, DAG);
2859 case ISD::SIGN_EXTEND:
2860 return LowerSIGN_EXTEND(Op, DAG);
2866 void SPUTargetLowering::ReplaceNodeResults(SDNode *N,
2867 SmallVectorImpl<SDValue>&Results,
2868 SelectionDAG &DAG) const
2871 unsigned Opc = (unsigned) N->getOpcode();
2872 EVT OpVT = N->getValueType(0);
2876 errs() << "SPUTargetLowering::ReplaceNodeResults(): need to fix this!\n";
2877 errs() << "Op.getOpcode() = " << Opc << "\n";
2878 errs() << "*Op.getNode():\n";
2886 /* Otherwise, return unchanged */
2889 //===----------------------------------------------------------------------===//
2890 // Target Optimization Hooks
2891 //===----------------------------------------------------------------------===//
2894 SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
2897 TargetMachine &TM = getTargetMachine();
2899 const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
2900 SelectionDAG &DAG = DCI.DAG;
2901 SDValue Op0 = N->getOperand(0); // everything has at least one operand
2902 EVT NodeVT = N->getValueType(0); // The node's value type
2903 EVT Op0VT = Op0.getValueType(); // The first operand's result
2904 SDValue Result; // Initially, empty result
2905 DebugLoc dl = N->getDebugLoc();
2907 switch (N->getOpcode()) {
2910 SDValue Op1 = N->getOperand(1);
2912 if (Op0.getOpcode() == SPUISD::IndirectAddr
2913 || Op1.getOpcode() == SPUISD::IndirectAddr) {
2914 // Normalize the operands to reduce repeated code
2915 SDValue IndirectArg = Op0, AddArg = Op1;
2917 if (Op1.getOpcode() == SPUISD::IndirectAddr) {
2922 if (isa<ConstantSDNode>(AddArg)) {
2923 ConstantSDNode *CN0 = cast<ConstantSDNode > (AddArg);
2924 SDValue IndOp1 = IndirectArg.getOperand(1);
2926 if (CN0->isNullValue()) {
2927 // (add (SPUindirect <arg>, <arg>), 0) ->
2928 // (SPUindirect <arg>, <arg>)
2930 #if !defined(NDEBUG)
2931 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2933 << "Replace: (add (SPUindirect <arg>, <arg>), 0)\n"
2934 << "With: (SPUindirect <arg>, <arg>)\n";
2939 } else if (isa<ConstantSDNode>(IndOp1)) {
2940 // (add (SPUindirect <arg>, <const>), <const>) ->
2941 // (SPUindirect <arg>, <const + const>)
2942 ConstantSDNode *CN1 = cast<ConstantSDNode > (IndOp1);
2943 int64_t combinedConst = CN0->getSExtValue() + CN1->getSExtValue();
2944 SDValue combinedValue = DAG.getConstant(combinedConst, Op0VT);
2946 #if !defined(NDEBUG)
2947 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2949 << "Replace: (add (SPUindirect <arg>, " << CN1->getSExtValue()
2950 << "), " << CN0->getSExtValue() << ")\n"
2951 << "With: (SPUindirect <arg>, "
2952 << combinedConst << ")\n";
2956 return DAG.getNode(SPUISD::IndirectAddr, dl, Op0VT,
2957 IndirectArg, combinedValue);
2963 case ISD::SIGN_EXTEND:
2964 case ISD::ZERO_EXTEND:
2965 case ISD::ANY_EXTEND: {
2966 if (Op0.getOpcode() == SPUISD::VEC2PREFSLOT && NodeVT == Op0VT) {
2967 // (any_extend (SPUextract_elt0 <arg>)) ->
2968 // (SPUextract_elt0 <arg>)
2969 // Types must match, however...
2970 #if !defined(NDEBUG)
2971 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
2972 errs() << "\nReplace: ";
2974 errs() << "\nWith: ";
2975 Op0.getNode()->dump(&DAG);
2984 case SPUISD::IndirectAddr: {
2985 if (!ST->usingLargeMem() && Op0.getOpcode() == SPUISD::AFormAddr) {
2986 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N->getOperand(1));
2987 if (CN != 0 && CN->isNullValue()) {
2988 // (SPUindirect (SPUaform <addr>, 0), 0) ->
2989 // (SPUaform <addr>, 0)
2991 DEBUG(errs() << "Replace: ");
2992 DEBUG(N->dump(&DAG));
2993 DEBUG(errs() << "\nWith: ");
2994 DEBUG(Op0.getNode()->dump(&DAG));
2995 DEBUG(errs() << "\n");
2999 } else if (Op0.getOpcode() == ISD::ADD) {
3000 SDValue Op1 = N->getOperand(1);
3001 if (ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(Op1)) {
3002 // (SPUindirect (add <arg>, <arg>), 0) ->
3003 // (SPUindirect <arg>, <arg>)
3004 if (CN1->isNullValue()) {
3006 #if !defined(NDEBUG)
3007 if (DebugFlag && isCurrentDebugType(DEBUG_TYPE)) {
3009 << "Replace: (SPUindirect (add <arg>, <arg>), 0)\n"
3010 << "With: (SPUindirect <arg>, <arg>)\n";
3014 return DAG.getNode(SPUISD::IndirectAddr, dl, Op0VT,
3015 Op0.getOperand(0), Op0.getOperand(1));
3021 case SPUISD::SHL_BITS:
3022 case SPUISD::SHL_BYTES:
3023 case SPUISD::ROTBYTES_LEFT: {
3024 SDValue Op1 = N->getOperand(1);
3026 // Kill degenerate vector shifts:
3027 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op1)) {
3028 if (CN->isNullValue()) {
3034 case SPUISD::PREFSLOT2VEC: {
3035 switch (Op0.getOpcode()) {
3038 case ISD::ANY_EXTEND:
3039 case ISD::ZERO_EXTEND:
3040 case ISD::SIGN_EXTEND: {
3041 // (SPUprefslot2vec (any|zero|sign_extend (SPUvec2prefslot <arg>))) ->
3043 // but only if the SPUprefslot2vec and <arg> types match.
3044 SDValue Op00 = Op0.getOperand(0);
3045 if (Op00.getOpcode() == SPUISD::VEC2PREFSLOT) {
3046 SDValue Op000 = Op00.getOperand(0);
3047 if (Op000.getValueType() == NodeVT) {
3053 case SPUISD::VEC2PREFSLOT: {
3054 // (SPUprefslot2vec (SPUvec2prefslot <arg>)) ->
3056 Result = Op0.getOperand(0);
3064 // Otherwise, return unchanged.
3066 if (Result.getNode()) {
3067 DEBUG(errs() << "\nReplace.SPU: ");
3068 DEBUG(N->dump(&DAG));
3069 DEBUG(errs() << "\nWith: ");
3070 DEBUG(Result.getNode()->dump(&DAG));
3071 DEBUG(errs() << "\n");
3078 //===----------------------------------------------------------------------===//
3079 // Inline Assembly Support
3080 //===----------------------------------------------------------------------===//
3082 /// getConstraintType - Given a constraint letter, return the type of
3083 /// constraint it is for this target.
3084 SPUTargetLowering::ConstraintType
3085 SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
3086 if (ConstraintLetter.size() == 1) {
3087 switch (ConstraintLetter[0]) {
3094 return C_RegisterClass;
3097 return TargetLowering::getConstraintType(ConstraintLetter);
3100 /// Examine constraint type and operand type and determine a weight value.
3101 /// This object must already have been set up with the operand type
3102 /// and the current alternative constraint selected.
3103 TargetLowering::ConstraintWeight
3104 SPUTargetLowering::getSingleConstraintMatchWeight(
3105 AsmOperandInfo &info, const char *constraint) const {
3106 ConstraintWeight weight = CW_Invalid;
3107 Value *CallOperandVal = info.CallOperandVal;
3108 // If we don't have a value, we can't do a match,
3109 // but allow it at the lowest weight.
3110 if (CallOperandVal == NULL)
3112 // Look at the constraint type.
3113 switch (*constraint) {
3115 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3117 //FIXME: Seems like the supported constraint letters were just copied
3118 // from PPC, as the following doesn't correspond to the GCC docs.
3119 // I'm leaving it so until someone adds the corresponding lowering support.
3126 weight = CW_Register;
3132 std::pair<unsigned, const TargetRegisterClass*>
3133 SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3136 if (Constraint.size() == 1) {
3137 // GCC RS6000 Constraint Letters
3138 switch (Constraint[0]) {
3142 return std::make_pair(0U, &SPU::R64CRegClass);
3143 return std::make_pair(0U, &SPU::R32CRegClass);
3146 return std::make_pair(0U, &SPU::R32FPRegClass);
3148 return std::make_pair(0U, &SPU::R64FPRegClass);
3151 return std::make_pair(0U, &SPU::GPRCRegClass);
3155 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3158 //! Compute used/known bits for a SPU operand
3160 SPUTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
3163 const SelectionDAG &DAG,
3164 unsigned Depth ) const {
3166 const uint64_t uint64_sizebits = sizeof(uint64_t) * CHAR_BIT;
3168 switch (Op.getOpcode()) {
3170 // KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
3176 case SPUISD::PREFSLOT2VEC:
3177 case SPUISD::LDRESULT:
3178 case SPUISD::VEC2PREFSLOT:
3179 case SPUISD::SHLQUAD_L_BITS:
3180 case SPUISD::SHLQUAD_L_BYTES:
3181 case SPUISD::VEC_ROTL:
3182 case SPUISD::VEC_ROTR:
3183 case SPUISD::ROTBYTES_LEFT:
3184 case SPUISD::SELECT_MASK:
3191 SPUTargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
3192 unsigned Depth) const {
3193 switch (Op.getOpcode()) {
3198 EVT VT = Op.getValueType();
3200 if (VT != MVT::i8 && VT != MVT::i16 && VT != MVT::i32) {
3203 return VT.getSizeInBits();
3208 // LowerAsmOperandForConstraint
3210 SPUTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3211 std::string &Constraint,
3212 std::vector<SDValue> &Ops,
3213 SelectionDAG &DAG) const {
3214 // Default, for the time being, to the base class handler
3215 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3218 /// isLegalAddressImmediate - Return true if the integer value can be used
3219 /// as the offset of the target addressing mode.
3220 bool SPUTargetLowering::isLegalAddressImmediate(int64_t V,
3222 // SPU's addresses are 256K:
3223 return (V > -(1 << 18) && V < (1 << 18) - 1);
3226 bool SPUTargetLowering::isLegalAddressImmediate(GlobalValue* GV) const {
3231 SPUTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3232 // The SPU target isn't yet aware of offsets.
3236 // can we compare to Imm without writing it into a register?
3237 bool SPUTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
3238 //ceqi, cgti, etc. all take s10 operand
3239 return isInt<10>(Imm);
3243 SPUTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3246 // A-form: 18bit absolute address.
3247 if (AM.BaseGV && !AM.HasBaseReg && AM.Scale == 0 && AM.BaseOffs == 0)
3250 // D-form: reg + 14bit offset
3251 if (AM.BaseGV ==0 && AM.HasBaseReg && AM.Scale == 0 && isInt<14>(AM.BaseOffs))
3255 if (AM.BaseGV == 0 && AM.HasBaseReg && AM.Scale == 1 && AM.BaseOffs ==0)