1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Chris Lattner and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the PPCISelLowering class.
12 //===----------------------------------------------------------------------===//
14 #include "PPCISelLowering.h"
15 #include "PPCTargetMachine.h"
16 #include "PPCPerfectShuffle.h"
17 #include "llvm/ADT/VectorExtras.h"
18 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
19 #include "llvm/CodeGen/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/CodeGen/SSARegMap.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Function.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/Support/MathExtras.h"
28 #include "llvm/Target/TargetOptions.h"
31 PPCTargetLowering::PPCTargetLowering(TargetMachine &TM)
32 : TargetLowering(TM) {
34 // Fold away setcc operations if possible.
35 setSetCCIsExpensive();
38 // Use _setjmp/_longjmp instead of setjmp/longjmp.
39 setUseUnderscoreSetJmpLongJmp(true);
41 // Set up the register classes.
42 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
43 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
44 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
46 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
47 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
49 // PowerPC has no intrinsics for these particular operations
50 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
51 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
52 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
54 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
55 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
56 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
58 // PowerPC has no SREM/UREM instructions
59 setOperationAction(ISD::SREM, MVT::i32, Expand);
60 setOperationAction(ISD::UREM, MVT::i32, Expand);
62 // We don't support sin/cos/sqrt/fmod
63 setOperationAction(ISD::FSIN , MVT::f64, Expand);
64 setOperationAction(ISD::FCOS , MVT::f64, Expand);
65 setOperationAction(ISD::FREM , MVT::f64, Expand);
66 setOperationAction(ISD::FSIN , MVT::f32, Expand);
67 setOperationAction(ISD::FCOS , MVT::f32, Expand);
68 setOperationAction(ISD::FREM , MVT::f32, Expand);
70 // If we're enabling GP optimizations, use hardware square root
71 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
72 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
73 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
76 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
77 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
79 // PowerPC does not have BSWAP, CTPOP or CTTZ
80 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
81 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
82 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
84 // PowerPC does not have ROTR
85 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
87 // PowerPC does not have Select
88 setOperationAction(ISD::SELECT, MVT::i32, Expand);
89 setOperationAction(ISD::SELECT, MVT::f32, Expand);
90 setOperationAction(ISD::SELECT, MVT::f64, Expand);
92 // PowerPC wants to turn select_cc of FP into fsel when possible.
93 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
94 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
96 // PowerPC wants to optimize integer setcc a bit
97 setOperationAction(ISD::SETCC, MVT::i32, Custom);
99 // PowerPC does not have BRCOND which requires SetCC
100 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
102 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
103 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
105 // PowerPC does not have [U|S]INT_TO_FP
106 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
107 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
109 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
110 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
112 // PowerPC does not have truncstore for i1.
113 setOperationAction(ISD::TRUNCSTORE, MVT::i1, Promote);
115 // Support label based line numbers.
116 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
117 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
118 // FIXME - use subtarget debug flags
119 if (!TM.getSubtarget<PPCSubtarget>().isDarwin())
120 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
122 // We want to legalize GlobalAddress and ConstantPool nodes into the
123 // appropriate instructions to materialize the address.
124 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
125 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
127 // RET must be custom lowered, to meet ABI requirements
128 setOperationAction(ISD::RET , MVT::Other, Custom);
130 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
131 setOperationAction(ISD::VASTART , MVT::Other, Custom);
133 // Use the default implementation.
134 setOperationAction(ISD::VAARG , MVT::Other, Expand);
135 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
136 setOperationAction(ISD::VAEND , MVT::Other, Expand);
137 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
138 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
139 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
141 // We want to custom lower some of our intrinsics.
142 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
144 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
145 // They also have instructions for converting between i64 and fp.
146 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
147 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
149 // FIXME: disable this lowered code. This generates 64-bit register values,
150 // and we don't model the fact that the top part is clobbered by calls. We
151 // need to flag these together so that the value isn't live across a call.
152 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
154 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
155 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
157 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
158 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
161 if (TM.getSubtarget<PPCSubtarget>().has64BitRegs()) {
162 // 64 bit PowerPC implementations can support i64 types directly
163 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
164 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
165 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
167 // 32 bit PowerPC wants to expand i64 shifts itself.
168 setOperationAction(ISD::SHL, MVT::i64, Custom);
169 setOperationAction(ISD::SRL, MVT::i64, Custom);
170 setOperationAction(ISD::SRA, MVT::i64, Custom);
173 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
174 // First set operation action for all vector types to expand. Then we
175 // will selectively turn on ones that can be effectively codegen'd.
176 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
177 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
178 // add/sub are legal for all supported vector VT's.
179 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
180 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
182 // We promote all shuffles to v16i8.
183 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
184 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
186 // We promote all non-typed operations to v4i32.
187 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
188 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
189 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
190 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
191 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
192 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
193 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
194 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
195 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
196 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
197 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
198 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
200 // No other operations are legal.
201 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
202 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
203 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
204 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
205 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
206 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
207 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
208 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
210 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
213 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
214 // with merges, splats, etc.
215 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
217 setOperationAction(ISD::AND , MVT::v4i32, Legal);
218 setOperationAction(ISD::OR , MVT::v4i32, Legal);
219 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
220 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
221 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
222 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
224 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
225 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
226 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
227 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
229 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
230 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
232 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
233 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
235 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
236 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
237 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
238 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
241 setSetCCResultContents(ZeroOrOneSetCCResult);
242 setStackPointerRegisterToSaveRestore(PPC::R1);
244 // We have target-specific dag combine patterns for the following nodes:
245 setTargetDAGCombine(ISD::SINT_TO_FP);
246 setTargetDAGCombine(ISD::STORE);
248 computeRegisterProperties();
251 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
254 case PPCISD::FSEL: return "PPCISD::FSEL";
255 case PPCISD::FCFID: return "PPCISD::FCFID";
256 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
257 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
258 case PPCISD::STFIWX: return "PPCISD::STFIWX";
259 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
260 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
261 case PPCISD::VPERM: return "PPCISD::VPERM";
262 case PPCISD::Hi: return "PPCISD::Hi";
263 case PPCISD::Lo: return "PPCISD::Lo";
264 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
265 case PPCISD::SRL: return "PPCISD::SRL";
266 case PPCISD::SRA: return "PPCISD::SRA";
267 case PPCISD::SHL: return "PPCISD::SHL";
268 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
269 case PPCISD::STD_32: return "PPCISD::STD_32";
270 case PPCISD::CALL: return "PPCISD::CALL";
271 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
272 case PPCISD::MFCR: return "PPCISD::MFCR";
273 case PPCISD::VCMP: return "PPCISD::VCMP";
274 case PPCISD::VCMPo: return "PPCISD::VCMPo";
278 //===----------------------------------------------------------------------===//
279 // Node matching predicates, for use by the tblgen matching code.
280 //===----------------------------------------------------------------------===//
282 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
283 static bool isFloatingPointZero(SDOperand Op) {
284 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
285 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
286 else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
287 // Maybe this has already been legalized into the constant pool?
288 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
289 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
290 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
295 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
296 /// true if Op is undef or if it matches the specified value.
297 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
298 return Op.getOpcode() == ISD::UNDEF ||
299 cast<ConstantSDNode>(Op)->getValue() == Val;
302 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
303 /// VPKUHUM instruction.
304 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
306 for (unsigned i = 0; i != 16; ++i)
307 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
310 for (unsigned i = 0; i != 8; ++i)
311 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
312 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
318 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
319 /// VPKUWUM instruction.
320 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
322 for (unsigned i = 0; i != 16; i += 2)
323 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
324 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
327 for (unsigned i = 0; i != 8; i += 2)
328 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
329 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
330 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
331 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
337 /// isVMerge - Common function, used to match vmrg* shuffles.
339 static bool isVMerge(SDNode *N, unsigned UnitSize,
340 unsigned LHSStart, unsigned RHSStart) {
341 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
342 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
343 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
344 "Unsupported merge size!");
346 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
347 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
348 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
349 LHSStart+j+i*UnitSize) ||
350 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
351 RHSStart+j+i*UnitSize))
357 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
358 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
359 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
361 return isVMerge(N, UnitSize, 8, 24);
362 return isVMerge(N, UnitSize, 8, 8);
365 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
366 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
367 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
369 return isVMerge(N, UnitSize, 0, 16);
370 return isVMerge(N, UnitSize, 0, 0);
374 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
375 /// amount, otherwise return -1.
376 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
377 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
378 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
379 // Find the first non-undef value in the shuffle mask.
381 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
384 if (i == 16) return -1; // all undef.
386 // Otherwise, check to see if the rest of the elements are consequtively
387 // numbered from this value.
388 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
389 if (ShiftAmt < i) return -1;
393 // Check the rest of the elements to see if they are consequtive.
394 for (++i; i != 16; ++i)
395 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
398 // Check the rest of the elements to see if they are consequtive.
399 for (++i; i != 16; ++i)
400 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
407 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
408 /// specifies a splat of a single element that is suitable for input to
409 /// VSPLTB/VSPLTH/VSPLTW.
410 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
411 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
412 N->getNumOperands() == 16 &&
413 (EltSize == 1 || EltSize == 2 || EltSize == 4));
415 // This is a splat operation if each element of the permute is the same, and
416 // if the value doesn't reference the second vector.
417 unsigned ElementBase = 0;
418 SDOperand Elt = N->getOperand(0);
419 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
420 ElementBase = EltV->getValue();
422 return false; // FIXME: Handle UNDEF elements too!
424 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
427 // Check that they are consequtive.
428 for (unsigned i = 1; i != EltSize; ++i) {
429 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
430 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
434 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
435 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
436 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
437 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
438 "Invalid VECTOR_SHUFFLE mask!");
439 for (unsigned j = 0; j != EltSize; ++j)
440 if (N->getOperand(i+j) != N->getOperand(j))
447 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
448 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
449 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
450 assert(isSplatShuffleMask(N, EltSize));
451 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
454 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
455 /// by using a vspltis[bhw] instruction of the specified element size, return
456 /// the constant being splatted. The ByteSize field indicates the number of
457 /// bytes of each element [124] -> [bhw].
458 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
459 SDOperand OpVal(0, 0);
461 // If ByteSize of the splat is bigger than the element size of the
462 // build_vector, then we have a case where we are checking for a splat where
463 // multiple elements of the buildvector are folded together into a single
464 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
465 unsigned EltSize = 16/N->getNumOperands();
466 if (EltSize < ByteSize) {
467 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
468 SDOperand UniquedVals[4];
469 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
471 // See if all of the elements in the buildvector agree across.
472 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
473 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
474 // If the element isn't a constant, bail fully out.
475 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
478 if (UniquedVals[i&(Multiple-1)].Val == 0)
479 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
480 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
481 return SDOperand(); // no match.
484 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
485 // either constant or undef values that are identical for each chunk. See
486 // if these chunks can form into a larger vspltis*.
488 // Check to see if all of the leading entries are either 0 or -1. If
489 // neither, then this won't fit into the immediate field.
490 bool LeadingZero = true;
491 bool LeadingOnes = true;
492 for (unsigned i = 0; i != Multiple-1; ++i) {
493 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
495 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
496 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
498 // Finally, check the least significant entry.
500 if (UniquedVals[Multiple-1].Val == 0)
501 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
502 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
504 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
507 if (UniquedVals[Multiple-1].Val == 0)
508 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
509 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
510 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
511 return DAG.getTargetConstant(Val, MVT::i32);
517 // Check to see if this buildvec has a single non-undef value in its elements.
518 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
519 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
521 OpVal = N->getOperand(i);
522 else if (OpVal != N->getOperand(i))
526 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
528 unsigned ValSizeInBytes = 0;
530 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
531 Value = CN->getValue();
532 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
533 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
534 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
535 Value = FloatToBits(CN->getValue());
539 // If the splat value is larger than the element value, then we can never do
540 // this splat. The only case that we could fit the replicated bits into our
541 // immediate field for would be zero, and we prefer to use vxor for it.
542 if (ValSizeInBytes < ByteSize) return SDOperand();
544 // If the element value is larger than the splat value, cut it in half and
545 // check to see if the two halves are equal. Continue doing this until we
546 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
547 while (ValSizeInBytes > ByteSize) {
548 ValSizeInBytes >>= 1;
550 // If the top half equals the bottom half, we're still ok.
551 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
552 (Value & ((1 << (8*ValSizeInBytes))-1)))
556 // Properly sign extend the value.
557 int ShAmt = (4-ByteSize)*8;
558 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
560 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
561 if (MaskVal == 0) return SDOperand();
563 // Finally, if this value fits in a 5 bit sext field, return it
564 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
565 return DAG.getTargetConstant(MaskVal, MVT::i32);
569 //===----------------------------------------------------------------------===//
570 // LowerOperation implementation
571 //===----------------------------------------------------------------------===//
573 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
574 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
575 Constant *C = CP->get();
576 SDOperand CPI = DAG.getTargetConstantPool(C, MVT::i32, CP->getAlignment());
577 SDOperand Zero = DAG.getConstant(0, MVT::i32);
579 const TargetMachine &TM = DAG.getTarget();
581 // If this is a non-darwin platform, we don't support non-static relo models
583 if (TM.getRelocationModel() == Reloc::Static ||
584 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
585 // Generate non-pic code that has direct accesses to the constant pool.
586 // The address of the global is just (hi(&g)+lo(&g)).
587 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
588 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
589 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
592 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
593 if (TM.getRelocationModel() == Reloc::PIC) {
594 // With PIC, the first instruction is actually "GR+hi(&G)".
595 Hi = DAG.getNode(ISD::ADD, MVT::i32,
596 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
599 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
600 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
604 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
605 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
606 GlobalValue *GV = GSDN->getGlobal();
607 SDOperand GA = DAG.getTargetGlobalAddress(GV, MVT::i32, GSDN->getOffset());
608 SDOperand Zero = DAG.getConstant(0, MVT::i32);
610 const TargetMachine &TM = DAG.getTarget();
612 // If this is a non-darwin platform, we don't support non-static relo models
614 if (TM.getRelocationModel() == Reloc::Static ||
615 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
616 // Generate non-pic code that has direct accesses to globals.
617 // The address of the global is just (hi(&g)+lo(&g)).
618 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
619 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
620 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
623 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
624 if (TM.getRelocationModel() == Reloc::PIC) {
625 // With PIC, the first instruction is actually "GR+hi(&G)".
626 Hi = DAG.getNode(ISD::ADD, MVT::i32,
627 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
630 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
631 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
633 if (!GV->hasWeakLinkage() && !GV->hasLinkOnceLinkage() &&
634 (!GV->isExternal() || GV->hasNotBeenReadFromBytecode()))
637 // If the global is weak or external, we have to go through the lazy
639 return DAG.getLoad(MVT::i32, DAG.getEntryNode(), Lo, DAG.getSrcValue(0));
642 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
643 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
645 // If we're comparing for equality to zero, expose the fact that this is
646 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
647 // fold the new nodes.
648 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
649 if (C->isNullValue() && CC == ISD::SETEQ) {
650 MVT::ValueType VT = Op.getOperand(0).getValueType();
651 SDOperand Zext = Op.getOperand(0);
654 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
656 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
657 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
658 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
659 DAG.getConstant(Log2b, MVT::i32));
660 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
662 // Leave comparisons against 0 and -1 alone for now, since they're usually
663 // optimized. FIXME: revisit this when we can custom lower all setcc
665 if (C->isAllOnesValue() || C->isNullValue())
669 // If we have an integer seteq/setne, turn it into a compare against zero
670 // by subtracting the rhs from the lhs, which is faster than setting a
671 // condition register, reading it back out, and masking the correct bit.
672 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
673 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
674 MVT::ValueType VT = Op.getValueType();
675 SDOperand Sub = DAG.getNode(ISD::SUB, LHSVT, Op.getOperand(0),
677 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
682 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
683 unsigned VarArgsFrameIndex) {
684 // vastart just stores the address of the VarArgsFrameIndex slot into the
685 // memory location argument.
686 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
687 return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
688 Op.getOperand(1), Op.getOperand(2));
691 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG) {
693 switch(Op.getNumOperands()) {
695 assert(0 && "Do not know how to return this many arguments!");
698 return SDOperand(); // ret void is legal
700 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
702 if (MVT::isVector(ArgVT))
704 else if (MVT::isInteger(ArgVT))
707 assert(MVT::isFloatingPoint(ArgVT));
711 Copy = DAG.getCopyToReg(Op.getOperand(0), ArgReg, Op.getOperand(1),
714 // If we haven't noted the R3/F1 are live out, do so now.
715 if (DAG.getMachineFunction().liveout_empty())
716 DAG.getMachineFunction().addLiveOut(ArgReg);
720 Copy = DAG.getCopyToReg(Op.getOperand(0), PPC::R3, Op.getOperand(2),
722 Copy = DAG.getCopyToReg(Copy, PPC::R4, Op.getOperand(1),Copy.getValue(1));
723 // If we haven't noted the R3+R4 are live out, do so now.
724 if (DAG.getMachineFunction().liveout_empty()) {
725 DAG.getMachineFunction().addLiveOut(PPC::R3);
726 DAG.getMachineFunction().addLiveOut(PPC::R4);
730 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
733 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
735 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
736 // Not FP? Not a fsel.
737 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
738 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
741 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
743 // Cannot handle SETEQ/SETNE.
744 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
746 MVT::ValueType ResVT = Op.getValueType();
747 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
748 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
749 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
751 // If the RHS of the comparison is a 0.0, we don't need to do the
752 // subtraction at all.
753 if (isFloatingPointZero(RHS))
755 default: break; // SETUO etc aren't handled by fsel.
758 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
761 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
762 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
763 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
766 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
769 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
770 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
771 return DAG.getNode(PPCISD::FSEL, ResVT,
772 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
777 default: break; // SETUO etc aren't handled by fsel.
780 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
781 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
782 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
783 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
786 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
787 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
788 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
789 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
792 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
793 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
794 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
795 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
798 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
799 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
800 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
801 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
806 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
807 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
808 SDOperand Src = Op.getOperand(0);
809 if (Src.getValueType() == MVT::f32)
810 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
813 switch (Op.getValueType()) {
814 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
816 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
819 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
823 // Convert the FP value to an int value through memory.
824 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
825 if (Op.getValueType() == MVT::i32)
826 Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
830 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
831 if (Op.getOperand(0).getValueType() == MVT::i64) {
832 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
833 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
834 if (Op.getValueType() == MVT::f32)
835 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
839 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
840 "Unhandled SINT_TO_FP type in custom expander!");
841 // Since we only generate this in 64-bit mode, we can take advantage of
842 // 64-bit registers. In particular, sign extend the input value into the
843 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
844 // then lfd it and fcfid it.
845 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
846 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
847 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
849 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
852 // STD the extended value into the stack slot.
853 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
854 DAG.getEntryNode(), Ext64, FIdx,
855 DAG.getSrcValue(NULL));
856 // Load the value as a double.
857 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
859 // FCFID it and return it.
860 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
861 if (Op.getValueType() == MVT::f32)
862 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
866 static SDOperand LowerSHL(SDOperand Op, SelectionDAG &DAG) {
867 assert(Op.getValueType() == MVT::i64 &&
868 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
869 // The generic code does a fine job expanding shift by a constant.
870 if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
872 // Otherwise, expand into a bunch of logical ops. Note that these ops
873 // depend on the PPC behavior for oversized shift amounts.
874 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
875 DAG.getConstant(0, MVT::i32));
876 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
877 DAG.getConstant(1, MVT::i32));
878 SDOperand Amt = Op.getOperand(1);
880 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
881 DAG.getConstant(32, MVT::i32), Amt);
882 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
883 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
884 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
885 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
886 DAG.getConstant(-32U, MVT::i32));
887 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
888 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
889 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
890 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
893 static SDOperand LowerSRL(SDOperand Op, SelectionDAG &DAG) {
894 assert(Op.getValueType() == MVT::i64 &&
895 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
896 // The generic code does a fine job expanding shift by a constant.
897 if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
899 // Otherwise, expand into a bunch of logical ops. Note that these ops
900 // depend on the PPC behavior for oversized shift amounts.
901 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
902 DAG.getConstant(0, MVT::i32));
903 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
904 DAG.getConstant(1, MVT::i32));
905 SDOperand Amt = Op.getOperand(1);
907 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
908 DAG.getConstant(32, MVT::i32), Amt);
909 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
910 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
911 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
912 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
913 DAG.getConstant(-32U, MVT::i32));
914 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
915 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
916 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
917 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
920 static SDOperand LowerSRA(SDOperand Op, SelectionDAG &DAG) {
921 assert(Op.getValueType() == MVT::i64 &&
922 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
923 // The generic code does a fine job expanding shift by a constant.
924 if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
926 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
927 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
928 DAG.getConstant(0, MVT::i32));
929 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
930 DAG.getConstant(1, MVT::i32));
931 SDOperand Amt = Op.getOperand(1);
933 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
934 DAG.getConstant(32, MVT::i32), Amt);
935 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
936 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
937 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
938 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
939 DAG.getConstant(-32U, MVT::i32));
940 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
941 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
942 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
943 Tmp4, Tmp6, ISD::SETLE);
944 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
947 //===----------------------------------------------------------------------===//
948 // Vector related lowering.
951 // If this is a vector of constants or undefs, get the bits. A bit in
952 // UndefBits is set if the corresponding element of the vector is an
953 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
954 // zero. Return true if this is not an array of constants, false if it is.
956 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
957 uint64_t UndefBits[2]) {
958 // Start with zero'd results.
959 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
961 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
962 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
963 SDOperand OpVal = BV->getOperand(i);
965 unsigned PartNo = i >= e/2; // In the upper 128 bits?
966 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
968 uint64_t EltBits = 0;
969 if (OpVal.getOpcode() == ISD::UNDEF) {
970 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
971 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
973 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
974 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
975 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
976 assert(CN->getValueType(0) == MVT::f32 &&
977 "Only one legal FP vector type!");
978 EltBits = FloatToBits(CN->getValue());
980 // Nonconstant element.
984 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
987 //printf("%llx %llx %llx %llx\n",
988 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
992 // If this is a splat (repetition) of a value across the whole vector, return
993 // the smallest size that splats it. For example, "0x01010101010101..." is a
994 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
995 // SplatSize = 1 byte.
996 static bool isConstantSplat(const uint64_t Bits128[2],
997 const uint64_t Undef128[2],
998 unsigned &SplatBits, unsigned &SplatUndef,
999 unsigned &SplatSize) {
1001 // Don't let undefs prevent splats from matching. See if the top 64-bits are
1002 // the same as the lower 64-bits, ignoring undefs.
1003 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
1004 return false; // Can't be a splat if two pieces don't match.
1006 uint64_t Bits64 = Bits128[0] | Bits128[1];
1007 uint64_t Undef64 = Undef128[0] & Undef128[1];
1009 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
1011 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
1012 return false; // Can't be a splat if two pieces don't match.
1014 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
1015 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
1017 // If the top 16-bits are different than the lower 16-bits, ignoring
1018 // undefs, we have an i32 splat.
1019 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
1021 SplatUndef = Undef32;
1026 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
1027 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
1029 // If the top 8-bits are different than the lower 8-bits, ignoring
1030 // undefs, we have an i16 splat.
1031 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
1033 SplatUndef = Undef16;
1038 // Otherwise, we have an 8-bit splat.
1039 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
1040 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
1045 /// BuildSplatI - Build a canonical splati of Val with an element size of
1046 /// SplatSize. Cast the result to VT.
1047 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
1048 SelectionDAG &DAG) {
1049 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
1051 // Force vspltis[hw] -1 to vspltisb -1.
1052 if (Val == -1) SplatSize = 1;
1054 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
1055 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
1057 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
1059 // Build a canonical splat for this value.
1060 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorBaseType(CanonicalVT));
1061 std::vector<SDOperand> Ops(MVT::getVectorNumElements(CanonicalVT), Elt);
1062 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT, Ops);
1063 return DAG.getNode(ISD::BIT_CONVERT, VT, Res);
1066 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
1067 /// specified intrinsic ID.
1068 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
1070 MVT::ValueType DestVT = MVT::Other) {
1071 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
1072 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
1073 DAG.getConstant(IID, MVT::i32), LHS, RHS);
1076 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
1077 /// specified intrinsic ID.
1078 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
1079 SDOperand Op2, SelectionDAG &DAG,
1080 MVT::ValueType DestVT = MVT::Other) {
1081 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
1082 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
1083 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
1087 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
1088 /// amount. The result has the specified value type.
1089 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
1090 MVT::ValueType VT, SelectionDAG &DAG) {
1091 // Force LHS/RHS to be the right type.
1092 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
1093 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
1095 std::vector<SDOperand> Ops;
1096 for (unsigned i = 0; i != 16; ++i)
1097 Ops.push_back(DAG.getConstant(i+Amt, MVT::i32));
1098 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
1099 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops));
1100 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
1103 // If this is a case we can't handle, return null and let the default
1104 // expansion code take care of it. If we CAN select this case, and if it
1105 // selects to a single instruction, return Op. Otherwise, if we can codegen
1106 // this case more efficiently than a constant pool load, lower it to the
1107 // sequence of ops that should be used.
1108 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1109 // If this is a vector of constants or undefs, get the bits. A bit in
1110 // UndefBits is set if the corresponding element of the vector is an
1111 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
1113 uint64_t VectorBits[2];
1114 uint64_t UndefBits[2];
1115 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
1116 return SDOperand(); // Not a constant vector.
1118 // If this is a splat (repetition) of a value across the whole vector, return
1119 // the smallest size that splats it. For example, "0x01010101010101..." is a
1120 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
1121 // SplatSize = 1 byte.
1122 unsigned SplatBits, SplatUndef, SplatSize;
1123 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
1124 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
1126 // First, handle single instruction cases.
1129 if (SplatBits == 0) {
1130 // Canonicalize all zero vectors to be v4i32.
1131 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
1132 SDOperand Z = DAG.getConstant(0, MVT::i32);
1133 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
1134 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
1139 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
1140 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
1141 if (SextVal >= -16 && SextVal <= 15)
1142 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
1145 // Two instruction sequences.
1147 // If this value is in the range [-32,30] and is even, use:
1148 // tmp = VSPLTI[bhw], result = add tmp, tmp
1149 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
1150 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
1151 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
1154 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
1155 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
1157 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
1158 // Make -1 and vspltisw -1:
1159 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
1161 // Make the VSLW intrinsic, computing 0x8000_0000.
1162 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
1165 // xor by OnesV to invert it.
1166 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
1167 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
1170 // Check to see if this is a wide variety of vsplti*, binop self cases.
1171 unsigned SplatBitSize = SplatSize*8;
1172 static const char SplatCsts[] = {
1173 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
1174 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
1176 for (unsigned idx = 0; idx < sizeof(SplatCsts)/sizeof(SplatCsts[0]); ++idx){
1177 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
1178 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
1179 int i = SplatCsts[idx];
1181 // Figure out what shift amount will be used by altivec if shifted by i in
1183 unsigned TypeShiftAmt = i & (SplatBitSize-1);
1185 // vsplti + shl self.
1186 if (SextVal == (i << (int)TypeShiftAmt)) {
1187 Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
1188 static const unsigned IIDs[] = { // Intrinsic to use for each size.
1189 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
1190 Intrinsic::ppc_altivec_vslw
1192 return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
1195 // vsplti + srl self.
1196 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
1197 Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
1198 static const unsigned IIDs[] = { // Intrinsic to use for each size.
1199 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
1200 Intrinsic::ppc_altivec_vsrw
1202 return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
1205 // vsplti + sra self.
1206 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
1207 Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
1208 static const unsigned IIDs[] = { // Intrinsic to use for each size.
1209 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
1210 Intrinsic::ppc_altivec_vsraw
1212 return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
1215 // vsplti + rol self.
1216 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
1217 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
1218 Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
1219 static const unsigned IIDs[] = { // Intrinsic to use for each size.
1220 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
1221 Intrinsic::ppc_altivec_vrlw
1223 return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
1226 // t = vsplti c, result = vsldoi t, t, 1
1227 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
1228 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
1229 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
1231 // t = vsplti c, result = vsldoi t, t, 2
1232 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
1233 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
1234 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
1236 // t = vsplti c, result = vsldoi t, t, 3
1237 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
1238 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
1239 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
1243 // Three instruction sequences.
1245 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
1246 if (SextVal >= 0 && SextVal <= 31) {
1247 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, Op.getValueType(),DAG);
1248 SDOperand RHS = BuildSplatI(-16, SplatSize, Op.getValueType(), DAG);
1249 return DAG.getNode(ISD::SUB, Op.getValueType(), LHS, RHS);
1251 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
1252 if (SextVal >= -31 && SextVal <= 0) {
1253 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, Op.getValueType(),DAG);
1254 SDOperand RHS = BuildSplatI(-16, SplatSize, Op.getValueType(), DAG);
1255 return DAG.getNode(ISD::ADD, Op.getValueType(), LHS, RHS);
1262 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
1263 /// the specified operations to build the shuffle.
1264 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
1265 SDOperand RHS, SelectionDAG &DAG) {
1266 unsigned OpNum = (PFEntry >> 26) & 0x0F;
1267 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
1268 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
1271 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
1283 if (OpNum == OP_COPY) {
1284 if (LHSID == (1*9+2)*9+3) return LHS;
1285 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
1289 SDOperand OpLHS, OpRHS;
1290 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
1291 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
1293 unsigned ShufIdxs[16];
1295 default: assert(0 && "Unknown i32 permute!");
1297 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
1298 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
1299 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
1300 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
1303 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
1304 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
1305 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
1306 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
1309 for (unsigned i = 0; i != 16; ++i)
1310 ShufIdxs[i] = (i&3)+0;
1313 for (unsigned i = 0; i != 16; ++i)
1314 ShufIdxs[i] = (i&3)+4;
1317 for (unsigned i = 0; i != 16; ++i)
1318 ShufIdxs[i] = (i&3)+8;
1321 for (unsigned i = 0; i != 16; ++i)
1322 ShufIdxs[i] = (i&3)+12;
1325 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
1327 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
1329 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
1331 std::vector<SDOperand> Ops;
1332 for (unsigned i = 0; i != 16; ++i)
1333 Ops.push_back(DAG.getConstant(ShufIdxs[i], MVT::i32));
1335 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
1336 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops));
1339 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
1340 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
1341 /// return the code it can be lowered into. Worst case, it can always be
1342 /// lowered into a vperm.
1343 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
1344 SDOperand V1 = Op.getOperand(0);
1345 SDOperand V2 = Op.getOperand(1);
1346 SDOperand PermMask = Op.getOperand(2);
1348 // Cases that are handled by instructions that take permute immediates
1349 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
1350 // selected by the instruction selector.
1351 if (V2.getOpcode() == ISD::UNDEF) {
1352 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
1353 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
1354 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
1355 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
1356 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
1357 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
1358 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
1359 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
1360 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
1361 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
1362 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
1363 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
1368 // Altivec has a variety of "shuffle immediates" that take two vector inputs
1369 // and produce a fixed permutation. If any of these match, do not lower to
1371 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
1372 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
1373 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
1374 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
1375 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
1376 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
1377 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
1378 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
1379 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
1382 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
1383 // perfect shuffle table to emit an optimal matching sequence.
1384 unsigned PFIndexes[4];
1385 bool isFourElementShuffle = true;
1386 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
1387 unsigned EltNo = 8; // Start out undef.
1388 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
1389 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
1390 continue; // Undef, ignore it.
1392 unsigned ByteSource =
1393 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
1394 if ((ByteSource & 3) != j) {
1395 isFourElementShuffle = false;
1400 EltNo = ByteSource/4;
1401 } else if (EltNo != ByteSource/4) {
1402 isFourElementShuffle = false;
1406 PFIndexes[i] = EltNo;
1409 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
1410 // perfect shuffle vector to determine if it is cost effective to do this as
1411 // discrete instructions, or whether we should use a vperm.
1412 if (isFourElementShuffle) {
1413 // Compute the index in the perfect shuffle table.
1414 unsigned PFTableIndex =
1415 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
1417 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
1418 unsigned Cost = (PFEntry >> 30);
1420 // Determining when to avoid vperm is tricky. Many things affect the cost
1421 // of vperm, particularly how many times the perm mask needs to be computed.
1422 // For example, if the perm mask can be hoisted out of a loop or is already
1423 // used (perhaps because there are multiple permutes with the same shuffle
1424 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
1425 // the loop requires an extra register.
1427 // As a compromise, we only emit discrete instructions if the shuffle can be
1428 // generated in 3 or fewer operations. When we have loop information
1429 // available, if this block is within a loop, we should avoid using vperm
1430 // for 3-operation perms and use a constant pool load instead.
1432 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
1435 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
1436 // vector that will get spilled to the constant pool.
1437 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1439 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
1440 // that it is in input element units, not in bytes. Convert now.
1441 MVT::ValueType EltVT = MVT::getVectorBaseType(V1.getValueType());
1442 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
1444 std::vector<SDOperand> ResultMask;
1445 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
1447 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
1450 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1452 for (unsigned j = 0; j != BytesPerElement; ++j)
1453 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
1457 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, ResultMask);
1458 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
1461 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
1462 /// lower, do it, otherwise return null.
1463 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
1464 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
1466 // If this is a lowered altivec predicate compare, CompareOpc is set to the
1467 // opcode number of the comparison.
1468 int CompareOpc = -1;
1471 default: return SDOperand(); // Don't custom lower most intrinsics.
1472 // Comparison predicates.
1473 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
1474 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
1475 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
1476 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
1477 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
1478 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
1479 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
1480 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
1481 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
1482 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
1483 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
1484 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
1485 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
1487 // Normal Comparisons.
1488 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
1489 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
1490 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
1491 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
1492 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
1493 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
1494 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
1495 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
1496 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
1497 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
1498 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
1499 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
1500 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
1503 assert(CompareOpc>0 && "We only lower altivec predicate compares so far!");
1505 // If this is a non-dot comparison, make the VCMP node.
1507 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
1508 Op.getOperand(1), Op.getOperand(2),
1509 DAG.getConstant(CompareOpc, MVT::i32));
1510 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
1513 // Create the PPCISD altivec 'dot' comparison node.
1514 std::vector<SDOperand> Ops;
1515 std::vector<MVT::ValueType> VTs;
1516 Ops.push_back(Op.getOperand(2)); // LHS
1517 Ops.push_back(Op.getOperand(3)); // RHS
1518 Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
1519 VTs.push_back(Op.getOperand(2).getValueType());
1520 VTs.push_back(MVT::Flag);
1521 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
1523 // Now that we have the comparison, emit a copy from the CR to a GPR.
1524 // This is flagged to the above dot comparison.
1525 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
1526 DAG.getRegister(PPC::CR6, MVT::i32),
1527 CompNode.getValue(1));
1529 // Unpack the result based on how the target uses it.
1530 unsigned BitNo; // Bit # of CR6.
1531 bool InvertBit; // Invert result?
1532 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1533 default: // Can't happen, don't crash on invalid number though.
1534 case 0: // Return the value of the EQ bit of CR6.
1535 BitNo = 0; InvertBit = false;
1537 case 1: // Return the inverted value of the EQ bit of CR6.
1538 BitNo = 0; InvertBit = true;
1540 case 2: // Return the value of the LT bit of CR6.
1541 BitNo = 2; InvertBit = false;
1543 case 3: // Return the inverted value of the LT bit of CR6.
1544 BitNo = 2; InvertBit = true;
1548 // Shift the bit into the low position.
1549 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
1550 DAG.getConstant(8-(3-BitNo), MVT::i32));
1552 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
1553 DAG.getConstant(1, MVT::i32));
1555 // If we are supposed to, toggle the bit.
1557 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
1558 DAG.getConstant(1, MVT::i32));
1562 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1563 // Create a stack slot that is 16-byte aligned.
1564 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
1565 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
1566 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
1568 // Store the input value into Value#0 of the stack slot.
1569 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
1570 Op.getOperand(0), FIdx,DAG.getSrcValue(NULL));
1572 return DAG.getLoad(Op.getValueType(), Store, FIdx, DAG.getSrcValue(NULL));
1575 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
1576 assert(Op.getValueType() == MVT::v4i32 && "Unknown mul to lower!");
1577 SDOperand LHS = Op.getOperand(0);
1578 SDOperand RHS = Op.getOperand(1);
1580 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
1581 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
1583 SDOperand RHSSwap = // = vrlw RHS, 16
1584 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
1586 // Shrinkify inputs to v8i16.
1587 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
1588 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
1589 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
1591 // Low parts multiplied together, generating 32-bit results (we ignore the top
1593 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
1594 LHS, RHS, DAG, MVT::v4i32);
1596 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
1597 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
1598 // Shift the high parts up 16 bits.
1599 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
1600 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
1603 /// LowerOperation - Provide custom lowering hooks for some operations.
1605 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
1606 switch (Op.getOpcode()) {
1607 default: assert(0 && "Wasn't expecting to be able to lower this!");
1608 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1609 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
1610 case ISD::SETCC: return LowerSETCC(Op, DAG);
1611 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1612 case ISD::RET: return LowerRET(Op, DAG);
1614 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
1615 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
1616 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
1618 // Lower 64-bit shifts.
1619 case ISD::SHL: return LowerSHL(Op, DAG);
1620 case ISD::SRL: return LowerSRL(Op, DAG);
1621 case ISD::SRA: return LowerSRA(Op, DAG);
1623 // Vector-related lowering.
1624 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
1625 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
1626 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
1627 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
1628 case ISD::MUL: return LowerMUL(Op, DAG);
1633 //===----------------------------------------------------------------------===//
1634 // Other Lowering Code
1635 //===----------------------------------------------------------------------===//
1637 std::vector<SDOperand>
1638 PPCTargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
1640 // add beautiful description of PPC stack frame format, or at least some docs
1642 MachineFunction &MF = DAG.getMachineFunction();
1643 MachineFrameInfo *MFI = MF.getFrameInfo();
1644 MachineBasicBlock& BB = MF.front();
1645 SSARegMap *RegMap = MF.getSSARegMap();
1646 std::vector<SDOperand> ArgValues;
1648 unsigned ArgOffset = 24;
1649 unsigned GPR_remaining = 8;
1650 unsigned FPR_remaining = 13;
1651 unsigned GPR_idx = 0, FPR_idx = 0;
1652 static const unsigned GPR[] = {
1653 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1654 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1656 static const unsigned FPR[] = {
1657 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1658 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1661 // Add DAG nodes to load the arguments... On entry to a function on PPC,
1662 // the arguments start at offset 24, although they are likely to be passed
1664 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
1665 SDOperand newroot, argt;
1667 bool needsLoad = false;
1668 bool ArgLive = !I->use_empty();
1669 MVT::ValueType ObjectVT = getValueType(I->getType());
1672 default: assert(0 && "Unhandled argument type!");
1678 if (!ArgLive) break;
1679 if (GPR_remaining > 0) {
1680 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1681 MF.addLiveIn(GPR[GPR_idx], VReg);
1682 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1683 if (ObjectVT != MVT::i32) {
1684 unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
1686 argt = DAG.getNode(AssertOp, MVT::i32, argt,
1687 DAG.getValueType(ObjectVT));
1688 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, argt);
1696 if (!ArgLive) break;
1697 if (GPR_remaining > 0) {
1698 SDOperand argHi, argLo;
1699 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1700 MF.addLiveIn(GPR[GPR_idx], VReg);
1701 argHi = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1702 // If we have two or more remaining argument registers, then both halves
1703 // of the i64 can be sourced from there. Otherwise, the lower half will
1704 // have to come off the stack. This can happen when an i64 is preceded
1705 // by 28 bytes of arguments.
1706 if (GPR_remaining > 1) {
1707 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1708 MF.addLiveIn(GPR[GPR_idx+1], VReg);
1709 argLo = DAG.getCopyFromReg(argHi, VReg, MVT::i32);
1711 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
1712 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
1713 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
1714 DAG.getSrcValue(NULL));
1716 // Build the outgoing arg thingy
1717 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
1725 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
1727 if (FPR_remaining > 0) {
1733 if (FPR_remaining > 0) {
1735 if (ObjectVT == MVT::f32)
1736 VReg = RegMap->createVirtualRegister(&PPC::F4RCRegClass);
1738 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1739 MF.addLiveIn(FPR[FPR_idx], VReg);
1740 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, ObjectVT);
1749 // We need to load the argument to a virtual register if we determined above
1750 // that we ran out of physical registers of the appropriate type
1752 unsigned SubregOffset = 0;
1753 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
1754 if (ObjectVT == MVT::i16) SubregOffset = 2;
1755 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1756 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
1757 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
1758 DAG.getConstant(SubregOffset, MVT::i32));
1759 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
1760 DAG.getSrcValue(NULL));
1763 // Every 4 bytes of argument space consumes one of the GPRs available for
1764 // argument passing.
1765 if (GPR_remaining > 0) {
1766 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
1767 GPR_remaining -= delta;
1770 ArgOffset += ObjSize;
1772 DAG.setRoot(newroot.getValue(1));
1774 ArgValues.push_back(argt);
1777 // If the function takes variable number of arguments, make a frame index for
1778 // the start of the first vararg value... for expansion of llvm.va_start.
1780 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1781 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
1782 // If this function is vararg, store any remaining integer argument regs
1783 // to their spots on the stack so that they may be loaded by deferencing the
1784 // result of va_next.
1785 std::vector<SDOperand> MemOps;
1786 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
1787 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1788 MF.addLiveIn(GPR[GPR_idx], VReg);
1789 SDOperand Val = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1790 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
1791 Val, FIN, DAG.getSrcValue(NULL));
1792 MemOps.push_back(Store);
1793 // Increment the address by four for the next argument to store
1794 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
1795 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
1797 if (!MemOps.empty()) {
1798 MemOps.push_back(DAG.getRoot());
1799 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
1806 std::pair<SDOperand, SDOperand>
1807 PPCTargetLowering::LowerCallTo(SDOperand Chain,
1808 const Type *RetTy, bool isVarArg,
1809 unsigned CallingConv, bool isTailCall,
1810 SDOperand Callee, ArgListTy &Args,
1811 SelectionDAG &DAG) {
1812 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1813 // SelectExpr to use to put the arguments in the appropriate registers.
1814 std::vector<SDOperand> args_to_use;
1816 // Count how many bytes are to be pushed on the stack, including the linkage
1817 // area, and parameter passing area.
1818 unsigned NumBytes = 24;
1821 Chain = DAG.getCALLSEQ_START(Chain,
1822 DAG.getConstant(NumBytes, getPointerTy()));
1824 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1825 switch (getValueType(Args[i].second)) {
1826 default: assert(0 && "Unknown value type!");
1841 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
1842 // plus 32 bytes of argument space in case any called code gets funky on us.
1843 // (Required by ABI to support var arg)
1844 if (NumBytes < 56) NumBytes = 56;
1846 // Adjust the stack pointer for the new arguments...
1847 // These operations are automatically eliminated by the prolog/epilog pass
1848 Chain = DAG.getCALLSEQ_START(Chain,
1849 DAG.getConstant(NumBytes, getPointerTy()));
1851 // Set up a copy of the stack pointer for use loading and storing any
1852 // arguments that may not fit in the registers available for argument
1854 SDOperand StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1856 // Figure out which arguments are going to go in registers, and which in
1857 // memory. Also, if this is a vararg function, floating point operations
1858 // must be stored to our stack, and loaded into integer regs as well, if
1859 // any integer regs are available for argument passing.
1860 unsigned ArgOffset = 24;
1861 unsigned GPR_remaining = 8;
1862 unsigned FPR_remaining = 13;
1864 std::vector<SDOperand> MemOps;
1865 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1866 // PtrOff will be used to store the current argument to the stack if a
1867 // register cannot be found for it.
1868 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1869 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
1870 MVT::ValueType ArgVT = getValueType(Args[i].second);
1873 default: assert(0 && "Unexpected ValueType for argument!");
1877 // Promote the integer to 32 bits. If the input type is signed use a
1878 // sign extend, otherwise use a zero extend.
1879 if (Args[i].second->isSigned())
1880 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
1882 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
1885 if (GPR_remaining > 0) {
1886 args_to_use.push_back(Args[i].first);
1889 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1890 Args[i].first, PtrOff,
1891 DAG.getSrcValue(NULL)));
1896 // If we have one free GPR left, we can place the upper half of the i64
1897 // in it, and store the other half to the stack. If we have two or more
1898 // free GPRs, then we can pass both halves of the i64 in registers.
1899 if (GPR_remaining > 0) {
1900 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1901 Args[i].first, DAG.getConstant(1, MVT::i32));
1902 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1903 Args[i].first, DAG.getConstant(0, MVT::i32));
1904 args_to_use.push_back(Hi);
1906 if (GPR_remaining > 0) {
1907 args_to_use.push_back(Lo);
1910 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1911 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1912 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1913 Lo, PtrOff, DAG.getSrcValue(NULL)));
1916 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1917 Args[i].first, PtrOff,
1918 DAG.getSrcValue(NULL)));
1924 if (FPR_remaining > 0) {
1925 args_to_use.push_back(Args[i].first);
1928 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
1929 Args[i].first, PtrOff,
1930 DAG.getSrcValue(NULL));
1931 MemOps.push_back(Store);
1932 // Float varargs are always shadowed in available integer registers
1933 if (GPR_remaining > 0) {
1934 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1935 DAG.getSrcValue(NULL));
1936 MemOps.push_back(Load.getValue(1));
1937 args_to_use.push_back(Load);
1940 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1941 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1942 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1943 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1944 DAG.getSrcValue(NULL));
1945 MemOps.push_back(Load.getValue(1));
1946 args_to_use.push_back(Load);
1950 // If we have any FPRs remaining, we may also have GPRs remaining.
1951 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1953 if (GPR_remaining > 0) {
1954 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1957 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1958 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1963 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1964 Args[i].first, PtrOff,
1965 DAG.getSrcValue(NULL)));
1967 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
1971 if (!MemOps.empty())
1972 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
1975 std::vector<MVT::ValueType> RetVals;
1976 MVT::ValueType RetTyVT = getValueType(RetTy);
1977 MVT::ValueType ActualRetTyVT = RetTyVT;
1978 if (RetTyVT >= MVT::i1 && RetTyVT <= MVT::i16)
1979 ActualRetTyVT = MVT::i32; // Promote result to i32.
1981 if (RetTyVT == MVT::i64) {
1982 RetVals.push_back(MVT::i32);
1983 RetVals.push_back(MVT::i32);
1984 } else if (RetTyVT != MVT::isVoid) {
1985 RetVals.push_back(ActualRetTyVT);
1987 RetVals.push_back(MVT::Other);
1989 // If the callee is a GlobalAddress node (quite common, every direct call is)
1990 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1991 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1992 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i32);
1994 std::vector<SDOperand> Ops;
1995 Ops.push_back(Chain);
1996 Ops.push_back(Callee);
1997 Ops.insert(Ops.end(), args_to_use.begin(), args_to_use.end());
1998 SDOperand TheCall = DAG.getNode(PPCISD::CALL, RetVals, Ops);
1999 Chain = TheCall.getValue(TheCall.Val->getNumValues()-1);
2000 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
2001 DAG.getConstant(NumBytes, getPointerTy()));
2002 SDOperand RetVal = TheCall;
2004 // If the result is a small value, add a note so that we keep track of the
2005 // information about whether it is sign or zero extended.
2006 if (RetTyVT != ActualRetTyVT) {
2007 RetVal = DAG.getNode(RetTy->isSigned() ? ISD::AssertSext : ISD::AssertZext,
2008 MVT::i32, RetVal, DAG.getValueType(RetTyVT));
2009 RetVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, RetVal);
2010 } else if (RetTyVT == MVT::i64) {
2011 RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, RetVal, RetVal.getValue(1));
2014 return std::make_pair(RetVal, Chain);
2018 PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
2019 MachineBasicBlock *BB) {
2020 assert((MI->getOpcode() == PPC::SELECT_CC_Int ||
2021 MI->getOpcode() == PPC::SELECT_CC_F4 ||
2022 MI->getOpcode() == PPC::SELECT_CC_F8 ||
2023 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
2024 "Unexpected instr type to insert");
2026 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
2027 // control-flow pattern. The incoming instruction knows the destination vreg
2028 // to set, the condition code register to branch on, the true/false values to
2029 // select between, and a branch opcode to use.
2030 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2031 ilist<MachineBasicBlock>::iterator It = BB;
2037 // cmpTY ccX, r1, r2
2039 // fallthrough --> copy0MBB
2040 MachineBasicBlock *thisMBB = BB;
2041 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
2042 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
2043 BuildMI(BB, MI->getOperand(4).getImmedValue(), 2)
2044 .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
2045 MachineFunction *F = BB->getParent();
2046 F->getBasicBlockList().insert(It, copy0MBB);
2047 F->getBasicBlockList().insert(It, sinkMBB);
2048 // Update machine-CFG edges by first adding all successors of the current
2049 // block to the new block which will contain the Phi node for the select.
2050 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2051 e = BB->succ_end(); i != e; ++i)
2052 sinkMBB->addSuccessor(*i);
2053 // Next, remove all successors of the current block, and add the true
2054 // and fallthrough blocks as its successors.
2055 while(!BB->succ_empty())
2056 BB->removeSuccessor(BB->succ_begin());
2057 BB->addSuccessor(copy0MBB);
2058 BB->addSuccessor(sinkMBB);
2061 // %FalseValue = ...
2062 // # fallthrough to sinkMBB
2065 // Update machine-CFG edges
2066 BB->addSuccessor(sinkMBB);
2069 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
2072 BuildMI(BB, PPC::PHI, 4, MI->getOperand(0).getReg())
2073 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
2074 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
2076 delete MI; // The pseudo instruction is gone now.
2080 //===----------------------------------------------------------------------===//
2081 // Target Optimization Hooks
2082 //===----------------------------------------------------------------------===//
2084 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
2085 DAGCombinerInfo &DCI) const {
2086 TargetMachine &TM = getTargetMachine();
2087 SelectionDAG &DAG = DCI.DAG;
2088 switch (N->getOpcode()) {
2090 case ISD::SINT_TO_FP:
2091 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
2092 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
2093 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
2094 // We allow the src/dst to be either f32/f64, but the intermediate
2095 // type must be i64.
2096 if (N->getOperand(0).getValueType() == MVT::i64) {
2097 SDOperand Val = N->getOperand(0).getOperand(0);
2098 if (Val.getValueType() == MVT::f32) {
2099 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
2100 DCI.AddToWorklist(Val.Val);
2103 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
2104 DCI.AddToWorklist(Val.Val);
2105 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
2106 DCI.AddToWorklist(Val.Val);
2107 if (N->getValueType(0) == MVT::f32) {
2108 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
2109 DCI.AddToWorklist(Val.Val);
2112 } else if (N->getOperand(0).getValueType() == MVT::i32) {
2113 // If the intermediate type is i32, we can avoid the load/store here
2120 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
2121 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
2122 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
2123 N->getOperand(1).getValueType() == MVT::i32) {
2124 SDOperand Val = N->getOperand(1).getOperand(0);
2125 if (Val.getValueType() == MVT::f32) {
2126 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
2127 DCI.AddToWorklist(Val.Val);
2129 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
2130 DCI.AddToWorklist(Val.Val);
2132 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
2133 N->getOperand(2), N->getOperand(3));
2134 DCI.AddToWorklist(Val.Val);
2138 case PPCISD::VCMP: {
2139 // If a VCMPo node already exists with exactly the same operands as this
2140 // node, use its result instead of this node (VCMPo computes both a CR6 and
2141 // a normal output).
2143 if (!N->getOperand(0).hasOneUse() &&
2144 !N->getOperand(1).hasOneUse() &&
2145 !N->getOperand(2).hasOneUse()) {
2147 // Scan all of the users of the LHS, looking for VCMPo's that match.
2148 SDNode *VCMPoNode = 0;
2150 SDNode *LHSN = N->getOperand(0).Val;
2151 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
2153 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
2154 (*UI)->getOperand(1) == N->getOperand(1) &&
2155 (*UI)->getOperand(2) == N->getOperand(2) &&
2156 (*UI)->getOperand(0) == N->getOperand(0)) {
2161 // If there are non-zero uses of the flag value, use the VCMPo node!
2162 if (VCMPoNode && !VCMPoNode->hasNUsesOfValue(0, 1))
2163 return SDOperand(VCMPoNode, 0);
2172 //===----------------------------------------------------------------------===//
2173 // Inline Assembly Support
2174 //===----------------------------------------------------------------------===//
2176 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
2178 uint64_t &KnownZero,
2180 unsigned Depth) const {
2183 switch (Op.getOpcode()) {
2185 case ISD::INTRINSIC_WO_CHAIN: {
2186 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
2188 case Intrinsic::ppc_altivec_vcmpbfp_p:
2189 case Intrinsic::ppc_altivec_vcmpeqfp_p:
2190 case Intrinsic::ppc_altivec_vcmpequb_p:
2191 case Intrinsic::ppc_altivec_vcmpequh_p:
2192 case Intrinsic::ppc_altivec_vcmpequw_p:
2193 case Intrinsic::ppc_altivec_vcmpgefp_p:
2194 case Intrinsic::ppc_altivec_vcmpgtfp_p:
2195 case Intrinsic::ppc_altivec_vcmpgtsb_p:
2196 case Intrinsic::ppc_altivec_vcmpgtsh_p:
2197 case Intrinsic::ppc_altivec_vcmpgtsw_p:
2198 case Intrinsic::ppc_altivec_vcmpgtub_p:
2199 case Intrinsic::ppc_altivec_vcmpgtuh_p:
2200 case Intrinsic::ppc_altivec_vcmpgtuw_p:
2201 KnownZero = ~1U; // All bits but the low one are known to be zero.
2209 /// getConstraintType - Given a constraint letter, return the type of
2210 /// constraint it is for this target.
2211 PPCTargetLowering::ConstraintType
2212 PPCTargetLowering::getConstraintType(char ConstraintLetter) const {
2213 switch (ConstraintLetter) {
2220 return C_RegisterClass;
2222 return TargetLowering::getConstraintType(ConstraintLetter);
2226 std::vector<unsigned> PPCTargetLowering::
2227 getRegClassForInlineAsmConstraint(const std::string &Constraint,
2228 MVT::ValueType VT) const {
2229 if (Constraint.size() == 1) {
2230 switch (Constraint[0]) { // GCC RS6000 Constraint Letters
2231 default: break; // Unknown constriant letter
2233 return make_vector<unsigned>(/*no R0*/ PPC::R1 , PPC::R2 , PPC::R3 ,
2234 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
2235 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
2236 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
2237 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
2238 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
2239 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
2240 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
2243 return make_vector<unsigned>(PPC::R0 , PPC::R1 , PPC::R2 , PPC::R3 ,
2244 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
2245 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
2246 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
2247 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
2248 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
2249 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
2250 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
2253 return make_vector<unsigned>(PPC::F0 , PPC::F1 , PPC::F2 , PPC::F3 ,
2254 PPC::F4 , PPC::F5 , PPC::F6 , PPC::F7 ,
2255 PPC::F8 , PPC::F9 , PPC::F10, PPC::F11,
2256 PPC::F12, PPC::F13, PPC::F14, PPC::F15,
2257 PPC::F16, PPC::F17, PPC::F18, PPC::F19,
2258 PPC::F20, PPC::F21, PPC::F22, PPC::F23,
2259 PPC::F24, PPC::F25, PPC::F26, PPC::F27,
2260 PPC::F28, PPC::F29, PPC::F30, PPC::F31,
2263 return make_vector<unsigned>(PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 ,
2264 PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 ,
2265 PPC::V8 , PPC::V9 , PPC::V10, PPC::V11,
2266 PPC::V12, PPC::V13, PPC::V14, PPC::V15,
2267 PPC::V16, PPC::V17, PPC::V18, PPC::V19,
2268 PPC::V20, PPC::V21, PPC::V22, PPC::V23,
2269 PPC::V24, PPC::V25, PPC::V26, PPC::V27,
2270 PPC::V28, PPC::V29, PPC::V30, PPC::V31,
2273 return make_vector<unsigned>(PPC::CR0, PPC::CR1, PPC::CR2, PPC::CR3,
2274 PPC::CR4, PPC::CR5, PPC::CR6, PPC::CR7,
2279 return std::vector<unsigned>();
2282 // isOperandValidForConstraint
2283 bool PPCTargetLowering::
2284 isOperandValidForConstraint(SDOperand Op, char Letter) {
2295 if (!isa<ConstantSDNode>(Op)) return false; // Must be an immediate.
2296 unsigned Value = cast<ConstantSDNode>(Op)->getValue();
2298 default: assert(0 && "Unknown constraint letter!");
2299 case 'I': // "I" is a signed 16-bit constant.
2300 return (short)Value == (int)Value;
2301 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
2302 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
2303 return (short)Value == 0;
2304 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
2305 return (Value >> 16) == 0;
2306 case 'M': // "M" is a constant that is greater than 31.
2308 case 'N': // "N" is a positive constant that is an exact power of two.
2309 return (int)Value > 0 && isPowerOf2_32(Value);
2310 case 'O': // "O" is the constant zero.
2312 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
2313 return (short)-Value == (int)-Value;
2319 // Handle standard constraint letters.
2320 return TargetLowering::isOperandValidForConstraint(Op, Letter);
2323 /// isLegalAddressImmediate - Return true if the integer value can be used
2324 /// as the offset of the target addressing mode.
2325 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V) const {
2326 // PPC allows a sign-extended 16-bit immediate field.
2327 return (V > -(1 << 16) && V < (1 << 16)-1);