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 "llvm/ADT/VectorExtras.h"
17 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/SelectionDAG.h"
22 #include "llvm/CodeGen/SSARegMap.h"
23 #include "llvm/Constants.h"
24 #include "llvm/Function.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/Support/MathExtras.h"
27 #include "llvm/Target/TargetOptions.h"
30 PPCTargetLowering::PPCTargetLowering(TargetMachine &TM)
31 : TargetLowering(TM) {
33 // Fold away setcc operations if possible.
34 setSetCCIsExpensive();
37 // Use _setjmp/_longjmp instead of setjmp/longjmp.
38 setUseUnderscoreSetJmpLongJmp(true);
40 // Set up the register classes.
41 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
42 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
43 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
45 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
46 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
48 // PowerPC has no intrinsics for these particular operations
49 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
50 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
51 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
53 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
54 setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
55 setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
57 // PowerPC has no SREM/UREM instructions
58 setOperationAction(ISD::SREM, MVT::i32, Expand);
59 setOperationAction(ISD::UREM, MVT::i32, Expand);
61 // We don't support sin/cos/sqrt/fmod
62 setOperationAction(ISD::FSIN , MVT::f64, Expand);
63 setOperationAction(ISD::FCOS , MVT::f64, Expand);
64 setOperationAction(ISD::FREM , MVT::f64, Expand);
65 setOperationAction(ISD::FSIN , MVT::f32, Expand);
66 setOperationAction(ISD::FCOS , MVT::f32, Expand);
67 setOperationAction(ISD::FREM , MVT::f32, Expand);
69 // If we're enabling GP optimizations, use hardware square root
70 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
71 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
72 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
75 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
76 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
78 // PowerPC does not have BSWAP, CTPOP or CTTZ
79 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
80 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
81 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
83 // PowerPC does not have ROTR
84 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
86 // PowerPC does not have Select
87 setOperationAction(ISD::SELECT, MVT::i32, Expand);
88 setOperationAction(ISD::SELECT, MVT::f32, Expand);
89 setOperationAction(ISD::SELECT, MVT::f64, Expand);
90 setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
91 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
92 setOperationAction(ISD::SELECT, MVT::v8i16, Expand);
93 setOperationAction(ISD::SELECT, MVT::v16i8, Expand);
95 // PowerPC wants to turn select_cc of FP into fsel when possible.
96 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
97 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
99 // PowerPC wants to optimize integer setcc a bit
100 setOperationAction(ISD::SETCC, MVT::i32, Custom);
102 // PowerPC does not have BRCOND which requires SetCC
103 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
105 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
106 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
108 // PowerPC does not have [U|S]INT_TO_FP
109 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
110 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
112 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
113 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
115 // PowerPC does not have truncstore for i1.
116 setOperationAction(ISD::TRUNCSTORE, MVT::i1, Promote);
118 // Support label based line numbers.
119 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
120 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
121 // FIXME - use subtarget debug flags
122 if (!TM.getSubtarget<PPCSubtarget>().isDarwin())
123 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
125 // We want to legalize GlobalAddress and ConstantPool nodes into the
126 // appropriate instructions to materialize the address.
127 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
128 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
130 // RET must be custom lowered, to meet ABI requirements
131 setOperationAction(ISD::RET , MVT::Other, Custom);
133 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
134 setOperationAction(ISD::VASTART , MVT::Other, Custom);
136 // Use the default implementation.
137 setOperationAction(ISD::VAARG , MVT::Other, Expand);
138 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
139 setOperationAction(ISD::VAEND , MVT::Other, Expand);
140 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
141 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
142 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
144 // We want to custom lower some of our intrinsics.
145 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
147 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
148 // They also have instructions for converting between i64 and fp.
149 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
150 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
152 // FIXME: disable this lowered code. This generates 64-bit register values,
153 // and we don't model the fact that the top part is clobbered by calls. We
154 // need to flag these together so that the value isn't live across a call.
155 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
157 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
158 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
160 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
161 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
164 if (TM.getSubtarget<PPCSubtarget>().has64BitRegs()) {
165 // 64 bit PowerPC implementations can support i64 types directly
166 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
167 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
168 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
170 // 32 bit PowerPC wants to expand i64 shifts itself.
171 setOperationAction(ISD::SHL, MVT::i64, Custom);
172 setOperationAction(ISD::SRL, MVT::i64, Custom);
173 setOperationAction(ISD::SRA, MVT::i64, Custom);
176 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
177 // First set operation action for all vector types to expand. Then we
178 // will selectively turn on ones that can be effectively codegen'd.
179 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
180 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
181 // add/sub/and/or/xor are legal for all supported vector VT's.
182 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
183 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
184 setOperationAction(ISD::AND , (MVT::ValueType)VT, Legal);
185 setOperationAction(ISD::OR , (MVT::ValueType)VT, Legal);
186 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Legal);
188 // We promote all shuffles to v16i8.
189 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
190 AddPromotedToType(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
192 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
193 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
194 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
195 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
196 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
197 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
198 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
199 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
201 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
204 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
205 // with merges, splats, etc.
206 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
208 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
209 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
210 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
211 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
213 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
215 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
216 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
218 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
219 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
220 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
221 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
224 setSetCCResultContents(ZeroOrOneSetCCResult);
225 setStackPointerRegisterToSaveRestore(PPC::R1);
227 // We have target-specific dag combine patterns for the following nodes:
228 setTargetDAGCombine(ISD::SINT_TO_FP);
229 setTargetDAGCombine(ISD::STORE);
231 computeRegisterProperties();
234 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
237 case PPCISD::FSEL: return "PPCISD::FSEL";
238 case PPCISD::FCFID: return "PPCISD::FCFID";
239 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
240 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
241 case PPCISD::STFIWX: return "PPCISD::STFIWX";
242 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
243 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
244 case PPCISD::VPERM: return "PPCISD::VPERM";
245 case PPCISD::Hi: return "PPCISD::Hi";
246 case PPCISD::Lo: return "PPCISD::Lo";
247 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
248 case PPCISD::SRL: return "PPCISD::SRL";
249 case PPCISD::SRA: return "PPCISD::SRA";
250 case PPCISD::SHL: return "PPCISD::SHL";
251 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
252 case PPCISD::STD_32: return "PPCISD::STD_32";
253 case PPCISD::CALL: return "PPCISD::CALL";
254 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
255 case PPCISD::MFCR: return "PPCISD::MFCR";
256 case PPCISD::VCMP: return "PPCISD::VCMP";
257 case PPCISD::VCMPo: return "PPCISD::VCMPo";
261 //===----------------------------------------------------------------------===//
262 // Node matching predicates, for use by the tblgen matching code.
263 //===----------------------------------------------------------------------===//
265 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
266 static bool isFloatingPointZero(SDOperand Op) {
267 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
268 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
269 else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
270 // Maybe this has already been legalized into the constant pool?
271 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
272 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
273 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
278 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
279 /// true if Op is undef or if it matches the specified value.
280 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
281 return Op.getOpcode() == ISD::UNDEF ||
282 cast<ConstantSDNode>(Op)->getValue() == Val;
285 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
286 /// VPKUHUM instruction.
287 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
289 for (unsigned i = 0; i != 16; ++i)
290 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
293 for (unsigned i = 0; i != 8; ++i)
294 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
295 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
301 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
302 /// VPKUWUM instruction.
303 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
305 for (unsigned i = 0; i != 16; i += 2)
306 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
307 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
310 for (unsigned i = 0; i != 8; i += 2)
311 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
312 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
313 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
314 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
320 /// isVMerge - Common function, used to match vmrg* shuffles.
322 static bool isVMerge(SDNode *N, unsigned UnitSize,
323 unsigned LHSStart, unsigned RHSStart) {
324 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
325 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
326 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
327 "Unsupported merge size!");
329 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
330 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
331 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
332 LHSStart+j+i*UnitSize) ||
333 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
334 RHSStart+j+i*UnitSize))
340 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
341 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
342 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
344 return isVMerge(N, UnitSize, 8, 24);
345 return isVMerge(N, UnitSize, 8, 8);
348 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
349 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
350 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
352 return isVMerge(N, UnitSize, 0, 16);
353 return isVMerge(N, UnitSize, 0, 0);
357 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
358 /// amount, otherwise return -1.
359 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
360 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
361 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
362 // Find the first non-undef value in the shuffle mask.
364 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
367 if (i == 16) return -1; // all undef.
369 // Otherwise, check to see if the rest of the elements are consequtively
370 // numbered from this value.
371 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
372 if (ShiftAmt < i) return -1;
376 // Check the rest of the elements to see if they are consequtive.
377 for (++i; i != 16; ++i)
378 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
381 // Check the rest of the elements to see if they are consequtive.
382 for (++i; i != 16; ++i)
383 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
390 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
391 /// specifies a splat of a single element that is suitable for input to
392 /// VSPLTB/VSPLTH/VSPLTW.
393 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
394 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
395 N->getNumOperands() == 16 &&
396 (EltSize == 1 || EltSize == 2 || EltSize == 4));
398 // This is a splat operation if each element of the permute is the same, and
399 // if the value doesn't reference the second vector.
400 unsigned ElementBase = 0;
401 SDOperand Elt = N->getOperand(0);
402 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
403 ElementBase = EltV->getValue();
405 return false; // FIXME: Handle UNDEF elements too!
407 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
410 // Check that they are consequtive.
411 for (unsigned i = 1; i != EltSize; ++i) {
412 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
413 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
417 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
418 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
419 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
420 "Invalid VECTOR_SHUFFLE mask!");
421 for (unsigned j = 0; j != EltSize; ++j)
422 if (N->getOperand(i+j) != N->getOperand(j))
429 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
430 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
431 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
432 assert(isSplatShuffleMask(N, EltSize));
433 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
436 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
437 /// by using a vspltis[bhw] instruction of the specified element size, return
438 /// the constant being splatted. The ByteSize field indicates the number of
439 /// bytes of each element [124] -> [bhw].
440 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
441 SDOperand OpVal(0, 0);
443 // If ByteSize of the splat is bigger than the element size of the
444 // build_vector, then we have a case where we are checking for a splat where
445 // multiple elements of the buildvector are folded together into a single
446 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
447 unsigned EltSize = 16/N->getNumOperands();
448 if (EltSize < ByteSize) {
449 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
450 SDOperand UniquedVals[4];
451 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
453 // See if all of the elements in the buildvector agree across.
454 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
455 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
456 // If the element isn't a constant, bail fully out.
457 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
460 if (UniquedVals[i&(Multiple-1)].Val == 0)
461 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
462 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
463 return SDOperand(); // no match.
466 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
467 // either constant or undef values that are identical for each chunk. See
468 // if these chunks can form into a larger vspltis*.
470 // Check to see if all of the leading entries are either 0 or -1. If
471 // neither, then this won't fit into the immediate field.
472 bool LeadingZero = true;
473 bool LeadingOnes = true;
474 for (unsigned i = 0; i != Multiple-1; ++i) {
475 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
477 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
478 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
480 // Finally, check the least significant entry.
482 if (UniquedVals[Multiple-1].Val == 0)
483 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
484 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
486 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
489 if (UniquedVals[Multiple-1].Val == 0)
490 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
491 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
492 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
493 return DAG.getTargetConstant(Val, MVT::i32);
499 // Check to see if this buildvec has a single non-undef value in its elements.
500 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
501 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
503 OpVal = N->getOperand(i);
504 else if (OpVal != N->getOperand(i))
508 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
510 unsigned ValSizeInBytes = 0;
512 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
513 Value = CN->getValue();
514 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
515 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
516 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
517 Value = FloatToBits(CN->getValue());
521 // If the splat value is larger than the element value, then we can never do
522 // this splat. The only case that we could fit the replicated bits into our
523 // immediate field for would be zero, and we prefer to use vxor for it.
524 if (ValSizeInBytes < ByteSize) return SDOperand();
526 // If the element value is larger than the splat value, cut it in half and
527 // check to see if the two halves are equal. Continue doing this until we
528 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
529 while (ValSizeInBytes > ByteSize) {
530 ValSizeInBytes >>= 1;
532 // If the top half equals the bottom half, we're still ok.
533 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
534 (Value & ((1 << (8*ValSizeInBytes))-1)))
538 // Properly sign extend the value.
539 int ShAmt = (4-ByteSize)*8;
540 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
542 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
543 if (MaskVal == 0) return SDOperand();
545 // Finally, if this value fits in a 5 bit sext field, return it
546 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
547 return DAG.getTargetConstant(MaskVal, MVT::i32);
551 //===----------------------------------------------------------------------===//
552 // LowerOperation implementation
553 //===----------------------------------------------------------------------===//
555 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
556 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
557 Constant *C = CP->get();
558 SDOperand CPI = DAG.getTargetConstantPool(C, MVT::i32, CP->getAlignment());
559 SDOperand Zero = DAG.getConstant(0, MVT::i32);
561 const TargetMachine &TM = DAG.getTarget();
563 // If this is a non-darwin platform, we don't support non-static relo models
565 if (TM.getRelocationModel() == Reloc::Static ||
566 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
567 // Generate non-pic code that has direct accesses to the constant pool.
568 // The address of the global is just (hi(&g)+lo(&g)).
569 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
570 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
571 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
574 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
575 if (TM.getRelocationModel() == Reloc::PIC) {
576 // With PIC, the first instruction is actually "GR+hi(&G)".
577 Hi = DAG.getNode(ISD::ADD, MVT::i32,
578 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
581 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
582 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
586 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
587 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
588 GlobalValue *GV = GSDN->getGlobal();
589 SDOperand GA = DAG.getTargetGlobalAddress(GV, MVT::i32, GSDN->getOffset());
590 SDOperand Zero = DAG.getConstant(0, MVT::i32);
592 const TargetMachine &TM = DAG.getTarget();
594 // If this is a non-darwin platform, we don't support non-static relo models
596 if (TM.getRelocationModel() == Reloc::Static ||
597 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
598 // Generate non-pic code that has direct accesses to globals.
599 // The address of the global is just (hi(&g)+lo(&g)).
600 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
601 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
602 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
605 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
606 if (TM.getRelocationModel() == Reloc::PIC) {
607 // With PIC, the first instruction is actually "GR+hi(&G)".
608 Hi = DAG.getNode(ISD::ADD, MVT::i32,
609 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
612 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
613 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
615 if (!GV->hasWeakLinkage() && !GV->hasLinkOnceLinkage() &&
616 (!GV->isExternal() || GV->hasNotBeenReadFromBytecode()))
619 // If the global is weak or external, we have to go through the lazy
621 return DAG.getLoad(MVT::i32, DAG.getEntryNode(), Lo, DAG.getSrcValue(0));
624 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
625 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
627 // If we're comparing for equality to zero, expose the fact that this is
628 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
629 // fold the new nodes.
630 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
631 if (C->isNullValue() && CC == ISD::SETEQ) {
632 MVT::ValueType VT = Op.getOperand(0).getValueType();
633 SDOperand Zext = Op.getOperand(0);
636 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
638 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
639 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
640 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
641 DAG.getConstant(Log2b, MVT::i32));
642 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
644 // Leave comparisons against 0 and -1 alone for now, since they're usually
645 // optimized. FIXME: revisit this when we can custom lower all setcc
647 if (C->isAllOnesValue() || C->isNullValue())
651 // If we have an integer seteq/setne, turn it into a compare against zero
652 // by subtracting the rhs from the lhs, which is faster than setting a
653 // condition register, reading it back out, and masking the correct bit.
654 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
655 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
656 MVT::ValueType VT = Op.getValueType();
657 SDOperand Sub = DAG.getNode(ISD::SUB, LHSVT, Op.getOperand(0),
659 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
664 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
665 unsigned VarArgsFrameIndex) {
666 // vastart just stores the address of the VarArgsFrameIndex slot into the
667 // memory location argument.
668 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
669 return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
670 Op.getOperand(1), Op.getOperand(2));
673 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG) {
675 switch(Op.getNumOperands()) {
677 assert(0 && "Do not know how to return this many arguments!");
680 return SDOperand(); // ret void is legal
682 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
684 if (MVT::isVector(ArgVT))
686 else if (MVT::isInteger(ArgVT))
689 assert(MVT::isFloatingPoint(ArgVT));
693 Copy = DAG.getCopyToReg(Op.getOperand(0), ArgReg, Op.getOperand(1),
696 // If we haven't noted the R3/F1 are live out, do so now.
697 if (DAG.getMachineFunction().liveout_empty())
698 DAG.getMachineFunction().addLiveOut(ArgReg);
702 Copy = DAG.getCopyToReg(Op.getOperand(0), PPC::R3, Op.getOperand(2),
704 Copy = DAG.getCopyToReg(Copy, PPC::R4, Op.getOperand(1),Copy.getValue(1));
705 // If we haven't noted the R3+R4 are live out, do so now.
706 if (DAG.getMachineFunction().liveout_empty()) {
707 DAG.getMachineFunction().addLiveOut(PPC::R3);
708 DAG.getMachineFunction().addLiveOut(PPC::R4);
712 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
715 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
717 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
718 // Not FP? Not a fsel.
719 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
720 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
723 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
725 // Cannot handle SETEQ/SETNE.
726 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
728 MVT::ValueType ResVT = Op.getValueType();
729 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
730 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
731 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
733 // If the RHS of the comparison is a 0.0, we don't need to do the
734 // subtraction at all.
735 if (isFloatingPointZero(RHS))
737 default: break; // SETUO etc aren't handled by fsel.
740 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
743 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
744 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
745 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
748 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
751 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
752 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
753 return DAG.getNode(PPCISD::FSEL, ResVT,
754 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
759 default: break; // SETUO etc aren't handled by fsel.
762 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
763 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
764 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
765 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
768 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
769 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
770 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
771 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
774 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
775 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
776 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
777 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
780 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
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, TV, FV);
788 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
789 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
790 SDOperand Src = Op.getOperand(0);
791 if (Src.getValueType() == MVT::f32)
792 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
795 switch (Op.getValueType()) {
796 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
798 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
801 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
805 // Convert the FP value to an int value through memory.
806 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
807 if (Op.getValueType() == MVT::i32)
808 Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
812 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
813 if (Op.getOperand(0).getValueType() == MVT::i64) {
814 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
815 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
816 if (Op.getValueType() == MVT::f32)
817 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
821 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
822 "Unhandled SINT_TO_FP type in custom expander!");
823 // Since we only generate this in 64-bit mode, we can take advantage of
824 // 64-bit registers. In particular, sign extend the input value into the
825 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
826 // then lfd it and fcfid it.
827 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
828 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
829 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
831 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
834 // STD the extended value into the stack slot.
835 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
836 DAG.getEntryNode(), Ext64, FIdx,
837 DAG.getSrcValue(NULL));
838 // Load the value as a double.
839 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
841 // FCFID it and return it.
842 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
843 if (Op.getValueType() == MVT::f32)
844 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
848 static SDOperand LowerSHL(SDOperand Op, SelectionDAG &DAG) {
849 assert(Op.getValueType() == MVT::i64 &&
850 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
851 // The generic code does a fine job expanding shift by a constant.
852 if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
854 // Otherwise, expand into a bunch of logical ops. Note that these ops
855 // depend on the PPC behavior for oversized shift amounts.
856 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
857 DAG.getConstant(0, MVT::i32));
858 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
859 DAG.getConstant(1, MVT::i32));
860 SDOperand Amt = Op.getOperand(1);
862 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
863 DAG.getConstant(32, MVT::i32), Amt);
864 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
865 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
866 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
867 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
868 DAG.getConstant(-32U, MVT::i32));
869 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
870 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
871 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
872 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
875 static SDOperand LowerSRL(SDOperand Op, SelectionDAG &DAG) {
876 assert(Op.getValueType() == MVT::i64 &&
877 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
878 // The generic code does a fine job expanding shift by a constant.
879 if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
881 // Otherwise, expand into a bunch of logical ops. Note that these ops
882 // depend on the PPC behavior for oversized shift amounts.
883 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
884 DAG.getConstant(0, MVT::i32));
885 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
886 DAG.getConstant(1, MVT::i32));
887 SDOperand Amt = Op.getOperand(1);
889 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
890 DAG.getConstant(32, MVT::i32), Amt);
891 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
892 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
893 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
894 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
895 DAG.getConstant(-32U, MVT::i32));
896 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
897 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
898 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
899 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
902 static SDOperand LowerSRA(SDOperand Op, SelectionDAG &DAG) {
903 assert(Op.getValueType() == MVT::i64 &&
904 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
905 // The generic code does a fine job expanding shift by a constant.
906 if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
908 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
909 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
910 DAG.getConstant(0, MVT::i32));
911 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
912 DAG.getConstant(1, MVT::i32));
913 SDOperand Amt = Op.getOperand(1);
915 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
916 DAG.getConstant(32, MVT::i32), Amt);
917 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
918 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
919 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
920 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
921 DAG.getConstant(-32U, MVT::i32));
922 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
923 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
924 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
925 Tmp4, Tmp6, ISD::SETLE);
926 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
929 //===----------------------------------------------------------------------===//
930 // Vector related lowering.
933 // If this is a vector of constants or undefs, get the bits. A bit in
934 // UndefBits is set if the corresponding element of the vector is an
935 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
936 // zero. Return true if this is not an array of constants, false if it is.
938 // Note that VectorBits/UndefBits are returned in 'little endian' form, so
939 // elements 0,1 go in VectorBits[0] and 2,3 go in VectorBits[1] for a v4i32.
940 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
941 uint64_t UndefBits[2]) {
942 // Start with zero'd results.
943 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
945 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
946 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
947 SDOperand OpVal = BV->getOperand(i);
949 unsigned PartNo = i >= e/2; // In the upper 128 bits?
950 unsigned SlotNo = i & (e/2-1); // Which subpiece of the uint64_t it is.
952 uint64_t EltBits = 0;
953 if (OpVal.getOpcode() == ISD::UNDEF) {
954 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
955 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
957 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
958 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
959 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
960 assert(CN->getValueType(0) == MVT::f32 &&
961 "Only one legal FP vector type!");
962 EltBits = FloatToBits(CN->getValue());
964 // Nonconstant element.
968 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
971 //printf("%llx %llx %llx %llx\n",
972 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
976 // If this is a case we can't handle, return null and let the default
977 // expansion code take care of it. If we CAN select this case, and if it
978 // selects to a single instruction, return Op. Otherwise, if we can codegen
979 // this case more efficiently than a constant pool load, lower it to the
980 // sequence of ops that should be used.
981 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
982 // If this is a vector of constants or undefs, get the bits. A bit in
983 // UndefBits is set if the corresponding element of the vector is an
984 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
986 uint64_t VectorBits[2];
987 uint64_t UndefBits[2];
988 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
989 return SDOperand(); // Not a constant vector.
991 // See if this is all zeros.
992 if ((VectorBits[0] | VectorBits[1]) == 0) {
993 // Canonicalize all zero vectors to be v4i32.
994 if (Op.getValueType() != MVT::v4i32) {
995 SDOperand Z = DAG.getConstant(0, MVT::i32);
996 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
997 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
1002 // Check to see if this is something we can use VSPLTI* to form.
1003 MVT::ValueType CanonicalVT = MVT::Other;
1006 if ((CST = PPC::get_VSPLTI_elt(Op.Val, 4, DAG).Val)) // vspltisw
1007 CanonicalVT = MVT::v4i32;
1008 else if ((CST = PPC::get_VSPLTI_elt(Op.Val, 2, DAG).Val)) // vspltish
1009 CanonicalVT = MVT::v8i16;
1010 else if ((CST = PPC::get_VSPLTI_elt(Op.Val, 1, DAG).Val)) // vspltisb
1011 CanonicalVT = MVT::v16i8;
1013 // If this matches one of the vsplti* patterns, force it to the canonical
1014 // type for the pattern.
1016 if (Op.getValueType() != CanonicalVT) {
1017 // Convert the splatted element to the right element type.
1018 SDOperand Elt = DAG.getNode(ISD::TRUNCATE,
1019 MVT::getVectorBaseType(CanonicalVT),
1021 std::vector<SDOperand> Ops(MVT::getVectorNumElements(CanonicalVT), Elt);
1022 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT, Ops);
1023 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
1028 // If this is some other splat of 4-byte elements, see if we can handle it
1030 // FIXME: Make this more undef happy and work with other widths (1,2 bytes).
1031 if (VectorBits[0] == VectorBits[1] &&
1032 unsigned(VectorBits[0]) == unsigned(VectorBits[0] >> 32)) {
1033 unsigned Bits = unsigned(VectorBits[0]);
1035 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
1036 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). These are important
1038 if (Bits == 0x80000000 || Bits == 0x7FFFFFFF) {
1039 // Make -1 and vspltisw -1:
1040 SDOperand OnesI = DAG.getConstant(~0U, MVT::i32);
1041 SDOperand OnesV = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
1042 OnesI, OnesI, OnesI, OnesI);
1044 // Make the VSLW intrinsic, computing 0x8000_0000.
1046 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, MVT::v4i32,
1047 DAG.getConstant(Intrinsic::ppc_altivec_vslw, MVT::i32),
1050 // If this is 0x7FFF_FFFF, xor by OnesV to invert it.
1051 if (Bits == 0x7FFFFFFF)
1052 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
1054 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
1061 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
1062 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
1063 /// return the code it can be lowered into. Worst case, it can always be
1064 /// lowered into a vperm.
1065 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
1066 SDOperand V1 = Op.getOperand(0);
1067 SDOperand V2 = Op.getOperand(1);
1068 SDOperand PermMask = Op.getOperand(2);
1070 // Cases that are handled by instructions that take permute immediates
1071 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
1072 // selected by the instruction selector.
1073 if (V2.getOpcode() == ISD::UNDEF) {
1074 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
1075 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
1076 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
1077 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
1078 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
1079 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
1080 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
1081 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
1082 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
1083 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
1084 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
1085 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
1090 // Altivec has a variety of "shuffle immediates" that take two vector inputs
1091 // and produce a fixed permutation. If any of these match, do not lower to
1093 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
1094 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
1095 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
1096 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
1097 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
1098 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
1099 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
1100 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
1101 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
1104 // TODO: Handle more cases, and also handle cases that are cheaper to do as
1105 // multiple such instructions than as a constant pool load/vperm pair.
1107 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
1108 // vector that will get spilled to the constant pool.
1109 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
1111 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
1112 // that it is in input element units, not in bytes. Convert now.
1113 MVT::ValueType EltVT = MVT::getVectorBaseType(V1.getValueType());
1114 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
1116 std::vector<SDOperand> ResultMask;
1117 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
1118 unsigned SrcElt =cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
1120 for (unsigned j = 0; j != BytesPerElement; ++j)
1121 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
1125 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, ResultMask);
1126 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
1129 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
1130 /// lower, do it, otherwise return null.
1131 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
1132 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
1134 // If this is a lowered altivec predicate compare, CompareOpc is set to the
1135 // opcode number of the comparison.
1136 int CompareOpc = -1;
1139 default: return SDOperand(); // Don't custom lower most intrinsics.
1140 // Comparison predicates.
1141 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
1142 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
1143 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
1144 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
1145 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
1146 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
1147 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
1148 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
1149 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
1150 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
1151 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
1152 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
1153 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
1155 // Normal Comparisons.
1156 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
1157 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
1158 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
1159 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
1160 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
1161 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
1162 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
1163 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
1164 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
1165 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
1166 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
1167 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
1168 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
1171 assert(CompareOpc>0 && "We only lower altivec predicate compares so far!");
1173 // If this is a non-dot comparison, make the VCMP node.
1175 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
1176 Op.getOperand(1), Op.getOperand(2),
1177 DAG.getConstant(CompareOpc, MVT::i32));
1178 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
1181 // Create the PPCISD altivec 'dot' comparison node.
1182 std::vector<SDOperand> Ops;
1183 std::vector<MVT::ValueType> VTs;
1184 Ops.push_back(Op.getOperand(2)); // LHS
1185 Ops.push_back(Op.getOperand(3)); // RHS
1186 Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
1187 VTs.push_back(Op.getOperand(2).getValueType());
1188 VTs.push_back(MVT::Flag);
1189 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
1191 // Now that we have the comparison, emit a copy from the CR to a GPR.
1192 // This is flagged to the above dot comparison.
1193 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
1194 DAG.getRegister(PPC::CR6, MVT::i32),
1195 CompNode.getValue(1));
1197 // Unpack the result based on how the target uses it.
1198 unsigned BitNo; // Bit # of CR6.
1199 bool InvertBit; // Invert result?
1200 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
1201 default: // Can't happen, don't crash on invalid number though.
1202 case 0: // Return the value of the EQ bit of CR6.
1203 BitNo = 0; InvertBit = false;
1205 case 1: // Return the inverted value of the EQ bit of CR6.
1206 BitNo = 0; InvertBit = true;
1208 case 2: // Return the value of the LT bit of CR6.
1209 BitNo = 2; InvertBit = false;
1211 case 3: // Return the inverted value of the LT bit of CR6.
1212 BitNo = 2; InvertBit = true;
1216 // Shift the bit into the low position.
1217 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
1218 DAG.getConstant(8-(3-BitNo), MVT::i32));
1220 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
1221 DAG.getConstant(1, MVT::i32));
1223 // If we are supposed to, toggle the bit.
1225 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
1226 DAG.getConstant(1, MVT::i32));
1230 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
1231 // Create a stack slot that is 16-byte aligned.
1232 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
1233 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
1234 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
1236 // Store the input value into Value#0 of the stack slot.
1237 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
1238 Op.getOperand(0), FIdx,DAG.getSrcValue(NULL));
1240 return DAG.getLoad(Op.getValueType(), Store, FIdx, DAG.getSrcValue(NULL));
1243 /// LowerOperation - Provide custom lowering hooks for some operations.
1245 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
1246 switch (Op.getOpcode()) {
1247 default: assert(0 && "Wasn't expecting to be able to lower this!");
1248 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
1249 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
1250 case ISD::SETCC: return LowerSETCC(Op, DAG);
1251 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
1252 case ISD::RET: return LowerRET(Op, DAG);
1254 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
1255 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
1256 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
1258 // Lower 64-bit shifts.
1259 case ISD::SHL: return LowerSHL(Op, DAG);
1260 case ISD::SRL: return LowerSRL(Op, DAG);
1261 case ISD::SRA: return LowerSRA(Op, DAG);
1263 // Vector-related lowering.
1264 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
1265 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
1266 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
1267 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
1272 //===----------------------------------------------------------------------===//
1273 // Other Lowering Code
1274 //===----------------------------------------------------------------------===//
1276 std::vector<SDOperand>
1277 PPCTargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
1279 // add beautiful description of PPC stack frame format, or at least some docs
1281 MachineFunction &MF = DAG.getMachineFunction();
1282 MachineFrameInfo *MFI = MF.getFrameInfo();
1283 MachineBasicBlock& BB = MF.front();
1284 SSARegMap *RegMap = MF.getSSARegMap();
1285 std::vector<SDOperand> ArgValues;
1287 unsigned ArgOffset = 24;
1288 unsigned GPR_remaining = 8;
1289 unsigned FPR_remaining = 13;
1290 unsigned GPR_idx = 0, FPR_idx = 0;
1291 static const unsigned GPR[] = {
1292 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1293 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1295 static const unsigned FPR[] = {
1296 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1297 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1300 // Add DAG nodes to load the arguments... On entry to a function on PPC,
1301 // the arguments start at offset 24, although they are likely to be passed
1303 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
1304 SDOperand newroot, argt;
1306 bool needsLoad = false;
1307 bool ArgLive = !I->use_empty();
1308 MVT::ValueType ObjectVT = getValueType(I->getType());
1311 default: assert(0 && "Unhandled argument type!");
1317 if (!ArgLive) break;
1318 if (GPR_remaining > 0) {
1319 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1320 MF.addLiveIn(GPR[GPR_idx], VReg);
1321 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1322 if (ObjectVT != MVT::i32) {
1323 unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
1325 argt = DAG.getNode(AssertOp, MVT::i32, argt,
1326 DAG.getValueType(ObjectVT));
1327 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, argt);
1335 if (!ArgLive) break;
1336 if (GPR_remaining > 0) {
1337 SDOperand argHi, argLo;
1338 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1339 MF.addLiveIn(GPR[GPR_idx], VReg);
1340 argHi = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1341 // If we have two or more remaining argument registers, then both halves
1342 // of the i64 can be sourced from there. Otherwise, the lower half will
1343 // have to come off the stack. This can happen when an i64 is preceded
1344 // by 28 bytes of arguments.
1345 if (GPR_remaining > 1) {
1346 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1347 MF.addLiveIn(GPR[GPR_idx+1], VReg);
1348 argLo = DAG.getCopyFromReg(argHi, VReg, MVT::i32);
1350 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
1351 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
1352 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
1353 DAG.getSrcValue(NULL));
1355 // Build the outgoing arg thingy
1356 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
1364 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
1366 if (FPR_remaining > 0) {
1372 if (FPR_remaining > 0) {
1374 if (ObjectVT == MVT::f32)
1375 VReg = RegMap->createVirtualRegister(&PPC::F4RCRegClass);
1377 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1378 MF.addLiveIn(FPR[FPR_idx], VReg);
1379 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, ObjectVT);
1388 // We need to load the argument to a virtual register if we determined above
1389 // that we ran out of physical registers of the appropriate type
1391 unsigned SubregOffset = 0;
1392 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
1393 if (ObjectVT == MVT::i16) SubregOffset = 2;
1394 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1395 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
1396 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
1397 DAG.getConstant(SubregOffset, MVT::i32));
1398 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
1399 DAG.getSrcValue(NULL));
1402 // Every 4 bytes of argument space consumes one of the GPRs available for
1403 // argument passing.
1404 if (GPR_remaining > 0) {
1405 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
1406 GPR_remaining -= delta;
1409 ArgOffset += ObjSize;
1411 DAG.setRoot(newroot.getValue(1));
1413 ArgValues.push_back(argt);
1416 // If the function takes variable number of arguments, make a frame index for
1417 // the start of the first vararg value... for expansion of llvm.va_start.
1419 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1420 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
1421 // If this function is vararg, store any remaining integer argument regs
1422 // to their spots on the stack so that they may be loaded by deferencing the
1423 // result of va_next.
1424 std::vector<SDOperand> MemOps;
1425 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
1426 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1427 MF.addLiveIn(GPR[GPR_idx], VReg);
1428 SDOperand Val = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1429 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
1430 Val, FIN, DAG.getSrcValue(NULL));
1431 MemOps.push_back(Store);
1432 // Increment the address by four for the next argument to store
1433 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
1434 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
1436 if (!MemOps.empty()) {
1437 MemOps.push_back(DAG.getRoot());
1438 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
1445 std::pair<SDOperand, SDOperand>
1446 PPCTargetLowering::LowerCallTo(SDOperand Chain,
1447 const Type *RetTy, bool isVarArg,
1448 unsigned CallingConv, bool isTailCall,
1449 SDOperand Callee, ArgListTy &Args,
1450 SelectionDAG &DAG) {
1451 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1452 // SelectExpr to use to put the arguments in the appropriate registers.
1453 std::vector<SDOperand> args_to_use;
1455 // Count how many bytes are to be pushed on the stack, including the linkage
1456 // area, and parameter passing area.
1457 unsigned NumBytes = 24;
1460 Chain = DAG.getCALLSEQ_START(Chain,
1461 DAG.getConstant(NumBytes, getPointerTy()));
1463 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1464 switch (getValueType(Args[i].second)) {
1465 default: assert(0 && "Unknown value type!");
1480 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
1481 // plus 32 bytes of argument space in case any called code gets funky on us.
1482 // (Required by ABI to support var arg)
1483 if (NumBytes < 56) NumBytes = 56;
1485 // Adjust the stack pointer for the new arguments...
1486 // These operations are automatically eliminated by the prolog/epilog pass
1487 Chain = DAG.getCALLSEQ_START(Chain,
1488 DAG.getConstant(NumBytes, getPointerTy()));
1490 // Set up a copy of the stack pointer for use loading and storing any
1491 // arguments that may not fit in the registers available for argument
1493 SDOperand StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1495 // Figure out which arguments are going to go in registers, and which in
1496 // memory. Also, if this is a vararg function, floating point operations
1497 // must be stored to our stack, and loaded into integer regs as well, if
1498 // any integer regs are available for argument passing.
1499 unsigned ArgOffset = 24;
1500 unsigned GPR_remaining = 8;
1501 unsigned FPR_remaining = 13;
1503 std::vector<SDOperand> MemOps;
1504 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1505 // PtrOff will be used to store the current argument to the stack if a
1506 // register cannot be found for it.
1507 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1508 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
1509 MVT::ValueType ArgVT = getValueType(Args[i].second);
1512 default: assert(0 && "Unexpected ValueType for argument!");
1516 // Promote the integer to 32 bits. If the input type is signed use a
1517 // sign extend, otherwise use a zero extend.
1518 if (Args[i].second->isSigned())
1519 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
1521 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
1524 if (GPR_remaining > 0) {
1525 args_to_use.push_back(Args[i].first);
1528 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1529 Args[i].first, PtrOff,
1530 DAG.getSrcValue(NULL)));
1535 // If we have one free GPR left, we can place the upper half of the i64
1536 // in it, and store the other half to the stack. If we have two or more
1537 // free GPRs, then we can pass both halves of the i64 in registers.
1538 if (GPR_remaining > 0) {
1539 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1540 Args[i].first, DAG.getConstant(1, MVT::i32));
1541 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1542 Args[i].first, DAG.getConstant(0, MVT::i32));
1543 args_to_use.push_back(Hi);
1545 if (GPR_remaining > 0) {
1546 args_to_use.push_back(Lo);
1549 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1550 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1551 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1552 Lo, PtrOff, DAG.getSrcValue(NULL)));
1555 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1556 Args[i].first, PtrOff,
1557 DAG.getSrcValue(NULL)));
1563 if (FPR_remaining > 0) {
1564 args_to_use.push_back(Args[i].first);
1567 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
1568 Args[i].first, PtrOff,
1569 DAG.getSrcValue(NULL));
1570 MemOps.push_back(Store);
1571 // Float varargs are always shadowed in available integer registers
1572 if (GPR_remaining > 0) {
1573 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1574 DAG.getSrcValue(NULL));
1575 MemOps.push_back(Load.getValue(1));
1576 args_to_use.push_back(Load);
1579 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1580 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1581 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1582 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1583 DAG.getSrcValue(NULL));
1584 MemOps.push_back(Load.getValue(1));
1585 args_to_use.push_back(Load);
1589 // If we have any FPRs remaining, we may also have GPRs remaining.
1590 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1592 if (GPR_remaining > 0) {
1593 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1596 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1597 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1602 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1603 Args[i].first, PtrOff,
1604 DAG.getSrcValue(NULL)));
1606 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
1610 if (!MemOps.empty())
1611 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
1614 std::vector<MVT::ValueType> RetVals;
1615 MVT::ValueType RetTyVT = getValueType(RetTy);
1616 MVT::ValueType ActualRetTyVT = RetTyVT;
1617 if (RetTyVT >= MVT::i1 && RetTyVT <= MVT::i16)
1618 ActualRetTyVT = MVT::i32; // Promote result to i32.
1620 if (RetTyVT == MVT::i64) {
1621 RetVals.push_back(MVT::i32);
1622 RetVals.push_back(MVT::i32);
1623 } else if (RetTyVT != MVT::isVoid) {
1624 RetVals.push_back(ActualRetTyVT);
1626 RetVals.push_back(MVT::Other);
1628 // If the callee is a GlobalAddress node (quite common, every direct call is)
1629 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1630 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1631 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i32);
1633 std::vector<SDOperand> Ops;
1634 Ops.push_back(Chain);
1635 Ops.push_back(Callee);
1636 Ops.insert(Ops.end(), args_to_use.begin(), args_to_use.end());
1637 SDOperand TheCall = DAG.getNode(PPCISD::CALL, RetVals, Ops);
1638 Chain = TheCall.getValue(TheCall.Val->getNumValues()-1);
1639 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
1640 DAG.getConstant(NumBytes, getPointerTy()));
1641 SDOperand RetVal = TheCall;
1643 // If the result is a small value, add a note so that we keep track of the
1644 // information about whether it is sign or zero extended.
1645 if (RetTyVT != ActualRetTyVT) {
1646 RetVal = DAG.getNode(RetTy->isSigned() ? ISD::AssertSext : ISD::AssertZext,
1647 MVT::i32, RetVal, DAG.getValueType(RetTyVT));
1648 RetVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, RetVal);
1649 } else if (RetTyVT == MVT::i64) {
1650 RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, RetVal, RetVal.getValue(1));
1653 return std::make_pair(RetVal, Chain);
1657 PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1658 MachineBasicBlock *BB) {
1659 assert((MI->getOpcode() == PPC::SELECT_CC_Int ||
1660 MI->getOpcode() == PPC::SELECT_CC_F4 ||
1661 MI->getOpcode() == PPC::SELECT_CC_F8 ||
1662 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
1663 "Unexpected instr type to insert");
1665 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
1666 // control-flow pattern. The incoming instruction knows the destination vreg
1667 // to set, the condition code register to branch on, the true/false values to
1668 // select between, and a branch opcode to use.
1669 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1670 ilist<MachineBasicBlock>::iterator It = BB;
1676 // cmpTY ccX, r1, r2
1678 // fallthrough --> copy0MBB
1679 MachineBasicBlock *thisMBB = BB;
1680 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1681 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1682 BuildMI(BB, MI->getOperand(4).getImmedValue(), 2)
1683 .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
1684 MachineFunction *F = BB->getParent();
1685 F->getBasicBlockList().insert(It, copy0MBB);
1686 F->getBasicBlockList().insert(It, sinkMBB);
1687 // Update machine-CFG edges by first adding all successors of the current
1688 // block to the new block which will contain the Phi node for the select.
1689 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1690 e = BB->succ_end(); i != e; ++i)
1691 sinkMBB->addSuccessor(*i);
1692 // Next, remove all successors of the current block, and add the true
1693 // and fallthrough blocks as its successors.
1694 while(!BB->succ_empty())
1695 BB->removeSuccessor(BB->succ_begin());
1696 BB->addSuccessor(copy0MBB);
1697 BB->addSuccessor(sinkMBB);
1700 // %FalseValue = ...
1701 // # fallthrough to sinkMBB
1704 // Update machine-CFG edges
1705 BB->addSuccessor(sinkMBB);
1708 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1711 BuildMI(BB, PPC::PHI, 4, MI->getOperand(0).getReg())
1712 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
1713 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1715 delete MI; // The pseudo instruction is gone now.
1719 //===----------------------------------------------------------------------===//
1720 // Target Optimization Hooks
1721 //===----------------------------------------------------------------------===//
1723 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
1724 DAGCombinerInfo &DCI) const {
1725 TargetMachine &TM = getTargetMachine();
1726 SelectionDAG &DAG = DCI.DAG;
1727 switch (N->getOpcode()) {
1729 case ISD::SINT_TO_FP:
1730 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
1731 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
1732 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
1733 // We allow the src/dst to be either f32/f64, but the intermediate
1734 // type must be i64.
1735 if (N->getOperand(0).getValueType() == MVT::i64) {
1736 SDOperand Val = N->getOperand(0).getOperand(0);
1737 if (Val.getValueType() == MVT::f32) {
1738 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
1739 DCI.AddToWorklist(Val.Val);
1742 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
1743 DCI.AddToWorklist(Val.Val);
1744 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
1745 DCI.AddToWorklist(Val.Val);
1746 if (N->getValueType(0) == MVT::f32) {
1747 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
1748 DCI.AddToWorklist(Val.Val);
1751 } else if (N->getOperand(0).getValueType() == MVT::i32) {
1752 // If the intermediate type is i32, we can avoid the load/store here
1759 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
1760 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
1761 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
1762 N->getOperand(1).getValueType() == MVT::i32) {
1763 SDOperand Val = N->getOperand(1).getOperand(0);
1764 if (Val.getValueType() == MVT::f32) {
1765 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
1766 DCI.AddToWorklist(Val.Val);
1768 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
1769 DCI.AddToWorklist(Val.Val);
1771 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
1772 N->getOperand(2), N->getOperand(3));
1773 DCI.AddToWorklist(Val.Val);
1777 case PPCISD::VCMP: {
1778 // If a VCMPo node already exists with exactly the same operands as this
1779 // node, use its result instead of this node (VCMPo computes both a CR6 and
1780 // a normal output).
1782 if (!N->getOperand(0).hasOneUse() &&
1783 !N->getOperand(1).hasOneUse() &&
1784 !N->getOperand(2).hasOneUse()) {
1786 // Scan all of the users of the LHS, looking for VCMPo's that match.
1787 SDNode *VCMPoNode = 0;
1789 SDNode *LHSN = N->getOperand(0).Val;
1790 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
1792 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
1793 (*UI)->getOperand(1) == N->getOperand(1) &&
1794 (*UI)->getOperand(2) == N->getOperand(2) &&
1795 (*UI)->getOperand(0) == N->getOperand(0)) {
1800 // If there are non-zero uses of the flag value, use the VCMPo node!
1801 if (VCMPoNode && !VCMPoNode->hasNUsesOfValue(0, 1))
1802 return SDOperand(VCMPoNode, 0);
1811 //===----------------------------------------------------------------------===//
1812 // Inline Assembly Support
1813 //===----------------------------------------------------------------------===//
1815 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1817 uint64_t &KnownZero,
1819 unsigned Depth) const {
1822 switch (Op.getOpcode()) {
1824 case ISD::INTRINSIC_WO_CHAIN: {
1825 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
1827 case Intrinsic::ppc_altivec_vcmpbfp_p:
1828 case Intrinsic::ppc_altivec_vcmpeqfp_p:
1829 case Intrinsic::ppc_altivec_vcmpequb_p:
1830 case Intrinsic::ppc_altivec_vcmpequh_p:
1831 case Intrinsic::ppc_altivec_vcmpequw_p:
1832 case Intrinsic::ppc_altivec_vcmpgefp_p:
1833 case Intrinsic::ppc_altivec_vcmpgtfp_p:
1834 case Intrinsic::ppc_altivec_vcmpgtsb_p:
1835 case Intrinsic::ppc_altivec_vcmpgtsh_p:
1836 case Intrinsic::ppc_altivec_vcmpgtsw_p:
1837 case Intrinsic::ppc_altivec_vcmpgtub_p:
1838 case Intrinsic::ppc_altivec_vcmpgtuh_p:
1839 case Intrinsic::ppc_altivec_vcmpgtuw_p:
1840 KnownZero = ~1U; // All bits but the low one are known to be zero.
1848 /// getConstraintType - Given a constraint letter, return the type of
1849 /// constraint it is for this target.
1850 PPCTargetLowering::ConstraintType
1851 PPCTargetLowering::getConstraintType(char ConstraintLetter) const {
1852 switch (ConstraintLetter) {
1859 return C_RegisterClass;
1861 return TargetLowering::getConstraintType(ConstraintLetter);
1865 std::vector<unsigned> PPCTargetLowering::
1866 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1867 MVT::ValueType VT) const {
1868 if (Constraint.size() == 1) {
1869 switch (Constraint[0]) { // GCC RS6000 Constraint Letters
1870 default: break; // Unknown constriant letter
1872 return make_vector<unsigned>(/*no R0*/ PPC::R1 , PPC::R2 , PPC::R3 ,
1873 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
1874 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
1875 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
1876 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
1877 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
1878 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
1879 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
1882 return make_vector<unsigned>(PPC::R0 , PPC::R1 , PPC::R2 , PPC::R3 ,
1883 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
1884 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
1885 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
1886 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
1887 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
1888 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
1889 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
1892 return make_vector<unsigned>(PPC::F0 , PPC::F1 , PPC::F2 , PPC::F3 ,
1893 PPC::F4 , PPC::F5 , PPC::F6 , PPC::F7 ,
1894 PPC::F8 , PPC::F9 , PPC::F10, PPC::F11,
1895 PPC::F12, PPC::F13, PPC::F14, PPC::F15,
1896 PPC::F16, PPC::F17, PPC::F18, PPC::F19,
1897 PPC::F20, PPC::F21, PPC::F22, PPC::F23,
1898 PPC::F24, PPC::F25, PPC::F26, PPC::F27,
1899 PPC::F28, PPC::F29, PPC::F30, PPC::F31,
1902 return make_vector<unsigned>(PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 ,
1903 PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 ,
1904 PPC::V8 , PPC::V9 , PPC::V10, PPC::V11,
1905 PPC::V12, PPC::V13, PPC::V14, PPC::V15,
1906 PPC::V16, PPC::V17, PPC::V18, PPC::V19,
1907 PPC::V20, PPC::V21, PPC::V22, PPC::V23,
1908 PPC::V24, PPC::V25, PPC::V26, PPC::V27,
1909 PPC::V28, PPC::V29, PPC::V30, PPC::V31,
1912 return make_vector<unsigned>(PPC::CR0, PPC::CR1, PPC::CR2, PPC::CR3,
1913 PPC::CR4, PPC::CR5, PPC::CR6, PPC::CR7,
1918 return std::vector<unsigned>();
1921 // isOperandValidForConstraint
1922 bool PPCTargetLowering::
1923 isOperandValidForConstraint(SDOperand Op, char Letter) {
1934 if (!isa<ConstantSDNode>(Op)) return false; // Must be an immediate.
1935 unsigned Value = cast<ConstantSDNode>(Op)->getValue();
1937 default: assert(0 && "Unknown constraint letter!");
1938 case 'I': // "I" is a signed 16-bit constant.
1939 return (short)Value == (int)Value;
1940 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
1941 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
1942 return (short)Value == 0;
1943 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
1944 return (Value >> 16) == 0;
1945 case 'M': // "M" is a constant that is greater than 31.
1947 case 'N': // "N" is a positive constant that is an exact power of two.
1948 return (int)Value > 0 && isPowerOf2_32(Value);
1949 case 'O': // "O" is the constant zero.
1951 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
1952 return (short)-Value == (int)-Value;
1958 // Handle standard constraint letters.
1959 return TargetLowering::isOperandValidForConstraint(Op, Letter);
1962 /// isLegalAddressImmediate - Return true if the integer value can be used
1963 /// as the offset of the target addressing mode.
1964 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V) const {
1965 // PPC allows a sign-extended 16-bit immediate field.
1966 return (V > -(1 << 16) && V < (1 << 16)-1);