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);
91 // PowerPC wants to turn select_cc of FP into fsel when possible.
92 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
93 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
95 // PowerPC wants to optimize integer setcc a bit
96 setOperationAction(ISD::SETCC, MVT::i32, Custom);
98 // PowerPC does not have BRCOND which requires SetCC
99 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
101 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
102 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
104 // PowerPC does not have [U|S]INT_TO_FP
105 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
106 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
108 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
109 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
111 // PowerPC does not have truncstore for i1.
112 setOperationAction(ISD::TRUNCSTORE, MVT::i1, Promote);
114 // Support label based line numbers.
115 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
116 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
117 // FIXME - use subtarget debug flags
118 if (!TM.getSubtarget<PPCSubtarget>().isDarwin())
119 setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
121 // We want to legalize GlobalAddress and ConstantPool nodes into the
122 // appropriate instructions to materialize the address.
123 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
124 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
126 // RET must be custom lowered, to meet ABI requirements
127 setOperationAction(ISD::RET , MVT::Other, Custom);
129 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
130 setOperationAction(ISD::VASTART , MVT::Other, Custom);
132 // Use the default implementation.
133 setOperationAction(ISD::VAARG , MVT::Other, Expand);
134 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
135 setOperationAction(ISD::VAEND , MVT::Other, Expand);
136 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
137 setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
138 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
140 // We want to custom lower some of our intrinsics.
141 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
143 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
144 // They also have instructions for converting between i64 and fp.
145 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
146 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
148 // FIXME: disable this lowered code. This generates 64-bit register values,
149 // and we don't model the fact that the top part is clobbered by calls. We
150 // need to flag these together so that the value isn't live across a call.
151 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
153 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
154 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
156 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
157 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
160 if (TM.getSubtarget<PPCSubtarget>().has64BitRegs()) {
161 // 64 bit PowerPC implementations can support i64 types directly
162 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
163 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
164 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
166 // 32 bit PowerPC wants to expand i64 shifts itself.
167 setOperationAction(ISD::SHL, MVT::i64, Custom);
168 setOperationAction(ISD::SRL, MVT::i64, Custom);
169 setOperationAction(ISD::SRA, MVT::i64, Custom);
172 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
173 // First set operation action for all vector types to expand. Then we
174 // will selectively turn on ones that can be effectively codegen'd.
175 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
176 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
177 // add/sub/and/or/xor are legal for all supported vector VT's.
178 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
179 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
180 setOperationAction(ISD::AND , (MVT::ValueType)VT, Legal);
181 setOperationAction(ISD::OR , (MVT::ValueType)VT, Legal);
182 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Legal);
184 // We promote all shuffles to v16i8.
185 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
186 AddPromotedToType(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
188 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
189 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
190 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
191 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
192 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
193 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
194 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
195 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
197 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
200 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
201 // with merges, splats, etc.
202 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
204 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
205 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
206 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
207 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
209 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
211 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
212 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
214 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
215 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
216 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
217 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
220 setSetCCResultContents(ZeroOrOneSetCCResult);
221 setStackPointerRegisterToSaveRestore(PPC::R1);
223 // We have target-specific dag combine patterns for the following nodes:
224 setTargetDAGCombine(ISD::SINT_TO_FP);
225 setTargetDAGCombine(ISD::STORE);
227 computeRegisterProperties();
230 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
233 case PPCISD::FSEL: return "PPCISD::FSEL";
234 case PPCISD::FCFID: return "PPCISD::FCFID";
235 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
236 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
237 case PPCISD::STFIWX: return "PPCISD::STFIWX";
238 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
239 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
240 case PPCISD::VPERM: return "PPCISD::VPERM";
241 case PPCISD::Hi: return "PPCISD::Hi";
242 case PPCISD::Lo: return "PPCISD::Lo";
243 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
244 case PPCISD::SRL: return "PPCISD::SRL";
245 case PPCISD::SRA: return "PPCISD::SRA";
246 case PPCISD::SHL: return "PPCISD::SHL";
247 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
248 case PPCISD::STD_32: return "PPCISD::STD_32";
249 case PPCISD::CALL: return "PPCISD::CALL";
250 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
251 case PPCISD::MFCR: return "PPCISD::MFCR";
252 case PPCISD::VCMP: return "PPCISD::VCMP";
253 case PPCISD::VCMPo: return "PPCISD::VCMPo";
257 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
258 static bool isFloatingPointZero(SDOperand Op) {
259 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
260 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
261 else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
262 // Maybe this has already been legalized into the constant pool?
263 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
264 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
265 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
270 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
271 /// true if Op is undef or if it matches the specified value.
272 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
273 return Op.getOpcode() == ISD::UNDEF ||
274 cast<ConstantSDNode>(Op)->getValue() == Val;
277 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
278 /// VPKUHUM instruction.
279 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
281 for (unsigned i = 0; i != 16; ++i)
282 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
285 for (unsigned i = 0; i != 8; ++i)
286 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
287 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
293 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
294 /// VPKUWUM instruction.
295 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
297 for (unsigned i = 0; i != 16; i += 2)
298 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
299 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
302 for (unsigned i = 0; i != 8; i += 2)
303 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
304 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
305 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
306 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
312 /// isVMerge - Common function, used to match vmrg* shuffles.
314 static bool isVMerge(SDNode *N, unsigned UnitSize,
315 unsigned LHSStart, unsigned RHSStart) {
316 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
317 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
318 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
319 "Unsupported merge size!");
321 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
322 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
323 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
324 LHSStart+j+i*UnitSize) ||
325 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
326 RHSStart+j+i*UnitSize))
332 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
333 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
334 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
336 return isVMerge(N, UnitSize, 8, 24);
337 return isVMerge(N, UnitSize, 8, 8);
340 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
341 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
342 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
344 return isVMerge(N, UnitSize, 0, 16);
345 return isVMerge(N, UnitSize, 0, 0);
349 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
350 /// amount, otherwise return -1.
351 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
352 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
353 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
354 // Find the first non-undef value in the shuffle mask.
356 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
359 if (i == 16) return -1; // all undef.
361 // Otherwise, check to see if the rest of the elements are consequtively
362 // numbered from this value.
363 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
364 if (ShiftAmt < i) return -1;
368 // Check the rest of the elements to see if they are consequtive.
369 for (++i; i != 16; ++i)
370 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
373 // Check the rest of the elements to see if they are consequtive.
374 for (++i; i != 16; ++i)
375 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
382 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
383 /// specifies a splat of a single element that is suitable for input to
384 /// VSPLTB/VSPLTH/VSPLTW.
385 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
386 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
387 N->getNumOperands() == 16 &&
388 (EltSize == 1 || EltSize == 2 || EltSize == 4));
390 // This is a splat operation if each element of the permute is the same, and
391 // if the value doesn't reference the second vector.
392 unsigned ElementBase = 0;
393 SDOperand Elt = N->getOperand(0);
394 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
395 ElementBase = EltV->getValue();
397 return false; // FIXME: Handle UNDEF elements too!
399 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
402 // Check that they are consequtive.
403 for (unsigned i = 1; i != EltSize; ++i) {
404 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
405 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
409 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
410 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
411 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
412 "Invalid VECTOR_SHUFFLE mask!");
413 for (unsigned j = 0; j != EltSize; ++j)
414 if (N->getOperand(i+j) != N->getOperand(j))
421 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
422 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
423 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
424 assert(isSplatShuffleMask(N, EltSize));
425 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
428 /// get_VSPLI_elt - If this is a build_vector of constants which can be formed
429 /// by using a vspltis[bhw] instruction of the specified element size, return
430 /// the constant being splatted. The ByteSize field indicates the number of
431 /// bytes of each element [124] -> [bhw].
432 SDOperand PPC::get_VSPLI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
433 SDOperand OpVal(0, 0);
434 // Check to see if this buildvec has a single non-undef value in its elements.
435 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
436 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
438 OpVal = N->getOperand(i);
439 else if (OpVal != N->getOperand(i))
443 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
445 unsigned ValSizeInBytes = 0;
447 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
448 Value = CN->getValue();
449 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
450 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
451 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
452 Value = FloatToBits(CN->getValue());
456 // If the splat value is larger than the element value, then we can never do
457 // this splat. The only case that we could fit the replicated bits into our
458 // immediate field for would be zero, and we prefer to use vxor for it.
459 if (ValSizeInBytes < ByteSize) return SDOperand();
461 // If the element value is larger than the splat value, cut it in half and
462 // check to see if the two halves are equal. Continue doing this until we
463 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
464 while (ValSizeInBytes > ByteSize) {
465 ValSizeInBytes >>= 1;
467 // If the top half equals the bottom half, we're still ok.
468 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
469 (Value & ((1 << (8*ValSizeInBytes))-1)))
473 // Properly sign extend the value.
474 int ShAmt = (4-ByteSize)*8;
475 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
477 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
478 if (MaskVal == 0) return SDOperand();
480 // Finally, if this value fits in a 5 bit sext field, return it
481 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
482 return DAG.getTargetConstant(MaskVal, MVT::i32);
487 /// LowerOperation - Provide custom lowering hooks for some operations.
489 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
490 switch (Op.getOpcode()) {
491 default: assert(0 && "Wasn't expecting to be able to lower this!");
492 case ISD::FP_TO_SINT: {
493 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
494 SDOperand Src = Op.getOperand(0);
495 if (Src.getValueType() == MVT::f32)
496 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
499 switch (Op.getValueType()) {
500 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
502 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
505 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
509 // Convert the FP value to an int value through memory.
510 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
511 if (Op.getValueType() == MVT::i32)
512 Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
515 case ISD::SINT_TO_FP:
516 if (Op.getOperand(0).getValueType() == MVT::i64) {
517 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
518 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
519 if (Op.getValueType() == MVT::f32)
520 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
523 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
524 "Unhandled SINT_TO_FP type in custom expander!");
525 // Since we only generate this in 64-bit mode, we can take advantage of
526 // 64-bit registers. In particular, sign extend the input value into the
527 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
528 // then lfd it and fcfid it.
529 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
530 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
531 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
533 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
536 // STD the extended value into the stack slot.
537 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
538 DAG.getEntryNode(), Ext64, FIdx,
539 DAG.getSrcValue(NULL));
540 // Load the value as a double.
541 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
543 // FCFID it and return it.
544 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
545 if (Op.getValueType() == MVT::f32)
546 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
551 case ISD::SELECT_CC: {
552 // Turn FP only select_cc's into fsel instructions.
553 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
554 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
557 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
559 // Cannot handle SETEQ/SETNE.
560 if (CC == ISD::SETEQ || CC == ISD::SETNE) break;
562 MVT::ValueType ResVT = Op.getValueType();
563 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
564 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
565 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
567 // If the RHS of the comparison is a 0.0, we don't need to do the
568 // subtraction at all.
569 if (isFloatingPointZero(RHS))
571 default: break; // SETUO etc aren't handled by fsel.
574 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
577 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
578 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
579 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
582 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
585 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
586 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
587 return DAG.getNode(PPCISD::FSEL, ResVT,
588 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
593 default: break; // SETUO etc aren't handled by fsel.
596 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
597 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
598 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
599 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
602 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
603 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
604 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
605 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
608 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
609 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
610 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
611 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
614 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
615 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
616 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
617 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
622 assert(Op.getValueType() == MVT::i64 &&
623 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
624 // The generic code does a fine job expanding shift by a constant.
625 if (isa<ConstantSDNode>(Op.getOperand(1))) break;
627 // Otherwise, expand into a bunch of logical ops. Note that these ops
628 // depend on the PPC behavior for oversized shift amounts.
629 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
630 DAG.getConstant(0, MVT::i32));
631 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
632 DAG.getConstant(1, MVT::i32));
633 SDOperand Amt = Op.getOperand(1);
635 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
636 DAG.getConstant(32, MVT::i32), Amt);
637 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
638 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
639 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
640 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
641 DAG.getConstant(-32U, MVT::i32));
642 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
643 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
644 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
645 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
648 assert(Op.getValueType() == MVT::i64 &&
649 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
650 // The generic code does a fine job expanding shift by a constant.
651 if (isa<ConstantSDNode>(Op.getOperand(1))) break;
653 // Otherwise, expand into a bunch of logical ops. Note that these ops
654 // depend on the PPC behavior for oversized shift amounts.
655 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
656 DAG.getConstant(0, MVT::i32));
657 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
658 DAG.getConstant(1, MVT::i32));
659 SDOperand Amt = Op.getOperand(1);
661 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
662 DAG.getConstant(32, MVT::i32), Amt);
663 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
664 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
665 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
666 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
667 DAG.getConstant(-32U, MVT::i32));
668 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
669 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
670 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
671 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
674 assert(Op.getValueType() == MVT::i64 &&
675 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
676 // The generic code does a fine job expanding shift by a constant.
677 if (isa<ConstantSDNode>(Op.getOperand(1))) break;
679 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
680 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
681 DAG.getConstant(0, MVT::i32));
682 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
683 DAG.getConstant(1, MVT::i32));
684 SDOperand Amt = Op.getOperand(1);
686 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
687 DAG.getConstant(32, MVT::i32), Amt);
688 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
689 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
690 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
691 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
692 DAG.getConstant(-32U, MVT::i32));
693 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
694 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
695 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
696 Tmp4, Tmp6, ISD::SETLE);
697 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
699 case ISD::ConstantPool: {
700 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
701 Constant *C = CP->get();
702 SDOperand CPI = DAG.getTargetConstantPool(C, MVT::i32, CP->getAlignment());
703 SDOperand Zero = DAG.getConstant(0, MVT::i32);
705 if (getTargetMachine().getRelocationModel() == Reloc::Static) {
706 // Generate non-pic code that has direct accesses to the constant pool.
707 // The address of the global is just (hi(&g)+lo(&g)).
708 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
709 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
710 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
713 // Only lower ConstantPool on Darwin.
714 if (!getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin()) break;
715 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
716 if (getTargetMachine().getRelocationModel() == Reloc::PIC) {
717 // With PIC, the first instruction is actually "GR+hi(&G)".
718 Hi = DAG.getNode(ISD::ADD, MVT::i32,
719 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
722 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
723 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
726 case ISD::GlobalAddress: {
727 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
728 GlobalValue *GV = GSDN->getGlobal();
729 SDOperand GA = DAG.getTargetGlobalAddress(GV, MVT::i32, GSDN->getOffset());
730 SDOperand Zero = DAG.getConstant(0, MVT::i32);
732 if (getTargetMachine().getRelocationModel() == Reloc::Static) {
733 // Generate non-pic code that has direct accesses to globals.
734 // The address of the global is just (hi(&g)+lo(&g)).
735 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
736 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
737 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
740 // Only lower GlobalAddress on Darwin.
741 if (!getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin()) break;
743 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
744 if (getTargetMachine().getRelocationModel() == Reloc::PIC) {
745 // With PIC, the first instruction is actually "GR+hi(&G)".
746 Hi = DAG.getNode(ISD::ADD, MVT::i32,
747 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
750 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
751 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
753 if (!GV->hasWeakLinkage() && !GV->hasLinkOnceLinkage() &&
754 (!GV->isExternal() || GV->hasNotBeenReadFromBytecode()))
757 // If the global is weak or external, we have to go through the lazy
759 return DAG.getLoad(MVT::i32, DAG.getEntryNode(), Lo, DAG.getSrcValue(0));
762 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
764 // If we're comparing for equality to zero, expose the fact that this is
765 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
766 // fold the new nodes.
767 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
768 if (C->isNullValue() && CC == ISD::SETEQ) {
769 MVT::ValueType VT = Op.getOperand(0).getValueType();
770 SDOperand Zext = Op.getOperand(0);
773 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
775 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
776 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
777 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
778 DAG.getConstant(Log2b, getShiftAmountTy()));
779 return DAG.getNode(ISD::TRUNCATE, getSetCCResultTy(), Scc);
781 // Leave comparisons against 0 and -1 alone for now, since they're usually
782 // optimized. FIXME: revisit this when we can custom lower all setcc
784 if (C->isAllOnesValue() || C->isNullValue())
788 // If we have an integer seteq/setne, turn it into a compare against zero
789 // by subtracting the rhs from the lhs, which is faster than setting a
790 // condition register, reading it back out, and masking the correct bit.
791 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
792 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
793 MVT::ValueType VT = Op.getValueType();
794 SDOperand Sub = DAG.getNode(ISD::SUB, LHSVT, Op.getOperand(0),
796 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
801 // vastart just stores the address of the VarArgsFrameIndex slot into the
802 // memory location argument.
803 // FIXME: Replace MVT::i32 with PointerTy
804 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
805 return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
806 Op.getOperand(1), Op.getOperand(2));
811 switch(Op.getNumOperands()) {
813 assert(0 && "Do not know how to return this many arguments!");
816 return SDOperand(); // ret void is legal
818 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
819 unsigned ArgReg = MVT::isInteger(ArgVT) ? PPC::R3 : PPC::F1;
820 Copy = DAG.getCopyToReg(Op.getOperand(0), ArgReg, Op.getOperand(1),
825 Copy = DAG.getCopyToReg(Op.getOperand(0), PPC::R3, Op.getOperand(2),
827 Copy = DAG.getCopyToReg(Copy, PPC::R4, Op.getOperand(1),Copy.getValue(1));
830 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
832 case ISD::SCALAR_TO_VECTOR: {
833 // Create a stack slot that is 16-byte aligned.
834 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
835 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
836 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
838 // Store the input value into Value#0 of the stack slot.
839 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
840 Op.getOperand(0), FIdx,DAG.getSrcValue(NULL));
842 return DAG.getLoad(Op.getValueType(), Store, FIdx, DAG.getSrcValue(NULL));
844 case ISD::BUILD_VECTOR:
845 // If this is a case we can't handle, return null and let the default
846 // expansion code take care of it. If we CAN select this case, return Op.
848 // See if this is all zeros.
849 // FIXME: We should handle splat(-0.0), and other cases here.
850 if (ISD::isBuildVectorAllZeros(Op.Val))
853 if (PPC::get_VSPLI_elt(Op.Val, 1, DAG).Val || // vspltisb
854 PPC::get_VSPLI_elt(Op.Val, 2, DAG).Val || // vspltish
855 PPC::get_VSPLI_elt(Op.Val, 4, DAG).Val) // vspltisw
860 case ISD::VECTOR_SHUFFLE: {
861 SDOperand V1 = Op.getOperand(0);
862 SDOperand V2 = Op.getOperand(1);
863 SDOperand PermMask = Op.getOperand(2);
865 // Cases that are handled by instructions that take permute immediates
866 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
867 // selected by the instruction selector.
868 if (V2.getOpcode() == ISD::UNDEF) {
869 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
870 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
871 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
872 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
873 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
874 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
875 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
876 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
877 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
878 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
879 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
880 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
885 // Altivec has a variety of "shuffle immediates" that take two vector inputs
886 // and produce a fixed permutation. If any of these match, do not lower to
888 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
889 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
890 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
891 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
892 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
893 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
894 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
895 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
896 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
899 // TODO: Handle more cases, and also handle cases that are cheaper to do as
900 // multiple such instructions than as a constant pool load/vperm pair.
902 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
903 // vector that will get spilled to the constant pool.
904 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
906 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
907 // that it is in input element units, not in bytes. Convert now.
908 MVT::ValueType EltVT = MVT::getVectorBaseType(V1.getValueType());
909 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
911 std::vector<SDOperand> ResultMask;
912 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
913 unsigned SrcElt =cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
915 for (unsigned j = 0; j != BytesPerElement; ++j)
916 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
920 SDOperand VPermMask =DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, ResultMask);
921 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
923 case ISD::INTRINSIC_WO_CHAIN: {
924 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
926 // If this is a lowered altivec predicate compare, CompareOpc is set to the
927 // opcode number of the comparison.
931 default: return SDOperand(); // Don't custom lower most intrinsics.
932 // Comparison predicates.
933 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
934 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
935 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
936 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
937 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
938 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
939 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
940 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
941 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
942 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
943 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
944 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
945 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
947 // Normal Comparisons.
948 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
949 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
950 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
951 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
952 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
953 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
954 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
955 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
956 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
957 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
958 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
959 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
960 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
963 assert(CompareOpc>0 && "We only lower altivec predicate compares so far!");
965 // If this is a non-dot comparison, make the VCMP node.
967 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
968 Op.getOperand(1), Op.getOperand(2),
969 DAG.getConstant(CompareOpc, MVT::i32));
970 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
973 // Create the PPCISD altivec 'dot' comparison node.
974 std::vector<SDOperand> Ops;
975 std::vector<MVT::ValueType> VTs;
976 Ops.push_back(Op.getOperand(2)); // LHS
977 Ops.push_back(Op.getOperand(3)); // RHS
978 Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
979 VTs.push_back(Op.getOperand(2).getValueType());
980 VTs.push_back(MVT::Flag);
981 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
983 // Now that we have the comparison, emit a copy from the CR to a GPR.
984 // This is flagged to the above dot comparison.
985 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
986 DAG.getRegister(PPC::CR6, MVT::i32),
987 CompNode.getValue(1));
989 // Unpack the result based on how the target uses it.
990 unsigned BitNo; // Bit # of CR6.
991 bool InvertBit; // Invert result?
992 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
993 default: // Can't happen, don't crash on invalid number though.
994 case 0: // Return the value of the EQ bit of CR6.
995 BitNo = 0; InvertBit = false;
997 case 1: // Return the inverted value of the EQ bit of CR6.
998 BitNo = 0; InvertBit = true;
1000 case 2: // Return the value of the LT bit of CR6.
1001 BitNo = 2; InvertBit = false;
1003 case 3: // Return the inverted value of the LT bit of CR6.
1004 BitNo = 2; InvertBit = true;
1008 // Shift the bit into the low position.
1009 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
1010 DAG.getConstant(8-(3-BitNo), MVT::i32));
1012 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
1013 DAG.getConstant(1, MVT::i32));
1015 // If we are supposed to, toggle the bit.
1017 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
1018 DAG.getConstant(1, MVT::i32));
1025 std::vector<SDOperand>
1026 PPCTargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
1028 // add beautiful description of PPC stack frame format, or at least some docs
1030 MachineFunction &MF = DAG.getMachineFunction();
1031 MachineFrameInfo *MFI = MF.getFrameInfo();
1032 MachineBasicBlock& BB = MF.front();
1033 SSARegMap *RegMap = MF.getSSARegMap();
1034 std::vector<SDOperand> ArgValues;
1036 unsigned ArgOffset = 24;
1037 unsigned GPR_remaining = 8;
1038 unsigned FPR_remaining = 13;
1039 unsigned GPR_idx = 0, FPR_idx = 0;
1040 static const unsigned GPR[] = {
1041 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1042 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1044 static const unsigned FPR[] = {
1045 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1046 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1049 // Add DAG nodes to load the arguments... On entry to a function on PPC,
1050 // the arguments start at offset 24, although they are likely to be passed
1052 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
1053 SDOperand newroot, argt;
1055 bool needsLoad = false;
1056 bool ArgLive = !I->use_empty();
1057 MVT::ValueType ObjectVT = getValueType(I->getType());
1060 default: assert(0 && "Unhandled argument type!");
1066 if (!ArgLive) break;
1067 if (GPR_remaining > 0) {
1068 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1069 MF.addLiveIn(GPR[GPR_idx], VReg);
1070 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1071 if (ObjectVT != MVT::i32) {
1072 unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
1074 argt = DAG.getNode(AssertOp, MVT::i32, argt,
1075 DAG.getValueType(ObjectVT));
1076 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, argt);
1084 if (!ArgLive) break;
1085 if (GPR_remaining > 0) {
1086 SDOperand argHi, argLo;
1087 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1088 MF.addLiveIn(GPR[GPR_idx], VReg);
1089 argHi = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1090 // If we have two or more remaining argument registers, then both halves
1091 // of the i64 can be sourced from there. Otherwise, the lower half will
1092 // have to come off the stack. This can happen when an i64 is preceded
1093 // by 28 bytes of arguments.
1094 if (GPR_remaining > 1) {
1095 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1096 MF.addLiveIn(GPR[GPR_idx+1], VReg);
1097 argLo = DAG.getCopyFromReg(argHi, VReg, MVT::i32);
1099 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
1100 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
1101 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
1102 DAG.getSrcValue(NULL));
1104 // Build the outgoing arg thingy
1105 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
1113 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
1115 if (FPR_remaining > 0) {
1121 if (FPR_remaining > 0) {
1123 if (ObjectVT == MVT::f32)
1124 VReg = RegMap->createVirtualRegister(&PPC::F4RCRegClass);
1126 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1127 MF.addLiveIn(FPR[FPR_idx], VReg);
1128 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, ObjectVT);
1137 // We need to load the argument to a virtual register if we determined above
1138 // that we ran out of physical registers of the appropriate type
1140 unsigned SubregOffset = 0;
1141 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
1142 if (ObjectVT == MVT::i16) SubregOffset = 2;
1143 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
1144 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
1145 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
1146 DAG.getConstant(SubregOffset, MVT::i32));
1147 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
1148 DAG.getSrcValue(NULL));
1151 // Every 4 bytes of argument space consumes one of the GPRs available for
1152 // argument passing.
1153 if (GPR_remaining > 0) {
1154 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
1155 GPR_remaining -= delta;
1158 ArgOffset += ObjSize;
1160 DAG.setRoot(newroot.getValue(1));
1162 ArgValues.push_back(argt);
1165 // If the function takes variable number of arguments, make a frame index for
1166 // the start of the first vararg value... for expansion of llvm.va_start.
1168 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1169 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
1170 // If this function is vararg, store any remaining integer argument regs
1171 // to their spots on the stack so that they may be loaded by deferencing the
1172 // result of va_next.
1173 std::vector<SDOperand> MemOps;
1174 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
1175 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1176 MF.addLiveIn(GPR[GPR_idx], VReg);
1177 SDOperand Val = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1178 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
1179 Val, FIN, DAG.getSrcValue(NULL));
1180 MemOps.push_back(Store);
1181 // Increment the address by four for the next argument to store
1182 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
1183 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
1185 if (!MemOps.empty()) {
1186 MemOps.push_back(DAG.getRoot());
1187 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
1191 // Finally, inform the code generator which regs we return values in.
1192 switch (getValueType(F.getReturnType())) {
1193 default: assert(0 && "Unknown type!");
1194 case MVT::isVoid: break;
1199 MF.addLiveOut(PPC::R3);
1202 MF.addLiveOut(PPC::R3);
1203 MF.addLiveOut(PPC::R4);
1207 MF.addLiveOut(PPC::F1);
1214 std::pair<SDOperand, SDOperand>
1215 PPCTargetLowering::LowerCallTo(SDOperand Chain,
1216 const Type *RetTy, bool isVarArg,
1217 unsigned CallingConv, bool isTailCall,
1218 SDOperand Callee, ArgListTy &Args,
1219 SelectionDAG &DAG) {
1220 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1221 // SelectExpr to use to put the arguments in the appropriate registers.
1222 std::vector<SDOperand> args_to_use;
1224 // Count how many bytes are to be pushed on the stack, including the linkage
1225 // area, and parameter passing area.
1226 unsigned NumBytes = 24;
1229 Chain = DAG.getCALLSEQ_START(Chain,
1230 DAG.getConstant(NumBytes, getPointerTy()));
1232 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1233 switch (getValueType(Args[i].second)) {
1234 default: assert(0 && "Unknown value type!");
1249 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
1250 // plus 32 bytes of argument space in case any called code gets funky on us.
1251 // (Required by ABI to support var arg)
1252 if (NumBytes < 56) NumBytes = 56;
1254 // Adjust the stack pointer for the new arguments...
1255 // These operations are automatically eliminated by the prolog/epilog pass
1256 Chain = DAG.getCALLSEQ_START(Chain,
1257 DAG.getConstant(NumBytes, getPointerTy()));
1259 // Set up a copy of the stack pointer for use loading and storing any
1260 // arguments that may not fit in the registers available for argument
1262 SDOperand StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1264 // Figure out which arguments are going to go in registers, and which in
1265 // memory. Also, if this is a vararg function, floating point operations
1266 // must be stored to our stack, and loaded into integer regs as well, if
1267 // any integer regs are available for argument passing.
1268 unsigned ArgOffset = 24;
1269 unsigned GPR_remaining = 8;
1270 unsigned FPR_remaining = 13;
1272 std::vector<SDOperand> MemOps;
1273 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1274 // PtrOff will be used to store the current argument to the stack if a
1275 // register cannot be found for it.
1276 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1277 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
1278 MVT::ValueType ArgVT = getValueType(Args[i].second);
1281 default: assert(0 && "Unexpected ValueType for argument!");
1285 // Promote the integer to 32 bits. If the input type is signed use a
1286 // sign extend, otherwise use a zero extend.
1287 if (Args[i].second->isSigned())
1288 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
1290 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
1293 if (GPR_remaining > 0) {
1294 args_to_use.push_back(Args[i].first);
1297 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1298 Args[i].first, PtrOff,
1299 DAG.getSrcValue(NULL)));
1304 // If we have one free GPR left, we can place the upper half of the i64
1305 // in it, and store the other half to the stack. If we have two or more
1306 // free GPRs, then we can pass both halves of the i64 in registers.
1307 if (GPR_remaining > 0) {
1308 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1309 Args[i].first, DAG.getConstant(1, MVT::i32));
1310 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1311 Args[i].first, DAG.getConstant(0, MVT::i32));
1312 args_to_use.push_back(Hi);
1314 if (GPR_remaining > 0) {
1315 args_to_use.push_back(Lo);
1318 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1319 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1320 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1321 Lo, PtrOff, DAG.getSrcValue(NULL)));
1324 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1325 Args[i].first, PtrOff,
1326 DAG.getSrcValue(NULL)));
1332 if (FPR_remaining > 0) {
1333 args_to_use.push_back(Args[i].first);
1336 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
1337 Args[i].first, PtrOff,
1338 DAG.getSrcValue(NULL));
1339 MemOps.push_back(Store);
1340 // Float varargs are always shadowed in available integer registers
1341 if (GPR_remaining > 0) {
1342 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1343 DAG.getSrcValue(NULL));
1344 MemOps.push_back(Load.getValue(1));
1345 args_to_use.push_back(Load);
1348 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1349 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1350 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1351 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1352 DAG.getSrcValue(NULL));
1353 MemOps.push_back(Load.getValue(1));
1354 args_to_use.push_back(Load);
1358 // If we have any FPRs remaining, we may also have GPRs remaining.
1359 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1361 if (GPR_remaining > 0) {
1362 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1365 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1366 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1371 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1372 Args[i].first, PtrOff,
1373 DAG.getSrcValue(NULL)));
1375 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
1379 if (!MemOps.empty())
1380 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
1383 std::vector<MVT::ValueType> RetVals;
1384 MVT::ValueType RetTyVT = getValueType(RetTy);
1385 MVT::ValueType ActualRetTyVT = RetTyVT;
1386 if (RetTyVT >= MVT::i1 && RetTyVT <= MVT::i16)
1387 ActualRetTyVT = MVT::i32; // Promote result to i32.
1389 if (RetTyVT == MVT::i64) {
1390 RetVals.push_back(MVT::i32);
1391 RetVals.push_back(MVT::i32);
1392 } else if (RetTyVT != MVT::isVoid) {
1393 RetVals.push_back(ActualRetTyVT);
1395 RetVals.push_back(MVT::Other);
1397 // If the callee is a GlobalAddress node (quite common, every direct call is)
1398 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1399 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1400 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i32);
1402 std::vector<SDOperand> Ops;
1403 Ops.push_back(Chain);
1404 Ops.push_back(Callee);
1405 Ops.insert(Ops.end(), args_to_use.begin(), args_to_use.end());
1406 SDOperand TheCall = DAG.getNode(PPCISD::CALL, RetVals, Ops);
1407 Chain = TheCall.getValue(TheCall.Val->getNumValues()-1);
1408 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
1409 DAG.getConstant(NumBytes, getPointerTy()));
1410 SDOperand RetVal = TheCall;
1412 // If the result is a small value, add a note so that we keep track of the
1413 // information about whether it is sign or zero extended.
1414 if (RetTyVT != ActualRetTyVT) {
1415 RetVal = DAG.getNode(RetTy->isSigned() ? ISD::AssertSext : ISD::AssertZext,
1416 MVT::i32, RetVal, DAG.getValueType(RetTyVT));
1417 RetVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, RetVal);
1418 } else if (RetTyVT == MVT::i64) {
1419 RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, RetVal, RetVal.getValue(1));
1422 return std::make_pair(RetVal, Chain);
1426 PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1427 MachineBasicBlock *BB) {
1428 assert((MI->getOpcode() == PPC::SELECT_CC_Int ||
1429 MI->getOpcode() == PPC::SELECT_CC_F4 ||
1430 MI->getOpcode() == PPC::SELECT_CC_F8) &&
1431 "Unexpected instr type to insert");
1433 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
1434 // control-flow pattern. The incoming instruction knows the destination vreg
1435 // to set, the condition code register to branch on, the true/false values to
1436 // select between, and a branch opcode to use.
1437 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1438 ilist<MachineBasicBlock>::iterator It = BB;
1444 // cmpTY ccX, r1, r2
1446 // fallthrough --> copy0MBB
1447 MachineBasicBlock *thisMBB = BB;
1448 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1449 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1450 BuildMI(BB, MI->getOperand(4).getImmedValue(), 2)
1451 .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
1452 MachineFunction *F = BB->getParent();
1453 F->getBasicBlockList().insert(It, copy0MBB);
1454 F->getBasicBlockList().insert(It, sinkMBB);
1455 // Update machine-CFG edges by first adding all successors of the current
1456 // block to the new block which will contain the Phi node for the select.
1457 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1458 e = BB->succ_end(); i != e; ++i)
1459 sinkMBB->addSuccessor(*i);
1460 // Next, remove all successors of the current block, and add the true
1461 // and fallthrough blocks as its successors.
1462 while(!BB->succ_empty())
1463 BB->removeSuccessor(BB->succ_begin());
1464 BB->addSuccessor(copy0MBB);
1465 BB->addSuccessor(sinkMBB);
1468 // %FalseValue = ...
1469 // # fallthrough to sinkMBB
1472 // Update machine-CFG edges
1473 BB->addSuccessor(sinkMBB);
1476 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1479 BuildMI(BB, PPC::PHI, 4, MI->getOperand(0).getReg())
1480 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
1481 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1483 delete MI; // The pseudo instruction is gone now.
1487 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
1488 DAGCombinerInfo &DCI) const {
1489 TargetMachine &TM = getTargetMachine();
1490 SelectionDAG &DAG = DCI.DAG;
1491 switch (N->getOpcode()) {
1493 case ISD::SINT_TO_FP:
1494 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
1495 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
1496 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
1497 // We allow the src/dst to be either f32/f64, but the intermediate
1498 // type must be i64.
1499 if (N->getOperand(0).getValueType() == MVT::i64) {
1500 SDOperand Val = N->getOperand(0).getOperand(0);
1501 if (Val.getValueType() == MVT::f32) {
1502 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
1503 DCI.AddToWorklist(Val.Val);
1506 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
1507 DCI.AddToWorklist(Val.Val);
1508 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
1509 DCI.AddToWorklist(Val.Val);
1510 if (N->getValueType(0) == MVT::f32) {
1511 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
1512 DCI.AddToWorklist(Val.Val);
1515 } else if (N->getOperand(0).getValueType() == MVT::i32) {
1516 // If the intermediate type is i32, we can avoid the load/store here
1523 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
1524 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
1525 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
1526 N->getOperand(1).getValueType() == MVT::i32) {
1527 SDOperand Val = N->getOperand(1).getOperand(0);
1528 if (Val.getValueType() == MVT::f32) {
1529 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
1530 DCI.AddToWorklist(Val.Val);
1532 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
1533 DCI.AddToWorklist(Val.Val);
1535 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
1536 N->getOperand(2), N->getOperand(3));
1537 DCI.AddToWorklist(Val.Val);
1541 case PPCISD::VCMP: {
1542 // If a VCMPo node already exists with exactly the same operands as this
1543 // node, use its result instead of this node (VCMPo computes both a CR6 and
1544 // a normal output).
1546 if (!N->getOperand(0).hasOneUse() &&
1547 !N->getOperand(1).hasOneUse() &&
1548 !N->getOperand(2).hasOneUse()) {
1550 // Scan all of the users of the LHS, looking for VCMPo's that match.
1551 SDNode *VCMPoNode = 0;
1553 SDNode *LHSN = N->getOperand(0).Val;
1554 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
1556 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
1557 (*UI)->getOperand(1) == N->getOperand(1) &&
1558 (*UI)->getOperand(2) == N->getOperand(2) &&
1559 (*UI)->getOperand(0) == N->getOperand(0)) {
1564 // If there are non-zero uses of the flag value, use the VCMPo node!
1565 if (VCMPoNode && !VCMPoNode->hasNUsesOfValue(0, 1))
1566 return SDOperand(VCMPoNode, 0);
1575 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1577 uint64_t &KnownZero,
1579 unsigned Depth) const {
1582 switch (Op.getOpcode()) {
1584 case ISD::INTRINSIC_WO_CHAIN: {
1585 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
1587 case Intrinsic::ppc_altivec_vcmpbfp_p:
1588 case Intrinsic::ppc_altivec_vcmpeqfp_p:
1589 case Intrinsic::ppc_altivec_vcmpequb_p:
1590 case Intrinsic::ppc_altivec_vcmpequh_p:
1591 case Intrinsic::ppc_altivec_vcmpequw_p:
1592 case Intrinsic::ppc_altivec_vcmpgefp_p:
1593 case Intrinsic::ppc_altivec_vcmpgtfp_p:
1594 case Intrinsic::ppc_altivec_vcmpgtsb_p:
1595 case Intrinsic::ppc_altivec_vcmpgtsh_p:
1596 case Intrinsic::ppc_altivec_vcmpgtsw_p:
1597 case Intrinsic::ppc_altivec_vcmpgtub_p:
1598 case Intrinsic::ppc_altivec_vcmpgtuh_p:
1599 case Intrinsic::ppc_altivec_vcmpgtuw_p:
1600 KnownZero = ~1U; // All bits but the low one are known to be zero.
1608 /// getConstraintType - Given a constraint letter, return the type of
1609 /// constraint it is for this target.
1610 PPCTargetLowering::ConstraintType
1611 PPCTargetLowering::getConstraintType(char ConstraintLetter) const {
1612 switch (ConstraintLetter) {
1619 return C_RegisterClass;
1621 return TargetLowering::getConstraintType(ConstraintLetter);
1625 std::vector<unsigned> PPCTargetLowering::
1626 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1627 MVT::ValueType VT) const {
1628 if (Constraint.size() == 1) {
1629 switch (Constraint[0]) { // GCC RS6000 Constraint Letters
1630 default: break; // Unknown constriant letter
1632 return make_vector<unsigned>(/*no R0*/ PPC::R1 , PPC::R2 , PPC::R3 ,
1633 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
1634 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
1635 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
1636 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
1637 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
1638 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
1639 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
1642 return make_vector<unsigned>(PPC::R0 , PPC::R1 , PPC::R2 , PPC::R3 ,
1643 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
1644 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
1645 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
1646 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
1647 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
1648 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
1649 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
1652 return make_vector<unsigned>(PPC::F0 , PPC::F1 , PPC::F2 , PPC::F3 ,
1653 PPC::F4 , PPC::F5 , PPC::F6 , PPC::F7 ,
1654 PPC::F8 , PPC::F9 , PPC::F10, PPC::F11,
1655 PPC::F12, PPC::F13, PPC::F14, PPC::F15,
1656 PPC::F16, PPC::F17, PPC::F18, PPC::F19,
1657 PPC::F20, PPC::F21, PPC::F22, PPC::F23,
1658 PPC::F24, PPC::F25, PPC::F26, PPC::F27,
1659 PPC::F28, PPC::F29, PPC::F30, PPC::F31,
1662 return make_vector<unsigned>(PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 ,
1663 PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 ,
1664 PPC::V8 , PPC::V9 , PPC::V10, PPC::V11,
1665 PPC::V12, PPC::V13, PPC::V14, PPC::V15,
1666 PPC::V16, PPC::V17, PPC::V18, PPC::V19,
1667 PPC::V20, PPC::V21, PPC::V22, PPC::V23,
1668 PPC::V24, PPC::V25, PPC::V26, PPC::V27,
1669 PPC::V28, PPC::V29, PPC::V30, PPC::V31,
1672 return make_vector<unsigned>(PPC::CR0, PPC::CR1, PPC::CR2, PPC::CR3,
1673 PPC::CR4, PPC::CR5, PPC::CR6, PPC::CR7,
1678 return std::vector<unsigned>();
1681 // isOperandValidForConstraint
1682 bool PPCTargetLowering::
1683 isOperandValidForConstraint(SDOperand Op, char Letter) {
1694 if (!isa<ConstantSDNode>(Op)) return false; // Must be an immediate.
1695 unsigned Value = cast<ConstantSDNode>(Op)->getValue();
1697 default: assert(0 && "Unknown constraint letter!");
1698 case 'I': // "I" is a signed 16-bit constant.
1699 return (short)Value == (int)Value;
1700 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
1701 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
1702 return (short)Value == 0;
1703 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
1704 return (Value >> 16) == 0;
1705 case 'M': // "M" is a constant that is greater than 31.
1707 case 'N': // "N" is a positive constant that is an exact power of two.
1708 return (int)Value > 0 && isPowerOf2_32(Value);
1709 case 'O': // "O" is the constant zero.
1711 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
1712 return (short)-Value == (int)-Value;
1718 // Handle standard constraint letters.
1719 return TargetLowering::isOperandValidForConstraint(Op, Letter);
1722 /// isLegalAddressImmediate - Return true if the integer value can be used
1723 /// as the offset of the target addressing mode.
1724 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V) const {
1725 // PPC allows a sign-extended 16-bit immediate field.
1726 return (V > -(1 << 16) && V < (1 << 16)-1);