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 can custom expand all VECTOR_SHUFFLEs to VPERM.
185 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
187 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
188 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
189 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
190 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
191 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
192 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
193 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
194 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
197 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
198 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
199 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
200 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
202 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
204 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
205 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
207 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
208 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
209 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
210 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
213 setSetCCResultContents(ZeroOrOneSetCCResult);
214 setStackPointerRegisterToSaveRestore(PPC::R1);
216 // We have target-specific dag combine patterns for the following nodes:
217 setTargetDAGCombine(ISD::SINT_TO_FP);
218 setTargetDAGCombine(ISD::STORE);
220 computeRegisterProperties();
223 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
226 case PPCISD::FSEL: return "PPCISD::FSEL";
227 case PPCISD::FCFID: return "PPCISD::FCFID";
228 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
229 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
230 case PPCISD::STFIWX: return "PPCISD::STFIWX";
231 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
232 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
233 case PPCISD::VPERM: return "PPCISD::VPERM";
234 case PPCISD::Hi: return "PPCISD::Hi";
235 case PPCISD::Lo: return "PPCISD::Lo";
236 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
237 case PPCISD::SRL: return "PPCISD::SRL";
238 case PPCISD::SRA: return "PPCISD::SRA";
239 case PPCISD::SHL: return "PPCISD::SHL";
240 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
241 case PPCISD::STD_32: return "PPCISD::STD_32";
242 case PPCISD::CALL: return "PPCISD::CALL";
243 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
244 case PPCISD::MFCR: return "PPCISD::MFCR";
245 case PPCISD::VCMP: return "PPCISD::VCMP";
246 case PPCISD::VCMPo: return "PPCISD::VCMPo";
250 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
251 static bool isFloatingPointZero(SDOperand Op) {
252 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
253 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
254 else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
255 // Maybe this has already been legalized into the constant pool?
256 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
257 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
258 return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
264 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
265 /// specifies a splat of a single element that is suitable for input to
266 /// VSPLTB/VSPLTH/VSPLTW.
267 bool PPC::isSplatShuffleMask(SDNode *N) {
268 assert(N->getOpcode() == ISD::BUILD_VECTOR);
270 // We can only splat 8-bit, 16-bit, and 32-bit quantities.
271 if (N->getNumOperands() != 4 && N->getNumOperands() != 8 &&
272 N->getNumOperands() != 16)
275 // This is a splat operation if each element of the permute is the same, and
276 // if the value doesn't reference the second vector.
277 SDOperand Elt = N->getOperand(0);
278 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
279 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i) {
280 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
281 "Invalid VECTOR_SHUFFLE mask!");
282 if (N->getOperand(i) != Elt) return false;
285 // Make sure it is a splat of the first vector operand.
286 return cast<ConstantSDNode>(Elt)->getValue() < N->getNumOperands();
289 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
290 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
291 unsigned PPC::getVSPLTImmediate(SDNode *N) {
292 assert(isSplatShuffleMask(N));
293 return cast<ConstantSDNode>(N->getOperand(0))->getValue();
296 /// isVecSplatImm - Return true if this is a build_vector of constants which
297 /// can be formed by using a vspltis[bhw] instruction. The ByteSize field
298 /// indicates the number of bytes of each element [124] -> [bhw].
299 bool PPC::isVecSplatImm(SDNode *N, unsigned ByteSize, char *Val) {
300 SDOperand OpVal(0, 0);
301 // Check to see if this buildvec has a single non-undef value in its elements.
302 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
303 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
305 OpVal = N->getOperand(i);
306 else if (OpVal != N->getOperand(i))
310 if (OpVal.Val == 0) return false; // All UNDEF: use implicit def.
312 unsigned ValSizeInBytes = 0;
314 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
315 Value = CN->getValue();
316 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
317 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
318 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
319 Value = FloatToBits(CN->getValue());
323 // If the splat value is larger than the element value, then we can never do
324 // this splat. The only case that we could fit the replicated bits into our
325 // immediate field for would be zero, and we prefer to use vxor for it.
326 if (ValSizeInBytes < ByteSize) return false;
328 // If the element value is larger than the splat value, cut it in half and
329 // check to see if the two halves are equal. Continue doing this until we
330 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
331 while (ValSizeInBytes > ByteSize) {
332 ValSizeInBytes >>= 1;
334 // If the top half equals the bottom half, we're still ok.
335 if (((Value >> (ValSizeInBytes*8)) & ((8 << ValSizeInBytes)-1)) !=
336 (Value & ((8 << ValSizeInBytes)-1)))
340 // Properly sign extend the value.
341 int ShAmt = (4-ByteSize)*8;
342 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
344 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
345 if (MaskVal == 0) return false;
347 if (Val) *Val = MaskVal;
349 // Finally, if this value fits in a 5 bit sext field, return true.
350 return ((MaskVal << (32-5)) >> (32-5)) == MaskVal;
354 /// LowerOperation - Provide custom lowering hooks for some operations.
356 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
357 switch (Op.getOpcode()) {
358 default: assert(0 && "Wasn't expecting to be able to lower this!");
359 case ISD::FP_TO_SINT: {
360 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
361 SDOperand Src = Op.getOperand(0);
362 if (Src.getValueType() == MVT::f32)
363 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
366 switch (Op.getValueType()) {
367 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
369 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
372 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
376 // Convert the FP value to an int value through memory.
377 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
378 if (Op.getValueType() == MVT::i32)
379 Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
382 case ISD::SINT_TO_FP:
383 if (Op.getOperand(0).getValueType() == MVT::i64) {
384 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
385 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
386 if (Op.getValueType() == MVT::f32)
387 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
390 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
391 "Unhandled SINT_TO_FP type in custom expander!");
392 // Since we only generate this in 64-bit mode, we can take advantage of
393 // 64-bit registers. In particular, sign extend the input value into the
394 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
395 // then lfd it and fcfid it.
396 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
397 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
398 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
400 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
403 // STD the extended value into the stack slot.
404 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
405 DAG.getEntryNode(), Ext64, FIdx,
406 DAG.getSrcValue(NULL));
407 // Load the value as a double.
408 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
410 // FCFID it and return it.
411 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
412 if (Op.getValueType() == MVT::f32)
413 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
418 case ISD::SELECT_CC: {
419 // Turn FP only select_cc's into fsel instructions.
420 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
421 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
424 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
426 // Cannot handle SETEQ/SETNE.
427 if (CC == ISD::SETEQ || CC == ISD::SETNE) break;
429 MVT::ValueType ResVT = Op.getValueType();
430 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
431 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
432 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
434 // If the RHS of the comparison is a 0.0, we don't need to do the
435 // subtraction at all.
436 if (isFloatingPointZero(RHS))
438 default: break; // SETUO etc aren't handled by fsel.
441 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
444 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
445 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
446 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
449 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
452 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
453 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
454 return DAG.getNode(PPCISD::FSEL, ResVT,
455 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
460 default: break; // SETUO etc aren't handled by fsel.
463 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
464 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
465 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
466 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
469 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
470 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
471 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
472 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
475 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
476 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
477 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
478 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
481 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
482 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
483 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
484 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
489 assert(Op.getValueType() == MVT::i64 &&
490 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
491 // The generic code does a fine job expanding shift by a constant.
492 if (isa<ConstantSDNode>(Op.getOperand(1))) break;
494 // Otherwise, expand into a bunch of logical ops. Note that these ops
495 // depend on the PPC behavior for oversized shift amounts.
496 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
497 DAG.getConstant(0, MVT::i32));
498 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
499 DAG.getConstant(1, MVT::i32));
500 SDOperand Amt = Op.getOperand(1);
502 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
503 DAG.getConstant(32, MVT::i32), Amt);
504 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
505 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
506 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
507 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
508 DAG.getConstant(-32U, MVT::i32));
509 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
510 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
511 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
512 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
515 assert(Op.getValueType() == MVT::i64 &&
516 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
517 // The generic code does a fine job expanding shift by a constant.
518 if (isa<ConstantSDNode>(Op.getOperand(1))) break;
520 // Otherwise, expand into a bunch of logical ops. Note that these ops
521 // depend on the PPC behavior for oversized shift amounts.
522 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
523 DAG.getConstant(0, MVT::i32));
524 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
525 DAG.getConstant(1, MVT::i32));
526 SDOperand Amt = Op.getOperand(1);
528 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
529 DAG.getConstant(32, MVT::i32), Amt);
530 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
531 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
532 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
533 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
534 DAG.getConstant(-32U, MVT::i32));
535 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
536 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
537 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
538 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
541 assert(Op.getValueType() == MVT::i64 &&
542 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
543 // The generic code does a fine job expanding shift by a constant.
544 if (isa<ConstantSDNode>(Op.getOperand(1))) break;
546 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
547 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
548 DAG.getConstant(0, MVT::i32));
549 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
550 DAG.getConstant(1, MVT::i32));
551 SDOperand Amt = Op.getOperand(1);
553 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
554 DAG.getConstant(32, MVT::i32), Amt);
555 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
556 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
557 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
558 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
559 DAG.getConstant(-32U, MVT::i32));
560 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
561 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
562 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
563 Tmp4, Tmp6, ISD::SETLE);
564 return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
566 case ISD::ConstantPool: {
567 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
568 Constant *C = CP->get();
569 SDOperand CPI = DAG.getTargetConstantPool(C, MVT::i32, CP->getAlignment());
570 SDOperand Zero = DAG.getConstant(0, MVT::i32);
572 if (getTargetMachine().getRelocationModel() == Reloc::Static) {
573 // Generate non-pic code that has direct accesses to the constant pool.
574 // The address of the global is just (hi(&g)+lo(&g)).
575 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
576 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
577 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
580 // Only lower ConstantPool on Darwin.
581 if (!getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin()) break;
582 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, CPI, Zero);
583 if (getTargetMachine().getRelocationModel() == Reloc::PIC) {
584 // With PIC, the first instruction is actually "GR+hi(&G)".
585 Hi = DAG.getNode(ISD::ADD, MVT::i32,
586 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
589 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, CPI, Zero);
590 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
593 case ISD::GlobalAddress: {
594 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
595 GlobalValue *GV = GSDN->getGlobal();
596 SDOperand GA = DAG.getTargetGlobalAddress(GV, MVT::i32, GSDN->getOffset());
597 SDOperand Zero = DAG.getConstant(0, MVT::i32);
599 if (getTargetMachine().getRelocationModel() == Reloc::Static) {
600 // Generate non-pic code that has direct accesses to globals.
601 // The address of the global is just (hi(&g)+lo(&g)).
602 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
603 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
604 return DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
607 // Only lower GlobalAddress on Darwin.
608 if (!getTargetMachine().getSubtarget<PPCSubtarget>().isDarwin()) break;
610 SDOperand Hi = DAG.getNode(PPCISD::Hi, MVT::i32, GA, Zero);
611 if (getTargetMachine().getRelocationModel() == Reloc::PIC) {
612 // With PIC, the first instruction is actually "GR+hi(&G)".
613 Hi = DAG.getNode(ISD::ADD, MVT::i32,
614 DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
617 SDOperand Lo = DAG.getNode(PPCISD::Lo, MVT::i32, GA, Zero);
618 Lo = DAG.getNode(ISD::ADD, MVT::i32, Hi, Lo);
620 if (!GV->hasWeakLinkage() && !GV->hasLinkOnceLinkage() &&
621 (!GV->isExternal() || GV->hasNotBeenReadFromBytecode()))
624 // If the global is weak or external, we have to go through the lazy
626 return DAG.getLoad(MVT::i32, DAG.getEntryNode(), Lo, DAG.getSrcValue(0));
629 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
631 // If we're comparing for equality to zero, expose the fact that this is
632 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
633 // fold the new nodes.
634 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
635 if (C->isNullValue() && CC == ISD::SETEQ) {
636 MVT::ValueType VT = Op.getOperand(0).getValueType();
637 SDOperand Zext = Op.getOperand(0);
640 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
642 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
643 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
644 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
645 DAG.getConstant(Log2b, getShiftAmountTy()));
646 return DAG.getNode(ISD::TRUNCATE, getSetCCResultTy(), Scc);
648 // Leave comparisons against 0 and -1 alone for now, since they're usually
649 // optimized. FIXME: revisit this when we can custom lower all setcc
651 if (C->isAllOnesValue() || C->isNullValue())
655 // If we have an integer seteq/setne, turn it into a compare against zero
656 // by subtracting the rhs from the lhs, which is faster than setting a
657 // condition register, reading it back out, and masking the correct bit.
658 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
659 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
660 MVT::ValueType VT = Op.getValueType();
661 SDOperand Sub = DAG.getNode(ISD::SUB, LHSVT, Op.getOperand(0),
663 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
668 // vastart just stores the address of the VarArgsFrameIndex slot into the
669 // memory location argument.
670 // FIXME: Replace MVT::i32 with PointerTy
671 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
672 return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
673 Op.getOperand(1), Op.getOperand(2));
678 switch(Op.getNumOperands()) {
680 assert(0 && "Do not know how to return this many arguments!");
683 return SDOperand(); // ret void is legal
685 MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
686 unsigned ArgReg = MVT::isInteger(ArgVT) ? PPC::R3 : PPC::F1;
687 Copy = DAG.getCopyToReg(Op.getOperand(0), ArgReg, Op.getOperand(1),
692 Copy = DAG.getCopyToReg(Op.getOperand(0), PPC::R3, Op.getOperand(2),
694 Copy = DAG.getCopyToReg(Copy, PPC::R4, Op.getOperand(1),Copy.getValue(1));
697 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
699 case ISD::SCALAR_TO_VECTOR: {
700 // Create a stack slot that is 16-byte aligned.
701 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
702 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
703 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
705 // Store the input value into Value#0 of the stack slot.
706 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
707 Op.getOperand(0), FIdx,DAG.getSrcValue(NULL));
709 return DAG.getLoad(Op.getValueType(), Store, FIdx, DAG.getSrcValue(NULL));
711 case ISD::BUILD_VECTOR:
712 // If this is a case we can't handle, return null and let the default
713 // expansion code take care of it. If we CAN select this case, return Op.
715 // See if this is all zeros.
716 // FIXME: We should handle splat(-0.0), and other cases here.
717 if (ISD::isBuildVectorAllZeros(Op.Val))
720 if (PPC::isVecSplatImm(Op.Val, 1) || // vspltisb
721 PPC::isVecSplatImm(Op.Val, 2) || // vspltish
722 PPC::isVecSplatImm(Op.Val, 4)) // vspltisw
727 case ISD::VECTOR_SHUFFLE: {
728 SDOperand V1 = Op.getOperand(0);
729 SDOperand V2 = Op.getOperand(1);
730 SDOperand PermMask = Op.getOperand(2);
732 // Cases that are handled by instructions that take permute immediates
733 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
734 // selected by the instruction selector.
735 if (PPC::isSplatShuffleMask(PermMask.Val) && V2.getOpcode() == ISD::UNDEF)
738 // TODO: Handle more cases, and also handle cases that are cheaper to do as
739 // multiple such instructions than as a constant pool load/vperm pair.
741 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
742 // vector that will get spilled to the constant pool.
743 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
745 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
746 // that it is in input element units, not in bytes. Convert now.
747 MVT::ValueType EltVT = MVT::getVectorBaseType(V1.getValueType());
748 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
750 std::vector<SDOperand> ResultMask;
751 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
752 unsigned SrcElt =cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
754 for (unsigned j = 0; j != BytesPerElement; ++j)
755 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
759 SDOperand VPermMask =DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, ResultMask);
760 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
762 case ISD::INTRINSIC_WO_CHAIN: {
763 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
765 // If this is a lowered altivec predicate compare, CompareOpc is set to the
766 // opcode number of the comparison.
770 default: return SDOperand(); // Don't custom lower most intrinsics.
771 // Comparison predicates.
772 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
773 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
774 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
775 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
776 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
777 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
778 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
779 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
780 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
781 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
782 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
783 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
784 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
786 // Normal Comparisons.
787 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
788 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
789 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
790 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
791 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
792 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
793 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
794 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
795 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
796 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
797 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
798 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
799 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
802 assert(CompareOpc>0 && "We only lower altivec predicate compares so far!");
804 // If this is a non-dot comparison, make the VCMP node.
806 return DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
807 Op.getOperand(1), Op.getOperand(2),
808 DAG.getConstant(CompareOpc, MVT::i32));
810 // Create the PPCISD altivec 'dot' comparison node.
811 std::vector<SDOperand> Ops;
812 std::vector<MVT::ValueType> VTs;
813 Ops.push_back(Op.getOperand(2)); // LHS
814 Ops.push_back(Op.getOperand(3)); // RHS
815 Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
816 VTs.push_back(Op.getOperand(2).getValueType());
817 VTs.push_back(MVT::Flag);
818 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
820 // Now that we have the comparison, emit a copy from the CR to a GPR.
821 // This is flagged to the above dot comparison.
822 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
823 DAG.getRegister(PPC::CR6, MVT::i32),
824 CompNode.getValue(1));
826 // Unpack the result based on how the target uses it.
827 unsigned BitNo; // Bit # of CR6.
828 bool InvertBit; // Invert result?
829 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
830 default: // Can't happen, don't crash on invalid number though.
831 case 0: // Return the value of the EQ bit of CR6.
832 BitNo = 0; InvertBit = false;
834 case 1: // Return the inverted value of the EQ bit of CR6.
835 BitNo = 0; InvertBit = true;
837 case 2: // Return the value of the LT bit of CR6.
838 BitNo = 2; InvertBit = false;
840 case 3: // Return the inverted value of the LT bit of CR6.
841 BitNo = 2; InvertBit = true;
845 // Shift the bit into the low position.
846 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
847 DAG.getConstant(8-(3-BitNo), MVT::i32));
849 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
850 DAG.getConstant(1, MVT::i32));
852 // If we are supposed to, toggle the bit.
854 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
855 DAG.getConstant(1, MVT::i32));
862 std::vector<SDOperand>
863 PPCTargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
865 // add beautiful description of PPC stack frame format, or at least some docs
867 MachineFunction &MF = DAG.getMachineFunction();
868 MachineFrameInfo *MFI = MF.getFrameInfo();
869 MachineBasicBlock& BB = MF.front();
870 SSARegMap *RegMap = MF.getSSARegMap();
871 std::vector<SDOperand> ArgValues;
873 unsigned ArgOffset = 24;
874 unsigned GPR_remaining = 8;
875 unsigned FPR_remaining = 13;
876 unsigned GPR_idx = 0, FPR_idx = 0;
877 static const unsigned GPR[] = {
878 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
879 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
881 static const unsigned FPR[] = {
882 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
883 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
886 // Add DAG nodes to load the arguments... On entry to a function on PPC,
887 // the arguments start at offset 24, although they are likely to be passed
889 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
890 SDOperand newroot, argt;
892 bool needsLoad = false;
893 bool ArgLive = !I->use_empty();
894 MVT::ValueType ObjectVT = getValueType(I->getType());
897 default: assert(0 && "Unhandled argument type!");
904 if (GPR_remaining > 0) {
905 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
906 MF.addLiveIn(GPR[GPR_idx], VReg);
907 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
908 if (ObjectVT != MVT::i32) {
909 unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
911 argt = DAG.getNode(AssertOp, MVT::i32, argt,
912 DAG.getValueType(ObjectVT));
913 argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, argt);
922 if (GPR_remaining > 0) {
923 SDOperand argHi, argLo;
924 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
925 MF.addLiveIn(GPR[GPR_idx], VReg);
926 argHi = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
927 // If we have two or more remaining argument registers, then both halves
928 // of the i64 can be sourced from there. Otherwise, the lower half will
929 // have to come off the stack. This can happen when an i64 is preceded
930 // by 28 bytes of arguments.
931 if (GPR_remaining > 1) {
932 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
933 MF.addLiveIn(GPR[GPR_idx+1], VReg);
934 argLo = DAG.getCopyFromReg(argHi, VReg, MVT::i32);
936 int FI = MFI->CreateFixedObject(4, ArgOffset+4);
937 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
938 argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
939 DAG.getSrcValue(NULL));
941 // Build the outgoing arg thingy
942 argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
950 ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
952 if (FPR_remaining > 0) {
958 if (FPR_remaining > 0) {
960 if (ObjectVT == MVT::f32)
961 VReg = RegMap->createVirtualRegister(&PPC::F4RCRegClass);
963 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
964 MF.addLiveIn(FPR[FPR_idx], VReg);
965 argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, ObjectVT);
974 // We need to load the argument to a virtual register if we determined above
975 // that we ran out of physical registers of the appropriate type
977 unsigned SubregOffset = 0;
978 if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
979 if (ObjectVT == MVT::i16) SubregOffset = 2;
980 int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
981 SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
982 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
983 DAG.getConstant(SubregOffset, MVT::i32));
984 argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
985 DAG.getSrcValue(NULL));
988 // Every 4 bytes of argument space consumes one of the GPRs available for
990 if (GPR_remaining > 0) {
991 unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
992 GPR_remaining -= delta;
995 ArgOffset += ObjSize;
997 DAG.setRoot(newroot.getValue(1));
999 ArgValues.push_back(argt);
1002 // If the function takes variable number of arguments, make a frame index for
1003 // the start of the first vararg value... for expansion of llvm.va_start.
1005 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
1006 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
1007 // If this function is vararg, store any remaining integer argument regs
1008 // to their spots on the stack so that they may be loaded by deferencing the
1009 // result of va_next.
1010 std::vector<SDOperand> MemOps;
1011 for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
1012 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1013 MF.addLiveIn(GPR[GPR_idx], VReg);
1014 SDOperand Val = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
1015 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
1016 Val, FIN, DAG.getSrcValue(NULL));
1017 MemOps.push_back(Store);
1018 // Increment the address by four for the next argument to store
1019 SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
1020 FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
1022 if (!MemOps.empty()) {
1023 MemOps.push_back(DAG.getRoot());
1024 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
1028 // Finally, inform the code generator which regs we return values in.
1029 switch (getValueType(F.getReturnType())) {
1030 default: assert(0 && "Unknown type!");
1031 case MVT::isVoid: break;
1036 MF.addLiveOut(PPC::R3);
1039 MF.addLiveOut(PPC::R3);
1040 MF.addLiveOut(PPC::R4);
1044 MF.addLiveOut(PPC::F1);
1051 std::pair<SDOperand, SDOperand>
1052 PPCTargetLowering::LowerCallTo(SDOperand Chain,
1053 const Type *RetTy, bool isVarArg,
1054 unsigned CallingConv, bool isTailCall,
1055 SDOperand Callee, ArgListTy &Args,
1056 SelectionDAG &DAG) {
1057 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1058 // SelectExpr to use to put the arguments in the appropriate registers.
1059 std::vector<SDOperand> args_to_use;
1061 // Count how many bytes are to be pushed on the stack, including the linkage
1062 // area, and parameter passing area.
1063 unsigned NumBytes = 24;
1066 Chain = DAG.getCALLSEQ_START(Chain,
1067 DAG.getConstant(NumBytes, getPointerTy()));
1069 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1070 switch (getValueType(Args[i].second)) {
1071 default: assert(0 && "Unknown value type!");
1086 // Just to be safe, we'll always reserve the full 24 bytes of linkage area
1087 // plus 32 bytes of argument space in case any called code gets funky on us.
1088 // (Required by ABI to support var arg)
1089 if (NumBytes < 56) NumBytes = 56;
1091 // Adjust the stack pointer for the new arguments...
1092 // These operations are automatically eliminated by the prolog/epilog pass
1093 Chain = DAG.getCALLSEQ_START(Chain,
1094 DAG.getConstant(NumBytes, getPointerTy()));
1096 // Set up a copy of the stack pointer for use loading and storing any
1097 // arguments that may not fit in the registers available for argument
1099 SDOperand StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1101 // Figure out which arguments are going to go in registers, and which in
1102 // memory. Also, if this is a vararg function, floating point operations
1103 // must be stored to our stack, and loaded into integer regs as well, if
1104 // any integer regs are available for argument passing.
1105 unsigned ArgOffset = 24;
1106 unsigned GPR_remaining = 8;
1107 unsigned FPR_remaining = 13;
1109 std::vector<SDOperand> MemOps;
1110 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1111 // PtrOff will be used to store the current argument to the stack if a
1112 // register cannot be found for it.
1113 SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
1114 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
1115 MVT::ValueType ArgVT = getValueType(Args[i].second);
1118 default: assert(0 && "Unexpected ValueType for argument!");
1122 // Promote the integer to 32 bits. If the input type is signed use a
1123 // sign extend, otherwise use a zero extend.
1124 if (Args[i].second->isSigned())
1125 Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
1127 Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
1130 if (GPR_remaining > 0) {
1131 args_to_use.push_back(Args[i].first);
1134 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1135 Args[i].first, PtrOff,
1136 DAG.getSrcValue(NULL)));
1141 // If we have one free GPR left, we can place the upper half of the i64
1142 // in it, and store the other half to the stack. If we have two or more
1143 // free GPRs, then we can pass both halves of the i64 in registers.
1144 if (GPR_remaining > 0) {
1145 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1146 Args[i].first, DAG.getConstant(1, MVT::i32));
1147 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
1148 Args[i].first, DAG.getConstant(0, MVT::i32));
1149 args_to_use.push_back(Hi);
1151 if (GPR_remaining > 0) {
1152 args_to_use.push_back(Lo);
1155 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1156 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1157 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1158 Lo, PtrOff, DAG.getSrcValue(NULL)));
1161 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1162 Args[i].first, PtrOff,
1163 DAG.getSrcValue(NULL)));
1169 if (FPR_remaining > 0) {
1170 args_to_use.push_back(Args[i].first);
1173 SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
1174 Args[i].first, PtrOff,
1175 DAG.getSrcValue(NULL));
1176 MemOps.push_back(Store);
1177 // Float varargs are always shadowed in available integer registers
1178 if (GPR_remaining > 0) {
1179 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1180 DAG.getSrcValue(NULL));
1181 MemOps.push_back(Load.getValue(1));
1182 args_to_use.push_back(Load);
1185 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1186 SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
1187 PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
1188 SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
1189 DAG.getSrcValue(NULL));
1190 MemOps.push_back(Load.getValue(1));
1191 args_to_use.push_back(Load);
1195 // If we have any FPRs remaining, we may also have GPRs remaining.
1196 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1198 if (GPR_remaining > 0) {
1199 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1202 if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
1203 args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
1208 MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
1209 Args[i].first, PtrOff,
1210 DAG.getSrcValue(NULL)));
1212 ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
1216 if (!MemOps.empty())
1217 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
1220 std::vector<MVT::ValueType> RetVals;
1221 MVT::ValueType RetTyVT = getValueType(RetTy);
1222 MVT::ValueType ActualRetTyVT = RetTyVT;
1223 if (RetTyVT >= MVT::i1 && RetTyVT <= MVT::i16)
1224 ActualRetTyVT = MVT::i32; // Promote result to i32.
1226 if (RetTyVT == MVT::i64) {
1227 RetVals.push_back(MVT::i32);
1228 RetVals.push_back(MVT::i32);
1229 } else if (RetTyVT != MVT::isVoid) {
1230 RetVals.push_back(ActualRetTyVT);
1232 RetVals.push_back(MVT::Other);
1234 // If the callee is a GlobalAddress node (quite common, every direct call is)
1235 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1236 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1237 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), MVT::i32);
1239 std::vector<SDOperand> Ops;
1240 Ops.push_back(Chain);
1241 Ops.push_back(Callee);
1242 Ops.insert(Ops.end(), args_to_use.begin(), args_to_use.end());
1243 SDOperand TheCall = DAG.getNode(PPCISD::CALL, RetVals, Ops);
1244 Chain = TheCall.getValue(TheCall.Val->getNumValues()-1);
1245 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
1246 DAG.getConstant(NumBytes, getPointerTy()));
1247 SDOperand RetVal = TheCall;
1249 // If the result is a small value, add a note so that we keep track of the
1250 // information about whether it is sign or zero extended.
1251 if (RetTyVT != ActualRetTyVT) {
1252 RetVal = DAG.getNode(RetTy->isSigned() ? ISD::AssertSext : ISD::AssertZext,
1253 MVT::i32, RetVal, DAG.getValueType(RetTyVT));
1254 RetVal = DAG.getNode(ISD::TRUNCATE, RetTyVT, RetVal);
1255 } else if (RetTyVT == MVT::i64) {
1256 RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, RetVal, RetVal.getValue(1));
1259 return std::make_pair(RetVal, Chain);
1263 PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
1264 MachineBasicBlock *BB) {
1265 assert((MI->getOpcode() == PPC::SELECT_CC_Int ||
1266 MI->getOpcode() == PPC::SELECT_CC_F4 ||
1267 MI->getOpcode() == PPC::SELECT_CC_F8) &&
1268 "Unexpected instr type to insert");
1270 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
1271 // control-flow pattern. The incoming instruction knows the destination vreg
1272 // to set, the condition code register to branch on, the true/false values to
1273 // select between, and a branch opcode to use.
1274 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1275 ilist<MachineBasicBlock>::iterator It = BB;
1281 // cmpTY ccX, r1, r2
1283 // fallthrough --> copy0MBB
1284 MachineBasicBlock *thisMBB = BB;
1285 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
1286 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
1287 BuildMI(BB, MI->getOperand(4).getImmedValue(), 2)
1288 .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
1289 MachineFunction *F = BB->getParent();
1290 F->getBasicBlockList().insert(It, copy0MBB);
1291 F->getBasicBlockList().insert(It, sinkMBB);
1292 // Update machine-CFG edges by first adding all successors of the current
1293 // block to the new block which will contain the Phi node for the select.
1294 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
1295 e = BB->succ_end(); i != e; ++i)
1296 sinkMBB->addSuccessor(*i);
1297 // Next, remove all successors of the current block, and add the true
1298 // and fallthrough blocks as its successors.
1299 while(!BB->succ_empty())
1300 BB->removeSuccessor(BB->succ_begin());
1301 BB->addSuccessor(copy0MBB);
1302 BB->addSuccessor(sinkMBB);
1305 // %FalseValue = ...
1306 // # fallthrough to sinkMBB
1309 // Update machine-CFG edges
1310 BB->addSuccessor(sinkMBB);
1313 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
1316 BuildMI(BB, PPC::PHI, 4, MI->getOperand(0).getReg())
1317 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
1318 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
1320 delete MI; // The pseudo instruction is gone now.
1324 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
1325 DAGCombinerInfo &DCI) const {
1326 TargetMachine &TM = getTargetMachine();
1327 SelectionDAG &DAG = DCI.DAG;
1328 switch (N->getOpcode()) {
1330 case ISD::SINT_TO_FP:
1331 if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
1332 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
1333 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
1334 // We allow the src/dst to be either f32/f64, but the intermediate
1335 // type must be i64.
1336 if (N->getOperand(0).getValueType() == MVT::i64) {
1337 SDOperand Val = N->getOperand(0).getOperand(0);
1338 if (Val.getValueType() == MVT::f32) {
1339 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
1340 DCI.AddToWorklist(Val.Val);
1343 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
1344 DCI.AddToWorklist(Val.Val);
1345 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
1346 DCI.AddToWorklist(Val.Val);
1347 if (N->getValueType(0) == MVT::f32) {
1348 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
1349 DCI.AddToWorklist(Val.Val);
1352 } else if (N->getOperand(0).getValueType() == MVT::i32) {
1353 // If the intermediate type is i32, we can avoid the load/store here
1360 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
1361 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
1362 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
1363 N->getOperand(1).getValueType() == MVT::i32) {
1364 SDOperand Val = N->getOperand(1).getOperand(0);
1365 if (Val.getValueType() == MVT::f32) {
1366 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
1367 DCI.AddToWorklist(Val.Val);
1369 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
1370 DCI.AddToWorklist(Val.Val);
1372 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
1373 N->getOperand(2), N->getOperand(3));
1374 DCI.AddToWorklist(Val.Val);
1378 case PPCISD::VCMP: {
1379 // If a VCMPo node already exists with exactly the same operands as this
1380 // node, use its result instead of this node (VCMPo computes both a CR6 and
1381 // a normal output).
1383 if (!N->getOperand(0).hasOneUse() &&
1384 !N->getOperand(1).hasOneUse() &&
1385 !N->getOperand(2).hasOneUse()) {
1387 // Scan all of the users of the LHS, looking for VCMPo's that match.
1388 SDNode *VCMPoNode = 0;
1390 SDNode *LHSN = N->getOperand(0).Val;
1391 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
1393 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
1394 (*UI)->getOperand(1) == N->getOperand(1) &&
1395 (*UI)->getOperand(2) == N->getOperand(2) &&
1396 (*UI)->getOperand(0) == N->getOperand(0)) {
1401 // If there are non-zero uses of the flag value, use the VCMPo node!
1402 if (VCMPoNode && !VCMPoNode->hasNUsesOfValue(0, 1))
1403 return SDOperand(VCMPoNode, 0);
1412 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
1414 uint64_t &KnownZero,
1416 unsigned Depth) const {
1419 switch (Op.getOpcode()) {
1421 case ISD::INTRINSIC_WO_CHAIN: {
1422 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
1424 case Intrinsic::ppc_altivec_vcmpbfp_p:
1425 case Intrinsic::ppc_altivec_vcmpeqfp_p:
1426 case Intrinsic::ppc_altivec_vcmpequb_p:
1427 case Intrinsic::ppc_altivec_vcmpequh_p:
1428 case Intrinsic::ppc_altivec_vcmpequw_p:
1429 case Intrinsic::ppc_altivec_vcmpgefp_p:
1430 case Intrinsic::ppc_altivec_vcmpgtfp_p:
1431 case Intrinsic::ppc_altivec_vcmpgtsb_p:
1432 case Intrinsic::ppc_altivec_vcmpgtsh_p:
1433 case Intrinsic::ppc_altivec_vcmpgtsw_p:
1434 case Intrinsic::ppc_altivec_vcmpgtub_p:
1435 case Intrinsic::ppc_altivec_vcmpgtuh_p:
1436 case Intrinsic::ppc_altivec_vcmpgtuw_p:
1437 KnownZero = ~1U; // All bits but the low one are known to be zero.
1445 /// getConstraintType - Given a constraint letter, return the type of
1446 /// constraint it is for this target.
1447 PPCTargetLowering::ConstraintType
1448 PPCTargetLowering::getConstraintType(char ConstraintLetter) const {
1449 switch (ConstraintLetter) {
1456 return C_RegisterClass;
1458 return TargetLowering::getConstraintType(ConstraintLetter);
1462 std::vector<unsigned> PPCTargetLowering::
1463 getRegClassForInlineAsmConstraint(const std::string &Constraint,
1464 MVT::ValueType VT) const {
1465 if (Constraint.size() == 1) {
1466 switch (Constraint[0]) { // GCC RS6000 Constraint Letters
1467 default: break; // Unknown constriant letter
1469 return make_vector<unsigned>(/*no R0*/ PPC::R1 , PPC::R2 , PPC::R3 ,
1470 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
1471 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
1472 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
1473 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
1474 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
1475 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
1476 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
1479 return make_vector<unsigned>(PPC::R0 , PPC::R1 , PPC::R2 , PPC::R3 ,
1480 PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
1481 PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
1482 PPC::R12, PPC::R13, PPC::R14, PPC::R15,
1483 PPC::R16, PPC::R17, PPC::R18, PPC::R19,
1484 PPC::R20, PPC::R21, PPC::R22, PPC::R23,
1485 PPC::R24, PPC::R25, PPC::R26, PPC::R27,
1486 PPC::R28, PPC::R29, PPC::R30, PPC::R31,
1489 return make_vector<unsigned>(PPC::F0 , PPC::F1 , PPC::F2 , PPC::F3 ,
1490 PPC::F4 , PPC::F5 , PPC::F6 , PPC::F7 ,
1491 PPC::F8 , PPC::F9 , PPC::F10, PPC::F11,
1492 PPC::F12, PPC::F13, PPC::F14, PPC::F15,
1493 PPC::F16, PPC::F17, PPC::F18, PPC::F19,
1494 PPC::F20, PPC::F21, PPC::F22, PPC::F23,
1495 PPC::F24, PPC::F25, PPC::F26, PPC::F27,
1496 PPC::F28, PPC::F29, PPC::F30, PPC::F31,
1499 return make_vector<unsigned>(PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 ,
1500 PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 ,
1501 PPC::V8 , PPC::V9 , PPC::V10, PPC::V11,
1502 PPC::V12, PPC::V13, PPC::V14, PPC::V15,
1503 PPC::V16, PPC::V17, PPC::V18, PPC::V19,
1504 PPC::V20, PPC::V21, PPC::V22, PPC::V23,
1505 PPC::V24, PPC::V25, PPC::V26, PPC::V27,
1506 PPC::V28, PPC::V29, PPC::V30, PPC::V31,
1509 return make_vector<unsigned>(PPC::CR0, PPC::CR1, PPC::CR2, PPC::CR3,
1510 PPC::CR4, PPC::CR5, PPC::CR6, PPC::CR7,
1515 return std::vector<unsigned>();
1518 // isOperandValidForConstraint
1519 bool PPCTargetLowering::
1520 isOperandValidForConstraint(SDOperand Op, char Letter) {
1531 if (!isa<ConstantSDNode>(Op)) return false; // Must be an immediate.
1532 unsigned Value = cast<ConstantSDNode>(Op)->getValue();
1534 default: assert(0 && "Unknown constraint letter!");
1535 case 'I': // "I" is a signed 16-bit constant.
1536 return (short)Value == (int)Value;
1537 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
1538 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
1539 return (short)Value == 0;
1540 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
1541 return (Value >> 16) == 0;
1542 case 'M': // "M" is a constant that is greater than 31.
1544 case 'N': // "N" is a positive constant that is an exact power of two.
1545 return (int)Value > 0 && isPowerOf2_32(Value);
1546 case 'O': // "O" is the constant zero.
1548 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
1549 return (short)-Value == (int)-Value;
1555 // Handle standard constraint letters.
1556 return TargetLowering::isOperandValidForConstraint(Op, Letter);
1559 /// isLegalAddressImmediate - Return true if the integer value can be used
1560 /// as the offset of the target addressing mode.
1561 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V) const {
1562 // PPC allows a sign-extended 16-bit immediate field.
1563 return (V > -(1 << 16) && V < (1 << 16)-1);