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 "PPCMachineFunctionInfo.h"
16 #include "PPCPredicates.h"
17 #include "PPCTargetMachine.h"
18 #include "PPCPerfectShuffle.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/VectorExtras.h"
21 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SSARegMap.h"
28 #include "llvm/Constants.h"
29 #include "llvm/Function.h"
30 #include "llvm/Intrinsics.h"
31 #include "llvm/Support/MathExtras.h"
32 #include "llvm/Target/TargetOptions.h"
33 #include "llvm/Support/CommandLine.h"
36 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
37 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
40 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
41 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
45 // Use _setjmp/_longjmp instead of setjmp/longjmp.
46 setUseUnderscoreSetJmp(true);
47 setUseUnderscoreLongJmp(true);
49 // Set up the register classes.
50 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
51 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
52 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
54 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
55 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
56 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Expand);
58 // PowerPC does not have truncstore for i1.
59 setStoreXAction(MVT::i1, Promote);
61 // PowerPC has pre-inc load and store's.
62 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
63 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
64 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
65 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
66 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
67 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
68 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
69 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
70 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
71 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
73 setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
74 setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
76 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
77 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
78 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
80 // PowerPC has no intrinsics for these particular operations
81 setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
82 setOperationAction(ISD::MEMSET, MVT::Other, Expand);
83 setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
85 // PowerPC has no SREM/UREM instructions
86 setOperationAction(ISD::SREM, MVT::i32, Expand);
87 setOperationAction(ISD::UREM, MVT::i32, Expand);
88 setOperationAction(ISD::SREM, MVT::i64, Expand);
89 setOperationAction(ISD::UREM, MVT::i64, Expand);
91 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
92 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
93 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
94 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
95 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
96 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
97 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
98 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
99 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
101 // We don't support sin/cos/sqrt/fmod
102 setOperationAction(ISD::FSIN , MVT::f64, Expand);
103 setOperationAction(ISD::FCOS , MVT::f64, Expand);
104 setOperationAction(ISD::FREM , MVT::f64, Expand);
105 setOperationAction(ISD::FSIN , MVT::f32, Expand);
106 setOperationAction(ISD::FCOS , MVT::f32, Expand);
107 setOperationAction(ISD::FREM , MVT::f32, Expand);
109 // If we're enabling GP optimizations, use hardware square root
110 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
111 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
112 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
115 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
116 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
118 // PowerPC does not have BSWAP, CTPOP or CTTZ
119 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
120 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
121 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
122 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
123 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
124 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
126 // PowerPC does not have ROTR
127 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
129 // PowerPC does not have Select
130 setOperationAction(ISD::SELECT, MVT::i32, Expand);
131 setOperationAction(ISD::SELECT, MVT::i64, Expand);
132 setOperationAction(ISD::SELECT, MVT::f32, Expand);
133 setOperationAction(ISD::SELECT, MVT::f64, Expand);
135 // PowerPC wants to turn select_cc of FP into fsel when possible.
136 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
137 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
139 // PowerPC wants to optimize integer setcc a bit
140 setOperationAction(ISD::SETCC, MVT::i32, Custom);
142 // PowerPC does not have BRCOND which requires SetCC
143 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
145 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
147 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
148 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
150 // PowerPC does not have [U|S]INT_TO_FP
151 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
152 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
154 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
155 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
156 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
157 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
159 // We cannot sextinreg(i1). Expand to shifts.
160 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
162 // Support label based line numbers.
163 setOperationAction(ISD::LOCATION, MVT::Other, Expand);
164 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
165 if (!TM.getSubtarget<PPCSubtarget>().isDarwin()) {
166 setOperationAction(ISD::LABEL, MVT::Other, Expand);
168 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
169 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
170 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
171 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
174 // We want to legalize GlobalAddress and ConstantPool nodes into the
175 // appropriate instructions to materialize the address.
176 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
177 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
178 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
179 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
180 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
181 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
182 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
183 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
185 // RET must be custom lowered, to meet ABI requirements
186 setOperationAction(ISD::RET , MVT::Other, Custom);
188 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
189 setOperationAction(ISD::VASTART , MVT::Other, Custom);
191 // VAARG is custom lowered with ELF 32 ABI
192 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
193 setOperationAction(ISD::VAARG, MVT::Other, Custom);
195 setOperationAction(ISD::VAARG, MVT::Other, Expand);
197 // Use the default implementation.
198 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
199 setOperationAction(ISD::VAEND , MVT::Other, Expand);
200 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
201 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
202 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
203 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
205 // We want to custom lower some of our intrinsics.
206 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
208 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
209 // They also have instructions for converting between i64 and fp.
210 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
211 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
212 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
213 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
214 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
216 // FIXME: disable this lowered code. This generates 64-bit register values,
217 // and we don't model the fact that the top part is clobbered by calls. We
218 // need to flag these together so that the value isn't live across a call.
219 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
221 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
222 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
224 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
225 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
228 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
229 // 64 bit PowerPC implementations can support i64 types directly
230 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
231 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
232 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
234 // 32 bit PowerPC wants to expand i64 shifts itself.
235 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
236 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
237 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
240 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
241 // First set operation action for all vector types to expand. Then we
242 // will selectively turn on ones that can be effectively codegen'd.
243 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
244 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
245 // add/sub are legal for all supported vector VT's.
246 setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
247 setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
249 // We promote all shuffles to v16i8.
250 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Promote);
251 AddPromotedToType (ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, MVT::v16i8);
253 // We promote all non-typed operations to v4i32.
254 setOperationAction(ISD::AND , (MVT::ValueType)VT, Promote);
255 AddPromotedToType (ISD::AND , (MVT::ValueType)VT, MVT::v4i32);
256 setOperationAction(ISD::OR , (MVT::ValueType)VT, Promote);
257 AddPromotedToType (ISD::OR , (MVT::ValueType)VT, MVT::v4i32);
258 setOperationAction(ISD::XOR , (MVT::ValueType)VT, Promote);
259 AddPromotedToType (ISD::XOR , (MVT::ValueType)VT, MVT::v4i32);
260 setOperationAction(ISD::LOAD , (MVT::ValueType)VT, Promote);
261 AddPromotedToType (ISD::LOAD , (MVT::ValueType)VT, MVT::v4i32);
262 setOperationAction(ISD::SELECT, (MVT::ValueType)VT, Promote);
263 AddPromotedToType (ISD::SELECT, (MVT::ValueType)VT, MVT::v4i32);
264 setOperationAction(ISD::STORE, (MVT::ValueType)VT, Promote);
265 AddPromotedToType (ISD::STORE, (MVT::ValueType)VT, MVT::v4i32);
267 // No other operations are legal.
268 setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
269 setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
270 setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
271 setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
272 setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
273 setOperationAction(ISD::FDIV, (MVT::ValueType)VT, Expand);
274 setOperationAction(ISD::FNEG, (MVT::ValueType)VT, Expand);
275 setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
276 setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
277 setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
278 setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
279 setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
280 setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
281 setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
283 setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
286 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
287 // with merges, splats, etc.
288 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
290 setOperationAction(ISD::AND , MVT::v4i32, Legal);
291 setOperationAction(ISD::OR , MVT::v4i32, Legal);
292 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
293 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
294 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
295 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
297 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
298 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
299 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
300 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
302 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
303 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
304 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
305 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
307 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
308 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
310 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
311 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
312 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
313 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
316 setSetCCResultType(MVT::i32);
317 setShiftAmountType(MVT::i32);
318 setSetCCResultContents(ZeroOrOneSetCCResult);
320 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
321 setStackPointerRegisterToSaveRestore(PPC::X1);
322 setExceptionPointerRegister(PPC::X3);
323 setExceptionSelectorRegister(PPC::X4);
325 setStackPointerRegisterToSaveRestore(PPC::R1);
326 setExceptionPointerRegister(PPC::R3);
327 setExceptionSelectorRegister(PPC::R4);
330 // We have target-specific dag combine patterns for the following nodes:
331 setTargetDAGCombine(ISD::SINT_TO_FP);
332 setTargetDAGCombine(ISD::STORE);
333 setTargetDAGCombine(ISD::BR_CC);
334 setTargetDAGCombine(ISD::BSWAP);
336 computeRegisterProperties();
339 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
342 case PPCISD::FSEL: return "PPCISD::FSEL";
343 case PPCISD::FCFID: return "PPCISD::FCFID";
344 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
345 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
346 case PPCISD::STFIWX: return "PPCISD::STFIWX";
347 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
348 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
349 case PPCISD::VPERM: return "PPCISD::VPERM";
350 case PPCISD::Hi: return "PPCISD::Hi";
351 case PPCISD::Lo: return "PPCISD::Lo";
352 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
353 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
354 case PPCISD::SRL: return "PPCISD::SRL";
355 case PPCISD::SRA: return "PPCISD::SRA";
356 case PPCISD::SHL: return "PPCISD::SHL";
357 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
358 case PPCISD::STD_32: return "PPCISD::STD_32";
359 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
360 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
361 case PPCISD::MTCTR: return "PPCISD::MTCTR";
362 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
363 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
364 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
365 case PPCISD::MFCR: return "PPCISD::MFCR";
366 case PPCISD::VCMP: return "PPCISD::VCMP";
367 case PPCISD::VCMPo: return "PPCISD::VCMPo";
368 case PPCISD::LBRX: return "PPCISD::LBRX";
369 case PPCISD::STBRX: return "PPCISD::STBRX";
370 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
374 //===----------------------------------------------------------------------===//
375 // Node matching predicates, for use by the tblgen matching code.
376 //===----------------------------------------------------------------------===//
378 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
379 static bool isFloatingPointZero(SDOperand Op) {
380 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
381 return CFP->getValueAPF().isZero();
382 else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
383 // Maybe this has already been legalized into the constant pool?
384 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
385 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
386 return CFP->getValueAPF().isZero();
391 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
392 /// true if Op is undef or if it matches the specified value.
393 static bool isConstantOrUndef(SDOperand Op, unsigned Val) {
394 return Op.getOpcode() == ISD::UNDEF ||
395 cast<ConstantSDNode>(Op)->getValue() == Val;
398 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
399 /// VPKUHUM instruction.
400 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
402 for (unsigned i = 0; i != 16; ++i)
403 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
406 for (unsigned i = 0; i != 8; ++i)
407 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
408 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
414 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
415 /// VPKUWUM instruction.
416 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
418 for (unsigned i = 0; i != 16; i += 2)
419 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
420 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
423 for (unsigned i = 0; i != 8; i += 2)
424 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
425 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
426 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
427 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
433 /// isVMerge - Common function, used to match vmrg* shuffles.
435 static bool isVMerge(SDNode *N, unsigned UnitSize,
436 unsigned LHSStart, unsigned RHSStart) {
437 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
438 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
439 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
440 "Unsupported merge size!");
442 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
443 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
444 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
445 LHSStart+j+i*UnitSize) ||
446 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
447 RHSStart+j+i*UnitSize))
453 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
454 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
455 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
457 return isVMerge(N, UnitSize, 8, 24);
458 return isVMerge(N, UnitSize, 8, 8);
461 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
462 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
463 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
465 return isVMerge(N, UnitSize, 0, 16);
466 return isVMerge(N, UnitSize, 0, 0);
470 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
471 /// amount, otherwise return -1.
472 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
473 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
474 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
475 // Find the first non-undef value in the shuffle mask.
477 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
480 if (i == 16) return -1; // all undef.
482 // Otherwise, check to see if the rest of the elements are consequtively
483 // numbered from this value.
484 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getValue();
485 if (ShiftAmt < i) return -1;
489 // Check the rest of the elements to see if they are consequtive.
490 for (++i; i != 16; ++i)
491 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
494 // Check the rest of the elements to see if they are consequtive.
495 for (++i; i != 16; ++i)
496 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
503 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
504 /// specifies a splat of a single element that is suitable for input to
505 /// VSPLTB/VSPLTH/VSPLTW.
506 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
507 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
508 N->getNumOperands() == 16 &&
509 (EltSize == 1 || EltSize == 2 || EltSize == 4));
511 // This is a splat operation if each element of the permute is the same, and
512 // if the value doesn't reference the second vector.
513 unsigned ElementBase = 0;
514 SDOperand Elt = N->getOperand(0);
515 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
516 ElementBase = EltV->getValue();
518 return false; // FIXME: Handle UNDEF elements too!
520 if (cast<ConstantSDNode>(Elt)->getValue() >= 16)
523 // Check that they are consequtive.
524 for (unsigned i = 1; i != EltSize; ++i) {
525 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
526 cast<ConstantSDNode>(N->getOperand(i))->getValue() != i+ElementBase)
530 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
531 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
532 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
533 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
534 "Invalid VECTOR_SHUFFLE mask!");
535 for (unsigned j = 0; j != EltSize; ++j)
536 if (N->getOperand(i+j) != N->getOperand(j))
543 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
545 bool PPC::isAllNegativeZeroVector(SDNode *N) {
546 assert(N->getOpcode() == ISD::BUILD_VECTOR);
547 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
548 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
549 return CFP->getValueAPF().isNegZero();
553 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
554 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
555 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
556 assert(isSplatShuffleMask(N, EltSize));
557 return cast<ConstantSDNode>(N->getOperand(0))->getValue() / EltSize;
560 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
561 /// by using a vspltis[bhw] instruction of the specified element size, return
562 /// the constant being splatted. The ByteSize field indicates the number of
563 /// bytes of each element [124] -> [bhw].
564 SDOperand PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
565 SDOperand OpVal(0, 0);
567 // If ByteSize of the splat is bigger than the element size of the
568 // build_vector, then we have a case where we are checking for a splat where
569 // multiple elements of the buildvector are folded together into a single
570 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
571 unsigned EltSize = 16/N->getNumOperands();
572 if (EltSize < ByteSize) {
573 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
574 SDOperand UniquedVals[4];
575 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
577 // See if all of the elements in the buildvector agree across.
578 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
579 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
580 // If the element isn't a constant, bail fully out.
581 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDOperand();
584 if (UniquedVals[i&(Multiple-1)].Val == 0)
585 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
586 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
587 return SDOperand(); // no match.
590 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
591 // either constant or undef values that are identical for each chunk. See
592 // if these chunks can form into a larger vspltis*.
594 // Check to see if all of the leading entries are either 0 or -1. If
595 // neither, then this won't fit into the immediate field.
596 bool LeadingZero = true;
597 bool LeadingOnes = true;
598 for (unsigned i = 0; i != Multiple-1; ++i) {
599 if (UniquedVals[i].Val == 0) continue; // Must have been undefs.
601 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
602 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
604 // Finally, check the least significant entry.
606 if (UniquedVals[Multiple-1].Val == 0)
607 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
608 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getValue();
610 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
613 if (UniquedVals[Multiple-1].Val == 0)
614 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
615 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
616 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
617 return DAG.getTargetConstant(Val, MVT::i32);
623 // Check to see if this buildvec has a single non-undef value in its elements.
624 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
625 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
627 OpVal = N->getOperand(i);
628 else if (OpVal != N->getOperand(i))
632 if (OpVal.Val == 0) return SDOperand(); // All UNDEF: use implicit def.
634 unsigned ValSizeInBytes = 0;
636 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
637 Value = CN->getValue();
638 ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
639 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
640 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
641 Value = FloatToBits(CN->getValueAPF().convertToFloat());
645 // If the splat value is larger than the element value, then we can never do
646 // this splat. The only case that we could fit the replicated bits into our
647 // immediate field for would be zero, and we prefer to use vxor for it.
648 if (ValSizeInBytes < ByteSize) return SDOperand();
650 // If the element value is larger than the splat value, cut it in half and
651 // check to see if the two halves are equal. Continue doing this until we
652 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
653 while (ValSizeInBytes > ByteSize) {
654 ValSizeInBytes >>= 1;
656 // If the top half equals the bottom half, we're still ok.
657 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
658 (Value & ((1 << (8*ValSizeInBytes))-1)))
662 // Properly sign extend the value.
663 int ShAmt = (4-ByteSize)*8;
664 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
666 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
667 if (MaskVal == 0) return SDOperand();
669 // Finally, if this value fits in a 5 bit sext field, return it
670 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
671 return DAG.getTargetConstant(MaskVal, MVT::i32);
675 //===----------------------------------------------------------------------===//
676 // Addressing Mode Selection
677 //===----------------------------------------------------------------------===//
679 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
680 /// or 64-bit immediate, and if the value can be accurately represented as a
681 /// sign extension from a 16-bit value. If so, this returns true and the
683 static bool isIntS16Immediate(SDNode *N, short &Imm) {
684 if (N->getOpcode() != ISD::Constant)
687 Imm = (short)cast<ConstantSDNode>(N)->getValue();
688 if (N->getValueType(0) == MVT::i32)
689 return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
691 return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
693 static bool isIntS16Immediate(SDOperand Op, short &Imm) {
694 return isIntS16Immediate(Op.Val, Imm);
698 /// SelectAddressRegReg - Given the specified addressed, check to see if it
699 /// can be represented as an indexed [r+r] operation. Returns false if it
700 /// can be more efficiently represented with [r+imm].
701 bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
705 if (N.getOpcode() == ISD::ADD) {
706 if (isIntS16Immediate(N.getOperand(1), imm))
708 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
711 Base = N.getOperand(0);
712 Index = N.getOperand(1);
714 } else if (N.getOpcode() == ISD::OR) {
715 if (isIntS16Immediate(N.getOperand(1), imm))
716 return false; // r+i can fold it if we can.
718 // If this is an or of disjoint bitfields, we can codegen this as an add
719 // (for better address arithmetic) if the LHS and RHS of the OR are provably
721 uint64_t LHSKnownZero, LHSKnownOne;
722 uint64_t RHSKnownZero, RHSKnownOne;
723 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
726 DAG.ComputeMaskedBits(N.getOperand(1), ~0U, RHSKnownZero, RHSKnownOne);
727 // If all of the bits are known zero on the LHS or RHS, the add won't
729 if ((LHSKnownZero | RHSKnownZero) == ~0U) {
730 Base = N.getOperand(0);
731 Index = N.getOperand(1);
740 /// Returns true if the address N can be represented by a base register plus
741 /// a signed 16-bit displacement [r+imm], and if it is not better
742 /// represented as reg+reg.
743 bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
744 SDOperand &Base, SelectionDAG &DAG){
745 // If this can be more profitably realized as r+r, fail.
746 if (SelectAddressRegReg(N, Disp, Base, DAG))
749 if (N.getOpcode() == ISD::ADD) {
751 if (isIntS16Immediate(N.getOperand(1), imm)) {
752 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
753 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
754 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
756 Base = N.getOperand(0);
758 return true; // [r+i]
759 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
760 // Match LOAD (ADD (X, Lo(G))).
761 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
762 && "Cannot handle constant offsets yet!");
763 Disp = N.getOperand(1).getOperand(0); // The global address.
764 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
765 Disp.getOpcode() == ISD::TargetConstantPool ||
766 Disp.getOpcode() == ISD::TargetJumpTable);
767 Base = N.getOperand(0);
768 return true; // [&g+r]
770 } else if (N.getOpcode() == ISD::OR) {
772 if (isIntS16Immediate(N.getOperand(1), imm)) {
773 // If this is an or of disjoint bitfields, we can codegen this as an add
774 // (for better address arithmetic) if the LHS and RHS of the OR are
775 // provably disjoint.
776 uint64_t LHSKnownZero, LHSKnownOne;
777 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
778 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
779 // If all of the bits are known zero on the LHS or RHS, the add won't
781 Base = N.getOperand(0);
782 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
786 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
787 // Loading from a constant address.
789 // If this address fits entirely in a 16-bit sext immediate field, codegen
792 if (isIntS16Immediate(CN, Imm)) {
793 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
794 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
798 // Handle 32-bit sext immediates with LIS + addr mode.
799 if (CN->getValueType(0) == MVT::i32 ||
800 (int64_t)CN->getValue() == (int)CN->getValue()) {
801 int Addr = (int)CN->getValue();
803 // Otherwise, break this down into an LIS + disp.
804 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
806 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
807 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
808 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
813 Disp = DAG.getTargetConstant(0, getPointerTy());
814 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
815 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
818 return true; // [r+0]
821 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
822 /// represented as an indexed [r+r] operation.
823 bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
826 // Check to see if we can easily represent this as an [r+r] address. This
827 // will fail if it thinks that the address is more profitably represented as
828 // reg+imm, e.g. where imm = 0.
829 if (SelectAddressRegReg(N, Base, Index, DAG))
832 // If the operand is an addition, always emit this as [r+r], since this is
833 // better (for code size, and execution, as the memop does the add for free)
834 // than emitting an explicit add.
835 if (N.getOpcode() == ISD::ADD) {
836 Base = N.getOperand(0);
837 Index = N.getOperand(1);
841 // Otherwise, do it the hard way, using R0 as the base register.
842 Base = DAG.getRegister(PPC::R0, N.getValueType());
847 /// SelectAddressRegImmShift - Returns true if the address N can be
848 /// represented by a base register plus a signed 14-bit displacement
849 /// [r+imm*4]. Suitable for use by STD and friends.
850 bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
853 // If this can be more profitably realized as r+r, fail.
854 if (SelectAddressRegReg(N, Disp, Base, DAG))
857 if (N.getOpcode() == ISD::ADD) {
859 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
860 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
861 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
862 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
864 Base = N.getOperand(0);
866 return true; // [r+i]
867 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
868 // Match LOAD (ADD (X, Lo(G))).
869 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
870 && "Cannot handle constant offsets yet!");
871 Disp = N.getOperand(1).getOperand(0); // The global address.
872 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
873 Disp.getOpcode() == ISD::TargetConstantPool ||
874 Disp.getOpcode() == ISD::TargetJumpTable);
875 Base = N.getOperand(0);
876 return true; // [&g+r]
878 } else if (N.getOpcode() == ISD::OR) {
880 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
881 // If this is an or of disjoint bitfields, we can codegen this as an add
882 // (for better address arithmetic) if the LHS and RHS of the OR are
883 // provably disjoint.
884 uint64_t LHSKnownZero, LHSKnownOne;
885 DAG.ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
886 if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
887 // If all of the bits are known zero on the LHS or RHS, the add won't
889 Base = N.getOperand(0);
890 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
894 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
895 // Loading from a constant address. Verify low two bits are clear.
896 if ((CN->getValue() & 3) == 0) {
897 // If this address fits entirely in a 14-bit sext immediate field, codegen
900 if (isIntS16Immediate(CN, Imm)) {
901 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
902 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
906 // Fold the low-part of 32-bit absolute addresses into addr mode.
907 if (CN->getValueType(0) == MVT::i32 ||
908 (int64_t)CN->getValue() == (int)CN->getValue()) {
909 int Addr = (int)CN->getValue();
911 // Otherwise, break this down into an LIS + disp.
912 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
914 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
915 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
916 Base = SDOperand(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
922 Disp = DAG.getTargetConstant(0, getPointerTy());
923 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
924 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
927 return true; // [r+0]
931 /// getPreIndexedAddressParts - returns true by value, base pointer and
932 /// offset pointer and addressing mode by reference if the node's address
933 /// can be legally represented as pre-indexed load / store address.
934 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
936 ISD::MemIndexedMode &AM,
938 // Disabled by default for now.
939 if (!EnablePPCPreinc) return false;
943 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
944 Ptr = LD->getBasePtr();
945 VT = LD->getLoadedVT();
947 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
949 Ptr = ST->getBasePtr();
950 VT = ST->getStoredVT();
954 // PowerPC doesn't have preinc load/store instructions for vectors.
955 if (MVT::isVector(VT))
958 // TODO: Check reg+reg first.
960 // LDU/STU use reg+imm*4, others use reg+imm.
961 if (VT != MVT::i64) {
963 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
967 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
971 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
972 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
973 // sext i32 to i64 when addr mode is r+i.
974 if (LD->getValueType(0) == MVT::i64 && LD->getLoadedVT() == MVT::i32 &&
975 LD->getExtensionType() == ISD::SEXTLOAD &&
976 isa<ConstantSDNode>(Offset))
984 //===----------------------------------------------------------------------===//
985 // LowerOperation implementation
986 //===----------------------------------------------------------------------===//
988 static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
989 MVT::ValueType PtrVT = Op.getValueType();
990 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
991 Constant *C = CP->getConstVal();
992 SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
993 SDOperand Zero = DAG.getConstant(0, PtrVT);
995 const TargetMachine &TM = DAG.getTarget();
997 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
998 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1000 // If this is a non-darwin platform, we don't support non-static relo models
1002 if (TM.getRelocationModel() == Reloc::Static ||
1003 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1004 // Generate non-pic code that has direct accesses to the constant pool.
1005 // The address of the global is just (hi(&g)+lo(&g)).
1006 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1009 if (TM.getRelocationModel() == Reloc::PIC_) {
1010 // With PIC, the first instruction is actually "GR+hi(&G)".
1011 Hi = DAG.getNode(ISD::ADD, PtrVT,
1012 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1015 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1019 static SDOperand LowerJumpTable(SDOperand Op, SelectionDAG &DAG) {
1020 MVT::ValueType PtrVT = Op.getValueType();
1021 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1022 SDOperand JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1023 SDOperand Zero = DAG.getConstant(0, PtrVT);
1025 const TargetMachine &TM = DAG.getTarget();
1027 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1028 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1030 // If this is a non-darwin platform, we don't support non-static relo models
1032 if (TM.getRelocationModel() == Reloc::Static ||
1033 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1034 // Generate non-pic code that has direct accesses to the constant pool.
1035 // The address of the global is just (hi(&g)+lo(&g)).
1036 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1039 if (TM.getRelocationModel() == Reloc::PIC_) {
1040 // With PIC, the first instruction is actually "GR+hi(&G)".
1041 Hi = DAG.getNode(ISD::ADD, PtrVT,
1042 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1045 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1049 static SDOperand LowerGlobalTLSAddress(SDOperand Op, SelectionDAG &DAG) {
1050 assert(0 && "TLS not implemented for PPC.");
1053 static SDOperand LowerGlobalAddress(SDOperand Op, SelectionDAG &DAG) {
1054 MVT::ValueType PtrVT = Op.getValueType();
1055 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1056 GlobalValue *GV = GSDN->getGlobal();
1057 SDOperand GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1058 SDOperand Zero = DAG.getConstant(0, PtrVT);
1060 const TargetMachine &TM = DAG.getTarget();
1062 SDOperand Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1063 SDOperand Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1065 // If this is a non-darwin platform, we don't support non-static relo models
1067 if (TM.getRelocationModel() == Reloc::Static ||
1068 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1069 // Generate non-pic code that has direct accesses to globals.
1070 // The address of the global is just (hi(&g)+lo(&g)).
1071 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1074 if (TM.getRelocationModel() == Reloc::PIC_) {
1075 // With PIC, the first instruction is actually "GR+hi(&G)".
1076 Hi = DAG.getNode(ISD::ADD, PtrVT,
1077 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1080 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1082 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1085 // If the global is weak or external, we have to go through the lazy
1087 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1090 static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
1091 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1093 // If we're comparing for equality to zero, expose the fact that this is
1094 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1095 // fold the new nodes.
1096 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1097 if (C->isNullValue() && CC == ISD::SETEQ) {
1098 MVT::ValueType VT = Op.getOperand(0).getValueType();
1099 SDOperand Zext = Op.getOperand(0);
1100 if (VT < MVT::i32) {
1102 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1104 unsigned Log2b = Log2_32(MVT::getSizeInBits(VT));
1105 SDOperand Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1106 SDOperand Scc = DAG.getNode(ISD::SRL, VT, Clz,
1107 DAG.getConstant(Log2b, MVT::i32));
1108 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1110 // Leave comparisons against 0 and -1 alone for now, since they're usually
1111 // optimized. FIXME: revisit this when we can custom lower all setcc
1113 if (C->isAllOnesValue() || C->isNullValue())
1117 // If we have an integer seteq/setne, turn it into a compare against zero
1118 // by xor'ing the rhs with the lhs, which is faster than setting a
1119 // condition register, reading it back out, and masking the correct bit. The
1120 // normal approach here uses sub to do this instead of xor. Using xor exposes
1121 // the result to other bit-twiddling opportunities.
1122 MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
1123 if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1124 MVT::ValueType VT = Op.getValueType();
1125 SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1127 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1132 static SDOperand LowerVAARG(SDOperand Op, SelectionDAG &DAG,
1133 int VarArgsFrameIndex,
1134 int VarArgsStackOffset,
1135 unsigned VarArgsNumGPR,
1136 unsigned VarArgsNumFPR,
1137 const PPCSubtarget &Subtarget) {
1139 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1142 static SDOperand LowerVASTART(SDOperand Op, SelectionDAG &DAG,
1143 int VarArgsFrameIndex,
1144 int VarArgsStackOffset,
1145 unsigned VarArgsNumGPR,
1146 unsigned VarArgsNumFPR,
1147 const PPCSubtarget &Subtarget) {
1149 if (Subtarget.isMachoABI()) {
1150 // vastart just stores the address of the VarArgsFrameIndex slot into the
1151 // memory location argument.
1152 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1153 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1154 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
1155 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV->getValue(),
1159 // For ELF 32 ABI we follow the layout of the va_list struct.
1160 // We suppose the given va_list is already allocated.
1163 // char gpr; /* index into the array of 8 GPRs
1164 // * stored in the register save area
1165 // * gpr=0 corresponds to r3,
1166 // * gpr=1 to r4, etc.
1168 // char fpr; /* index into the array of 8 FPRs
1169 // * stored in the register save area
1170 // * fpr=0 corresponds to f1,
1171 // * fpr=1 to f2, etc.
1173 // char *overflow_arg_area;
1174 // /* location on stack that holds
1175 // * the next overflow argument
1177 // char *reg_save_area;
1178 // /* where r3:r10 and f1:f8 (if saved)
1184 SDOperand ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1185 SDOperand ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1188 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1190 SDOperand StackOffset = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1191 SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1193 SDOperand ConstFrameOffset = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8,
1195 SDOperand ConstStackOffset = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8 - 1,
1197 SDOperand ConstFPROffset = DAG.getConstant(1, PtrVT);
1199 SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
1201 // Store first byte : number of int regs
1202 SDOperand firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1203 Op.getOperand(1), SV->getValue(),
1205 SDOperand nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1208 // Store second byte : number of float regs
1209 SDOperand secondStore = DAG.getStore(firstStore, ArgFPR, nextPtr,
1210 SV->getValue(), SV->getOffset());
1211 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1213 // Store second word : arguments given on stack
1214 SDOperand thirdStore = DAG.getStore(secondStore, StackOffset, nextPtr,
1215 SV->getValue(), SV->getOffset());
1216 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1218 // Store third word : arguments given in registers
1219 return DAG.getStore(thirdStore, FR, nextPtr, SV->getValue(),
1224 #include "PPCGenCallingConv.inc"
1226 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1227 /// depending on which subtarget is selected.
1228 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1229 if (Subtarget.isMachoABI()) {
1230 static const unsigned FPR[] = {
1231 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1232 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1238 static const unsigned FPR[] = {
1239 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1245 static SDOperand LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG,
1246 int &VarArgsFrameIndex,
1247 int &VarArgsStackOffset,
1248 unsigned &VarArgsNumGPR,
1249 unsigned &VarArgsNumFPR,
1250 const PPCSubtarget &Subtarget) {
1251 // TODO: add description of PPC stack frame format, or at least some docs.
1253 MachineFunction &MF = DAG.getMachineFunction();
1254 MachineFrameInfo *MFI = MF.getFrameInfo();
1255 SSARegMap *RegMap = MF.getSSARegMap();
1256 SmallVector<SDOperand, 8> ArgValues;
1257 SDOperand Root = Op.getOperand(0);
1259 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1260 bool isPPC64 = PtrVT == MVT::i64;
1261 bool isMachoABI = Subtarget.isMachoABI();
1262 bool isELF32_ABI = Subtarget.isELF32_ABI();
1263 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1265 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1267 static const unsigned GPR_32[] = { // 32-bit registers.
1268 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1269 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1271 static const unsigned GPR_64[] = { // 64-bit registers.
1272 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1273 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1276 static const unsigned *FPR = GetFPR(Subtarget);
1278 static const unsigned VR[] = {
1279 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1280 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1283 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1284 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1285 const unsigned Num_VR_Regs = array_lengthof( VR);
1287 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1289 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1291 // Add DAG nodes to load the arguments or copy them out of registers. On
1292 // entry to a function on PPC, the arguments start after the linkage area,
1293 // although the first ones are often in registers.
1295 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1296 // represented with two words (long long or double) must be copied to an
1297 // even GPR_idx value or to an even ArgOffset value.
1299 for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
1301 bool needsLoad = false;
1302 MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
1303 unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
1304 unsigned ArgSize = ObjSize;
1305 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(ArgNo+3))->getValue();
1306 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1307 // See if next argument requires stack alignment in ELF
1308 bool Expand = (ObjectVT == MVT::f64) || ((ArgNo + 1 < e) &&
1309 (cast<ConstantSDNode>(Op.getOperand(ArgNo+4))->getValue() & AlignFlag) &&
1310 (!(Flags & AlignFlag)));
1312 unsigned CurArgOffset = ArgOffset;
1314 default: assert(0 && "Unhandled argument type!");
1316 // Double word align in ELF
1317 if (Expand && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1318 if (GPR_idx != Num_GPR_Regs) {
1319 unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1320 MF.addLiveIn(GPR[GPR_idx], VReg);
1321 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1325 ArgSize = PtrByteSize;
1327 // Stack align in ELF
1328 if (needsLoad && Expand && isELF32_ABI)
1329 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1330 // All int arguments reserve stack space in Macho ABI.
1331 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1334 case MVT::i64: // PPC64
1335 if (GPR_idx != Num_GPR_Regs) {
1336 unsigned VReg = RegMap->createVirtualRegister(&PPC::G8RCRegClass);
1337 MF.addLiveIn(GPR[GPR_idx], VReg);
1338 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1343 // All int arguments reserve stack space in Macho ABI.
1344 if (isMachoABI || needsLoad) ArgOffset += 8;
1349 // Every 4 bytes of argument space consumes one of the GPRs available for
1350 // argument passing.
1351 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1353 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1356 if (FPR_idx != Num_FPR_Regs) {
1358 if (ObjectVT == MVT::f32)
1359 VReg = RegMap->createVirtualRegister(&PPC::F4RCRegClass);
1361 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1362 MF.addLiveIn(FPR[FPR_idx], VReg);
1363 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1369 // Stack align in ELF
1370 if (needsLoad && Expand && isELF32_ABI)
1371 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1372 // All FP arguments reserve stack space in Macho ABI.
1373 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1379 // Note that vector arguments in registers don't reserve stack space.
1380 if (VR_idx != Num_VR_Regs) {
1381 unsigned VReg = RegMap->createVirtualRegister(&PPC::VRRCRegClass);
1382 MF.addLiveIn(VR[VR_idx], VReg);
1383 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1386 // This should be simple, but requires getting 16-byte aligned stack
1388 assert(0 && "Loading VR argument not implemented yet!");
1394 // We need to load the argument to a virtual register if we determined above
1395 // that we ran out of physical registers of the appropriate type
1397 // If the argument is actually used, emit a load from the right stack
1399 if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
1400 int FI = MFI->CreateFixedObject(ObjSize,
1401 CurArgOffset + (ArgSize - ObjSize));
1402 SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
1403 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1405 // Don't emit a dead load.
1406 ArgVal = DAG.getNode(ISD::UNDEF, ObjectVT);
1410 ArgValues.push_back(ArgVal);
1413 // If the function takes variable number of arguments, make a frame index for
1414 // the start of the first vararg value... for expansion of llvm.va_start.
1415 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1420 VarArgsNumGPR = GPR_idx;
1421 VarArgsNumFPR = FPR_idx;
1423 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1425 depth = -(Num_GPR_Regs * MVT::getSizeInBits(PtrVT)/8 +
1426 Num_FPR_Regs * MVT::getSizeInBits(MVT::f64)/8 +
1427 MVT::getSizeInBits(PtrVT)/8);
1429 VarArgsStackOffset = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1436 VarArgsFrameIndex = MFI->CreateFixedObject(MVT::getSizeInBits(PtrVT)/8,
1438 SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1440 SmallVector<SDOperand, 8> MemOps;
1442 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1443 // stored to the VarArgsFrameIndex on the stack.
1445 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1446 SDOperand Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1447 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1448 MemOps.push_back(Store);
1449 // Increment the address by four for the next argument to store
1450 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1451 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1455 // If this function is vararg, store any remaining integer argument regs
1456 // to their spots on the stack so that they may be loaded by deferencing the
1457 // result of va_next.
1458 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1461 VReg = RegMap->createVirtualRegister(&PPC::G8RCRegClass);
1463 VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
1465 MF.addLiveIn(GPR[GPR_idx], VReg);
1466 SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1467 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1468 MemOps.push_back(Store);
1469 // Increment the address by four for the next argument to store
1470 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
1471 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1474 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1477 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1478 SDOperand Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1479 SDOperand Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1480 MemOps.push_back(Store);
1481 // Increment the address by eight for the next argument to store
1482 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1484 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1487 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1489 VReg = RegMap->createVirtualRegister(&PPC::F8RCRegClass);
1491 MF.addLiveIn(FPR[FPR_idx], VReg);
1492 SDOperand Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1493 SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1494 MemOps.push_back(Store);
1495 // Increment the address by eight for the next argument to store
1496 SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(MVT::f64)/8,
1498 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1502 if (!MemOps.empty())
1503 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1506 ArgValues.push_back(Root);
1508 // Return the new list of results.
1509 std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
1510 Op.Val->value_end());
1511 return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
1514 /// isCallCompatibleAddress - Return the immediate to use if the specified
1515 /// 32-bit value is representable in the immediate field of a BxA instruction.
1516 static SDNode *isBLACompatibleAddress(SDOperand Op, SelectionDAG &DAG) {
1517 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1520 int Addr = C->getValue();
1521 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1522 (Addr << 6 >> 6) != Addr)
1523 return 0; // Top 6 bits have to be sext of immediate.
1525 return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).Val;
1529 static SDOperand LowerCALL(SDOperand Op, SelectionDAG &DAG,
1530 const PPCSubtarget &Subtarget) {
1531 SDOperand Chain = Op.getOperand(0);
1532 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
1533 SDOperand Callee = Op.getOperand(4);
1534 unsigned NumOps = (Op.getNumOperands() - 5) / 2;
1536 bool isMachoABI = Subtarget.isMachoABI();
1537 bool isELF32_ABI = Subtarget.isELF32_ABI();
1539 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1540 bool isPPC64 = PtrVT == MVT::i64;
1541 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1543 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
1544 // SelectExpr to use to put the arguments in the appropriate registers.
1545 std::vector<SDOperand> args_to_use;
1547 // Count how many bytes are to be pushed on the stack, including the linkage
1548 // area, and parameter passing area. We start with 24/48 bytes, which is
1549 // prereserved space for [SP][CR][LR][3 x unused].
1550 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1552 // Add up all the space actually used.
1553 for (unsigned i = 0; i != NumOps; ++i) {
1554 unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
1555 ArgSize = std::max(ArgSize, PtrByteSize);
1556 NumBytes += ArgSize;
1559 // The prolog code of the callee may store up to 8 GPR argument registers to
1560 // the stack, allowing va_start to index over them in memory if its varargs.
1561 // Because we cannot tell if this is needed on the caller side, we have to
1562 // conservatively assume that it is needed. As such, make sure we have at
1563 // least enough stack space for the caller to store the 8 GPRs.
1564 NumBytes = std::max(NumBytes,
1565 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1567 // Adjust the stack pointer for the new arguments...
1568 // These operations are automatically eliminated by the prolog/epilog pass
1569 Chain = DAG.getCALLSEQ_START(Chain,
1570 DAG.getConstant(NumBytes, PtrVT));
1572 // Set up a copy of the stack pointer for use loading and storing any
1573 // arguments that may not fit in the registers available for argument
1577 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
1579 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
1581 // Figure out which arguments are going to go in registers, and which in
1582 // memory. Also, if this is a vararg function, floating point operations
1583 // must be stored to our stack, and loaded into integer regs as well, if
1584 // any integer regs are available for argument passing.
1585 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1586 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1588 static const unsigned GPR_32[] = { // 32-bit registers.
1589 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1590 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1592 static const unsigned GPR_64[] = { // 64-bit registers.
1593 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1594 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1596 static const unsigned *FPR = GetFPR(Subtarget);
1598 static const unsigned VR[] = {
1599 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1600 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1602 const unsigned NumGPRs = array_lengthof(GPR_32);
1603 const unsigned NumFPRs = isMachoABI ? 13 : 8;
1604 const unsigned NumVRs = array_lengthof( VR);
1606 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1608 std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
1609 SmallVector<SDOperand, 8> MemOpChains;
1610 for (unsigned i = 0; i != NumOps; ++i) {
1612 SDOperand Arg = Op.getOperand(5+2*i);
1613 unsigned Flags = cast<ConstantSDNode>(Op.getOperand(5+2*i+1))->getValue();
1614 unsigned AlignFlag = 1 << ISD::ParamFlags::OrigAlignmentOffs;
1615 // See if next argument requires stack alignment in ELF
1616 unsigned next = 5+2*(i+1)+1;
1617 bool Expand = (Arg.getValueType() == MVT::f64) || ((i + 1 < NumOps) &&
1618 (cast<ConstantSDNode>(Op.getOperand(next))->getValue() & AlignFlag) &&
1619 (!(Flags & AlignFlag)));
1621 // PtrOff will be used to store the current argument to the stack if a
1622 // register cannot be found for it.
1625 // Stack align in ELF 32
1626 if (isELF32_ABI && Expand)
1627 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
1628 StackPtr.getValueType());
1630 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
1632 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
1634 // On PPC64, promote integers to 64-bit values.
1635 if (isPPC64 && Arg.getValueType() == MVT::i32) {
1636 unsigned ExtOp = (Flags & 1) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
1638 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
1641 switch (Arg.getValueType()) {
1642 default: assert(0 && "Unexpected ValueType for argument!");
1645 // Double word align in ELF
1646 if (isELF32_ABI && Expand) GPR_idx += (GPR_idx % 2);
1647 if (GPR_idx != NumGPRs) {
1648 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
1650 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1653 if (inMem || isMachoABI) {
1654 // Stack align in ELF
1655 if (isELF32_ABI && Expand)
1656 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1658 ArgOffset += PtrByteSize;
1664 // Float varargs need to be promoted to double.
1665 if (Arg.getValueType() == MVT::f32)
1666 Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
1669 if (FPR_idx != NumFPRs) {
1670 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
1673 SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
1674 MemOpChains.push_back(Store);
1676 // Float varargs are always shadowed in available integer registers
1677 if (GPR_idx != NumGPRs) {
1678 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1679 MemOpChains.push_back(Load.getValue(1));
1680 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1683 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
1684 SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
1685 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
1686 SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
1687 MemOpChains.push_back(Load.getValue(1));
1688 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
1692 // If we have any FPRs remaining, we may also have GPRs remaining.
1693 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
1696 if (GPR_idx != NumGPRs)
1698 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
1699 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
1704 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
1707 if (inMem || isMachoABI) {
1708 // Stack align in ELF
1709 if (isELF32_ABI && Expand)
1710 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1714 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
1721 assert(!isVarArg && "Don't support passing vectors to varargs yet!");
1722 assert(VR_idx != NumVRs &&
1723 "Don't support passing more than 12 vector args yet!");
1724 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
1728 if (!MemOpChains.empty())
1729 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
1730 &MemOpChains[0], MemOpChains.size());
1732 // Build a sequence of copy-to-reg nodes chained together with token chain
1733 // and flag operands which copy the outgoing args into the appropriate regs.
1735 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1736 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
1738 InFlag = Chain.getValue(1);
1741 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
1742 if (isVarArg && isELF32_ABI) {
1743 SDOperand SetCR(DAG.getTargetNode(PPC::SETCR, MVT::i32), 0);
1744 Chain = DAG.getCopyToReg(Chain, PPC::CR6, SetCR, InFlag);
1745 InFlag = Chain.getValue(1);
1748 std::vector<MVT::ValueType> NodeTys;
1749 NodeTys.push_back(MVT::Other); // Returns a chain
1750 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
1752 SmallVector<SDOperand, 8> Ops;
1753 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
1755 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1756 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1757 // node so that legalize doesn't hack it.
1758 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1759 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
1760 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
1761 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
1762 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
1763 // If this is an absolute destination address, use the munged value.
1764 Callee = SDOperand(Dest, 0);
1766 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
1767 // to do the call, we can't use PPCISD::CALL.
1768 SDOperand MTCTROps[] = {Chain, Callee, InFlag};
1769 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
1770 InFlag = Chain.getValue(1);
1772 // Copy the callee address into R12 on darwin.
1774 Chain = DAG.getCopyToReg(Chain, PPC::R12, Callee, InFlag);
1775 InFlag = Chain.getValue(1);
1779 NodeTys.push_back(MVT::Other);
1780 NodeTys.push_back(MVT::Flag);
1781 Ops.push_back(Chain);
1782 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
1786 // If this is a direct call, pass the chain and the callee.
1788 Ops.push_back(Chain);
1789 Ops.push_back(Callee);
1792 // Add argument registers to the end of the list so that they are known live
1794 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1795 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1796 RegsToPass[i].second.getValueType()));
1799 Ops.push_back(InFlag);
1800 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
1801 InFlag = Chain.getValue(1);
1803 SDOperand ResultVals[3];
1804 unsigned NumResults = 0;
1807 // If the call has results, copy the values out of the ret val registers.
1808 switch (Op.Val->getValueType(0)) {
1809 default: assert(0 && "Unexpected ret value!");
1810 case MVT::Other: break;
1812 if (Op.Val->getValueType(1) == MVT::i32) {
1813 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1814 ResultVals[0] = Chain.getValue(0);
1815 Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32,
1816 Chain.getValue(2)).getValue(1);
1817 ResultVals[1] = Chain.getValue(0);
1819 NodeTys.push_back(MVT::i32);
1821 Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
1822 ResultVals[0] = Chain.getValue(0);
1825 NodeTys.push_back(MVT::i32);
1828 Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
1829 ResultVals[0] = Chain.getValue(0);
1831 NodeTys.push_back(MVT::i64);
1834 if (Op.Val->getValueType(1) == MVT::f64) {
1835 Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
1836 ResultVals[0] = Chain.getValue(0);
1837 Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
1838 Chain.getValue(2)).getValue(1);
1839 ResultVals[1] = Chain.getValue(0);
1841 NodeTys.push_back(MVT::f64);
1842 NodeTys.push_back(MVT::f64);
1845 // else fall through
1847 Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
1848 InFlag).getValue(1);
1849 ResultVals[0] = Chain.getValue(0);
1851 NodeTys.push_back(Op.Val->getValueType(0));
1857 Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
1858 InFlag).getValue(1);
1859 ResultVals[0] = Chain.getValue(0);
1861 NodeTys.push_back(Op.Val->getValueType(0));
1865 Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
1866 DAG.getConstant(NumBytes, PtrVT));
1867 NodeTys.push_back(MVT::Other);
1869 // If the function returns void, just return the chain.
1870 if (NumResults == 0)
1873 // Otherwise, merge everything together with a MERGE_VALUES node.
1874 ResultVals[NumResults++] = Chain;
1875 SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
1876 ResultVals, NumResults);
1877 return Res.getValue(Op.ResNo);
1880 static SDOperand LowerRET(SDOperand Op, SelectionDAG &DAG, TargetMachine &TM) {
1881 SmallVector<CCValAssign, 16> RVLocs;
1882 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1883 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1884 CCState CCInfo(CC, isVarArg, TM, RVLocs);
1885 CCInfo.AnalyzeReturn(Op.Val, RetCC_PPC);
1887 // If this is the first return lowered for this function, add the regs to the
1888 // liveout set for the function.
1889 if (DAG.getMachineFunction().liveout_empty()) {
1890 for (unsigned i = 0; i != RVLocs.size(); ++i)
1891 DAG.getMachineFunction().addLiveOut(RVLocs[i].getLocReg());
1894 SDOperand Chain = Op.getOperand(0);
1897 // Copy the result values into the output registers.
1898 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1899 CCValAssign &VA = RVLocs[i];
1900 assert(VA.isRegLoc() && "Can only return in registers!");
1901 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
1902 Flag = Chain.getValue(1);
1906 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
1908 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
1911 static SDOperand LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
1912 const PPCSubtarget &Subtarget) {
1913 // When we pop the dynamic allocation we need to restore the SP link.
1915 // Get the corect type for pointers.
1916 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1918 // Construct the stack pointer operand.
1919 bool IsPPC64 = Subtarget.isPPC64();
1920 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
1921 SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
1923 // Get the operands for the STACKRESTORE.
1924 SDOperand Chain = Op.getOperand(0);
1925 SDOperand SaveSP = Op.getOperand(1);
1927 // Load the old link SP.
1928 SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
1930 // Restore the stack pointer.
1931 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
1933 // Store the old link SP.
1934 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
1937 static SDOperand LowerDYNAMIC_STACKALLOC(SDOperand Op, SelectionDAG &DAG,
1938 const PPCSubtarget &Subtarget) {
1939 MachineFunction &MF = DAG.getMachineFunction();
1940 bool IsPPC64 = Subtarget.isPPC64();
1941 bool isMachoABI = Subtarget.isMachoABI();
1943 // Get current frame pointer save index. The users of this index will be
1944 // primarily DYNALLOC instructions.
1945 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1946 int FPSI = FI->getFramePointerSaveIndex();
1948 // If the frame pointer save index hasn't been defined yet.
1950 // Find out what the fix offset of the frame pointer save area.
1951 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
1953 // Allocate the frame index for frame pointer save area.
1954 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
1956 FI->setFramePointerSaveIndex(FPSI);
1960 SDOperand Chain = Op.getOperand(0);
1961 SDOperand Size = Op.getOperand(1);
1963 // Get the corect type for pointers.
1964 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1966 SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
1967 DAG.getConstant(0, PtrVT), Size);
1968 // Construct a node for the frame pointer save index.
1969 SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
1970 // Build a DYNALLOC node.
1971 SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
1972 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
1973 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
1977 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
1979 static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
1980 // Not FP? Not a fsel.
1981 if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
1982 !MVT::isFloatingPoint(Op.getOperand(2).getValueType()))
1985 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
1987 // Cannot handle SETEQ/SETNE.
1988 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDOperand();
1990 MVT::ValueType ResVT = Op.getValueType();
1991 MVT::ValueType CmpVT = Op.getOperand(0).getValueType();
1992 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
1993 SDOperand TV = Op.getOperand(2), FV = Op.getOperand(3);
1995 // If the RHS of the comparison is a 0.0, we don't need to do the
1996 // subtraction at all.
1997 if (isFloatingPointZero(RHS))
1999 default: break; // SETUO etc aren't handled by fsel.
2003 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2007 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2008 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2009 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2013 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2017 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2018 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2019 return DAG.getNode(PPCISD::FSEL, ResVT,
2020 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2025 default: break; // SETUO etc aren't handled by fsel.
2029 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2030 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2031 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2032 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2036 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2037 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2038 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2039 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2043 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2044 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2045 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2046 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2050 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2051 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2052 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2053 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2058 static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) {
2059 assert(MVT::isFloatingPoint(Op.getOperand(0).getValueType()));
2060 SDOperand Src = Op.getOperand(0);
2061 if (Src.getValueType() == MVT::f32)
2062 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2065 switch (Op.getValueType()) {
2066 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2068 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2071 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2075 // Convert the FP value to an int value through memory.
2076 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
2077 if (Op.getValueType() == MVT::i32)
2078 Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
2082 static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
2083 if (Op.getOperand(0).getValueType() == MVT::i64) {
2084 SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2085 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2086 if (Op.getValueType() == MVT::f32)
2087 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
2091 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2092 "Unhandled SINT_TO_FP type in custom expander!");
2093 // Since we only generate this in 64-bit mode, we can take advantage of
2094 // 64-bit registers. In particular, sign extend the input value into the
2095 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2096 // then lfd it and fcfid it.
2097 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2098 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2099 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2100 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2102 SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2105 // STD the extended value into the stack slot.
2106 SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2107 DAG.getEntryNode(), Ext64, FIdx,
2108 DAG.getSrcValue(NULL));
2109 // Load the value as a double.
2110 SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2112 // FCFID it and return it.
2113 SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2114 if (Op.getValueType() == MVT::f32)
2115 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
2119 static SDOperand LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2120 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2121 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
2123 // Expand into a bunch of logical ops. Note that these ops
2124 // depend on the PPC behavior for oversized shift amounts.
2125 SDOperand Lo = Op.getOperand(0);
2126 SDOperand Hi = Op.getOperand(1);
2127 SDOperand Amt = Op.getOperand(2);
2129 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2130 DAG.getConstant(32, MVT::i32), Amt);
2131 SDOperand Tmp2 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Amt);
2132 SDOperand Tmp3 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Tmp1);
2133 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2134 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2135 DAG.getConstant(-32U, MVT::i32));
2136 SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
2137 SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2138 SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
2139 SDOperand OutOps[] = { OutLo, OutHi };
2140 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2144 static SDOperand LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
2145 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2146 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRL!");
2148 // Otherwise, expand into a bunch of logical ops. Note that these ops
2149 // depend on the PPC behavior for oversized shift amounts.
2150 SDOperand Lo = Op.getOperand(0);
2151 SDOperand Hi = Op.getOperand(1);
2152 SDOperand Amt = Op.getOperand(2);
2154 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2155 DAG.getConstant(32, MVT::i32), Amt);
2156 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2157 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2158 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2159 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2160 DAG.getConstant(-32U, MVT::i32));
2161 SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
2162 SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
2163 SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
2164 SDOperand OutOps[] = { OutLo, OutHi };
2165 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2169 static SDOperand LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
2170 assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
2171 Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
2173 // Otherwise, expand into a bunch of logical ops, followed by a select_cc.
2174 SDOperand Lo = Op.getOperand(0);
2175 SDOperand Hi = Op.getOperand(1);
2176 SDOperand Amt = Op.getOperand(2);
2178 SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
2179 DAG.getConstant(32, MVT::i32), Amt);
2180 SDOperand Tmp2 = DAG.getNode(PPCISD::SRL, MVT::i32, Lo, Amt);
2181 SDOperand Tmp3 = DAG.getNode(PPCISD::SHL, MVT::i32, Hi, Tmp1);
2182 SDOperand Tmp4 = DAG.getNode(ISD::OR , MVT::i32, Tmp2, Tmp3);
2183 SDOperand Tmp5 = DAG.getNode(ISD::ADD, MVT::i32, Amt,
2184 DAG.getConstant(-32U, MVT::i32));
2185 SDOperand Tmp6 = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Tmp5);
2186 SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
2187 SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
2188 Tmp4, Tmp6, ISD::SETLE);
2189 SDOperand OutOps[] = { OutLo, OutHi };
2190 return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
2194 //===----------------------------------------------------------------------===//
2195 // Vector related lowering.
2198 // If this is a vector of constants or undefs, get the bits. A bit in
2199 // UndefBits is set if the corresponding element of the vector is an
2200 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2201 // zero. Return true if this is not an array of constants, false if it is.
2203 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
2204 uint64_t UndefBits[2]) {
2205 // Start with zero'd results.
2206 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
2208 unsigned EltBitSize = MVT::getSizeInBits(BV->getOperand(0).getValueType());
2209 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
2210 SDOperand OpVal = BV->getOperand(i);
2212 unsigned PartNo = i >= e/2; // In the upper 128 bits?
2213 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
2215 uint64_t EltBits = 0;
2216 if (OpVal.getOpcode() == ISD::UNDEF) {
2217 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
2218 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
2220 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
2221 EltBits = CN->getValue() & (~0U >> (32-EltBitSize));
2222 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
2223 assert(CN->getValueType(0) == MVT::f32 &&
2224 "Only one legal FP vector type!");
2225 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
2227 // Nonconstant element.
2231 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
2234 //printf("%llx %llx %llx %llx\n",
2235 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
2239 // If this is a splat (repetition) of a value across the whole vector, return
2240 // the smallest size that splats it. For example, "0x01010101010101..." is a
2241 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2242 // SplatSize = 1 byte.
2243 static bool isConstantSplat(const uint64_t Bits128[2],
2244 const uint64_t Undef128[2],
2245 unsigned &SplatBits, unsigned &SplatUndef,
2246 unsigned &SplatSize) {
2248 // Don't let undefs prevent splats from matching. See if the top 64-bits are
2249 // the same as the lower 64-bits, ignoring undefs.
2250 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
2251 return false; // Can't be a splat if two pieces don't match.
2253 uint64_t Bits64 = Bits128[0] | Bits128[1];
2254 uint64_t Undef64 = Undef128[0] & Undef128[1];
2256 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
2258 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
2259 return false; // Can't be a splat if two pieces don't match.
2261 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
2262 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
2264 // If the top 16-bits are different than the lower 16-bits, ignoring
2265 // undefs, we have an i32 splat.
2266 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
2268 SplatUndef = Undef32;
2273 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
2274 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
2276 // If the top 8-bits are different than the lower 8-bits, ignoring
2277 // undefs, we have an i16 splat.
2278 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
2280 SplatUndef = Undef16;
2285 // Otherwise, we have an 8-bit splat.
2286 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
2287 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
2292 /// BuildSplatI - Build a canonical splati of Val with an element size of
2293 /// SplatSize. Cast the result to VT.
2294 static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
2295 SelectionDAG &DAG) {
2296 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
2298 static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
2299 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
2302 MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
2304 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
2308 MVT::ValueType CanonicalVT = VTys[SplatSize-1];
2310 // Build a canonical splat for this value.
2311 SDOperand Elt = DAG.getConstant(Val, MVT::getVectorElementType(CanonicalVT));
2312 SmallVector<SDOperand, 8> Ops;
2313 Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
2314 SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
2315 &Ops[0], Ops.size());
2316 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
2319 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
2320 /// specified intrinsic ID.
2321 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand LHS, SDOperand RHS,
2323 MVT::ValueType DestVT = MVT::Other) {
2324 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
2325 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2326 DAG.getConstant(IID, MVT::i32), LHS, RHS);
2329 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
2330 /// specified intrinsic ID.
2331 static SDOperand BuildIntrinsicOp(unsigned IID, SDOperand Op0, SDOperand Op1,
2332 SDOperand Op2, SelectionDAG &DAG,
2333 MVT::ValueType DestVT = MVT::Other) {
2334 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
2335 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
2336 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
2340 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
2341 /// amount. The result has the specified value type.
2342 static SDOperand BuildVSLDOI(SDOperand LHS, SDOperand RHS, unsigned Amt,
2343 MVT::ValueType VT, SelectionDAG &DAG) {
2344 // Force LHS/RHS to be the right type.
2345 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
2346 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
2349 for (unsigned i = 0; i != 16; ++i)
2350 Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
2351 SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
2352 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
2353 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
2356 // If this is a case we can't handle, return null and let the default
2357 // expansion code take care of it. If we CAN select this case, and if it
2358 // selects to a single instruction, return Op. Otherwise, if we can codegen
2359 // this case more efficiently than a constant pool load, lower it to the
2360 // sequence of ops that should be used.
2361 static SDOperand LowerBUILD_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2362 // If this is a vector of constants or undefs, get the bits. A bit in
2363 // UndefBits is set if the corresponding element of the vector is an
2364 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
2366 uint64_t VectorBits[2];
2367 uint64_t UndefBits[2];
2368 if (GetConstantBuildVectorBits(Op.Val, VectorBits, UndefBits))
2369 return SDOperand(); // Not a constant vector.
2371 // If this is a splat (repetition) of a value across the whole vector, return
2372 // the smallest size that splats it. For example, "0x01010101010101..." is a
2373 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
2374 // SplatSize = 1 byte.
2375 unsigned SplatBits, SplatUndef, SplatSize;
2376 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
2377 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
2379 // First, handle single instruction cases.
2382 if (SplatBits == 0) {
2383 // Canonicalize all zero vectors to be v4i32.
2384 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
2385 SDOperand Z = DAG.getConstant(0, MVT::i32);
2386 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
2387 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
2392 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
2393 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
2394 if (SextVal >= -16 && SextVal <= 15)
2395 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
2398 // Two instruction sequences.
2400 // If this value is in the range [-32,30] and is even, use:
2401 // tmp = VSPLTI[bhw], result = add tmp, tmp
2402 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
2403 Op = BuildSplatI(SextVal >> 1, SplatSize, Op.getValueType(), DAG);
2404 return DAG.getNode(ISD::ADD, Op.getValueType(), Op, Op);
2407 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
2408 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
2410 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
2411 // Make -1 and vspltisw -1:
2412 SDOperand OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
2414 // Make the VSLW intrinsic, computing 0x8000_0000.
2415 SDOperand Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
2418 // xor by OnesV to invert it.
2419 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
2420 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2423 // Check to see if this is a wide variety of vsplti*, binop self cases.
2424 unsigned SplatBitSize = SplatSize*8;
2425 static const signed char SplatCsts[] = {
2426 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
2427 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
2430 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
2431 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
2432 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
2433 int i = SplatCsts[idx];
2435 // Figure out what shift amount will be used by altivec if shifted by i in
2437 unsigned TypeShiftAmt = i & (SplatBitSize-1);
2439 // vsplti + shl self.
2440 if (SextVal == (i << (int)TypeShiftAmt)) {
2441 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2442 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2443 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
2444 Intrinsic::ppc_altivec_vslw
2446 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2447 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2450 // vsplti + srl self.
2451 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2452 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2453 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2454 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
2455 Intrinsic::ppc_altivec_vsrw
2457 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2458 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2461 // vsplti + sra self.
2462 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
2463 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2464 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2465 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
2466 Intrinsic::ppc_altivec_vsraw
2468 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2469 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2472 // vsplti + rol self.
2473 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
2474 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
2475 SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
2476 static const unsigned IIDs[] = { // Intrinsic to use for each size.
2477 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
2478 Intrinsic::ppc_altivec_vrlw
2480 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
2481 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
2484 // t = vsplti c, result = vsldoi t, t, 1
2485 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
2486 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2487 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
2489 // t = vsplti c, result = vsldoi t, t, 2
2490 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
2491 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2492 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
2494 // t = vsplti c, result = vsldoi t, t, 3
2495 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
2496 SDOperand T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
2497 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
2501 // Three instruction sequences.
2503 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
2504 if (SextVal >= 0 && SextVal <= 31) {
2505 SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
2506 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2507 LHS = DAG.getNode(ISD::SUB, Op.getValueType(), LHS, RHS);
2508 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2510 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
2511 if (SextVal >= -31 && SextVal <= 0) {
2512 SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
2513 SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
2514 LHS = DAG.getNode(ISD::ADD, Op.getValueType(), LHS, RHS);
2515 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
2522 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
2523 /// the specified operations to build the shuffle.
2524 static SDOperand GeneratePerfectShuffle(unsigned PFEntry, SDOperand LHS,
2525 SDOperand RHS, SelectionDAG &DAG) {
2526 unsigned OpNum = (PFEntry >> 26) & 0x0F;
2527 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
2528 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
2531 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
2543 if (OpNum == OP_COPY) {
2544 if (LHSID == (1*9+2)*9+3) return LHS;
2545 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
2549 SDOperand OpLHS, OpRHS;
2550 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
2551 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
2553 unsigned ShufIdxs[16];
2555 default: assert(0 && "Unknown i32 permute!");
2557 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
2558 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
2559 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
2560 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
2563 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
2564 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
2565 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
2566 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
2569 for (unsigned i = 0; i != 16; ++i)
2570 ShufIdxs[i] = (i&3)+0;
2573 for (unsigned i = 0; i != 16; ++i)
2574 ShufIdxs[i] = (i&3)+4;
2577 for (unsigned i = 0; i != 16; ++i)
2578 ShufIdxs[i] = (i&3)+8;
2581 for (unsigned i = 0; i != 16; ++i)
2582 ShufIdxs[i] = (i&3)+12;
2585 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
2587 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
2589 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
2592 for (unsigned i = 0; i != 16; ++i)
2593 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
2595 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
2596 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2599 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
2600 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
2601 /// return the code it can be lowered into. Worst case, it can always be
2602 /// lowered into a vperm.
2603 static SDOperand LowerVECTOR_SHUFFLE(SDOperand Op, SelectionDAG &DAG) {
2604 SDOperand V1 = Op.getOperand(0);
2605 SDOperand V2 = Op.getOperand(1);
2606 SDOperand PermMask = Op.getOperand(2);
2608 // Cases that are handled by instructions that take permute immediates
2609 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
2610 // selected by the instruction selector.
2611 if (V2.getOpcode() == ISD::UNDEF) {
2612 if (PPC::isSplatShuffleMask(PermMask.Val, 1) ||
2613 PPC::isSplatShuffleMask(PermMask.Val, 2) ||
2614 PPC::isSplatShuffleMask(PermMask.Val, 4) ||
2615 PPC::isVPKUWUMShuffleMask(PermMask.Val, true) ||
2616 PPC::isVPKUHUMShuffleMask(PermMask.Val, true) ||
2617 PPC::isVSLDOIShuffleMask(PermMask.Val, true) != -1 ||
2618 PPC::isVMRGLShuffleMask(PermMask.Val, 1, true) ||
2619 PPC::isVMRGLShuffleMask(PermMask.Val, 2, true) ||
2620 PPC::isVMRGLShuffleMask(PermMask.Val, 4, true) ||
2621 PPC::isVMRGHShuffleMask(PermMask.Val, 1, true) ||
2622 PPC::isVMRGHShuffleMask(PermMask.Val, 2, true) ||
2623 PPC::isVMRGHShuffleMask(PermMask.Val, 4, true)) {
2628 // Altivec has a variety of "shuffle immediates" that take two vector inputs
2629 // and produce a fixed permutation. If any of these match, do not lower to
2631 if (PPC::isVPKUWUMShuffleMask(PermMask.Val, false) ||
2632 PPC::isVPKUHUMShuffleMask(PermMask.Val, false) ||
2633 PPC::isVSLDOIShuffleMask(PermMask.Val, false) != -1 ||
2634 PPC::isVMRGLShuffleMask(PermMask.Val, 1, false) ||
2635 PPC::isVMRGLShuffleMask(PermMask.Val, 2, false) ||
2636 PPC::isVMRGLShuffleMask(PermMask.Val, 4, false) ||
2637 PPC::isVMRGHShuffleMask(PermMask.Val, 1, false) ||
2638 PPC::isVMRGHShuffleMask(PermMask.Val, 2, false) ||
2639 PPC::isVMRGHShuffleMask(PermMask.Val, 4, false))
2642 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
2643 // perfect shuffle table to emit an optimal matching sequence.
2644 unsigned PFIndexes[4];
2645 bool isFourElementShuffle = true;
2646 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
2647 unsigned EltNo = 8; // Start out undef.
2648 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
2649 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
2650 continue; // Undef, ignore it.
2652 unsigned ByteSource =
2653 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getValue();
2654 if ((ByteSource & 3) != j) {
2655 isFourElementShuffle = false;
2660 EltNo = ByteSource/4;
2661 } else if (EltNo != ByteSource/4) {
2662 isFourElementShuffle = false;
2666 PFIndexes[i] = EltNo;
2669 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
2670 // perfect shuffle vector to determine if it is cost effective to do this as
2671 // discrete instructions, or whether we should use a vperm.
2672 if (isFourElementShuffle) {
2673 // Compute the index in the perfect shuffle table.
2674 unsigned PFTableIndex =
2675 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
2677 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
2678 unsigned Cost = (PFEntry >> 30);
2680 // Determining when to avoid vperm is tricky. Many things affect the cost
2681 // of vperm, particularly how many times the perm mask needs to be computed.
2682 // For example, if the perm mask can be hoisted out of a loop or is already
2683 // used (perhaps because there are multiple permutes with the same shuffle
2684 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
2685 // the loop requires an extra register.
2687 // As a compromise, we only emit discrete instructions if the shuffle can be
2688 // generated in 3 or fewer operations. When we have loop information
2689 // available, if this block is within a loop, we should avoid using vperm
2690 // for 3-operation perms and use a constant pool load instead.
2692 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
2695 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
2696 // vector that will get spilled to the constant pool.
2697 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
2699 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
2700 // that it is in input element units, not in bytes. Convert now.
2701 MVT::ValueType EltVT = MVT::getVectorElementType(V1.getValueType());
2702 unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
2704 SmallVector<SDOperand, 16> ResultMask;
2705 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
2707 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
2710 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
2712 for (unsigned j = 0; j != BytesPerElement; ++j)
2713 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
2717 SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
2718 &ResultMask[0], ResultMask.size());
2719 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
2722 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
2723 /// altivec comparison. If it is, return true and fill in Opc/isDot with
2724 /// information about the intrinsic.
2725 static bool getAltivecCompareInfo(SDOperand Intrin, int &CompareOpc,
2727 unsigned IntrinsicID = cast<ConstantSDNode>(Intrin.getOperand(0))->getValue();
2730 switch (IntrinsicID) {
2731 default: return false;
2732 // Comparison predicates.
2733 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
2734 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
2735 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
2736 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
2737 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
2738 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
2739 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
2740 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
2741 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
2742 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
2743 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
2744 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
2745 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
2747 // Normal Comparisons.
2748 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
2749 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
2750 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
2751 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
2752 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
2753 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
2754 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
2755 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
2756 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
2757 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
2758 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
2759 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
2760 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
2765 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
2766 /// lower, do it, otherwise return null.
2767 static SDOperand LowerINTRINSIC_WO_CHAIN(SDOperand Op, SelectionDAG &DAG) {
2768 // If this is a lowered altivec predicate compare, CompareOpc is set to the
2769 // opcode number of the comparison.
2772 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
2773 return SDOperand(); // Don't custom lower most intrinsics.
2775 // If this is a non-dot comparison, make the VCMP node and we are done.
2777 SDOperand Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
2778 Op.getOperand(1), Op.getOperand(2),
2779 DAG.getConstant(CompareOpc, MVT::i32));
2780 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
2783 // Create the PPCISD altivec 'dot' comparison node.
2785 Op.getOperand(2), // LHS
2786 Op.getOperand(3), // RHS
2787 DAG.getConstant(CompareOpc, MVT::i32)
2789 std::vector<MVT::ValueType> VTs;
2790 VTs.push_back(Op.getOperand(2).getValueType());
2791 VTs.push_back(MVT::Flag);
2792 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
2794 // Now that we have the comparison, emit a copy from the CR to a GPR.
2795 // This is flagged to the above dot comparison.
2796 SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
2797 DAG.getRegister(PPC::CR6, MVT::i32),
2798 CompNode.getValue(1));
2800 // Unpack the result based on how the target uses it.
2801 unsigned BitNo; // Bit # of CR6.
2802 bool InvertBit; // Invert result?
2803 switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
2804 default: // Can't happen, don't crash on invalid number though.
2805 case 0: // Return the value of the EQ bit of CR6.
2806 BitNo = 0; InvertBit = false;
2808 case 1: // Return the inverted value of the EQ bit of CR6.
2809 BitNo = 0; InvertBit = true;
2811 case 2: // Return the value of the LT bit of CR6.
2812 BitNo = 2; InvertBit = false;
2814 case 3: // Return the inverted value of the LT bit of CR6.
2815 BitNo = 2; InvertBit = true;
2819 // Shift the bit into the low position.
2820 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
2821 DAG.getConstant(8-(3-BitNo), MVT::i32));
2823 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
2824 DAG.getConstant(1, MVT::i32));
2826 // If we are supposed to, toggle the bit.
2828 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
2829 DAG.getConstant(1, MVT::i32));
2833 static SDOperand LowerSCALAR_TO_VECTOR(SDOperand Op, SelectionDAG &DAG) {
2834 // Create a stack slot that is 16-byte aligned.
2835 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2836 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
2837 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2838 SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2840 // Store the input value into Value#0 of the stack slot.
2841 SDOperand Store = DAG.getStore(DAG.getEntryNode(),
2842 Op.getOperand(0), FIdx, NULL, 0);
2844 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
2847 static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
2848 if (Op.getValueType() == MVT::v4i32) {
2849 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2851 SDOperand Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
2852 SDOperand Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
2854 SDOperand RHSSwap = // = vrlw RHS, 16
2855 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
2857 // Shrinkify inputs to v8i16.
2858 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
2859 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
2860 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
2862 // Low parts multiplied together, generating 32-bit results (we ignore the
2864 SDOperand LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
2865 LHS, RHS, DAG, MVT::v4i32);
2867 SDOperand HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
2868 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
2869 // Shift the high parts up 16 bits.
2870 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
2871 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
2872 } else if (Op.getValueType() == MVT::v8i16) {
2873 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2875 SDOperand Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
2877 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
2878 LHS, RHS, Zero, DAG);
2879 } else if (Op.getValueType() == MVT::v16i8) {
2880 SDOperand LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2882 // Multiply the even 8-bit parts, producing 16-bit sums.
2883 SDOperand EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
2884 LHS, RHS, DAG, MVT::v8i16);
2885 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
2887 // Multiply the odd 8-bit parts, producing 16-bit sums.
2888 SDOperand OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
2889 LHS, RHS, DAG, MVT::v8i16);
2890 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
2892 // Merge the results together.
2894 for (unsigned i = 0; i != 8; ++i) {
2895 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
2896 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
2898 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
2899 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
2901 assert(0 && "Unknown mul to lower!");
2906 /// LowerOperation - Provide custom lowering hooks for some operations.
2908 SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
2909 switch (Op.getOpcode()) {
2910 default: assert(0 && "Wasn't expecting to be able to lower this!");
2911 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2912 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
2913 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2914 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
2915 case ISD::SETCC: return LowerSETCC(Op, DAG);
2917 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
2918 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
2921 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
2922 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
2924 case ISD::FORMAL_ARGUMENTS:
2925 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
2926 VarArgsStackOffset, VarArgsNumGPR,
2927 VarArgsNumFPR, PPCSubTarget);
2929 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget);
2930 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
2931 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
2932 case ISD::DYNAMIC_STACKALLOC:
2933 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
2935 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
2936 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
2937 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
2939 // Lower 64-bit shifts.
2940 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
2941 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
2942 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
2944 // Vector-related lowering.
2945 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2946 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2947 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2948 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
2949 case ISD::MUL: return LowerMUL(Op, DAG);
2951 // Frame & Return address. Currently unimplemented
2952 case ISD::RETURNADDR: break;
2953 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2958 //===----------------------------------------------------------------------===//
2959 // Other Lowering Code
2960 //===----------------------------------------------------------------------===//
2963 PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
2964 MachineBasicBlock *BB) {
2965 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
2966 assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
2967 MI->getOpcode() == PPC::SELECT_CC_I8 ||
2968 MI->getOpcode() == PPC::SELECT_CC_F4 ||
2969 MI->getOpcode() == PPC::SELECT_CC_F8 ||
2970 MI->getOpcode() == PPC::SELECT_CC_VRRC) &&
2971 "Unexpected instr type to insert");
2973 // To "insert" a SELECT_CC instruction, we actually have to insert the diamond
2974 // control-flow pattern. The incoming instruction knows the destination vreg
2975 // to set, the condition code register to branch on, the true/false values to
2976 // select between, and a branch opcode to use.
2977 const BasicBlock *LLVM_BB = BB->getBasicBlock();
2978 ilist<MachineBasicBlock>::iterator It = BB;
2984 // cmpTY ccX, r1, r2
2986 // fallthrough --> copy0MBB
2987 MachineBasicBlock *thisMBB = BB;
2988 MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
2989 MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
2990 unsigned SelectPred = MI->getOperand(4).getImm();
2991 BuildMI(BB, TII->get(PPC::BCC))
2992 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
2993 MachineFunction *F = BB->getParent();
2994 F->getBasicBlockList().insert(It, copy0MBB);
2995 F->getBasicBlockList().insert(It, sinkMBB);
2996 // Update machine-CFG edges by first adding all successors of the current
2997 // block to the new block which will contain the Phi node for the select.
2998 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
2999 e = BB->succ_end(); i != e; ++i)
3000 sinkMBB->addSuccessor(*i);
3001 // Next, remove all successors of the current block, and add the true
3002 // and fallthrough blocks as its successors.
3003 while(!BB->succ_empty())
3004 BB->removeSuccessor(BB->succ_begin());
3005 BB->addSuccessor(copy0MBB);
3006 BB->addSuccessor(sinkMBB);
3009 // %FalseValue = ...
3010 // # fallthrough to sinkMBB
3013 // Update machine-CFG edges
3014 BB->addSuccessor(sinkMBB);
3017 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
3020 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
3021 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
3022 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
3024 delete MI; // The pseudo instruction is gone now.
3028 //===----------------------------------------------------------------------===//
3029 // Target Optimization Hooks
3030 //===----------------------------------------------------------------------===//
3032 SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
3033 DAGCombinerInfo &DCI) const {
3034 TargetMachine &TM = getTargetMachine();
3035 SelectionDAG &DAG = DCI.DAG;
3036 switch (N->getOpcode()) {
3039 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3040 if (C->getValue() == 0) // 0 << V -> 0.
3041 return N->getOperand(0);
3045 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3046 if (C->getValue() == 0) // 0 >>u V -> 0.
3047 return N->getOperand(0);
3051 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
3052 if (C->getValue() == 0 || // 0 >>s V -> 0.
3053 C->isAllOnesValue()) // -1 >>s V -> -1.
3054 return N->getOperand(0);
3058 case ISD::SINT_TO_FP:
3059 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
3060 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
3061 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
3062 // We allow the src/dst to be either f32/f64, but the intermediate
3063 // type must be i64.
3064 if (N->getOperand(0).getValueType() == MVT::i64) {
3065 SDOperand Val = N->getOperand(0).getOperand(0);
3066 if (Val.getValueType() == MVT::f32) {
3067 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3068 DCI.AddToWorklist(Val.Val);
3071 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
3072 DCI.AddToWorklist(Val.Val);
3073 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
3074 DCI.AddToWorklist(Val.Val);
3075 if (N->getValueType(0) == MVT::f32) {
3076 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
3077 DCI.AddToWorklist(Val.Val);
3080 } else if (N->getOperand(0).getValueType() == MVT::i32) {
3081 // If the intermediate type is i32, we can avoid the load/store here
3088 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
3089 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
3090 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
3091 N->getOperand(1).getValueType() == MVT::i32) {
3092 SDOperand Val = N->getOperand(1).getOperand(0);
3093 if (Val.getValueType() == MVT::f32) {
3094 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
3095 DCI.AddToWorklist(Val.Val);
3097 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
3098 DCI.AddToWorklist(Val.Val);
3100 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
3101 N->getOperand(2), N->getOperand(3));
3102 DCI.AddToWorklist(Val.Val);
3106 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
3107 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
3108 N->getOperand(1).Val->hasOneUse() &&
3109 (N->getOperand(1).getValueType() == MVT::i32 ||
3110 N->getOperand(1).getValueType() == MVT::i16)) {
3111 SDOperand BSwapOp = N->getOperand(1).getOperand(0);
3112 // Do an any-extend to 32-bits if this is a half-word input.
3113 if (BSwapOp.getValueType() == MVT::i16)
3114 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
3116 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
3117 N->getOperand(2), N->getOperand(3),
3118 DAG.getValueType(N->getOperand(1).getValueType()));
3122 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
3123 if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
3124 N->getOperand(0).hasOneUse() &&
3125 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
3126 SDOperand Load = N->getOperand(0);
3127 LoadSDNode *LD = cast<LoadSDNode>(Load);
3128 // Create the byte-swapping load.
3129 std::vector<MVT::ValueType> VTs;
3130 VTs.push_back(MVT::i32);
3131 VTs.push_back(MVT::Other);
3132 SDOperand SV = DAG.getSrcValue(LD->getSrcValue(), LD->getSrcValueOffset());
3134 LD->getChain(), // Chain
3135 LD->getBasePtr(), // Ptr
3137 DAG.getValueType(N->getValueType(0)) // VT
3139 SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
3141 // If this is an i16 load, insert the truncate.
3142 SDOperand ResVal = BSLoad;
3143 if (N->getValueType(0) == MVT::i16)
3144 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
3146 // First, combine the bswap away. This makes the value produced by the
3148 DCI.CombineTo(N, ResVal);
3150 // Next, combine the load away, we give it a bogus result value but a real
3151 // chain result. The result value is dead because the bswap is dead.
3152 DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
3154 // Return N so it doesn't get rechecked!
3155 return SDOperand(N, 0);
3159 case PPCISD::VCMP: {
3160 // If a VCMPo node already exists with exactly the same operands as this
3161 // node, use its result instead of this node (VCMPo computes both a CR6 and
3162 // a normal output).
3164 if (!N->getOperand(0).hasOneUse() &&
3165 !N->getOperand(1).hasOneUse() &&
3166 !N->getOperand(2).hasOneUse()) {
3168 // Scan all of the users of the LHS, looking for VCMPo's that match.
3169 SDNode *VCMPoNode = 0;
3171 SDNode *LHSN = N->getOperand(0).Val;
3172 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
3174 if ((*UI)->getOpcode() == PPCISD::VCMPo &&
3175 (*UI)->getOperand(1) == N->getOperand(1) &&
3176 (*UI)->getOperand(2) == N->getOperand(2) &&
3177 (*UI)->getOperand(0) == N->getOperand(0)) {
3182 // If there is no VCMPo node, or if the flag value has a single use, don't
3184 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
3187 // Look at the (necessarily single) use of the flag value. If it has a
3188 // chain, this transformation is more complex. Note that multiple things
3189 // could use the value result, which we should ignore.
3190 SDNode *FlagUser = 0;
3191 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
3192 FlagUser == 0; ++UI) {
3193 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
3195 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
3196 if (User->getOperand(i) == SDOperand(VCMPoNode, 1)) {
3203 // If the user is a MFCR instruction, we know this is safe. Otherwise we
3204 // give up for right now.
3205 if (FlagUser->getOpcode() == PPCISD::MFCR)
3206 return SDOperand(VCMPoNode, 0);
3211 // If this is a branch on an altivec predicate comparison, lower this so
3212 // that we don't have to do a MFCR: instead, branch directly on CR6. This
3213 // lowering is done pre-legalize, because the legalizer lowers the predicate
3214 // compare down to code that is difficult to reassemble.
3215 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
3216 SDOperand LHS = N->getOperand(2), RHS = N->getOperand(3);
3220 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
3221 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
3222 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
3223 assert(isDot && "Can't compare against a vector result!");
3225 // If this is a comparison against something other than 0/1, then we know
3226 // that the condition is never/always true.
3227 unsigned Val = cast<ConstantSDNode>(RHS)->getValue();
3228 if (Val != 0 && Val != 1) {
3229 if (CC == ISD::SETEQ) // Cond never true, remove branch.
3230 return N->getOperand(0);
3231 // Always !=, turn it into an unconditional branch.
3232 return DAG.getNode(ISD::BR, MVT::Other,
3233 N->getOperand(0), N->getOperand(4));
3236 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
3238 // Create the PPCISD altivec 'dot' comparison node.
3239 std::vector<MVT::ValueType> VTs;
3241 LHS.getOperand(2), // LHS of compare
3242 LHS.getOperand(3), // RHS of compare
3243 DAG.getConstant(CompareOpc, MVT::i32)
3245 VTs.push_back(LHS.getOperand(2).getValueType());
3246 VTs.push_back(MVT::Flag);
3247 SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3249 // Unpack the result based on how the target uses it.
3250 PPC::Predicate CompOpc;
3251 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
3252 default: // Can't happen, don't crash on invalid number though.
3253 case 0: // Branch on the value of the EQ bit of CR6.
3254 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
3256 case 1: // Branch on the inverted value of the EQ bit of CR6.
3257 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
3259 case 2: // Branch on the value of the LT bit of CR6.
3260 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
3262 case 3: // Branch on the inverted value of the LT bit of CR6.
3263 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
3267 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
3268 DAG.getConstant(CompOpc, MVT::i32),
3269 DAG.getRegister(PPC::CR6, MVT::i32),
3270 N->getOperand(4), CompNode.getValue(1));
3279 //===----------------------------------------------------------------------===//
3280 // Inline Assembly Support
3281 //===----------------------------------------------------------------------===//
3283 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
3285 uint64_t &KnownZero,
3287 const SelectionDAG &DAG,
3288 unsigned Depth) const {
3291 switch (Op.getOpcode()) {
3293 case PPCISD::LBRX: {
3294 // lhbrx is known to have the top bits cleared out.
3295 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
3296 KnownZero = 0xFFFF0000;
3299 case ISD::INTRINSIC_WO_CHAIN: {
3300 switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
3302 case Intrinsic::ppc_altivec_vcmpbfp_p:
3303 case Intrinsic::ppc_altivec_vcmpeqfp_p:
3304 case Intrinsic::ppc_altivec_vcmpequb_p:
3305 case Intrinsic::ppc_altivec_vcmpequh_p:
3306 case Intrinsic::ppc_altivec_vcmpequw_p:
3307 case Intrinsic::ppc_altivec_vcmpgefp_p:
3308 case Intrinsic::ppc_altivec_vcmpgtfp_p:
3309 case Intrinsic::ppc_altivec_vcmpgtsb_p:
3310 case Intrinsic::ppc_altivec_vcmpgtsh_p:
3311 case Intrinsic::ppc_altivec_vcmpgtsw_p:
3312 case Intrinsic::ppc_altivec_vcmpgtub_p:
3313 case Intrinsic::ppc_altivec_vcmpgtuh_p:
3314 case Intrinsic::ppc_altivec_vcmpgtuw_p:
3315 KnownZero = ~1U; // All bits but the low one are known to be zero.
3323 /// getConstraintType - Given a constraint, return the type of
3324 /// constraint it is for this target.
3325 PPCTargetLowering::ConstraintType
3326 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
3327 if (Constraint.size() == 1) {
3328 switch (Constraint[0]) {
3335 return C_RegisterClass;
3338 return TargetLowering::getConstraintType(Constraint);
3341 std::pair<unsigned, const TargetRegisterClass*>
3342 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3343 MVT::ValueType VT) const {
3344 if (Constraint.size() == 1) {
3345 // GCC RS6000 Constraint Letters
3346 switch (Constraint[0]) {
3349 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
3350 return std::make_pair(0U, PPC::G8RCRegisterClass);
3351 return std::make_pair(0U, PPC::GPRCRegisterClass);
3354 return std::make_pair(0U, PPC::F4RCRegisterClass);
3355 else if (VT == MVT::f64)
3356 return std::make_pair(0U, PPC::F8RCRegisterClass);
3359 return std::make_pair(0U, PPC::VRRCRegisterClass);
3361 return std::make_pair(0U, PPC::CRRCRegisterClass);
3365 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3369 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3370 /// vector. If it is invalid, don't add anything to Ops.
3371 void PPCTargetLowering::LowerAsmOperandForConstraint(SDOperand Op, char Letter,
3372 std::vector<SDOperand>&Ops,
3373 SelectionDAG &DAG) {
3374 SDOperand Result(0,0);
3385 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
3386 if (!CST) return; // Must be an immediate to match.
3387 unsigned Value = CST->getValue();
3389 default: assert(0 && "Unknown constraint letter!");
3390 case 'I': // "I" is a signed 16-bit constant.
3391 if ((short)Value == (int)Value)
3392 Result = DAG.getTargetConstant(Value, Op.getValueType());
3394 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
3395 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
3396 if ((short)Value == 0)
3397 Result = DAG.getTargetConstant(Value, Op.getValueType());
3399 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
3400 if ((Value >> 16) == 0)
3401 Result = DAG.getTargetConstant(Value, Op.getValueType());
3403 case 'M': // "M" is a constant that is greater than 31.
3405 Result = DAG.getTargetConstant(Value, Op.getValueType());
3407 case 'N': // "N" is a positive constant that is an exact power of two.
3408 if ((int)Value > 0 && isPowerOf2_32(Value))
3409 Result = DAG.getTargetConstant(Value, Op.getValueType());
3411 case 'O': // "O" is the constant zero.
3413 Result = DAG.getTargetConstant(Value, Op.getValueType());
3415 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
3416 if ((short)-Value == (int)-Value)
3417 Result = DAG.getTargetConstant(Value, Op.getValueType());
3425 Ops.push_back(Result);
3429 // Handle standard constraint letters.
3430 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
3433 // isLegalAddressingMode - Return true if the addressing mode represented
3434 // by AM is legal for this target, for a load/store of the specified type.
3435 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3436 const Type *Ty) const {
3437 // FIXME: PPC does not allow r+i addressing modes for vectors!
3439 // PPC allows a sign-extended 16-bit immediate field.
3440 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
3443 // No global is ever allowed as a base.
3447 // PPC only support r+r,
3449 case 0: // "r+i" or just "i", depending on HasBaseReg.
3452 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
3454 // Otherwise we have r+r or r+i.
3457 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
3459 // Allow 2*r as r+r.
3462 // No other scales are supported.
3469 /// isLegalAddressImmediate - Return true if the integer value can be used
3470 /// as the offset of the target addressing mode for load / store of the
3472 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
3473 // PPC allows a sign-extended 16-bit immediate field.
3474 return (V > -(1 << 16) && V < (1 << 16)-1);
3477 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
3481 SDOperand PPCTargetLowering::LowerFRAMEADDR(SDOperand Op, SelectionDAG &DAG)
3483 // Depths > 0 not supported yet!
3484 if (cast<ConstantSDNode>(Op.getOperand(0))->getValue() > 0)
3487 MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3488 bool isPPC64 = PtrVT == MVT::i64;
3490 MachineFunction &MF = DAG.getMachineFunction();
3491 MachineFrameInfo *MFI = MF.getFrameInfo();
3492 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
3493 && MFI->getStackSize();
3496 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
3499 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,