1 //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/PseudoSourceValue.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CallingConv.h"
29 #include "llvm/Constants.h"
30 #include "llvm/Function.h"
31 #include "llvm/Intrinsics.h"
32 #include "llvm/ParameterAttributes.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetOptions.h"
35 #include "llvm/Support/CommandLine.h"
38 static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc",
39 cl::desc("enable preincrement load/store generation on PPC (experimental)"),
42 PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
43 : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
47 // Use _setjmp/_longjmp instead of setjmp/longjmp.
48 setUseUnderscoreSetJmp(true);
49 setUseUnderscoreLongJmp(true);
51 // Set up the register classes.
52 addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
53 addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
54 addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
56 // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
57 setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
58 setLoadXAction(ISD::SEXTLOAD, MVT::i8, Expand);
60 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
62 // PowerPC has pre-inc load and store's.
63 setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
64 setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
65 setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
66 setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
67 setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
68 setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
69 setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
70 setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
71 setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
72 setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
74 // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
75 setConvertAction(MVT::ppcf128, MVT::f64, Expand);
76 setConvertAction(MVT::ppcf128, MVT::f32, Expand);
77 // This is used in the ppcf128->int sequence. Note it has different semantics
78 // from FP_ROUND: that rounds to nearest, this rounds to zero.
79 setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
81 // PowerPC has no SREM/UREM instructions
82 setOperationAction(ISD::SREM, MVT::i32, Expand);
83 setOperationAction(ISD::UREM, MVT::i32, Expand);
84 setOperationAction(ISD::SREM, MVT::i64, Expand);
85 setOperationAction(ISD::UREM, MVT::i64, Expand);
87 // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
88 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
89 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
90 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
91 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
92 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
93 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
94 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
95 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
97 // We don't support sin/cos/sqrt/fmod/pow
98 setOperationAction(ISD::FSIN , MVT::f64, Expand);
99 setOperationAction(ISD::FCOS , MVT::f64, Expand);
100 setOperationAction(ISD::FREM , MVT::f64, Expand);
101 setOperationAction(ISD::FPOW , MVT::f64, Expand);
102 setOperationAction(ISD::FSIN , MVT::f32, Expand);
103 setOperationAction(ISD::FCOS , MVT::f32, Expand);
104 setOperationAction(ISD::FREM , MVT::f32, Expand);
105 setOperationAction(ISD::FPOW , MVT::f32, Expand);
107 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
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);
128 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
130 // PowerPC does not have Select
131 setOperationAction(ISD::SELECT, MVT::i32, Expand);
132 setOperationAction(ISD::SELECT, MVT::i64, Expand);
133 setOperationAction(ISD::SELECT, MVT::f32, Expand);
134 setOperationAction(ISD::SELECT, MVT::f64, Expand);
136 // PowerPC wants to turn select_cc of FP into fsel when possible.
137 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
138 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
140 // PowerPC wants to optimize integer setcc a bit
141 setOperationAction(ISD::SETCC, MVT::i32, Custom);
143 // PowerPC does not have BRCOND which requires SetCC
144 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
146 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
148 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
149 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
151 // PowerPC does not have [U|S]INT_TO_FP
152 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
153 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
155 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
156 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
157 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
158 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
160 // We cannot sextinreg(i1). Expand to shifts.
161 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
163 // Support label based line numbers.
164 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
165 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
167 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
168 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
169 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
170 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
173 // We want to legalize GlobalAddress and ConstantPool nodes into the
174 // appropriate instructions to materialize the address.
175 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
176 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
177 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
178 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
179 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
180 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
181 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
182 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
184 // RET must be custom lowered, to meet ABI requirements.
185 setOperationAction(ISD::RET , MVT::Other, Custom);
188 setOperationAction(ISD::TRAP, MVT::Other, Legal);
190 // TRAMPOLINE is custom lowered.
191 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
193 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
194 setOperationAction(ISD::VASTART , MVT::Other, Custom);
196 // VAARG is custom lowered with ELF 32 ABI
197 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
198 setOperationAction(ISD::VAARG, MVT::Other, Custom);
200 setOperationAction(ISD::VAARG, MVT::Other, Expand);
202 // Use the default implementation.
203 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
204 setOperationAction(ISD::VAEND , MVT::Other, Expand);
205 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
206 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
207 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
208 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
210 // We want to custom lower some of our intrinsics.
211 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
213 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
214 // They also have instructions for converting between i64 and fp.
215 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
216 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
217 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
218 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
219 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
221 // FIXME: disable this lowered code. This generates 64-bit register values,
222 // and we don't model the fact that the top part is clobbered by calls. We
223 // need to flag these together so that the value isn't live across a call.
224 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
226 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
227 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
229 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
230 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
233 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
234 // 64-bit PowerPC implementations can support i64 types directly
235 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
236 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
237 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
238 // 64-bit PowerPC wants to expand i128 shifts itself.
239 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
240 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
241 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
243 // 32-bit PowerPC wants to expand i64 shifts itself.
244 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
245 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
246 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
249 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
250 // First set operation action for all vector types to expand. Then we
251 // will selectively turn on ones that can be effectively codegen'd.
252 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
253 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
254 MVT VT = (MVT::SimpleValueType)i;
256 // add/sub are legal for all supported vector VT's.
257 setOperationAction(ISD::ADD , VT, Legal);
258 setOperationAction(ISD::SUB , VT, Legal);
260 // We promote all shuffles to v16i8.
261 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
262 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
264 // We promote all non-typed operations to v4i32.
265 setOperationAction(ISD::AND , VT, Promote);
266 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
267 setOperationAction(ISD::OR , VT, Promote);
268 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
269 setOperationAction(ISD::XOR , VT, Promote);
270 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
271 setOperationAction(ISD::LOAD , VT, Promote);
272 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
273 setOperationAction(ISD::SELECT, VT, Promote);
274 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
275 setOperationAction(ISD::STORE, VT, Promote);
276 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
278 // No other operations are legal.
279 setOperationAction(ISD::MUL , VT, Expand);
280 setOperationAction(ISD::SDIV, VT, Expand);
281 setOperationAction(ISD::SREM, VT, Expand);
282 setOperationAction(ISD::UDIV, VT, Expand);
283 setOperationAction(ISD::UREM, VT, Expand);
284 setOperationAction(ISD::FDIV, VT, Expand);
285 setOperationAction(ISD::FNEG, VT, Expand);
286 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
287 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
288 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
289 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
290 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
291 setOperationAction(ISD::UDIVREM, VT, Expand);
292 setOperationAction(ISD::SDIVREM, VT, Expand);
293 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
294 setOperationAction(ISD::FPOW, VT, Expand);
295 setOperationAction(ISD::CTPOP, VT, Expand);
296 setOperationAction(ISD::CTLZ, VT, Expand);
297 setOperationAction(ISD::CTTZ, VT, Expand);
300 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
301 // with merges, splats, etc.
302 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
304 setOperationAction(ISD::AND , MVT::v4i32, Legal);
305 setOperationAction(ISD::OR , MVT::v4i32, Legal);
306 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
307 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
308 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
309 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
311 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
312 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
313 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
314 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
316 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
317 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
318 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
319 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
321 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
322 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
324 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
325 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
326 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
327 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
330 setShiftAmountType(MVT::i32);
331 setSetCCResultContents(ZeroOrOneSetCCResult);
333 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
334 setStackPointerRegisterToSaveRestore(PPC::X1);
335 setExceptionPointerRegister(PPC::X3);
336 setExceptionSelectorRegister(PPC::X4);
338 setStackPointerRegisterToSaveRestore(PPC::R1);
339 setExceptionPointerRegister(PPC::R3);
340 setExceptionSelectorRegister(PPC::R4);
343 // We have target-specific dag combine patterns for the following nodes:
344 setTargetDAGCombine(ISD::SINT_TO_FP);
345 setTargetDAGCombine(ISD::STORE);
346 setTargetDAGCombine(ISD::BR_CC);
347 setTargetDAGCombine(ISD::BSWAP);
349 // Darwin long double math library functions have $LDBL128 appended.
350 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
351 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
352 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
353 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
354 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
355 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
356 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
357 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
358 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
359 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
360 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
363 computeRegisterProperties();
366 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
367 /// function arguments in the caller parameter area.
368 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
369 TargetMachine &TM = getTargetMachine();
370 // Darwin passes everything on 4 byte boundary.
371 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
377 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
380 case PPCISD::FSEL: return "PPCISD::FSEL";
381 case PPCISD::FCFID: return "PPCISD::FCFID";
382 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
383 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
384 case PPCISD::STFIWX: return "PPCISD::STFIWX";
385 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
386 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
387 case PPCISD::VPERM: return "PPCISD::VPERM";
388 case PPCISD::Hi: return "PPCISD::Hi";
389 case PPCISD::Lo: return "PPCISD::Lo";
390 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
391 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
392 case PPCISD::SRL: return "PPCISD::SRL";
393 case PPCISD::SRA: return "PPCISD::SRA";
394 case PPCISD::SHL: return "PPCISD::SHL";
395 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
396 case PPCISD::STD_32: return "PPCISD::STD_32";
397 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
398 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
399 case PPCISD::MTCTR: return "PPCISD::MTCTR";
400 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
401 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
402 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
403 case PPCISD::MFCR: return "PPCISD::MFCR";
404 case PPCISD::VCMP: return "PPCISD::VCMP";
405 case PPCISD::VCMPo: return "PPCISD::VCMPo";
406 case PPCISD::LBRX: return "PPCISD::LBRX";
407 case PPCISD::STBRX: return "PPCISD::STBRX";
408 case PPCISD::LARX: return "PPCISD::LARX";
409 case PPCISD::STCX: return "PPCISD::STCX";
410 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
411 case PPCISD::MFFS: return "PPCISD::MFFS";
412 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
413 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
414 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
415 case PPCISD::MTFSF: return "PPCISD::MTFSF";
416 case PPCISD::TAILCALL: return "PPCISD::TAILCALL";
417 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
422 MVT PPCTargetLowering::getSetCCResultType(const SDValue &) const {
427 //===----------------------------------------------------------------------===//
428 // Node matching predicates, for use by the tblgen matching code.
429 //===----------------------------------------------------------------------===//
431 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
432 static bool isFloatingPointZero(SDValue Op) {
433 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
434 return CFP->getValueAPF().isZero();
435 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
436 // Maybe this has already been legalized into the constant pool?
437 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
438 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
439 return CFP->getValueAPF().isZero();
444 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
445 /// true if Op is undef or if it matches the specified value.
446 static bool isConstantOrUndef(SDValue Op, unsigned Val) {
447 return Op.getOpcode() == ISD::UNDEF ||
448 cast<ConstantSDNode>(Op)->getZExtValue() == Val;
451 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
452 /// VPKUHUM instruction.
453 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
455 for (unsigned i = 0; i != 16; ++i)
456 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
459 for (unsigned i = 0; i != 8; ++i)
460 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
461 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
467 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
468 /// VPKUWUM instruction.
469 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
471 for (unsigned i = 0; i != 16; i += 2)
472 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
473 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
476 for (unsigned i = 0; i != 8; i += 2)
477 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
478 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
479 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
480 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
486 /// isVMerge - Common function, used to match vmrg* shuffles.
488 static bool isVMerge(SDNode *N, unsigned UnitSize,
489 unsigned LHSStart, unsigned RHSStart) {
490 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
491 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
492 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
493 "Unsupported merge size!");
495 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
496 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
497 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
498 LHSStart+j+i*UnitSize) ||
499 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
500 RHSStart+j+i*UnitSize))
506 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
507 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
508 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
510 return isVMerge(N, UnitSize, 8, 24);
511 return isVMerge(N, UnitSize, 8, 8);
514 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
515 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
516 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
518 return isVMerge(N, UnitSize, 0, 16);
519 return isVMerge(N, UnitSize, 0, 0);
523 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
524 /// amount, otherwise return -1.
525 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
526 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
527 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
528 // Find the first non-undef value in the shuffle mask.
530 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
533 if (i == 16) return -1; // all undef.
535 // Otherwise, check to see if the rest of the elements are consequtively
536 // numbered from this value.
537 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getZExtValue();
538 if (ShiftAmt < i) return -1;
542 // Check the rest of the elements to see if they are consequtive.
543 for (++i; i != 16; ++i)
544 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
547 // Check the rest of the elements to see if they are consequtive.
548 for (++i; i != 16; ++i)
549 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
556 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
557 /// specifies a splat of a single element that is suitable for input to
558 /// VSPLTB/VSPLTH/VSPLTW.
559 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
560 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
561 N->getNumOperands() == 16 &&
562 (EltSize == 1 || EltSize == 2 || EltSize == 4));
564 // This is a splat operation if each element of the permute is the same, and
565 // if the value doesn't reference the second vector.
566 unsigned ElementBase = 0;
567 SDValue Elt = N->getOperand(0);
568 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
569 ElementBase = EltV->getZExtValue();
571 return false; // FIXME: Handle UNDEF elements too!
573 if (cast<ConstantSDNode>(Elt)->getZExtValue() >= 16)
576 // Check that they are consequtive.
577 for (unsigned i = 1; i != EltSize; ++i) {
578 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
579 cast<ConstantSDNode>(N->getOperand(i))->getZExtValue() != i+ElementBase)
583 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
584 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
585 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
586 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
587 "Invalid VECTOR_SHUFFLE mask!");
588 for (unsigned j = 0; j != EltSize; ++j)
589 if (N->getOperand(i+j) != N->getOperand(j))
596 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
598 bool PPC::isAllNegativeZeroVector(SDNode *N) {
599 assert(N->getOpcode() == ISD::BUILD_VECTOR);
600 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
601 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
602 return CFP->getValueAPF().isNegZero();
606 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
607 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
608 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
609 assert(isSplatShuffleMask(N, EltSize));
610 return cast<ConstantSDNode>(N->getOperand(0))->getZExtValue() / EltSize;
613 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
614 /// by using a vspltis[bhw] instruction of the specified element size, return
615 /// the constant being splatted. The ByteSize field indicates the number of
616 /// bytes of each element [124] -> [bhw].
617 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
620 // If ByteSize of the splat is bigger than the element size of the
621 // build_vector, then we have a case where we are checking for a splat where
622 // multiple elements of the buildvector are folded together into a single
623 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
624 unsigned EltSize = 16/N->getNumOperands();
625 if (EltSize < ByteSize) {
626 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
627 SDValue UniquedVals[4];
628 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
630 // See if all of the elements in the buildvector agree across.
631 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
632 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
633 // If the element isn't a constant, bail fully out.
634 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
637 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
638 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
639 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
640 return SDValue(); // no match.
643 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
644 // either constant or undef values that are identical for each chunk. See
645 // if these chunks can form into a larger vspltis*.
647 // Check to see if all of the leading entries are either 0 or -1. If
648 // neither, then this won't fit into the immediate field.
649 bool LeadingZero = true;
650 bool LeadingOnes = true;
651 for (unsigned i = 0; i != Multiple-1; ++i) {
652 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
654 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
655 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
657 // Finally, check the least significant entry.
659 if (UniquedVals[Multiple-1].getNode() == 0)
660 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
661 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
663 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
666 if (UniquedVals[Multiple-1].getNode() == 0)
667 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
668 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
669 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
670 return DAG.getTargetConstant(Val, MVT::i32);
676 // Check to see if this buildvec has a single non-undef value in its elements.
677 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
678 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
679 if (OpVal.getNode() == 0)
680 OpVal = N->getOperand(i);
681 else if (OpVal != N->getOperand(i))
685 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
687 unsigned ValSizeInBytes = 0;
689 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
690 Value = CN->getZExtValue();
691 ValSizeInBytes = CN->getValueType(0).getSizeInBits()/8;
692 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
693 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
694 Value = FloatToBits(CN->getValueAPF().convertToFloat());
698 // If the splat value is larger than the element value, then we can never do
699 // this splat. The only case that we could fit the replicated bits into our
700 // immediate field for would be zero, and we prefer to use vxor for it.
701 if (ValSizeInBytes < ByteSize) return SDValue();
703 // If the element value is larger than the splat value, cut it in half and
704 // check to see if the two halves are equal. Continue doing this until we
705 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
706 while (ValSizeInBytes > ByteSize) {
707 ValSizeInBytes >>= 1;
709 // If the top half equals the bottom half, we're still ok.
710 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
711 (Value & ((1 << (8*ValSizeInBytes))-1)))
715 // Properly sign extend the value.
716 int ShAmt = (4-ByteSize)*8;
717 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
719 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
720 if (MaskVal == 0) return SDValue();
722 // Finally, if this value fits in a 5 bit sext field, return it
723 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
724 return DAG.getTargetConstant(MaskVal, MVT::i32);
728 //===----------------------------------------------------------------------===//
729 // Addressing Mode Selection
730 //===----------------------------------------------------------------------===//
732 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
733 /// or 64-bit immediate, and if the value can be accurately represented as a
734 /// sign extension from a 16-bit value. If so, this returns true and the
736 static bool isIntS16Immediate(SDNode *N, short &Imm) {
737 if (N->getOpcode() != ISD::Constant)
740 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
741 if (N->getValueType(0) == MVT::i32)
742 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
744 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
746 static bool isIntS16Immediate(SDValue Op, short &Imm) {
747 return isIntS16Immediate(Op.getNode(), Imm);
751 /// SelectAddressRegReg - Given the specified addressed, check to see if it
752 /// can be represented as an indexed [r+r] operation. Returns false if it
753 /// can be more efficiently represented with [r+imm].
754 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
758 if (N.getOpcode() == ISD::ADD) {
759 if (isIntS16Immediate(N.getOperand(1), imm))
761 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
764 Base = N.getOperand(0);
765 Index = N.getOperand(1);
767 } else if (N.getOpcode() == ISD::OR) {
768 if (isIntS16Immediate(N.getOperand(1), imm))
769 return false; // r+i can fold it if we can.
771 // If this is an or of disjoint bitfields, we can codegen this as an add
772 // (for better address arithmetic) if the LHS and RHS of the OR are provably
774 APInt LHSKnownZero, LHSKnownOne;
775 APInt RHSKnownZero, RHSKnownOne;
776 DAG.ComputeMaskedBits(N.getOperand(0),
777 APInt::getAllOnesValue(N.getOperand(0)
778 .getValueSizeInBits()),
779 LHSKnownZero, LHSKnownOne);
781 if (LHSKnownZero.getBoolValue()) {
782 DAG.ComputeMaskedBits(N.getOperand(1),
783 APInt::getAllOnesValue(N.getOperand(1)
784 .getValueSizeInBits()),
785 RHSKnownZero, RHSKnownOne);
786 // If all of the bits are known zero on the LHS or RHS, the add won't
788 if (~(LHSKnownZero | RHSKnownZero) == 0) {
789 Base = N.getOperand(0);
790 Index = N.getOperand(1);
799 /// Returns true if the address N can be represented by a base register plus
800 /// a signed 16-bit displacement [r+imm], and if it is not better
801 /// represented as reg+reg.
802 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
803 SDValue &Base, SelectionDAG &DAG){
804 // If this can be more profitably realized as r+r, fail.
805 if (SelectAddressRegReg(N, Disp, Base, DAG))
808 if (N.getOpcode() == ISD::ADD) {
810 if (isIntS16Immediate(N.getOperand(1), imm)) {
811 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
812 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
813 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
815 Base = N.getOperand(0);
817 return true; // [r+i]
818 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
819 // Match LOAD (ADD (X, Lo(G))).
820 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
821 && "Cannot handle constant offsets yet!");
822 Disp = N.getOperand(1).getOperand(0); // The global address.
823 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
824 Disp.getOpcode() == ISD::TargetConstantPool ||
825 Disp.getOpcode() == ISD::TargetJumpTable);
826 Base = N.getOperand(0);
827 return true; // [&g+r]
829 } else if (N.getOpcode() == ISD::OR) {
831 if (isIntS16Immediate(N.getOperand(1), imm)) {
832 // If this is an or of disjoint bitfields, we can codegen this as an add
833 // (for better address arithmetic) if the LHS and RHS of the OR are
834 // provably disjoint.
835 APInt LHSKnownZero, LHSKnownOne;
836 DAG.ComputeMaskedBits(N.getOperand(0),
837 APInt::getAllOnesValue(N.getOperand(0)
838 .getValueSizeInBits()),
839 LHSKnownZero, LHSKnownOne);
841 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
842 // If all of the bits are known zero on the LHS or RHS, the add won't
844 Base = N.getOperand(0);
845 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
849 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
850 // Loading from a constant address.
852 // If this address fits entirely in a 16-bit sext immediate field, codegen
855 if (isIntS16Immediate(CN, Imm)) {
856 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
857 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
861 // Handle 32-bit sext immediates with LIS + addr mode.
862 if (CN->getValueType(0) == MVT::i32 ||
863 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
864 int Addr = (int)CN->getZExtValue();
866 // Otherwise, break this down into an LIS + disp.
867 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
869 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
870 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
871 Base = SDValue(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
876 Disp = DAG.getTargetConstant(0, getPointerTy());
877 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
878 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
881 return true; // [r+0]
884 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
885 /// represented as an indexed [r+r] operation.
886 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
889 // Check to see if we can easily represent this as an [r+r] address. This
890 // will fail if it thinks that the address is more profitably represented as
891 // reg+imm, e.g. where imm = 0.
892 if (SelectAddressRegReg(N, Base, Index, DAG))
895 // If the operand is an addition, always emit this as [r+r], since this is
896 // better (for code size, and execution, as the memop does the add for free)
897 // than emitting an explicit add.
898 if (N.getOpcode() == ISD::ADD) {
899 Base = N.getOperand(0);
900 Index = N.getOperand(1);
904 // Otherwise, do it the hard way, using R0 as the base register.
905 Base = DAG.getRegister(PPC::R0, N.getValueType());
910 /// SelectAddressRegImmShift - Returns true if the address N can be
911 /// represented by a base register plus a signed 14-bit displacement
912 /// [r+imm*4]. Suitable for use by STD and friends.
913 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
916 // If this can be more profitably realized as r+r, fail.
917 if (SelectAddressRegReg(N, Disp, Base, DAG))
920 if (N.getOpcode() == ISD::ADD) {
922 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
923 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
924 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
925 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
927 Base = N.getOperand(0);
929 return true; // [r+i]
930 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
931 // Match LOAD (ADD (X, Lo(G))).
932 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
933 && "Cannot handle constant offsets yet!");
934 Disp = N.getOperand(1).getOperand(0); // The global address.
935 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
936 Disp.getOpcode() == ISD::TargetConstantPool ||
937 Disp.getOpcode() == ISD::TargetJumpTable);
938 Base = N.getOperand(0);
939 return true; // [&g+r]
941 } else if (N.getOpcode() == ISD::OR) {
943 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
944 // If this is an or of disjoint bitfields, we can codegen this as an add
945 // (for better address arithmetic) if the LHS and RHS of the OR are
946 // provably disjoint.
947 APInt LHSKnownZero, LHSKnownOne;
948 DAG.ComputeMaskedBits(N.getOperand(0),
949 APInt::getAllOnesValue(N.getOperand(0)
950 .getValueSizeInBits()),
951 LHSKnownZero, LHSKnownOne);
952 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
953 // If all of the bits are known zero on the LHS or RHS, the add won't
955 Base = N.getOperand(0);
956 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
960 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
961 // Loading from a constant address. Verify low two bits are clear.
962 if ((CN->getZExtValue() & 3) == 0) {
963 // If this address fits entirely in a 14-bit sext immediate field, codegen
966 if (isIntS16Immediate(CN, Imm)) {
967 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
968 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
972 // Fold the low-part of 32-bit absolute addresses into addr mode.
973 if (CN->getValueType(0) == MVT::i32 ||
974 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
975 int Addr = (int)CN->getZExtValue();
977 // Otherwise, break this down into an LIS + disp.
978 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
980 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
981 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
982 Base = SDValue(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
988 Disp = DAG.getTargetConstant(0, getPointerTy());
989 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
990 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
993 return true; // [r+0]
997 /// getPreIndexedAddressParts - returns true by value, base pointer and
998 /// offset pointer and addressing mode by reference if the node's address
999 /// can be legally represented as pre-indexed load / store address.
1000 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1002 ISD::MemIndexedMode &AM,
1003 SelectionDAG &DAG) {
1004 // Disabled by default for now.
1005 if (!EnablePPCPreinc) return false;
1009 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1010 Ptr = LD->getBasePtr();
1011 VT = LD->getMemoryVT();
1013 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1015 Ptr = ST->getBasePtr();
1016 VT = ST->getMemoryVT();
1020 // PowerPC doesn't have preinc load/store instructions for vectors.
1024 // TODO: Check reg+reg first.
1026 // LDU/STU use reg+imm*4, others use reg+imm.
1027 if (VT != MVT::i64) {
1029 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1033 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1037 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1038 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1039 // sext i32 to i64 when addr mode is r+i.
1040 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1041 LD->getExtensionType() == ISD::SEXTLOAD &&
1042 isa<ConstantSDNode>(Offset))
1050 //===----------------------------------------------------------------------===//
1051 // LowerOperation implementation
1052 //===----------------------------------------------------------------------===//
1054 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1055 SelectionDAG &DAG) {
1056 MVT PtrVT = Op.getValueType();
1057 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1058 Constant *C = CP->getConstVal();
1059 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1060 SDValue Zero = DAG.getConstant(0, PtrVT);
1062 const TargetMachine &TM = DAG.getTarget();
1064 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1065 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1067 // If this is a non-darwin platform, we don't support non-static relo models
1069 if (TM.getRelocationModel() == Reloc::Static ||
1070 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1071 // Generate non-pic code that has direct accesses to the constant pool.
1072 // The address of the global is just (hi(&g)+lo(&g)).
1073 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1076 if (TM.getRelocationModel() == Reloc::PIC_) {
1077 // With PIC, the first instruction is actually "GR+hi(&G)".
1078 Hi = DAG.getNode(ISD::ADD, PtrVT,
1079 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1082 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1086 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1087 MVT PtrVT = Op.getValueType();
1088 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1089 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1090 SDValue Zero = DAG.getConstant(0, PtrVT);
1092 const TargetMachine &TM = DAG.getTarget();
1094 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1095 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1097 // If this is a non-darwin platform, we don't support non-static relo models
1099 if (TM.getRelocationModel() == Reloc::Static ||
1100 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1101 // Generate non-pic code that has direct accesses to the constant pool.
1102 // The address of the global is just (hi(&g)+lo(&g)).
1103 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1106 if (TM.getRelocationModel() == Reloc::PIC_) {
1107 // With PIC, the first instruction is actually "GR+hi(&G)".
1108 Hi = DAG.getNode(ISD::ADD, PtrVT,
1109 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1112 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1116 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1117 SelectionDAG &DAG) {
1118 assert(0 && "TLS not implemented for PPC.");
1119 return SDValue(); // Not reached
1122 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1123 SelectionDAG &DAG) {
1124 MVT PtrVT = Op.getValueType();
1125 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1126 GlobalValue *GV = GSDN->getGlobal();
1127 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1128 // If it's a debug information descriptor, don't mess with it.
1129 if (DAG.isVerifiedDebugInfoDesc(Op))
1131 SDValue Zero = DAG.getConstant(0, PtrVT);
1133 const TargetMachine &TM = DAG.getTarget();
1135 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1136 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1138 // If this is a non-darwin platform, we don't support non-static relo models
1140 if (TM.getRelocationModel() == Reloc::Static ||
1141 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1142 // Generate non-pic code that has direct accesses to globals.
1143 // The address of the global is just (hi(&g)+lo(&g)).
1144 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1147 if (TM.getRelocationModel() == Reloc::PIC_) {
1148 // With PIC, the first instruction is actually "GR+hi(&G)".
1149 Hi = DAG.getNode(ISD::ADD, PtrVT,
1150 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1153 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1155 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1158 // If the global is weak or external, we have to go through the lazy
1160 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1163 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1164 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1166 // If we're comparing for equality to zero, expose the fact that this is
1167 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1168 // fold the new nodes.
1169 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1170 if (C->isNullValue() && CC == ISD::SETEQ) {
1171 MVT VT = Op.getOperand(0).getValueType();
1172 SDValue Zext = Op.getOperand(0);
1173 if (VT.bitsLT(MVT::i32)) {
1175 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1177 unsigned Log2b = Log2_32(VT.getSizeInBits());
1178 SDValue Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1179 SDValue Scc = DAG.getNode(ISD::SRL, VT, Clz,
1180 DAG.getConstant(Log2b, MVT::i32));
1181 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1183 // Leave comparisons against 0 and -1 alone for now, since they're usually
1184 // optimized. FIXME: revisit this when we can custom lower all setcc
1186 if (C->isAllOnesValue() || C->isNullValue())
1190 // If we have an integer seteq/setne, turn it into a compare against zero
1191 // by xor'ing the rhs with the lhs, which is faster than setting a
1192 // condition register, reading it back out, and masking the correct bit. The
1193 // normal approach here uses sub to do this instead of xor. Using xor exposes
1194 // the result to other bit-twiddling opportunities.
1195 MVT LHSVT = Op.getOperand(0).getValueType();
1196 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1197 MVT VT = Op.getValueType();
1198 SDValue Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1200 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1205 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1206 int VarArgsFrameIndex,
1207 int VarArgsStackOffset,
1208 unsigned VarArgsNumGPR,
1209 unsigned VarArgsNumFPR,
1210 const PPCSubtarget &Subtarget) {
1212 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1213 return SDValue(); // Not reached
1216 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
1217 SDValue Chain = Op.getOperand(0);
1218 SDValue Trmp = Op.getOperand(1); // trampoline
1219 SDValue FPtr = Op.getOperand(2); // nested function
1220 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1222 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1223 bool isPPC64 = (PtrVT == MVT::i64);
1224 const Type *IntPtrTy =
1225 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType();
1227 TargetLowering::ArgListTy Args;
1228 TargetLowering::ArgListEntry Entry;
1230 Entry.Ty = IntPtrTy;
1231 Entry.Node = Trmp; Args.push_back(Entry);
1233 // TrampSize == (isPPC64 ? 48 : 40);
1234 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1235 isPPC64 ? MVT::i64 : MVT::i32);
1236 Args.push_back(Entry);
1238 Entry.Node = FPtr; Args.push_back(Entry);
1239 Entry.Node = Nest; Args.push_back(Entry);
1241 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1242 std::pair<SDValue, SDValue> CallResult =
1243 LowerCallTo(Chain, Op.getValueType().getTypeForMVT(), false, false,
1244 false, CallingConv::C, false,
1245 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1249 { CallResult.first, CallResult.second };
1251 return DAG.getMergeValues(Ops, 2, false);
1254 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1255 int VarArgsFrameIndex,
1256 int VarArgsStackOffset,
1257 unsigned VarArgsNumGPR,
1258 unsigned VarArgsNumFPR,
1259 const PPCSubtarget &Subtarget) {
1261 if (Subtarget.isMachoABI()) {
1262 // vastart just stores the address of the VarArgsFrameIndex slot into the
1263 // memory location argument.
1264 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1265 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1266 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1267 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1270 // For ELF 32 ABI we follow the layout of the va_list struct.
1271 // We suppose the given va_list is already allocated.
1274 // char gpr; /* index into the array of 8 GPRs
1275 // * stored in the register save area
1276 // * gpr=0 corresponds to r3,
1277 // * gpr=1 to r4, etc.
1279 // char fpr; /* index into the array of 8 FPRs
1280 // * stored in the register save area
1281 // * fpr=0 corresponds to f1,
1282 // * fpr=1 to f2, etc.
1284 // char *overflow_arg_area;
1285 // /* location on stack that holds
1286 // * the next overflow argument
1288 // char *reg_save_area;
1289 // /* where r3:r10 and f1:f8 (if saved)
1295 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1296 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1299 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1301 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1302 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1304 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1305 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1307 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1308 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1310 uint64_t FPROffset = 1;
1311 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1313 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1315 // Store first byte : number of int regs
1316 SDValue firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1317 Op.getOperand(1), SV, 0);
1318 uint64_t nextOffset = FPROffset;
1319 SDValue nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1322 // Store second byte : number of float regs
1323 SDValue secondStore =
1324 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1325 nextOffset += StackOffset;
1326 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1328 // Store second word : arguments given on stack
1329 SDValue thirdStore =
1330 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1331 nextOffset += FrameOffset;
1332 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1334 // Store third word : arguments given in registers
1335 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1339 #include "PPCGenCallingConv.inc"
1341 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1342 /// depending on which subtarget is selected.
1343 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1344 if (Subtarget.isMachoABI()) {
1345 static const unsigned FPR[] = {
1346 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1347 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1353 static const unsigned FPR[] = {
1354 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1360 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1362 static unsigned CalculateStackSlotSize(SDValue Arg, ISD::ArgFlagsTy Flags,
1363 bool isVarArg, unsigned PtrByteSize) {
1364 MVT ArgVT = Arg.getValueType();
1365 unsigned ArgSize =ArgVT.getSizeInBits()/8;
1366 if (Flags.isByVal())
1367 ArgSize = Flags.getByValSize();
1368 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1374 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op,
1376 int &VarArgsFrameIndex,
1377 int &VarArgsStackOffset,
1378 unsigned &VarArgsNumGPR,
1379 unsigned &VarArgsNumFPR,
1380 const PPCSubtarget &Subtarget) {
1381 // TODO: add description of PPC stack frame format, or at least some docs.
1383 MachineFunction &MF = DAG.getMachineFunction();
1384 MachineFrameInfo *MFI = MF.getFrameInfo();
1385 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1386 SmallVector<SDValue, 8> ArgValues;
1387 SDValue Root = Op.getOperand(0);
1388 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1390 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1391 bool isPPC64 = PtrVT == MVT::i64;
1392 bool isMachoABI = Subtarget.isMachoABI();
1393 bool isELF32_ABI = Subtarget.isELF32_ABI();
1394 // Potential tail calls could cause overwriting of argument stack slots.
1395 unsigned CC = MF.getFunction()->getCallingConv();
1396 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1397 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1399 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1400 // Area that is at least reserved in caller of this function.
1401 unsigned MinReservedArea = ArgOffset;
1403 static const unsigned GPR_32[] = { // 32-bit registers.
1404 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1405 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1407 static const unsigned GPR_64[] = { // 64-bit registers.
1408 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1409 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1412 static const unsigned *FPR = GetFPR(Subtarget);
1414 static const unsigned VR[] = {
1415 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1416 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1419 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1420 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1421 const unsigned Num_VR_Regs = array_lengthof( VR);
1423 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1425 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1427 // In 32-bit non-varargs functions, the stack space for vectors is after the
1428 // stack space for non-vectors. We do not use this space unless we have
1429 // too many vectors to fit in registers, something that only occurs in
1430 // constructed examples:), but we have to walk the arglist to figure
1431 // that out...for the pathological case, compute VecArgOffset as the
1432 // start of the vector parameter area. Computing VecArgOffset is the
1433 // entire point of the following loop.
1434 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1435 // to handle Elf here.
1436 unsigned VecArgOffset = ArgOffset;
1437 if (!isVarArg && !isPPC64) {
1438 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e;
1440 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1441 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1442 ISD::ArgFlagsTy Flags =
1443 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1445 if (Flags.isByVal()) {
1446 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1447 ObjSize = Flags.getByValSize();
1449 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1450 VecArgOffset += ArgSize;
1454 switch(ObjectVT.getSimpleVT()) {
1455 default: assert(0 && "Unhandled argument type!");
1458 VecArgOffset += isPPC64 ? 8 : 4;
1460 case MVT::i64: // PPC64
1468 // Nothing to do, we're only looking at Nonvector args here.
1473 // We've found where the vector parameter area in memory is. Skip the
1474 // first 12 parameters; these don't use that memory.
1475 VecArgOffset = ((VecArgOffset+15)/16)*16;
1476 VecArgOffset += 12*16;
1478 // Add DAG nodes to load the arguments or copy them out of registers. On
1479 // entry to a function on PPC, the arguments start after the linkage area,
1480 // although the first ones are often in registers.
1482 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1483 // represented with two words (long long or double) must be copied to an
1484 // even GPR_idx value or to an even ArgOffset value.
1486 SmallVector<SDValue, 8> MemOps;
1487 unsigned nAltivecParamsAtEnd = 0;
1488 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues() - 1;
1489 ArgNo != e; ++ArgNo) {
1491 bool needsLoad = false;
1492 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1493 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1494 unsigned ArgSize = ObjSize;
1495 ISD::ArgFlagsTy Flags =
1496 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1497 // See if next argument requires stack alignment in ELF
1498 bool Align = Flags.isSplit();
1500 unsigned CurArgOffset = ArgOffset;
1502 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1503 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1504 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1505 if (isVarArg || isPPC64) {
1506 MinReservedArea = ((MinReservedArea+15)/16)*16;
1507 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1511 } else nAltivecParamsAtEnd++;
1513 // Calculate min reserved area.
1514 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1519 // FIXME alignment for ELF may not be right
1520 // FIXME the codegen can be much improved in some cases.
1521 // We do not have to keep everything in memory.
1522 if (Flags.isByVal()) {
1523 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1524 ObjSize = Flags.getByValSize();
1525 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1526 // Double word align in ELF
1527 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1528 // Objects of size 1 and 2 are right justified, everything else is
1529 // left justified. This means the memory address is adjusted forwards.
1530 if (ObjSize==1 || ObjSize==2) {
1531 CurArgOffset = CurArgOffset + (4 - ObjSize);
1533 // The value of the object is its address.
1534 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1535 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1536 ArgValues.push_back(FIN);
1537 if (ObjSize==1 || ObjSize==2) {
1538 if (GPR_idx != Num_GPR_Regs) {
1539 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1540 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1541 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1542 SDValue Store = DAG.getTruncStore(Val.getValue(1), Val, FIN,
1543 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1544 MemOps.push_back(Store);
1546 if (isMachoABI) ArgOffset += PtrByteSize;
1548 ArgOffset += PtrByteSize;
1552 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1553 // Store whatever pieces of the object are in registers
1554 // to memory. ArgVal will be address of the beginning of
1556 if (GPR_idx != Num_GPR_Regs) {
1557 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1558 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1559 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1560 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1561 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1562 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1563 MemOps.push_back(Store);
1565 if (isMachoABI) ArgOffset += PtrByteSize;
1567 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1574 switch (ObjectVT.getSimpleVT()) {
1575 default: assert(0 && "Unhandled argument type!");
1578 // Double word align in ELF
1579 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1581 if (GPR_idx != Num_GPR_Regs) {
1582 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1583 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1584 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1588 ArgSize = PtrByteSize;
1590 // Stack align in ELF
1591 if (needsLoad && Align && isELF32_ABI)
1592 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1593 // All int arguments reserve stack space in Macho ABI.
1594 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1598 case MVT::i64: // PPC64
1599 if (GPR_idx != Num_GPR_Regs) {
1600 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1601 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1602 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1604 if (ObjectVT == MVT::i32) {
1605 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1606 // value to MVT::i64 and then truncate to the correct register size.
1608 ArgVal = DAG.getNode(ISD::AssertSext, MVT::i64, ArgVal,
1609 DAG.getValueType(ObjectVT));
1610 else if (Flags.isZExt())
1611 ArgVal = DAG.getNode(ISD::AssertZext, MVT::i64, ArgVal,
1612 DAG.getValueType(ObjectVT));
1614 ArgVal = DAG.getNode(ISD::TRUNCATE, MVT::i32, ArgVal);
1620 ArgSize = PtrByteSize;
1622 // All int arguments reserve stack space in Macho ABI.
1623 if (isMachoABI || needsLoad) ArgOffset += 8;
1628 // Every 4 bytes of argument space consumes one of the GPRs available for
1629 // argument passing.
1630 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1632 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1635 if (FPR_idx != Num_FPR_Regs) {
1637 if (ObjectVT == MVT::f32)
1638 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1640 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1641 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1642 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1648 // Stack align in ELF
1649 if (needsLoad && Align && isELF32_ABI)
1650 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1651 // All FP arguments reserve stack space in Macho ABI.
1652 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1658 // Note that vector arguments in registers don't reserve stack space,
1659 // except in varargs functions.
1660 if (VR_idx != Num_VR_Regs) {
1661 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1662 RegInfo.addLiveIn(VR[VR_idx], VReg);
1663 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1665 while ((ArgOffset % 16) != 0) {
1666 ArgOffset += PtrByteSize;
1667 if (GPR_idx != Num_GPR_Regs)
1671 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1675 if (!isVarArg && !isPPC64) {
1676 // Vectors go after all the nonvectors.
1677 CurArgOffset = VecArgOffset;
1680 // Vectors are aligned.
1681 ArgOffset = ((ArgOffset+15)/16)*16;
1682 CurArgOffset = ArgOffset;
1690 // We need to load the argument to a virtual register if we determined above
1691 // that we ran out of physical registers of the appropriate type.
1693 int FI = MFI->CreateFixedObject(ObjSize,
1694 CurArgOffset + (ArgSize - ObjSize),
1696 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1697 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1700 ArgValues.push_back(ArgVal);
1703 // Set the size that is at least reserved in caller of this function. Tail
1704 // call optimized function's reserved stack space needs to be aligned so that
1705 // taking the difference between two stack areas will result in an aligned
1707 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1708 // Add the Altivec parameters at the end, if needed.
1709 if (nAltivecParamsAtEnd) {
1710 MinReservedArea = ((MinReservedArea+15)/16)*16;
1711 MinReservedArea += 16*nAltivecParamsAtEnd;
1714 std::max(MinReservedArea,
1715 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1716 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1717 getStackAlignment();
1718 unsigned AlignMask = TargetAlign-1;
1719 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1720 FI->setMinReservedArea(MinReservedArea);
1722 // If the function takes variable number of arguments, make a frame index for
1723 // the start of the first vararg value... for expansion of llvm.va_start.
1728 VarArgsNumGPR = GPR_idx;
1729 VarArgsNumFPR = FPR_idx;
1731 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1733 depth = -(Num_GPR_Regs * PtrVT.getSizeInBits()/8 +
1734 Num_FPR_Regs * MVT(MVT::f64).getSizeInBits()/8 +
1735 PtrVT.getSizeInBits()/8);
1737 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1744 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1746 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1748 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1749 // stored to the VarArgsFrameIndex on the stack.
1751 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1752 SDValue Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1753 SDValue Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1754 MemOps.push_back(Store);
1755 // Increment the address by four for the next argument to store
1756 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1757 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1761 // If this function is vararg, store any remaining integer argument regs
1762 // to their spots on the stack so that they may be loaded by deferencing the
1763 // result of va_next.
1764 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1767 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1769 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1771 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1772 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1773 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1774 MemOps.push_back(Store);
1775 // Increment the address by four for the next argument to store
1776 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1777 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1780 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1783 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1784 SDValue Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1785 SDValue Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1786 MemOps.push_back(Store);
1787 // Increment the address by eight for the next argument to store
1788 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1790 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1793 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1795 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1797 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1798 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1799 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1800 MemOps.push_back(Store);
1801 // Increment the address by eight for the next argument to store
1802 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1804 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1809 if (!MemOps.empty())
1810 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1812 ArgValues.push_back(Root);
1814 // Return the new list of results.
1815 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1819 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
1822 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
1827 CallSDNode *TheCall,
1828 unsigned &nAltivecParamsAtEnd) {
1829 // Count how many bytes are to be pushed on the stack, including the linkage
1830 // area, and parameter passing area. We start with 24/48 bytes, which is
1831 // prereserved space for [SP][CR][LR][3 x unused].
1832 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1833 unsigned NumOps = TheCall->getNumArgs();
1834 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1836 // Add up all the space actually used.
1837 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1838 // they all go in registers, but we must reserve stack space for them for
1839 // possible use by the caller. In varargs or 64-bit calls, parameters are
1840 // assigned stack space in order, with padding so Altivec parameters are
1842 nAltivecParamsAtEnd = 0;
1843 for (unsigned i = 0; i != NumOps; ++i) {
1844 SDValue Arg = TheCall->getArg(i);
1845 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1846 MVT ArgVT = Arg.getValueType();
1847 // Varargs Altivec parameters are padded to a 16 byte boundary.
1848 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1849 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1850 if (!isVarArg && !isPPC64) {
1851 // Non-varargs Altivec parameters go after all the non-Altivec
1852 // parameters; handle those later so we know how much padding we need.
1853 nAltivecParamsAtEnd++;
1856 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1857 NumBytes = ((NumBytes+15)/16)*16;
1859 NumBytes += CalculateStackSlotSize(Arg, Flags, isVarArg, PtrByteSize);
1862 // Allow for Altivec parameters at the end, if needed.
1863 if (nAltivecParamsAtEnd) {
1864 NumBytes = ((NumBytes+15)/16)*16;
1865 NumBytes += 16*nAltivecParamsAtEnd;
1868 // The prolog code of the callee may store up to 8 GPR argument registers to
1869 // the stack, allowing va_start to index over them in memory if its varargs.
1870 // Because we cannot tell if this is needed on the caller side, we have to
1871 // conservatively assume that it is needed. As such, make sure we have at
1872 // least enough stack space for the caller to store the 8 GPRs.
1873 NumBytes = std::max(NumBytes,
1874 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1876 // Tail call needs the stack to be aligned.
1877 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1878 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1879 getStackAlignment();
1880 unsigned AlignMask = TargetAlign-1;
1881 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
1887 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
1888 /// adjusted to accomodate the arguments for the tailcall.
1889 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
1890 unsigned ParamSize) {
1892 if (!IsTailCall) return 0;
1894 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
1895 unsigned CallerMinReservedArea = FI->getMinReservedArea();
1896 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
1897 // Remember only if the new adjustement is bigger.
1898 if (SPDiff < FI->getTailCallSPDelta())
1899 FI->setTailCallSPDelta(SPDiff);
1904 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1905 /// following the call is a return. A function is eligible if caller/callee
1906 /// calling conventions match, currently only fastcc supports tail calls, and
1907 /// the function CALL is immediatly followed by a RET.
1909 PPCTargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1911 SelectionDAG& DAG) const {
1912 // Variable argument functions are not supported.
1913 if (!PerformTailCallOpt || TheCall->isVarArg())
1916 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1917 MachineFunction &MF = DAG.getMachineFunction();
1918 unsigned CallerCC = MF.getFunction()->getCallingConv();
1919 unsigned CalleeCC = TheCall->getCallingConv();
1920 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1921 // Functions containing by val parameters are not supported.
1922 for (unsigned i = 0; i != TheCall->getNumArgs(); i++) {
1923 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1924 if (Flags.isByVal()) return false;
1927 SDValue Callee = TheCall->getCallee();
1928 // Non PIC/GOT tail calls are supported.
1929 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1932 // At the moment we can only do local tail calls (in same module, hidden
1933 // or protected) if we are generating PIC.
1934 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1935 return G->getGlobal()->hasHiddenVisibility()
1936 || G->getGlobal()->hasProtectedVisibility();
1943 /// isCallCompatibleAddress - Return the immediate to use if the specified
1944 /// 32-bit value is representable in the immediate field of a BxA instruction.
1945 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
1946 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1949 int Addr = C->getZExtValue();
1950 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1951 (Addr << 6 >> 6) != Addr)
1952 return 0; // Top 6 bits have to be sext of immediate.
1954 return DAG.getConstant((int)C->getZExtValue() >> 2,
1955 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
1960 struct TailCallArgumentInfo {
1965 TailCallArgumentInfo() : FrameIdx(0) {}
1970 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
1972 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
1974 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
1975 SmallVector<SDValue, 8> &MemOpChains) {
1976 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
1977 SDValue Arg = TailCallArgs[i].Arg;
1978 SDValue FIN = TailCallArgs[i].FrameIdxOp;
1979 int FI = TailCallArgs[i].FrameIdx;
1980 // Store relative to framepointer.
1981 MemOpChains.push_back(DAG.getStore(Chain, Arg, FIN,
1982 PseudoSourceValue::getFixedStack(FI),
1987 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
1988 /// the appropriate stack slot for the tail call optimized function call.
1989 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
1990 MachineFunction &MF,
1998 // Calculate the new stack slot for the return address.
1999 int SlotSize = isPPC64 ? 8 : 4;
2000 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
2002 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2004 int NewFPLoc = SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64,
2006 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc);
2008 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2009 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2010 Chain = DAG.getStore(Chain, OldRetAddr, NewRetAddrFrIdx,
2011 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
2012 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2013 Chain = DAG.getStore(Chain, OldFP, NewFramePtrIdx,
2014 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
2019 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2020 /// the position of the argument.
2022 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2023 SDValue Arg, int SPDiff, unsigned ArgOffset,
2024 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2025 int Offset = ArgOffset + SPDiff;
2026 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2027 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
2028 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2029 SDValue FIN = DAG.getFrameIndex(FI, VT);
2030 TailCallArgumentInfo Info;
2032 Info.FrameIdxOp = FIN;
2034 TailCallArguments.push_back(Info);
2037 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2038 /// stack slot. Returns the chain as result and the loaded frame pointers in
2039 /// LROpOut/FPOpout. Used when tail calling.
2040 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2046 // Load the LR and FP stack slot for later adjusting.
2047 MVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2048 LROpOut = getReturnAddrFrameIndex(DAG);
2049 LROpOut = DAG.getLoad(VT, Chain, LROpOut, NULL, 0);
2050 Chain = SDValue(LROpOut.getNode(), 1);
2051 FPOpOut = getFramePointerFrameIndex(DAG);
2052 FPOpOut = DAG.getLoad(VT, Chain, FPOpOut, NULL, 0);
2053 Chain = SDValue(FPOpOut.getNode(), 1);
2058 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2059 /// by "Src" to address "Dst" of size "Size". Alignment information is
2060 /// specified by the specific parameter attribute. The copy will be passed as
2061 /// a byval function parameter.
2062 /// Sometimes what we are copying is the end of a larger object, the part that
2063 /// does not fit in registers.
2065 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2066 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2068 SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
2069 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(), false,
2073 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2076 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2077 SDValue Arg, SDValue PtrOff, int SPDiff,
2078 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2079 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2080 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2081 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2086 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2088 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2089 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2090 DAG.getConstant(ArgOffset, PtrVT));
2092 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
2093 // Calculate and remember argument location.
2094 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2098 SDValue PPCTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG,
2099 const PPCSubtarget &Subtarget,
2100 TargetMachine &TM) {
2101 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
2102 SDValue Chain = TheCall->getChain();
2103 bool isVarArg = TheCall->isVarArg();
2104 unsigned CC = TheCall->getCallingConv();
2105 bool isTailCall = TheCall->isTailCall()
2106 && CC == CallingConv::Fast && PerformTailCallOpt;
2107 SDValue Callee = TheCall->getCallee();
2108 unsigned NumOps = TheCall->getNumArgs();
2110 bool isMachoABI = Subtarget.isMachoABI();
2111 bool isELF32_ABI = Subtarget.isELF32_ABI();
2113 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2114 bool isPPC64 = PtrVT == MVT::i64;
2115 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2117 MachineFunction &MF = DAG.getMachineFunction();
2119 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
2120 // SelectExpr to use to put the arguments in the appropriate registers.
2121 std::vector<SDValue> args_to_use;
2123 // Mark this function as potentially containing a function that contains a
2124 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2125 // and restoring the callers stack pointer in this functions epilog. This is
2126 // done because by tail calling the called function might overwrite the value
2127 // in this function's (MF) stack pointer stack slot 0(SP).
2128 if (PerformTailCallOpt && CC==CallingConv::Fast)
2129 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2131 unsigned nAltivecParamsAtEnd = 0;
2133 // Count how many bytes are to be pushed on the stack, including the linkage
2134 // area, and parameter passing area. We start with 24/48 bytes, which is
2135 // prereserved space for [SP][CR][LR][3 x unused].
2137 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isMachoABI, isVarArg, CC,
2138 TheCall, nAltivecParamsAtEnd);
2140 // Calculate by how many bytes the stack has to be adjusted in case of tail
2141 // call optimization.
2142 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2144 // Adjust the stack pointer for the new arguments...
2145 // These operations are automatically eliminated by the prolog/epilog pass
2146 Chain = DAG.getCALLSEQ_START(Chain,
2147 DAG.getConstant(NumBytes, PtrVT));
2148 SDValue CallSeqStart = Chain;
2150 // Load the return address and frame pointer so it can be move somewhere else
2153 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp);
2155 // Set up a copy of the stack pointer for use loading and storing any
2156 // arguments that may not fit in the registers available for argument
2160 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2162 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2164 // Figure out which arguments are going to go in registers, and which in
2165 // memory. Also, if this is a vararg function, floating point operations
2166 // must be stored to our stack, and loaded into integer regs as well, if
2167 // any integer regs are available for argument passing.
2168 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2169 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2171 static const unsigned GPR_32[] = { // 32-bit registers.
2172 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2173 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2175 static const unsigned GPR_64[] = { // 64-bit registers.
2176 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2177 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2179 static const unsigned *FPR = GetFPR(Subtarget);
2181 static const unsigned VR[] = {
2182 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2183 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2185 const unsigned NumGPRs = array_lengthof(GPR_32);
2186 const unsigned NumFPRs = isMachoABI ? 13 : 8;
2187 const unsigned NumVRs = array_lengthof( VR);
2189 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2191 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
2192 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2194 SmallVector<SDValue, 8> MemOpChains;
2195 for (unsigned i = 0; i != NumOps; ++i) {
2197 SDValue Arg = TheCall->getArg(i);
2198 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2199 // See if next argument requires stack alignment in ELF
2200 bool Align = Flags.isSplit();
2202 // PtrOff will be used to store the current argument to the stack if a
2203 // register cannot be found for it.
2206 // Stack align in ELF 32
2207 if (isELF32_ABI && Align)
2208 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
2209 StackPtr.getValueType());
2211 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2213 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
2215 // On PPC64, promote integers to 64-bit values.
2216 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2217 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2218 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2219 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
2222 // FIXME Elf untested, what are alignment rules?
2223 // FIXME memcpy is used way more than necessary. Correctness first.
2224 if (Flags.isByVal()) {
2225 unsigned Size = Flags.getByValSize();
2226 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2227 if (Size==1 || Size==2) {
2228 // Very small objects are passed right-justified.
2229 // Everything else is passed left-justified.
2230 MVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2231 if (GPR_idx != NumGPRs) {
2232 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
2234 MemOpChains.push_back(Load.getValue(1));
2235 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2237 ArgOffset += PtrByteSize;
2239 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2240 SDValue AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
2241 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2242 CallSeqStart.getNode()->getOperand(0),
2244 // This must go outside the CALLSEQ_START..END.
2245 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2246 CallSeqStart.getNode()->getOperand(1));
2247 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2248 NewCallSeqStart.getNode());
2249 Chain = CallSeqStart = NewCallSeqStart;
2250 ArgOffset += PtrByteSize;
2254 // Copy entire object into memory. There are cases where gcc-generated
2255 // code assumes it is there, even if it could be put entirely into
2256 // registers. (This is not what the doc says.)
2257 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2258 CallSeqStart.getNode()->getOperand(0),
2260 // This must go outside the CALLSEQ_START..END.
2261 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2262 CallSeqStart.getNode()->getOperand(1));
2263 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2264 Chain = CallSeqStart = NewCallSeqStart;
2265 // And copy the pieces of it that fit into registers.
2266 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2267 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2268 SDValue AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
2269 if (GPR_idx != NumGPRs) {
2270 SDValue Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
2271 MemOpChains.push_back(Load.getValue(1));
2272 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2274 ArgOffset += PtrByteSize;
2276 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2283 switch (Arg.getValueType().getSimpleVT()) {
2284 default: assert(0 && "Unexpected ValueType for argument!");
2287 // Double word align in ELF
2288 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2289 if (GPR_idx != NumGPRs) {
2290 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2292 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2293 isPPC64, isTailCall, false, MemOpChains,
2297 if (inMem || isMachoABI) {
2298 // Stack align in ELF
2299 if (isELF32_ABI && Align)
2300 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2302 ArgOffset += PtrByteSize;
2307 if (FPR_idx != NumFPRs) {
2308 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
2311 SDValue Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2312 MemOpChains.push_back(Store);
2314 // Float varargs are always shadowed in available integer registers
2315 if (GPR_idx != NumGPRs) {
2316 SDValue Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
2317 MemOpChains.push_back(Load.getValue(1));
2318 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2321 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
2322 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
2323 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
2324 SDValue Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
2325 MemOpChains.push_back(Load.getValue(1));
2326 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2330 // If we have any FPRs remaining, we may also have GPRs remaining.
2331 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
2334 if (GPR_idx != NumGPRs)
2336 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
2337 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
2342 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2343 isPPC64, isTailCall, false, MemOpChains,
2347 if (inMem || isMachoABI) {
2348 // Stack align in ELF
2349 if (isELF32_ABI && Align)
2350 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2354 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2362 // These go aligned on the stack, or in the corresponding R registers
2363 // when within range. The Darwin PPC ABI doc claims they also go in
2364 // V registers; in fact gcc does this only for arguments that are
2365 // prototyped, not for those that match the ... We do it for all
2366 // arguments, seems to work.
2367 while (ArgOffset % 16 !=0) {
2368 ArgOffset += PtrByteSize;
2369 if (GPR_idx != NumGPRs)
2372 // We could elide this store in the case where the object fits
2373 // entirely in R registers. Maybe later.
2374 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2375 DAG.getConstant(ArgOffset, PtrVT));
2376 SDValue Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2377 MemOpChains.push_back(Store);
2378 if (VR_idx != NumVRs) {
2379 SDValue Load = DAG.getLoad(MVT::v4f32, Store, PtrOff, NULL, 0);
2380 MemOpChains.push_back(Load.getValue(1));
2381 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2384 for (unsigned i=0; i<16; i+=PtrByteSize) {
2385 if (GPR_idx == NumGPRs)
2387 SDValue Ix = DAG.getNode(ISD::ADD, PtrVT, PtrOff,
2388 DAG.getConstant(i, PtrVT));
2389 SDValue Load = DAG.getLoad(PtrVT, Store, Ix, NULL, 0);
2390 MemOpChains.push_back(Load.getValue(1));
2391 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2396 // Non-varargs Altivec params generally go in registers, but have
2397 // stack space allocated at the end.
2398 if (VR_idx != NumVRs) {
2399 // Doesn't have GPR space allocated.
2400 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2401 } else if (nAltivecParamsAtEnd==0) {
2402 // We are emitting Altivec params in order.
2403 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2404 isPPC64, isTailCall, true, MemOpChains,
2411 // If all Altivec parameters fit in registers, as they usually do,
2412 // they get stack space following the non-Altivec parameters. We
2413 // don't track this here because nobody below needs it.
2414 // If there are more Altivec parameters than fit in registers emit
2416 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2418 // Offset is aligned; skip 1st 12 params which go in V registers.
2419 ArgOffset = ((ArgOffset+15)/16)*16;
2421 for (unsigned i = 0; i != NumOps; ++i) {
2422 SDValue Arg = TheCall->getArg(i);
2423 MVT ArgType = Arg.getValueType();
2424 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2425 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2428 // We are emitting Altivec params in order.
2429 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2430 isPPC64, isTailCall, true, MemOpChains,
2438 if (!MemOpChains.empty())
2439 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2440 &MemOpChains[0], MemOpChains.size());
2442 // Build a sequence of copy-to-reg nodes chained together with token chain
2443 // and flag operands which copy the outgoing args into the appropriate regs.
2445 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2446 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
2448 InFlag = Chain.getValue(1);
2451 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2452 if (isVarArg && isELF32_ABI) {
2453 SDValue SetCR(DAG.getTargetNode(PPC::CRSET, MVT::i32), 0);
2454 Chain = DAG.getCopyToReg(Chain, PPC::CR1EQ, SetCR, InFlag);
2455 InFlag = Chain.getValue(1);
2458 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2459 // might overwrite each other in case of tail call optimization.
2461 SmallVector<SDValue, 8> MemOpChains2;
2462 // Do not flag preceeding copytoreg stuff together with the following stuff.
2464 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2466 if (!MemOpChains2.empty())
2467 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2468 &MemOpChains2[0], MemOpChains2.size());
2470 // Store the return address to the appropriate stack slot.
2471 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2472 isPPC64, isMachoABI);
2475 // Emit callseq_end just before tailcall node.
2477 SmallVector<SDValue, 8> CallSeqOps;
2478 SDVTList CallSeqNodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2479 CallSeqOps.push_back(Chain);
2480 CallSeqOps.push_back(DAG.getIntPtrConstant(NumBytes));
2481 CallSeqOps.push_back(DAG.getIntPtrConstant(0));
2482 if (InFlag.getNode())
2483 CallSeqOps.push_back(InFlag);
2484 Chain = DAG.getNode(ISD::CALLSEQ_END, CallSeqNodeTys, &CallSeqOps[0],
2486 InFlag = Chain.getValue(1);
2489 std::vector<MVT> NodeTys;
2490 NodeTys.push_back(MVT::Other); // Returns a chain
2491 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2493 SmallVector<SDValue, 8> Ops;
2494 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2496 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2497 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2498 // node so that legalize doesn't hack it.
2499 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2500 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2501 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2502 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2503 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2504 // If this is an absolute destination address, use the munged value.
2505 Callee = SDValue(Dest, 0);
2507 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2508 // to do the call, we can't use PPCISD::CALL.
2509 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2510 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps,
2511 2 + (InFlag.getNode() != 0));
2512 InFlag = Chain.getValue(1);
2514 // Copy the callee address into R12/X12 on darwin.
2516 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2517 Chain = DAG.getCopyToReg(Chain, Reg, Callee, InFlag);
2518 InFlag = Chain.getValue(1);
2522 NodeTys.push_back(MVT::Other);
2523 NodeTys.push_back(MVT::Flag);
2524 Ops.push_back(Chain);
2525 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2527 // Add CTR register as callee so a bctr can be emitted later.
2529 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
2532 // If this is a direct call, pass the chain and the callee.
2533 if (Callee.getNode()) {
2534 Ops.push_back(Chain);
2535 Ops.push_back(Callee);
2537 // If this is a tail call add stack pointer delta.
2539 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2541 // Add argument registers to the end of the list so that they are known live
2543 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2544 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2545 RegsToPass[i].second.getValueType()));
2547 // When performing tail call optimization the callee pops its arguments off
2548 // the stack. Account for this here so these bytes can be pushed back on in
2549 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2550 int BytesCalleePops =
2551 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2553 if (InFlag.getNode())
2554 Ops.push_back(InFlag);
2558 assert(InFlag.getNode() &&
2559 "Flag must be set. Depend on flag being set in LowerRET");
2560 Chain = DAG.getNode(PPCISD::TAILCALL,
2561 TheCall->getVTList(), &Ops[0], Ops.size());
2562 return SDValue(Chain.getNode(), Op.getResNo());
2565 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
2566 InFlag = Chain.getValue(1);
2568 Chain = DAG.getCALLSEQ_END(Chain,
2569 DAG.getConstant(NumBytes, PtrVT),
2570 DAG.getConstant(BytesCalleePops, PtrVT),
2572 if (TheCall->getValueType(0) != MVT::Other)
2573 InFlag = Chain.getValue(1);
2575 SmallVector<SDValue, 16> ResultVals;
2576 SmallVector<CCValAssign, 16> RVLocs;
2577 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
2578 CCState CCInfo(CallerCC, isVarArg, TM, RVLocs);
2579 CCInfo.AnalyzeCallResult(TheCall, RetCC_PPC);
2581 // Copy all of the result registers out of their specified physreg.
2582 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2583 CCValAssign &VA = RVLocs[i];
2584 MVT VT = VA.getValVT();
2585 assert(VA.isRegLoc() && "Can only return in registers!");
2586 Chain = DAG.getCopyFromReg(Chain, VA.getLocReg(), VT, InFlag).getValue(1);
2587 ResultVals.push_back(Chain.getValue(0));
2588 InFlag = Chain.getValue(2);
2591 // If the function returns void, just return the chain.
2595 // Otherwise, merge everything together with a MERGE_VALUES node.
2596 ResultVals.push_back(Chain);
2597 SDValue Res = DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
2599 return Res.getValue(Op.getResNo());
2602 SDValue PPCTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG,
2603 TargetMachine &TM) {
2604 SmallVector<CCValAssign, 16> RVLocs;
2605 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2606 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2607 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2608 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_PPC);
2610 // If this is the first return lowered for this function, add the regs to the
2611 // liveout set for the function.
2612 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2613 for (unsigned i = 0; i != RVLocs.size(); ++i)
2614 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2617 SDValue Chain = Op.getOperand(0);
2619 Chain = GetPossiblePreceedingTailCall(Chain, PPCISD::TAILCALL);
2620 if (Chain.getOpcode() == PPCISD::TAILCALL) {
2621 SDValue TailCall = Chain;
2622 SDValue TargetAddress = TailCall.getOperand(1);
2623 SDValue StackAdjustment = TailCall.getOperand(2);
2625 assert(((TargetAddress.getOpcode() == ISD::Register &&
2626 cast<RegisterSDNode>(TargetAddress)->getReg() == PPC::CTR) ||
2627 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
2628 TargetAddress.getOpcode() == ISD::TargetGlobalAddress ||
2629 isa<ConstantSDNode>(TargetAddress)) &&
2630 "Expecting an global address, external symbol, absolute value or register");
2632 assert(StackAdjustment.getOpcode() == ISD::Constant &&
2633 "Expecting a const value");
2635 SmallVector<SDValue,8> Operands;
2636 Operands.push_back(Chain.getOperand(0));
2637 Operands.push_back(TargetAddress);
2638 Operands.push_back(StackAdjustment);
2639 // Copy registers used by the call. Last operand is a flag so it is not
2641 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
2642 Operands.push_back(Chain.getOperand(i));
2644 return DAG.getNode(PPCISD::TC_RETURN, MVT::Other, &Operands[0],
2650 // Copy the result values into the output registers.
2651 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2652 CCValAssign &VA = RVLocs[i];
2653 assert(VA.isRegLoc() && "Can only return in registers!");
2654 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2655 Flag = Chain.getValue(1);
2659 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2661 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2664 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
2665 const PPCSubtarget &Subtarget) {
2666 // When we pop the dynamic allocation we need to restore the SP link.
2668 // Get the corect type for pointers.
2669 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2671 // Construct the stack pointer operand.
2672 bool IsPPC64 = Subtarget.isPPC64();
2673 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2674 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
2676 // Get the operands for the STACKRESTORE.
2677 SDValue Chain = Op.getOperand(0);
2678 SDValue SaveSP = Op.getOperand(1);
2680 // Load the old link SP.
2681 SDValue LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2683 // Restore the stack pointer.
2684 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2686 // Store the old link SP.
2687 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2693 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
2694 MachineFunction &MF = DAG.getMachineFunction();
2695 bool IsPPC64 = PPCSubTarget.isPPC64();
2696 bool isMachoABI = PPCSubTarget.isMachoABI();
2697 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2699 // Get current frame pointer save index. The users of this index will be
2700 // primarily DYNALLOC instructions.
2701 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2702 int RASI = FI->getReturnAddrSaveIndex();
2704 // If the frame pointer save index hasn't been defined yet.
2706 // Find out what the fix offset of the frame pointer save area.
2707 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isMachoABI);
2708 // Allocate the frame index for frame pointer save area.
2709 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset);
2711 FI->setReturnAddrSaveIndex(RASI);
2713 return DAG.getFrameIndex(RASI, PtrVT);
2717 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
2718 MachineFunction &MF = DAG.getMachineFunction();
2719 bool IsPPC64 = PPCSubTarget.isPPC64();
2720 bool isMachoABI = PPCSubTarget.isMachoABI();
2721 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2723 // Get current frame pointer save index. The users of this index will be
2724 // primarily DYNALLOC instructions.
2725 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2726 int FPSI = FI->getFramePointerSaveIndex();
2728 // If the frame pointer save index hasn't been defined yet.
2730 // Find out what the fix offset of the frame pointer save area.
2731 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2733 // Allocate the frame index for frame pointer save area.
2734 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2736 FI->setFramePointerSaveIndex(FPSI);
2738 return DAG.getFrameIndex(FPSI, PtrVT);
2741 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
2743 const PPCSubtarget &Subtarget) {
2745 SDValue Chain = Op.getOperand(0);
2746 SDValue Size = Op.getOperand(1);
2748 // Get the corect type for pointers.
2749 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2751 SDValue NegSize = DAG.getNode(ISD::SUB, PtrVT,
2752 DAG.getConstant(0, PtrVT), Size);
2753 // Construct a node for the frame pointer save index.
2754 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
2755 // Build a DYNALLOC node.
2756 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
2757 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2758 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2761 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2763 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
2764 // Not FP? Not a fsel.
2765 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
2766 !Op.getOperand(2).getValueType().isFloatingPoint())
2769 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2771 // Cannot handle SETEQ/SETNE.
2772 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDValue();
2774 MVT ResVT = Op.getValueType();
2775 MVT CmpVT = Op.getOperand(0).getValueType();
2776 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2777 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
2779 // If the RHS of the comparison is a 0.0, we don't need to do the
2780 // subtraction at all.
2781 if (isFloatingPointZero(RHS))
2783 default: break; // SETUO etc aren't handled by fsel.
2786 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2789 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2790 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2791 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2794 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2797 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2798 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2799 return DAG.getNode(PPCISD::FSEL, ResVT,
2800 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2805 default: break; // SETUO etc aren't handled by fsel.
2808 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2809 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2810 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2811 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2814 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2815 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2816 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2817 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2820 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2821 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2822 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2823 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2826 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2827 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2828 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2829 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2834 // FIXME: Split this code up when LegalizeDAGTypes lands.
2835 SDValue PPCTargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
2836 assert(Op.getOperand(0).getValueType().isFloatingPoint());
2837 SDValue Src = Op.getOperand(0);
2838 if (Src.getValueType() == MVT::f32)
2839 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2842 switch (Op.getValueType().getSimpleVT()) {
2843 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2845 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2848 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2852 // Convert the FP value to an int value through memory.
2853 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
2855 // Emit a store to the stack slot.
2856 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2858 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2860 if (Op.getValueType() == MVT::i32)
2861 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2862 DAG.getConstant(4, FIPtr.getValueType()));
2863 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2866 SDValue PPCTargetLowering::LowerFP_ROUND_INREG(SDValue Op,
2867 SelectionDAG &DAG) {
2868 assert(Op.getValueType() == MVT::ppcf128);
2869 SDNode *Node = Op.getNode();
2870 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2871 assert(Node->getOperand(0).getNode()->getOpcode() == ISD::BUILD_PAIR);
2872 SDValue Lo = Node->getOperand(0).getNode()->getOperand(0);
2873 SDValue Hi = Node->getOperand(0).getNode()->getOperand(1);
2875 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2876 // of the long double, and puts FPSCR back the way it was. We do not
2877 // actually model FPSCR.
2878 std::vector<MVT> NodeTys;
2879 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
2881 NodeTys.push_back(MVT::f64); // Return register
2882 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2883 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2884 MFFSreg = Result.getValue(0);
2885 InFlag = Result.getValue(1);
2888 NodeTys.push_back(MVT::Flag); // Returns a flag
2889 Ops[0] = DAG.getConstant(31, MVT::i32);
2891 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2892 InFlag = Result.getValue(0);
2895 NodeTys.push_back(MVT::Flag); // Returns a flag
2896 Ops[0] = DAG.getConstant(30, MVT::i32);
2898 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2899 InFlag = Result.getValue(0);
2902 NodeTys.push_back(MVT::f64); // result of add
2903 NodeTys.push_back(MVT::Flag); // Returns a flag
2907 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2908 FPreg = Result.getValue(0);
2909 InFlag = Result.getValue(1);
2912 NodeTys.push_back(MVT::f64);
2913 Ops[0] = DAG.getConstant(1, MVT::i32);
2917 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2918 FPreg = Result.getValue(0);
2920 // We know the low half is about to be thrown away, so just use something
2922 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2925 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2926 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2927 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2930 if (Op.getOperand(0).getValueType() == MVT::i64) {
2931 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2932 SDValue FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2933 if (Op.getValueType() == MVT::f32)
2934 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2938 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2939 "Unhandled SINT_TO_FP type in custom expander!");
2940 // Since we only generate this in 64-bit mode, we can take advantage of
2941 // 64-bit registers. In particular, sign extend the input value into the
2942 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2943 // then lfd it and fcfid it.
2944 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2945 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2946 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2947 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2949 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2952 // STD the extended value into the stack slot.
2953 MachineMemOperand MO(PseudoSourceValue::getFixedStack(FrameIdx),
2954 MachineMemOperand::MOStore, 0, 8, 8);
2955 SDValue Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2956 DAG.getEntryNode(), Ext64, FIdx,
2957 DAG.getMemOperand(MO));
2958 // Load the value as a double.
2959 SDValue Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2961 // FCFID it and return it.
2962 SDValue FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2963 if (Op.getValueType() == MVT::f32)
2964 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2968 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
2970 The rounding mode is in bits 30:31 of FPSR, and has the following
2977 FLT_ROUNDS, on the other hand, expects the following:
2984 To perform the conversion, we do:
2985 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2988 MachineFunction &MF = DAG.getMachineFunction();
2989 MVT VT = Op.getValueType();
2990 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2991 std::vector<MVT> NodeTys;
2992 SDValue MFFSreg, InFlag;
2994 // Save FP Control Word to register
2995 NodeTys.push_back(MVT::f64); // return register
2996 NodeTys.push_back(MVT::Flag); // unused in this context
2997 SDValue Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2999 // Save FP register to stack slot
3000 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3001 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3002 SDValue Store = DAG.getStore(DAG.getEntryNode(), Chain,
3003 StackSlot, NULL, 0);
3005 // Load FP Control Word from low 32 bits of stack slot.
3006 SDValue Four = DAG.getConstant(4, PtrVT);
3007 SDValue Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
3008 SDValue CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
3010 // Transform as necessary
3012 DAG.getNode(ISD::AND, MVT::i32,
3013 CWD, DAG.getConstant(3, MVT::i32));
3015 DAG.getNode(ISD::SRL, MVT::i32,
3016 DAG.getNode(ISD::AND, MVT::i32,
3017 DAG.getNode(ISD::XOR, MVT::i32,
3018 CWD, DAG.getConstant(3, MVT::i32)),
3019 DAG.getConstant(3, MVT::i32)),
3020 DAG.getConstant(1, MVT::i8));
3023 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
3025 return DAG.getNode((VT.getSizeInBits() < 16 ?
3026 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
3029 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3030 MVT VT = Op.getValueType();
3031 unsigned BitWidth = VT.getSizeInBits();
3032 assert(Op.getNumOperands() == 3 &&
3033 VT == Op.getOperand(1).getValueType() &&
3036 // Expand into a bunch of logical ops. Note that these ops
3037 // depend on the PPC behavior for oversized shift amounts.
3038 SDValue Lo = Op.getOperand(0);
3039 SDValue Hi = Op.getOperand(1);
3040 SDValue Amt = Op.getOperand(2);
3041 MVT AmtVT = Amt.getValueType();
3043 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3044 DAG.getConstant(BitWidth, AmtVT), Amt);
3045 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
3046 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
3047 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3048 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3049 DAG.getConstant(-BitWidth, AmtVT));
3050 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
3051 SDValue OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
3052 SDValue OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
3053 SDValue OutOps[] = { OutLo, OutHi };
3054 return DAG.getMergeValues(OutOps, 2);
3057 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3058 MVT VT = Op.getValueType();
3059 unsigned BitWidth = VT.getSizeInBits();
3060 assert(Op.getNumOperands() == 3 &&
3061 VT == Op.getOperand(1).getValueType() &&
3064 // Expand into a bunch of logical ops. Note that these ops
3065 // depend on the PPC behavior for oversized shift amounts.
3066 SDValue Lo = Op.getOperand(0);
3067 SDValue Hi = Op.getOperand(1);
3068 SDValue Amt = Op.getOperand(2);
3069 MVT AmtVT = Amt.getValueType();
3071 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3072 DAG.getConstant(BitWidth, AmtVT), Amt);
3073 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
3074 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
3075 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3076 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3077 DAG.getConstant(-BitWidth, AmtVT));
3078 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
3079 SDValue OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
3080 SDValue OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
3081 SDValue OutOps[] = { OutLo, OutHi };
3082 return DAG.getMergeValues(OutOps, 2);
3085 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3086 MVT VT = Op.getValueType();
3087 unsigned BitWidth = VT.getSizeInBits();
3088 assert(Op.getNumOperands() == 3 &&
3089 VT == Op.getOperand(1).getValueType() &&
3092 // Expand into a bunch of logical ops, followed by a select_cc.
3093 SDValue Lo = Op.getOperand(0);
3094 SDValue Hi = Op.getOperand(1);
3095 SDValue Amt = Op.getOperand(2);
3096 MVT AmtVT = Amt.getValueType();
3098 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3099 DAG.getConstant(BitWidth, AmtVT), Amt);
3100 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
3101 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
3102 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3103 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3104 DAG.getConstant(-BitWidth, AmtVT));
3105 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
3106 SDValue OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
3107 SDValue OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
3108 Tmp4, Tmp6, ISD::SETLE);
3109 SDValue OutOps[] = { OutLo, OutHi };
3110 return DAG.getMergeValues(OutOps, 2);
3113 //===----------------------------------------------------------------------===//
3114 // Vector related lowering.
3117 // If this is a vector of constants or undefs, get the bits. A bit in
3118 // UndefBits is set if the corresponding element of the vector is an
3119 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
3120 // zero. Return true if this is not an array of constants, false if it is.
3122 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
3123 uint64_t UndefBits[2]) {
3124 // Start with zero'd results.
3125 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
3127 unsigned EltBitSize = BV->getOperand(0).getValueType().getSizeInBits();
3128 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
3129 SDValue OpVal = BV->getOperand(i);
3131 unsigned PartNo = i >= e/2; // In the upper 128 bits?
3132 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
3134 uint64_t EltBits = 0;
3135 if (OpVal.getOpcode() == ISD::UNDEF) {
3136 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
3137 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
3139 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
3140 EltBits = CN->getZExtValue() & (~0U >> (32-EltBitSize));
3141 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
3142 assert(CN->getValueType(0) == MVT::f32 &&
3143 "Only one legal FP vector type!");
3144 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
3146 // Nonconstant element.
3150 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
3153 //printf("%llx %llx %llx %llx\n",
3154 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
3158 // If this is a splat (repetition) of a value across the whole vector, return
3159 // the smallest size that splats it. For example, "0x01010101010101..." is a
3160 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
3161 // SplatSize = 1 byte.
3162 static bool isConstantSplat(const uint64_t Bits128[2],
3163 const uint64_t Undef128[2],
3164 unsigned &SplatBits, unsigned &SplatUndef,
3165 unsigned &SplatSize) {
3167 // Don't let undefs prevent splats from matching. See if the top 64-bits are
3168 // the same as the lower 64-bits, ignoring undefs.
3169 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
3170 return false; // Can't be a splat if two pieces don't match.
3172 uint64_t Bits64 = Bits128[0] | Bits128[1];
3173 uint64_t Undef64 = Undef128[0] & Undef128[1];
3175 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
3177 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
3178 return false; // Can't be a splat if two pieces don't match.
3180 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
3181 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
3183 // If the top 16-bits are different than the lower 16-bits, ignoring
3184 // undefs, we have an i32 splat.
3185 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
3187 SplatUndef = Undef32;
3192 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
3193 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
3195 // If the top 8-bits are different than the lower 8-bits, ignoring
3196 // undefs, we have an i16 splat.
3197 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
3199 SplatUndef = Undef16;
3204 // Otherwise, we have an 8-bit splat.
3205 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
3206 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
3211 /// BuildSplatI - Build a canonical splati of Val with an element size of
3212 /// SplatSize. Cast the result to VT.
3213 static SDValue BuildSplatI(int Val, unsigned SplatSize, MVT VT,
3214 SelectionDAG &DAG) {
3215 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3217 static const MVT VTys[] = { // canonical VT to use for each size.
3218 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3221 MVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3223 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3227 MVT CanonicalVT = VTys[SplatSize-1];
3229 // Build a canonical splat for this value.
3230 SDValue Elt = DAG.getConstant(Val, CanonicalVT.getVectorElementType());
3231 SmallVector<SDValue, 8> Ops;
3232 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3233 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
3234 &Ops[0], Ops.size());
3235 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
3238 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3239 /// specified intrinsic ID.
3240 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3242 MVT DestVT = MVT::Other) {
3243 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3244 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
3245 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3248 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3249 /// specified intrinsic ID.
3250 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3251 SDValue Op2, SelectionDAG &DAG,
3252 MVT DestVT = MVT::Other) {
3253 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3254 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
3255 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3259 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3260 /// amount. The result has the specified value type.
3261 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3262 MVT VT, SelectionDAG &DAG) {
3263 // Force LHS/RHS to be the right type.
3264 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
3265 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
3268 for (unsigned i = 0; i != 16; ++i)
3269 Ops[i] = DAG.getConstant(i+Amt, MVT::i8);
3270 SDValue T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
3271 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
3272 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
3275 // If this is a case we can't handle, return null and let the default
3276 // expansion code take care of it. If we CAN select this case, and if it
3277 // selects to a single instruction, return Op. Otherwise, if we can codegen
3278 // this case more efficiently than a constant pool load, lower it to the
3279 // sequence of ops that should be used.
3280 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3281 SelectionDAG &DAG) {
3282 // If this is a vector of constants or undefs, get the bits. A bit in
3283 // UndefBits is set if the corresponding element of the vector is an
3284 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
3286 uint64_t VectorBits[2];
3287 uint64_t UndefBits[2];
3288 if (GetConstantBuildVectorBits(Op.getNode(), VectorBits, UndefBits))
3289 return SDValue(); // Not a constant vector.
3291 // If this is a splat (repetition) of a value across the whole vector, return
3292 // the smallest size that splats it. For example, "0x01010101010101..." is a
3293 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
3294 // SplatSize = 1 byte.
3295 unsigned SplatBits, SplatUndef, SplatSize;
3296 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
3297 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
3299 // First, handle single instruction cases.
3302 if (SplatBits == 0) {
3303 // Canonicalize all zero vectors to be v4i32.
3304 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3305 SDValue Z = DAG.getConstant(0, MVT::i32);
3306 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
3307 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
3312 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3313 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
3314 if (SextVal >= -16 && SextVal <= 15)
3315 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
3318 // Two instruction sequences.
3320 // If this value is in the range [-32,30] and is even, use:
3321 // tmp = VSPLTI[bhw], result = add tmp, tmp
3322 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3323 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG);
3324 Res = DAG.getNode(ISD::ADD, Res.getValueType(), Res, Res);
3325 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3328 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3329 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3331 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3332 // Make -1 and vspltisw -1:
3333 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
3335 // Make the VSLW intrinsic, computing 0x8000_0000.
3336 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3339 // xor by OnesV to invert it.
3340 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
3341 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3344 // Check to see if this is a wide variety of vsplti*, binop self cases.
3345 unsigned SplatBitSize = SplatSize*8;
3346 static const signed char SplatCsts[] = {
3347 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3348 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3351 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3352 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3353 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3354 int i = SplatCsts[idx];
3356 // Figure out what shift amount will be used by altivec if shifted by i in
3358 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3360 // vsplti + shl self.
3361 if (SextVal == (i << (int)TypeShiftAmt)) {
3362 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3363 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3364 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3365 Intrinsic::ppc_altivec_vslw
3367 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3368 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3371 // vsplti + srl self.
3372 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3373 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3374 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3375 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3376 Intrinsic::ppc_altivec_vsrw
3378 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3379 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3382 // vsplti + sra self.
3383 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3384 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3385 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3386 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3387 Intrinsic::ppc_altivec_vsraw
3389 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3390 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3393 // vsplti + rol self.
3394 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3395 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3396 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3397 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3398 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3399 Intrinsic::ppc_altivec_vrlw
3401 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3402 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3405 // t = vsplti c, result = vsldoi t, t, 1
3406 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3407 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3408 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
3410 // t = vsplti c, result = vsldoi t, t, 2
3411 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3412 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3413 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
3415 // t = vsplti c, result = vsldoi t, t, 3
3416 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3417 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3418 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
3422 // Three instruction sequences.
3424 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3425 if (SextVal >= 0 && SextVal <= 31) {
3426 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
3427 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
3428 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
3429 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
3431 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3432 if (SextVal >= -31 && SextVal <= 0) {
3433 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
3434 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
3435 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
3436 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
3443 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3444 /// the specified operations to build the shuffle.
3445 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3446 SDValue RHS, SelectionDAG &DAG) {
3447 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3448 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3449 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3452 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3464 if (OpNum == OP_COPY) {
3465 if (LHSID == (1*9+2)*9+3) return LHS;
3466 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3470 SDValue OpLHS, OpRHS;
3471 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
3472 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
3474 unsigned ShufIdxs[16];
3476 default: assert(0 && "Unknown i32 permute!");
3478 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3479 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3480 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3481 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3484 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3485 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3486 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3487 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3490 for (unsigned i = 0; i != 16; ++i)
3491 ShufIdxs[i] = (i&3)+0;
3494 for (unsigned i = 0; i != 16; ++i)
3495 ShufIdxs[i] = (i&3)+4;
3498 for (unsigned i = 0; i != 16; ++i)
3499 ShufIdxs[i] = (i&3)+8;
3502 for (unsigned i = 0; i != 16; ++i)
3503 ShufIdxs[i] = (i&3)+12;
3506 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
3508 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
3510 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
3513 for (unsigned i = 0; i != 16; ++i)
3514 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i8);
3516 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
3517 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3520 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3521 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3522 /// return the code it can be lowered into. Worst case, it can always be
3523 /// lowered into a vperm.
3524 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
3525 SelectionDAG &DAG) {
3526 SDValue V1 = Op.getOperand(0);
3527 SDValue V2 = Op.getOperand(1);
3528 SDValue PermMask = Op.getOperand(2);
3530 // Cases that are handled by instructions that take permute immediates
3531 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3532 // selected by the instruction selector.
3533 if (V2.getOpcode() == ISD::UNDEF) {
3534 if (PPC::isSplatShuffleMask(PermMask.getNode(), 1) ||
3535 PPC::isSplatShuffleMask(PermMask.getNode(), 2) ||
3536 PPC::isSplatShuffleMask(PermMask.getNode(), 4) ||
3537 PPC::isVPKUWUMShuffleMask(PermMask.getNode(), true) ||
3538 PPC::isVPKUHUMShuffleMask(PermMask.getNode(), true) ||
3539 PPC::isVSLDOIShuffleMask(PermMask.getNode(), true) != -1 ||
3540 PPC::isVMRGLShuffleMask(PermMask.getNode(), 1, true) ||
3541 PPC::isVMRGLShuffleMask(PermMask.getNode(), 2, true) ||
3542 PPC::isVMRGLShuffleMask(PermMask.getNode(), 4, true) ||
3543 PPC::isVMRGHShuffleMask(PermMask.getNode(), 1, true) ||
3544 PPC::isVMRGHShuffleMask(PermMask.getNode(), 2, true) ||
3545 PPC::isVMRGHShuffleMask(PermMask.getNode(), 4, true)) {
3550 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3551 // and produce a fixed permutation. If any of these match, do not lower to
3553 if (PPC::isVPKUWUMShuffleMask(PermMask.getNode(), false) ||
3554 PPC::isVPKUHUMShuffleMask(PermMask.getNode(), false) ||
3555 PPC::isVSLDOIShuffleMask(PermMask.getNode(), false) != -1 ||
3556 PPC::isVMRGLShuffleMask(PermMask.getNode(), 1, false) ||
3557 PPC::isVMRGLShuffleMask(PermMask.getNode(), 2, false) ||
3558 PPC::isVMRGLShuffleMask(PermMask.getNode(), 4, false) ||
3559 PPC::isVMRGHShuffleMask(PermMask.getNode(), 1, false) ||
3560 PPC::isVMRGHShuffleMask(PermMask.getNode(), 2, false) ||
3561 PPC::isVMRGHShuffleMask(PermMask.getNode(), 4, false))
3564 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3565 // perfect shuffle table to emit an optimal matching sequence.
3566 unsigned PFIndexes[4];
3567 bool isFourElementShuffle = true;
3568 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3569 unsigned EltNo = 8; // Start out undef.
3570 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3571 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
3572 continue; // Undef, ignore it.
3574 unsigned ByteSource =
3575 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getZExtValue();
3576 if ((ByteSource & 3) != j) {
3577 isFourElementShuffle = false;
3582 EltNo = ByteSource/4;
3583 } else if (EltNo != ByteSource/4) {
3584 isFourElementShuffle = false;
3588 PFIndexes[i] = EltNo;
3591 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3592 // perfect shuffle vector to determine if it is cost effective to do this as
3593 // discrete instructions, or whether we should use a vperm.
3594 if (isFourElementShuffle) {
3595 // Compute the index in the perfect shuffle table.
3596 unsigned PFTableIndex =
3597 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3599 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3600 unsigned Cost = (PFEntry >> 30);
3602 // Determining when to avoid vperm is tricky. Many things affect the cost
3603 // of vperm, particularly how many times the perm mask needs to be computed.
3604 // For example, if the perm mask can be hoisted out of a loop or is already
3605 // used (perhaps because there are multiple permutes with the same shuffle
3606 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3607 // the loop requires an extra register.
3609 // As a compromise, we only emit discrete instructions if the shuffle can be
3610 // generated in 3 or fewer operations. When we have loop information
3611 // available, if this block is within a loop, we should avoid using vperm
3612 // for 3-operation perms and use a constant pool load instead.
3614 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3617 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3618 // vector that will get spilled to the constant pool.
3619 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3621 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3622 // that it is in input element units, not in bytes. Convert now.
3623 MVT EltVT = V1.getValueType().getVectorElementType();
3624 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
3626 SmallVector<SDValue, 16> ResultMask;
3627 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3629 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3632 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getZExtValue();
3634 for (unsigned j = 0; j != BytesPerElement; ++j)
3635 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3639 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3640 &ResultMask[0], ResultMask.size());
3641 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3644 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3645 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3646 /// information about the intrinsic.
3647 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
3649 unsigned IntrinsicID =
3650 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
3653 switch (IntrinsicID) {
3654 default: return false;
3655 // Comparison predicates.
3656 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3657 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3658 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3659 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3660 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3661 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3662 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3663 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3664 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3665 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3666 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3667 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3668 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3670 // Normal Comparisons.
3671 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3672 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3673 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3674 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3675 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3676 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3677 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3678 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3679 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3680 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3681 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3682 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3683 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3688 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3689 /// lower, do it, otherwise return null.
3690 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
3691 SelectionDAG &DAG) {
3692 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3693 // opcode number of the comparison.
3696 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3697 return SDValue(); // Don't custom lower most intrinsics.
3699 // If this is a non-dot comparison, make the VCMP node and we are done.
3701 SDValue Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3702 Op.getOperand(1), Op.getOperand(2),
3703 DAG.getConstant(CompareOpc, MVT::i32));
3704 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3707 // Create the PPCISD altivec 'dot' comparison node.
3709 Op.getOperand(2), // LHS
3710 Op.getOperand(3), // RHS
3711 DAG.getConstant(CompareOpc, MVT::i32)
3713 std::vector<MVT> VTs;
3714 VTs.push_back(Op.getOperand(2).getValueType());
3715 VTs.push_back(MVT::Flag);
3716 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3718 // Now that we have the comparison, emit a copy from the CR to a GPR.
3719 // This is flagged to the above dot comparison.
3720 SDValue Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3721 DAG.getRegister(PPC::CR6, MVT::i32),
3722 CompNode.getValue(1));
3724 // Unpack the result based on how the target uses it.
3725 unsigned BitNo; // Bit # of CR6.
3726 bool InvertBit; // Invert result?
3727 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
3728 default: // Can't happen, don't crash on invalid number though.
3729 case 0: // Return the value of the EQ bit of CR6.
3730 BitNo = 0; InvertBit = false;
3732 case 1: // Return the inverted value of the EQ bit of CR6.
3733 BitNo = 0; InvertBit = true;
3735 case 2: // Return the value of the LT bit of CR6.
3736 BitNo = 2; InvertBit = false;
3738 case 3: // Return the inverted value of the LT bit of CR6.
3739 BitNo = 2; InvertBit = true;
3743 // Shift the bit into the low position.
3744 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3745 DAG.getConstant(8-(3-BitNo), MVT::i32));
3747 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3748 DAG.getConstant(1, MVT::i32));
3750 // If we are supposed to, toggle the bit.
3752 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3753 DAG.getConstant(1, MVT::i32));
3757 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
3758 SelectionDAG &DAG) {
3759 // Create a stack slot that is 16-byte aligned.
3760 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3761 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3762 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3763 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3765 // Store the input value into Value#0 of the stack slot.
3766 SDValue Store = DAG.getStore(DAG.getEntryNode(),
3767 Op.getOperand(0), FIdx, NULL, 0);
3769 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3772 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
3773 if (Op.getValueType() == MVT::v4i32) {
3774 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3776 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3777 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3779 SDValue RHSSwap = // = vrlw RHS, 16
3780 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3782 // Shrinkify inputs to v8i16.
3783 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3784 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3785 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3787 // Low parts multiplied together, generating 32-bit results (we ignore the
3789 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3790 LHS, RHS, DAG, MVT::v4i32);
3792 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3793 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3794 // Shift the high parts up 16 bits.
3795 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3796 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3797 } else if (Op.getValueType() == MVT::v8i16) {
3798 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3800 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3802 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3803 LHS, RHS, Zero, DAG);
3804 } else if (Op.getValueType() == MVT::v16i8) {
3805 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3807 // Multiply the even 8-bit parts, producing 16-bit sums.
3808 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3809 LHS, RHS, DAG, MVT::v8i16);
3810 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3812 // Multiply the odd 8-bit parts, producing 16-bit sums.
3813 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3814 LHS, RHS, DAG, MVT::v8i16);
3815 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3817 // Merge the results together.
3819 for (unsigned i = 0; i != 8; ++i) {
3820 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3821 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3823 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3824 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3826 assert(0 && "Unknown mul to lower!");
3831 /// LowerOperation - Provide custom lowering hooks for some operations.
3833 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3834 switch (Op.getOpcode()) {
3835 default: assert(0 && "Wasn't expecting to be able to lower this!");
3836 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3837 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3838 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3839 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3840 case ISD::SETCC: return LowerSETCC(Op, DAG);
3841 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
3843 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3844 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3847 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3848 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3850 case ISD::FORMAL_ARGUMENTS:
3851 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3852 VarArgsStackOffset, VarArgsNumGPR,
3853 VarArgsNumFPR, PPCSubTarget);
3855 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3856 getTargetMachine());
3857 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3858 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3859 case ISD::DYNAMIC_STACKALLOC:
3860 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3862 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3863 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3864 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3865 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3866 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3868 // Lower 64-bit shifts.
3869 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3870 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3871 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3873 // Vector-related lowering.
3874 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3875 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3876 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3877 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3878 case ISD::MUL: return LowerMUL(Op, DAG);
3880 // Frame & Return address.
3881 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3882 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3887 SDNode *PPCTargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
3888 switch (N->getOpcode()) {
3889 default: assert(0 && "Wasn't expecting to be able to lower this!");
3890 case ISD::FP_TO_SINT: {
3891 SDValue Res = LowerFP_TO_SINT(SDValue(N, 0), DAG);
3892 // Use MERGE_VALUES to drop the chain result value and get a node with one
3893 // result. This requires turning off getMergeValues simplification, since
3894 // otherwise it will give us Res back.
3895 return DAG.getMergeValues(&Res, 1, false).getNode();
3901 //===----------------------------------------------------------------------===//
3902 // Other Lowering Code
3903 //===----------------------------------------------------------------------===//
3906 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3907 bool is64bit, unsigned BinOpcode) {
3908 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3909 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3911 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3912 MachineFunction *F = BB->getParent();
3913 MachineFunction::iterator It = BB;
3916 unsigned dest = MI->getOperand(0).getReg();
3917 unsigned ptrA = MI->getOperand(1).getReg();
3918 unsigned ptrB = MI->getOperand(2).getReg();
3919 unsigned incr = MI->getOperand(3).getReg();
3921 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3922 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3923 F->insert(It, loopMBB);
3924 F->insert(It, exitMBB);
3925 exitMBB->transferSuccessors(BB);
3927 MachineRegisterInfo &RegInfo = F->getRegInfo();
3928 unsigned TmpReg = (!BinOpcode) ? incr :
3929 RegInfo.createVirtualRegister(
3930 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3931 (const TargetRegisterClass *) &PPC::GPRCRegClass);
3935 // fallthrough --> loopMBB
3936 BB->addSuccessor(loopMBB);
3939 // l[wd]arx dest, ptr
3940 // add r0, dest, incr
3941 // st[wd]cx. r0, ptr
3943 // fallthrough --> exitMBB
3945 BuildMI(BB, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
3946 .addReg(ptrA).addReg(ptrB);
3948 BuildMI(BB, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
3949 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
3950 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
3951 BuildMI(BB, TII->get(PPC::BCC))
3952 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
3953 BB->addSuccessor(loopMBB);
3954 BB->addSuccessor(exitMBB);
3963 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
3964 MachineBasicBlock *BB,
3965 bool is8bit, // operation
3966 unsigned BinOpcode) {
3967 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3968 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3969 // In 64 bit mode we have to use 64 bits for addresses, even though the
3970 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
3971 // registers without caring whether they're 32 or 64, but here we're
3972 // doing actual arithmetic on the addresses.
3973 bool is64bit = PPCSubTarget.isPPC64();
3975 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3976 MachineFunction *F = BB->getParent();
3977 MachineFunction::iterator It = BB;
3980 unsigned dest = MI->getOperand(0).getReg();
3981 unsigned ptrA = MI->getOperand(1).getReg();
3982 unsigned ptrB = MI->getOperand(2).getReg();
3983 unsigned incr = MI->getOperand(3).getReg();
3985 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3986 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3987 F->insert(It, loopMBB);
3988 F->insert(It, exitMBB);
3989 exitMBB->transferSuccessors(BB);
3991 MachineRegisterInfo &RegInfo = F->getRegInfo();
3992 const TargetRegisterClass *RC =
3993 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3994 (const TargetRegisterClass *) &PPC::GPRCRegClass;
3995 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
3996 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
3997 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
3998 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
3999 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4000 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4001 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4002 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4003 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4004 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4005 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4007 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4011 // fallthrough --> loopMBB
4012 BB->addSuccessor(loopMBB);
4014 // The 4-byte load must be aligned, while a char or short may be
4015 // anywhere in the word. Hence all this nasty bookkeeping code.
4016 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4017 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4018 // xori shift, shift1, 24 [16]
4019 // rlwinm ptr, ptr1, 0, 0, 29
4020 // slw incr2, incr, shift
4021 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4022 // slw mask, mask2, shift
4024 // lwarx tmpDest, ptr
4025 // add tmp, tmpDest, incr2
4026 // andc tmp2, tmpDest, mask
4027 // and tmp3, tmp, mask
4028 // or tmp4, tmp3, tmp2
4031 // fallthrough --> exitMBB
4032 // srw dest, tmpDest, shift
4034 if (ptrA!=PPC::R0) {
4035 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4036 BuildMI(BB, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4037 .addReg(ptrA).addReg(ptrB);
4041 BuildMI(BB, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4042 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4043 BuildMI(BB, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4044 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4046 BuildMI(BB, TII->get(PPC::RLDICR), PtrReg)
4047 .addReg(Ptr1Reg).addImm(0).addImm(61);
4049 BuildMI(BB, TII->get(PPC::RLWINM), PtrReg)
4050 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4051 BuildMI(BB, TII->get(PPC::SLW), Incr2Reg)
4052 .addReg(incr).addReg(ShiftReg);
4054 BuildMI(BB, TII->get(PPC::LI), Mask2Reg).addImm(255);
4056 BuildMI(BB, TII->get(PPC::LI), Mask3Reg).addImm(0);
4057 BuildMI(BB, TII->get(PPC::ORI), Mask2Reg).addReg(Mask3Reg).addImm(65535);
4059 BuildMI(BB, TII->get(PPC::SLW), MaskReg)
4060 .addReg(Mask2Reg).addReg(ShiftReg);
4063 BuildMI(BB, TII->get(PPC::LWARX), TmpDestReg)
4064 .addReg(PPC::R0).addReg(PtrReg);
4066 BuildMI(BB, TII->get(BinOpcode), TmpReg)
4067 .addReg(Incr2Reg).addReg(TmpDestReg);
4068 BuildMI(BB, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4069 .addReg(TmpDestReg).addReg(MaskReg);
4070 BuildMI(BB, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4071 .addReg(TmpReg).addReg(MaskReg);
4072 BuildMI(BB, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4073 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4074 BuildMI(BB, TII->get(PPC::STWCX))
4075 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4076 BuildMI(BB, TII->get(PPC::BCC))
4077 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4078 BB->addSuccessor(loopMBB);
4079 BB->addSuccessor(exitMBB);
4084 BuildMI(BB, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4089 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4090 MachineBasicBlock *BB) {
4091 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4093 // To "insert" these instructions we actually have to insert their
4094 // control-flow patterns.
4095 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4096 MachineFunction::iterator It = BB;
4099 MachineFunction *F = BB->getParent();
4101 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4102 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4103 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4104 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4105 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4107 // The incoming instruction knows the destination vreg to set, the
4108 // condition code register to branch on, the true/false values to
4109 // select between, and a branch opcode to use.
4114 // cmpTY ccX, r1, r2
4116 // fallthrough --> copy0MBB
4117 MachineBasicBlock *thisMBB = BB;
4118 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4119 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4120 unsigned SelectPred = MI->getOperand(4).getImm();
4121 BuildMI(BB, TII->get(PPC::BCC))
4122 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4123 F->insert(It, copy0MBB);
4124 F->insert(It, sinkMBB);
4125 // Update machine-CFG edges by transferring all successors of the current
4126 // block to the new block which will contain the Phi node for the select.
4127 sinkMBB->transferSuccessors(BB);
4128 // Next, add the true and fallthrough blocks as its successors.
4129 BB->addSuccessor(copy0MBB);
4130 BB->addSuccessor(sinkMBB);
4133 // %FalseValue = ...
4134 // # fallthrough to sinkMBB
4137 // Update machine-CFG edges
4138 BB->addSuccessor(sinkMBB);
4141 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4144 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4145 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4146 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4148 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4149 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4150 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4151 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4152 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4153 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4154 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4155 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4157 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4158 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4159 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4160 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4161 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4162 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4163 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4164 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4166 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4167 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4168 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4169 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4170 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4171 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4172 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4173 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4175 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4176 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4177 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4178 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4179 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4180 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4181 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4182 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4184 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4185 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4186 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4187 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4188 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4189 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4190 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4191 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4193 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4194 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4195 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4196 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4197 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4198 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4199 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4200 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4202 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4203 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4204 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4205 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4206 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4207 BB = EmitAtomicBinary(MI, BB, false, 0);
4208 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4209 BB = EmitAtomicBinary(MI, BB, true, 0);
4211 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4212 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4213 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4215 unsigned dest = MI->getOperand(0).getReg();
4216 unsigned ptrA = MI->getOperand(1).getReg();
4217 unsigned ptrB = MI->getOperand(2).getReg();
4218 unsigned oldval = MI->getOperand(3).getReg();
4219 unsigned newval = MI->getOperand(4).getReg();
4221 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4222 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4223 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4224 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4225 F->insert(It, loop1MBB);
4226 F->insert(It, loop2MBB);
4227 F->insert(It, midMBB);
4228 F->insert(It, exitMBB);
4229 exitMBB->transferSuccessors(BB);
4233 // fallthrough --> loopMBB
4234 BB->addSuccessor(loop1MBB);
4237 // l[wd]arx dest, ptr
4238 // cmp[wd] dest, oldval
4241 // st[wd]cx. newval, ptr
4245 // st[wd]cx. dest, ptr
4248 BuildMI(BB, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4249 .addReg(ptrA).addReg(ptrB);
4250 BuildMI(BB, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4251 .addReg(oldval).addReg(dest);
4252 BuildMI(BB, TII->get(PPC::BCC))
4253 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4254 BB->addSuccessor(loop2MBB);
4255 BB->addSuccessor(midMBB);
4258 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4259 .addReg(newval).addReg(ptrA).addReg(ptrB);
4260 BuildMI(BB, TII->get(PPC::BCC))
4261 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4262 BuildMI(BB, TII->get(PPC::B)).addMBB(exitMBB);
4263 BB->addSuccessor(loop1MBB);
4264 BB->addSuccessor(exitMBB);
4267 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4268 .addReg(dest).addReg(ptrA).addReg(ptrB);
4269 BB->addSuccessor(exitMBB);
4274 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4275 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4276 // We must use 64-bit registers for addresses when targeting 64-bit,
4277 // since we're actually doing arithmetic on them. Other registers
4279 bool is64bit = PPCSubTarget.isPPC64();
4280 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4282 unsigned dest = MI->getOperand(0).getReg();
4283 unsigned ptrA = MI->getOperand(1).getReg();
4284 unsigned ptrB = MI->getOperand(2).getReg();
4285 unsigned oldval = MI->getOperand(3).getReg();
4286 unsigned newval = MI->getOperand(4).getReg();
4288 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4289 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4290 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4291 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4292 F->insert(It, loop1MBB);
4293 F->insert(It, loop2MBB);
4294 F->insert(It, midMBB);
4295 F->insert(It, exitMBB);
4296 exitMBB->transferSuccessors(BB);
4298 MachineRegisterInfo &RegInfo = F->getRegInfo();
4299 const TargetRegisterClass *RC =
4300 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4301 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4302 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4303 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4304 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4305 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4306 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4307 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4308 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4309 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4310 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4311 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4312 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4313 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4314 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4316 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4319 // fallthrough --> loopMBB
4320 BB->addSuccessor(loop1MBB);
4322 // The 4-byte load must be aligned, while a char or short may be
4323 // anywhere in the word. Hence all this nasty bookkeeping code.
4324 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4325 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4326 // xori shift, shift1, 24 [16]
4327 // rlwinm ptr, ptr1, 0, 0, 29
4328 // slw newval2, newval, shift
4329 // slw oldval2, oldval,shift
4330 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4331 // slw mask, mask2, shift
4332 // and newval3, newval2, mask
4333 // and oldval3, oldval2, mask
4335 // lwarx tmpDest, ptr
4336 // and tmp, tmpDest, mask
4337 // cmpw tmp, oldval3
4340 // andc tmp2, tmpDest, mask
4341 // or tmp4, tmp2, newval3
4346 // stwcx. tmpDest, ptr
4348 // srw dest, tmpDest, shift
4349 if (ptrA!=PPC::R0) {
4350 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4351 BuildMI(BB, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4352 .addReg(ptrA).addReg(ptrB);
4356 BuildMI(BB, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4357 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4358 BuildMI(BB, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4359 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4361 BuildMI(BB, TII->get(PPC::RLDICR), PtrReg)
4362 .addReg(Ptr1Reg).addImm(0).addImm(61);
4364 BuildMI(BB, TII->get(PPC::RLWINM), PtrReg)
4365 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4366 BuildMI(BB, TII->get(PPC::SLW), NewVal2Reg)
4367 .addReg(newval).addReg(ShiftReg);
4368 BuildMI(BB, TII->get(PPC::SLW), OldVal2Reg)
4369 .addReg(oldval).addReg(ShiftReg);
4371 BuildMI(BB, TII->get(PPC::LI), Mask2Reg).addImm(255);
4373 BuildMI(BB, TII->get(PPC::LI), Mask3Reg).addImm(0);
4374 BuildMI(BB, TII->get(PPC::ORI), Mask2Reg).addReg(Mask3Reg).addImm(65535);
4376 BuildMI(BB, TII->get(PPC::SLW), MaskReg)
4377 .addReg(Mask2Reg).addReg(ShiftReg);
4378 BuildMI(BB, TII->get(PPC::AND), NewVal3Reg)
4379 .addReg(NewVal2Reg).addReg(MaskReg);
4380 BuildMI(BB, TII->get(PPC::AND), OldVal3Reg)
4381 .addReg(OldVal2Reg).addReg(MaskReg);
4384 BuildMI(BB, TII->get(PPC::LWARX), TmpDestReg)
4385 .addReg(PPC::R0).addReg(PtrReg);
4386 BuildMI(BB, TII->get(PPC::AND),TmpReg).addReg(TmpDestReg).addReg(MaskReg);
4387 BuildMI(BB, TII->get(PPC::CMPW), PPC::CR0)
4388 .addReg(TmpReg).addReg(OldVal3Reg);
4389 BuildMI(BB, TII->get(PPC::BCC))
4390 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4391 BB->addSuccessor(loop2MBB);
4392 BB->addSuccessor(midMBB);
4395 BuildMI(BB, TII->get(PPC::ANDC),Tmp2Reg).addReg(TmpDestReg).addReg(MaskReg);
4396 BuildMI(BB, TII->get(PPC::OR),Tmp4Reg).addReg(Tmp2Reg).addReg(NewVal3Reg);
4397 BuildMI(BB, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4398 .addReg(PPC::R0).addReg(PtrReg);
4399 BuildMI(BB, TII->get(PPC::BCC))
4400 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4401 BuildMI(BB, TII->get(PPC::B)).addMBB(exitMBB);
4402 BB->addSuccessor(loop1MBB);
4403 BB->addSuccessor(exitMBB);
4406 BuildMI(BB, TII->get(PPC::STWCX)).addReg(TmpDestReg)
4407 .addReg(PPC::R0).addReg(PtrReg);
4408 BB->addSuccessor(exitMBB);
4413 BuildMI(BB, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
4415 assert(0 && "Unexpected instr type to insert");
4418 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
4422 //===----------------------------------------------------------------------===//
4423 // Target Optimization Hooks
4424 //===----------------------------------------------------------------------===//
4426 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
4427 DAGCombinerInfo &DCI) const {
4428 TargetMachine &TM = getTargetMachine();
4429 SelectionDAG &DAG = DCI.DAG;
4430 switch (N->getOpcode()) {
4433 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4434 if (C->getZExtValue() == 0) // 0 << V -> 0.
4435 return N->getOperand(0);
4439 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4440 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
4441 return N->getOperand(0);
4445 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4446 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
4447 C->isAllOnesValue()) // -1 >>s V -> -1.
4448 return N->getOperand(0);
4452 case ISD::SINT_TO_FP:
4453 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
4454 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
4455 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
4456 // We allow the src/dst to be either f32/f64, but the intermediate
4457 // type must be i64.
4458 if (N->getOperand(0).getValueType() == MVT::i64 &&
4459 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
4460 SDValue Val = N->getOperand(0).getOperand(0);
4461 if (Val.getValueType() == MVT::f32) {
4462 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
4463 DCI.AddToWorklist(Val.getNode());
4466 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
4467 DCI.AddToWorklist(Val.getNode());
4468 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
4469 DCI.AddToWorklist(Val.getNode());
4470 if (N->getValueType(0) == MVT::f32) {
4471 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
4472 DAG.getIntPtrConstant(0));
4473 DCI.AddToWorklist(Val.getNode());
4476 } else if (N->getOperand(0).getValueType() == MVT::i32) {
4477 // If the intermediate type is i32, we can avoid the load/store here
4484 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
4485 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
4486 !cast<StoreSDNode>(N)->isTruncatingStore() &&
4487 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
4488 N->getOperand(1).getValueType() == MVT::i32 &&
4489 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
4490 SDValue Val = N->getOperand(1).getOperand(0);
4491 if (Val.getValueType() == MVT::f32) {
4492 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
4493 DCI.AddToWorklist(Val.getNode());
4495 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
4496 DCI.AddToWorklist(Val.getNode());
4498 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
4499 N->getOperand(2), N->getOperand(3));
4500 DCI.AddToWorklist(Val.getNode());
4504 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
4505 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
4506 N->getOperand(1).getNode()->hasOneUse() &&
4507 (N->getOperand(1).getValueType() == MVT::i32 ||
4508 N->getOperand(1).getValueType() == MVT::i16)) {
4509 SDValue BSwapOp = N->getOperand(1).getOperand(0);
4510 // Do an any-extend to 32-bits if this is a half-word input.
4511 if (BSwapOp.getValueType() == MVT::i16)
4512 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
4514 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
4515 N->getOperand(2), N->getOperand(3),
4516 DAG.getValueType(N->getOperand(1).getValueType()));
4520 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
4521 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
4522 N->getOperand(0).hasOneUse() &&
4523 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
4524 SDValue Load = N->getOperand(0);
4525 LoadSDNode *LD = cast<LoadSDNode>(Load);
4526 // Create the byte-swapping load.
4527 std::vector<MVT> VTs;
4528 VTs.push_back(MVT::i32);
4529 VTs.push_back(MVT::Other);
4530 SDValue MO = DAG.getMemOperand(LD->getMemOperand());
4532 LD->getChain(), // Chain
4533 LD->getBasePtr(), // Ptr
4535 DAG.getValueType(N->getValueType(0)) // VT
4537 SDValue BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
4539 // If this is an i16 load, insert the truncate.
4540 SDValue ResVal = BSLoad;
4541 if (N->getValueType(0) == MVT::i16)
4542 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
4544 // First, combine the bswap away. This makes the value produced by the
4546 DCI.CombineTo(N, ResVal);
4548 // Next, combine the load away, we give it a bogus result value but a real
4549 // chain result. The result value is dead because the bswap is dead.
4550 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
4552 // Return N so it doesn't get rechecked!
4553 return SDValue(N, 0);
4557 case PPCISD::VCMP: {
4558 // If a VCMPo node already exists with exactly the same operands as this
4559 // node, use its result instead of this node (VCMPo computes both a CR6 and
4560 // a normal output).
4562 if (!N->getOperand(0).hasOneUse() &&
4563 !N->getOperand(1).hasOneUse() &&
4564 !N->getOperand(2).hasOneUse()) {
4566 // Scan all of the users of the LHS, looking for VCMPo's that match.
4567 SDNode *VCMPoNode = 0;
4569 SDNode *LHSN = N->getOperand(0).getNode();
4570 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
4572 if (UI->getOpcode() == PPCISD::VCMPo &&
4573 UI->getOperand(1) == N->getOperand(1) &&
4574 UI->getOperand(2) == N->getOperand(2) &&
4575 UI->getOperand(0) == N->getOperand(0)) {
4580 // If there is no VCMPo node, or if the flag value has a single use, don't
4582 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
4585 // Look at the (necessarily single) use of the flag value. If it has a
4586 // chain, this transformation is more complex. Note that multiple things
4587 // could use the value result, which we should ignore.
4588 SDNode *FlagUser = 0;
4589 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
4590 FlagUser == 0; ++UI) {
4591 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
4593 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
4594 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
4601 // If the user is a MFCR instruction, we know this is safe. Otherwise we
4602 // give up for right now.
4603 if (FlagUser->getOpcode() == PPCISD::MFCR)
4604 return SDValue(VCMPoNode, 0);
4609 // If this is a branch on an altivec predicate comparison, lower this so
4610 // that we don't have to do a MFCR: instead, branch directly on CR6. This
4611 // lowering is done pre-legalize, because the legalizer lowers the predicate
4612 // compare down to code that is difficult to reassemble.
4613 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
4614 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
4618 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
4619 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
4620 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
4621 assert(isDot && "Can't compare against a vector result!");
4623 // If this is a comparison against something other than 0/1, then we know
4624 // that the condition is never/always true.
4625 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
4626 if (Val != 0 && Val != 1) {
4627 if (CC == ISD::SETEQ) // Cond never true, remove branch.
4628 return N->getOperand(0);
4629 // Always !=, turn it into an unconditional branch.
4630 return DAG.getNode(ISD::BR, MVT::Other,
4631 N->getOperand(0), N->getOperand(4));
4634 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
4636 // Create the PPCISD altivec 'dot' comparison node.
4637 std::vector<MVT> VTs;
4639 LHS.getOperand(2), // LHS of compare
4640 LHS.getOperand(3), // RHS of compare
4641 DAG.getConstant(CompareOpc, MVT::i32)
4643 VTs.push_back(LHS.getOperand(2).getValueType());
4644 VTs.push_back(MVT::Flag);
4645 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
4647 // Unpack the result based on how the target uses it.
4648 PPC::Predicate CompOpc;
4649 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
4650 default: // Can't happen, don't crash on invalid number though.
4651 case 0: // Branch on the value of the EQ bit of CR6.
4652 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
4654 case 1: // Branch on the inverted value of the EQ bit of CR6.
4655 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
4657 case 2: // Branch on the value of the LT bit of CR6.
4658 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
4660 case 3: // Branch on the inverted value of the LT bit of CR6.
4661 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
4665 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
4666 DAG.getConstant(CompOpc, MVT::i32),
4667 DAG.getRegister(PPC::CR6, MVT::i32),
4668 N->getOperand(4), CompNode.getValue(1));
4677 //===----------------------------------------------------------------------===//
4678 // Inline Assembly Support
4679 //===----------------------------------------------------------------------===//
4681 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4685 const SelectionDAG &DAG,
4686 unsigned Depth) const {
4687 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4688 switch (Op.getOpcode()) {
4690 case PPCISD::LBRX: {
4691 // lhbrx is known to have the top bits cleared out.
4692 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
4693 KnownZero = 0xFFFF0000;
4696 case ISD::INTRINSIC_WO_CHAIN: {
4697 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
4699 case Intrinsic::ppc_altivec_vcmpbfp_p:
4700 case Intrinsic::ppc_altivec_vcmpeqfp_p:
4701 case Intrinsic::ppc_altivec_vcmpequb_p:
4702 case Intrinsic::ppc_altivec_vcmpequh_p:
4703 case Intrinsic::ppc_altivec_vcmpequw_p:
4704 case Intrinsic::ppc_altivec_vcmpgefp_p:
4705 case Intrinsic::ppc_altivec_vcmpgtfp_p:
4706 case Intrinsic::ppc_altivec_vcmpgtsb_p:
4707 case Intrinsic::ppc_altivec_vcmpgtsh_p:
4708 case Intrinsic::ppc_altivec_vcmpgtsw_p:
4709 case Intrinsic::ppc_altivec_vcmpgtub_p:
4710 case Intrinsic::ppc_altivec_vcmpgtuh_p:
4711 case Intrinsic::ppc_altivec_vcmpgtuw_p:
4712 KnownZero = ~1U; // All bits but the low one are known to be zero.
4720 /// getConstraintType - Given a constraint, return the type of
4721 /// constraint it is for this target.
4722 PPCTargetLowering::ConstraintType
4723 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
4724 if (Constraint.size() == 1) {
4725 switch (Constraint[0]) {
4732 return C_RegisterClass;
4735 return TargetLowering::getConstraintType(Constraint);
4738 std::pair<unsigned, const TargetRegisterClass*>
4739 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4741 if (Constraint.size() == 1) {
4742 // GCC RS6000 Constraint Letters
4743 switch (Constraint[0]) {
4746 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
4747 return std::make_pair(0U, PPC::G8RCRegisterClass);
4748 return std::make_pair(0U, PPC::GPRCRegisterClass);
4751 return std::make_pair(0U, PPC::F4RCRegisterClass);
4752 else if (VT == MVT::f64)
4753 return std::make_pair(0U, PPC::F8RCRegisterClass);
4756 return std::make_pair(0U, PPC::VRRCRegisterClass);
4758 return std::make_pair(0U, PPC::CRRCRegisterClass);
4762 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4766 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4767 /// vector. If it is invalid, don't add anything to Ops.
4768 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
4769 std::vector<SDValue>&Ops,
4770 SelectionDAG &DAG) const {
4771 SDValue Result(0,0);
4782 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
4783 if (!CST) return; // Must be an immediate to match.
4784 unsigned Value = CST->getZExtValue();
4786 default: assert(0 && "Unknown constraint letter!");
4787 case 'I': // "I" is a signed 16-bit constant.
4788 if ((short)Value == (int)Value)
4789 Result = DAG.getTargetConstant(Value, Op.getValueType());
4791 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
4792 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
4793 if ((short)Value == 0)
4794 Result = DAG.getTargetConstant(Value, Op.getValueType());
4796 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
4797 if ((Value >> 16) == 0)
4798 Result = DAG.getTargetConstant(Value, Op.getValueType());
4800 case 'M': // "M" is a constant that is greater than 31.
4802 Result = DAG.getTargetConstant(Value, Op.getValueType());
4804 case 'N': // "N" is a positive constant that is an exact power of two.
4805 if ((int)Value > 0 && isPowerOf2_32(Value))
4806 Result = DAG.getTargetConstant(Value, Op.getValueType());
4808 case 'O': // "O" is the constant zero.
4810 Result = DAG.getTargetConstant(Value, Op.getValueType());
4812 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
4813 if ((short)-Value == (int)-Value)
4814 Result = DAG.getTargetConstant(Value, Op.getValueType());
4821 if (Result.getNode()) {
4822 Ops.push_back(Result);
4826 // Handle standard constraint letters.
4827 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
4830 // isLegalAddressingMode - Return true if the addressing mode represented
4831 // by AM is legal for this target, for a load/store of the specified type.
4832 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4833 const Type *Ty) const {
4834 // FIXME: PPC does not allow r+i addressing modes for vectors!
4836 // PPC allows a sign-extended 16-bit immediate field.
4837 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
4840 // No global is ever allowed as a base.
4844 // PPC only support r+r,
4846 case 0: // "r+i" or just "i", depending on HasBaseReg.
4849 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
4851 // Otherwise we have r+r or r+i.
4854 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
4856 // Allow 2*r as r+r.
4859 // No other scales are supported.
4866 /// isLegalAddressImmediate - Return true if the integer value can be used
4867 /// as the offset of the target addressing mode for load / store of the
4869 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
4870 // PPC allows a sign-extended 16-bit immediate field.
4871 return (V > -(1 << 16) && V < (1 << 16)-1);
4874 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
4878 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
4879 // Depths > 0 not supported yet!
4880 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4883 MachineFunction &MF = DAG.getMachineFunction();
4884 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4886 // Just load the return address off the stack.
4887 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
4889 // Make sure the function really does not optimize away the store of the RA
4891 FuncInfo->setLRStoreRequired();
4892 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4895 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
4896 // Depths > 0 not supported yet!
4897 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4900 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4901 bool isPPC64 = PtrVT == MVT::i64;
4903 MachineFunction &MF = DAG.getMachineFunction();
4904 MachineFrameInfo *MFI = MF.getFrameInfo();
4905 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4906 && MFI->getStackSize();
4909 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
4912 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,