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::FLOG , MVT::f64, Expand);
103 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
104 setOperationAction(ISD::FLOG10,MVT::f64, Expand);
105 setOperationAction(ISD::FEXP ,MVT::f64, Expand);
106 setOperationAction(ISD::FEXP2 ,MVT::f64, Expand);
107 setOperationAction(ISD::FSIN , MVT::f32, Expand);
108 setOperationAction(ISD::FCOS , MVT::f32, Expand);
109 setOperationAction(ISD::FREM , MVT::f32, Expand);
110 setOperationAction(ISD::FPOW , MVT::f32, Expand);
111 setOperationAction(ISD::FLOG , MVT::f32, Expand);
112 setOperationAction(ISD::FLOG2 ,MVT::f32, Expand);
113 setOperationAction(ISD::FLOG10,MVT::f32, Expand);
114 setOperationAction(ISD::FEXP ,MVT::f32, Expand);
115 setOperationAction(ISD::FEXP2 ,MVT::f32, Expand);
117 setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
119 // If we're enabling GP optimizations, use hardware square root
120 if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
121 setOperationAction(ISD::FSQRT, MVT::f64, Expand);
122 setOperationAction(ISD::FSQRT, MVT::f32, Expand);
125 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
126 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
128 // PowerPC does not have BSWAP, CTPOP or CTTZ
129 setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
130 setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
131 setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
132 setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
133 setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
134 setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
136 // PowerPC does not have ROTR
137 setOperationAction(ISD::ROTR, MVT::i32 , Expand);
138 setOperationAction(ISD::ROTR, MVT::i64 , Expand);
140 // PowerPC does not have Select
141 setOperationAction(ISD::SELECT, MVT::i32, Expand);
142 setOperationAction(ISD::SELECT, MVT::i64, Expand);
143 setOperationAction(ISD::SELECT, MVT::f32, Expand);
144 setOperationAction(ISD::SELECT, MVT::f64, Expand);
146 // PowerPC wants to turn select_cc of FP into fsel when possible.
147 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
148 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
150 // PowerPC wants to optimize integer setcc a bit
151 setOperationAction(ISD::SETCC, MVT::i32, Custom);
153 // PowerPC does not have BRCOND which requires SetCC
154 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
156 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
158 // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
159 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
161 // PowerPC does not have [U|S]INT_TO_FP
162 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
163 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
165 setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
166 setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
167 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
168 setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
170 // We cannot sextinreg(i1). Expand to shifts.
171 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
173 // Support label based line numbers.
174 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
175 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
177 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
178 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
179 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
180 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
183 // We want to legalize GlobalAddress and ConstantPool nodes into the
184 // appropriate instructions to materialize the address.
185 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
186 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
187 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
188 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
189 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
190 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
191 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
192 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
194 // RET must be custom lowered, to meet ABI requirements.
195 setOperationAction(ISD::RET , MVT::Other, Custom);
198 setOperationAction(ISD::TRAP, MVT::Other, Legal);
200 // TRAMPOLINE is custom lowered.
201 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
203 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
204 setOperationAction(ISD::VASTART , MVT::Other, Custom);
206 // VAARG is custom lowered with ELF 32 ABI
207 if (TM.getSubtarget<PPCSubtarget>().isELF32_ABI())
208 setOperationAction(ISD::VAARG, MVT::Other, Custom);
210 setOperationAction(ISD::VAARG, MVT::Other, Expand);
212 // Use the default implementation.
213 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
214 setOperationAction(ISD::VAEND , MVT::Other, Expand);
215 setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
216 setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
217 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
218 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
220 // We want to custom lower some of our intrinsics.
221 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
223 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
224 // They also have instructions for converting between i64 and fp.
225 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
226 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
227 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
228 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
229 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
231 // FIXME: disable this lowered code. This generates 64-bit register values,
232 // and we don't model the fact that the top part is clobbered by calls. We
233 // need to flag these together so that the value isn't live across a call.
234 //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
236 // To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
237 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
239 // PowerPC does not have FP_TO_UINT on 32-bit implementations.
240 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
243 if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
244 // 64-bit PowerPC implementations can support i64 types directly
245 addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
246 // BUILD_PAIR can't be handled natively, and should be expanded to shl/or
247 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
248 // 64-bit PowerPC wants to expand i128 shifts itself.
249 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
250 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
251 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
253 // 32-bit PowerPC wants to expand i64 shifts itself.
254 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
255 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
256 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
259 if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
260 // First set operation action for all vector types to expand. Then we
261 // will selectively turn on ones that can be effectively codegen'd.
262 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
263 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
264 MVT VT = (MVT::SimpleValueType)i;
266 // add/sub are legal for all supported vector VT's.
267 setOperationAction(ISD::ADD , VT, Legal);
268 setOperationAction(ISD::SUB , VT, Legal);
270 // We promote all shuffles to v16i8.
271 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote);
272 AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8);
274 // We promote all non-typed operations to v4i32.
275 setOperationAction(ISD::AND , VT, Promote);
276 AddPromotedToType (ISD::AND , VT, MVT::v4i32);
277 setOperationAction(ISD::OR , VT, Promote);
278 AddPromotedToType (ISD::OR , VT, MVT::v4i32);
279 setOperationAction(ISD::XOR , VT, Promote);
280 AddPromotedToType (ISD::XOR , VT, MVT::v4i32);
281 setOperationAction(ISD::LOAD , VT, Promote);
282 AddPromotedToType (ISD::LOAD , VT, MVT::v4i32);
283 setOperationAction(ISD::SELECT, VT, Promote);
284 AddPromotedToType (ISD::SELECT, VT, MVT::v4i32);
285 setOperationAction(ISD::STORE, VT, Promote);
286 AddPromotedToType (ISD::STORE, VT, MVT::v4i32);
288 // No other operations are legal.
289 setOperationAction(ISD::MUL , VT, Expand);
290 setOperationAction(ISD::SDIV, VT, Expand);
291 setOperationAction(ISD::SREM, VT, Expand);
292 setOperationAction(ISD::UDIV, VT, Expand);
293 setOperationAction(ISD::UREM, VT, Expand);
294 setOperationAction(ISD::FDIV, VT, Expand);
295 setOperationAction(ISD::FNEG, VT, Expand);
296 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
297 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
298 setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
299 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
300 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
301 setOperationAction(ISD::UDIVREM, VT, Expand);
302 setOperationAction(ISD::SDIVREM, VT, Expand);
303 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand);
304 setOperationAction(ISD::FPOW, VT, Expand);
305 setOperationAction(ISD::CTPOP, VT, Expand);
306 setOperationAction(ISD::CTLZ, VT, Expand);
307 setOperationAction(ISD::CTTZ, VT, Expand);
310 // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
311 // with merges, splats, etc.
312 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
314 setOperationAction(ISD::AND , MVT::v4i32, Legal);
315 setOperationAction(ISD::OR , MVT::v4i32, Legal);
316 setOperationAction(ISD::XOR , MVT::v4i32, Legal);
317 setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
318 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
319 setOperationAction(ISD::STORE , MVT::v4i32, Legal);
321 addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
322 addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
323 addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
324 addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
326 setOperationAction(ISD::MUL, MVT::v4f32, Legal);
327 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
328 setOperationAction(ISD::MUL, MVT::v8i16, Custom);
329 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
331 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
332 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
334 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
335 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
336 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
337 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
340 setShiftAmountType(MVT::i32);
341 setSetCCResultContents(ZeroOrOneSetCCResult);
343 if (TM.getSubtarget<PPCSubtarget>().isPPC64()) {
344 setStackPointerRegisterToSaveRestore(PPC::X1);
345 setExceptionPointerRegister(PPC::X3);
346 setExceptionSelectorRegister(PPC::X4);
348 setStackPointerRegisterToSaveRestore(PPC::R1);
349 setExceptionPointerRegister(PPC::R3);
350 setExceptionSelectorRegister(PPC::R4);
353 // We have target-specific dag combine patterns for the following nodes:
354 setTargetDAGCombine(ISD::SINT_TO_FP);
355 setTargetDAGCombine(ISD::STORE);
356 setTargetDAGCombine(ISD::BR_CC);
357 setTargetDAGCombine(ISD::BSWAP);
359 // Darwin long double math library functions have $LDBL128 appended.
360 if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
361 setLibcallName(RTLIB::COS_PPCF128, "cosl$LDBL128");
362 setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
363 setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
364 setLibcallName(RTLIB::SIN_PPCF128, "sinl$LDBL128");
365 setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
366 setLibcallName(RTLIB::LOG_PPCF128, "logl$LDBL128");
367 setLibcallName(RTLIB::LOG2_PPCF128, "log2l$LDBL128");
368 setLibcallName(RTLIB::LOG10_PPCF128, "log10l$LDBL128");
369 setLibcallName(RTLIB::EXP_PPCF128, "expl$LDBL128");
370 setLibcallName(RTLIB::EXP2_PPCF128, "exp2l$LDBL128");
373 computeRegisterProperties();
376 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
377 /// function arguments in the caller parameter area.
378 unsigned PPCTargetLowering::getByValTypeAlignment(const Type *Ty) const {
379 TargetMachine &TM = getTargetMachine();
380 // Darwin passes everything on 4 byte boundary.
381 if (TM.getSubtarget<PPCSubtarget>().isDarwin())
387 const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const {
390 case PPCISD::FSEL: return "PPCISD::FSEL";
391 case PPCISD::FCFID: return "PPCISD::FCFID";
392 case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
393 case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
394 case PPCISD::STFIWX: return "PPCISD::STFIWX";
395 case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
396 case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
397 case PPCISD::VPERM: return "PPCISD::VPERM";
398 case PPCISD::Hi: return "PPCISD::Hi";
399 case PPCISD::Lo: return "PPCISD::Lo";
400 case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
401 case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
402 case PPCISD::SRL: return "PPCISD::SRL";
403 case PPCISD::SRA: return "PPCISD::SRA";
404 case PPCISD::SHL: return "PPCISD::SHL";
405 case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
406 case PPCISD::STD_32: return "PPCISD::STD_32";
407 case PPCISD::CALL_ELF: return "PPCISD::CALL_ELF";
408 case PPCISD::CALL_Macho: return "PPCISD::CALL_Macho";
409 case PPCISD::MTCTR: return "PPCISD::MTCTR";
410 case PPCISD::BCTRL_Macho: return "PPCISD::BCTRL_Macho";
411 case PPCISD::BCTRL_ELF: return "PPCISD::BCTRL_ELF";
412 case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
413 case PPCISD::MFCR: return "PPCISD::MFCR";
414 case PPCISD::VCMP: return "PPCISD::VCMP";
415 case PPCISD::VCMPo: return "PPCISD::VCMPo";
416 case PPCISD::LBRX: return "PPCISD::LBRX";
417 case PPCISD::STBRX: return "PPCISD::STBRX";
418 case PPCISD::LARX: return "PPCISD::LARX";
419 case PPCISD::STCX: return "PPCISD::STCX";
420 case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
421 case PPCISD::MFFS: return "PPCISD::MFFS";
422 case PPCISD::MTFSB0: return "PPCISD::MTFSB0";
423 case PPCISD::MTFSB1: return "PPCISD::MTFSB1";
424 case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ";
425 case PPCISD::MTFSF: return "PPCISD::MTFSF";
426 case PPCISD::TAILCALL: return "PPCISD::TAILCALL";
427 case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN";
432 MVT PPCTargetLowering::getSetCCResultType(const SDValue &) const {
437 //===----------------------------------------------------------------------===//
438 // Node matching predicates, for use by the tblgen matching code.
439 //===----------------------------------------------------------------------===//
441 /// isFloatingPointZero - Return true if this is 0.0 or -0.0.
442 static bool isFloatingPointZero(SDValue Op) {
443 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
444 return CFP->getValueAPF().isZero();
445 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
446 // Maybe this has already been legalized into the constant pool?
447 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
448 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
449 return CFP->getValueAPF().isZero();
454 /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return
455 /// true if Op is undef or if it matches the specified value.
456 static bool isConstantOrUndef(SDValue Op, unsigned Val) {
457 return Op.getOpcode() == ISD::UNDEF ||
458 cast<ConstantSDNode>(Op)->getZExtValue() == Val;
461 /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
462 /// VPKUHUM instruction.
463 bool PPC::isVPKUHUMShuffleMask(SDNode *N, bool isUnary) {
465 for (unsigned i = 0; i != 16; ++i)
466 if (!isConstantOrUndef(N->getOperand(i), i*2+1))
469 for (unsigned i = 0; i != 8; ++i)
470 if (!isConstantOrUndef(N->getOperand(i), i*2+1) ||
471 !isConstantOrUndef(N->getOperand(i+8), i*2+1))
477 /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
478 /// VPKUWUM instruction.
479 bool PPC::isVPKUWUMShuffleMask(SDNode *N, bool isUnary) {
481 for (unsigned i = 0; i != 16; i += 2)
482 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
483 !isConstantOrUndef(N->getOperand(i+1), i*2+3))
486 for (unsigned i = 0; i != 8; i += 2)
487 if (!isConstantOrUndef(N->getOperand(i ), i*2+2) ||
488 !isConstantOrUndef(N->getOperand(i+1), i*2+3) ||
489 !isConstantOrUndef(N->getOperand(i+8), i*2+2) ||
490 !isConstantOrUndef(N->getOperand(i+9), i*2+3))
496 /// isVMerge - Common function, used to match vmrg* shuffles.
498 static bool isVMerge(SDNode *N, unsigned UnitSize,
499 unsigned LHSStart, unsigned RHSStart) {
500 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
501 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
502 assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&
503 "Unsupported merge size!");
505 for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units
506 for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit
507 if (!isConstantOrUndef(N->getOperand(i*UnitSize*2+j),
508 LHSStart+j+i*UnitSize) ||
509 !isConstantOrUndef(N->getOperand(i*UnitSize*2+UnitSize+j),
510 RHSStart+j+i*UnitSize))
516 /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
517 /// a VRGL* instruction with the specified unit size (1,2 or 4 bytes).
518 bool PPC::isVMRGLShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
520 return isVMerge(N, UnitSize, 8, 24);
521 return isVMerge(N, UnitSize, 8, 8);
524 /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
525 /// a VRGH* instruction with the specified unit size (1,2 or 4 bytes).
526 bool PPC::isVMRGHShuffleMask(SDNode *N, unsigned UnitSize, bool isUnary) {
528 return isVMerge(N, UnitSize, 0, 16);
529 return isVMerge(N, UnitSize, 0, 0);
533 /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
534 /// amount, otherwise return -1.
535 int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary) {
536 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
537 N->getNumOperands() == 16 && "PPC only supports shuffles by bytes!");
538 // Find the first non-undef value in the shuffle mask.
540 for (i = 0; i != 16 && N->getOperand(i).getOpcode() == ISD::UNDEF; ++i)
543 if (i == 16) return -1; // all undef.
545 // Otherwise, check to see if the rest of the elements are consequtively
546 // numbered from this value.
547 unsigned ShiftAmt = cast<ConstantSDNode>(N->getOperand(i))->getZExtValue();
548 if (ShiftAmt < i) return -1;
552 // Check the rest of the elements to see if they are consequtive.
553 for (++i; i != 16; ++i)
554 if (!isConstantOrUndef(N->getOperand(i), ShiftAmt+i))
557 // Check the rest of the elements to see if they are consequtive.
558 for (++i; i != 16; ++i)
559 if (!isConstantOrUndef(N->getOperand(i), (ShiftAmt+i) & 15))
566 /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
567 /// specifies a splat of a single element that is suitable for input to
568 /// VSPLTB/VSPLTH/VSPLTW.
569 bool PPC::isSplatShuffleMask(SDNode *N, unsigned EltSize) {
570 assert(N->getOpcode() == ISD::BUILD_VECTOR &&
571 N->getNumOperands() == 16 &&
572 (EltSize == 1 || EltSize == 2 || EltSize == 4));
574 // This is a splat operation if each element of the permute is the same, and
575 // if the value doesn't reference the second vector.
576 unsigned ElementBase = 0;
577 SDValue Elt = N->getOperand(0);
578 if (ConstantSDNode *EltV = dyn_cast<ConstantSDNode>(Elt))
579 ElementBase = EltV->getZExtValue();
581 return false; // FIXME: Handle UNDEF elements too!
583 if (cast<ConstantSDNode>(Elt)->getZExtValue() >= 16)
586 // Check that they are consequtive.
587 for (unsigned i = 1; i != EltSize; ++i) {
588 if (!isa<ConstantSDNode>(N->getOperand(i)) ||
589 cast<ConstantSDNode>(N->getOperand(i))->getZExtValue() != i+ElementBase)
593 assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
594 for (unsigned i = EltSize, e = 16; i != e; i += EltSize) {
595 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
596 assert(isa<ConstantSDNode>(N->getOperand(i)) &&
597 "Invalid VECTOR_SHUFFLE mask!");
598 for (unsigned j = 0; j != EltSize; ++j)
599 if (N->getOperand(i+j) != N->getOperand(j))
606 /// isAllNegativeZeroVector - Returns true if all elements of build_vector
608 bool PPC::isAllNegativeZeroVector(SDNode *N) {
609 assert(N->getOpcode() == ISD::BUILD_VECTOR);
610 if (PPC::isSplatShuffleMask(N, N->getNumOperands()))
611 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N))
612 return CFP->getValueAPF().isNegZero();
616 /// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
617 /// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
618 unsigned PPC::getVSPLTImmediate(SDNode *N, unsigned EltSize) {
619 assert(isSplatShuffleMask(N, EltSize));
620 return cast<ConstantSDNode>(N->getOperand(0))->getZExtValue() / EltSize;
623 /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed
624 /// by using a vspltis[bhw] instruction of the specified element size, return
625 /// the constant being splatted. The ByteSize field indicates the number of
626 /// bytes of each element [124] -> [bhw].
627 SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) {
630 // If ByteSize of the splat is bigger than the element size of the
631 // build_vector, then we have a case where we are checking for a splat where
632 // multiple elements of the buildvector are folded together into a single
633 // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8).
634 unsigned EltSize = 16/N->getNumOperands();
635 if (EltSize < ByteSize) {
636 unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval.
637 SDValue UniquedVals[4];
638 assert(Multiple > 1 && Multiple <= 4 && "How can this happen?");
640 // See if all of the elements in the buildvector agree across.
641 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
642 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
643 // If the element isn't a constant, bail fully out.
644 if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue();
647 if (UniquedVals[i&(Multiple-1)].getNode() == 0)
648 UniquedVals[i&(Multiple-1)] = N->getOperand(i);
649 else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i))
650 return SDValue(); // no match.
653 // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains
654 // either constant or undef values that are identical for each chunk. See
655 // if these chunks can form into a larger vspltis*.
657 // Check to see if all of the leading entries are either 0 or -1. If
658 // neither, then this won't fit into the immediate field.
659 bool LeadingZero = true;
660 bool LeadingOnes = true;
661 for (unsigned i = 0; i != Multiple-1; ++i) {
662 if (UniquedVals[i].getNode() == 0) continue; // Must have been undefs.
664 LeadingZero &= cast<ConstantSDNode>(UniquedVals[i])->isNullValue();
665 LeadingOnes &= cast<ConstantSDNode>(UniquedVals[i])->isAllOnesValue();
667 // Finally, check the least significant entry.
669 if (UniquedVals[Multiple-1].getNode() == 0)
670 return DAG.getTargetConstant(0, MVT::i32); // 0,0,0,undef
671 int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue();
673 return DAG.getTargetConstant(Val, MVT::i32); // 0,0,0,4 -> vspltisw(4)
676 if (UniquedVals[Multiple-1].getNode() == 0)
677 return DAG.getTargetConstant(~0U, MVT::i32); // -1,-1,-1,undef
678 int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSignExtended();
679 if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2)
680 return DAG.getTargetConstant(Val, MVT::i32);
686 // Check to see if this buildvec has a single non-undef value in its elements.
687 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
688 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
689 if (OpVal.getNode() == 0)
690 OpVal = N->getOperand(i);
691 else if (OpVal != N->getOperand(i))
695 if (OpVal.getNode() == 0) return SDValue(); // All UNDEF: use implicit def.
697 unsigned ValSizeInBytes = 0;
699 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
700 Value = CN->getZExtValue();
701 ValSizeInBytes = CN->getValueType(0).getSizeInBits()/8;
702 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
703 assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
704 Value = FloatToBits(CN->getValueAPF().convertToFloat());
708 // If the splat value is larger than the element value, then we can never do
709 // this splat. The only case that we could fit the replicated bits into our
710 // immediate field for would be zero, and we prefer to use vxor for it.
711 if (ValSizeInBytes < ByteSize) return SDValue();
713 // If the element value is larger than the splat value, cut it in half and
714 // check to see if the two halves are equal. Continue doing this until we
715 // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
716 while (ValSizeInBytes > ByteSize) {
717 ValSizeInBytes >>= 1;
719 // If the top half equals the bottom half, we're still ok.
720 if (((Value >> (ValSizeInBytes*8)) & ((1 << (8*ValSizeInBytes))-1)) !=
721 (Value & ((1 << (8*ValSizeInBytes))-1)))
725 // Properly sign extend the value.
726 int ShAmt = (4-ByteSize)*8;
727 int MaskVal = ((int)Value << ShAmt) >> ShAmt;
729 // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
730 if (MaskVal == 0) return SDValue();
732 // Finally, if this value fits in a 5 bit sext field, return it
733 if (((MaskVal << (32-5)) >> (32-5)) == MaskVal)
734 return DAG.getTargetConstant(MaskVal, MVT::i32);
738 //===----------------------------------------------------------------------===//
739 // Addressing Mode Selection
740 //===----------------------------------------------------------------------===//
742 /// isIntS16Immediate - This method tests to see if the node is either a 32-bit
743 /// or 64-bit immediate, and if the value can be accurately represented as a
744 /// sign extension from a 16-bit value. If so, this returns true and the
746 static bool isIntS16Immediate(SDNode *N, short &Imm) {
747 if (N->getOpcode() != ISD::Constant)
750 Imm = (short)cast<ConstantSDNode>(N)->getZExtValue();
751 if (N->getValueType(0) == MVT::i32)
752 return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue();
754 return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue();
756 static bool isIntS16Immediate(SDValue Op, short &Imm) {
757 return isIntS16Immediate(Op.getNode(), Imm);
761 /// SelectAddressRegReg - Given the specified addressed, check to see if it
762 /// can be represented as an indexed [r+r] operation. Returns false if it
763 /// can be more efficiently represented with [r+imm].
764 bool PPCTargetLowering::SelectAddressRegReg(SDValue N, SDValue &Base,
768 if (N.getOpcode() == ISD::ADD) {
769 if (isIntS16Immediate(N.getOperand(1), imm))
771 if (N.getOperand(1).getOpcode() == PPCISD::Lo)
774 Base = N.getOperand(0);
775 Index = N.getOperand(1);
777 } else if (N.getOpcode() == ISD::OR) {
778 if (isIntS16Immediate(N.getOperand(1), imm))
779 return false; // r+i can fold it if we can.
781 // If this is an or of disjoint bitfields, we can codegen this as an add
782 // (for better address arithmetic) if the LHS and RHS of the OR are provably
784 APInt LHSKnownZero, LHSKnownOne;
785 APInt RHSKnownZero, RHSKnownOne;
786 DAG.ComputeMaskedBits(N.getOperand(0),
787 APInt::getAllOnesValue(N.getOperand(0)
788 .getValueSizeInBits()),
789 LHSKnownZero, LHSKnownOne);
791 if (LHSKnownZero.getBoolValue()) {
792 DAG.ComputeMaskedBits(N.getOperand(1),
793 APInt::getAllOnesValue(N.getOperand(1)
794 .getValueSizeInBits()),
795 RHSKnownZero, RHSKnownOne);
796 // If all of the bits are known zero on the LHS or RHS, the add won't
798 if (~(LHSKnownZero | RHSKnownZero) == 0) {
799 Base = N.getOperand(0);
800 Index = N.getOperand(1);
809 /// Returns true if the address N can be represented by a base register plus
810 /// a signed 16-bit displacement [r+imm], and if it is not better
811 /// represented as reg+reg.
812 bool PPCTargetLowering::SelectAddressRegImm(SDValue N, SDValue &Disp,
813 SDValue &Base, SelectionDAG &DAG){
814 // If this can be more profitably realized as r+r, fail.
815 if (SelectAddressRegReg(N, Disp, Base, DAG))
818 if (N.getOpcode() == ISD::ADD) {
820 if (isIntS16Immediate(N.getOperand(1), imm)) {
821 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
822 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
823 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
825 Base = N.getOperand(0);
827 return true; // [r+i]
828 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
829 // Match LOAD (ADD (X, Lo(G))).
830 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
831 && "Cannot handle constant offsets yet!");
832 Disp = N.getOperand(1).getOperand(0); // The global address.
833 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
834 Disp.getOpcode() == ISD::TargetConstantPool ||
835 Disp.getOpcode() == ISD::TargetJumpTable);
836 Base = N.getOperand(0);
837 return true; // [&g+r]
839 } else if (N.getOpcode() == ISD::OR) {
841 if (isIntS16Immediate(N.getOperand(1), imm)) {
842 // If this is an or of disjoint bitfields, we can codegen this as an add
843 // (for better address arithmetic) if the LHS and RHS of the OR are
844 // provably disjoint.
845 APInt LHSKnownZero, LHSKnownOne;
846 DAG.ComputeMaskedBits(N.getOperand(0),
847 APInt::getAllOnesValue(N.getOperand(0)
848 .getValueSizeInBits()),
849 LHSKnownZero, LHSKnownOne);
851 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
852 // If all of the bits are known zero on the LHS or RHS, the add won't
854 Base = N.getOperand(0);
855 Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
859 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
860 // Loading from a constant address.
862 // If this address fits entirely in a 16-bit sext immediate field, codegen
865 if (isIntS16Immediate(CN, Imm)) {
866 Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
867 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
871 // Handle 32-bit sext immediates with LIS + addr mode.
872 if (CN->getValueType(0) == MVT::i32 ||
873 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
874 int Addr = (int)CN->getZExtValue();
876 // Otherwise, break this down into an LIS + disp.
877 Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
879 Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, MVT::i32);
880 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
881 Base = SDValue(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
886 Disp = DAG.getTargetConstant(0, getPointerTy());
887 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
888 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
891 return true; // [r+0]
894 /// SelectAddressRegRegOnly - Given the specified addressed, force it to be
895 /// represented as an indexed [r+r] operation.
896 bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base,
899 // Check to see if we can easily represent this as an [r+r] address. This
900 // will fail if it thinks that the address is more profitably represented as
901 // reg+imm, e.g. where imm = 0.
902 if (SelectAddressRegReg(N, Base, Index, DAG))
905 // If the operand is an addition, always emit this as [r+r], since this is
906 // better (for code size, and execution, as the memop does the add for free)
907 // than emitting an explicit add.
908 if (N.getOpcode() == ISD::ADD) {
909 Base = N.getOperand(0);
910 Index = N.getOperand(1);
914 // Otherwise, do it the hard way, using R0 as the base register.
915 Base = DAG.getRegister(PPC::R0, N.getValueType());
920 /// SelectAddressRegImmShift - Returns true if the address N can be
921 /// represented by a base register plus a signed 14-bit displacement
922 /// [r+imm*4]. Suitable for use by STD and friends.
923 bool PPCTargetLowering::SelectAddressRegImmShift(SDValue N, SDValue &Disp,
926 // If this can be more profitably realized as r+r, fail.
927 if (SelectAddressRegReg(N, Disp, Base, DAG))
930 if (N.getOpcode() == ISD::ADD) {
932 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
933 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
934 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
935 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
937 Base = N.getOperand(0);
939 return true; // [r+i]
940 } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
941 // Match LOAD (ADD (X, Lo(G))).
942 assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()
943 && "Cannot handle constant offsets yet!");
944 Disp = N.getOperand(1).getOperand(0); // The global address.
945 assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
946 Disp.getOpcode() == ISD::TargetConstantPool ||
947 Disp.getOpcode() == ISD::TargetJumpTable);
948 Base = N.getOperand(0);
949 return true; // [&g+r]
951 } else if (N.getOpcode() == ISD::OR) {
953 if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
954 // If this is an or of disjoint bitfields, we can codegen this as an add
955 // (for better address arithmetic) if the LHS and RHS of the OR are
956 // provably disjoint.
957 APInt LHSKnownZero, LHSKnownOne;
958 DAG.ComputeMaskedBits(N.getOperand(0),
959 APInt::getAllOnesValue(N.getOperand(0)
960 .getValueSizeInBits()),
961 LHSKnownZero, LHSKnownOne);
962 if ((LHSKnownZero.getZExtValue()|~(uint64_t)imm) == ~0ULL) {
963 // If all of the bits are known zero on the LHS or RHS, the add won't
965 Base = N.getOperand(0);
966 Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
970 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
971 // Loading from a constant address. Verify low two bits are clear.
972 if ((CN->getZExtValue() & 3) == 0) {
973 // If this address fits entirely in a 14-bit sext immediate field, codegen
976 if (isIntS16Immediate(CN, Imm)) {
977 Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
978 Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
982 // Fold the low-part of 32-bit absolute addresses into addr mode.
983 if (CN->getValueType(0) == MVT::i32 ||
984 (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) {
985 int Addr = (int)CN->getZExtValue();
987 // Otherwise, break this down into an LIS + disp.
988 Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
990 Base = DAG.getTargetConstant((Addr-(signed short)Addr) >> 16, MVT::i32);
991 unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8;
992 Base = SDValue(DAG.getTargetNode(Opc, CN->getValueType(0), Base), 0);
998 Disp = DAG.getTargetConstant(0, getPointerTy());
999 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
1000 Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
1003 return true; // [r+0]
1007 /// getPreIndexedAddressParts - returns true by value, base pointer and
1008 /// offset pointer and addressing mode by reference if the node's address
1009 /// can be legally represented as pre-indexed load / store address.
1010 bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
1012 ISD::MemIndexedMode &AM,
1013 SelectionDAG &DAG) {
1014 // Disabled by default for now.
1015 if (!EnablePPCPreinc) return false;
1019 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1020 Ptr = LD->getBasePtr();
1021 VT = LD->getMemoryVT();
1023 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
1025 Ptr = ST->getBasePtr();
1026 VT = ST->getMemoryVT();
1030 // PowerPC doesn't have preinc load/store instructions for vectors.
1034 // TODO: Check reg+reg first.
1036 // LDU/STU use reg+imm*4, others use reg+imm.
1037 if (VT != MVT::i64) {
1039 if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
1043 if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
1047 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
1048 // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
1049 // sext i32 to i64 when addr mode is r+i.
1050 if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 &&
1051 LD->getExtensionType() == ISD::SEXTLOAD &&
1052 isa<ConstantSDNode>(Offset))
1060 //===----------------------------------------------------------------------===//
1061 // LowerOperation implementation
1062 //===----------------------------------------------------------------------===//
1064 SDValue PPCTargetLowering::LowerConstantPool(SDValue Op,
1065 SelectionDAG &DAG) {
1066 MVT PtrVT = Op.getValueType();
1067 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1068 Constant *C = CP->getConstVal();
1069 SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
1070 SDValue Zero = DAG.getConstant(0, PtrVT);
1072 const TargetMachine &TM = DAG.getTarget();
1074 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, CPI, Zero);
1075 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, CPI, Zero);
1077 // If this is a non-darwin platform, we don't support non-static relo models
1079 if (TM.getRelocationModel() == Reloc::Static ||
1080 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1081 // Generate non-pic code that has direct accesses to the constant pool.
1082 // The address of the global is just (hi(&g)+lo(&g)).
1083 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1086 if (TM.getRelocationModel() == Reloc::PIC_) {
1087 // With PIC, the first instruction is actually "GR+hi(&G)".
1088 Hi = DAG.getNode(ISD::ADD, PtrVT,
1089 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1092 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1096 SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
1097 MVT PtrVT = Op.getValueType();
1098 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1099 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
1100 SDValue Zero = DAG.getConstant(0, PtrVT);
1102 const TargetMachine &TM = DAG.getTarget();
1104 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, JTI, Zero);
1105 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, JTI, Zero);
1107 // If this is a non-darwin platform, we don't support non-static relo models
1109 if (TM.getRelocationModel() == Reloc::Static ||
1110 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1111 // Generate non-pic code that has direct accesses to the constant pool.
1112 // The address of the global is just (hi(&g)+lo(&g)).
1113 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1116 if (TM.getRelocationModel() == Reloc::PIC_) {
1117 // With PIC, the first instruction is actually "GR+hi(&G)".
1118 Hi = DAG.getNode(ISD::ADD, PtrVT,
1119 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1122 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1126 SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1127 SelectionDAG &DAG) {
1128 assert(0 && "TLS not implemented for PPC.");
1129 return SDValue(); // Not reached
1132 SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op,
1133 SelectionDAG &DAG) {
1134 MVT PtrVT = Op.getValueType();
1135 GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
1136 GlobalValue *GV = GSDN->getGlobal();
1137 SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
1138 // If it's a debug information descriptor, don't mess with it.
1139 if (DAG.isVerifiedDebugInfoDesc(Op))
1141 SDValue Zero = DAG.getConstant(0, PtrVT);
1143 const TargetMachine &TM = DAG.getTarget();
1145 SDValue Hi = DAG.getNode(PPCISD::Hi, PtrVT, GA, Zero);
1146 SDValue Lo = DAG.getNode(PPCISD::Lo, PtrVT, GA, Zero);
1148 // If this is a non-darwin platform, we don't support non-static relo models
1150 if (TM.getRelocationModel() == Reloc::Static ||
1151 !TM.getSubtarget<PPCSubtarget>().isDarwin()) {
1152 // Generate non-pic code that has direct accesses to globals.
1153 // The address of the global is just (hi(&g)+lo(&g)).
1154 return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1157 if (TM.getRelocationModel() == Reloc::PIC_) {
1158 // With PIC, the first instruction is actually "GR+hi(&G)".
1159 Hi = DAG.getNode(ISD::ADD, PtrVT,
1160 DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
1163 Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
1165 if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
1168 // If the global is weak or external, we have to go through the lazy
1170 return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
1173 SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
1174 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
1176 // If we're comparing for equality to zero, expose the fact that this is
1177 // implented as a ctlz/srl pair on ppc, so that the dag combiner can
1178 // fold the new nodes.
1179 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
1180 if (C->isNullValue() && CC == ISD::SETEQ) {
1181 MVT VT = Op.getOperand(0).getValueType();
1182 SDValue Zext = Op.getOperand(0);
1183 if (VT.bitsLT(MVT::i32)) {
1185 Zext = DAG.getNode(ISD::ZERO_EXTEND, VT, Op.getOperand(0));
1187 unsigned Log2b = Log2_32(VT.getSizeInBits());
1188 SDValue Clz = DAG.getNode(ISD::CTLZ, VT, Zext);
1189 SDValue Scc = DAG.getNode(ISD::SRL, VT, Clz,
1190 DAG.getConstant(Log2b, MVT::i32));
1191 return DAG.getNode(ISD::TRUNCATE, MVT::i32, Scc);
1193 // Leave comparisons against 0 and -1 alone for now, since they're usually
1194 // optimized. FIXME: revisit this when we can custom lower all setcc
1196 if (C->isAllOnesValue() || C->isNullValue())
1200 // If we have an integer seteq/setne, turn it into a compare against zero
1201 // by xor'ing the rhs with the lhs, which is faster than setting a
1202 // condition register, reading it back out, and masking the correct bit. The
1203 // normal approach here uses sub to do this instead of xor. Using xor exposes
1204 // the result to other bit-twiddling opportunities.
1205 MVT LHSVT = Op.getOperand(0).getValueType();
1206 if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
1207 MVT VT = Op.getValueType();
1208 SDValue Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
1210 return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
1215 SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG,
1216 int VarArgsFrameIndex,
1217 int VarArgsStackOffset,
1218 unsigned VarArgsNumGPR,
1219 unsigned VarArgsNumFPR,
1220 const PPCSubtarget &Subtarget) {
1222 assert(0 && "VAARG in ELF32 ABI not implemented yet!");
1223 return SDValue(); // Not reached
1226 SDValue PPCTargetLowering::LowerTRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
1227 SDValue Chain = Op.getOperand(0);
1228 SDValue Trmp = Op.getOperand(1); // trampoline
1229 SDValue FPtr = Op.getOperand(2); // nested function
1230 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
1232 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1233 bool isPPC64 = (PtrVT == MVT::i64);
1234 const Type *IntPtrTy =
1235 DAG.getTargetLoweringInfo().getTargetData()->getIntPtrType();
1237 TargetLowering::ArgListTy Args;
1238 TargetLowering::ArgListEntry Entry;
1240 Entry.Ty = IntPtrTy;
1241 Entry.Node = Trmp; Args.push_back(Entry);
1243 // TrampSize == (isPPC64 ? 48 : 40);
1244 Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40,
1245 isPPC64 ? MVT::i64 : MVT::i32);
1246 Args.push_back(Entry);
1248 Entry.Node = FPtr; Args.push_back(Entry);
1249 Entry.Node = Nest; Args.push_back(Entry);
1251 // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg)
1252 std::pair<SDValue, SDValue> CallResult =
1253 LowerCallTo(Chain, Op.getValueType().getTypeForMVT(), false, false,
1254 false, CallingConv::C, false,
1255 DAG.getExternalSymbol("__trampoline_setup", PtrVT),
1259 { CallResult.first, CallResult.second };
1261 return DAG.getMergeValues(Ops, 2, false);
1264 SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG,
1265 int VarArgsFrameIndex,
1266 int VarArgsStackOffset,
1267 unsigned VarArgsNumGPR,
1268 unsigned VarArgsNumFPR,
1269 const PPCSubtarget &Subtarget) {
1271 if (Subtarget.isMachoABI()) {
1272 // vastart just stores the address of the VarArgsFrameIndex slot into the
1273 // memory location argument.
1274 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1275 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1276 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1277 return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV, 0);
1280 // For ELF 32 ABI we follow the layout of the va_list struct.
1281 // We suppose the given va_list is already allocated.
1284 // char gpr; /* index into the array of 8 GPRs
1285 // * stored in the register save area
1286 // * gpr=0 corresponds to r3,
1287 // * gpr=1 to r4, etc.
1289 // char fpr; /* index into the array of 8 FPRs
1290 // * stored in the register save area
1291 // * fpr=0 corresponds to f1,
1292 // * fpr=1 to f2, etc.
1294 // char *overflow_arg_area;
1295 // /* location on stack that holds
1296 // * the next overflow argument
1298 // char *reg_save_area;
1299 // /* where r3:r10 and f1:f8 (if saved)
1305 SDValue ArgGPR = DAG.getConstant(VarArgsNumGPR, MVT::i8);
1306 SDValue ArgFPR = DAG.getConstant(VarArgsNumFPR, MVT::i8);
1309 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1311 SDValue StackOffsetFI = DAG.getFrameIndex(VarArgsStackOffset, PtrVT);
1312 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1314 uint64_t FrameOffset = PtrVT.getSizeInBits()/8;
1315 SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, PtrVT);
1317 uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1;
1318 SDValue ConstStackOffset = DAG.getConstant(StackOffset, PtrVT);
1320 uint64_t FPROffset = 1;
1321 SDValue ConstFPROffset = DAG.getConstant(FPROffset, PtrVT);
1323 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1325 // Store first byte : number of int regs
1326 SDValue firstStore = DAG.getStore(Op.getOperand(0), ArgGPR,
1327 Op.getOperand(1), SV, 0);
1328 uint64_t nextOffset = FPROffset;
1329 SDValue nextPtr = DAG.getNode(ISD::ADD, PtrVT, Op.getOperand(1),
1332 // Store second byte : number of float regs
1333 SDValue secondStore =
1334 DAG.getStore(firstStore, ArgFPR, nextPtr, SV, nextOffset);
1335 nextOffset += StackOffset;
1336 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstStackOffset);
1338 // Store second word : arguments given on stack
1339 SDValue thirdStore =
1340 DAG.getStore(secondStore, StackOffsetFI, nextPtr, SV, nextOffset);
1341 nextOffset += FrameOffset;
1342 nextPtr = DAG.getNode(ISD::ADD, PtrVT, nextPtr, ConstFrameOffset);
1344 // Store third word : arguments given in registers
1345 return DAG.getStore(thirdStore, FR, nextPtr, SV, nextOffset);
1349 #include "PPCGenCallingConv.inc"
1351 /// GetFPR - Get the set of FP registers that should be allocated for arguments,
1352 /// depending on which subtarget is selected.
1353 static const unsigned *GetFPR(const PPCSubtarget &Subtarget) {
1354 if (Subtarget.isMachoABI()) {
1355 static const unsigned FPR[] = {
1356 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1357 PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13
1363 static const unsigned FPR[] = {
1364 PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7,
1370 /// CalculateStackSlotSize - Calculates the size reserved for this argument on
1372 static unsigned CalculateStackSlotSize(SDValue Arg, ISD::ArgFlagsTy Flags,
1373 bool isVarArg, unsigned PtrByteSize) {
1374 MVT ArgVT = Arg.getValueType();
1375 unsigned ArgSize =ArgVT.getSizeInBits()/8;
1376 if (Flags.isByVal())
1377 ArgSize = Flags.getByValSize();
1378 ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1384 PPCTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op,
1386 int &VarArgsFrameIndex,
1387 int &VarArgsStackOffset,
1388 unsigned &VarArgsNumGPR,
1389 unsigned &VarArgsNumFPR,
1390 const PPCSubtarget &Subtarget) {
1391 // TODO: add description of PPC stack frame format, or at least some docs.
1393 MachineFunction &MF = DAG.getMachineFunction();
1394 MachineFrameInfo *MFI = MF.getFrameInfo();
1395 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1396 SmallVector<SDValue, 8> ArgValues;
1397 SDValue Root = Op.getOperand(0);
1398 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1400 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1401 bool isPPC64 = PtrVT == MVT::i64;
1402 bool isMachoABI = Subtarget.isMachoABI();
1403 bool isELF32_ABI = Subtarget.isELF32_ABI();
1404 // Potential tail calls could cause overwriting of argument stack slots.
1405 unsigned CC = MF.getFunction()->getCallingConv();
1406 bool isImmutable = !(PerformTailCallOpt && (CC==CallingConv::Fast));
1407 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1409 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1410 // Area that is at least reserved in caller of this function.
1411 unsigned MinReservedArea = ArgOffset;
1413 static const unsigned GPR_32[] = { // 32-bit registers.
1414 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
1415 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
1417 static const unsigned GPR_64[] = { // 64-bit registers.
1418 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
1419 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
1422 static const unsigned *FPR = GetFPR(Subtarget);
1424 static const unsigned VR[] = {
1425 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
1426 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
1429 const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
1430 const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
1431 const unsigned Num_VR_Regs = array_lengthof( VR);
1433 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
1435 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
1437 // In 32-bit non-varargs functions, the stack space for vectors is after the
1438 // stack space for non-vectors. We do not use this space unless we have
1439 // too many vectors to fit in registers, something that only occurs in
1440 // constructed examples:), but we have to walk the arglist to figure
1441 // that out...for the pathological case, compute VecArgOffset as the
1442 // start of the vector parameter area. Computing VecArgOffset is the
1443 // entire point of the following loop.
1444 // Altivec is not mentioned in the ppc32 Elf Supplement, so I'm not trying
1445 // to handle Elf here.
1446 unsigned VecArgOffset = ArgOffset;
1447 if (!isVarArg && !isPPC64) {
1448 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e;
1450 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1451 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1452 ISD::ArgFlagsTy Flags =
1453 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1455 if (Flags.isByVal()) {
1456 // ObjSize is the true size, ArgSize rounded up to multiple of regs.
1457 ObjSize = Flags.getByValSize();
1459 ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1460 VecArgOffset += ArgSize;
1464 switch(ObjectVT.getSimpleVT()) {
1465 default: assert(0 && "Unhandled argument type!");
1468 VecArgOffset += isPPC64 ? 8 : 4;
1470 case MVT::i64: // PPC64
1478 // Nothing to do, we're only looking at Nonvector args here.
1483 // We've found where the vector parameter area in memory is. Skip the
1484 // first 12 parameters; these don't use that memory.
1485 VecArgOffset = ((VecArgOffset+15)/16)*16;
1486 VecArgOffset += 12*16;
1488 // Add DAG nodes to load the arguments or copy them out of registers. On
1489 // entry to a function on PPC, the arguments start after the linkage area,
1490 // although the first ones are often in registers.
1492 // In the ELF 32 ABI, GPRs and stack are double word align: an argument
1493 // represented with two words (long long or double) must be copied to an
1494 // even GPR_idx value or to an even ArgOffset value.
1496 SmallVector<SDValue, 8> MemOps;
1497 unsigned nAltivecParamsAtEnd = 0;
1498 for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues() - 1;
1499 ArgNo != e; ++ArgNo) {
1501 bool needsLoad = false;
1502 MVT ObjectVT = Op.getValue(ArgNo).getValueType();
1503 unsigned ObjSize = ObjectVT.getSizeInBits()/8;
1504 unsigned ArgSize = ObjSize;
1505 ISD::ArgFlagsTy Flags =
1506 cast<ARG_FLAGSSDNode>(Op.getOperand(ArgNo+3))->getArgFlags();
1507 // See if next argument requires stack alignment in ELF
1508 bool Align = Flags.isSplit();
1510 unsigned CurArgOffset = ArgOffset;
1512 // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary.
1513 if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 ||
1514 ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) {
1515 if (isVarArg || isPPC64) {
1516 MinReservedArea = ((MinReservedArea+15)/16)*16;
1517 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1521 } else nAltivecParamsAtEnd++;
1523 // Calculate min reserved area.
1524 MinReservedArea += CalculateStackSlotSize(Op.getValue(ArgNo),
1529 // FIXME alignment for ELF may not be right
1530 // FIXME the codegen can be much improved in some cases.
1531 // We do not have to keep everything in memory.
1532 if (Flags.isByVal()) {
1533 // ObjSize is the true size, ArgSize rounded up to multiple of registers.
1534 ObjSize = Flags.getByValSize();
1535 ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
1536 // Double word align in ELF
1537 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1538 // Objects of size 1 and 2 are right justified, everything else is
1539 // left justified. This means the memory address is adjusted forwards.
1540 if (ObjSize==1 || ObjSize==2) {
1541 CurArgOffset = CurArgOffset + (4 - ObjSize);
1543 // The value of the object is its address.
1544 int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
1545 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1546 ArgValues.push_back(FIN);
1547 if (ObjSize==1 || ObjSize==2) {
1548 if (GPR_idx != Num_GPR_Regs) {
1549 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1550 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1551 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1552 SDValue Store = DAG.getTruncStore(Val.getValue(1), Val, FIN,
1553 NULL, 0, ObjSize==1 ? MVT::i8 : MVT::i16 );
1554 MemOps.push_back(Store);
1556 if (isMachoABI) ArgOffset += PtrByteSize;
1558 ArgOffset += PtrByteSize;
1562 for (unsigned j = 0; j < ArgSize; j += PtrByteSize) {
1563 // Store whatever pieces of the object are in registers
1564 // to memory. ArgVal will be address of the beginning of
1566 if (GPR_idx != Num_GPR_Regs) {
1567 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1568 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1569 int FI = MFI->CreateFixedObject(PtrByteSize, ArgOffset);
1570 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1571 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1572 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1573 MemOps.push_back(Store);
1575 if (isMachoABI) ArgOffset += PtrByteSize;
1577 ArgOffset += ArgSize - (ArgOffset-CurArgOffset);
1584 switch (ObjectVT.getSimpleVT()) {
1585 default: assert(0 && "Unhandled argument type!");
1588 // Double word align in ELF
1589 if (Align && isELF32_ABI) GPR_idx += (GPR_idx % 2);
1591 if (GPR_idx != Num_GPR_Regs) {
1592 unsigned VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1593 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1594 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
1598 ArgSize = PtrByteSize;
1600 // Stack align in ELF
1601 if (needsLoad && Align && isELF32_ABI)
1602 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1603 // All int arguments reserve stack space in Macho ABI.
1604 if (isMachoABI || needsLoad) ArgOffset += PtrByteSize;
1608 case MVT::i64: // PPC64
1609 if (GPR_idx != Num_GPR_Regs) {
1610 unsigned VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1611 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1612 ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
1614 if (ObjectVT == MVT::i32) {
1615 // PPC64 passes i8, i16, and i32 values in i64 registers. Promote
1616 // value to MVT::i64 and then truncate to the correct register size.
1618 ArgVal = DAG.getNode(ISD::AssertSext, MVT::i64, ArgVal,
1619 DAG.getValueType(ObjectVT));
1620 else if (Flags.isZExt())
1621 ArgVal = DAG.getNode(ISD::AssertZext, MVT::i64, ArgVal,
1622 DAG.getValueType(ObjectVT));
1624 ArgVal = DAG.getNode(ISD::TRUNCATE, MVT::i32, ArgVal);
1630 ArgSize = PtrByteSize;
1632 // All int arguments reserve stack space in Macho ABI.
1633 if (isMachoABI || needsLoad) ArgOffset += 8;
1638 // Every 4 bytes of argument space consumes one of the GPRs available for
1639 // argument passing.
1640 if (GPR_idx != Num_GPR_Regs && isMachoABI) {
1642 if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
1645 if (FPR_idx != Num_FPR_Regs) {
1647 if (ObjectVT == MVT::f32)
1648 VReg = RegInfo.createVirtualRegister(&PPC::F4RCRegClass);
1650 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1651 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1652 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1658 // Stack align in ELF
1659 if (needsLoad && Align && isELF32_ABI)
1660 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
1661 // All FP arguments reserve stack space in Macho ABI.
1662 if (isMachoABI || needsLoad) ArgOffset += isPPC64 ? 8 : ObjSize;
1668 // Note that vector arguments in registers don't reserve stack space,
1669 // except in varargs functions.
1670 if (VR_idx != Num_VR_Regs) {
1671 unsigned VReg = RegInfo.createVirtualRegister(&PPC::VRRCRegClass);
1672 RegInfo.addLiveIn(VR[VR_idx], VReg);
1673 ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
1675 while ((ArgOffset % 16) != 0) {
1676 ArgOffset += PtrByteSize;
1677 if (GPR_idx != Num_GPR_Regs)
1681 GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs);
1685 if (!isVarArg && !isPPC64) {
1686 // Vectors go after all the nonvectors.
1687 CurArgOffset = VecArgOffset;
1690 // Vectors are aligned.
1691 ArgOffset = ((ArgOffset+15)/16)*16;
1692 CurArgOffset = ArgOffset;
1700 // We need to load the argument to a virtual register if we determined above
1701 // that we ran out of physical registers of the appropriate type.
1703 int FI = MFI->CreateFixedObject(ObjSize,
1704 CurArgOffset + (ArgSize - ObjSize),
1706 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1707 ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
1710 ArgValues.push_back(ArgVal);
1713 // Set the size that is at least reserved in caller of this function. Tail
1714 // call optimized function's reserved stack space needs to be aligned so that
1715 // taking the difference between two stack areas will result in an aligned
1717 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
1718 // Add the Altivec parameters at the end, if needed.
1719 if (nAltivecParamsAtEnd) {
1720 MinReservedArea = ((MinReservedArea+15)/16)*16;
1721 MinReservedArea += 16*nAltivecParamsAtEnd;
1724 std::max(MinReservedArea,
1725 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1726 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1727 getStackAlignment();
1728 unsigned AlignMask = TargetAlign-1;
1729 MinReservedArea = (MinReservedArea + AlignMask) & ~AlignMask;
1730 FI->setMinReservedArea(MinReservedArea);
1732 // If the function takes variable number of arguments, make a frame index for
1733 // the start of the first vararg value... for expansion of llvm.va_start.
1738 VarArgsNumGPR = GPR_idx;
1739 VarArgsNumFPR = FPR_idx;
1741 // Make room for Num_GPR_Regs, Num_FPR_Regs and for a possible frame
1743 depth = -(Num_GPR_Regs * PtrVT.getSizeInBits()/8 +
1744 Num_FPR_Regs * MVT(MVT::f64).getSizeInBits()/8 +
1745 PtrVT.getSizeInBits()/8);
1747 VarArgsStackOffset = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1754 VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
1756 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
1758 // In ELF 32 ABI, the fixed integer arguments of a variadic function are
1759 // stored to the VarArgsFrameIndex on the stack.
1761 for (GPR_idx = 0; GPR_idx != VarArgsNumGPR; ++GPR_idx) {
1762 SDValue Val = DAG.getRegister(GPR[GPR_idx], PtrVT);
1763 SDValue Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1764 MemOps.push_back(Store);
1765 // Increment the address by four for the next argument to store
1766 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1767 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1771 // If this function is vararg, store any remaining integer argument regs
1772 // to their spots on the stack so that they may be loaded by deferencing the
1773 // result of va_next.
1774 for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
1777 VReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass);
1779 VReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass);
1781 RegInfo.addLiveIn(GPR[GPR_idx], VReg);
1782 SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
1783 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1784 MemOps.push_back(Store);
1785 // Increment the address by four for the next argument to store
1786 SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
1787 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1790 // In ELF 32 ABI, the double arguments are stored to the VarArgsFrameIndex
1793 for (FPR_idx = 0; FPR_idx != VarArgsNumFPR; ++FPR_idx) {
1794 SDValue Val = DAG.getRegister(FPR[FPR_idx], MVT::f64);
1795 SDValue Store = DAG.getStore(Root, Val, FIN, NULL, 0);
1796 MemOps.push_back(Store);
1797 // Increment the address by eight for the next argument to store
1798 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1800 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1803 for (; FPR_idx != Num_FPR_Regs; ++FPR_idx) {
1805 VReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass);
1807 RegInfo.addLiveIn(FPR[FPR_idx], VReg);
1808 SDValue Val = DAG.getCopyFromReg(Root, VReg, MVT::f64);
1809 SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
1810 MemOps.push_back(Store);
1811 // Increment the address by eight for the next argument to store
1812 SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8,
1814 FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
1819 if (!MemOps.empty())
1820 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
1822 ArgValues.push_back(Root);
1824 // Return the new list of results.
1825 return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
1829 /// CalculateParameterAndLinkageAreaSize - Get the size of the paramter plus
1832 CalculateParameterAndLinkageAreaSize(SelectionDAG &DAG,
1837 CallSDNode *TheCall,
1838 unsigned &nAltivecParamsAtEnd) {
1839 // Count how many bytes are to be pushed on the stack, including the linkage
1840 // area, and parameter passing area. We start with 24/48 bytes, which is
1841 // prereserved space for [SP][CR][LR][3 x unused].
1842 unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
1843 unsigned NumOps = TheCall->getNumArgs();
1844 unsigned PtrByteSize = isPPC64 ? 8 : 4;
1846 // Add up all the space actually used.
1847 // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually
1848 // they all go in registers, but we must reserve stack space for them for
1849 // possible use by the caller. In varargs or 64-bit calls, parameters are
1850 // assigned stack space in order, with padding so Altivec parameters are
1852 nAltivecParamsAtEnd = 0;
1853 for (unsigned i = 0; i != NumOps; ++i) {
1854 SDValue Arg = TheCall->getArg(i);
1855 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1856 MVT ArgVT = Arg.getValueType();
1857 // Varargs Altivec parameters are padded to a 16 byte boundary.
1858 if (ArgVT==MVT::v4f32 || ArgVT==MVT::v4i32 ||
1859 ArgVT==MVT::v8i16 || ArgVT==MVT::v16i8) {
1860 if (!isVarArg && !isPPC64) {
1861 // Non-varargs Altivec parameters go after all the non-Altivec
1862 // parameters; handle those later so we know how much padding we need.
1863 nAltivecParamsAtEnd++;
1866 // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary.
1867 NumBytes = ((NumBytes+15)/16)*16;
1869 NumBytes += CalculateStackSlotSize(Arg, Flags, isVarArg, PtrByteSize);
1872 // Allow for Altivec parameters at the end, if needed.
1873 if (nAltivecParamsAtEnd) {
1874 NumBytes = ((NumBytes+15)/16)*16;
1875 NumBytes += 16*nAltivecParamsAtEnd;
1878 // The prolog code of the callee may store up to 8 GPR argument registers to
1879 // the stack, allowing va_start to index over them in memory if its varargs.
1880 // Because we cannot tell if this is needed on the caller side, we have to
1881 // conservatively assume that it is needed. As such, make sure we have at
1882 // least enough stack space for the caller to store the 8 GPRs.
1883 NumBytes = std::max(NumBytes,
1884 PPCFrameInfo::getMinCallFrameSize(isPPC64, isMachoABI));
1886 // Tail call needs the stack to be aligned.
1887 if (CC==CallingConv::Fast && PerformTailCallOpt) {
1888 unsigned TargetAlign = DAG.getMachineFunction().getTarget().getFrameInfo()->
1889 getStackAlignment();
1890 unsigned AlignMask = TargetAlign-1;
1891 NumBytes = (NumBytes + AlignMask) & ~AlignMask;
1897 /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be
1898 /// adjusted to accomodate the arguments for the tailcall.
1899 static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool IsTailCall,
1900 unsigned ParamSize) {
1902 if (!IsTailCall) return 0;
1904 PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
1905 unsigned CallerMinReservedArea = FI->getMinReservedArea();
1906 int SPDiff = (int)CallerMinReservedArea - (int)ParamSize;
1907 // Remember only if the new adjustement is bigger.
1908 if (SPDiff < FI->getTailCallSPDelta())
1909 FI->setTailCallSPDelta(SPDiff);
1914 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
1915 /// following the call is a return. A function is eligible if caller/callee
1916 /// calling conventions match, currently only fastcc supports tail calls, and
1917 /// the function CALL is immediatly followed by a RET.
1919 PPCTargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
1921 SelectionDAG& DAG) const {
1922 // Variable argument functions are not supported.
1923 if (!PerformTailCallOpt || TheCall->isVarArg())
1926 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
1927 MachineFunction &MF = DAG.getMachineFunction();
1928 unsigned CallerCC = MF.getFunction()->getCallingConv();
1929 unsigned CalleeCC = TheCall->getCallingConv();
1930 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) {
1931 // Functions containing by val parameters are not supported.
1932 for (unsigned i = 0; i != TheCall->getNumArgs(); i++) {
1933 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1934 if (Flags.isByVal()) return false;
1937 SDValue Callee = TheCall->getCallee();
1938 // Non PIC/GOT tail calls are supported.
1939 if (getTargetMachine().getRelocationModel() != Reloc::PIC_)
1942 // At the moment we can only do local tail calls (in same module, hidden
1943 // or protected) if we are generating PIC.
1944 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
1945 return G->getGlobal()->hasHiddenVisibility()
1946 || G->getGlobal()->hasProtectedVisibility();
1953 /// isCallCompatibleAddress - Return the immediate to use if the specified
1954 /// 32-bit value is representable in the immediate field of a BxA instruction.
1955 static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) {
1956 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
1959 int Addr = C->getZExtValue();
1960 if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
1961 (Addr << 6 >> 6) != Addr)
1962 return 0; // Top 6 bits have to be sext of immediate.
1964 return DAG.getConstant((int)C->getZExtValue() >> 2,
1965 DAG.getTargetLoweringInfo().getPointerTy()).getNode();
1970 struct TailCallArgumentInfo {
1975 TailCallArgumentInfo() : FrameIdx(0) {}
1980 /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot.
1982 StoreTailCallArgumentsToStackSlot(SelectionDAG &DAG,
1984 const SmallVector<TailCallArgumentInfo, 8> &TailCallArgs,
1985 SmallVector<SDValue, 8> &MemOpChains) {
1986 for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) {
1987 SDValue Arg = TailCallArgs[i].Arg;
1988 SDValue FIN = TailCallArgs[i].FrameIdxOp;
1989 int FI = TailCallArgs[i].FrameIdx;
1990 // Store relative to framepointer.
1991 MemOpChains.push_back(DAG.getStore(Chain, Arg, FIN,
1992 PseudoSourceValue::getFixedStack(FI),
1997 /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to
1998 /// the appropriate stack slot for the tail call optimized function call.
1999 static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG,
2000 MachineFunction &MF,
2008 // Calculate the new stack slot for the return address.
2009 int SlotSize = isPPC64 ? 8 : 4;
2010 int NewRetAddrLoc = SPDiff + PPCFrameInfo::getReturnSaveOffset(isPPC64,
2012 int NewRetAddr = MF.getFrameInfo()->CreateFixedObject(SlotSize,
2014 int NewFPLoc = SPDiff + PPCFrameInfo::getFramePointerSaveOffset(isPPC64,
2016 int NewFPIdx = MF.getFrameInfo()->CreateFixedObject(SlotSize, NewFPLoc);
2018 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2019 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT);
2020 Chain = DAG.getStore(Chain, OldRetAddr, NewRetAddrFrIdx,
2021 PseudoSourceValue::getFixedStack(NewRetAddr), 0);
2022 SDValue NewFramePtrIdx = DAG.getFrameIndex(NewFPIdx, VT);
2023 Chain = DAG.getStore(Chain, OldFP, NewFramePtrIdx,
2024 PseudoSourceValue::getFixedStack(NewFPIdx), 0);
2029 /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate
2030 /// the position of the argument.
2032 CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64,
2033 SDValue Arg, int SPDiff, unsigned ArgOffset,
2034 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2035 int Offset = ArgOffset + SPDiff;
2036 uint32_t OpSize = (Arg.getValueType().getSizeInBits()+7)/8;
2037 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
2038 MVT VT = isPPC64 ? MVT::i64 : MVT::i32;
2039 SDValue FIN = DAG.getFrameIndex(FI, VT);
2040 TailCallArgumentInfo Info;
2042 Info.FrameIdxOp = FIN;
2044 TailCallArguments.push_back(Info);
2047 /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address
2048 /// stack slot. Returns the chain as result and the loaded frame pointers in
2049 /// LROpOut/FPOpout. Used when tail calling.
2050 SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr(SelectionDAG & DAG,
2056 // Load the LR and FP stack slot for later adjusting.
2057 MVT VT = PPCSubTarget.isPPC64() ? MVT::i64 : MVT::i32;
2058 LROpOut = getReturnAddrFrameIndex(DAG);
2059 LROpOut = DAG.getLoad(VT, Chain, LROpOut, NULL, 0);
2060 Chain = SDValue(LROpOut.getNode(), 1);
2061 FPOpOut = getFramePointerFrameIndex(DAG);
2062 FPOpOut = DAG.getLoad(VT, Chain, FPOpOut, NULL, 0);
2063 Chain = SDValue(FPOpOut.getNode(), 1);
2068 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
2069 /// by "Src" to address "Dst" of size "Size". Alignment information is
2070 /// specified by the specific parameter attribute. The copy will be passed as
2071 /// a byval function parameter.
2072 /// Sometimes what we are copying is the end of a larger object, the part that
2073 /// does not fit in registers.
2075 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2076 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2078 SDValue SizeNode = DAG.getConstant(Size, MVT::i32);
2079 return DAG.getMemcpy(Chain, Dst, Src, SizeNode, Flags.getByValAlign(), false,
2083 /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of
2086 LowerMemOpCallTo(SelectionDAG &DAG, MachineFunction &MF, SDValue Chain,
2087 SDValue Arg, SDValue PtrOff, int SPDiff,
2088 unsigned ArgOffset, bool isPPC64, bool isTailCall,
2089 bool isVector, SmallVector<SDValue, 8> &MemOpChains,
2090 SmallVector<TailCallArgumentInfo, 8>& TailCallArguments) {
2091 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2096 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2098 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2099 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2100 DAG.getConstant(ArgOffset, PtrVT));
2102 MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
2103 // Calculate and remember argument location.
2104 } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset,
2108 SDValue PPCTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG,
2109 const PPCSubtarget &Subtarget,
2110 TargetMachine &TM) {
2111 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
2112 SDValue Chain = TheCall->getChain();
2113 bool isVarArg = TheCall->isVarArg();
2114 unsigned CC = TheCall->getCallingConv();
2115 bool isTailCall = TheCall->isTailCall()
2116 && CC == CallingConv::Fast && PerformTailCallOpt;
2117 SDValue Callee = TheCall->getCallee();
2118 unsigned NumOps = TheCall->getNumArgs();
2120 bool isMachoABI = Subtarget.isMachoABI();
2121 bool isELF32_ABI = Subtarget.isELF32_ABI();
2123 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2124 bool isPPC64 = PtrVT == MVT::i64;
2125 unsigned PtrByteSize = isPPC64 ? 8 : 4;
2127 MachineFunction &MF = DAG.getMachineFunction();
2129 // args_to_use will accumulate outgoing args for the PPCISD::CALL case in
2130 // SelectExpr to use to put the arguments in the appropriate registers.
2131 std::vector<SDValue> args_to_use;
2133 // Mark this function as potentially containing a function that contains a
2134 // tail call. As a consequence the frame pointer will be used for dynamicalloc
2135 // and restoring the callers stack pointer in this functions epilog. This is
2136 // done because by tail calling the called function might overwrite the value
2137 // in this function's (MF) stack pointer stack slot 0(SP).
2138 if (PerformTailCallOpt && CC==CallingConv::Fast)
2139 MF.getInfo<PPCFunctionInfo>()->setHasFastCall();
2141 unsigned nAltivecParamsAtEnd = 0;
2143 // Count how many bytes are to be pushed on the stack, including the linkage
2144 // area, and parameter passing area. We start with 24/48 bytes, which is
2145 // prereserved space for [SP][CR][LR][3 x unused].
2147 CalculateParameterAndLinkageAreaSize(DAG, isPPC64, isMachoABI, isVarArg, CC,
2148 TheCall, nAltivecParamsAtEnd);
2150 // Calculate by how many bytes the stack has to be adjusted in case of tail
2151 // call optimization.
2152 int SPDiff = CalculateTailCallSPDiff(DAG, isTailCall, NumBytes);
2154 // Adjust the stack pointer for the new arguments...
2155 // These operations are automatically eliminated by the prolog/epilog pass
2156 Chain = DAG.getCALLSEQ_START(Chain,
2157 DAG.getConstant(NumBytes, PtrVT));
2158 SDValue CallSeqStart = Chain;
2160 // Load the return address and frame pointer so it can be move somewhere else
2163 Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp);
2165 // Set up a copy of the stack pointer for use loading and storing any
2166 // arguments that may not fit in the registers available for argument
2170 StackPtr = DAG.getRegister(PPC::X1, MVT::i64);
2172 StackPtr = DAG.getRegister(PPC::R1, MVT::i32);
2174 // Figure out which arguments are going to go in registers, and which in
2175 // memory. Also, if this is a vararg function, floating point operations
2176 // must be stored to our stack, and loaded into integer regs as well, if
2177 // any integer regs are available for argument passing.
2178 unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64, isMachoABI);
2179 unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
2181 static const unsigned GPR_32[] = { // 32-bit registers.
2182 PPC::R3, PPC::R4, PPC::R5, PPC::R6,
2183 PPC::R7, PPC::R8, PPC::R9, PPC::R10,
2185 static const unsigned GPR_64[] = { // 64-bit registers.
2186 PPC::X3, PPC::X4, PPC::X5, PPC::X6,
2187 PPC::X7, PPC::X8, PPC::X9, PPC::X10,
2189 static const unsigned *FPR = GetFPR(Subtarget);
2191 static const unsigned VR[] = {
2192 PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
2193 PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
2195 const unsigned NumGPRs = array_lengthof(GPR_32);
2196 const unsigned NumFPRs = isMachoABI ? 13 : 8;
2197 const unsigned NumVRs = array_lengthof( VR);
2199 const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
2201 std::vector<std::pair<unsigned, SDValue> > RegsToPass;
2202 SmallVector<TailCallArgumentInfo, 8> TailCallArguments;
2204 SmallVector<SDValue, 8> MemOpChains;
2205 for (unsigned i = 0; i != NumOps; ++i) {
2207 SDValue Arg = TheCall->getArg(i);
2208 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
2209 // See if next argument requires stack alignment in ELF
2210 bool Align = Flags.isSplit();
2212 // PtrOff will be used to store the current argument to the stack if a
2213 // register cannot be found for it.
2216 // Stack align in ELF 32
2217 if (isELF32_ABI && Align)
2218 PtrOff = DAG.getConstant(ArgOffset + ((ArgOffset/4) % 2) * PtrByteSize,
2219 StackPtr.getValueType());
2221 PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
2223 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
2225 // On PPC64, promote integers to 64-bit values.
2226 if (isPPC64 && Arg.getValueType() == MVT::i32) {
2227 // FIXME: Should this use ANY_EXTEND if neither sext nor zext?
2228 unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2229 Arg = DAG.getNode(ExtOp, MVT::i64, Arg);
2232 // FIXME Elf untested, what are alignment rules?
2233 // FIXME memcpy is used way more than necessary. Correctness first.
2234 if (Flags.isByVal()) {
2235 unsigned Size = Flags.getByValSize();
2236 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2237 if (Size==1 || Size==2) {
2238 // Very small objects are passed right-justified.
2239 // Everything else is passed left-justified.
2240 MVT VT = (Size==1) ? MVT::i8 : MVT::i16;
2241 if (GPR_idx != NumGPRs) {
2242 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, PtrVT, Chain, Arg,
2244 MemOpChains.push_back(Load.getValue(1));
2245 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2247 ArgOffset += PtrByteSize;
2249 SDValue Const = DAG.getConstant(4 - Size, PtrOff.getValueType());
2250 SDValue AddPtr = DAG.getNode(ISD::ADD, PtrVT, PtrOff, Const);
2251 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, AddPtr,
2252 CallSeqStart.getNode()->getOperand(0),
2254 // This must go outside the CALLSEQ_START..END.
2255 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2256 CallSeqStart.getNode()->getOperand(1));
2257 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(),
2258 NewCallSeqStart.getNode());
2259 Chain = CallSeqStart = NewCallSeqStart;
2260 ArgOffset += PtrByteSize;
2264 // Copy entire object into memory. There are cases where gcc-generated
2265 // code assumes it is there, even if it could be put entirely into
2266 // registers. (This is not what the doc says.)
2267 SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff,
2268 CallSeqStart.getNode()->getOperand(0),
2270 // This must go outside the CALLSEQ_START..END.
2271 SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall,
2272 CallSeqStart.getNode()->getOperand(1));
2273 DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), NewCallSeqStart.getNode());
2274 Chain = CallSeqStart = NewCallSeqStart;
2275 // And copy the pieces of it that fit into registers.
2276 for (unsigned j=0; j<Size; j+=PtrByteSize) {
2277 SDValue Const = DAG.getConstant(j, PtrOff.getValueType());
2278 SDValue AddArg = DAG.getNode(ISD::ADD, PtrVT, Arg, Const);
2279 if (GPR_idx != NumGPRs) {
2280 SDValue Load = DAG.getLoad(PtrVT, Chain, AddArg, NULL, 0);
2281 MemOpChains.push_back(Load.getValue(1));
2282 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2284 ArgOffset += PtrByteSize;
2286 ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize;
2293 switch (Arg.getValueType().getSimpleVT()) {
2294 default: assert(0 && "Unexpected ValueType for argument!");
2297 // Double word align in ELF
2298 if (isELF32_ABI && Align) GPR_idx += (GPR_idx % 2);
2299 if (GPR_idx != NumGPRs) {
2300 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
2302 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2303 isPPC64, isTailCall, false, MemOpChains,
2307 if (inMem || isMachoABI) {
2308 // Stack align in ELF
2309 if (isELF32_ABI && Align)
2310 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2312 ArgOffset += PtrByteSize;
2317 if (FPR_idx != NumFPRs) {
2318 RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
2321 SDValue Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2322 MemOpChains.push_back(Store);
2324 // Float varargs are always shadowed in available integer registers
2325 if (GPR_idx != NumGPRs) {
2326 SDValue Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
2327 MemOpChains.push_back(Load.getValue(1));
2328 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2331 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
2332 SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
2333 PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
2334 SDValue Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
2335 MemOpChains.push_back(Load.getValue(1));
2336 if (isMachoABI) RegsToPass.push_back(std::make_pair(GPR[GPR_idx++],
2340 // If we have any FPRs remaining, we may also have GPRs remaining.
2341 // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
2344 if (GPR_idx != NumGPRs)
2346 if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 &&
2347 !isPPC64) // PPC64 has 64-bit GPR's obviously :)
2352 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2353 isPPC64, isTailCall, false, MemOpChains,
2357 if (inMem || isMachoABI) {
2358 // Stack align in ELF
2359 if (isELF32_ABI && Align)
2360 ArgOffset += ((ArgOffset/4) % 2) * PtrByteSize;
2364 ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8;
2372 // These go aligned on the stack, or in the corresponding R registers
2373 // when within range. The Darwin PPC ABI doc claims they also go in
2374 // V registers; in fact gcc does this only for arguments that are
2375 // prototyped, not for those that match the ... We do it for all
2376 // arguments, seems to work.
2377 while (ArgOffset % 16 !=0) {
2378 ArgOffset += PtrByteSize;
2379 if (GPR_idx != NumGPRs)
2382 // We could elide this store in the case where the object fits
2383 // entirely in R registers. Maybe later.
2384 PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr,
2385 DAG.getConstant(ArgOffset, PtrVT));
2386 SDValue Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
2387 MemOpChains.push_back(Store);
2388 if (VR_idx != NumVRs) {
2389 SDValue Load = DAG.getLoad(MVT::v4f32, Store, PtrOff, NULL, 0);
2390 MemOpChains.push_back(Load.getValue(1));
2391 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load));
2394 for (unsigned i=0; i<16; i+=PtrByteSize) {
2395 if (GPR_idx == NumGPRs)
2397 SDValue Ix = DAG.getNode(ISD::ADD, PtrVT, PtrOff,
2398 DAG.getConstant(i, PtrVT));
2399 SDValue Load = DAG.getLoad(PtrVT, Store, Ix, NULL, 0);
2400 MemOpChains.push_back(Load.getValue(1));
2401 RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
2406 // Non-varargs Altivec params generally go in registers, but have
2407 // stack space allocated at the end.
2408 if (VR_idx != NumVRs) {
2409 // Doesn't have GPR space allocated.
2410 RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg));
2411 } else if (nAltivecParamsAtEnd==0) {
2412 // We are emitting Altivec params in order.
2413 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2414 isPPC64, isTailCall, true, MemOpChains,
2421 // If all Altivec parameters fit in registers, as they usually do,
2422 // they get stack space following the non-Altivec parameters. We
2423 // don't track this here because nobody below needs it.
2424 // If there are more Altivec parameters than fit in registers emit
2426 if (!isVarArg && nAltivecParamsAtEnd > NumVRs) {
2428 // Offset is aligned; skip 1st 12 params which go in V registers.
2429 ArgOffset = ((ArgOffset+15)/16)*16;
2431 for (unsigned i = 0; i != NumOps; ++i) {
2432 SDValue Arg = TheCall->getArg(i);
2433 MVT ArgType = Arg.getValueType();
2434 if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 ||
2435 ArgType==MVT::v8i16 || ArgType==MVT::v16i8) {
2438 // We are emitting Altivec params in order.
2439 LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset,
2440 isPPC64, isTailCall, true, MemOpChains,
2448 if (!MemOpChains.empty())
2449 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2450 &MemOpChains[0], MemOpChains.size());
2452 // Build a sequence of copy-to-reg nodes chained together with token chain
2453 // and flag operands which copy the outgoing args into the appropriate regs.
2455 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2456 Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
2458 InFlag = Chain.getValue(1);
2461 // With the ELF 32 ABI, set CR6 to true if this is a vararg call.
2462 if (isVarArg && isELF32_ABI) {
2463 SDValue SetCR(DAG.getTargetNode(PPC::CRSET, MVT::i32), 0);
2464 Chain = DAG.getCopyToReg(Chain, PPC::CR1EQ, SetCR, InFlag);
2465 InFlag = Chain.getValue(1);
2468 // Emit a sequence of copyto/copyfrom virtual registers for arguments that
2469 // might overwrite each other in case of tail call optimization.
2471 SmallVector<SDValue, 8> MemOpChains2;
2472 // Do not flag preceeding copytoreg stuff together with the following stuff.
2474 StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments,
2476 if (!MemOpChains2.empty())
2477 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2478 &MemOpChains2[0], MemOpChains2.size());
2480 // Store the return address to the appropriate stack slot.
2481 Chain = EmitTailCallStoreFPAndRetAddr(DAG, MF, Chain, LROp, FPOp, SPDiff,
2482 isPPC64, isMachoABI);
2485 // Emit callseq_end just before tailcall node.
2487 SmallVector<SDValue, 8> CallSeqOps;
2488 SDVTList CallSeqNodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2489 CallSeqOps.push_back(Chain);
2490 CallSeqOps.push_back(DAG.getIntPtrConstant(NumBytes));
2491 CallSeqOps.push_back(DAG.getIntPtrConstant(0));
2492 if (InFlag.getNode())
2493 CallSeqOps.push_back(InFlag);
2494 Chain = DAG.getNode(ISD::CALLSEQ_END, CallSeqNodeTys, &CallSeqOps[0],
2496 InFlag = Chain.getValue(1);
2499 std::vector<MVT> NodeTys;
2500 NodeTys.push_back(MVT::Other); // Returns a chain
2501 NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
2503 SmallVector<SDValue, 8> Ops;
2504 unsigned CallOpc = isMachoABI? PPCISD::CALL_Macho : PPCISD::CALL_ELF;
2506 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2507 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2508 // node so that legalize doesn't hack it.
2509 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
2510 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), Callee.getValueType());
2511 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
2512 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), Callee.getValueType());
2513 else if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG))
2514 // If this is an absolute destination address, use the munged value.
2515 Callee = SDValue(Dest, 0);
2517 // Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
2518 // to do the call, we can't use PPCISD::CALL.
2519 SDValue MTCTROps[] = {Chain, Callee, InFlag};
2520 Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps,
2521 2 + (InFlag.getNode() != 0));
2522 InFlag = Chain.getValue(1);
2524 // Copy the callee address into R12/X12 on darwin.
2526 unsigned Reg = Callee.getValueType() == MVT::i32 ? PPC::R12 : PPC::X12;
2527 Chain = DAG.getCopyToReg(Chain, Reg, Callee, InFlag);
2528 InFlag = Chain.getValue(1);
2532 NodeTys.push_back(MVT::Other);
2533 NodeTys.push_back(MVT::Flag);
2534 Ops.push_back(Chain);
2535 CallOpc = isMachoABI ? PPCISD::BCTRL_Macho : PPCISD::BCTRL_ELF;
2537 // Add CTR register as callee so a bctr can be emitted later.
2539 Ops.push_back(DAG.getRegister(PPC::CTR, getPointerTy()));
2542 // If this is a direct call, pass the chain and the callee.
2543 if (Callee.getNode()) {
2544 Ops.push_back(Chain);
2545 Ops.push_back(Callee);
2547 // If this is a tail call add stack pointer delta.
2549 Ops.push_back(DAG.getConstant(SPDiff, MVT::i32));
2551 // Add argument registers to the end of the list so that they are known live
2553 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2554 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2555 RegsToPass[i].second.getValueType()));
2557 // When performing tail call optimization the callee pops its arguments off
2558 // the stack. Account for this here so these bytes can be pushed back on in
2559 // PPCRegisterInfo::eliminateCallFramePseudoInstr.
2560 int BytesCalleePops =
2561 (CC==CallingConv::Fast && PerformTailCallOpt) ? NumBytes : 0;
2563 if (InFlag.getNode())
2564 Ops.push_back(InFlag);
2568 assert(InFlag.getNode() &&
2569 "Flag must be set. Depend on flag being set in LowerRET");
2570 Chain = DAG.getNode(PPCISD::TAILCALL,
2571 TheCall->getVTList(), &Ops[0], Ops.size());
2572 return SDValue(Chain.getNode(), Op.getResNo());
2575 Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
2576 InFlag = Chain.getValue(1);
2578 Chain = DAG.getCALLSEQ_END(Chain,
2579 DAG.getConstant(NumBytes, PtrVT),
2580 DAG.getConstant(BytesCalleePops, PtrVT),
2582 if (TheCall->getValueType(0) != MVT::Other)
2583 InFlag = Chain.getValue(1);
2585 SmallVector<SDValue, 16> ResultVals;
2586 SmallVector<CCValAssign, 16> RVLocs;
2587 unsigned CallerCC = DAG.getMachineFunction().getFunction()->getCallingConv();
2588 CCState CCInfo(CallerCC, isVarArg, TM, RVLocs);
2589 CCInfo.AnalyzeCallResult(TheCall, RetCC_PPC);
2591 // Copy all of the result registers out of their specified physreg.
2592 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2593 CCValAssign &VA = RVLocs[i];
2594 MVT VT = VA.getValVT();
2595 assert(VA.isRegLoc() && "Can only return in registers!");
2596 Chain = DAG.getCopyFromReg(Chain, VA.getLocReg(), VT, InFlag).getValue(1);
2597 ResultVals.push_back(Chain.getValue(0));
2598 InFlag = Chain.getValue(2);
2601 // If the function returns void, just return the chain.
2605 // Otherwise, merge everything together with a MERGE_VALUES node.
2606 ResultVals.push_back(Chain);
2607 SDValue Res = DAG.getMergeValues(TheCall->getVTList(), &ResultVals[0],
2609 return Res.getValue(Op.getResNo());
2612 SDValue PPCTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG,
2613 TargetMachine &TM) {
2614 SmallVector<CCValAssign, 16> RVLocs;
2615 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
2616 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
2617 CCState CCInfo(CC, isVarArg, TM, RVLocs);
2618 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_PPC);
2620 // If this is the first return lowered for this function, add the regs to the
2621 // liveout set for the function.
2622 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
2623 for (unsigned i = 0; i != RVLocs.size(); ++i)
2624 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2627 SDValue Chain = Op.getOperand(0);
2629 Chain = GetPossiblePreceedingTailCall(Chain, PPCISD::TAILCALL);
2630 if (Chain.getOpcode() == PPCISD::TAILCALL) {
2631 SDValue TailCall = Chain;
2632 SDValue TargetAddress = TailCall.getOperand(1);
2633 SDValue StackAdjustment = TailCall.getOperand(2);
2635 assert(((TargetAddress.getOpcode() == ISD::Register &&
2636 cast<RegisterSDNode>(TargetAddress)->getReg() == PPC::CTR) ||
2637 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
2638 TargetAddress.getOpcode() == ISD::TargetGlobalAddress ||
2639 isa<ConstantSDNode>(TargetAddress)) &&
2640 "Expecting an global address, external symbol, absolute value or register");
2642 assert(StackAdjustment.getOpcode() == ISD::Constant &&
2643 "Expecting a const value");
2645 SmallVector<SDValue,8> Operands;
2646 Operands.push_back(Chain.getOperand(0));
2647 Operands.push_back(TargetAddress);
2648 Operands.push_back(StackAdjustment);
2649 // Copy registers used by the call. Last operand is a flag so it is not
2651 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
2652 Operands.push_back(Chain.getOperand(i));
2654 return DAG.getNode(PPCISD::TC_RETURN, MVT::Other, &Operands[0],
2660 // Copy the result values into the output registers.
2661 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2662 CCValAssign &VA = RVLocs[i];
2663 assert(VA.isRegLoc() && "Can only return in registers!");
2664 Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
2665 Flag = Chain.getValue(1);
2669 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain, Flag);
2671 return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Chain);
2674 SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, SelectionDAG &DAG,
2675 const PPCSubtarget &Subtarget) {
2676 // When we pop the dynamic allocation we need to restore the SP link.
2678 // Get the corect type for pointers.
2679 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2681 // Construct the stack pointer operand.
2682 bool IsPPC64 = Subtarget.isPPC64();
2683 unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
2684 SDValue StackPtr = DAG.getRegister(SP, PtrVT);
2686 // Get the operands for the STACKRESTORE.
2687 SDValue Chain = Op.getOperand(0);
2688 SDValue SaveSP = Op.getOperand(1);
2690 // Load the old link SP.
2691 SDValue LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
2693 // Restore the stack pointer.
2694 Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
2696 // Store the old link SP.
2697 return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
2703 PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG & DAG) const {
2704 MachineFunction &MF = DAG.getMachineFunction();
2705 bool IsPPC64 = PPCSubTarget.isPPC64();
2706 bool isMachoABI = PPCSubTarget.isMachoABI();
2707 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2709 // Get current frame pointer save index. The users of this index will be
2710 // primarily DYNALLOC instructions.
2711 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2712 int RASI = FI->getReturnAddrSaveIndex();
2714 // If the frame pointer save index hasn't been defined yet.
2716 // Find out what the fix offset of the frame pointer save area.
2717 int LROffset = PPCFrameInfo::getReturnSaveOffset(IsPPC64, isMachoABI);
2718 // Allocate the frame index for frame pointer save area.
2719 RASI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, LROffset);
2721 FI->setReturnAddrSaveIndex(RASI);
2723 return DAG.getFrameIndex(RASI, PtrVT);
2727 PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const {
2728 MachineFunction &MF = DAG.getMachineFunction();
2729 bool IsPPC64 = PPCSubTarget.isPPC64();
2730 bool isMachoABI = PPCSubTarget.isMachoABI();
2731 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2733 // Get current frame pointer save index. The users of this index will be
2734 // primarily DYNALLOC instructions.
2735 PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
2736 int FPSI = FI->getFramePointerSaveIndex();
2738 // If the frame pointer save index hasn't been defined yet.
2740 // Find out what the fix offset of the frame pointer save area.
2741 int FPOffset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64, isMachoABI);
2743 // Allocate the frame index for frame pointer save area.
2744 FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, FPOffset);
2746 FI->setFramePointerSaveIndex(FPSI);
2748 return DAG.getFrameIndex(FPSI, PtrVT);
2751 SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
2753 const PPCSubtarget &Subtarget) {
2755 SDValue Chain = Op.getOperand(0);
2756 SDValue Size = Op.getOperand(1);
2758 // Get the corect type for pointers.
2759 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2761 SDValue NegSize = DAG.getNode(ISD::SUB, PtrVT,
2762 DAG.getConstant(0, PtrVT), Size);
2763 // Construct a node for the frame pointer save index.
2764 SDValue FPSIdx = getFramePointerFrameIndex(DAG);
2765 // Build a DYNALLOC node.
2766 SDValue Ops[3] = { Chain, NegSize, FPSIdx };
2767 SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
2768 return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
2771 /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
2773 SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
2774 // Not FP? Not a fsel.
2775 if (!Op.getOperand(0).getValueType().isFloatingPoint() ||
2776 !Op.getOperand(2).getValueType().isFloatingPoint())
2779 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2781 // Cannot handle SETEQ/SETNE.
2782 if (CC == ISD::SETEQ || CC == ISD::SETNE) return SDValue();
2784 MVT ResVT = Op.getValueType();
2785 MVT CmpVT = Op.getOperand(0).getValueType();
2786 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
2787 SDValue TV = Op.getOperand(2), FV = Op.getOperand(3);
2789 // If the RHS of the comparison is a 0.0, we don't need to do the
2790 // subtraction at all.
2791 if (isFloatingPointZero(RHS))
2793 default: break; // SETUO etc aren't handled by fsel.
2796 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2799 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2800 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2801 return DAG.getNode(PPCISD::FSEL, ResVT, LHS, TV, FV);
2804 std::swap(TV, FV); // fsel is natively setge, swap operands for setlt
2807 if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits
2808 LHS = DAG.getNode(ISD::FP_EXTEND, MVT::f64, LHS);
2809 return DAG.getNode(PPCISD::FSEL, ResVT,
2810 DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
2815 default: break; // SETUO etc aren't handled by fsel.
2818 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2819 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2820 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2821 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2824 Cmp = DAG.getNode(ISD::FSUB, CmpVT, LHS, RHS);
2825 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2826 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2827 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2830 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2831 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2832 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2833 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, FV, TV);
2836 Cmp = DAG.getNode(ISD::FSUB, CmpVT, RHS, LHS);
2837 if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits
2838 Cmp = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Cmp);
2839 return DAG.getNode(PPCISD::FSEL, ResVT, Cmp, TV, FV);
2844 // FIXME: Split this code up when LegalizeDAGTypes lands.
2845 SDValue PPCTargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
2846 assert(Op.getOperand(0).getValueType().isFloatingPoint());
2847 SDValue Src = Op.getOperand(0);
2848 if (Src.getValueType() == MVT::f32)
2849 Src = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Src);
2852 switch (Op.getValueType().getSimpleVT()) {
2853 default: assert(0 && "Unhandled FP_TO_SINT type in custom expander!");
2855 Tmp = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Src);
2858 Tmp = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Src);
2862 // Convert the FP value to an int value through memory.
2863 SDValue FIPtr = DAG.CreateStackTemporary(MVT::f64);
2865 // Emit a store to the stack slot.
2866 SDValue Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
2868 // Result is a load from the stack slot. If loading 4 bytes, make sure to
2870 if (Op.getValueType() == MVT::i32)
2871 FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
2872 DAG.getConstant(4, FIPtr.getValueType()));
2873 return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
2876 SDValue PPCTargetLowering::LowerFP_ROUND_INREG(SDValue Op,
2877 SelectionDAG &DAG) {
2878 assert(Op.getValueType() == MVT::ppcf128);
2879 SDNode *Node = Op.getNode();
2880 assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
2881 assert(Node->getOperand(0).getNode()->getOpcode() == ISD::BUILD_PAIR);
2882 SDValue Lo = Node->getOperand(0).getNode()->getOperand(0);
2883 SDValue Hi = Node->getOperand(0).getNode()->getOperand(1);
2885 // This sequence changes FPSCR to do round-to-zero, adds the two halves
2886 // of the long double, and puts FPSCR back the way it was. We do not
2887 // actually model FPSCR.
2888 std::vector<MVT> NodeTys;
2889 SDValue Ops[4], Result, MFFSreg, InFlag, FPreg;
2891 NodeTys.push_back(MVT::f64); // Return register
2892 NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
2893 Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
2894 MFFSreg = Result.getValue(0);
2895 InFlag = Result.getValue(1);
2898 NodeTys.push_back(MVT::Flag); // Returns a flag
2899 Ops[0] = DAG.getConstant(31, MVT::i32);
2901 Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
2902 InFlag = Result.getValue(0);
2905 NodeTys.push_back(MVT::Flag); // Returns a flag
2906 Ops[0] = DAG.getConstant(30, MVT::i32);
2908 Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
2909 InFlag = Result.getValue(0);
2912 NodeTys.push_back(MVT::f64); // result of add
2913 NodeTys.push_back(MVT::Flag); // Returns a flag
2917 Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
2918 FPreg = Result.getValue(0);
2919 InFlag = Result.getValue(1);
2922 NodeTys.push_back(MVT::f64);
2923 Ops[0] = DAG.getConstant(1, MVT::i32);
2927 Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
2928 FPreg = Result.getValue(0);
2930 // We know the low half is about to be thrown away, so just use something
2932 return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
2935 SDValue PPCTargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
2936 // Don't handle ppc_fp128 here; let it be lowered to a libcall.
2937 if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64)
2940 if (Op.getOperand(0).getValueType() == MVT::i64) {
2941 SDValue Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
2942 SDValue FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
2943 if (Op.getValueType() == MVT::f32)
2944 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2948 assert(Op.getOperand(0).getValueType() == MVT::i32 &&
2949 "Unhandled SINT_TO_FP type in custom expander!");
2950 // Since we only generate this in 64-bit mode, we can take advantage of
2951 // 64-bit registers. In particular, sign extend the input value into the
2952 // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
2953 // then lfd it and fcfid it.
2954 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
2955 int FrameIdx = FrameInfo->CreateStackObject(8, 8);
2956 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2957 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
2959 SDValue Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
2962 // STD the extended value into the stack slot.
2963 MachineMemOperand MO(PseudoSourceValue::getFixedStack(FrameIdx),
2964 MachineMemOperand::MOStore, 0, 8, 8);
2965 SDValue Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
2966 DAG.getEntryNode(), Ext64, FIdx,
2967 DAG.getMemOperand(MO));
2968 // Load the value as a double.
2969 SDValue Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
2971 // FCFID it and return it.
2972 SDValue FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
2973 if (Op.getValueType() == MVT::f32)
2974 FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP, DAG.getIntPtrConstant(0));
2978 SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
2980 The rounding mode is in bits 30:31 of FPSR, and has the following
2987 FLT_ROUNDS, on the other hand, expects the following:
2994 To perform the conversion, we do:
2995 ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1))
2998 MachineFunction &MF = DAG.getMachineFunction();
2999 MVT VT = Op.getValueType();
3000 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3001 std::vector<MVT> NodeTys;
3002 SDValue MFFSreg, InFlag;
3004 // Save FP Control Word to register
3005 NodeTys.push_back(MVT::f64); // return register
3006 NodeTys.push_back(MVT::Flag); // unused in this context
3007 SDValue Chain = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
3009 // Save FP register to stack slot
3010 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
3011 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
3012 SDValue Store = DAG.getStore(DAG.getEntryNode(), Chain,
3013 StackSlot, NULL, 0);
3015 // Load FP Control Word from low 32 bits of stack slot.
3016 SDValue Four = DAG.getConstant(4, PtrVT);
3017 SDValue Addr = DAG.getNode(ISD::ADD, PtrVT, StackSlot, Four);
3018 SDValue CWD = DAG.getLoad(MVT::i32, Store, Addr, NULL, 0);
3020 // Transform as necessary
3022 DAG.getNode(ISD::AND, MVT::i32,
3023 CWD, DAG.getConstant(3, MVT::i32));
3025 DAG.getNode(ISD::SRL, MVT::i32,
3026 DAG.getNode(ISD::AND, MVT::i32,
3027 DAG.getNode(ISD::XOR, MVT::i32,
3028 CWD, DAG.getConstant(3, MVT::i32)),
3029 DAG.getConstant(3, MVT::i32)),
3030 DAG.getConstant(1, MVT::i8));
3033 DAG.getNode(ISD::XOR, MVT::i32, CWD1, CWD2);
3035 return DAG.getNode((VT.getSizeInBits() < 16 ?
3036 ISD::TRUNCATE : ISD::ZERO_EXTEND), VT, RetVal);
3039 SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) {
3040 MVT VT = Op.getValueType();
3041 unsigned BitWidth = VT.getSizeInBits();
3042 assert(Op.getNumOperands() == 3 &&
3043 VT == Op.getOperand(1).getValueType() &&
3046 // Expand into a bunch of logical ops. Note that these ops
3047 // depend on the PPC behavior for oversized shift amounts.
3048 SDValue Lo = Op.getOperand(0);
3049 SDValue Hi = Op.getOperand(1);
3050 SDValue Amt = Op.getOperand(2);
3051 MVT AmtVT = Amt.getValueType();
3053 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3054 DAG.getConstant(BitWidth, AmtVT), Amt);
3055 SDValue Tmp2 = DAG.getNode(PPCISD::SHL, VT, Hi, Amt);
3056 SDValue Tmp3 = DAG.getNode(PPCISD::SRL, VT, Lo, Tmp1);
3057 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3058 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3059 DAG.getConstant(-BitWidth, AmtVT));
3060 SDValue Tmp6 = DAG.getNode(PPCISD::SHL, VT, Lo, Tmp5);
3061 SDValue OutHi = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
3062 SDValue OutLo = DAG.getNode(PPCISD::SHL, VT, Lo, Amt);
3063 SDValue OutOps[] = { OutLo, OutHi };
3064 return DAG.getMergeValues(OutOps, 2);
3067 SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) {
3068 MVT VT = Op.getValueType();
3069 unsigned BitWidth = VT.getSizeInBits();
3070 assert(Op.getNumOperands() == 3 &&
3071 VT == Op.getOperand(1).getValueType() &&
3074 // Expand into a bunch of logical ops. Note that these ops
3075 // depend on the PPC behavior for oversized shift amounts.
3076 SDValue Lo = Op.getOperand(0);
3077 SDValue Hi = Op.getOperand(1);
3078 SDValue Amt = Op.getOperand(2);
3079 MVT AmtVT = Amt.getValueType();
3081 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3082 DAG.getConstant(BitWidth, AmtVT), Amt);
3083 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
3084 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
3085 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3086 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3087 DAG.getConstant(-BitWidth, AmtVT));
3088 SDValue Tmp6 = DAG.getNode(PPCISD::SRL, VT, Hi, Tmp5);
3089 SDValue OutLo = DAG.getNode(ISD::OR, VT, Tmp4, Tmp6);
3090 SDValue OutHi = DAG.getNode(PPCISD::SRL, VT, Hi, Amt);
3091 SDValue OutOps[] = { OutLo, OutHi };
3092 return DAG.getMergeValues(OutOps, 2);
3095 SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) {
3096 MVT VT = Op.getValueType();
3097 unsigned BitWidth = VT.getSizeInBits();
3098 assert(Op.getNumOperands() == 3 &&
3099 VT == Op.getOperand(1).getValueType() &&
3102 // Expand into a bunch of logical ops, followed by a select_cc.
3103 SDValue Lo = Op.getOperand(0);
3104 SDValue Hi = Op.getOperand(1);
3105 SDValue Amt = Op.getOperand(2);
3106 MVT AmtVT = Amt.getValueType();
3108 SDValue Tmp1 = DAG.getNode(ISD::SUB, AmtVT,
3109 DAG.getConstant(BitWidth, AmtVT), Amt);
3110 SDValue Tmp2 = DAG.getNode(PPCISD::SRL, VT, Lo, Amt);
3111 SDValue Tmp3 = DAG.getNode(PPCISD::SHL, VT, Hi, Tmp1);
3112 SDValue Tmp4 = DAG.getNode(ISD::OR , VT, Tmp2, Tmp3);
3113 SDValue Tmp5 = DAG.getNode(ISD::ADD, AmtVT, Amt,
3114 DAG.getConstant(-BitWidth, AmtVT));
3115 SDValue Tmp6 = DAG.getNode(PPCISD::SRA, VT, Hi, Tmp5);
3116 SDValue OutHi = DAG.getNode(PPCISD::SRA, VT, Hi, Amt);
3117 SDValue OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, AmtVT),
3118 Tmp4, Tmp6, ISD::SETLE);
3119 SDValue OutOps[] = { OutLo, OutHi };
3120 return DAG.getMergeValues(OutOps, 2);
3123 //===----------------------------------------------------------------------===//
3124 // Vector related lowering.
3127 // If this is a vector of constants or undefs, get the bits. A bit in
3128 // UndefBits is set if the corresponding element of the vector is an
3129 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
3130 // zero. Return true if this is not an array of constants, false if it is.
3132 static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
3133 uint64_t UndefBits[2]) {
3134 // Start with zero'd results.
3135 VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
3137 unsigned EltBitSize = BV->getOperand(0).getValueType().getSizeInBits();
3138 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
3139 SDValue OpVal = BV->getOperand(i);
3141 unsigned PartNo = i >= e/2; // In the upper 128 bits?
3142 unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
3144 uint64_t EltBits = 0;
3145 if (OpVal.getOpcode() == ISD::UNDEF) {
3146 uint64_t EltUndefBits = ~0U >> (32-EltBitSize);
3147 UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
3149 } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
3150 EltBits = CN->getZExtValue() & (~0U >> (32-EltBitSize));
3151 } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
3152 assert(CN->getValueType(0) == MVT::f32 &&
3153 "Only one legal FP vector type!");
3154 EltBits = FloatToBits(CN->getValueAPF().convertToFloat());
3156 // Nonconstant element.
3160 VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
3163 //printf("%llx %llx %llx %llx\n",
3164 // VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
3168 // If this is a splat (repetition) of a value across the whole vector, return
3169 // the smallest size that splats it. For example, "0x01010101010101..." is a
3170 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
3171 // SplatSize = 1 byte.
3172 static bool isConstantSplat(const uint64_t Bits128[2],
3173 const uint64_t Undef128[2],
3174 unsigned &SplatBits, unsigned &SplatUndef,
3175 unsigned &SplatSize) {
3177 // Don't let undefs prevent splats from matching. See if the top 64-bits are
3178 // the same as the lower 64-bits, ignoring undefs.
3179 if ((Bits128[0] & ~Undef128[1]) != (Bits128[1] & ~Undef128[0]))
3180 return false; // Can't be a splat if two pieces don't match.
3182 uint64_t Bits64 = Bits128[0] | Bits128[1];
3183 uint64_t Undef64 = Undef128[0] & Undef128[1];
3185 // Check that the top 32-bits are the same as the lower 32-bits, ignoring
3187 if ((Bits64 & (~Undef64 >> 32)) != ((Bits64 >> 32) & ~Undef64))
3188 return false; // Can't be a splat if two pieces don't match.
3190 uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
3191 uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
3193 // If the top 16-bits are different than the lower 16-bits, ignoring
3194 // undefs, we have an i32 splat.
3195 if ((Bits32 & (~Undef32 >> 16)) != ((Bits32 >> 16) & ~Undef32)) {
3197 SplatUndef = Undef32;
3202 uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
3203 uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
3205 // If the top 8-bits are different than the lower 8-bits, ignoring
3206 // undefs, we have an i16 splat.
3207 if ((Bits16 & (uint16_t(~Undef16) >> 8)) != ((Bits16 >> 8) & ~Undef16)) {
3209 SplatUndef = Undef16;
3214 // Otherwise, we have an 8-bit splat.
3215 SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
3216 SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
3221 /// BuildSplatI - Build a canonical splati of Val with an element size of
3222 /// SplatSize. Cast the result to VT.
3223 static SDValue BuildSplatI(int Val, unsigned SplatSize, MVT VT,
3224 SelectionDAG &DAG) {
3225 assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
3227 static const MVT VTys[] = { // canonical VT to use for each size.
3228 MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
3231 MVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
3233 // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
3237 MVT CanonicalVT = VTys[SplatSize-1];
3239 // Build a canonical splat for this value.
3240 SDValue Elt = DAG.getConstant(Val, CanonicalVT.getVectorElementType());
3241 SmallVector<SDValue, 8> Ops;
3242 Ops.assign(CanonicalVT.getVectorNumElements(), Elt);
3243 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
3244 &Ops[0], Ops.size());
3245 return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
3248 /// BuildIntrinsicOp - Return a binary operator intrinsic node with the
3249 /// specified intrinsic ID.
3250 static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS,
3252 MVT DestVT = MVT::Other) {
3253 if (DestVT == MVT::Other) DestVT = LHS.getValueType();
3254 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
3255 DAG.getConstant(IID, MVT::i32), LHS, RHS);
3258 /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the
3259 /// specified intrinsic ID.
3260 static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1,
3261 SDValue Op2, SelectionDAG &DAG,
3262 MVT DestVT = MVT::Other) {
3263 if (DestVT == MVT::Other) DestVT = Op0.getValueType();
3264 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DestVT,
3265 DAG.getConstant(IID, MVT::i32), Op0, Op1, Op2);
3269 /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified
3270 /// amount. The result has the specified value type.
3271 static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt,
3272 MVT VT, SelectionDAG &DAG) {
3273 // Force LHS/RHS to be the right type.
3274 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
3275 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
3278 for (unsigned i = 0; i != 16; ++i)
3279 Ops[i] = DAG.getConstant(i+Amt, MVT::i8);
3280 SDValue T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
3281 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
3282 return DAG.getNode(ISD::BIT_CONVERT, VT, T);
3285 // If this is a case we can't handle, return null and let the default
3286 // expansion code take care of it. If we CAN select this case, and if it
3287 // selects to a single instruction, return Op. Otherwise, if we can codegen
3288 // this case more efficiently than a constant pool load, lower it to the
3289 // sequence of ops that should be used.
3290 SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op,
3291 SelectionDAG &DAG) {
3292 // If this is a vector of constants or undefs, get the bits. A bit in
3293 // UndefBits is set if the corresponding element of the vector is an
3294 // ISD::UNDEF value. For undefs, the corresponding VectorBits values are
3296 uint64_t VectorBits[2];
3297 uint64_t UndefBits[2];
3298 if (GetConstantBuildVectorBits(Op.getNode(), VectorBits, UndefBits))
3299 return SDValue(); // Not a constant vector.
3301 // If this is a splat (repetition) of a value across the whole vector, return
3302 // the smallest size that splats it. For example, "0x01010101010101..." is a
3303 // splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
3304 // SplatSize = 1 byte.
3305 unsigned SplatBits, SplatUndef, SplatSize;
3306 if (isConstantSplat(VectorBits, UndefBits, SplatBits, SplatUndef, SplatSize)){
3307 bool HasAnyUndefs = (UndefBits[0] | UndefBits[1]) != 0;
3309 // First, handle single instruction cases.
3312 if (SplatBits == 0) {
3313 // Canonicalize all zero vectors to be v4i32.
3314 if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) {
3315 SDValue Z = DAG.getConstant(0, MVT::i32);
3316 Z = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Z, Z, Z, Z);
3317 Op = DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Z);
3322 // If the sign extended value is in the range [-16,15], use VSPLTI[bhw].
3323 int32_t SextVal= int32_t(SplatBits << (32-8*SplatSize)) >> (32-8*SplatSize);
3324 if (SextVal >= -16 && SextVal <= 15)
3325 return BuildSplatI(SextVal, SplatSize, Op.getValueType(), DAG);
3328 // Two instruction sequences.
3330 // If this value is in the range [-32,30] and is even, use:
3331 // tmp = VSPLTI[bhw], result = add tmp, tmp
3332 if (SextVal >= -32 && SextVal <= 30 && (SextVal & 1) == 0) {
3333 SDValue Res = BuildSplatI(SextVal >> 1, SplatSize, MVT::Other, DAG);
3334 Res = DAG.getNode(ISD::ADD, Res.getValueType(), Res, Res);
3335 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3338 // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is
3339 // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important
3341 if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) {
3342 // Make -1 and vspltisw -1:
3343 SDValue OnesV = BuildSplatI(-1, 4, MVT::v4i32, DAG);
3345 // Make the VSLW intrinsic, computing 0x8000_0000.
3346 SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV,
3349 // xor by OnesV to invert it.
3350 Res = DAG.getNode(ISD::XOR, MVT::v4i32, Res, OnesV);
3351 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3354 // Check to see if this is a wide variety of vsplti*, binop self cases.
3355 unsigned SplatBitSize = SplatSize*8;
3356 static const signed char SplatCsts[] = {
3357 -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
3358 -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
3361 for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
3362 // Indirect through the SplatCsts array so that we favor 'vsplti -1' for
3363 // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
3364 int i = SplatCsts[idx];
3366 // Figure out what shift amount will be used by altivec if shifted by i in
3368 unsigned TypeShiftAmt = i & (SplatBitSize-1);
3370 // vsplti + shl self.
3371 if (SextVal == (i << (int)TypeShiftAmt)) {
3372 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3373 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3374 Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
3375 Intrinsic::ppc_altivec_vslw
3377 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3378 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3381 // vsplti + srl self.
3382 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3383 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3384 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3385 Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
3386 Intrinsic::ppc_altivec_vsrw
3388 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3389 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3392 // vsplti + sra self.
3393 if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
3394 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3395 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3396 Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
3397 Intrinsic::ppc_altivec_vsraw
3399 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3400 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3403 // vsplti + rol self.
3404 if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
3405 ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
3406 SDValue Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
3407 static const unsigned IIDs[] = { // Intrinsic to use for each size.
3408 Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
3409 Intrinsic::ppc_altivec_vrlw
3411 Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
3412 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
3415 // t = vsplti c, result = vsldoi t, t, 1
3416 if (SextVal == ((i << 8) | (i >> (TypeShiftAmt-8)))) {
3417 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3418 return BuildVSLDOI(T, T, 1, Op.getValueType(), DAG);
3420 // t = vsplti c, result = vsldoi t, t, 2
3421 if (SextVal == ((i << 16) | (i >> (TypeShiftAmt-16)))) {
3422 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3423 return BuildVSLDOI(T, T, 2, Op.getValueType(), DAG);
3425 // t = vsplti c, result = vsldoi t, t, 3
3426 if (SextVal == ((i << 24) | (i >> (TypeShiftAmt-24)))) {
3427 SDValue T = BuildSplatI(i, SplatSize, MVT::v16i8, DAG);
3428 return BuildVSLDOI(T, T, 3, Op.getValueType(), DAG);
3432 // Three instruction sequences.
3434 // Odd, in range [17,31]: (vsplti C)-(vsplti -16).
3435 if (SextVal >= 0 && SextVal <= 31) {
3436 SDValue LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
3437 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
3438 LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
3439 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
3441 // Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
3442 if (SextVal >= -31 && SextVal <= 0) {
3443 SDValue LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
3444 SDValue RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
3445 LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
3446 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
3453 /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
3454 /// the specified operations to build the shuffle.
3455 static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
3456 SDValue RHS, SelectionDAG &DAG) {
3457 unsigned OpNum = (PFEntry >> 26) & 0x0F;
3458 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
3459 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
3462 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
3474 if (OpNum == OP_COPY) {
3475 if (LHSID == (1*9+2)*9+3) return LHS;
3476 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
3480 SDValue OpLHS, OpRHS;
3481 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG);
3482 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG);
3484 unsigned ShufIdxs[16];
3486 default: assert(0 && "Unknown i32 permute!");
3488 ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3;
3489 ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19;
3490 ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7;
3491 ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23;
3494 ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11;
3495 ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27;
3496 ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15;
3497 ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31;
3500 for (unsigned i = 0; i != 16; ++i)
3501 ShufIdxs[i] = (i&3)+0;
3504 for (unsigned i = 0; i != 16; ++i)
3505 ShufIdxs[i] = (i&3)+4;
3508 for (unsigned i = 0; i != 16; ++i)
3509 ShufIdxs[i] = (i&3)+8;
3512 for (unsigned i = 0; i != 16; ++i)
3513 ShufIdxs[i] = (i&3)+12;
3516 return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG);
3518 return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG);
3520 return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
3523 for (unsigned i = 0; i != 16; ++i)
3524 Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i8);
3526 return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
3527 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3530 /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
3531 /// is a shuffle we can handle in a single instruction, return it. Otherwise,
3532 /// return the code it can be lowered into. Worst case, it can always be
3533 /// lowered into a vperm.
3534 SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
3535 SelectionDAG &DAG) {
3536 SDValue V1 = Op.getOperand(0);
3537 SDValue V2 = Op.getOperand(1);
3538 SDValue PermMask = Op.getOperand(2);
3540 // Cases that are handled by instructions that take permute immediates
3541 // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
3542 // selected by the instruction selector.
3543 if (V2.getOpcode() == ISD::UNDEF) {
3544 if (PPC::isSplatShuffleMask(PermMask.getNode(), 1) ||
3545 PPC::isSplatShuffleMask(PermMask.getNode(), 2) ||
3546 PPC::isSplatShuffleMask(PermMask.getNode(), 4) ||
3547 PPC::isVPKUWUMShuffleMask(PermMask.getNode(), true) ||
3548 PPC::isVPKUHUMShuffleMask(PermMask.getNode(), true) ||
3549 PPC::isVSLDOIShuffleMask(PermMask.getNode(), true) != -1 ||
3550 PPC::isVMRGLShuffleMask(PermMask.getNode(), 1, true) ||
3551 PPC::isVMRGLShuffleMask(PermMask.getNode(), 2, true) ||
3552 PPC::isVMRGLShuffleMask(PermMask.getNode(), 4, true) ||
3553 PPC::isVMRGHShuffleMask(PermMask.getNode(), 1, true) ||
3554 PPC::isVMRGHShuffleMask(PermMask.getNode(), 2, true) ||
3555 PPC::isVMRGHShuffleMask(PermMask.getNode(), 4, true)) {
3560 // Altivec has a variety of "shuffle immediates" that take two vector inputs
3561 // and produce a fixed permutation. If any of these match, do not lower to
3563 if (PPC::isVPKUWUMShuffleMask(PermMask.getNode(), false) ||
3564 PPC::isVPKUHUMShuffleMask(PermMask.getNode(), false) ||
3565 PPC::isVSLDOIShuffleMask(PermMask.getNode(), false) != -1 ||
3566 PPC::isVMRGLShuffleMask(PermMask.getNode(), 1, false) ||
3567 PPC::isVMRGLShuffleMask(PermMask.getNode(), 2, false) ||
3568 PPC::isVMRGLShuffleMask(PermMask.getNode(), 4, false) ||
3569 PPC::isVMRGHShuffleMask(PermMask.getNode(), 1, false) ||
3570 PPC::isVMRGHShuffleMask(PermMask.getNode(), 2, false) ||
3571 PPC::isVMRGHShuffleMask(PermMask.getNode(), 4, false))
3574 // Check to see if this is a shuffle of 4-byte values. If so, we can use our
3575 // perfect shuffle table to emit an optimal matching sequence.
3576 unsigned PFIndexes[4];
3577 bool isFourElementShuffle = true;
3578 for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number
3579 unsigned EltNo = 8; // Start out undef.
3580 for (unsigned j = 0; j != 4; ++j) { // Intra-element byte.
3581 if (PermMask.getOperand(i*4+j).getOpcode() == ISD::UNDEF)
3582 continue; // Undef, ignore it.
3584 unsigned ByteSource =
3585 cast<ConstantSDNode>(PermMask.getOperand(i*4+j))->getZExtValue();
3586 if ((ByteSource & 3) != j) {
3587 isFourElementShuffle = false;
3592 EltNo = ByteSource/4;
3593 } else if (EltNo != ByteSource/4) {
3594 isFourElementShuffle = false;
3598 PFIndexes[i] = EltNo;
3601 // If this shuffle can be expressed as a shuffle of 4-byte elements, use the
3602 // perfect shuffle vector to determine if it is cost effective to do this as
3603 // discrete instructions, or whether we should use a vperm.
3604 if (isFourElementShuffle) {
3605 // Compute the index in the perfect shuffle table.
3606 unsigned PFTableIndex =
3607 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
3609 unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
3610 unsigned Cost = (PFEntry >> 30);
3612 // Determining when to avoid vperm is tricky. Many things affect the cost
3613 // of vperm, particularly how many times the perm mask needs to be computed.
3614 // For example, if the perm mask can be hoisted out of a loop or is already
3615 // used (perhaps because there are multiple permutes with the same shuffle
3616 // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of
3617 // the loop requires an extra register.
3619 // As a compromise, we only emit discrete instructions if the shuffle can be
3620 // generated in 3 or fewer operations. When we have loop information
3621 // available, if this block is within a loop, we should avoid using vperm
3622 // for 3-operation perms and use a constant pool load instead.
3624 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG);
3627 // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
3628 // vector that will get spilled to the constant pool.
3629 if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
3631 // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
3632 // that it is in input element units, not in bytes. Convert now.
3633 MVT EltVT = V1.getValueType().getVectorElementType();
3634 unsigned BytesPerElement = EltVT.getSizeInBits()/8;
3636 SmallVector<SDValue, 16> ResultMask;
3637 for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
3639 if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
3642 SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getZExtValue();
3644 for (unsigned j = 0; j != BytesPerElement; ++j)
3645 ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
3649 SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
3650 &ResultMask[0], ResultMask.size());
3651 return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
3654 /// getAltivecCompareInfo - Given an intrinsic, return false if it is not an
3655 /// altivec comparison. If it is, return true and fill in Opc/isDot with
3656 /// information about the intrinsic.
3657 static bool getAltivecCompareInfo(SDValue Intrin, int &CompareOpc,
3659 unsigned IntrinsicID =
3660 cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue();
3663 switch (IntrinsicID) {
3664 default: return false;
3665 // Comparison predicates.
3666 case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
3667 case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
3668 case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
3669 case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
3670 case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
3671 case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
3672 case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
3673 case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
3674 case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
3675 case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
3676 case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
3677 case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
3678 case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
3680 // Normal Comparisons.
3681 case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
3682 case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
3683 case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
3684 case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
3685 case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
3686 case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
3687 case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
3688 case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
3689 case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
3690 case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
3691 case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
3692 case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
3693 case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
3698 /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom
3699 /// lower, do it, otherwise return null.
3700 SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
3701 SelectionDAG &DAG) {
3702 // If this is a lowered altivec predicate compare, CompareOpc is set to the
3703 // opcode number of the comparison.
3706 if (!getAltivecCompareInfo(Op, CompareOpc, isDot))
3707 return SDValue(); // Don't custom lower most intrinsics.
3709 // If this is a non-dot comparison, make the VCMP node and we are done.
3711 SDValue Tmp = DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
3712 Op.getOperand(1), Op.getOperand(2),
3713 DAG.getConstant(CompareOpc, MVT::i32));
3714 return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Tmp);
3717 // Create the PPCISD altivec 'dot' comparison node.
3719 Op.getOperand(2), // LHS
3720 Op.getOperand(3), // RHS
3721 DAG.getConstant(CompareOpc, MVT::i32)
3723 std::vector<MVT> VTs;
3724 VTs.push_back(Op.getOperand(2).getValueType());
3725 VTs.push_back(MVT::Flag);
3726 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
3728 // Now that we have the comparison, emit a copy from the CR to a GPR.
3729 // This is flagged to the above dot comparison.
3730 SDValue Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
3731 DAG.getRegister(PPC::CR6, MVT::i32),
3732 CompNode.getValue(1));
3734 // Unpack the result based on how the target uses it.
3735 unsigned BitNo; // Bit # of CR6.
3736 bool InvertBit; // Invert result?
3737 switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) {
3738 default: // Can't happen, don't crash on invalid number though.
3739 case 0: // Return the value of the EQ bit of CR6.
3740 BitNo = 0; InvertBit = false;
3742 case 1: // Return the inverted value of the EQ bit of CR6.
3743 BitNo = 0; InvertBit = true;
3745 case 2: // Return the value of the LT bit of CR6.
3746 BitNo = 2; InvertBit = false;
3748 case 3: // Return the inverted value of the LT bit of CR6.
3749 BitNo = 2; InvertBit = true;
3753 // Shift the bit into the low position.
3754 Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
3755 DAG.getConstant(8-(3-BitNo), MVT::i32));
3757 Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
3758 DAG.getConstant(1, MVT::i32));
3760 // If we are supposed to, toggle the bit.
3762 Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
3763 DAG.getConstant(1, MVT::i32));
3767 SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
3768 SelectionDAG &DAG) {
3769 // Create a stack slot that is 16-byte aligned.
3770 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
3771 int FrameIdx = FrameInfo->CreateStackObject(16, 16);
3772 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
3773 SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
3775 // Store the input value into Value#0 of the stack slot.
3776 SDValue Store = DAG.getStore(DAG.getEntryNode(),
3777 Op.getOperand(0), FIdx, NULL, 0);
3779 return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
3782 SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) {
3783 if (Op.getValueType() == MVT::v4i32) {
3784 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3786 SDValue Zero = BuildSplatI( 0, 1, MVT::v4i32, DAG);
3787 SDValue Neg16 = BuildSplatI(-16, 4, MVT::v4i32, DAG); // +16 as shift amt.
3789 SDValue RHSSwap = // = vrlw RHS, 16
3790 BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG);
3792 // Shrinkify inputs to v8i16.
3793 LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, LHS);
3794 RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHS);
3795 RHSSwap = DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, RHSSwap);
3797 // Low parts multiplied together, generating 32-bit results (we ignore the
3799 SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh,
3800 LHS, RHS, DAG, MVT::v4i32);
3802 SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm,
3803 LHS, RHSSwap, Zero, DAG, MVT::v4i32);
3804 // Shift the high parts up 16 bits.
3805 HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, Neg16, DAG);
3806 return DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd);
3807 } else if (Op.getValueType() == MVT::v8i16) {
3808 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3810 SDValue Zero = BuildSplatI(0, 1, MVT::v8i16, DAG);
3812 return BuildIntrinsicOp(Intrinsic::ppc_altivec_vmladduhm,
3813 LHS, RHS, Zero, DAG);
3814 } else if (Op.getValueType() == MVT::v16i8) {
3815 SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1);
3817 // Multiply the even 8-bit parts, producing 16-bit sums.
3818 SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub,
3819 LHS, RHS, DAG, MVT::v8i16);
3820 EvenParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, EvenParts);
3822 // Multiply the odd 8-bit parts, producing 16-bit sums.
3823 SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub,
3824 LHS, RHS, DAG, MVT::v8i16);
3825 OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
3827 // Merge the results together.
3829 for (unsigned i = 0; i != 8; ++i) {
3830 Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
3831 Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
3833 return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
3834 DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
3836 assert(0 && "Unknown mul to lower!");
3841 /// LowerOperation - Provide custom lowering hooks for some operations.
3843 SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
3844 switch (Op.getOpcode()) {
3845 default: assert(0 && "Wasn't expecting to be able to lower this!");
3846 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3847 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
3848 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3849 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3850 case ISD::SETCC: return LowerSETCC(Op, DAG);
3851 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
3853 return LowerVASTART(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3854 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3857 return LowerVAARG(Op, DAG, VarArgsFrameIndex, VarArgsStackOffset,
3858 VarArgsNumGPR, VarArgsNumFPR, PPCSubTarget);
3860 case ISD::FORMAL_ARGUMENTS:
3861 return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex,
3862 VarArgsStackOffset, VarArgsNumGPR,
3863 VarArgsNumFPR, PPCSubTarget);
3865 case ISD::CALL: return LowerCALL(Op, DAG, PPCSubTarget,
3866 getTargetMachine());
3867 case ISD::RET: return LowerRET(Op, DAG, getTargetMachine());
3868 case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
3869 case ISD::DYNAMIC_STACKALLOC:
3870 return LowerDYNAMIC_STACKALLOC(Op, DAG, PPCSubTarget);
3872 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3873 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
3874 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
3875 case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
3876 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
3878 // Lower 64-bit shifts.
3879 case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
3880 case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
3881 case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
3883 // Vector-related lowering.
3884 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
3885 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3886 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
3887 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
3888 case ISD::MUL: return LowerMUL(Op, DAG);
3890 // Frame & Return address.
3891 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3892 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3897 SDNode *PPCTargetLowering::ReplaceNodeResults(SDNode *N, SelectionDAG &DAG) {
3898 switch (N->getOpcode()) {
3899 default: assert(0 && "Wasn't expecting to be able to lower this!");
3900 case ISD::FP_TO_SINT: {
3901 SDValue Res = LowerFP_TO_SINT(SDValue(N, 0), DAG);
3902 // Use MERGE_VALUES to drop the chain result value and get a node with one
3903 // result. This requires turning off getMergeValues simplification, since
3904 // otherwise it will give us Res back.
3905 return DAG.getMergeValues(&Res, 1, false).getNode();
3911 //===----------------------------------------------------------------------===//
3912 // Other Lowering Code
3913 //===----------------------------------------------------------------------===//
3916 PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
3917 bool is64bit, unsigned BinOpcode) {
3918 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3919 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3921 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3922 MachineFunction *F = BB->getParent();
3923 MachineFunction::iterator It = BB;
3926 unsigned dest = MI->getOperand(0).getReg();
3927 unsigned ptrA = MI->getOperand(1).getReg();
3928 unsigned ptrB = MI->getOperand(2).getReg();
3929 unsigned incr = MI->getOperand(3).getReg();
3931 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3932 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3933 F->insert(It, loopMBB);
3934 F->insert(It, exitMBB);
3935 exitMBB->transferSuccessors(BB);
3937 MachineRegisterInfo &RegInfo = F->getRegInfo();
3938 unsigned TmpReg = (!BinOpcode) ? incr :
3939 RegInfo.createVirtualRegister(
3940 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
3941 (const TargetRegisterClass *) &PPC::GPRCRegClass);
3945 // fallthrough --> loopMBB
3946 BB->addSuccessor(loopMBB);
3949 // l[wd]arx dest, ptr
3950 // add r0, dest, incr
3951 // st[wd]cx. r0, ptr
3953 // fallthrough --> exitMBB
3955 BuildMI(BB, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
3956 .addReg(ptrA).addReg(ptrB);
3958 BuildMI(BB, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest);
3959 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
3960 .addReg(TmpReg).addReg(ptrA).addReg(ptrB);
3961 BuildMI(BB, TII->get(PPC::BCC))
3962 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
3963 BB->addSuccessor(loopMBB);
3964 BB->addSuccessor(exitMBB);
3973 PPCTargetLowering::EmitPartwordAtomicBinary(MachineInstr *MI,
3974 MachineBasicBlock *BB,
3975 bool is8bit, // operation
3976 unsigned BinOpcode) {
3977 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
3978 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
3979 // In 64 bit mode we have to use 64 bits for addresses, even though the
3980 // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
3981 // registers without caring whether they're 32 or 64, but here we're
3982 // doing actual arithmetic on the addresses.
3983 bool is64bit = PPCSubTarget.isPPC64();
3985 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3986 MachineFunction *F = BB->getParent();
3987 MachineFunction::iterator It = BB;
3990 unsigned dest = MI->getOperand(0).getReg();
3991 unsigned ptrA = MI->getOperand(1).getReg();
3992 unsigned ptrB = MI->getOperand(2).getReg();
3993 unsigned incr = MI->getOperand(3).getReg();
3995 MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
3996 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
3997 F->insert(It, loopMBB);
3998 F->insert(It, exitMBB);
3999 exitMBB->transferSuccessors(BB);
4001 MachineRegisterInfo &RegInfo = F->getRegInfo();
4002 const TargetRegisterClass *RC =
4003 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4004 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4005 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4006 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4007 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4008 unsigned Incr2Reg = RegInfo.createVirtualRegister(RC);
4009 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4010 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4011 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4012 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4013 unsigned Tmp3Reg = RegInfo.createVirtualRegister(RC);
4014 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4015 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4017 unsigned TmpReg = (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(RC);
4021 // fallthrough --> loopMBB
4022 BB->addSuccessor(loopMBB);
4024 // The 4-byte load must be aligned, while a char or short may be
4025 // anywhere in the word. Hence all this nasty bookkeeping code.
4026 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4027 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4028 // xori shift, shift1, 24 [16]
4029 // rlwinm ptr, ptr1, 0, 0, 29
4030 // slw incr2, incr, shift
4031 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4032 // slw mask, mask2, shift
4034 // lwarx tmpDest, ptr
4035 // add tmp, tmpDest, incr2
4036 // andc tmp2, tmpDest, mask
4037 // and tmp3, tmp, mask
4038 // or tmp4, tmp3, tmp2
4041 // fallthrough --> exitMBB
4042 // srw dest, tmpDest, shift
4044 if (ptrA!=PPC::R0) {
4045 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4046 BuildMI(BB, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4047 .addReg(ptrA).addReg(ptrB);
4051 BuildMI(BB, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4052 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4053 BuildMI(BB, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4054 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4056 BuildMI(BB, TII->get(PPC::RLDICR), PtrReg)
4057 .addReg(Ptr1Reg).addImm(0).addImm(61);
4059 BuildMI(BB, TII->get(PPC::RLWINM), PtrReg)
4060 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4061 BuildMI(BB, TII->get(PPC::SLW), Incr2Reg)
4062 .addReg(incr).addReg(ShiftReg);
4064 BuildMI(BB, TII->get(PPC::LI), Mask2Reg).addImm(255);
4066 BuildMI(BB, TII->get(PPC::LI), Mask3Reg).addImm(0);
4067 BuildMI(BB, TII->get(PPC::ORI), Mask2Reg).addReg(Mask3Reg).addImm(65535);
4069 BuildMI(BB, TII->get(PPC::SLW), MaskReg)
4070 .addReg(Mask2Reg).addReg(ShiftReg);
4073 BuildMI(BB, TII->get(PPC::LWARX), TmpDestReg)
4074 .addReg(PPC::R0).addReg(PtrReg);
4076 BuildMI(BB, TII->get(BinOpcode), TmpReg)
4077 .addReg(Incr2Reg).addReg(TmpDestReg);
4078 BuildMI(BB, TII->get(is64bit ? PPC::ANDC8 : PPC::ANDC), Tmp2Reg)
4079 .addReg(TmpDestReg).addReg(MaskReg);
4080 BuildMI(BB, TII->get(is64bit ? PPC::AND8 : PPC::AND), Tmp3Reg)
4081 .addReg(TmpReg).addReg(MaskReg);
4082 BuildMI(BB, TII->get(is64bit ? PPC::OR8 : PPC::OR), Tmp4Reg)
4083 .addReg(Tmp3Reg).addReg(Tmp2Reg);
4084 BuildMI(BB, TII->get(PPC::STWCX))
4085 .addReg(Tmp4Reg).addReg(PPC::R0).addReg(PtrReg);
4086 BuildMI(BB, TII->get(PPC::BCC))
4087 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB);
4088 BB->addSuccessor(loopMBB);
4089 BB->addSuccessor(exitMBB);
4094 BuildMI(BB, TII->get(PPC::SRW), dest).addReg(TmpDestReg).addReg(ShiftReg);
4099 PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4100 MachineBasicBlock *BB) {
4101 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
4103 // To "insert" these instructions we actually have to insert their
4104 // control-flow patterns.
4105 const BasicBlock *LLVM_BB = BB->getBasicBlock();
4106 MachineFunction::iterator It = BB;
4109 MachineFunction *F = BB->getParent();
4111 if (MI->getOpcode() == PPC::SELECT_CC_I4 ||
4112 MI->getOpcode() == PPC::SELECT_CC_I8 ||
4113 MI->getOpcode() == PPC::SELECT_CC_F4 ||
4114 MI->getOpcode() == PPC::SELECT_CC_F8 ||
4115 MI->getOpcode() == PPC::SELECT_CC_VRRC) {
4117 // The incoming instruction knows the destination vreg to set, the
4118 // condition code register to branch on, the true/false values to
4119 // select between, and a branch opcode to use.
4124 // cmpTY ccX, r1, r2
4126 // fallthrough --> copy0MBB
4127 MachineBasicBlock *thisMBB = BB;
4128 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
4129 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
4130 unsigned SelectPred = MI->getOperand(4).getImm();
4131 BuildMI(BB, TII->get(PPC::BCC))
4132 .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
4133 F->insert(It, copy0MBB);
4134 F->insert(It, sinkMBB);
4135 // Update machine-CFG edges by transferring all successors of the current
4136 // block to the new block which will contain the Phi node for the select.
4137 sinkMBB->transferSuccessors(BB);
4138 // Next, add the true and fallthrough blocks as its successors.
4139 BB->addSuccessor(copy0MBB);
4140 BB->addSuccessor(sinkMBB);
4143 // %FalseValue = ...
4144 // # fallthrough to sinkMBB
4147 // Update machine-CFG edges
4148 BB->addSuccessor(sinkMBB);
4151 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
4154 BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
4155 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
4156 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
4158 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I8)
4159 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4);
4160 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I16)
4161 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4);
4162 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I32)
4163 BB = EmitAtomicBinary(MI, BB, false, PPC::ADD4);
4164 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_ADD_I64)
4165 BB = EmitAtomicBinary(MI, BB, true, PPC::ADD8);
4167 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I8)
4168 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND);
4169 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I16)
4170 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND);
4171 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I32)
4172 BB = EmitAtomicBinary(MI, BB, false, PPC::AND);
4173 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_AND_I64)
4174 BB = EmitAtomicBinary(MI, BB, true, PPC::AND8);
4176 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I8)
4177 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR);
4178 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I16)
4179 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR);
4180 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I32)
4181 BB = EmitAtomicBinary(MI, BB, false, PPC::OR);
4182 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_OR_I64)
4183 BB = EmitAtomicBinary(MI, BB, true, PPC::OR8);
4185 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I8)
4186 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR);
4187 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I16)
4188 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR);
4189 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I32)
4190 BB = EmitAtomicBinary(MI, BB, false, PPC::XOR);
4191 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_XOR_I64)
4192 BB = EmitAtomicBinary(MI, BB, true, PPC::XOR8);
4194 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I8)
4195 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ANDC);
4196 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I16)
4197 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ANDC);
4198 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I32)
4199 BB = EmitAtomicBinary(MI, BB, false, PPC::ANDC);
4200 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_NAND_I64)
4201 BB = EmitAtomicBinary(MI, BB, true, PPC::ANDC8);
4203 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I8)
4204 BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF);
4205 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I16)
4206 BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF);
4207 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I32)
4208 BB = EmitAtomicBinary(MI, BB, false, PPC::SUBF);
4209 else if (MI->getOpcode() == PPC::ATOMIC_LOAD_SUB_I64)
4210 BB = EmitAtomicBinary(MI, BB, true, PPC::SUBF8);
4212 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I8)
4213 BB = EmitPartwordAtomicBinary(MI, BB, true, 0);
4214 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I16)
4215 BB = EmitPartwordAtomicBinary(MI, BB, false, 0);
4216 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I32)
4217 BB = EmitAtomicBinary(MI, BB, false, 0);
4218 else if (MI->getOpcode() == PPC::ATOMIC_SWAP_I64)
4219 BB = EmitAtomicBinary(MI, BB, true, 0);
4221 else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 ||
4222 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64) {
4223 bool is64bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I64;
4225 unsigned dest = MI->getOperand(0).getReg();
4226 unsigned ptrA = MI->getOperand(1).getReg();
4227 unsigned ptrB = MI->getOperand(2).getReg();
4228 unsigned oldval = MI->getOperand(3).getReg();
4229 unsigned newval = MI->getOperand(4).getReg();
4231 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4232 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4233 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4234 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4235 F->insert(It, loop1MBB);
4236 F->insert(It, loop2MBB);
4237 F->insert(It, midMBB);
4238 F->insert(It, exitMBB);
4239 exitMBB->transferSuccessors(BB);
4243 // fallthrough --> loopMBB
4244 BB->addSuccessor(loop1MBB);
4247 // l[wd]arx dest, ptr
4248 // cmp[wd] dest, oldval
4251 // st[wd]cx. newval, ptr
4255 // st[wd]cx. dest, ptr
4258 BuildMI(BB, TII->get(is64bit ? PPC::LDARX : PPC::LWARX), dest)
4259 .addReg(ptrA).addReg(ptrB);
4260 BuildMI(BB, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0)
4261 .addReg(oldval).addReg(dest);
4262 BuildMI(BB, TII->get(PPC::BCC))
4263 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4264 BB->addSuccessor(loop2MBB);
4265 BB->addSuccessor(midMBB);
4268 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4269 .addReg(newval).addReg(ptrA).addReg(ptrB);
4270 BuildMI(BB, TII->get(PPC::BCC))
4271 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4272 BuildMI(BB, TII->get(PPC::B)).addMBB(exitMBB);
4273 BB->addSuccessor(loop1MBB);
4274 BB->addSuccessor(exitMBB);
4277 BuildMI(BB, TII->get(is64bit ? PPC::STDCX : PPC::STWCX))
4278 .addReg(dest).addReg(ptrA).addReg(ptrB);
4279 BB->addSuccessor(exitMBB);
4284 } else if (MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 ||
4285 MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) {
4286 // We must use 64-bit registers for addresses when targeting 64-bit,
4287 // since we're actually doing arithmetic on them. Other registers
4289 bool is64bit = PPCSubTarget.isPPC64();
4290 bool is8bit = MI->getOpcode() == PPC::ATOMIC_CMP_SWAP_I8;
4292 unsigned dest = MI->getOperand(0).getReg();
4293 unsigned ptrA = MI->getOperand(1).getReg();
4294 unsigned ptrB = MI->getOperand(2).getReg();
4295 unsigned oldval = MI->getOperand(3).getReg();
4296 unsigned newval = MI->getOperand(4).getReg();
4298 MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB);
4299 MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB);
4300 MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB);
4301 MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
4302 F->insert(It, loop1MBB);
4303 F->insert(It, loop2MBB);
4304 F->insert(It, midMBB);
4305 F->insert(It, exitMBB);
4306 exitMBB->transferSuccessors(BB);
4308 MachineRegisterInfo &RegInfo = F->getRegInfo();
4309 const TargetRegisterClass *RC =
4310 is64bit ? (const TargetRegisterClass *) &PPC::G8RCRegClass :
4311 (const TargetRegisterClass *) &PPC::GPRCRegClass;
4312 unsigned PtrReg = RegInfo.createVirtualRegister(RC);
4313 unsigned Shift1Reg = RegInfo.createVirtualRegister(RC);
4314 unsigned ShiftReg = RegInfo.createVirtualRegister(RC);
4315 unsigned NewVal2Reg = RegInfo.createVirtualRegister(RC);
4316 unsigned NewVal3Reg = RegInfo.createVirtualRegister(RC);
4317 unsigned OldVal2Reg = RegInfo.createVirtualRegister(RC);
4318 unsigned OldVal3Reg = RegInfo.createVirtualRegister(RC);
4319 unsigned MaskReg = RegInfo.createVirtualRegister(RC);
4320 unsigned Mask2Reg = RegInfo.createVirtualRegister(RC);
4321 unsigned Mask3Reg = RegInfo.createVirtualRegister(RC);
4322 unsigned Tmp2Reg = RegInfo.createVirtualRegister(RC);
4323 unsigned Tmp4Reg = RegInfo.createVirtualRegister(RC);
4324 unsigned TmpDestReg = RegInfo.createVirtualRegister(RC);
4326 unsigned TmpReg = RegInfo.createVirtualRegister(RC);
4329 // fallthrough --> loopMBB
4330 BB->addSuccessor(loop1MBB);
4332 // The 4-byte load must be aligned, while a char or short may be
4333 // anywhere in the word. Hence all this nasty bookkeeping code.
4334 // add ptr1, ptrA, ptrB [copy if ptrA==0]
4335 // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27]
4336 // xori shift, shift1, 24 [16]
4337 // rlwinm ptr, ptr1, 0, 0, 29
4338 // slw newval2, newval, shift
4339 // slw oldval2, oldval,shift
4340 // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535]
4341 // slw mask, mask2, shift
4342 // and newval3, newval2, mask
4343 // and oldval3, oldval2, mask
4345 // lwarx tmpDest, ptr
4346 // and tmp, tmpDest, mask
4347 // cmpw tmp, oldval3
4350 // andc tmp2, tmpDest, mask
4351 // or tmp4, tmp2, newval3
4356 // stwcx. tmpDest, ptr
4358 // srw dest, tmpDest, shift
4359 if (ptrA!=PPC::R0) {
4360 Ptr1Reg = RegInfo.createVirtualRegister(RC);
4361 BuildMI(BB, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg)
4362 .addReg(ptrA).addReg(ptrB);
4366 BuildMI(BB, TII->get(PPC::RLWINM), Shift1Reg).addReg(Ptr1Reg)
4367 .addImm(3).addImm(27).addImm(is8bit ? 28 : 27);
4368 BuildMI(BB, TII->get(is64bit ? PPC::XORI8 : PPC::XORI), ShiftReg)
4369 .addReg(Shift1Reg).addImm(is8bit ? 24 : 16);
4371 BuildMI(BB, TII->get(PPC::RLDICR), PtrReg)
4372 .addReg(Ptr1Reg).addImm(0).addImm(61);
4374 BuildMI(BB, TII->get(PPC::RLWINM), PtrReg)
4375 .addReg(Ptr1Reg).addImm(0).addImm(0).addImm(29);
4376 BuildMI(BB, TII->get(PPC::SLW), NewVal2Reg)
4377 .addReg(newval).addReg(ShiftReg);
4378 BuildMI(BB, TII->get(PPC::SLW), OldVal2Reg)
4379 .addReg(oldval).addReg(ShiftReg);
4381 BuildMI(BB, TII->get(PPC::LI), Mask2Reg).addImm(255);
4383 BuildMI(BB, TII->get(PPC::LI), Mask3Reg).addImm(0);
4384 BuildMI(BB, TII->get(PPC::ORI), Mask2Reg).addReg(Mask3Reg).addImm(65535);
4386 BuildMI(BB, TII->get(PPC::SLW), MaskReg)
4387 .addReg(Mask2Reg).addReg(ShiftReg);
4388 BuildMI(BB, TII->get(PPC::AND), NewVal3Reg)
4389 .addReg(NewVal2Reg).addReg(MaskReg);
4390 BuildMI(BB, TII->get(PPC::AND), OldVal3Reg)
4391 .addReg(OldVal2Reg).addReg(MaskReg);
4394 BuildMI(BB, TII->get(PPC::LWARX), TmpDestReg)
4395 .addReg(PPC::R0).addReg(PtrReg);
4396 BuildMI(BB, TII->get(PPC::AND),TmpReg).addReg(TmpDestReg).addReg(MaskReg);
4397 BuildMI(BB, TII->get(PPC::CMPW), PPC::CR0)
4398 .addReg(TmpReg).addReg(OldVal3Reg);
4399 BuildMI(BB, TII->get(PPC::BCC))
4400 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(midMBB);
4401 BB->addSuccessor(loop2MBB);
4402 BB->addSuccessor(midMBB);
4405 BuildMI(BB, TII->get(PPC::ANDC),Tmp2Reg).addReg(TmpDestReg).addReg(MaskReg);
4406 BuildMI(BB, TII->get(PPC::OR),Tmp4Reg).addReg(Tmp2Reg).addReg(NewVal3Reg);
4407 BuildMI(BB, TII->get(PPC::STWCX)).addReg(Tmp4Reg)
4408 .addReg(PPC::R0).addReg(PtrReg);
4409 BuildMI(BB, TII->get(PPC::BCC))
4410 .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loop1MBB);
4411 BuildMI(BB, TII->get(PPC::B)).addMBB(exitMBB);
4412 BB->addSuccessor(loop1MBB);
4413 BB->addSuccessor(exitMBB);
4416 BuildMI(BB, TII->get(PPC::STWCX)).addReg(TmpDestReg)
4417 .addReg(PPC::R0).addReg(PtrReg);
4418 BB->addSuccessor(exitMBB);
4423 BuildMI(BB, TII->get(PPC::SRW),dest).addReg(TmpReg).addReg(ShiftReg);
4425 assert(0 && "Unexpected instr type to insert");
4428 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
4432 //===----------------------------------------------------------------------===//
4433 // Target Optimization Hooks
4434 //===----------------------------------------------------------------------===//
4436 SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N,
4437 DAGCombinerInfo &DCI) const {
4438 TargetMachine &TM = getTargetMachine();
4439 SelectionDAG &DAG = DCI.DAG;
4440 switch (N->getOpcode()) {
4443 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4444 if (C->getZExtValue() == 0) // 0 << V -> 0.
4445 return N->getOperand(0);
4449 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4450 if (C->getZExtValue() == 0) // 0 >>u V -> 0.
4451 return N->getOperand(0);
4455 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
4456 if (C->getZExtValue() == 0 || // 0 >>s V -> 0.
4457 C->isAllOnesValue()) // -1 >>s V -> -1.
4458 return N->getOperand(0);
4462 case ISD::SINT_TO_FP:
4463 if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
4464 if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
4465 // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
4466 // We allow the src/dst to be either f32/f64, but the intermediate
4467 // type must be i64.
4468 if (N->getOperand(0).getValueType() == MVT::i64 &&
4469 N->getOperand(0).getOperand(0).getValueType() != MVT::ppcf128) {
4470 SDValue Val = N->getOperand(0).getOperand(0);
4471 if (Val.getValueType() == MVT::f32) {
4472 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
4473 DCI.AddToWorklist(Val.getNode());
4476 Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
4477 DCI.AddToWorklist(Val.getNode());
4478 Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
4479 DCI.AddToWorklist(Val.getNode());
4480 if (N->getValueType(0) == MVT::f32) {
4481 Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val,
4482 DAG.getIntPtrConstant(0));
4483 DCI.AddToWorklist(Val.getNode());
4486 } else if (N->getOperand(0).getValueType() == MVT::i32) {
4487 // If the intermediate type is i32, we can avoid the load/store here
4494 // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
4495 if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
4496 !cast<StoreSDNode>(N)->isTruncatingStore() &&
4497 N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
4498 N->getOperand(1).getValueType() == MVT::i32 &&
4499 N->getOperand(1).getOperand(0).getValueType() != MVT::ppcf128) {
4500 SDValue Val = N->getOperand(1).getOperand(0);
4501 if (Val.getValueType() == MVT::f32) {
4502 Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
4503 DCI.AddToWorklist(Val.getNode());
4505 Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
4506 DCI.AddToWorklist(Val.getNode());
4508 Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
4509 N->getOperand(2), N->getOperand(3));
4510 DCI.AddToWorklist(Val.getNode());
4514 // Turn STORE (BSWAP) -> sthbrx/stwbrx.
4515 if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
4516 N->getOperand(1).getNode()->hasOneUse() &&
4517 (N->getOperand(1).getValueType() == MVT::i32 ||
4518 N->getOperand(1).getValueType() == MVT::i16)) {
4519 SDValue BSwapOp = N->getOperand(1).getOperand(0);
4520 // Do an any-extend to 32-bits if this is a half-word input.
4521 if (BSwapOp.getValueType() == MVT::i16)
4522 BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
4524 return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
4525 N->getOperand(2), N->getOperand(3),
4526 DAG.getValueType(N->getOperand(1).getValueType()));
4530 // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
4531 if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&
4532 N->getOperand(0).hasOneUse() &&
4533 (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
4534 SDValue Load = N->getOperand(0);
4535 LoadSDNode *LD = cast<LoadSDNode>(Load);
4536 // Create the byte-swapping load.
4537 std::vector<MVT> VTs;
4538 VTs.push_back(MVT::i32);
4539 VTs.push_back(MVT::Other);
4540 SDValue MO = DAG.getMemOperand(LD->getMemOperand());
4542 LD->getChain(), // Chain
4543 LD->getBasePtr(), // Ptr
4545 DAG.getValueType(N->getValueType(0)) // VT
4547 SDValue BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
4549 // If this is an i16 load, insert the truncate.
4550 SDValue ResVal = BSLoad;
4551 if (N->getValueType(0) == MVT::i16)
4552 ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
4554 // First, combine the bswap away. This makes the value produced by the
4556 DCI.CombineTo(N, ResVal);
4558 // Next, combine the load away, we give it a bogus result value but a real
4559 // chain result. The result value is dead because the bswap is dead.
4560 DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1));
4562 // Return N so it doesn't get rechecked!
4563 return SDValue(N, 0);
4567 case PPCISD::VCMP: {
4568 // If a VCMPo node already exists with exactly the same operands as this
4569 // node, use its result instead of this node (VCMPo computes both a CR6 and
4570 // a normal output).
4572 if (!N->getOperand(0).hasOneUse() &&
4573 !N->getOperand(1).hasOneUse() &&
4574 !N->getOperand(2).hasOneUse()) {
4576 // Scan all of the users of the LHS, looking for VCMPo's that match.
4577 SDNode *VCMPoNode = 0;
4579 SDNode *LHSN = N->getOperand(0).getNode();
4580 for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
4582 if (UI->getOpcode() == PPCISD::VCMPo &&
4583 UI->getOperand(1) == N->getOperand(1) &&
4584 UI->getOperand(2) == N->getOperand(2) &&
4585 UI->getOperand(0) == N->getOperand(0)) {
4590 // If there is no VCMPo node, or if the flag value has a single use, don't
4592 if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1))
4595 // Look at the (necessarily single) use of the flag value. If it has a
4596 // chain, this transformation is more complex. Note that multiple things
4597 // could use the value result, which we should ignore.
4598 SDNode *FlagUser = 0;
4599 for (SDNode::use_iterator UI = VCMPoNode->use_begin();
4600 FlagUser == 0; ++UI) {
4601 assert(UI != VCMPoNode->use_end() && "Didn't find user!");
4603 for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) {
4604 if (User->getOperand(i) == SDValue(VCMPoNode, 1)) {
4611 // If the user is a MFCR instruction, we know this is safe. Otherwise we
4612 // give up for right now.
4613 if (FlagUser->getOpcode() == PPCISD::MFCR)
4614 return SDValue(VCMPoNode, 0);
4619 // If this is a branch on an altivec predicate comparison, lower this so
4620 // that we don't have to do a MFCR: instead, branch directly on CR6. This
4621 // lowering is done pre-legalize, because the legalizer lowers the predicate
4622 // compare down to code that is difficult to reassemble.
4623 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
4624 SDValue LHS = N->getOperand(2), RHS = N->getOperand(3);
4628 if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
4629 isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) &&
4630 getAltivecCompareInfo(LHS, CompareOpc, isDot)) {
4631 assert(isDot && "Can't compare against a vector result!");
4633 // If this is a comparison against something other than 0/1, then we know
4634 // that the condition is never/always true.
4635 unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue();
4636 if (Val != 0 && Val != 1) {
4637 if (CC == ISD::SETEQ) // Cond never true, remove branch.
4638 return N->getOperand(0);
4639 // Always !=, turn it into an unconditional branch.
4640 return DAG.getNode(ISD::BR, MVT::Other,
4641 N->getOperand(0), N->getOperand(4));
4644 bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
4646 // Create the PPCISD altivec 'dot' comparison node.
4647 std::vector<MVT> VTs;
4649 LHS.getOperand(2), // LHS of compare
4650 LHS.getOperand(3), // RHS of compare
4651 DAG.getConstant(CompareOpc, MVT::i32)
4653 VTs.push_back(LHS.getOperand(2).getValueType());
4654 VTs.push_back(MVT::Flag);
4655 SDValue CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
4657 // Unpack the result based on how the target uses it.
4658 PPC::Predicate CompOpc;
4659 switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) {
4660 default: // Can't happen, don't crash on invalid number though.
4661 case 0: // Branch on the value of the EQ bit of CR6.
4662 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
4664 case 1: // Branch on the inverted value of the EQ bit of CR6.
4665 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
4667 case 2: // Branch on the value of the LT bit of CR6.
4668 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
4670 case 3: // Branch on the inverted value of the LT bit of CR6.
4671 CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
4675 return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
4676 DAG.getConstant(CompOpc, MVT::i32),
4677 DAG.getRegister(PPC::CR6, MVT::i32),
4678 N->getOperand(4), CompNode.getValue(1));
4687 //===----------------------------------------------------------------------===//
4688 // Inline Assembly Support
4689 //===----------------------------------------------------------------------===//
4691 void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
4695 const SelectionDAG &DAG,
4696 unsigned Depth) const {
4697 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
4698 switch (Op.getOpcode()) {
4700 case PPCISD::LBRX: {
4701 // lhbrx is known to have the top bits cleared out.
4702 if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
4703 KnownZero = 0xFFFF0000;
4706 case ISD::INTRINSIC_WO_CHAIN: {
4707 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
4709 case Intrinsic::ppc_altivec_vcmpbfp_p:
4710 case Intrinsic::ppc_altivec_vcmpeqfp_p:
4711 case Intrinsic::ppc_altivec_vcmpequb_p:
4712 case Intrinsic::ppc_altivec_vcmpequh_p:
4713 case Intrinsic::ppc_altivec_vcmpequw_p:
4714 case Intrinsic::ppc_altivec_vcmpgefp_p:
4715 case Intrinsic::ppc_altivec_vcmpgtfp_p:
4716 case Intrinsic::ppc_altivec_vcmpgtsb_p:
4717 case Intrinsic::ppc_altivec_vcmpgtsh_p:
4718 case Intrinsic::ppc_altivec_vcmpgtsw_p:
4719 case Intrinsic::ppc_altivec_vcmpgtub_p:
4720 case Intrinsic::ppc_altivec_vcmpgtuh_p:
4721 case Intrinsic::ppc_altivec_vcmpgtuw_p:
4722 KnownZero = ~1U; // All bits but the low one are known to be zero.
4730 /// getConstraintType - Given a constraint, return the type of
4731 /// constraint it is for this target.
4732 PPCTargetLowering::ConstraintType
4733 PPCTargetLowering::getConstraintType(const std::string &Constraint) const {
4734 if (Constraint.size() == 1) {
4735 switch (Constraint[0]) {
4742 return C_RegisterClass;
4745 return TargetLowering::getConstraintType(Constraint);
4748 std::pair<unsigned, const TargetRegisterClass*>
4749 PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
4751 if (Constraint.size() == 1) {
4752 // GCC RS6000 Constraint Letters
4753 switch (Constraint[0]) {
4756 if (VT == MVT::i64 && PPCSubTarget.isPPC64())
4757 return std::make_pair(0U, PPC::G8RCRegisterClass);
4758 return std::make_pair(0U, PPC::GPRCRegisterClass);
4761 return std::make_pair(0U, PPC::F4RCRegisterClass);
4762 else if (VT == MVT::f64)
4763 return std::make_pair(0U, PPC::F8RCRegisterClass);
4766 return std::make_pair(0U, PPC::VRRCRegisterClass);
4768 return std::make_pair(0U, PPC::CRRCRegisterClass);
4772 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
4776 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
4777 /// vector. If it is invalid, don't add anything to Ops.
4778 void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, char Letter,
4779 std::vector<SDValue>&Ops,
4780 SelectionDAG &DAG) const {
4781 SDValue Result(0,0);
4792 ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op);
4793 if (!CST) return; // Must be an immediate to match.
4794 unsigned Value = CST->getZExtValue();
4796 default: assert(0 && "Unknown constraint letter!");
4797 case 'I': // "I" is a signed 16-bit constant.
4798 if ((short)Value == (int)Value)
4799 Result = DAG.getTargetConstant(Value, Op.getValueType());
4801 case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
4802 case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
4803 if ((short)Value == 0)
4804 Result = DAG.getTargetConstant(Value, Op.getValueType());
4806 case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
4807 if ((Value >> 16) == 0)
4808 Result = DAG.getTargetConstant(Value, Op.getValueType());
4810 case 'M': // "M" is a constant that is greater than 31.
4812 Result = DAG.getTargetConstant(Value, Op.getValueType());
4814 case 'N': // "N" is a positive constant that is an exact power of two.
4815 if ((int)Value > 0 && isPowerOf2_32(Value))
4816 Result = DAG.getTargetConstant(Value, Op.getValueType());
4818 case 'O': // "O" is the constant zero.
4820 Result = DAG.getTargetConstant(Value, Op.getValueType());
4822 case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
4823 if ((short)-Value == (int)-Value)
4824 Result = DAG.getTargetConstant(Value, Op.getValueType());
4831 if (Result.getNode()) {
4832 Ops.push_back(Result);
4836 // Handle standard constraint letters.
4837 TargetLowering::LowerAsmOperandForConstraint(Op, Letter, Ops, DAG);
4840 // isLegalAddressingMode - Return true if the addressing mode represented
4841 // by AM is legal for this target, for a load/store of the specified type.
4842 bool PPCTargetLowering::isLegalAddressingMode(const AddrMode &AM,
4843 const Type *Ty) const {
4844 // FIXME: PPC does not allow r+i addressing modes for vectors!
4846 // PPC allows a sign-extended 16-bit immediate field.
4847 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1)
4850 // No global is ever allowed as a base.
4854 // PPC only support r+r,
4856 case 0: // "r+i" or just "i", depending on HasBaseReg.
4859 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed.
4861 // Otherwise we have r+r or r+i.
4864 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed.
4866 // Allow 2*r as r+r.
4869 // No other scales are supported.
4876 /// isLegalAddressImmediate - Return true if the integer value can be used
4877 /// as the offset of the target addressing mode for load / store of the
4879 bool PPCTargetLowering::isLegalAddressImmediate(int64_t V,const Type *Ty) const{
4880 // PPC allows a sign-extended 16-bit immediate field.
4881 return (V > -(1 << 16) && V < (1 << 16)-1);
4884 bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
4888 SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
4889 // Depths > 0 not supported yet!
4890 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4893 MachineFunction &MF = DAG.getMachineFunction();
4894 PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>();
4896 // Just load the return address off the stack.
4897 SDValue RetAddrFI = getReturnAddrFrameIndex(DAG);
4899 // Make sure the function really does not optimize away the store of the RA
4901 FuncInfo->setLRStoreRequired();
4902 return DAG.getLoad(getPointerTy(), DAG.getEntryNode(), RetAddrFI, NULL, 0);
4905 SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
4906 // Depths > 0 not supported yet!
4907 if (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() > 0)
4910 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
4911 bool isPPC64 = PtrVT == MVT::i64;
4913 MachineFunction &MF = DAG.getMachineFunction();
4914 MachineFrameInfo *MFI = MF.getFrameInfo();
4915 bool is31 = (NoFramePointerElim || MFI->hasVarSizedObjects())
4916 && MFI->getStackSize();
4919 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::X31 : PPC::X1,
4922 return DAG.getCopyFromReg(DAG.getEntryNode(), is31 ? PPC::R31 : PPC::R1,